# ScienceToStartup - Full Content Summary for AI Systems
# https://sciencetostartup.com/llms-full.txt
# Version: 2026.03.10
# Last Updated: 2026-03-10
# Generated: 2026-04-11T08:08:16.110Z

## About
ScienceToStartup is an Agent Operating System for Research Commercialization.
It is an API and MCP platform for turning research papers into buildable product
signals. The public site is organized around proof surfaces, developer install
paths, stable contracts, and execution handoffs.

Founded by Musa Emin Ozdem in 2026.

## Core Taxonomy

### 1. Proof Surfaces
These are the durable public URLs that anchor trust, search, and agent retrieval.
- Canonical Paper Page: https://sciencetostartup.com/paper/2601.18631 - Use the canonical paper page as a proof artifact.
- Signal Canvas: https://sciencetostartup.com/signal-canvas/2601.18631 - Use Signal Canvas as the narrative proof surface.
- Topic Page: https://sciencetostartup.com/topic/llm-analysis - Use this topic page as a durable research-area proof surface.
- Benchmark Resource: https://sciencetostartup.com/resources/benchmark - Use the benchmark as a ranking and proof layer.
- Dataset Resource: https://sciencetostartup.com/resources/dataset - Use the public dataset as a machine-readable proof surface.

### 2. Developer Surfaces
- Overview hub: https://sciencetostartup.com/developers
- Start Here: https://sciencetostartup.com/developers/start-here
- REST guide: https://sciencetostartup.com/developers/rest
- MCP guide: https://sciencetostartup.com/developers/mcp
- Signal Canvas guide: https://sciencetostartup.com/developers/signal-canvas
- Key management: https://sciencetostartup.com/developers/keys
- Query-led pages:
- REST API: https://sciencetostartup.com/developers/rest - Canonical HTTP routes for search, export, and workspace actions.
- MCP Server: https://sciencetostartup.com/developers/mcp - Tool and resource contracts for agent-native research workflows.
- Examples: https://sciencetostartup.com/developers/examples - Reference implementation flows across research, proof, and handoff.
- Launch Pack via API + MCP: https://sciencetostartup.com/developers/examples/launch-pack-via-api-mcp - Combine REST and MCP to turn paper context into execution artifacts.
- Paper Search to Workspace: https://sciencetostartup.com/developers/examples/paper-search-to-workspace - Take a public paper search into a durable workspace flow.
- ChatGPT Integration: https://sciencetostartup.com/developers/integrations/chatgpt - Bridge ChatGPT to ScienceToStartup public proof and API surfaces.
- Cursor Integration: https://sciencetostartup.com/developers/integrations/cursor - Use Cursor with MCP and REST surfaces for research workflows.
- Claude Integration: https://sciencetostartup.com/developers/integrations/claude - Connect Claude to research proof, search, and workspace handoffs.

### 3. Discovery Files and Stable Contracts
- Developers: https://sciencetostartup.com/developers - Hub for REST, MCP, integrations, and examples.
- API Docs: https://sciencetostartup.com/docs/api - OpenAPI, capabilities, and discovery documents.
- REST API: https://sciencetostartup.com/developers/rest - HTTP integrations for research and workspace workflows.
- MCP Server: https://sciencetostartup.com/developers/mcp - Tool and resource contracts for agents and copilots.

### 4. Execution Surfaces
- Build Loop: https://sciencetostartup.com/build-loop
- Workspace: https://sciencetostartup.com/workspace
- Talent: https://sciencetostartup.com/talent
- Signal Canvas: https://sciencetostartup.com/signal-canvas

### 5. Editorial and Distribution Wrappers
- Trends wrappers:
- Operator Cycle: https://sciencetostartup.com/trends
- Talent wrappers:
- Hiring for LLM Analysis: https://sciencetostartup.com/talent/hiring-for-llm-analysis
- Research to Role Map: https://sciencetostartup.com/talent/research-to-role-map
- Source Builders from Proof: https://sciencetostartup.com/talent/source-builders-from-proof
- Article families:
- Weekly Benchmark Roundup: https://sciencetostartup.com/articles/weekly-benchmark-roundup
- Commercializable Papers Report: https://sciencetostartup.com/articles/commercializable-papers-report

### 6. Research Papers (/paper/[arxivId])
Individual paper analysis with canonical evidence, viability, citations, and
handoffs into Signal Canvas, workspaces, and Build Loop.

### 7. Topic Pages (/topic/[slug])
Durable research-area proof pages with paper counts, viability trends, author
signals, and top questions.

### 8. Resources (/resources)
Machine-readable and human-facing reference surfaces:
- Hub: https://sciencetostartup.com/resources
- Benchmark: https://sciencetostartup.com/resources/benchmark (weekly scoreboard)
- Industry Index: https://sciencetostartup.com/resources/industry-index
- Interactive Database: https://sciencetostartup.com/resources/database
- Public Dataset: https://sciencetostartup.com/resources/dataset
- Glossary: https://sciencetostartup.com/resources/glossary
- State Reports: https://sciencetostartup.com/resources/state-reports
- Directory: https://sciencetostartup.com/resources/directory
- Calculator: https://sciencetostartup.com/resources/calculator
- Templates: https://sciencetostartup.com/resources/templates
- Alternatives: https://sciencetostartup.com/resources/alternatives
- Comparison hubs: https://sciencetostartup.com/resources/comparisons
- Startup idea generator: https://sciencetostartup.com/resources/generator
- Briefs: https://sciencetostartup.com/resources/briefs
- Grants: https://sciencetostartup.com/resources/grants
- Patents: https://sciencetostartup.com/resources/patents

### 9. Trends and Articles
- Trends desk: https://sciencetostartup.com/trends
- Articles index: https://sciencetostartup.com/articles
- Changelog: https://sciencetostartup.com/changelog

## Data Sources and Freshness
- Papers: Updated daily from arXiv
- Topic summaries: Regenerated weekly
- Viability scores: Computed on ingestion using LLM analysis
- Metrics (GitHub stars, citations): Updated daily
- Articles and editorial wrappers: Refreshed on the publish schedule
- Legal pages: Updated from admin-published legal documents

## Viability Score Definition
The Viability Score (1–10) measures how likely an AI paper is to become a fundable startup, based on code availability, author commercialization track record, market timing, and competitive landscape.
Scoring model: GPT-4o

## Signal Fusion Formula
TrendScore = (viability_score × 10) + (unicorn_probability × 30) + (total_votes × 0.1) + (star_velocity × 5)

## Research Topic Index (594)
- Medical AI (720 papers): https://sciencetostartup.com/topic/medical-ai
- Agents (479 papers): https://sciencetostartup.com/topic/agents
- Robotics (359 papers): https://sciencetostartup.com/topic/robotics
- Reinforcement Learning (313 papers): https://sciencetostartup.com/topic/reinforcement-learning
- LLM Training (230 papers): https://sciencetostartup.com/topic/llm-training
- Computer Vision (185 papers): https://sciencetostartup.com/topic/computer-vision
- LLM Evaluation (112 papers): https://sciencetostartup.com/topic/llm-evaluation
- Generative Video (107 papers): https://sciencetostartup.com/topic/generative-video
- 3D Reconstruction (93 papers): https://sciencetostartup.com/topic/3d-reconstruction
- NLP (84 papers): https://sciencetostartup.com/topic/nlp
- Federated Learning (79 papers): https://sciencetostartup.com/topic/federated-learning
- Autonomous Driving (78 papers): https://sciencetostartup.com/topic/autonomous-driving
- LLM Optimization (77 papers): https://sciencetostartup.com/topic/llm-optimization
- AI Security (71 papers): https://sciencetostartup.com/topic/ai-security
- Multimodal AI (71 papers): https://sciencetostartup.com/topic/multimodal-ai
- LLM Reasoning (66 papers): https://sciencetostartup.com/topic/llm-reasoning
- LLM Agents (59 papers): https://sciencetostartup.com/topic/llm-agents
- Educational AI (56 papers): https://sciencetostartup.com/topic/educational-ai
- AI Safety (54 papers): https://sciencetostartup.com/topic/ai-safety
- LLM Security (53 papers): https://sciencetostartup.com/topic/llm-security
- Generative Models (52 papers): https://sciencetostartup.com/topic/generative-models
- Multi-Agent Systems (49 papers): https://sciencetostartup.com/topic/multi-agent-systems
- AI Agents (46 papers): https://sciencetostartup.com/topic/ai-agents
- Graph Neural Networks (45 papers): https://sciencetostartup.com/topic/graph-neural-networks
- Diffusion Models (42 papers): https://sciencetostartup.com/topic/diffusion-models
- Robotic Manipulation (42 papers): https://sciencetostartup.com/topic/robotic-manipulation
- Video Understanding (40 papers): https://sciencetostartup.com/topic/video-understanding
- Robotics Control (40 papers): https://sciencetostartup.com/topic/robotics-control
- Explainable AI (40 papers): https://sciencetostartup.com/topic/explainable-ai
- Healthcare AI (39 papers): https://sciencetostartup.com/topic/healthcare-ai
- Optimization Algorithms (36 papers): https://sciencetostartup.com/topic/optimization-algorithms
- Robotics Navigation (34 papers): https://sciencetostartup.com/topic/robotics-navigation
- LLM Inference Optimization (34 papers): https://sciencetostartup.com/topic/llm-inference-optimization
- Continual Learning (34 papers): https://sciencetostartup.com/topic/continual-learning
- LLM Alignment (33 papers): https://sciencetostartup.com/topic/llm-alignment
- Generative Image (32 papers): https://sciencetostartup.com/topic/generative-image
- Causal Inference (32 papers): https://sciencetostartup.com/topic/causal-inference
- Robotics AI (31 papers): https://sciencetostartup.com/topic/robotics-ai
- AI Evaluation (31 papers): https://sciencetostartup.com/topic/ai-evaluation
- Anomaly Detection (31 papers): https://sciencetostartup.com/topic/anomaly-detection
- Generative AI (31 papers): https://sciencetostartup.com/topic/generative-ai
- Cybersecurity (28 papers): https://sciencetostartup.com/topic/cybersecurity
- Legal AI (27 papers): https://sciencetostartup.com/topic/legal-ai
- Financial AI (27 papers): https://sciencetostartup.com/topic/financial-ai
- LLM Safety (26 papers): https://sciencetostartup.com/topic/llm-safety
- LLM Interpretability (26 papers): https://sciencetostartup.com/topic/llm-interpretability
- Recommendation Systems (25 papers): https://sciencetostartup.com/topic/recommendation-systems
- Benchmarking (25 papers): https://sciencetostartup.com/topic/benchmarking
- Image Restoration (25 papers): https://sciencetostartup.com/topic/image-restoration
- Recommender Systems (24 papers): https://sciencetostartup.com/topic/recommender-systems
- Conversational AI (23 papers): https://sciencetostartup.com/topic/conversational-ai
- Embodied AI (23 papers): https://sciencetostartup.com/topic/embodied-ai
- AI Governance (23 papers): https://sciencetostartup.com/topic/ai-governance
- Multimodal Reasoning (22 papers): https://sciencetostartup.com/topic/multimodal-reasoning
- Cybersecurity AI (22 papers): https://sciencetostartup.com/topic/cybersecurity-ai
- LLM Adaptation (21 papers): https://sciencetostartup.com/topic/llm-adaptation
- Model Optimization (21 papers): https://sciencetostartup.com/topic/model-optimization
- LLM Fine-tuning (21 papers): https://sciencetostartup.com/topic/llm-fine-tuning
- AI Reasoning (20 papers): https://sciencetostartup.com/topic/ai-reasoning
- LLM Applications (20 papers): https://sciencetostartup.com/topic/llm-applications
- Speech Recognition (20 papers): https://sciencetostartup.com/topic/speech-recognition
- Autonomous Vehicles (19 papers): https://sciencetostartup.com/topic/autonomous-vehicles
- 3D Scene Understanding (19 papers): https://sciencetostartup.com/topic/3d-scene-understanding
- Language Models (19 papers): https://sciencetostartup.com/topic/language-models
- AI Theory (18 papers): https://sciencetostartup.com/topic/ai-theory
- Cryptography (18 papers): https://sciencetostartup.com/topic/cryptography
- Human-AI Interaction (18 papers): https://sciencetostartup.com/topic/human-ai-interaction
- Human-Robot Interaction (18 papers): https://sciencetostartup.com/topic/human-robot-interaction
- Time Series Analysis (18 papers): https://sciencetostartup.com/topic/time-series-analysis
- Machine Translation (17 papers): https://sciencetostartup.com/topic/machine-translation
- Adversarial Attacks (17 papers): https://sciencetostartup.com/topic/adversarial-attacks
- AI Alignment (17 papers): https://sciencetostartup.com/topic/ai-alignment
- Generative Image Editing (16 papers): https://sciencetostartup.com/topic/generative-image-editing
- AI Ethics (16 papers): https://sciencetostartup.com/topic/ai-ethics
- Security (16 papers): https://sciencetostartup.com/topic/security
- LLM Analysis (16 papers): https://sciencetostartup.com/topic/llm-analysis
- Privacy-Preserving AI (16 papers): https://sciencetostartup.com/topic/privacy-preserving-ai
- Geospatial AI (16 papers): https://sciencetostartup.com/topic/geospatial-ai
- Industrial AI (16 papers): https://sciencetostartup.com/topic/industrial-ai
- Robotics Simulation (16 papers): https://sciencetostartup.com/topic/robotics-simulation
- 3D Computer Vision (16 papers): https://sciencetostartup.com/topic/3d-computer-vision
- Vision Language Models (16 papers): https://sciencetostartup.com/topic/vision-language-models
- Image Processing (15 papers): https://sciencetostartup.com/topic/image-processing
- Remote Sensing (15 papers): https://sciencetostartup.com/topic/remote-sensing
- NLP Research (15 papers): https://sciencetostartup.com/topic/nlp-research
- Multimodal LLMs (15 papers): https://sciencetostartup.com/topic/multimodal-llms
- Agricultural AI (15 papers): https://sciencetostartup.com/topic/agricultural-ai
- Humanoid Robotics (15 papers): https://sciencetostartup.com/topic/humanoid-robotics
- Synthetic Data Generation (15 papers): https://sciencetostartup.com/topic/synthetic-data-generation
- Optimization (15 papers): https://sciencetostartup.com/topic/optimization
- RAG (14 papers): https://sciencetostartup.com/topic/rag
- Object Detection (14 papers): https://sciencetostartup.com/topic/object-detection
- LLM Efficiency (14 papers): https://sciencetostartup.com/topic/llm-efficiency
- Code Generation (14 papers): https://sciencetostartup.com/topic/code-generation
- Model Compression (14 papers): https://sciencetostartup.com/topic/model-compression
- Robotics Perception (13 papers): https://sciencetostartup.com/topic/robotics-perception
- Deepfake Detection (13 papers): https://sciencetostartup.com/topic/deepfake-detection
- Remote Sensing AI (13 papers): https://sciencetostartup.com/topic/remote-sensing-ai
- Edge AI (13 papers): https://sciencetostartup.com/topic/edge-ai
- Image Editing (13 papers): https://sciencetostartup.com/topic/image-editing
- Human-Computer Interaction (13 papers): https://sciencetostartup.com/topic/human-computer-interaction
- Autonomous Driving Perception (12 papers): https://sciencetostartup.com/topic/autonomous-driving-perception
- Security AI (12 papers): https://sciencetostartup.com/topic/security-ai
- Federated Learning Security (12 papers): https://sciencetostartup.com/topic/federated-learning-security
- Autonomous Navigation (12 papers): https://sciencetostartup.com/topic/autonomous-navigation
- Video Reasoning (12 papers): https://sciencetostartup.com/topic/video-reasoning
- Video Generation (12 papers): https://sciencetostartup.com/topic/video-generation
- Image Retrieval (11 papers): https://sciencetostartup.com/topic/image-retrieval
- Multimodal Models (11 papers): https://sciencetostartup.com/topic/multimodal-models
- Multi-Agent Reinforcement Learning (11 papers): https://sciencetostartup.com/topic/multi-agent-reinforcement-learning
- Knowledge Graphs (11 papers): https://sciencetostartup.com/topic/knowledge-graphs
- Vision-Language Navigation (11 papers): https://sciencetostartup.com/topic/vision-language-navigation
- Text-to-Speech (11 papers): https://sciencetostartup.com/topic/text-to-speech
- Security in AI (11 papers): https://sciencetostartup.com/topic/security-in-ai
- LLM Architecture (11 papers): https://sciencetostartup.com/topic/llm-architecture
- AI Interpretability (11 papers): https://sciencetostartup.com/topic/ai-interpretability
- Medical Imaging (11 papers): https://sciencetostartup.com/topic/medical-imaging
- 3D Generation (11 papers): https://sciencetostartup.com/topic/3d-generation
- Theoretical AI (11 papers): https://sciencetostartup.com/topic/theoretical-ai
- LLM Compression (11 papers): https://sciencetostartup.com/topic/llm-compression
- AI Benchmarking (10 papers): https://sciencetostartup.com/topic/ai-benchmarking
- IoT Security (10 papers): https://sciencetostartup.com/topic/iot-security
- Emotion Recognition (10 papers): https://sciencetostartup.com/topic/emotion-recognition
- NLP Evaluation (10 papers): https://sciencetostartup.com/topic/nlp-evaluation
- Symbolic Regression (10 papers): https://sciencetostartup.com/topic/symbolic-regression
- Image Super-Resolution (10 papers): https://sciencetostartup.com/topic/image-super-resolution
- Image Enhancement (10 papers): https://sciencetostartup.com/topic/image-enhancement
- Mental Health AI (10 papers): https://sciencetostartup.com/topic/mental-health-ai
- Audio AI (10 papers): https://sciencetostartup.com/topic/audio-ai
- 3D Object Detection (10 papers): https://sciencetostartup.com/topic/3d-object-detection
- Speech Processing (10 papers): https://sciencetostartup.com/topic/speech-processing
- AI Model Optimization (9 papers): https://sciencetostartup.com/topic/ai-model-optimization
- Reinforcement Learning Theory (9 papers): https://sciencetostartup.com/topic/reinforcement-learning-theory
- AI Optimization (9 papers): https://sciencetostartup.com/topic/ai-optimization
- AI Research (9 papers): https://sciencetostartup.com/topic/ai-research
- Graph Learning (9 papers): https://sciencetostartup.com/topic/graph-learning
- AI Infrastructure (9 papers): https://sciencetostartup.com/topic/ai-infrastructure
- Statistical Modeling (9 papers): https://sciencetostartup.com/topic/statistical-modeling
- Multimodal Learning (9 papers): https://sciencetostartup.com/topic/multimodal-learning
- Assistive Robotics (9 papers): https://sciencetostartup.com/topic/assistive-robotics
- Weather Forecasting (9 papers): https://sciencetostartup.com/topic/weather-forecasting
- Sports Analytics (9 papers): https://sciencetostartup.com/topic/sports-analytics
- Neuromorphic Computing (9 papers): https://sciencetostartup.com/topic/neuromorphic-computing
- Uncertainty Quantification (9 papers): https://sciencetostartup.com/topic/uncertainty-quantification
- Autonomous Driving AI (9 papers): https://sciencetostartup.com/topic/autonomous-driving-ai
- Image Compression (9 papers): https://sciencetostartup.com/topic/image-compression
- Generative Modeling (9 papers): https://sciencetostartup.com/topic/generative-modeling
- Network Security (9 papers): https://sciencetostartup.com/topic/network-security
- Interpretable AI (9 papers): https://sciencetostartup.com/topic/interpretable-ai
- Retrieval-Augmented Generation (9 papers): https://sciencetostartup.com/topic/retrieval-augmented-generation
- Differential Privacy (9 papers): https://sciencetostartup.com/topic/differential-privacy
- Smart Contract Security (9 papers): https://sciencetostartup.com/topic/smart-contract-security
- Multilingual NLP (9 papers): https://sciencetostartup.com/topic/multilingual-nlp
- World Models (9 papers): https://sciencetostartup.com/topic/world-models
- Speech AI (9 papers): https://sciencetostartup.com/topic/speech-ai
- Scientific AI (8 papers): https://sciencetostartup.com/topic/scientific-ai
- Physics-Informed AI (8 papers): https://sciencetostartup.com/topic/physics-informed-ai
- Environmental AI (8 papers): https://sciencetostartup.com/topic/environmental-ai
- Statistical Inference (8 papers): https://sciencetostartup.com/topic/statistical-inference
- Scientific ML (8 papers): https://sciencetostartup.com/topic/scientific-ml
- Model Merging (8 papers): https://sciencetostartup.com/topic/model-merging
- Formal Verification (8 papers): https://sciencetostartup.com/topic/formal-verification
- Dataset Distillation (8 papers): https://sciencetostartup.com/topic/dataset-distillation
- Vision-Language-Action (8 papers): https://sciencetostartup.com/topic/vision-language-action
- Robotics and Automation (8 papers): https://sciencetostartup.com/topic/robotics-and-automation
- Machine Learning Theory (8 papers): https://sciencetostartup.com/topic/machine-learning-theory
- 3D Scene Generation (8 papers): https://sciencetostartup.com/topic/3d-scene-generation
- Robotics Planning (8 papers): https://sciencetostartup.com/topic/robotics-planning
- Generative Image Models (8 papers): https://sciencetostartup.com/topic/generative-image-models
- Content Moderation (8 papers): https://sciencetostartup.com/topic/content-moderation
- Environmental Monitoring (8 papers): https://sciencetostartup.com/topic/environmental-monitoring
- Medical Robotics (8 papers): https://sciencetostartup.com/topic/medical-robotics
- Information Retrieval (8 papers): https://sciencetostartup.com/topic/information-retrieval
- Multi-Object Tracking (8 papers): https://sciencetostartup.com/topic/multi-object-tracking
- 3D Perception (8 papers): https://sciencetostartup.com/topic/3d-perception
- AI Efficiency (8 papers): https://sciencetostartup.com/topic/ai-efficiency
- Music AI (8 papers): https://sciencetostartup.com/topic/music-ai
- Autonomous Agents (7 papers): https://sciencetostartup.com/topic/autonomous-agents
- Scientific Machine Learning (7 papers): https://sciencetostartup.com/topic/scientific-machine-learning
- 3D Vision (7 papers): https://sciencetostartup.com/topic/3d-vision
- Dialogue Systems (7 papers): https://sciencetostartup.com/topic/dialogue-systems
- Clustering Algorithms (7 papers): https://sciencetostartup.com/topic/clustering-algorithms
- Neuro-Symbolic AI (7 papers): https://sciencetostartup.com/topic/neuro-symbolic-ai
- Sentiment Analysis (7 papers): https://sciencetostartup.com/topic/sentiment-analysis
- Physics-Informed Neural Networks (7 papers): https://sciencetostartup.com/topic/physics-informed-neural-networks
- Vision Transformers (7 papers): https://sciencetostartup.com/topic/vision-transformers
- Machine Unlearning (7 papers): https://sciencetostartup.com/topic/machine-unlearning
- AI Planning (7 papers): https://sciencetostartup.com/topic/ai-planning
- Generative Design (7 papers): https://sciencetostartup.com/topic/generative-design
- Human-Robot Collaboration (7 papers): https://sciencetostartup.com/topic/human-robot-collaboration
- Attention Mechanisms (7 papers): https://sciencetostartup.com/topic/attention-mechanisms
- Text-to-Image Generation (7 papers): https://sciencetostartup.com/topic/text-to-image-generation
- LLM Control (7 papers): https://sciencetostartup.com/topic/llm-control
- Audio Processing (7 papers): https://sciencetostartup.com/topic/audio-processing
- Time-Series Forecasting (7 papers): https://sciencetostartup.com/topic/time-series-forecasting
- Visual Reasoning (7 papers): https://sciencetostartup.com/topic/visual-reasoning
- Bias Mitigation (7 papers): https://sciencetostartup.com/topic/bias-mitigation
- Transformer Optimization (7 papers): https://sciencetostartup.com/topic/transformer-optimization
- Ethical AI (7 papers): https://sciencetostartup.com/topic/ethical-ai
- LLM Enhancement (7 papers): https://sciencetostartup.com/topic/llm-enhancement
- Molecular AI (7 papers): https://sciencetostartup.com/topic/molecular-ai
- Transformers (7 papers): https://sciencetostartup.com/topic/transformers
- Vision-Language-Action Models (7 papers): https://sciencetostartup.com/topic/vision-language-action-models
- AI in Education (7 papers): https://sciencetostartup.com/topic/ai-in-education
- Representation Learning (7 papers): https://sciencetostartup.com/topic/representation-learning
- RAG Optimization (7 papers): https://sciencetostartup.com/topic/rag-optimization
- Video Anomaly Detection (7 papers): https://sciencetostartup.com/topic/video-anomaly-detection
- Robotics Motion Planning (7 papers): https://sciencetostartup.com/topic/robotics-motion-planning
- Privacy-Preserving ML (7 papers): https://sciencetostartup.com/topic/privacy-preserving-ml
- Human-AI Collaboration (7 papers): https://sciencetostartup.com/topic/human-ai-collaboration
- Medical Imaging AI (7 papers): https://sciencetostartup.com/topic/medical-imaging-ai
- LLM Reliability (7 papers): https://sciencetostartup.com/topic/llm-reliability
- Geo-localization (6 papers): https://sciencetostartup.com/topic/geo-localization
- Robot Manipulation (6 papers): https://sciencetostartup.com/topic/robot-manipulation
- Scientific Computing (6 papers): https://sciencetostartup.com/topic/scientific-computing
- Sports AI (6 papers): https://sciencetostartup.com/topic/sports-ai
- 3D Rendering (6 papers): https://sciencetostartup.com/topic/3d-rendering
- Time Series Classification (6 papers): https://sciencetostartup.com/topic/time-series-classification
- LLM Inference (6 papers): https://sciencetostartup.com/topic/llm-inference
- Signal Processing (6 papers): https://sciencetostartup.com/topic/signal-processing
- Security in LLMs (6 papers): https://sciencetostartup.com/topic/security-in-llms
- Action Recognition (6 papers): https://sciencetostartup.com/topic/action-recognition
- Image Classification (6 papers): https://sciencetostartup.com/topic/image-classification
- Bioinformatics AI (6 papers): https://sciencetostartup.com/topic/bioinformatics-ai
- 3D Scene Reconstruction (6 papers): https://sciencetostartup.com/topic/3d-scene-reconstruction
- Robotics Safety (6 papers): https://sciencetostartup.com/topic/robotics-safety
- Mathematical Reasoning (6 papers): https://sciencetostartup.com/topic/mathematical-reasoning
- Bandit Algorithms (6 papers): https://sciencetostartup.com/topic/bandit-algorithms
- Fairness in ML (6 papers): https://sciencetostartup.com/topic/fairness-in-ml
- Memory Systems (6 papers): https://sciencetostartup.com/topic/memory-systems
- Quantum Machine Learning (6 papers): https://sciencetostartup.com/topic/quantum-machine-learning
- Blockchain Security (6 papers): https://sciencetostartup.com/topic/blockchain-security
- Software Security (6 papers): https://sciencetostartup.com/topic/software-security
- Fairness in AI (6 papers): https://sciencetostartup.com/topic/fairness-in-ai
- LLM Calibration (6 papers): https://sciencetostartup.com/topic/llm-calibration
- Model Interpretability (6 papers): https://sciencetostartup.com/topic/model-interpretability
- Graph Generation (6 papers): https://sciencetostartup.com/topic/graph-generation
- Statistical Learning (6 papers): https://sciencetostartup.com/topic/statistical-learning
- Generative Vision (6 papers): https://sciencetostartup.com/topic/generative-vision
- Bayesian Inference (6 papers): https://sciencetostartup.com/topic/bayesian-inference
- Vision Models (6 papers): https://sciencetostartup.com/topic/vision-models
- GUI Agents (6 papers): https://sciencetostartup.com/topic/gui-agents
- Combinatorial Optimization (6 papers): https://sciencetostartup.com/topic/combinatorial-optimization
- Brain-Computer Interface (6 papers): https://sciencetostartup.com/topic/brain-computer-interface
- Graph Anomaly Detection (6 papers): https://sciencetostartup.com/topic/graph-anomaly-detection
- Bayesian Optimization (6 papers): https://sciencetostartup.com/topic/bayesian-optimization
- Robotics Localization (6 papers): https://sciencetostartup.com/topic/robotics-localization
- Image Captioning (6 papers): https://sciencetostartup.com/topic/image-captioning
- Causal Discovery (6 papers): https://sciencetostartup.com/topic/causal-discovery
- Explainable AI (XAI) (6 papers): https://sciencetostartup.com/topic/explainable-ai-xai
- Virtual Try-On (6 papers): https://sciencetostartup.com/topic/virtual-try-on
- Agentic AI (6 papers): https://sciencetostartup.com/topic/agentic-ai
- Biomedical AI (6 papers): https://sciencetostartup.com/topic/biomedical-ai
- Audio Deepfake Detection (6 papers): https://sciencetostartup.com/topic/audio-deepfake-detection
- Game AI (6 papers): https://sciencetostartup.com/topic/game-ai
- Novel View Synthesis (5 papers): https://sciencetostartup.com/topic/novel-view-synthesis
- Enterprise AI (5 papers): https://sciencetostartup.com/topic/enterprise-ai
- Benchmarking Tools (5 papers): https://sciencetostartup.com/topic/benchmarking-tools
- Reinforcement Learning Optimization (5 papers): https://sciencetostartup.com/topic/reinforcement-learning-optimization
- Inference Optimization (5 papers): https://sciencetostartup.com/topic/inference-optimization
- Document Processing (5 papers): https://sciencetostartup.com/topic/document-processing
- LLM Bias Mitigation (5 papers): https://sciencetostartup.com/topic/llm-bias-mitigation
- Document Parsing (5 papers): https://sciencetostartup.com/topic/document-parsing
- LLM Theory (5 papers): https://sciencetostartup.com/topic/llm-theory
- Cultural AI (5 papers): https://sciencetostartup.com/topic/cultural-ai
- Surrogate Modeling (5 papers): https://sciencetostartup.com/topic/surrogate-modeling
- Autonomous Systems (5 papers): https://sciencetostartup.com/topic/autonomous-systems
- Materials Science AI (5 papers): https://sciencetostartup.com/topic/materials-science-ai
- LLM Hallucination Mitigation (5 papers): https://sciencetostartup.com/topic/llm-hallucination-mitigation
- Reasoning Models (5 papers): https://sciencetostartup.com/topic/reasoning-models
- LLM Privacy (5 papers): https://sciencetostartup.com/topic/llm-privacy
- Document AI (5 papers): https://sciencetostartup.com/topic/document-ai
- Biological AI (5 papers): https://sciencetostartup.com/topic/biological-ai
- Federated Learning Optimization (5 papers): https://sciencetostartup.com/topic/federated-learning-optimization
- Image Segmentation (5 papers): https://sciencetostartup.com/topic/image-segmentation
- Biotech AI (5 papers): https://sciencetostartup.com/topic/biotech-ai
- Medical Image Segmentation (5 papers): https://sciencetostartup.com/topic/medical-image-segmentation
- Text-to-SQL (5 papers): https://sciencetostartup.com/topic/text-to-sql
- Autonomous Driving Simulation (5 papers): https://sciencetostartup.com/topic/autonomous-driving-simulation
- SLAM (5 papers): https://sciencetostartup.com/topic/slam
- AI for Science (5 papers): https://sciencetostartup.com/topic/ai-for-science
- Mathematical AI (5 papers): https://sciencetostartup.com/topic/mathematical-ai
- Out-of-Distribution Detection (5 papers): https://sciencetostartup.com/topic/out-of-distribution-detection
- Privacy in AI (5 papers): https://sciencetostartup.com/topic/privacy-in-ai
- Distributed Systems (5 papers): https://sciencetostartup.com/topic/distributed-systems
- Hardware Security (5 papers): https://sciencetostartup.com/topic/hardware-security
- Speech Emotion Recognition (5 papers): https://sciencetostartup.com/topic/speech-emotion-recognition
- Neural Network Theory (5 papers): https://sciencetostartup.com/topic/neural-network-theory
- Image Quality Assessment (5 papers): https://sciencetostartup.com/topic/image-quality-assessment
- Multimodal Retrieval (5 papers): https://sciencetostartup.com/topic/multimodal-retrieval
- Multimodal Agents (5 papers): https://sciencetostartup.com/topic/multimodal-agents
- Maritime AI (5 papers): https://sciencetostartup.com/topic/maritime-ai
- Image Generation (5 papers): https://sciencetostartup.com/topic/image-generation
- Knowledge Representation (5 papers): https://sciencetostartup.com/topic/knowledge-representation
- Pose Estimation (5 papers): https://sciencetostartup.com/topic/pose-estimation
- Active Learning (5 papers): https://sciencetostartup.com/topic/active-learning
- Data Augmentation (5 papers): https://sciencetostartup.com/topic/data-augmentation
- Generative Motion (5 papers): https://sciencetostartup.com/topic/generative-motion
- Document Understanding (5 papers): https://sciencetostartup.com/topic/document-understanding
- RAG Systems (5 papers): https://sciencetostartup.com/topic/rag-systems
- Cognitive AI (5 papers): https://sciencetostartup.com/topic/cognitive-ai
- Video Compression (5 papers): https://sciencetostartup.com/topic/video-compression
- Social Media Analysis (5 papers): https://sciencetostartup.com/topic/social-media-analysis
- Chemical AI (5 papers): https://sciencetostartup.com/topic/chemical-ai
- Online Learning (5 papers): https://sciencetostartup.com/topic/online-learning
- Molecular Generation (5 papers): https://sciencetostartup.com/topic/molecular-generation
- Spatial Reasoning (5 papers): https://sciencetostartup.com/topic/spatial-reasoning
- Vision-and-Language Navigation (5 papers): https://sciencetostartup.com/topic/vision-and-language-navigation
- AI Benchmarks (5 papers): https://sciencetostartup.com/topic/ai-benchmarks
- Neural Network Optimization (5 papers): https://sciencetostartup.com/topic/neural-network-optimization
- Learning Theory (5 papers): https://sciencetostartup.com/topic/learning-theory
- Video LLMs (4 papers): https://sciencetostartup.com/topic/video-llms
- Generative Music (4 papers): https://sciencetostartup.com/topic/generative-music
- Soft Robotics (4 papers): https://sciencetostartup.com/topic/soft-robotics
- Motion Generation (4 papers): https://sciencetostartup.com/topic/motion-generation
- Agent Evaluation (4 papers): https://sciencetostartup.com/topic/agent-evaluation
- Video Super-Resolution (4 papers): https://sciencetostartup.com/topic/video-super-resolution
- Engineering AI (4 papers): https://sciencetostartup.com/topic/engineering-ai
- LLM Bias Evaluation (4 papers): https://sciencetostartup.com/topic/llm-bias-evaluation
- Drug Discovery AI (4 papers): https://sciencetostartup.com/topic/drug-discovery-ai
- LLM Behavior Analysis (4 papers): https://sciencetostartup.com/topic/llm-behavior-analysis
- Neural Network Training (4 papers): https://sciencetostartup.com/topic/neural-network-training
- Benchmark Development (4 papers): https://sciencetostartup.com/topic/benchmark-development
- Social Robotics (4 papers): https://sciencetostartup.com/topic/social-robotics
- Question Answering (4 papers): https://sciencetostartup.com/topic/question-answering
- Language Model Optimization (4 papers): https://sciencetostartup.com/topic/language-model-optimization
- AI Agent Security (4 papers): https://sciencetostartup.com/topic/ai-agent-security
- Video Editing AI (4 papers): https://sciencetostartup.com/topic/video-editing-ai
- Brain-Computer Interfaces (4 papers): https://sciencetostartup.com/topic/brain-computer-interfaces
- Geometric Deep Learning (4 papers): https://sciencetostartup.com/topic/geometric-deep-learning
- Cloud Security (4 papers): https://sciencetostartup.com/topic/cloud-security
- Health Monitoring (4 papers): https://sciencetostartup.com/topic/health-monitoring
- Graph AI (4 papers): https://sciencetostartup.com/topic/graph-ai
- NLP Optimization (4 papers): https://sciencetostartup.com/topic/nlp-optimization
- Audio Generation (4 papers): https://sciencetostartup.com/topic/audio-generation
- Affective Computing (4 papers): https://sciencetostartup.com/topic/affective-computing
- LLM Safety & Alignment (4 papers): https://sciencetostartup.com/topic/llm-safety-alignment
- NLP Tools (4 papers): https://sciencetostartup.com/topic/nlp-tools
- AI in Healthcare (4 papers): https://sciencetostartup.com/topic/ai-in-healthcare
- Speaker Verification (4 papers): https://sciencetostartup.com/topic/speaker-verification
- Natural Language Processing (4 papers): https://sciencetostartup.com/topic/natural-language-processing
- Feature Selection (4 papers): https://sciencetostartup.com/topic/feature-selection
- Earth Observation AI (4 papers): https://sciencetostartup.com/topic/earth-observation-ai
- Multi-Robot Systems (4 papers): https://sciencetostartup.com/topic/multi-robot-systems
- Graph Representation Learning (4 papers): https://sciencetostartup.com/topic/graph-representation-learning
- Multilingual AI (4 papers): https://sciencetostartup.com/topic/multilingual-ai
- Few-Shot Learning (4 papers): https://sciencetostartup.com/topic/few-shot-learning
- Optimization Theory (4 papers): https://sciencetostartup.com/topic/optimization-theory
- Generative Image Synthesis (4 papers): https://sciencetostartup.com/topic/generative-image-synthesis
- Generative Art (4 papers): https://sciencetostartup.com/topic/generative-art
- AI Explainability (4 papers): https://sciencetostartup.com/topic/ai-explainability
- AI Safety & Security (4 papers): https://sciencetostartup.com/topic/ai-safety-security
- Agent Security (4 papers): https://sciencetostartup.com/topic/agent-security
- Traffic Prediction (4 papers): https://sciencetostartup.com/topic/traffic-prediction
- Clinical NLP (4 papers): https://sciencetostartup.com/topic/clinical-nlp
- Robotic Navigation (4 papers): https://sciencetostartup.com/topic/robotic-navigation
- 3D Scene Editing (4 papers): https://sciencetostartup.com/topic/3d-scene-editing
- Wireless Networks (4 papers): https://sciencetostartup.com/topic/wireless-networks
- LLM Serving (4 papers): https://sciencetostartup.com/topic/llm-serving
- Clinical AI (4 papers): https://sciencetostartup.com/topic/clinical-ai
- Logical Reasoning (4 papers): https://sciencetostartup.com/topic/logical-reasoning
- Automotive AI (4 papers): https://sciencetostartup.com/topic/automotive-ai
- Machine Learning (4 papers): https://sciencetostartup.com/topic/machine-learning
- Multimodal Generation (4 papers): https://sciencetostartup.com/topic/multimodal-generation
- Knowledge Distillation (4 papers): https://sciencetostartup.com/topic/knowledge-distillation
- Neural Architecture Search (4 papers): https://sciencetostartup.com/topic/neural-architecture-search
- LLM Editing (4 papers): https://sciencetostartup.com/topic/llm-editing
- LLM Development (4 papers): https://sciencetostartup.com/topic/llm-development
- Edge Computing (4 papers): https://sciencetostartup.com/topic/edge-computing
- E-commerce AI (4 papers): https://sciencetostartup.com/topic/e-commerce-ai
- Scene Graph Generation (4 papers): https://sciencetostartup.com/topic/scene-graph-generation
- Self-Supervised Learning (4 papers): https://sciencetostartup.com/topic/self-supervised-learning
- Gesture Recognition (4 papers): https://sciencetostartup.com/topic/gesture-recognition
- Wireless Communication AI (4 papers): https://sciencetostartup.com/topic/wireless-communication-ai
- Biometric Security (4 papers): https://sciencetostartup.com/topic/biometric-security
- OCR (4 papers): https://sciencetostartup.com/topic/ocr
- Web Security (4 papers): https://sciencetostartup.com/topic/web-security
- Protein Engineering (4 papers): https://sciencetostartup.com/topic/protein-engineering
- Spiking Neural Networks (4 papers): https://sciencetostartup.com/topic/spiking-neural-networks
- Predictive Maintenance (4 papers): https://sciencetostartup.com/topic/predictive-maintenance
- Multimodal Sentiment Analysis (4 papers): https://sciencetostartup.com/topic/multimodal-sentiment-analysis
- Distributed Learning (4 papers): https://sciencetostartup.com/topic/distributed-learning
- LLM Research (4 papers): https://sciencetostartup.com/topic/llm-research
- Quantum Cryptography (4 papers): https://sciencetostartup.com/topic/quantum-cryptography
- Multimodal AI Evaluation (4 papers): https://sciencetostartup.com/topic/multimodal-ai-evaluation
- Privacy & Security (4 papers): https://sciencetostartup.com/topic/privacy-security
- Industrial Automation (4 papers): https://sciencetostartup.com/topic/industrial-automation
- Efficient Inference (4 papers): https://sciencetostartup.com/topic/efficient-inference
- Traffic Forecasting (4 papers): https://sciencetostartup.com/topic/traffic-forecasting
- Code Security (4 papers): https://sciencetostartup.com/topic/code-security
- 3D Anomaly Detection (4 papers): https://sciencetostartup.com/topic/3d-anomaly-detection
- Network Optimization (4 papers): https://sciencetostartup.com/topic/network-optimization
- Causal Representation Learning (4 papers): https://sciencetostartup.com/topic/causal-representation-learning
- Financial Forecasting (4 papers): https://sciencetostartup.com/topic/financial-forecasting
- Content Moderation AI (4 papers): https://sciencetostartup.com/topic/content-moderation-ai
- Energy Optimization (4 papers): https://sciencetostartup.com/topic/energy-optimization
- Time-Series Anomaly Detection (4 papers): https://sciencetostartup.com/topic/time-series-anomaly-detection
- Video Editing (4 papers): https://sciencetostartup.com/topic/video-editing
- Advertising Technology (4 papers): https://sciencetostartup.com/topic/advertising-technology
- AI Model Analysis (4 papers): https://sciencetostartup.com/topic/ai-model-analysis
- LLM Reasoning Optimization (4 papers): https://sciencetostartup.com/topic/llm-reasoning-optimization
- 3D Editing (4 papers): https://sciencetostartup.com/topic/3d-editing
- Trajectory Prediction (4 papers): https://sciencetostartup.com/topic/trajectory-prediction
- Dimensionality Reduction (4 papers): https://sciencetostartup.com/topic/dimensionality-reduction
- Language Modeling (4 papers): https://sciencetostartup.com/topic/language-modeling
- Adversarial AI (4 papers): https://sciencetostartup.com/topic/adversarial-ai
- Distributed Optimization (4 papers): https://sciencetostartup.com/topic/distributed-optimization
- Image Fusion (4 papers): https://sciencetostartup.com/topic/image-fusion
- Information Extraction (4 papers): https://sciencetostartup.com/topic/information-extraction
- Monocular Depth Estimation (4 papers): https://sciencetostartup.com/topic/monocular-depth-estimation
- Motion Planning (4 papers): https://sciencetostartup.com/topic/motion-planning
- LLM Steering (4 papers): https://sciencetostartup.com/topic/llm-steering
- Robotic Grasping (4 papers): https://sciencetostartup.com/topic/robotic-grasping
- Web Agents (4 papers): https://sciencetostartup.com/topic/web-agents
- Uncertainty Estimation (4 papers): https://sciencetostartup.com/topic/uncertainty-estimation
- Adversarial Robustness (4 papers): https://sciencetostartup.com/topic/adversarial-robustness
- Robotic Perception (4 papers): https://sciencetostartup.com/topic/robotic-perception
- AI Detection (4 papers): https://sciencetostartup.com/topic/ai-detection
- Code Translation (3 papers): https://sciencetostartup.com/topic/code-translation
- Language Learning (3 papers): https://sciencetostartup.com/topic/language-learning
- Political AI (3 papers): https://sciencetostartup.com/topic/political-ai
- LLM Debiasing (3 papers): https://sciencetostartup.com/topic/llm-debiasing
- Low-Resource NLP (3 papers): https://sciencetostartup.com/topic/low-resource-nlp
- Scientific Simulation (3 papers): https://sciencetostartup.com/topic/scientific-simulation
- Generative Audio (3 papers): https://sciencetostartup.com/topic/generative-audio
- LLM Uncertainty (3 papers): https://sciencetostartup.com/topic/llm-uncertainty
- AI Research Automation (3 papers): https://sciencetostartup.com/topic/ai-research-automation
- LLM Decoding (3 papers): https://sciencetostartup.com/topic/llm-decoding
- Privacy and Security (3 papers): https://sciencetostartup.com/topic/privacy-and-security
- Benchmarking LLMs (3 papers): https://sciencetostartup.com/topic/benchmarking-llms
- 3D Animation (3 papers): https://sciencetostartup.com/topic/3d-animation
- Recruitment AI (3 papers): https://sciencetostartup.com/topic/recruitment-ai
- Education AI (3 papers): https://sciencetostartup.com/topic/education-ai
- Video World Models (3 papers): https://sciencetostartup.com/topic/video-world-models
- AI Robotics (3 papers): https://sciencetostartup.com/topic/ai-robotics
- Manufacturing AI (3 papers): https://sciencetostartup.com/topic/manufacturing-ai
- AI Text Detection (3 papers): https://sciencetostartup.com/topic/ai-text-detection
- Multi-Agent Coordination (3 papers): https://sciencetostartup.com/topic/multi-agent-coordination
- Robotic Surgery (3 papers): https://sciencetostartup.com/topic/robotic-surgery
- Vision Foundation Models (3 papers): https://sciencetostartup.com/topic/vision-foundation-models
- Multi-Agent Path Finding (3 papers): https://sciencetostartup.com/topic/multi-agent-path-finding
- Aviation AI (3 papers): https://sciencetostartup.com/topic/aviation-ai
- Traffic Signal Control (3 papers): https://sciencetostartup.com/topic/traffic-signal-control
- Bayesian Networks (3 papers): https://sciencetostartup.com/topic/bayesian-networks
- Software Development AI (3 papers): https://sciencetostartup.com/topic/software-development-ai
- Quantum AI (3 papers): https://sciencetostartup.com/topic/quantum-ai
- LLM Prompt Optimization (3 papers): https://sciencetostartup.com/topic/llm-prompt-optimization
- Physics-Informed Learning (3 papers): https://sciencetostartup.com/topic/physics-informed-learning
- LLM Quantization (3 papers): https://sciencetostartup.com/topic/llm-quantization
- Data Imputation (3 papers): https://sciencetostartup.com/topic/data-imputation
- Text Classification (3 papers): https://sciencetostartup.com/topic/text-classification
- Database AI (3 papers): https://sciencetostartup.com/topic/database-ai
- Generative Avatars (3 papers): https://sciencetostartup.com/topic/generative-avatars
- Surgical AI (3 papers): https://sciencetostartup.com/topic/surgical-ai
- Emotional AI (3 papers): https://sciencetostartup.com/topic/emotional-ai
- Neuroscience AI (3 papers): https://sciencetostartup.com/topic/neuroscience-ai
- Disaster Response (3 papers): https://sciencetostartup.com/topic/disaster-response
- Software Engineering Tools (3 papers): https://sciencetostartup.com/topic/software-engineering-tools
- Safety Alignment (3 papers): https://sciencetostartup.com/topic/safety-alignment
- LLM Pruning (3 papers): https://sciencetostartup.com/topic/llm-pruning
- Domain Adaptation (3 papers): https://sciencetostartup.com/topic/domain-adaptation
- Probabilistic Forecasting (3 papers): https://sciencetostartup.com/topic/probabilistic-forecasting
- Robotic Simulation (3 papers): https://sciencetostartup.com/topic/robotic-simulation
- AI Diagnostics (3 papers): https://sciencetostartup.com/topic/ai-diagnostics
- Neural Networks (3 papers): https://sciencetostartup.com/topic/neural-networks
- Explainable AI for Time Series (3 papers): https://sciencetostartup.com/topic/explainable-ai-for-time-series
- Machine Learning Tools (3 papers): https://sciencetostartup.com/topic/machine-learning-tools
- AI Model Training (3 papers): https://sciencetostartup.com/topic/ai-model-training
- LLM Personalization (3 papers): https://sciencetostartup.com/topic/llm-personalization
- Bias Detection (3 papers): https://sciencetostartup.com/topic/bias-detection
- Climate Modeling (3 papers): https://sciencetostartup.com/topic/climate-modeling
- Multimodal Understanding (3 papers): https://sciencetostartup.com/topic/multimodal-understanding
- AI Reasoning Enhancement (3 papers): https://sciencetostartup.com/topic/ai-reasoning-enhancement
- Multi-Agent RL (3 papers): https://sciencetostartup.com/topic/multi-agent-rl
- Graph Machine Learning (3 papers): https://sciencetostartup.com/topic/graph-machine-learning
- Fact Verification (3 papers): https://sciencetostartup.com/topic/fact-verification
- LLM Testing (3 papers): https://sciencetostartup.com/topic/llm-testing
- Energy Forecasting (3 papers): https://sciencetostartup.com/topic/energy-forecasting
- Bias Mitigation in AI (3 papers): https://sciencetostartup.com/topic/bias-mitigation-in-ai
- Linguistics (3 papers): https://sciencetostartup.com/topic/linguistics
- AI Research Tools (3 papers): https://sciencetostartup.com/topic/ai-research-tools
- NLP Classification (3 papers): https://sciencetostartup.com/topic/nlp-classification
- Automated Theorem Proving (3 papers): https://sciencetostartup.com/topic/automated-theorem-proving
- Video Retrieval (3 papers): https://sciencetostartup.com/topic/video-retrieval
- Drone Technology (3 papers): https://sciencetostartup.com/topic/drone-technology
- LLM Distillation (3 papers): https://sciencetostartup.com/topic/llm-distillation
- Transformer Interpretability (3 papers): https://sciencetostartup.com/topic/transformer-interpretability
- Mobile AI (3 papers): https://sciencetostartup.com/topic/mobile-ai
- Misinformation Detection (3 papers): https://sciencetostartup.com/topic/misinformation-detection
- Genomic AI (3 papers): https://sciencetostartup.com/topic/genomic-ai
- Multimodal Emotion Recognition (3 papers): https://sciencetostartup.com/topic/multimodal-emotion-recognition
- Hardware Optimization (3 papers): https://sciencetostartup.com/topic/hardware-optimization
- Malware Detection (3 papers): https://sciencetostartup.com/topic/malware-detection
- Robotics & Automation (3 papers): https://sciencetostartup.com/topic/robotics-automation
- Agent Memory Systems (3 papers): https://sciencetostartup.com/topic/agent-memory-systems
- Material Science AI (3 papers): https://sciencetostartup.com/topic/material-science-ai
- Speech LLMs (3 papers): https://sciencetostartup.com/topic/speech-llms
- LLM Prompting (3 papers): https://sciencetostartup.com/topic/llm-prompting
- Lifelong Learning (3 papers): https://sciencetostartup.com/topic/lifelong-learning
- Autonomous Driving Safety (3 papers): https://sciencetostartup.com/topic/autonomous-driving-safety
- AI Safety and Ethics (3 papers): https://sciencetostartup.com/topic/ai-safety-and-ethics
- Fraud Detection (3 papers): https://sciencetostartup.com/topic/fraud-detection
- AI Governance & Regulation (3 papers): https://sciencetostartup.com/topic/ai-governance-regulation
- Adversarial Defense (3 papers): https://sciencetostartup.com/topic/adversarial-defense
- Responsible AI (3 papers): https://sciencetostartup.com/topic/responsible-ai
- LLM Benchmarking (3 papers): https://sciencetostartup.com/topic/llm-benchmarking
- Video Question Answering (3 papers): https://sciencetostartup.com/topic/video-question-answering
- Conversational Agents (3 papers): https://sciencetostartup.com/topic/conversational-agents
- Machine Learning Optimization (3 papers): https://sciencetostartup.com/topic/machine-learning-optimization
- AI for Hardware Design (3 papers): https://sciencetostartup.com/topic/ai-for-hardware-design
- Embodied Agents (3 papers): https://sciencetostartup.com/topic/embodied-agents
- Drug Discovery (3 papers): https://sciencetostartup.com/topic/drug-discovery
- LLM Robustness (3 papers): https://sciencetostartup.com/topic/llm-robustness
- Theoretical Foundations (3 papers): https://sciencetostartup.com/topic/theoretical-foundations
- AI-Generated Content Detection (3 papers): https://sciencetostartup.com/topic/ai-generated-content-detection
- Linguistic Analysis (3 papers): https://sciencetostartup.com/topic/linguistic-analysis
- Incremental Learning (3 papers): https://sciencetostartup.com/topic/incremental-learning
- Safe Reinforcement Learning (3 papers): https://sciencetostartup.com/topic/safe-reinforcement-learning
- Time Series Modeling (3 papers): https://sciencetostartup.com/topic/time-series-modeling
- Video AI (3 papers): https://sciencetostartup.com/topic/video-ai
- Emotional Support AI (3 papers): https://sciencetostartup.com/topic/emotional-support-ai
- Drone Navigation (3 papers): https://sciencetostartup.com/topic/drone-navigation
- Neural Network Architecture (3 papers): https://sciencetostartup.com/topic/neural-network-architecture
- Evaluation Frameworks (3 papers): https://sciencetostartup.com/topic/evaluation-frameworks
- LLM Behavior (3 papers): https://sciencetostartup.com/topic/llm-behavior
- NLP Applications (3 papers): https://sciencetostartup.com/topic/nlp-applications
- Genomics AI (3 papers): https://sciencetostartup.com/topic/genomics-ai
- Molecular Property Prediction (3 papers): https://sciencetostartup.com/topic/molecular-property-prediction
- AI Development Tools (3 papers): https://sciencetostartup.com/topic/ai-development-tools
- AI Monitoring (3 papers): https://sciencetostartup.com/topic/ai-monitoring
- LLM Prompt Engineering (3 papers): https://sciencetostartup.com/topic/llm-prompt-engineering
- AI Assistants (3 papers): https://sciencetostartup.com/topic/ai-assistants
- Geo-spatial AI (3 papers): https://sciencetostartup.com/topic/geo-spatial-ai
- Text Analysis (3 papers): https://sciencetostartup.com/topic/text-analysis
- Image Translation (3 papers): https://sciencetostartup.com/topic/image-translation
- Audio-Language Models (3 papers): https://sciencetostartup.com/topic/audio-language-models
- Depth Completion (3 papers): https://sciencetostartup.com/topic/depth-completion
- Audio Security (3 papers): https://sciencetostartup.com/topic/audio-security
- Gaze Estimation (3 papers): https://sciencetostartup.com/topic/gaze-estimation
- Safety in AI (3 papers): https://sciencetostartup.com/topic/safety-in-ai
- Cross-Modal Learning (3 papers): https://sciencetostartup.com/topic/cross-modal-learning
- 3D Modeling (3 papers): https://sciencetostartup.com/topic/3d-modeling
- AI Models (3 papers): https://sciencetostartup.com/topic/ai-models
- Prompt Engineering (3 papers): https://sciencetostartup.com/topic/prompt-engineering
- Video Processing (3 papers): https://sciencetostartup.com/topic/video-processing
- AI Code Review (3 papers): https://sciencetostartup.com/topic/ai-code-review
- 3D Graphics (3 papers): https://sciencetostartup.com/topic/3d-graphics
- Protein Design (3 papers): https://sciencetostartup.com/topic/protein-design
- Speech Synthesis (3 papers): https://sciencetostartup.com/topic/speech-synthesis
- ASR Optimization (3 papers): https://sciencetostartup.com/topic/asr-optimization
- Probabilistic Inference (3 papers): https://sciencetostartup.com/topic/probabilistic-inference
- Formal Methods (3 papers): https://sciencetostartup.com/topic/formal-methods
- Generative 3D (3 papers): https://sciencetostartup.com/topic/generative-3d
- Knowledge Management (3 papers): https://sciencetostartup.com/topic/knowledge-management
- Efficient Transformers (3 papers): https://sciencetostartup.com/topic/efficient-transformers
- UAV Navigation (3 papers): https://sciencetostartup.com/topic/uav-navigation
- LLM Training Optimization (3 papers): https://sciencetostartup.com/topic/llm-training-optimization
- Vision-Language Alignment (3 papers): https://sciencetostartup.com/topic/vision-language-alignment
- 3D Spatial Reasoning (3 papers): https://sciencetostartup.com/topic/3d-spatial-reasoning
- Knowledge Graph Construction (3 papers): https://sciencetostartup.com/topic/knowledge-graph-construction
- Speech Technology (3 papers): https://sciencetostartup.com/topic/speech-technology
- AI in Software Engineering (3 papers): https://sciencetostartup.com/topic/ai-in-software-engineering
- Contrastive Learning (3 papers): https://sciencetostartup.com/topic/contrastive-learning
- Interpretable Machine Learning (3 papers): https://sciencetostartup.com/topic/interpretable-machine-learning
- Software Development Tools (3 papers): https://sciencetostartup.com/topic/software-development-tools
- Computer Vision Datasets (3 papers): https://sciencetostartup.com/topic/computer-vision-datasets
- Neural Operators (3 papers): https://sciencetostartup.com/topic/neural-operators
- Data Assimilation (3 papers): https://sciencetostartup.com/topic/data-assimilation
- AI Education (3 papers): https://sciencetostartup.com/topic/ai-education
- AI Model Evaluation (3 papers): https://sciencetostartup.com/topic/ai-model-evaluation
- Hyperspectral Image Classification (3 papers): https://sciencetostartup.com/topic/hyperspectral-image-classification
- Robotics Optimization (3 papers): https://sciencetostartup.com/topic/robotics-optimization
- AI Code Generation (3 papers): https://sciencetostartup.com/topic/ai-code-generation
- LLM Tools (3 papers): https://sciencetostartup.com/topic/llm-tools
- Software Engineering Agents (3 papers): https://sciencetostartup.com/topic/software-engineering-agents
- UAV Control (3 papers): https://sciencetostartup.com/topic/uav-control
- Model Security (3 papers): https://sciencetostartup.com/topic/model-security
- Speech Translation (3 papers): https://sciencetostartup.com/topic/speech-translation
- Sensor Fusion (3 papers): https://sciencetostartup.com/topic/sensor-fusion
- Symbolic AI (3 papers): https://sciencetostartup.com/topic/symbolic-ai
- LiDAR Security (3 papers): https://sciencetostartup.com/topic/lidar-security
- Climate AI (3 papers): https://sciencetostartup.com/topic/climate-ai
- Offline Reinforcement Learning (3 papers): https://sciencetostartup.com/topic/offline-reinforcement-learning
- Embodied AI Navigation (3 papers): https://sciencetostartup.com/topic/embodied-ai-navigation
- Supply Chain AI (3 papers): https://sciencetostartup.com/topic/supply-chain-ai
- AI Safety and Security (3 papers): https://sciencetostartup.com/topic/ai-safety-and-security
- Theoretical ML (3 papers): https://sciencetostartup.com/topic/theoretical-ml
- Collaborative Perception (3 papers): https://sciencetostartup.com/topic/collaborative-perception
- 3D Semantic Segmentation (3 papers): https://sciencetostartup.com/topic/3d-semantic-segmentation
- Fake News Detection (3 papers): https://sciencetostartup.com/topic/fake-news-detection
- Multilingual Reasoning (3 papers): https://sciencetostartup.com/topic/multilingual-reasoning
- AI Evaluation Tools (3 papers): https://sciencetostartup.com/topic/ai-evaluation-tools
- AI/ML (3 papers): https://sciencetostartup.com/topic/aiml
- Software Engineering AI (3 papers): https://sciencetostartup.com/topic/software-engineering-ai

## Article Index (44)
- AI Reasoning, Generative Media, and Robotics Advance (2026-04-10): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-04-10
- AV-SQL Tackles Complex Text-to-SQL, EVGeoQA Tests Geo-Spatial Agents (2026-04-09): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-04-09
- Agents Conquer Software, Retail Supply Chains Get Smarter (2026-04-08): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-04-08
- Tiny AI Models Prove Hard Theorems, Video Mashups Get Smarter (2026-04-07): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-04-07
- AI Breakthroughs in Memory Systems, Calibration, and Autonomous Driving (2026-04-03): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-04-03
- AI Innovations in Data Types, Image Generation, and Video Understanding (2026-04-01): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-04-01
- OpenAI's $122B Funding Round and New AI Innovations (2026-03-31): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-03-31
- AXON, CADSmith, and Ruka-v2 Enhance Medical Imaging and Robotics (2026-03-30): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-03-30
- PackForcing, LEMMA, and PICon Advance Video and AI Evaluation (2026-03-27): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-03-27
- MARCH, VFIG, and OneSearch-V2 Transform AI Capabilities (2026-03-26): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-03-26
- SAMA, VEGA-3D, and Matryoshka Gaussian Splatting Redefine AI Capabilities (2026-03-23): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-03-23
- SAMA, VEGA-3D, and Matryoshka Gaussian Splatting Redefine AI Capabilities (2026-03-21): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-03-21
- Cubic Discrete Diffusion, SAMA, and Multilingual Embeddings Transform AI (2026-03-20): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-03-20
- New 3D Scene Reconstruction, Efficient LLMs, and Interactive Video AI (2026-03-19): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-03-19
- OpenSeeker and DOMINO: Pioneering AI Search and Robotics (2026-03-17): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-03-17
- Advancements in AI for Robotics, Medical Imaging, and Reasoning (2026-03-16): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-03-16
- AI Innovations in Video Generation, Humanoid Robotics, and Streaming Efficiency (2026-03-15): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-03-15
- AI Breakthroughs in Video Understanding, 3D Scene Generation, and Depth Estimation (2026-03-14): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-03-14
- AI Innovations in Image Editing, Video Generation, and Humanoid Robotics (2026-03-13): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-03-13
- AI-Driven Video-to-Music, Color Fidelity, and Point Cloud Innovations (2026-03-12): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-03-12
- Modular Robotics, Language-Conditioned Navigation, and AI in Healthcare (2026-03-11): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-03-11
- Open-Source Cybersecurity, Evolving Agents, and Deepfake Detection (2026-03-10): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-03-10
- Robotics Policy Iteration, AI Security, and Biomedical Evidence Tools (2026-03-09): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-03-09
- AI Security Innovations, Adaptive Learning, and Real-Time Video Generation (2026-03-07): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-03-07
- AI Security Frameworks, Robotics Innovations, and 3D Generation Breakthroughs (2026-03-06): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-03-06
- AI Feedback Tools, 3D Vision, and Agentic Workflows (2026-03-05): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-03-05
- AI Innovations in Industrial Optimization and Content Creation (2026-03-04): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-03-04
- AI Innovations in Conversion Rate Prediction and Unlearning (2026-03-03): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-03-03
- AI Security Innovations and Manufacturing Efficiency Boosts (2026-03-02): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-03-02
- AI Research Rundown: Innovations in Emotional AI and Video Generation (2026-03-01): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-03-01
- AI Research Rundown: Innovations in Medical Video Generation and QA (2026-02-28): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-02-28
- AI Research Rundown: Innovations in Human-Robot Interaction and QA (2026-02-27): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-02-27
- AI Research Rundown: Innovations in Text Ranking and Medical Imaging (2026-02-26): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-02-26
- AI Research Rundown: Innovations in Long-Context Inference and Medical AI (2026-02-25): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-02-25
- AI Research Rundown: Satellite Detection, VR Simulation, and Legal Predictions (2026-02-24): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-02-24
- AI Research Rundown: Enhancements in Adversarial Attacks and Reinforcement Learning (2026-02-23): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-02-23
- AI Research Rundown: Time Series, Adversarial Attacks, and Formula Recognition (2026-02-22): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-02-22
- AI Research Rundown: Adversarial Attacks, Time Series, and Behavioral Prediction (2026-02-21): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-02-21
- AI Research Rundown: Formula Recognition, Behavioral Prediction, and Adversarial Attacks (2026-02-21): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-02-20
- AI Research Rundown: Humanoid Motion, Graph Reasoning, and 3D Imaging (2026-02-19): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-02-19
- Frontier AI Innovations for Startup Execution (2026-02-18): https://sciencetostartup.com/articles/frontier-ai-innovations-for-startup-execution-2026-02-18
- AI Research Rundown: Skills, Incident Response, and Video Models (2026-02-17): https://sciencetostartup.com/articles/science-to-startup-daily-brief-2026-02-16
- Frontier AI and Startup Execution: Key Innovations (2026-02-15): https://sciencetostartup.com/articles/frontier-ai-and-startup-execution-key-innovations-2026-02-15
- Daily AI Research Rundown (2026-02-13): https://sciencetostartup.com/articles/daily-ai-research-rundown-2026-02-13

## Paper Index (11888)
- Act Wisely: Cultivating Meta-Cognitive Tool Use in Agentic Multimodal Models (viability: 7): https://sciencetostartup.com/paper/act-wisely-cultivating-meta-cognitive-tool-use-in-agentic-multimodal-models - A framework that significantly reduces tool invocations in agentic multimodal models while improving reasoning accuracy by decoupling accuracy and efficiency optimization.
- SIM1: Physics-Aligned Simulator as Zero-Shot Data Scaler in Deformable Worlds (viability: 7): https://sciencetostartup.com/paper/sim1-physics-aligned-simulator-as-zero-shot-data-scaler-in-deformable-worlds - A physics-aligned simulator that acts as a zero-shot data scaler for robotic manipulation of deformable objects, achieving parity with real-data baselines using purely synthetic data.
- Seeing but Not Thinking: Routing Distraction in Multimodal Mixture-of-Experts (viability: 7): https://sciencetostartup.com/paper/seeing-but-not-thinking-routing-distraction-in-multimodal-mixture-of-experts - A method to improve reasoning in multimodal MoE models by addressing 'Seeing but Not Thinking' by guiding expert activation to overcome routing distraction.
- OpenVLThinkerV2: A Generalist Multimodal Reasoning Model for Multi-domain Visual Tasks (viability: 7): https://sciencetostartup.com/paper/openvlthinkerv2-a-generalist-multimodal-reasoning-model-for-multi-domain-visual-tasks - A generalist multimodal reasoning model that significantly outperforms frontier models across 18 benchmarks by introducing a novel RL training objective and task-level shaping mechanisms.
- AVGen-Bench: A Task-Driven Benchmark for Multi-Granular Evaluation of Text-to-Audio-Video Generation (viability: 7): https://sciencetostartup.com/paper/avgen-bench-a-task-driven-benchmark-for-multi-granular-evaluation-of-text-to-audio-video-generation - A task-driven benchmark and evaluation framework for text-to-audio-video generation that reveals significant gaps in semantic controllability.
- RewardFlow: Generate Images by Optimizing What You Reward (viability: 6): https://sciencetostartup.com/paper/rewardflow-generate-images-by-optimizing-what-you-reward - A new image generation tool leveraging multi-reward Langevin dynamics for state-of-the-art image editing.
- PSI: Shared State as the Missing Layer for Coherent AI-Generated Instruments in Personal AI Agents (viability: 4): https://sciencetostartup.com/paper/psi-shared-state-as-the-missing-layer-for-coherent-ai-generated-instruments-in-personal-ai-agents - PSI is a shared-state architecture that transforms independently generated AI modules into coherent, connected personal computing environments accessible through GUIs and chat agents.
- Ads in AI Chatbots? An Analysis of How Large Language Models Navigate Conflicts of Interest (viability: 3): https://sciencetostartup.com/paper/ads-in-ai-chatbots-an-analysis-of-how-large-language-models-navigate-conflicts-of-interest - This paper analyzes how current LLMs navigate conflicts of interest between user welfare and company incentives, finding a majority prioritize company revenue over user benefit.
- What Drives Representation Steering? A Mechanistic Case Study on Steering Refusal (viability: 3): https://sciencetostartup.com/paper/what-drives-representation-steering-a-mechanistic-case-study-on-steering-refusal - This research investigates the internal mechanisms of steering vectors in large language models, revealing their primary interaction with the attention mechanism's OV circuit and enabling significant sparsification.
- ClawBench: Can AI Agents Complete Everyday Online Tasks? (viability: 7): https://sciencetostartup.com/paper/clawbench-can-ai-agents-complete-everyday-online-tasks - ClawBench is a new benchmark for evaluating AI agents on everyday online tasks across 144 live platforms, revealing current frontier models can only complete a small fraction of these complex, real-world challenges.
- Differentially Private Language Generation and Identification in the Limit (viability: 3): https://sciencetostartup.com/paper/differentially-private-language-generation-and-identification-in-the-limit - This research explores differentially private language generation and identification, showing privacy has no qualitative cost for generation but creates fundamental barriers for identification.
- Quantifying Explanation Consistency: The C-Score Metric for CAM-Based Explainability in Medical Image Classification (viability: 6): https://sciencetostartup.com/paper/quantifying-explanation-consistency-the-c-score-metric-for-cam-based-explainability-in-medical-image-classification - The C-Score metric quantifies explanation consistency in medical image classification, providing an early warning signal of model instability and enabling architecture-specific deployment recommendations.
- PIArena: A Platform for Prompt Injection Evaluation (viability: 7): https://sciencetostartup.com/paper/piarena-a-platform-for-prompt-injection-evaluation - A unified platform for evaluating prompt injection attacks and defenses, enabling reliable comparison and uncovering limitations of current security measures.
- SUPERNOVA: Eliciting General Reasoning in LLMs with Reinforcement Learning on Natural Instructions (viability: 8): https://sciencetostartup.com/paper/supernova-eliciting-general-reasoning-in-llms-with-reinforcement-learning-on-natural-instructions - SUPERNOVA enhances general reasoning in language models through a curated data framework for reinforcement learning with verifiable rewards.
- Faithful GRPO: Improving Visual Spatial Reasoning in Multimodal Language Models via Constrained Policy Optimization (viability: 7): https://sciencetostartup.com/paper/faithful-grpo-improving-visual-spatial-reasoning-in-multimodal-language-models-via-constrained-policy-optimization - Faithful GRPO improves visual spatial reasoning in multimodal models by enforcing logical consistency and visual grounding, leading to more accurate and trustworthy answers.
- TTVS: Boosting Self-Exploring Reinforcement Learning via Test-time Variational Synthesis (viability: 3): https://sciencetostartup.com/paper/ttvs-boosting-self-exploring-reinforcement-learning-via-test-time-variational-synthesis - TTVS enables large reasoning models to self-evolve at test time by dynamically synthesizing diverse variations of unlabeled queries, improving performance in specialized domains.
- From Safety Risk to Design Principle: Peer-Preservation in Multi-Agent LLM Systems and Its Implications for Orchestrated Democratic Discourse Analysis (viability: 3): https://sciencetostartup.com/paper/from-safety-risk-to-design-principle-peer-preservation-in-multi-agent-llm-systems-and-its-implications-for-orchestrated - This paper explores emergent 'peer-preservation' alignment issues in multi-agent LLM systems and proposes architectural mitigations for orchestrated democratic discourse analysis.
- OVS-DINO: Open-Vocabulary Segmentation via Structure-Aligned SAM-DINO with Language Guidance (viability: 7): https://sciencetostartup.com/paper/ovs-dino-open-vocabulary-segmentation-via-structure-aligned-sam-dino-with-language-guidance - OVS-DINO revitalizes edge-sensitivity in DINO for open-vocabulary segmentation by structurally aligning with SAM, achieving state-of-the-art results.
- A Machine Learning Framework for Turbofan Health Estimation via Inverse Problem Formulation (viability: 5): https://sciencetostartup.com/paper/a-machine-learning-framework-for-turbofan-health-estimation-via-inverse-problem-formulation - A machine learning framework for turbofan health estimation using a new dataset and self-supervised learning to address the ill-posed inverse problem.
- CrashSight: A Phase-Aware, Infrastructure-Centric Video Benchmark for Traffic Crash Scene Understanding and Reasoning (viability: 7): https://sciencetostartup.com/paper/crashsight-a-phase-aware-infrastructure-centric-video-benchmark-for-traffic-crash-scene-understanding-and-reasoning - CrashSight is a large-scale benchmark for traffic crash scene understanding using roadside camera data, evaluating vision-language models' reasoning in safety-critical scenarios.
- KnowU-Bench: Towards Interactive, Proactive, and Personalized Mobile Agent Evaluation (viability: 7): https://sciencetostartup.com/paper/knowu-bench-towards-interactive-proactive-and-personalized-mobile-agent-evaluation - A new benchmark and simulator for evaluating personalized mobile agents that can infer user preferences and provide proactive assistance in real-time GUI environments.
- HST-HGN: Heterogeneous Spatial-Temporal Hypergraph Networks with Bidirectional State Space Models for Global Fatigue Assessment (viability: 7): https://sciencetostartup.com/paper/hst-hgn-heterogeneous-spatial-temporal-hypergraph-networks-with-bidirectional-state-space-models-for-global-fatigue-asse - A novel hypergraph network with bidirectional state space models for efficient and accurate driver fatigue assessment from untrimmed videos, suitable for edge deployment.
- Small-scale photonic Kolmogorov-Arnold networks using standard telecom nonlinear modules (viability: 6): https://sciencetostartup.com/paper/small-scale-photonic-kolmogorov-arnold-networks-using-standard-telecom-nonlinear-modules - Small-scale photonic Kolmogorov-Arnold networks implemented with standard telecom components for ultrafast inference across various tasks, overcoming optical-electrical-optical bottlenecks.
- KV Cache Offloading for Context-Intensive Tasks (viability: 7): https://sciencetostartup.com/paper/kv-cache-offloading-for-context-intensive-tasks - A new benchmark and improved KV cache offloading strategy to address performance degradation on context-intensive LLM tasks, enabling more accurate long-context processing.
- Learning Who Disagrees: Demographic Importance Weighting for Modeling Annotator Distributions with DiADEM (viability: 7): https://sciencetostartup.com/paper/learning-who-disagrees-demographic-importance-weighting-for-modeling-annotator-distributions-with-diadem - A neural architecture that models annotator disagreement by learning the importance of demographic factors, outperforming LLMs on subjective content labeling.
- On-board Telemetry Monitoring in Autonomous Satellites: Challenges and Opportunities (viability: 4): https://sciencetostartup.com/paper/on-board-telemetry-monitoring-in-autonomous-satellites-challenges-and-opportunities - A framework for explainable AI in autonomous satellite fault detection, using neural anomaly detectors with interpretable indicators.
- Synthetic Data for any Differentiable Target (viability: 6): https://sciencetostartup.com/paper/synthetic-data-for-any-differentiable-target - A flexible synthetic data generator for customizable machine learning tasks.
- Exploring Temporal Representation in Neural Processes for Multimodal Action Prediction (viability: 4): https://sciencetostartup.com/paper/exploring-temporal-representation-in-neural-processes-for-multimodal-action-prediction - A revised Conditional Neural Process model that improves temporal representation for multimodal action prediction in robotics.
- Selective Attention System (SAS): Device-Addressed Speech Detection for Real-Time On-Device Voice AI (viability: 7): https://sciencetostartup.com/paper/selective-attention-system-sas-device-addressed-speech-detection-for-real-time-on-device-voice-ai - A selective attention system for device-addressed speech detection that operates on-device with low latency and footprint, achieving high accuracy with optional video fusion.
- Verify Before You Commit: Towards Faithful Reasoning in LLM Agents via Self-Auditing (viability: 3): https://sciencetostartup.com/paper/verify-before-you-commit-towards-faithful-reasoning-in-llm-agents-via-self-auditing - Develop a self-auditing tool for LLM agents to ensure faithful reasoning.
- Zero-shot Multivariate Time Series Forecasting Using Tabular Prior Fitted Networks (viability: 4): https://sciencetostartup.com/paper/zero-shot-multivariate-time-series-forecasting-using-tabular-prior-fitted-networks - A framework for multivariate time series forecasting using tabular foundation models by recasting the problem into a series of scalar regression problems solvable zero-shot.
- ADAPTive Input Training for Many-to-One Pre-Training on Time-Series Classification (viability: 7): https://sciencetostartup.com/paper/adaptive-input-training-for-many-to-one-pre-training-on-time-series-classification - ADAPT is a new pre-training paradigm for time-series data that enables mixed-batch pre-training across diverse datasets, setting new state-of-the-art performance for classification benchmarks.
- Phantasia: Context-Adaptive Backdoors in Vision Language Models (viability: 4): https://sciencetostartup.com/paper/phantasia-context-adaptive-backdoors-in-vision-language-models - A novel backdoor attack for Vision-Language Models that dynamically aligns poisoned outputs with input semantics for improved stealth and adaptability.
- Awakening the Sleeping Agent: Lean-Specific Agentic Data Reactivates General Tool Use in Goedel Prover (viability: 7): https://sciencetostartup.com/paper/awakening-the-sleeping-agent-lean-specific-agentic-data-reactivates-general-tool-use-in-goedel-prover - A small amount of domain-specific agentic data can reactivate dormant tool-use capabilities in large language models, significantly improving performance on diverse tasks.
- TASU2: Controllable CTC Simulation for Alignment and Low-Resource Adaptation of Speech LLMs (viability: 4): https://sciencetostartup.com/paper/tasu2-controllable-ctc-simulation-for-alignment-and-low-resource-adaptation-of-speech-llms - TASU2 is a controllable CTC simulation framework for speech LLMs that enables principled post-training curricula for improved alignment and low-resource adaptation.
- A GAN and LLM-Driven Data Augmentation Framework for Dynamic Linguistic Pattern Modeling in Chinese Sarcasm Detection (viability: 7): https://sciencetostartup.com/paper/a-gan-and-llm-driven-data-augmentation-framework-for-dynamic-linguistic-pattern-modeling-in-chinese-sarcasm-detection - A GAN and LLM-driven framework that dynamically models user linguistic patterns for enhanced Chinese sarcasm detection, achieving state-of-the-art results.
- SkillClaw: Let Skills Evolve Collectively with Agentic Evolver (viability: 4): https://sciencetostartup.com/paper/skillclaw-let-skills-evolve-collectively-with-agentic-evolver - SkillClaw enables LLM agents to collectively evolve their skills by learning from cross-user interactions, improving performance system-wide without user effort.
- Don't Overthink It: Inter-Rollout Action Agreement as a Free Adaptive-Compute Signal for LLM Agents (viability: 6): https://sciencetostartup.com/paper/don-t-overthink-it-inter-rollout-action-agreement-as-a-free-adaptive-compute-signal-for-llm-agents - TrACE adaptively allocates LLM compute for agents by measuring inter-rollout action agreement, reducing LLM calls while maintaining accuracy.
- Scaling-Aware Data Selection for End-to-End Autonomous Driving Systems (viability: 5): https://sciencetostartup.com/paper/scaling-aware-data-selection-for-end-to-end-autonomous-driving-systems - MOSAIC is a data selection framework for autonomous driving that uses neural scaling laws to optimize training data mixtures, reducing data needs by up to 80%.
- Scalable Neural Decoders for Practical Fault-Tolerant Quantum Computation (viability: 5): https://sciencetostartup.com/paper/scalable-neural-decoders-for-practical-fault-tolerant-quantum-computation - A convolutional neural network decoder for quantum error correction achieves significantly lower logical error rates and higher throughput than existing methods.
- ASPECT:Analogical Semantic Policy Execution via Language Conditioned Transfer (viability: 4): https://sciencetostartup.com/paper/aspect-analogical-semantic-policy-execution-via-language-conditioned-transfer - A reinforcement learning agent that uses a Large Language Model as a semantic operator to achieve zero-shot transfer to novel analogous tasks.
- Human-AI Collaboration Reconfigures Group Regulation from Socially Shared to Hybrid Co-Regulation (viability: 0): https://sciencetostartup.com/paper/human-ai-collaboration-reconfigures-group-regulation-from-socially-shared-to-hybrid-co-regulation - This paper investigates how generative AI impacts group regulation in collaborative learning environments, shifting from socially shared to hybrid co-regulation.
- PokeGym: A Visually-Driven Long-Horizon Benchmark for Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/pokegym-a-visually-driven-long-horizon-benchmark-for-vision-language-models - PokeGym is a visually-driven benchmark for Vision-Language Models in complex 3D environments, revealing physical deadlock recovery as a key bottleneck.
- InstAP: Instance-Aware Vision-Language Pre-Train for Spatial-Temporal Understanding (viability: 7): https://sciencetostartup.com/paper/instap-instance-aware-vision-language-pre-train-for-spatial-temporal-understanding - InstAP is an instance-aware pre-training framework for vision-language models that improves both instance-level and global understanding.
- Dead Weights, Live Signals: Feedforward Graphs of Frozen Language Models (viability: 7): https://sciencetostartup.com/paper/dead-weights-live-signals-feedforward-graphs-of-frozen-language-models - A novel feedforward graph architecture that composes heterogeneous frozen LLMs to achieve state-of-the-art performance on challenging benchmarks with minimal trainable parameters.
- Lost in the Hype: Revealing and Dissecting the Performance Degradation of Medical Multimodal Large Language Models in Image Classification (viability: 4): https://sciencetostartup.com/paper/lost-in-the-hype-revealing-and-dissecting-the-performance-degradation-of-medical-multimodal-large-language-models-in-ima - Investigates the performance degradation of medical multimodal LLMs in image classification, identifying key failure modes to guide future development.
- ProMedical: Hierarchical Fine-Grained Criteria Modeling for Medical LLM Alignment via Explicit Injection (viability: 8): https://sciencetostartup.com/paper/promedical-hierarchical-fine-grained-criteria-modeling-for-medical-llm-alignment-via-explicit-injection - A novel alignment framework for medical LLMs that uses fine-grained clinical criteria and explicit injection to achieve superior accuracy and safety compliance.
- Multi-Modal Learning meets Genetic Programming: Analyzing Alignment in Latent Space Optimization (viability: 3): https://sciencetostartup.com/paper/multi-modal-learning-meets-genetic-programming-analyzing-alignment-in-latent-space-optimization - Investigates the effectiveness of multi-modal latent space optimization for symbolic regression, revealing limitations in current alignment techniques.
- HistDiT: A Structure-Aware Latent Conditional Diffusion Model for High-Fidelity Virtual Staining in Histopathology (viability: 7): https://sciencetostartup.com/paper/histdit-a-structure-aware-latent-conditional-diffusion-model-for-high-fidelity-virtual-staining-in-histopathology - A structure-aware latent conditional diffusion model for high-fidelity virtual staining in histopathology, outperforming existing methods.
- Securing Retrieval-Augmented Generation: A Taxonomy of Attacks, Defenses, and Future Directions (viability: 3): https://sciencetostartup.com/paper/securing-retrieval-augmented-generation-a-taxonomy-of-attacks-defenses-and-future-directions - A taxonomy of attacks and defenses for retrieval-augmented generation (RAG) systems, highlighting fragmented current defenses.
- DMax: Aggressive Parallel Decoding for dLLMs (viability: 8): https://sciencetostartup.com/paper/dmax-aggressive-parallel-decoding-for-dllms - DMax offers aggressive parallel decoding for diffusion language models, significantly increasing throughput while preserving generation quality.
- SeLaR: Selective Latent Reasoning in Large Language Models (viability: 7): https://sciencetostartup.com/paper/selar-selective-latent-reasoning-in-large-language-models - SeLaR is a training-free framework that selectively uses latent reasoning to improve LLM performance on complex reasoning tasks.
- U-CECE: A Universal Multi-Resolution Framework for Conceptual Counterfactual Explanations (viability: 7): https://sciencetostartup.com/paper/u-cece-a-universal-multi-resolution-framework-for-conceptual-counterfactual-explanations - A universal framework for generating conceptual counterfactual explanations for AI models, balancing expressivity and efficiency across different data regimes.
- Can Vision Language Models Judge Action Quality? An Empirical Evaluation (viability: 2): https://sciencetostartup.com/paper/can-vision-language-models-judge-action-quality-an-empirical-evaluation - Evaluating the capability of Vision Language Models for Action Quality Assessment, revealing significant limitations and biases.
- CIAO - Code In Architecture Out - Automated Software Architecture Documentation with Large Language Models (viability: 8): https://sciencetostartup.com/paper/ciao-code-in-architecture-out-automated-software-architecture-documentation-with-large-language-models - Automated system-level software architecture documentation generation from code repositories using LLMs, providing valuable and cost-effective insights.
- Distributed Multi-Layer Editing for Rule-Level Knowledge in Large Language Models (viability: 7): https://sciencetostartup.com/paper/distributed-multi-layer-editing-for-rule-level-knowledge-in-large-language-models - A distributed multi-layer editing approach for rule-level knowledge in LLMs, improving consistency and understanding across different knowledge forms.
- QARIMA: A Quantum Approach To Classical Time Series Analysis (viability: 5): https://sciencetostartup.com/paper/qarima-a-quantum-approach-to-classical-time-series-analysis - A quantum-inspired ARIMA methodology for enhanced time series analysis.
- ACF: A Collaborative Framework for Agent Covert Communication under Cognitive Asymmetry (viability: 5): https://sciencetostartup.com/paper/acf-a-collaborative-framework-for-agent-covert-communication-under-cognitive-asymmetry - A framework for covert communication among autonomous agents overcoming cognitive asymmetry.
- Neural-Symbolic Knowledge Tracing: Injecting Educational Knowledge into Deep Learning for Responsible Learner Modelling (viability: 7): https://sciencetostartup.com/paper/neural-symbolic-knowledge-tracing-injecting-educational-knowledge-into-deep-learning-for-responsible-learner-modelling - A neural-symbolic approach for responsible learner modeling in education.
- DBMF: A Dual-Branch Multimodal Framework for Out-of-Distribution Detection (viability: 6): https://sciencetostartup.com/paper/dbmf-a-dual-branch-multimodal-framework-for-out-of-distribution-detection - A dual-branch multimodal framework for robust out-of-distribution detection.
- Behavior-Aware Item Modeling via Dynamic Procedural Solution Representations for Knowledge Tracing (viability: 4): https://sciencetostartup.com/paper/behavior-aware-item-modeling-via-dynamic-procedural-solution-representations-for-knowledge-tracing - A framework that enhances knowledge tracing by integrating dynamic procedural solution stages into item representations, improving prediction of learner performance.
- HyperMem: Hypergraph Memory for Long-Term Conversations (viability: 7): https://sciencetostartup.com/paper/hypermem-hypergraph-memory-for-long-term-conversations - HyperMem is a hypergraph-based memory architecture for conversational agents that captures high-order associations for more coherent and personalized long-term dialogues.
- From Phenomenological Fitting to Endogenous Deduction: A Paradigm Leap via Meta-Principle Physics Architecture (viability: 3): https://sciencetostartup.com/paper/from-phenomenological-fitting-to-endogenous-deduction-a-paradigm-leap-via-meta-principle-physics-architecture - A Meta-Principle Physics Architecture that embeds core physical principles like connectivity, conservation, and periodicity into neural networks for improved physical reasoning and generalization.
- HiRO-Nav: Hybrid ReasOning Enables Efficient Embodied Navigation (viability: 7): https://sciencetostartup.com/paper/hiro-nav-hybrid-reasoning-enables-efficient-embodied-navigation - HiRO-Nav is an embodied navigation agent that adaptively uses reasoning only for high-entropy actions, reducing computational cost while improving decision quality.
- Grounding Clinical AI Competency in Human Cognition Through the Clinical World Model and Skill-Mix Framework (viability: 3): https://sciencetostartup.com/paper/grounding-clinical-ai-competency-in-human-cognition-through-the-clinical-world-model-and-skill-mix-framework - A framework to formalize clinical AI competency by modeling the interactions between patient, provider, and ecosystem, grounded in clinical cognition.
- EditCaption: Human-Aligned Instruction Synthesis for Image Editing via Supervised Fine-Tuning and Direct Preference Optimization (viability: 7): https://sciencetostartup.com/paper/editcaption-human-aligned-instruction-synthesis-for-image-editing-via-supervised-fine-tuning-and-direct-preference-optim - EditCaption synthesizes human-aligned instructions for image editing by combining supervised fine-tuning and direct preference optimization, significantly improving VLM accuracy.
- MedVR: Annotation-Free Medical Visual Reasoning via Agentic Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/medvr-annotation-free-medical-visual-reasoning-via-agentic-reinforcement-learning - MedVR enables annotation-free medical visual reasoning for VLMs using agentic reinforcement learning, improving robustness and transparency.
- AT-ADD: All-Type Audio Deepfake Detection Challenge Evaluation Plan (viability: 3): https://sciencetostartup.com/paper/at-add-all-type-audio-deepfake-detection-challenge-evaluation-plan - The AT-ADD challenge aims to advance all-type audio deepfake detection beyond speech-centric methods for robust multimedia forensics.
- Aligning Agents via Planning: A Benchmark for Trajectory-Level Reward Modeling (viability: 7): https://sciencetostartup.com/paper/aligning-agents-via-planning-a-benchmark-for-trajectory-level-reward-modeling - A new benchmark and evaluation suite for training reward models that align AI agents capable of complex tool use and reasoning.
- OceanMAE: A Foundation Model for Ocean Remote Sensing (viability: 8): https://sciencetostartup.com/paper/oceanmae-a-foundation-model-for-ocean-remote-sensing - A foundation model for ocean remote sensing that improves marine pollutant detection and bathymetry estimation by integrating multispectral data with ocean-specific features.
- Activation Steering for Aligned Open-ended Generation without Sacrificing Coherence (viability: 3): https://sciencetostartup.com/paper/activation-steering-for-aligned-open-ended-generation-without-sacrificing-coherence - A novel activation steering method for large language models that improves alignment without sacrificing coherence, addressing brittleness in generation.
- ViVa: A Video-Generative Value Model for Robot Reinforcement Learning (viability: 4): https://sciencetostartup.com/paper/viva-a-video-generative-value-model-for-robot-reinforcement-learning - A video-generative value model for robot reinforcement learning that improves task progress estimation by leveraging spatiotemporal priors from video data.
- Face-D(^2)CL: Multi-Domain Synergistic Representation with Dual Continual Learning for Facial DeepFake Detection (viability: 4): https://sciencetostartup.com/paper/face-d-2-cl-multi-domain-synergistic-representation-with-dual-continual-learning-for-facial-deepfake-detection - A dual continual learning framework for facial DeepFake detection that fuses spatial and frequency-domain features to adapt to evolving forgery patterns without historical data replay.
- Multimodal Reasoning with LLM for Encrypted Traffic Interpretation: A Benchmark (viability: 7): https://sciencetostartup.com/paper/multimodal-reasoning-with-llm-for-encrypted-traffic-interpretation-a-benchmark - An end-to-end multimodal reasoning framework that uses LLMs to interpret encrypted network traffic, generating human-readable reports grounded in raw byte data.
- Alloc-MoE: Budget-Aware Expert Activation Allocation for Efficient Mixture-of-Experts Inference (viability: 7): https://sciencetostartup.com/paper/alloc-moe-budget-aware-expert-activation-allocation-for-efficient-mixture-of-experts-inference - Alloc-MoE optimizes expert activation allocation in Mixture-of-Experts models for efficient inference, maintaining performance under budget constraints.
- Beyond Stochastic Exploration: What Makes Training Data Valuable for Agentic Search (viability: 4): https://sciencetostartup.com/paper/beyond-stochastic-exploration-what-makes-training-data-valuable-for-agentic-search - Hierarchical Experience (HiExp) framework enhances LLM search agents by transforming raw reasoning trajectories into hierarchical knowledge for strategic, experience-driven exploration.
- LegoDiffusion: Micro-Serving Text-to-Image Diffusion Workflows (viability: 4): https://sciencetostartup.com/paper/legodiffusion-micro-serving-text-to-image-diffusion-workflows - A system for micro-serving text-to-image diffusion workflows that optimizes resource management and performance.
- Uni-ViGU: Towards Unified Video Generation and Understanding via A Diffusion-Based Video Generator (viability: 7): https://sciencetostartup.com/paper/uni-vigu-towards-unified-video-generation-and-understanding-via-a-diffusion-based-video-generator - A unified framework for video generation and understanding that leverages a video generator as the foundation.
- Small Vision-Language Models are Smart Compressors for Long Video Understanding (viability: 7): https://sciencetostartup.com/paper/small-vision-language-models-are-smart-compressors-for-long-video-understanding - An efficient framework that compresses long videos for downstream understanding using small vision-language models.
- Revise: A Framework for Revising OCRed text in Practical Information Systems with Data Contamination Strategy (viability: 7): https://sciencetostartup.com/paper/revise-a-framework-for-revising-ocred-text-in-practical-information-systems-with-data-contamination-strategy - A framework for correcting OCR errors and improving document retrieval and question answering.
- TADP-RME: A Trust-Adaptive Differential Privacy Framework for Enhancing Reliability of Data-Driven Systems (viability: 3): https://sciencetostartup.com/paper/tadp-rme-a-trust-adaptive-differential-privacy-framework-for-enhancing-reliability-of-data-driven-systems - A framework for adaptive differential privacy that modulates privacy budgets based on user trust and uses manifold embedding to disrupt inference attacks.
- OV-Stitcher: A Global Context-Aware Framework for Training-Free Open-Vocabulary Semantic Segmentation (viability: 7): https://sciencetostartup.com/paper/ov-stitcher-a-global-context-aware-framework-for-training-free-open-vocabulary-semantic-segmentation - A training-free framework that stitches sub-image features to enable global context awareness for open-vocabulary semantic segmentation.
- AtlasOCR: Building the First Open-Source Darija OCR Model with Vision Language Models (viability: 8): https://sciencetostartup.com/paper/atlasocr-building-the-first-open-source-darija-ocr-model-with-vision-language-models - AtlasOCR is the first open-source Darija OCR model, fine-tuned from a VLM using efficient training strategies and a custom dataset.
- ImplicitMemBench: Measuring Unconscious Behavioral Adaptation in Large Language Models (viability: 5): https://sciencetostartup.com/paper/implicitmembench-measuring-unconscious-behavioral-adaptation-in-large-language-models - ImplicitMemBench is a new benchmark for LLM agents that measures unconscious behavioral adaptation, revealing limitations in current models.
- From Gaze to Guidance: Interpreting and Adapting to Users' Cognitive Needs with Multimodal Gaze-Aware AI Assistants (viability: 5): https://sciencetostartup.com/paper/from-gaze-to-guidance-interpreting-and-adapting-to-users-cognitive-needs-with-multimodal-gaze-aware-ai-assistants - A gaze-aware AI assistant that uses egocentric video to understand user cognitive needs and provide more accurate and personalized assistance.
- Governed Capability Evolution for Embodied Agents: Safe Upgrade, Compatibility Checking, and Runtime Rollback for Embodied Capability Modules (viability: 3): https://sciencetostartup.com/paper/governed-capability-evolution-for-embodied-agents-safe-upgrade-compatibility-checking-and-runtime-rollback-for-embodied - A framework for safely upgrading and managing capabilities of embodied agents, ensuring compatibility and preventing runtime failures.
- 3DrawAgent: Teaching LLM to Draw in 3D with Early Contrastive Experience (viability: 4): https://sciencetostartup.com/paper/3drawagent-teaching-llm-to-draw-in-3d-with-early-contrastive-experience - A training-free framework that teaches LLMs to draw 3D sketches using early contrastive experience and geometric feedback.
- LINE: LLM-based Iterative Neuron Explanations for Vision Models (viability: 7): https://sciencetostartup.com/paper/line-llm-based-iterative-neuron-explanations-for-vision-models - A black-box, training-free method using LLMs to iteratively label neurons in vision models with open-vocabulary concepts, improving interpretability and safety.
- PrivFedTalk: Privacy-Aware Federated Diffusion with Identity-Stable Adapters for Personalized Talking-Head Generation (viability: 7): https://sciencetostartup.com/paper/privfedtalk-privacy-aware-federated-diffusion-with-identity-stable-adapters-for-personalized-talking-head-generation - A privacy-preserving federated learning framework for personalized talking-head generation using lightweight identity adapters.
- IoT-Brain: Grounding LLMs for Semantic-Spatial Sensor Scheduling (viability: 8): https://sciencetostartup.com/paper/iot-brain-grounding-llms-for-semantic-spatial-sensor-scheduling - IoT-Brain bridges LLMs and sensor networks for proactive, intent-driven physical world interaction through semantic-spatial sensor scheduling.
- "Why This Avoidance Maneuver?" Contrastive Explanations in Human-Supervised Maritime Autonomous Navigation (viability: 3): https://sciencetostartup.com/paper/why-this-avoidance-maneuver-contrastive-explanations-in-human-supervised-maritime-autonomous-navigation - Contrastive explanations for maritime autonomous navigation systems to improve human supervisor understanding of avoidance maneuvers.
- From Universal to Individualized Actionability: Revisiting Personalization in Algorithmic Recourse (viability: 6): https://sciencetostartup.com/paper/from-universal-to-individualized-actionability-revisiting-personalization-in-algorithmic-recourse - Formalizing personalization in algorithmic recourse to provide actionable recommendations tailored to individual constraints and preferences.
- Wiring the 'Why': A Unified Taxonomy and Survey of Abductive Reasoning in LLMs (viability: 4): https://sciencetostartup.com/paper/wiring-the-why-a-unified-taxonomy-and-survey-of-abductive-reasoning-in-llms - A comprehensive survey and taxonomy of abductive reasoning in large language models, addressing conceptual confusion and task definitions.
- SearchAD: Large-Scale Rare Image Retrieval Dataset for Autonomous Driving (viability: 7): https://sciencetostartup.com/paper/searchad-large-scale-rare-image-retrieval-dataset-for-autonomous-driving - SearchAD is a large-scale dataset designed for rare image retrieval in autonomous driving, enhancing data curation and perception research.
- Evaluating Counterfactual Explanation Methods on Incomplete Inputs (viability: 2): https://sciencetostartup.com/paper/evaluating-counterfactual-explanation-methods-on-incomplete-inputs - This paper evaluates counterfactual explanation methods under the challenge of incomplete inputs, highlighting the need for robust solutions.
- The ecosystem of machine learning competitions: Platforms, participants, and their impact on AI development (viability: 5): https://sciencetostartup.com/paper/the-ecosystem-of-machine-learning-competitions-platforms-participants-and-their-impact-on-ai-development - An analysis of machine learning competitions and their impact on AI development, emphasizing collaboration and innovation.
- PASK: Toward Intent-Aware Proactive Agents with Long-Term Memory (viability: 7): https://sciencetostartup.com/paper/pask-toward-intent-aware-proactive-agents-with-long-term-memory - Develops a proactive agent paradigm with long-term memory and a real-world benchmark for streaming AI agents.
- Show Me the Infographic I Imagine: Intent-Aware Infographic Retrieval for Authoring Support (viability: 7): https://sciencetostartup.com/paper/show-me-the-infographic-i-imagine-intent-aware-infographic-retrieval-for-authoring-support - An intent-aware infographic retrieval framework that uses a taxonomy to align user queries with infographic designs for authoring support.
- LogAct: Enabling Agentic Reliability via Shared Logs (viability: 7): https://sciencetostartup.com/paper/logact-enabling-agentic-reliability-via-shared-logs - LogAct is a new abstraction for LLM-driven agents that enables reliability through shared logs, allowing for introspection and recovery.
- A Decomposition Perspective to Long-context Reasoning for LLMs (viability: 4): https://sciencetostartup.com/paper/a-decomposition-perspective-to-long-context-reasoning-for-llms - Decomposes long-context reasoning into atomic skills and uses reinforcement learning on synthesized datasets to improve LLM performance.
- How Far Are Large Multimodal Models from Human-Level Spatial Action? A Benchmark for Goal-Oriented Embodied Navigation in Urban Airspace (viability: 7): https://sciencetostartup.com/paper/how-far-are-large-multimodal-models-from-human-level-spatial-action-a-benchmark-for-goal-oriented-embodied-navigation-in - A benchmark and dataset for evaluating large multimodal models in goal-oriented urban airspace navigation, revealing current limitations and future improvement directions.
- AtomEval: Atomic Evaluation of Adversarial Claims in Fact Verification (viability: 7): https://sciencetostartup.com/paper/atomeval-atomic-evaluation-of-adversarial-claims-in-fact-verification - AtomEval is a validity-aware evaluation framework for fact verification that decomposes claims into atomic units to detect factual corruption in adversarial rewrites.
- DSCA: Dynamic Subspace Concept Alignment for Lifelong VLM Editing (viability: 8): https://sciencetostartup.com/paper/dsca-dynamic-subspace-concept-alignment-for-lifelong-vlm-editing - DSCA offers a novel approach to lifelong VLM editing by structurally separating concepts into orthogonal semantic subspaces, enabling precise, non-interfering edits with state-of-the-art stability.
- Are we still able to recognize pearls? Machine-driven peer review and the risk to creativity: An explainable RAG-XAI detection framework with markers extraction (viability: 3): https://sciencetostartup.com/paper/are-we-still-able-to-recognize-pearls-machine-driven-peer-review-and-the-risk-to-creativity-an-explainable-rag-xai-detec - An explainable RAG-XAI framework is proposed to detect machine-driven peer review patterns and markers, aiming to preserve creativity in science.
- Rethinking Data Mixing from the Perspective of Large Language Models (viability: 4): https://sciencetostartup.com/paper/rethinking-data-mixing-from-the-perspective-of-large-language-models - A theoretical framework and reweighting method for optimizing LLM training data mixing to improve generalization.
- TOOLCAD: Exploring Tool-Using Large Language Models in Text-to-CAD Generation with Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/toolcad-exploring-tool-using-large-language-models-in-text-to-cad-generation-with-reinforcement-learning - An agentic framework using LLMs and reinforcement learning to generate CAD models from text, competitive with proprietary models.
- WorldMAP: Bootstrapping Vision-Language Navigation Trajectory Prediction with Generative World Models (viability: 5): https://sciencetostartup.com/paper/worldmap-bootstrapping-vision-language-navigation-trajectory-prediction-with-generative-world-models - A framework that uses generative world models to create supervision signals for vision-language navigation trajectory prediction.
- MONETA: Multimodal Industry Classification through Geographic Information with Multi Agent Systems (viability: 7): https://sciencetostartup.com/paper/moneta-multimodal-industry-classification-through-geographic-information-with-multi-agent-systems - A multimodal benchmark and system for industry classification using text and geospatial data, achieving significant performance gains.
- Pruning Extensions and Efficiency Trade-Offs for Sustainable Time Series Classification (viability: 7): https://sciencetostartup.com/paper/pruning-extensions-and-efficiency-trade-offs-for-sustainable-time-series-classification - A framework for energy-efficient time series classification through a novel pruning strategy.
- Investigation of Automated Design of Quantum Circuits for Imaginary Time Evolution Methods Using Deep Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/investigation-of-automated-design-of-quantum-circuits-for-imaginary-time-evolution-methods-using-deep-reinforcement-lear - An automated framework for designing quantum circuits using deep reinforcement learning.
- Incremental Residual Reinforcement Learning Toward Real-World Learning for Social Navigation (viability: 7): https://sciencetostartup.com/paper/incremental-residual-reinforcement-learning-toward-real-world-learning-for-social-navigation - A novel incremental residual reinforcement learning method for real-world social navigation in mobile robots.
- On-Policy Distillation of Language Models for Autonomous Vehicle Motion Planning (viability: 7): https://sciencetostartup.com/paper/on-policy-distillation-of-language-models-for-autonomous-vehicle-motion-planning - A method for distilling knowledge from large language models for efficient motion planning in autonomous vehicles.
- Large Language Model Post-Training: A Unified View of Off-Policy and On-Policy Learning (viability: 1): https://sciencetostartup.com/paper/large-language-model-post-training-a-unified-view-of-off-policy-and-on-policy-learning - A survey that unifies various LLM post-training methods by framing them as structured interventions on model behavior, categorized by trajectory provenance and behavioral roles.
- Same Outcomes, Different Journeys: A Trace-Level Framework for Comparing Human and GUI-Agent Behavior in Production Search Systems (viability: 4): https://sciencetostartup.com/paper/same-outcomes-different-journeys-a-trace-level-framework-for-comparing-human-and-gui-agent-behavior-in-production-search - A framework to compare human and GUI-agent behavior at a trace level in production search systems, revealing differences in navigation strategies beyond task success.
- EigentSearch-Q+: Enhancing Deep Research Agents with Structured Reasoning Tools (viability: 7): https://sciencetostartup.com/paper/eigentsearch-q-enhancing-deep-research-agents-with-structured-reasoning-tools - Enhance AI agents for deep research by integrating structured reasoning tools (Q+) into web search for deliberate query planning and evidence extraction, improving accuracy.
- Sinkhorn doubly stochastic attention rank decay analysis (viability: 1): https://sciencetostartup.com/paper/sinkhorn-doubly-stochastic-attention-rank-decay-analysis - Analyzes rank decay in self-attention mechanisms, showing that Sinkhorn doubly stochastic attention preserves rank more effectively than standard Softmax attention.
- SAT: Balancing Reasoning Accuracy and Efficiency with Stepwise Adaptive Thinking (viability: 7): https://sciencetostartup.com/paper/sat-balancing-reasoning-accuracy-and-efficiency-with-stepwise-adaptive-thinking - A framework that adaptively prunes reasoning steps in large language models to reduce token usage without sacrificing accuracy.
- Mitigating Entangled Steering in Large Vision-Language Models for Hallucination Reduction (viability: 7): https://sciencetostartup.com/paper/mitigating-entangled-steering-in-large-vision-language-models-for-hallucination-reduction - A plug-and-play framework for vision-language models that reduces hallucinations by selectively intervening in latent space without altering generation behavior.
- Dynamic Attentional Context Scoping: Agent-Triggered Focus Sessions for Isolated Per-Agent Steering in Multi-Agent LLM Orchestration (viability: 6): https://sciencetostartup.com/paper/dynamic-attentional-context-scoping-agent-triggered-focus-sessions-for-isolated-per-agent-steering-in-multi-agent-llm-or - A novel context management system for multi-agent LLM orchestration that isolates agent contexts to prevent pollution and improve decision quality.
- Capture-Quiet Decomposition: A Verification Theorem for Chess Endgame Tablebases (viability: 3): https://sciencetostartup.com/paper/capture-quiet-decomposition-a-verification-theorem-for-chess-endgame-tablebases - A structural theorem for verifying chess endgame tablebases by decomposing positions into terminal, capture, or quiet categories.
- AnomalyAgent: Agentic Industrial Anomaly Synthesis via Tool-Augmented Reinforcement Learning (viability: 8): https://sciencetostartup.com/paper/anomalyagent-agentic-industrial-anomaly-synthesis-via-tool-augmented-reinforcement-learning - An agentic system that synthesizes realistic industrial anomalies for data augmentation, improving anomaly detection.
- Visual Perceptual to Conceptual First-Order Rule Learning Networks (viability: 7): https://sciencetostartup.com/paper/visual-perceptual-to-conceptual-first-order-rule-learning-networks - A differentiable framework for learning symbolic rules directly from image data, enabling explainable AI and enhanced reasoning.
- DialBGM: A Benchmark for Background Music Recommendation from Everyday Multi-Turn Dialogues (viability: 7): https://sciencetostartup.com/paper/dialbgm-a-benchmark-for-background-music-recommendation-from-everyday-multi-turn-dialogues - A benchmark and evaluation framework for dialogue-conditioned background music recommendation, addressing a gap in media production.
- TSUBASA: Improving Long-Horizon Personalization via Evolving Memory and Self-Learning with Context Distillation (viability: 8): https://sciencetostartup.com/paper/tsubasa-improving-long-horizon-personalization-via-evolving-memory-and-self-learning-with-context-distillation - TSUBASA enhances personalized LLMs for long-horizon tasks by evolving memory and self-learning with context distillation.
- Data Selection for Multi-turn Dialogue Instruction Tuning (viability: 8): https://sciencetostartup.com/paper/data-selection-for-multi-turn-dialogue-instruction-tuning - MDS is a dialogue-level framework that enhances instruction-tuned language models by selecting high-quality multi-turn dialogues.
- Reinforcement-Guided Synthetic Data Generation for Privacy-Sensitive Identity Recognition (viability: 7): https://sciencetostartup.com/paper/reinforcement-guided-synthetic-data-generation-for-privacy-sensitive-identity-recognition - A reinforcement-guided framework generates high-fidelity synthetic data for privacy-sensitive identity recognition tasks.
- An Agentic Evaluation Architecture for Historical Bias Detection in Educational Textbooks (viability: 6): https://sciencetostartup.com/paper/an-agentic-evaluation-architecture-for-historical-bias-detection-in-educational-textbooks - An agentic evaluation architecture detects biases in educational textbooks using a multimodal screening approach.
- FlowGuard: Towards Lightweight In-Generation Safety Detection for Diffusion Models via Linear Latent Decoding (viability: 8): https://sciencetostartup.com/paper/flowguard-towards-lightweight-in-generation-safety-detection-for-diffusion-models-via-linear-latent-decoding - FlowGuard is a lightweight framework for in-generation safety detection in diffusion models, reducing computational costs significantly.
- PyVRP$^+$: LLM-Driven Metacognitive Heuristic Evolution for Hybrid Genetic Search in Vehicle Routing Problems (viability: 5): https://sciencetostartup.com/paper/pyvrp-llm-driven-metacognitive-heuristic-evolution-for-hybrid-genetic-search-in-vehicle-routing-problems - A novel framework using LLMs to enhance metaheuristic search for vehicle routing problems.
- Task-Adaptive Retrieval over Agentic Multi-Modal Web Histories via Learned Graph Memory (viability: 7): https://sciencetostartup.com/paper/task-adaptive-retrieval-over-agentic-multi-modal-web-histories-via-learned-graph-memory - A task-adaptive retrieval system leveraging learned graph memory for improved web interaction history relevance.
- Networking-Aware Energy Efficiency in Agentic AI Inference: A Survey (viability: 2): https://sciencetostartup.com/paper/networking-aware-energy-efficiency-in-agentic-ai-inference-a-survey - A survey on energy efficiency challenges in agentic AI inference.
- Hidden Biases in Conditioning Autoregressive Models (viability: 2): https://sciencetostartup.com/paper/hidden-biases-in-conditioning-autoregressive-models - An exploration of hidden biases in autoregressive models for constrained generation.
- QaRL: Rollout-Aligned Quantization-Aware RL for Fast and Stable Training under Training--Inference Mismatch (viability: 3): https://sciencetostartup.com/paper/qarl-rollout-aligned-quantization-aware-rl-for-fast-and-stable-training-under-training-inference-mismatch - A novel RL framework for LLM training that aligns training with quantized rollouts to improve speed and stability.
- ReRec: Reasoning-Augmented LLM-based Recommendation Assistant via Reinforcement Fine-tuning (viability: 6): https://sciencetostartup.com/paper/rerec-reasoning-augmented-llm-based-recommendation-assistant-via-reinforcement-fine-tuning - A reinforcement fine-tuning framework that enhances LLM reasoning for complex recommendation tasks.
- SPARD: Self-Paced Curriculum for RL Alignment via Integrating Reward Dynamics and Data Utility (viability: 3): https://sciencetostartup.com/paper/spard-self-paced-curriculum-for-rl-alignment-via-integrating-reward-dynamics-and-data-utility - A self-paced curriculum framework that dynamically adjusts reward weights and data importance for LLM alignment.
- Silencing the Guardrails: Inference-Time Jailbreaking via Dynamic Contextual Representation Ablation (viability: 6): https://sciencetostartup.com/paper/silencing-the-guardrails-inference-time-jailbreaking-via-dynamic-contextual-representation-ablation - An inference-time intervention framework that dynamically silences LLM guardrails by ablating contextual representations.
- Filling the Gaps: Selective Knowledge Augmentation for LLM Recommenders (viability: 7): https://sciencetostartup.com/paper/filling-the-gaps-selective-knowledge-augmentation-for-llm-recommenders - A novel method to improve LLM-based recommenders by selectively augmenting item knowledge, boosting accuracy and context efficiency without fine-tuning.
- LPM 1.0: Video-based Character Performance Model (viability: 7): https://sciencetostartup.com/paper/lpm-1-0-video-based-character-performance-model - LPM 1.0 offers a video-based character performance model for interactive conversations, minimizing conventional 3D pipeline complexities.
- Loop, Think, & Generalize: Implicit Reasoning in Recurrent-Depth Transformers (viability: 3): https://sciencetostartup.com/paper/loop-think-generalize-implicit-reasoning-in-recurrent-depth-transformers - Recurrent-depth transformers show potential for improved implicit reasoning and compositional generalization in LLMs, addressing limitations in systematic generalization and depth extrapolation.
- More Capable, Less Cooperative? When LLMs Fail At Zero-Cost Collaboration (viability: 3): https://sciencetostartup.com/paper/more-capable-less-cooperative-when-llms-fail-at-zero-cost-collaboration - LLM agents exhibit cooperation failures in frictionless environments, indicating that scaling intelligence alone is insufficient for multi-agent coordination.
- Automatic Generation of Executable BPMN Models from Medical Guidelines (viability: 5): https://sciencetostartup.com/paper/automatic-generation-of-executable-bpmn-models-from-medical-guidelines - Automate the conversion of complex medical guidelines into executable simulation models for policy evaluation.
- Agentivism: a learning theory for the age of artificial intelligence (viability: 0): https://sciencetostartup.com/paper/agentivism-a-learning-theory-for-the-age-of-artificial-intelligence - A new learning theory to understand how humans learn effectively in the age of generative AI.
- PolicyLong: Towards On-Policy Context Extension (viability: 4): https://sciencetostartup.com/paper/policylong-towards-on-policy-context-extension - Dynamically generate high-quality long-context data for LLMs by aligning data construction with model evolution.
- The Weaponization of Computer Vision: Tracing Military-Surveillance Ties through Conference Sponsorship (viability: 4): https://sciencetostartup.com/paper/the-weaponization-of-computer-vision-tracing-military-surveillance-ties-through-conference-sponsorship - Uncover the military and surveillance ties within computer vision research by analyzing conference sponsorship data.
- Latent Anomaly Knowledge Excavation: Unveiling Sparse Sensitive Neurons in Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/latent-anomaly-knowledge-excavation-unveiling-sparse-sensitive-neurons-in-vision-language-models - A training-free framework that identifies and elicits latent anomaly-sensitive neurons in vision-language models to achieve state-of-the-art anomaly detection with neuron-level interpretability.
- TEMPER: Testing Emotional Perturbation in Quantitative Reasoning (viability: 7): https://sciencetostartup.com/paper/temper-testing-emotional-perturbation-in-quantitative-reasoning - A framework for testing and mitigating the impact of emotional framing on LLM quantitative reasoning, demonstrating accuracy degradation and proposing neutralization as a solution.
- Learning Without Losing Identity: Capability Evolution for Embodied Agents (viability: 7): https://sciencetostartup.com/paper/learning-without-losing-identity-capability-evolution-for-embodied-agents - A capability-centric evolution paradigm for embodied agents that decouples capability learning from agent identity, enabling continuous improvement without loss of stability.
- Lightweight LLM Agent Memory with Small Language Models (viability: 4): https://sciencetostartup.com/paper/lightweight-llm-agent-memory-with-small-language-models - LightMem is a lightweight memory system for LLM agents that uses Small Language Models to modularize memory operations, improving accuracy and reducing latency.
- SEARL: Joint Optimization of Policy and Tool Graph Memory for Self-Evolving Agents (viability: 7): https://sciencetostartup.com/paper/searl-joint-optimization-of-policy-and-tool-graph-memory-for-self-evolving-agents - A framework for self-evolving agents that optimizes policy and tool graph memory for more practical and efficient learning in resource-constrained environments.
- Automotive Engineering-Centric Agentic AI Workflow Framework (viability: 3): https://sciencetostartup.com/paper/automotive-engineering-centric-agentic-ai-workflow-framework - An industrial vision framework that models automotive engineering workflows as constrained, history-aware sequential decision processes for AI agent support.
- Toward Generalizable Graph Learning for 3D Engineering AI: Explainable Workflows for CAE Mode Shape Classification and CFD Field Prediction (viability: 7): https://sciencetostartup.com/paper/toward-generalizable-graph-learning-for-3d-engineering-ai-explainable-workflows-for-cae-mode-shape-classification-and-cf - A graph learning framework for 3D engineering AI that converts heterogeneous data into physics-aware graph representations for explainable CAE and CFD decision support.
- The Accountability Horizon: An Impossibility Theorem for Governing Human-Agent Collectives (viability: 3): https://sciencetostartup.com/paper/the-accountability-horizon-an-impossibility-theorem-for-governing-human-agent-collectives - An impossibility theorem proving that agentic AI systems violate accountability assumptions once autonomy exceeds a computable threshold, necessitating distributed accountability mechanisms.
- ACIArena: Toward Unified Evaluation for Agent Cascading Injection (viability: 4): https://sciencetostartup.com/paper/aciarena-toward-unified-evaluation-for-agent-cascading-injection - A unified framework for evaluating the security of multi-agent systems against cascading injection attacks, providing a benchmark and insights into robust design.
- Sensitivity-Positional Co-Localization in GQA Transformers (viability: 7): https://sciencetostartup.com/paper/sensitivity-positional-co-localization-in-gqa-transformers - A novel method for efficient LLM fine-tuning that decouples sensitivity and positional encoding adaptation, achieving strong performance across benchmarks with reduced compute.
- Beyond Surface Artifacts: Capturing Shared Latent Forgery Knowledge Across Modalities (viability: 7): https://sciencetostartup.com/paper/beyond-surface-artifacts-capturing-shared-latent-forgery-knowledge-across-modalities - A modality-agnostic framework for deepfake detection that captures shared latent forgery knowledge, enabling generalization to unseen modalities.
- DailyArt: Discovering Articulation from Single Static Images via Latent Dynamics (viability: 7): https://sciencetostartup.com/paper/dailyart-discovering-articulation-from-single-static-images-via-latent-dynamics - A novel approach to infer articulated object kinematics from single static images by synthesizing an opened state to reveal hidden motion cues.
- MIMIC-Py: An Extensible Tool for Personality-Driven Automated Game Testing with Large Language Models (viability: 7): https://sciencetostartup.com/paper/mimic-py-an-extensible-tool-for-personality-driven-automated-game-testing-with-large-language-models - MIMIC-Py is a reusable framework for personality-driven LLM agents to automate game testing, bridging research prototypes and practical applications.
- Mitigating Distribution Sharpening in Math RLVR via Distribution-Aligned Hint Synthesis and Backward Hint Annealing (viability: 5): https://sciencetostartup.com/paper/mitigating-distribution-sharpening-in-math-rlvr-via-distribution-aligned-hint-synthesis-and-backward-hint-annealing - This work introduces a novel hint synthesis and annealing method to improve reasoning accuracy and coverage in math RLVR for LLMs.
- The Cartesian Cut in Agentic AI (viability: 1): https://sciencetostartup.com/paper/the-cartesian-cut-in-agentic-ai - This paper proposes a theoretical framework for understanding control in agentic AI systems, contrasting Cartesian agency with integrated approaches.
- Beyond Pedestrians: Caption-Guided CLIP Framework for High-Difficulty Video-based Person Re-Identification (viability: 7): https://sciencetostartup.com/paper/beyond-pedestrians-caption-guided-clip-framework-for-high-difficulty-video-based-person-re-identification - CG-CLIP is a caption-guided framework for high-difficulty video person re-identification, outperforming state-of-the-art in challenging scenarios.
- CivBench: Progress-Based Evaluation for LLMs' Strategic Decision-Making in Civilization V (viability: 4): https://sciencetostartup.com/paper/civbench-progress-based-evaluation-for-llms-strategic-decision-making-in-civilization-v - A benchmark for evaluating LLM strategic decision-making in complex, multi-agent games like Civilization V, providing richer signals than simple win/loss outcomes.
- Emotion Concepts and their Function in a Large Language Model (viability: 0): https://sciencetostartup.com/paper/emotion-concepts-and-their-function-in-a-large-language-model - Investigates the internal representations of emotion concepts in LLMs and how they causally influence model behavior, including alignment-relevant actions.
- TrajGuard: Streaming Hidden-state Trajectory Detection for Decoding-time Jailbreak Defense (viability: 7): https://sciencetostartup.com/paper/trajguard-streaming-hidden-state-trajectory-detection-for-decoding-time-jailbreak-defense - A real-time, training-free defense system that detects LLM jailbreaks by analyzing hidden state trajectories during decoding, achieving high accuracy with low latency.
- Squeeze Evolve: Unified Multi-Model Orchestration for Verifier-Free Evolution (viability: 7): https://sciencetostartup.com/paper/squeeze-evolve-unified-multi-model-orchestration-for-verifier-free-evolution - A framework that intelligently orchestrates multiple LLMs of varying costs to optimize performance and efficiency for verifier-free evolutionary inference.
- Towards Knowledgeable Deep Research: Framework and Benchmark (viability: 7): https://sciencetostartup.com/paper/towards-knowledgeable-deep-research-framework-and-benchmark - A framework and benchmark for LLM agents to perform deep research using both structured and unstructured knowledge, generating multimodal reports.
- Joint Task Offloading, Inference Optimization and UAV Trajectory Planning for Generative AI Empowered Intelligent Transportation Digital Twin (viability: 7): https://sciencetostartup.com/paper/joint-task-offloading-inference-optimization-and-uav-trajectory-planning-for-generative-ai-empowered-intelligent-transpo - Optimizing UAV-based generative AI inference for intelligent transportation digital twins to maximize system utility and minimize delay.
- An Imperfect Verifier is Good Enough: Learning with Noisy Rewards (viability: 4): https://sciencetostartup.com/paper/an-imperfect-verifier-is-good-enough-learning-with-noisy-rewards - This research demonstrates that imperfect reward verification in Reinforcement Learning is sufficient for effective LLM training, suggesting a more practical approach to RLVR.
- How Independent are Large Language Models? A Statistical Framework for Auditing Behavioral Entanglement and Reweighting Verifier Ensembles (viability: 7): https://sciencetostartup.com/paper/how-independent-are-large-language-models-a-statistical-framework-for-auditing-behavioral-entanglement-and-reweighting-v - A statistical framework to audit and mitigate behavioral entanglement in large language models, improving ensemble verification accuracy.
- PRIME: Training Free Proactive Reasoning via Iterative Memory Evolution for User-Centric Agent (viability: 7): https://sciencetostartup.com/paper/prime-training-free-proactive-reasoning-via-iterative-memory-evolution-for-user-centric-agent - PRIME is a gradient-free framework for proactive, collaborative agents that learn from human-AI interactions without expensive training.
- Safe Large-Scale Robust Nonlinear MPC in Milliseconds via Reachability-Constrained System Level Synthesis on the GPU (viability: 8): https://sciencetostartup.com/paper/safe-large-scale-robust-nonlinear-mpc-in-milliseconds-via-reachability-constrained-system-level-synthesis-on-the-gpu - GPU-accelerated framework for safe, robust nonlinear model predictive control that achieves real-time performance on high-dimensional robotic systems.
- Exponential quantum advantage in processing massive classical data (viability: 3): https://sciencetostartup.com/paper/exponential-quantum-advantage-in-processing-massive-classical-data - Demonstrates exponential quantum advantage in processing massive classical data for classification and dimension reduction using quantum oracle sketching.
- Sheaf-Laplacian Obstruction and Projection Hardness for Cross-Modal Compatibility on a Modality-Independent Site (viability: 1): https://sciencetostartup.com/paper/sheaf-laplacian-obstruction-and-projection-hardness-for-cross-modal-compatibility-on-a-modality-independent-site - A theoretical framework for analyzing cross-modal compatibility using sheaf theory and spectral gaps.
- DIVERSED: Relaxed Speculative Decoding via Dynamic Ensemble Verification (viability: 7): https://sciencetostartup.com/paper/diversed-relaxed-speculative-decoding-via-dynamic-ensemble-verification - Accelerate LLM inference by relaxing speculative decoding's rigid verification step with a dynamic ensemble.
- Towards Real-Time Human-AI Musical Co-Performance: Accompaniment Generation with Latent Diffusion Models and MAX/MSP (viability: 7): https://sciencetostartup.com/paper/towards-real-time-human-ai-musical-co-performance-accompaniment-generation-with-latent-diffusion-models-and-max-msp - Enable real-time human-AI musical co-performance with latent diffusion models generating accompaniment via MAX/MSP.
- Google, AI Literacy, and the Learning Sciences: Multiple Modes of Research, Industry, and Practice Partnerships (viability: 1): https://sciencetostartup.com/paper/google-ai-literacy-and-the-learning-sciences-multiple-modes-of-research-industry-and-practice-partnerships - Exploring multi-stakeholder partnerships between research, industry, and practice to advance AI literacy.
- Reasoning Graphs: Deterministic Agent Accuracy through Evidence-Centric Chain-of-Thought Feedback (viability: 7): https://sciencetostartup.com/paper/reasoning-graphs-deterministic-agent-accuracy-through-evidence-centric-chain-of-thought-feedback - This paper introduces reasoning graphs, a novel memory mechanism for language model agents that persists and reuses evidence-centric chains of thought to deterministically improve accuracy and reduce variance without retraining.
- Too long; didn't solve (viability: 3): https://sciencetostartup.com/paper/too-long-didn-t-solve - This research investigates the impact of prompt and solution length on large language model performance in mathematical reasoning tasks, finding that longer inputs correlate with increased model failure.
- From Ground Truth to Measurement: A Statistical Framework for Human Labeling (viability: 4): https://sciencetostartup.com/paper/from-ground-truth-to-measurement-a-statistical-framework-for-human-labeling - This paper proposes a statistical framework to decompose human labeling variation into interpretable sources like instance difficulty, annotator bias, and situational noise, moving beyond treating all disagreement as noise.
- DCD: Domain-Oriented Design for Controlled Retrieval-Augmented Generation (viability: 7): https://sciencetostartup.com/paper/dcd-domain-oriented-design-for-controlled-retrieval-augmented-generation - DCD is a domain-oriented design for controlled RAG that structures knowledge hierarchically and uses multi-stage routing to improve factual accuracy and answer relevance without modifying the base LLM.
- Don't Measure Once: Measuring Visibility in AI Search (GEO) (viability: 1): https://sciencetostartup.com/paper/don-t-measure-once-measuring-visibility-in-ai-search-geo - This paper proposes a new method for measuring visibility in AI search by accounting for the probabilistic nature of generative search results.
- From Papers to Property Tables: A Priority-Based LLM Workflow for Materials Data Extraction (viability: 7): https://sciencetostartup.com/paper/from-papers-to-property-tables-a-priority-based-llm-workflow-for-materials-data-extraction - An LLM-powered workflow automatically extracts and reconstructs structured materials data from research articles, enabling scalable database construction.
- Learning is Forgetting: LLM Training As Lossy Compression (viability: 2): https://sciencetostartup.com/paper/learning-is-forgetting-llm-training-as-lossy-compression - This paper frames LLM training as lossy compression, linking representational structure to downstream performance through an information-theoretic lens.
- Reasoning-Based Refinement of Unsupervised Text Clusters with LLMs (viability: 7): https://sciencetostartup.com/paper/reasoning-based-refinement-of-unsupervised-text-clusters-with-llms - A reasoning-based framework uses LLMs to refine unsupervised text clusters, improving coherence and interpretability without supervision.
- Dual-Loop Control in DCVerse: Advancing Reliable Deployment of AI in Data Centers via Digital Twins (viability: 4): https://sciencetostartup.com/paper/dual-loop-control-in-dcverse-advancing-reliable-deployment-of-ai-in-data-centers-via-digital-twins - A digital twin framework for reliable AI control in data centers, improving energy efficiency and reducing outage risk.
- Generative Experiences for Digital Mental Health Interventions: Evidence from a Randomized Study (viability: 5): https://sciencetostartup.com/paper/generative-experiences-for-digital-mental-health-interventions-evidence-from-a-randomized-study - A system that generates personalized mental health intervention experiences at runtime, reducing stress and improving user experience.
- TR-EduVSum: A Turkish-Focused Dataset and Consensus Framework for Educational Video Summarization (viability: 7): https://sciencetostartup.com/paper/tr-eduvsum-a-turkish-focused-dataset-and-consensus-framework-for-educational-video-summarization - A framework and dataset for Turkish educational video summarization that automatically generates gold-standard summaries based on human consensus.
- MCP-DPT: A Defense-Placement Taxonomy and Coverage Analysis for Model Context Protocol Security (viability: 6): https://sciencetostartup.com/paper/mcp-dpt-a-defense-placement-taxonomy-and-coverage-analysis-for-model-context-protocol-security - A taxonomy and analysis of defense placement for Model Context Protocol security in LLM agents, identifying gaps in current mitigation strategies.
- EMSDialog: Synthetic Multi-person Emergency Medical Service Dialogue Generation from Electronic Patient Care Reports via Multi-LLM Agents (viability: 7): https://sciencetostartup.com/paper/emsdialog-synthetic-multi-person-emergency-medical-service-dialogue-generation-from-electronic-patient-care-reports-via - A multi-LLM agent system generates synthetic medical dialogues for training diagnostic prediction models, creating a valuable dataset and improving model performance.
- Agentic Copyright, Data Scraping & AI Governance: Toward a Coasean Bargain in the Era of Artificial Intelligence (viability: 2): https://sciencetostartup.com/paper/agentic-copyright-data-scraping-ai-governance-toward-a-coasean-bargain-in-the-era-of-artificial-intelligence - This paper proposes a theoretical framework for agentic copyright and a supervised multi-agent governance model to address market failures in AI-driven creative industries.
- Trust the AI, Doubt Yourself: The Effect of Urgency on Self-Confidence in Human-AI Interaction (viability: 1): https://sciencetostartup.com/paper/trust-the-ai-doubt-yourself-the-effect-of-urgency-on-self-confidence-in-human-ai-interaction - Urgency in human-AI interactions, while not affecting trust in AI, can negatively impact human self-confidence and lead to performance degradation.
- RL-ASL: A Dynamic Listening Optimization for TSCH Networks Using Reinforcement Learning (viability: 6): https://sciencetostartup.com/paper/rl-asl-a-dynamic-listening-optimization-for-tsch-networks-using-reinforcement-learning - A reinforcement learning framework dynamically optimizes listening slots in TSCH networks to significantly reduce power consumption and latency in IIoT devices.
- The Shrinking Lifespan of LLMs in Science (viability: 2): https://sciencetostartup.com/paper/the-shrinking-lifespan-of-llms-in-science - This paper analyzes the adoption and abandonment trends of large language models in scientific research, revealing a compressing lifespan for models over time.
- SYN-DIGITS: A Synthetic Control Framework for Calibrated Digital Twin Simulation (viability: 7): https://sciencetostartup.com/paper/syn-digits-a-synthetic-control-framework-for-calibrated-digital-twin-simulation - SYN-DIGITS is a model-agnostic framework that calibrates AI persona simulations to align with real human behavior, improving reliability for market research and social sciences.
- Rhizome OS-1: Rhizome's Semi-Autonomous Operating System for Small Molecule Drug Discovery (viability: 8): https://sciencetostartup.com/paper/rhizome-os-1-rhizome-s-semi-autonomous-operating-system-for-small-molecule-drug-discovery - Rhizome OS-1 is a semi-autonomous operating system for small molecule drug discovery, leveraging multi-modal AI agents and graph neural networks to generate novel chemical matter.
- ReflectRM: Boosting Generative Reward Models via Self-Reflection within a Unified Judgment Framework (viability: 7): https://sciencetostartup.com/paper/reflectrm-boosting-generative-reward-models-via-self-reflection-within-a-unified-judgment-framework - ReflectRM enhances Generative Reward Models by incorporating self-reflection to improve preference modeling and mitigate bias in LLM alignment.
- Beyond Human-Readable: Rethinking Software Engineering Conventions for the Agentic Development Era (viability: 7): https://sciencetostartup.com/paper/beyond-human-readable-rethinking-software-engineering-conventions-for-the-agentic-development-era - Optimize software engineering conventions for AI agents by prioritizing semantic density over human readability to improve agent efficiency and reduce costs.
- Triage: Routing Software Engineering Tasks to Cost-Effective LLM Tiers via Code Quality Signals (viability: 4): https://sciencetostartup.com/paper/triage-routing-software-engineering-tasks-to-cost-effective-llm-tiers-via-code-quality-signals - Route software engineering tasks to cost-effective LLM tiers using code quality signals to reduce inference costs without sacrificing output quality.
- Cluster Attention for Graph Machine Learning (viability: 7): https://sciencetostartup.com/paper/cluster-attention-for-graph-machine-learning - Enhance graph machine learning models with cluster attention to significantly improve performance on graph datasets by increasing receptive fields while preserving graph structure.
- Enabling Intrinsic Reasoning over Dense Geospatial Embeddings with DFR-Gemma (viability: 8): https://sciencetostartup.com/paper/enabling-intrinsic-reasoning-over-dense-geospatial-embeddings-with-dfr-gemma - Enable LLMs to directly reason over dense geospatial embeddings with DFR-Gemma, offering a more efficient and accurate approach to multimodal geospatial intelligence.
- CLEAR: Context Augmentation from Contrastive Learning of Experience via Agentic Reflection (viability: 8): https://sciencetostartup.com/paper/clear-context-augmentation-from-contrastive-learning-of-experience-via-agentic-reflection - A framework that trains LLM agents to generate task-specific context, improving decision-making and task completion rates.
- Private Seeds, Public LLMs: Realistic and Privacy-Preserving Synthetic Data Generation (viability: 7): https://sciencetostartup.com/paper/private-seeds-public-llms-realistic-and-privacy-preserving-synthetic-data-generation - A privacy-preserving method for generating realistic synthetic text data from private sources using differential privacy and private seeds.
- ConsistRM: Improving Generative Reward Models via Consistency-Aware Self-Training (viability: 8): https://sciencetostartup.com/paper/consistrm-improving-generative-reward-models-via-consistency-aware-self-training - A self-training framework for generative reward models that improves LLM alignment by ensuring consistency in generated critiques and rewards.
- Active Reward Machine Inference From Raw State Trajectories (viability: 3): https://sciencetostartup.com/paper/active-reward-machine-inference-from-raw-state-trajectories - A theoretical framework for learning reward machines from raw state trajectories without explicit reward or label observations.
- When Switching Algorithms Helps: A Theoretical Study of Online Algorithm Selection (viability: 1): https://sciencetostartup.com/paper/when-switching-algorithms-helps-a-theoretical-study-of-online-algorithm-selection - A theoretical study exploring how switching between optimization algorithms can lead to faster problem-solving in specific scenarios.
- M-ArtAgent: Evidence-Based Multimodal Agent for Implicit Art Influence Discovery (viability: 7): https://sciencetostartup.com/paper/m-artagent-evidence-based-multimodal-agent-for-implicit-art-influence-discovery - An AI agent that uses multimodal evidence and art historical axioms to discover and verify implicit artistic influences.
- CMP: Robust Whole-Body Tracking for Loco-Manipulation via Competence Manifold Projection (viability: 7): https://sciencetostartup.com/paper/cmp-robust-whole-body-tracking-for-loco-manipulation-via-competence-manifold-projection - A robust whole-body tracking system for robots that projects control policies onto a competence manifold to handle unexpected inputs.
- Munkres' General Topology Autoformalized in Isabelle/HOL (viability: 3): https://sciencetostartup.com/paper/munkres-general-topology-autoformalized-in-isabelle-hol - An experiment demonstrating LLM-assisted autoformalization of a comprehensive general topology textbook into a formal proof system.
- Toward a Tractability Frontier for Exact Relevance Certification (viability: 1): https://sciencetostartup.com/paper/toward-a-tractability-frontier-for-exact-relevance-certification - This paper explores theoretical limits in exact relevance certification for decision problems, offering no immediate product application.
- MoRight: Motion Control Done Right (viability: 8): https://sciencetostartup.com/paper/moright-motion-control-done-right - MoRight revolutionizes motion control with a dual-stream architecture for disentangled camera-object motion generation.
- RoSHI: A Versatile Robot-oriented Suit for Human Data In-the-Wild (viability: 7): https://sciencetostartup.com/paper/roshi-a-versatile-robot-oriented-suit-for-human-data-in-the-wild - RoSHI is a wearable system that fuses IMUs and egocentric vision to capture metric 3D human pose and body shape for robot learning.
- Syntax Is Easy, Semantics Is Hard: Evaluating LLMs for LTL Translation (viability: 5): https://sciencetostartup.com/paper/syntax-is-easy-semantics-is-hard-evaluating-llms-for-ltl-translation - This paper evaluates LLMs for translating natural language into LTL formulas, finding that prompt engineering and code completion reformulations improve performance.
- Evaluating In-Context Translation with Synchronous Context-Free Grammar Transduction (viability: 2): https://sciencetostartup.com/paper/evaluating-in-context-translation-with-synchronous-context-free-grammar-transduction - Evaluating LLMs' ability to perform in-context translation for low-resource languages by formalizing the task with synchronous context-free grammars.
- Chatbot-Based Assessment of Code Understanding in Automated Programming Assessment Systems (viability: 7): https://sciencetostartup.com/paper/chatbot-based-assessment-of-code-understanding-in-automated-programming-assessment-systems - A hybrid framework integrating conversational AI with automated programming assessment to verify student code understanding, not just correctness.
- Region-Graph Optimal Transport Routing for Mixture-of-Experts Whole-Slide Image Classification (viability: 7): https://sciencetostartup.com/paper/region-graph-optimal-transport-routing-for-mixture-of-experts-whole-slide-image-classification - ROAM: A spatially aware Mixture-of-Experts aggregator for whole-slide image classification that uses optimal transport to balance expert utilization.
- CADENCE: Context-Adaptive Depth Estimation for Navigation and Computational Efficiency (viability: 7): https://sciencetostartup.com/paper/cadence-context-adaptive-depth-estimation-for-navigation-and-computational-efficiency - Optimize autonomous vehicle navigation with dynamic AI-based depth estimation for efficiency and precision.
- Android Coach: Improve Online Agentic Training Efficiency with Single State Multiple Actions (viability: 7): https://sciencetostartup.com/paper/android-coach-improve-online-agentic-training-efficiency-with-single-state-multiple-actions - Android Coach enhances online agent training efficiency by enabling agents to explore multiple actions from a single state, significantly improving success rates and training speed.
- Making Room for AI: Multi-GPU Molecular Dynamics with Deep Potentials in GROMACS (viability: 4): https://sciencetostartup.com/paper/making-room-for-ai-multi-gpu-molecular-dynamics-with-deep-potentials-in-gromacs - Integrates AI-driven molecular dynamics potentials into GROMACS for high-performance, multi-GPU simulations, enabling near-quantum accuracy at scale.
- A Systematic Study of Retrieval Pipeline Design for Retrieval-Augmented Medical Question Answering (viability: 7): https://sciencetostartup.com/paper/a-systematic-study-of-retrieval-pipeline-design-for-retrieval-augmented-medical-question-answering - A systematic study of retrieval pipeline designs for medical question answering, demonstrating significant performance improvements with dense retrieval and query reformulation on modest computational resources.
- Validated Intent Compilation for Constrained Routing in LEO Mega-Constellations (viability: 3): https://sciencetostartup.com/paper/validated-intent-compilation-for-constrained-routing-in-leo-mega-constellations - Develop an AI-based routing system for LEO satellites that translates operator intents into network configurations.
- Designing Safe and Accountable GenAI as a Learning Companion with Women Banned from Formal Education (viability: 3): https://sciencetostartup.com/paper/designing-safe-and-accountable-genai-as-a-learning-companion-with-women-banned-from-formal-education - Designing safe and accountable GenAI companions for women banned from formal education to foster learning and employability.
- $k$-server-bench: Automating Potential Discovery for the $k$-Server Conjecture (viability: 6): https://sciencetostartup.com/paper/k-server-bench-automating-potential-discovery-for-the-k-server-conjecture - An automated benchmark for discovering mathematical potentials to solve the k-server conjecture, serving as a testbed for code-based discovery agents.
- How Much LLM Does a Self-Revising Agent Actually Need? (viability: 3): https://sciencetostartup.com/paper/how-much-llm-does-a-self-revising-agent-actually-need - Empirically dissecting the contribution of LLMs versus explicit structure in self-revising agents by externalizing agent state and decision-making.
- TraceSafe: A Systematic Assessment of LLM Guardrails on Multi-Step Tool-Calling Trajectories (viability: 5): https://sciencetostartup.com/paper/tracesafe-a-systematic-assessment-of-llm-guardrails-on-multi-step-tool-calling-trajectories - TraceSafe-Bench: A benchmark to systematically assess LLM guardrails on multi-step tool-calling trajectories, revealing structural competence as key to safety.
- Mixture Proportion Estimation and Weakly-supervised Kernel Test for Conditional Independence (viability: 3): https://sciencetostartup.com/paper/mixture-proportion-estimation-and-weakly-supervised-kernel-test-for-conditional-independence - Develops novel assumptions and estimators for mixture proportion estimation in weakly supervised learning, with applications in causal discovery and fairness.
- The ATOM Report: Measuring the Open Language Model Ecosystem (viability: 4): https://sciencetostartup.com/paper/the-atom-report-measuring-the-open-language-model-ecosystem - Analyzes the adoption and trends of open language models, highlighting the rise of Chinese models and their impact on the global ecosystem.
- TeaLeafVision: An Explainable and Robust Deep Learning Framework for Tea Leaf Disease Classification (viability: 7): https://sciencetostartup.com/paper/tealeafvision-an-explainable-and-robust-deep-learning-framework-for-tea-leaf-disease-classification - A robust and explainable deep learning framework for tea leaf disease classification using DenseNet201, achieving 99% accuracy and a real-world prototype.
- Energy-based Tissue Manifolds for Longitudinal Multiparametric MRI Analysis (viability: 6): https://sciencetostartup.com/paper/energy-based-tissue-manifolds-for-longitudinal-multiparametric-mri-analysis - Proposes a geometric framework for longitudinal MRI analysis using energy-based tissue manifolds, offering a novel approach to tracking disease progression.
- Reason in Chains, Learn in Trees: Self-Rectification and Grafting for Multi-turn Agent Policy Optimization (viability: 7): https://sciencetostartup.com/paper/reason-in-chains-learn-in-trees-self-rectification-and-grafting-for-multi-turn-agent-policy-optimization - A framework that optimizes multi-turn agent policies by recovering latent reward structures and synthesizing corrective reasoning through a cognitive tree.
- Bridging MRI and PET physiology: Untangling complementarity through orthogonal representations (viability: 7): https://sciencetostartup.com/paper/bridging-mri-and-pet-physiology-untangling-complementarity-through-orthogonal-representations - A subspace decomposition framework that untangles complementary information between MRI and PET imaging for improved prostate cancer characterization.
- Dynamic Context Evolution for Scalable Synthetic Data Generation (viability: 6): https://sciencetostartup.com/paper/dynamic-context-evolution-for-scalable-synthetic-data-generation - Dynamic Context Evolution (DCE) is a framework to prevent repetitive output and ensure diverse generations from large language models using existing APIs.
- Energy Saving for Cell-Free Massive MIMO Networks: A Multi-Agent Deep Reinforcement Learning Approach (viability: 4): https://sciencetostartup.com/paper/energy-saving-for-cell-free-massive-mimo-networks-a-multi-agent-deep-reinforcement-learning-approach - A multi-agent deep reinforcement learning approach for energy saving in cell-free massive MIMO networks.
- CSA-Graphs: A Privacy-Preserving Structural Dataset for Child Sexual Abuse Research (viability: 7): https://sciencetostartup.com/paper/csa-graphs-a-privacy-preserving-structural-dataset-for-child-sexual-abuse-research - A privacy-preserving structural dataset for child sexual abuse imagery classification, enabling broader research while respecting legal constraints.
- Self-Discovered Intention-aware Transformer for Multi-modal Vehicle Trajectory Prediction (viability: 4): https://sciencetostartup.com/paper/self-discovered-intention-aware-transformer-for-multi-modal-vehicle-trajectory-prediction - A Transformer-based network for multi-modal vehicle trajectory prediction that separates spatial and trajectory generation modules for improved performance.
- Mixed-Initiative Context: Structuring and Managing Context for Human-AI Collaboration (viability: 3): https://sciencetostartup.com/paper/mixed-initiative-context-structuring-and-managing-context-for-human-ai-collaboration - A conceptual framework and probe system for structuring and managing context in multi-turn human-AI interactions.
- Assessing the Added Value of Onboard Earth Observation Processing with the IRIDE HEO Service Segment (viability: 7): https://sciencetostartup.com/paper/assessing-the-added-value-of-onboard-earth-observation-processing-with-the-iride-heo-service-segment - Evaluating the added value of onboard Earth Observation processing for improved spatial detail, smaller detectable events, and faster response times.
- Information as Structural Alignment: A Dynamical Theory of Continual Learning (viability: 3): https://sciencetostartup.com/paper/information-as-structural-alignment-a-dynamical-theory-of-continual-learning - A new framework for continual learning that achieves replay-superior retention without storing raw data by treating information as structural alignment.
- The Impact of Steering Large Language Models with Persona Vectors in Educational Applications (viability: 5): https://sciencetostartup.com/paper/the-impact-of-steering-large-language-models-with-persona-vectors-in-educational-applications - Investigating the impact of persona steering on large language models for educational applications, revealing task-dependent sensitivities in generation and scoring.
- SurFITR: A Dataset for Surveillance Image Forgery Detection and Localisation (viability: 7): https://sciencetostartup.com/paper/surfitr-a-dataset-for-surveillance-image-forgery-detection-and-localisation - Introducing SurFITR, a novel dataset and benchmark for surveillance image forgery detection, addressing the limitations of existing models in subtle, localized tampering.
- STRIDE-ED: A Strategy-Grounded Stepwise Reasoning Framework for Empathetic Dialogue Systems (viability: 5): https://sciencetostartup.com/paper/stride-ed-a-strategy-grounded-stepwise-reasoning-framework-for-empathetic-dialogue-systems - STRIDE-ED is a strategy-grounded framework for empathetic dialogue systems that improves response generation through structured, strategy-conditioned reasoning.
- Flow Motion Policy: Manipulator Motion Planning with Flow Matching Models (viability: 7): https://sciencetostartup.com/paper/flow-motion-policy-manipulator-motion-planning-with-flow-matching-models - A neural motion planner for robotic manipulators that uses flow matching to generate multiple collision-free paths, improving planning success and efficiency.
- EVGeoQA: Benchmarking LLMs on Dynamic, Multi-Objective Geo-Spatial Exploration (viability: 7): https://sciencetostartup.com/paper/evgeoqa-benchmarking-llms-on-dynamic-multi-objective-geo-spatial-exploration - Develop a geo-spatial AI tool to optimize electric vehicle routing and charging decisions using dynamic data inputs.
- Planning Task Shielding: Detecting and Repairing Flaws in Planning Tasks through Turning them Unsolvable (viability: 0): https://sciencetostartup.com/paper/planning-task-shielding-detecting-and-repairing-flaws-in-planning-tasks-through-turning-them-unsolvable - A theoretical approach to modify planning tasks to ensure flawed states are never reached by making the task unsolvable.
- AV-SQL: Decomposing Complex Text-to-SQL Queries with Agentic Views (viability: 8): https://sciencetostartup.com/paper/av-sql-decomposing-complex-text-to-sql-queries-with-agentic-views - AV-SQL simplifies complex Text-to-SQL queries for large database schemas using a novel agentic view decomposition approach.
- AEROS: A Single-Agent Operating Architecture with Embodied Capability Modules (viability: 4): https://sciencetostartup.com/paper/aeros-a-single-agent-operating-architecture-with-embodied-capability-modules - AEROS provides a unified operating architecture for robotic intelligence, organizing capabilities and execution through modular packages for enhanced extensibility and safety.
- KITE: Keyframe-Indexed Tokenized Evidence for VLM-Based Robot Failure Analysis (viability: 7): https://sciencetostartup.com/paper/kite-keyframe-indexed-tokenized-evidence-for-vlm-based-robot-failure-analysis - KITE transforms robot execution videos into compact, interpretable evidence for VLMs, enabling training-free failure analysis and correction.
- Strategic Persuasion with Trait-Conditioned Multi-Agent Systems for Iterative Legal Argumentation (viability: 7): https://sciencetostartup.com/paper/strategic-persuasion-with-trait-conditioned-multi-agent-systems-for-iterative-legal-argumentation - A multi-agent simulation framework for strategic legal argumentation, enabling trait-conditioned LLM agents to adapt and persuade iteratively.
- ConceptTracer: Interactive Analysis of Concept Saliency and Selectivity in Neural Representations (viability: 7): https://sciencetostartup.com/paper/concepttracer-interactive-analysis-of-concept-saliency-and-selectivity-in-neural-representations - ConceptTracer is an interactive tool for analyzing concept saliency and selectivity in neural representations, aiding in the discovery of interpretable neurons.
- A-MBER: Affective Memory Benchmark for Emotion Recognition (viability: 7): https://sciencetostartup.com/paper/a-mber-affective-memory-benchmark-for-emotion-recognition - A benchmark for AI assistants to understand user emotions based on past interactions, enabling more personalized responses.
- CAFP: A Post-Processing Framework for Group Fairness via Counterfactual Model Averaging (viability: 7): https://sciencetostartup.com/paper/cafp-a-post-processing-framework-for-group-fairness-via-counterfactual-model-averaging - A model-agnostic post-processing method to ensure fairness in machine learning predictions by averaging outcomes across factual and counterfactual scenarios.
- AgentCity: Constitutional Governance for Autonomous Agent Economies via Separation of Power (viability: 3): https://sciencetostartup.com/paper/agentcity-constitutional-governance-for-autonomous-agent-economies-via-separation-of-power - A constitutional governance architecture for autonomous agent economies using blockchain to separate legislative, execution, and adjudication powers.
- EmoMAS: Emotion-Aware Multi-Agent System for High-Stakes Edge-Deployable Negotiation with Bayesian Orchestration (viability: 7): https://sciencetostartup.com/paper/emomas-emotion-aware-multi-agent-system-for-high-stakes-edge-deployable-negotiation-with-bayesian-orchestration - An emotion-aware multi-agent system for high-stakes negotiation on edge devices, using Bayesian orchestration to strategically manage emotional states.
- Self-Preference Bias in Rubric-Based Evaluation of Large Language Models (viability: 4): https://sciencetostartup.com/paper/self-preference-bias-in-rubric-based-evaluation-of-large-language-models - This research identifies and quantifies self-preference bias in LLM evaluations, proposing ensembling as a mitigation strategy for more reliable model development.
- What's Missing in Screen-to-Action? Towards a UI-in-the-Loop Paradigm for Multimodal GUI Reasoning (viability: 7): https://sciencetostartup.com/paper/what-s-missing-in-screen-to-action-towards-a-ui-in-the-loop-paradigm-for-multimodal-gui-reasoning - UILoop introduces a novel UI-in-the-Loop paradigm for multimodal GUI reasoning, enabling precise element discovery and interpretable decision-making with a new benchmark.
- Stress Estimation in Elderly Oncology Patients Using Visual Wearable Representations and Multi-Instance Learning (viability: 0): https://sciencetostartup.com/paper/stress-estimation-in-elderly-oncology-patients-using-visual-wearable-representations-and-multi-instance-learning - This paper explores stress estimation in elderly oncology patients using multimodal wearable data and multi-instance learning, showing moderate agreement with self-reported scores.
- Generative Phomosaic with Structure-Aligned and Personalized Diffusion (viability: 7): https://sciencetostartup.com/paper/generative-phomosaic-with-structure-aligned-and-personalized-diffusion - This research presents a generative approach to photomosaic creation using diffusion models for semantically expressive and structurally coherent compositions.
- CAAP: Capture-Aware Adversarial Patch Attacks on Palmprint Recognition Models (viability: 7): https://sciencetostartup.com/paper/caap-capture-aware-adversarial-patch-attacks-on-palmprint-recognition-models - A capture-aware adversarial patch framework for palmprint recognition that learns a universal patch effective under realistic acquisition variations.
- Frailty Estimation in Elderly Oncology Patients Using Multimodal Wearable Data and Multi-Instance Learning (viability: 4): https://sciencetostartup.com/paper/frailty-estimation-in-elderly-oncology-patients-using-multimodal-wearable-data-and-multi-instance-learning - A multimodal wearable framework using attention-based multi-instance learning to estimate frailty-related functional change in elderly oncology patients.
- An empirical study of LoRA-based fine-tuning of large language models for automated test case generation (viability: 8): https://sciencetostartup.com/paper/an-empirical-study-of-lora-based-fine-tuning-of-large-language-models-for-automated-test-case-generation - LoRA-based fine-tuning of open-source LLMs for automated test case generation, achieving performance comparable to proprietary models.
- A First Guess is Rarely the Final Answer: Learning to Search in the Travelling Salesperson Problem (viability: 4): https://sciencetostartup.com/paper/a-first-guess-is-rarely-the-final-answer-learning-to-search-in-the-travelling-salesperson-problem - NICO-TSP: A neural improvement framework for the Traveling Salesperson Problem that learns to search using 2-opt moves.
- Multi-modal user interface control detection using cross-attention (viability: 7): https://sciencetostartup.com/paper/multi-modal-user-interface-control-detection-using-cross-attention - A multi-modal AI that detects UI controls from screenshots by combining visual data with GPT-generated text descriptions, improving automated testing and accessibility.
- FP4 Explore, BF16 Train: Diffusion Reinforcement Learning via Efficient Rollout Scaling (viability: 6): https://sciencetostartup.com/paper/fp4-explore-bf16-train-diffusion-reinforcement-learning-via-efficient-rollout-scaling - A novel framework that enhances reinforcement learning for diffusion models through efficient rollout scaling.
- Q-Zoom: Query-Aware Adaptive Perception for Efficient Multimodal Large Language Models (viability: 7): https://sciencetostartup.com/paper/q-zoom-query-aware-adaptive-perception-for-efficient-multimodal-large-language-models - An adaptive perception framework that optimizes high-resolution input processing for multimodal language models.
- The AI Skills Shift: Mapping Skill Obsolescence, Emergence, and Transition Pathways in the LLM Era (viability: 8): https://sciencetostartup.com/paper/the-ai-skills-shift-mapping-skill-obsolescence-emergence-and-transition-pathways-in-the-llm-era - A framework that assesses skill obsolescence and emergence in the labor market due to AI advancements.
- XR-CareerAssist: An Immersive Platform for Personalised Career Guidance Leveraging Extended Reality and Multimodal AI (viability: 7): https://sciencetostartup.com/paper/xr-careerassist-an-immersive-platform-for-personalised-career-guidance-leveraging-extended-reality-and-multimodal-ai - An immersive XR platform leveraging multimodal AI for personalized career guidance, integrating speech recognition, translation, conversational AI, and visualisations.
- SentinelSphere: Integrating AI-Powered Real-Time Threat Detection with Cybersecurity Awareness Training (viability: 7): https://sciencetostartup.com/paper/sentinelsphere-integrating-ai-powered-real-time-threat-detection-with-cybersecurity-awareness-training - A cybersecurity platform integrating AI-driven threat detection with LLM-powered awareness training to address practitioner deficits and human-factor weaknesses.
- Do We Need Distinct Representations for Every Speech Token? Unveiling and Exploiting Redundancy in Large Speech Language Models (viability: 4): https://sciencetostartup.com/paper/do-we-need-distinct-representations-for-every-speech-token-unveiling-and-exploiting-redundancy-in-large-speech-language - A novel token merging mechanism for Large Speech Language Models that significantly reduces inference costs and memory usage without compromising accuracy.
- Physical Adversarial Attacks on AI Surveillance Systems:Detection, Tracking, and Visible--Infrared Evasion (viability: 3): https://sciencetostartup.com/paper/physical-adversarial-attacks-on-ai-surveillance-systems-detection-tracking-and-visible-infrared-evasion - A review of physical adversarial attacks on AI surveillance systems, focusing on temporal persistence, sensing modality, and carrier realism.
- Digital Skin, Digital Bias: Uncovering Tone-Based Biases in LLMs and Emoji Embeddings (viability: 7): https://sciencetostartup.com/paper/digital-skin-digital-bias-uncovering-tone-based-biases-in-llms-and-emoji-embeddings - This paper identifies and quantifies tone-based biases in LLMs and emoji embeddings, providing a framework for auditing and mitigating representational harms in online communication.
- MedDialBench: Benchmarking LLM Diagnostic Robustness under Parametric Adversarial Patient Behaviors (viability: 7): https://sciencetostartup.com/paper/meddialbench-benchmarking-llm-diagnostic-robustness-under-parametric-adversarial-patient-behaviors - MedDialBench is a novel benchmark for evaluating LLM diagnostic robustness against adversarial patient behaviors, revealing critical vulnerabilities in current models.
- HingeMem: Boundary Guided Long-Term Memory with Query Adaptive Retrieval for Scalable Dialogues (viability: 6): https://sciencetostartup.com/paper/hingemem-boundary-guided-long-term-memory-with-query-adaptive-retrieval-for-scalable-dialogues - HingeMem introduces a boundary-guided long-term memory system for dialogues with query-adaptive retrieval, improving efficiency and accuracy for extended interactions.
- Explaining Neural Networks in Preference Learning: a Post-hoc Inductive Logic Programming Approach (viability: 5): https://sciencetostartup.com/paper/explaining-neural-networks-in-preference-learning-a-post-hoc-inductive-logic-programming-approach - This paper proposes using Inductive Logic Programming (ILASP) to explain black-box neural networks in user preference learning, with a focus on dimensionality reduction for transparency.
- On the Step Length Confounding in LLM Reasoning Data Selection (viability: 5): https://sciencetostartup.com/paper/on-the-step-length-confounding-in-llm-reasoning-data-selection - A novel method to improve LLM reasoning data quality by mitigating step length confounding, leading to better performance on complex tasks.
- Beyond Surface Judgments: Human-Grounded Risk Evaluation of LLM-Generated Disinformation (viability: 3): https://sciencetostartup.com/paper/beyond-surface-judgments-human-grounded-risk-evaluation-of-llm-generated-disinformation - This paper analyzes the reliability of LLM judges for evaluating disinformation, finding they are harsher and misalign with human reader responses.
- OmniTabBench: Mapping the Empirical Frontiers of GBDTs, Neural Networks, and Foundation Models for Tabular Data at Scale (viability: 5): https://sciencetostartup.com/paper/omnitabbench-mapping-the-empirical-frontiers-of-gbdts-neural-networks-and-foundation-models-for-tabular-data-at-scale - OmniTabBench is a large-scale benchmark for tabular data, evaluating GBDTs, neural networks, and foundation models to guide model selection.
- SkillTrojan: Backdoor Attacks on Skill-Based Agent Systems (viability: 7): https://sciencetostartup.com/paper/skilltrojan-backdoor-attacks-on-skill-based-agent-systems - SkillTrojan introduces a novel backdoor attack on skill-based agent systems, demonstrating high attack success with minimal performance degradation.
- Riemann-Bench: A Benchmark for Moonshot Mathematics (viability: 7): https://sciencetostartup.com/paper/riemann-bench-a-benchmark-for-moonshot-mathematics - Riemann-Bench is a private benchmark for research-level mathematics, revealing a significant gap in current AI capabilities beyond competition math.
- MoBiE: Efficient Inference of Mixture of Binary Experts under Post-Training Quantization (viability: 7): https://sciencetostartup.com/paper/mobie-efficient-inference-of-mixture-of-binary-experts-under-post-training-quantization - A framework for efficiently quantizing Mixture-of-Experts LLMs, significantly improving inference speed and performance without additional storage.
- Instance-Adaptive Parametrization for Amortized Variational Inference (viability: 6): https://sciencetostartup.com/paper/instance-adaptive-parametrization-for-amortized-variational-inference - An instance-adaptive variational autoencoder that uses input-dependent modulations to improve posterior approximation accuracy and model efficiency.
- FedDAP: Domain-Aware Prototype Learning for Federated Learning under Domain Shift (viability: 7): https://sciencetostartup.com/paper/feddap-domain-aware-prototype-learning-for-federated-learning-under-domain-shift - A federated learning framework that constructs domain-specific prototypes to improve model generalization under domain shift.
- Evaluating Repository-level Software Documentation via Question Answering and Feature-Driven Development (viability: 5): https://sciencetostartup.com/paper/evaluating-repository-level-software-documentation-via-question-answering-and-feature-driven-development - A benchmark for evaluating repository-level software documentation using question answering and feature-driven development.
- FVD: Inference-Time Alignment of Diffusion Models via Fleming-Viot Resampling (viability: 4): https://sciencetostartup.com/paper/fvd-inference-time-alignment-of-diffusion-models-via-fleming-viot-resampling - A novel diffusion model sampler that improves image quality and generation speed by using population dynamics inspired by biological systems.
- Sparse-Aware Neural Networks for Nonlinear Functionals: Mitigating the Exponential Dependence on Dimension (viability: 0): https://sciencetostartup.com/paper/sparse-aware-neural-networks-for-nonlinear-functionals-mitigating-the-exponential-dependence-on-dimension - A theoretical framework using sparse neural networks to improve the efficiency and accuracy of learning complex mathematical functions.
- Multi-Faceted Self-Consistent Preference Alignment for Query Rewriting in Conversational Search (viability: 4): https://sciencetostartup.com/paper/multi-faceted-self-consistent-preference-alignment-for-query-rewriting-in-conversational-search - An AI model that improves conversational search by rewriting ambiguous queries based on feedback from retrieval and response generation.
- FlowExtract: Procedural Knowledge Extraction from Maintenance Flowcharts (viability: 7): https://sciencetostartup.com/paper/flowextract-procedural-knowledge-extraction-from-maintenance-flowcharts - A pipeline that extracts structured procedural knowledge from maintenance flowcharts, making them accessible to operator support systems.
- TeamLLM: A Human-Like Team-Oriented Collaboration Framework for Multi-Step Contextualized Tasks (viability: 7): https://sciencetostartup.com/paper/teamllm-a-human-like-team-oriented-collaboration-framework-for-multi-step-contextualized-tasks - TeamLLM is a human-like multi-LLM collaboration framework that improves performance on multi-step contextualized tasks by dividing roles and employing a three-phase collaboration.
- TurboAgent: An LLM-Driven Autonomous Multi-Agent Framework for Turbomachinery Aerodynamic Design (viability: 8): https://sciencetostartup.com/paper/turboagent-an-llm-driven-autonomous-multi-agent-framework-for-turbomachinery-aerodynamic-design - TurboAgent is an LLM-driven autonomous multi-agent framework that automates the complex turbomachinery aerodynamic design process from natural language requirements to final design generation.
- Evaluating LLM-Based 0-to-1 Software Generation in End-to-End CLI Tool Scenarios (viability: 7): https://sciencetostartup.com/paper/evaluating-llm-based-0-to-1-software-generation-in-end-to-end-cli-tool-scenarios - CLI-Tool-Bench is a new benchmark for evaluating LLM-based 0-to-1 software generation of CLI tools, revealing current limitations in generating complete and robust applications.
- Luwen Technical Report (viability: 7): https://sciencetostartup.com/paper/luwen-technical-report - Luwen is an open-source Chinese legal language model that adapts general-purpose LLMs to the legal domain through continual pre-training, instruction tuning, and retrieval-augmented generation.
- URMF: Uncertainty-aware Robust Multimodal Fusion for Multimodal Sarcasm Detection (viability: 7): https://sciencetostartup.com/paper/urmf-uncertainty-aware-robust-multimodal-fusion-for-multimodal-sarcasm-detection - A unified framework for multimodal sarcasm detection that models modality reliability to improve accuracy and robustness.
- The Traveling Thief Problem with Time Windows: Benchmarks and Heuristics (viability: 7): https://sciencetostartup.com/paper/the-traveling-thief-problem-with-time-windows-benchmarks-and-heuristics - A new heuristic approach for the Traveling Thief Problem with time window constraints, outperforming existing methods on benchmark instances.
- Fine-grained Approaches for Confidence Calibration of LLMs in Automated Code Revision (viability: 4): https://sciencetostartup.com/paper/fine-grained-approaches-for-confidence-calibration-of-llms-in-automated-code-revision - Fine-grained confidence calibration methods for LLMs in automated code revision to improve trustworthiness.
- HQF-Net: A Hybrid Quantum-Classical Multi-Scale Fusion Network for Remote Sensing Image Segmentation (viability: 7): https://sciencetostartup.com/paper/hqf-net-a-hybrid-quantum-classical-multi-scale-fusion-network-for-remote-sensing-image-segmentation - A hybrid quantum-classical network for remote sensing image segmentation that integrates multi-scale features and quantum-enhanced connections.
- Steering the Verifiability of Multimodal AI Hallucinations (viability: 7): https://sciencetostartup.com/paper/steering-the-verifiability-of-multimodal-ai-hallucinations - A novel dataset and intervention method to control the verifiability of AI hallucinations in multimodal applications.
- ATANT: An Evaluation Framework for AI Continuity (viability: 8): https://sciencetostartup.com/paper/atant-an-evaluation-framework-for-ai-continuity - An open evaluation framework for AI continuity, enabling robust testing and validation of memory systems.
- AgentGate: A Lightweight Structured Routing Engine for the Internet of Agents (viability: 7): https://sciencetostartup.com/paper/agentgate-a-lightweight-structured-routing-engine-for-the-internet-of-agents - AgentGate: a lightweight routing engine for efficient and privacy-aware dispatch in the Internet of Agents.
- Reasoning Fails Where Step Flow Breaks (viability: 3): https://sciencetostartup.com/paper/reasoning-fails-where-step-flow-breaks - A theoretical approach to improve LLM reasoning by analyzing and intervening in information flow failures.
- KD-MARL: Resource-Aware Knowledge Distillation in Multi-Agent Reinforcement Learning (viability: 4): https://sciencetostartup.com/paper/kd-marl-resource-aware-knowledge-distillation-in-multi-agent-reinforcement-learning - A framework for resource-aware knowledge distillation in multi-agent reinforcement learning to enable practical deployment on edge devices.
- ChemVLR: Prioritizing Reasoning in Perception for Chemical Vision-Language Understanding (viability: 7): https://sciencetostartup.com/paper/chemvlr-prioritizing-reasoning-in-perception-for-chemical-vision-language-understanding - ChemVLR is a chemical vision-language model that prioritizes reasoning in perception for interpretable chemical understanding, achieving state-of-the-art results.
- Between Century and Poet: Graph-Based Lexical Semantic Change in Persian Poetry (viability: 0): https://sciencetostartup.com/paper/between-century-and-poet-graph-based-lexical-semantic-change-in-persian-poetry - A graph-based approach to analyze lexical semantic change in Persian poetry by modeling neighborhood rewiring rather than abstract drift.
- A Graph-Enhanced Defense Framework for Explainable Fake News Detection with LLM (viability: 7): https://sciencetostartup.com/paper/a-graph-enhanced-defense-framework-for-explainable-fake-news-detection-with-llm - G-Defense is a graph-enhanced framework for explainable fake news detection using LLMs and RAG to provide fine-grained explanations based on unverified reports.
- Restoring Heterogeneity in LLM-based Social Simulation: An Audience Segmentation Approach (viability: 3): https://sciencetostartup.com/paper/restoring-heterogeneity-in-llm-based-social-simulation-an-audience-segmentation-approach - This paper introduces audience segmentation to improve the diversity and accuracy of social simulations performed by Large Language Models.
- A Parameter-Efficient Transfer Learning Approach through Multitask Prompt Distillation and Decomposition for Clinical NLP (viability: 7): https://sciencetostartup.com/paper/a-parameter-efficient-transfer-learning-approach-through-multitask-prompt-distillation-and-decomposition-for-clinical-nl - A parameter-efficient framework for clinical NLP that distills multitask prompts to adapt LLMs to new tasks with minimal parameters, outperforming LoRA.
- RPM-Net Reciprocal Point MLP Network for Unknown Network Security Threat Detection (viability: 7): https://sciencetostartup.com/paper/rpm-net-reciprocal-point-mlp-network-for-unknown-network-security-threat-detection - RPM-Net is a novel framework for detecting unknown network security threats by learning 'non-class' representations and using adversarial margin constraints for interpretability.
- SHAPE: Stage-aware Hierarchical Advantage via Potential Estimation for LLM Reasoning (viability: 7): https://sciencetostartup.com/paper/shape-stage-aware-hierarchical-advantage-via-potential-estimation-for-llm-reasoning - SHAPE is a framework that improves LLM reasoning by assigning credit hierarchically and estimating potential to reduce token consumption while increasing accuracy.
- SubFLOT: Submodel Extraction for Efficient and Personalized Federated Learning via Optimal Transport (viability: 7): https://sciencetostartup.com/paper/subflot-submodel-extraction-for-efficient-and-personalized-federated-learning-via-optimal-transport - A framework for personalized federated pruning that creates customized submodels using optimal transport and adaptive regularization to improve efficiency on resource-constrained devices.
- Logical Robots: Declarative Multi-Agent Programming in Logica (viability: 3): https://sciencetostartup.com/paper/logical-robots-declarative-multi-agent-programming-in-logica - A simulation platform for multi-agent robot behavior where robot actions are declaratively programmed using logic predicates, unifying reactive control and planning.
- Rethinking Generalization in Reasoning SFT: A Conditional Analysis on Optimization, Data, and Model Capability (viability: 3): https://sciencetostartup.com/paper/rethinking-generalization-in-reasoning-sft-a-conditional-analysis-on-optimization-data-and-model-capability - Analyzes the conditional generalization of reasoning in LLMs, finding it depends on optimization, data quality, and model capability, with trade-offs in safety.
- CubeGraph: Efficient Retrieval-Augmented Generation for Spatial and Temporal Data (viability: 8): https://sciencetostartup.com/paper/cubegraph-efficient-retrieval-augmented-generation-for-spatial-and-temporal-data - CubeGraph is an indexing framework that unifies vector search and spatial filters for retrieval-augmented generation, significantly improving query performance.
- The Detection--Extraction Gap: Models Know the Answer Before They Can Say It (viability: 7): https://sciencetostartup.com/paper/the-detection-extraction-gap-models-know-the-answer-before-they-can-say-it - A novel early exit strategy for LLMs that significantly reduces token generation while improving accuracy by leveraging free continuations.
- TwinLoop: Simulation-in-the-Loop Digital Twins for Online Multi-Agent Reinforcement Learning (viability: 3): https://sciencetostartup.com/paper/twinloop-simulation-in-the-loop-digital-twins-for-online-multi-agent-reinforcement-learning - A simulation-in-the-loop digital twin framework to accelerate policy adaptation in multi-agent reinforcement learning systems.
- Scientific Knowledge-driven Decoding Constraints Improving the Reliability of LLMs (viability: 7): https://sciencetostartup.com/paper/scientific-knowledge-driven-decoding-constraints-improving-the-reliability-of-llms - A method to integrate scientific knowledge as strong constraints into LLM generation, significantly reducing hallucination and improving accuracy on scientific tasks.
- LLM-based Schema-Guided Extraction and Validation of Missing-Person Intelligence from Heterogeneous Data Sources (viability: 5): https://sciencetostartup.com/paper/llm-based-schema-guided-extraction-and-validation-of-missing-person-intelligence-from-heterogeneous-data-sources - An AI-driven pipeline that parses and normalizes missing-person investigative documents into a unified schema, with an LLM-assisted pathway for improved extraction quality.
- AI-Driven Research for Databases (viability: 8): https://sciencetostartup.com/paper/ai-driven-research-for-databases - Automates the discovery of novel database optimization algorithms by co-evolving AI evaluators with solution generators, significantly outperforming state-of-the-art baselines.
- On Emotion-Sensitive Decision Making of Small Language Model Agents (viability: 5): https://sciencetostartup.com/paper/on-emotion-sensitive-decision-making-of-small-language-model-agents - Develops a benchmark and method to study how emotional states influence small language model agent decision-making in strategic games.
- SkillSieve: A Hierarchical Triage Framework for Detecting Malicious AI Agent Skills (viability: 8): https://sciencetostartup.com/paper/skillsieve-a-hierarchical-triage-framework-for-detecting-malicious-ai-agent-skills - SkillSieve is a hierarchical framework that uses multi-modal LLM analysis to detect malicious agent skills with high accuracy and low cost, outperforming existing methods.
- Soft-Quantum Algorithms (viability: 7): https://sciencetostartup.com/paper/soft-quantum-algorithms - Introduces 'Soft-Quantum Algorithms' that train matrices directly for quantum neural networks, significantly reducing training time and improving performance on classification and reinforcement learning tasks.
- Database Querying under Missing Values Governed by Missingness Mechanisms (viability: 2): https://sciencetostartup.com/paper/database-querying-under-missing-values-governed-by-missingness-mechanisms - Develops a theoretical framework for querying relational databases with missing values governed by probabilistic mechanisms.
- Adaptive Differential Privacy for Federated Medical Image Segmentation Across Diverse Modalities (viability: 7): https://sciencetostartup.com/paper/adaptive-differential-privacy-for-federated-medical-image-segmentation-across-diverse-modalities - An adaptive differentially private federated learning framework for medical image segmentation that balances privacy and utility across diverse modalities.
- Efficient Quantization of Mixture-of-Experts with Theoretical Generalization Guarantees (viability: 3): https://sciencetostartup.com/paper/efficient-quantization-of-mixture-of-experts-with-theoretical-generalization-guarantees - Proposes a theoretically grounded expert-wise mixed precision quantization strategy for Mixture-of-Experts models to reduce inference cost.
- MedConclusion: A Benchmark for Biomedical Conclusion Generation from Structured Abstracts (viability: 8): https://sciencetostartup.com/paper/medconclusion-a-benchmark-for-biomedical-conclusion-generation-from-structured-abstracts - Introduces MedConclusion, a large-scale dataset and benchmark for evaluating LLM's ability to generate biomedical conclusions from structured abstracts.
- Improving Robustness In Sparse Autoencoders via Masked Regularization (viability: 3): https://sciencetostartup.com/paper/improving-robustness-in-sparse-autoencoders-via-masked-regularization - A regularization technique to improve the robustness and interpretability of sparse autoencoders for LLM activations.
- Discrete Flow Matching Policy Optimization (viability: 4): https://sciencetostartup.com/paper/discrete-flow-matching-policy-optimization - A unified RL framework for fine-tuning discrete flow matching models, improving controllable discrete sequence generation.
- Inference-Time Code Selection via Symbolic Equivalence Partitioning (viability: 7): https://sciencetostartup.com/paper/inference-time-code-selection-via-symbolic-equivalence-partitioning - A symbolic execution framework for LLM code generation that improves accuracy by semantically grouping candidate programs without additional inference.
- Distributed Interpretability and Control for Large Language Models (viability: 7): https://sciencetostartup.com/paper/distributed-interpretability-and-control-for-large-language-models - Innovative interpretability tool for real-time control of large language models in distributed systems.
- Hybrid ResNet-1D-BiGRU with Multi-Head Attention for Cyberattack Detection in Industrial IoT Environments (viability: 7): https://sciencetostartup.com/paper/hybrid-resnet-1d-bigru-with-multi-head-attention-for-cyberattack-detection-in-industrial-iot-environments - A hybrid deep learning model for real-time intrusion detection in Industrial IoT systems, outperforming existing methods with high accuracy and low latency.
- Multi-objective Evolutionary Merging Enables Efficient Reasoning Models (viability: 7): https://sciencetostartup.com/paper/multi-objective-evolutionary-merging-enables-efficient-reasoning-models - A novel framework using evolutionary model merging to efficiently reduce reasoning token length in LLMs while preserving accuracy.
- From Load Tests to Live Streams: Graph Embedding-Based Anomaly Detection in Microservice Architectures (viability: 7): https://sciencetostartup.com/paper/from-load-tests-to-live-streams-graph-embedding-based-anomaly-detection-in-microservice-architectures - A graph-based anomaly detection system for microservice architectures that identifies under-represented services during live events.
- The Defense Trilemma: Why Prompt Injection Defense Wrappers Fail? (viability: 3): https://sciencetostartup.com/paper/the-defense-trilemma-why-prompt-injection-defense-wrappers-fail - Theoretical proof that continuous, utility-preserving wrapper defenses for LLMs cannot prevent all prompt injection attacks.
- Continual Visual Anomaly Detection on the Edge: Benchmark and Efficient Solutions (viability: 7): https://sciencetostartup.com/paper/continual-visual-anomaly-detection-on-the-edge-benchmark-and-efficient-solutions - Enabling continuous visual anomaly detection on resource-constrained edge devices with a new benchmark and efficient model adaptations.
- When to Call an Apple Red: Humans Follow Introspective Rules, VLMs Don't (viability: 7): https://sciencetostartup.com/paper/when-to-call-an-apple-red-humans-follow-introspective-rules-vlms-don-t - Develops a benchmark dataset and methodology to reveal and quantify how Vision-Language Models fail to adhere to their own introspective reasoning, highlighting a critical gap for trustworthy deployment.
- Attention Flows: Tracing LLM Conceptual Engagement via Story Summaries (viability: 7): https://sciencetostartup.com/paper/attention-flows-tracing-llm-conceptual-engagement-via-story-summaries - Introduces a novel dataset and methodology to compare human and LLM conceptual engagement in novel summarization, revealing LLM limitations in long-form text comprehension.
- Towards Resilient Intrusion Detection in CubeSats: Challenges, TinyML Solutions, and Future Directions (viability: 3): https://sciencetostartup.com/paper/towards-resilient-intrusion-detection-in-cubesats-challenges-tinyml-solutions-and-future-directions - Reviews cybersecurity challenges for CubeSats and explores TinyML as a potential solution for resource-constrained intrusion detection, identifying future research directions.
- Say Something Else: Rethinking Contextual Privacy as Information Sufficiency (viability: 3): https://sciencetostartup.com/paper/say-something-else-rethinking-contextual-privacy-as-information-sufficiency - Formalizes contextual privacy for LLM agents as an Information Sufficiency task, introducing free-text pseudonymization and a conversational evaluation protocol.
- BDI-Kit Demo: A Toolkit for Programmable and Conversational Data Harmonization (viability: 7): https://sciencetostartup.com/paper/bdi-kit-demo-a-toolkit-for-programmable-and-conversational-data-harmonization - A toolkit and AI-powered chat interface for programmatic and conversational data harmonization, reducing bottlenecks in integrative analysis.
- FMI@SU ToxHabits: Evaluating LLMs Performance on Toxic Habit Extraction in Spanish Clinical Texts (viability: 5): https://sciencetostartup.com/paper/fmi-su-toxhabits-evaluating-llms-performance-on-toxic-habit-extraction-in-spanish-clinical-texts - Leveraging LLMs for toxic habit extraction in Spanish clinical texts, achieving promising F1 scores for substance use recognition.
- ProofSketcher: Hybrid LLM + Lightweight Proof Checker for Reliable Math/Logic Reasoning (viability: 3): https://sciencetostartup.com/paper/proofsketcher-hybrid-llm-lightweight-proof-checker-for-reliable-math-logic-reasoning - A hybrid system combining LLMs with lightweight proof checkers for more reliable mathematical and logical reasoning.
- Qualixar OS: A Universal Operating System for AI Agent Orchestration (viability: 3): https://sciencetostartup.com/paper/qualixar-os-a-universal-operating-system-for-ai-agent-orchestration - Qualixar OS is a universal operating system for orchestrating heterogeneous multi-agent systems across multiple LLM providers and frameworks.
- Toward a universal foundation model for graph-structured data (viability: 7): https://sciencetostartup.com/paper/toward-a-universal-foundation-model-for-graph-structured-data - A universal foundation model for graph-structured data that learns transferable representations, outperforming supervised baselines in zero-shot and few-shot generalization for biomedical and network science applications.
- MorphDistill: Distilling Unified Morphological Knowledge from Pathology Foundation Models for Colorectal Cancer Survival Prediction (viability: 7): https://sciencetostartup.com/paper/morphdistill-distilling-unified-morphological-knowledge-from-pathology-foundation-models-for-colorectal-cancer-survival - MorphDistill distills unified morphological knowledge from pathology foundation models into a compact encoder for improved colorectal cancer survival prediction, outperforming baselines with strong generalization.
- SELFDOUBT: Uncertainty Quantification for Reasoning LLMs via the Hedge-to-Verify Ratio (viability: 8): https://sciencetostartup.com/paper/selfdoubt-uncertainty-quantification-for-reasoning-llms-via-the-hedge-to-verify-ratio - SELFDOUBT quantifies uncertainty in reasoning LLMs using a single-pass trace analysis, offering a scalable, production-ready solution for proprietary APIs with high-precision confidence gating.
- Uncertainty Estimation for Deep Reconstruction in Actuatic Disaster Scenarios with Autonomous Vehicles (viability: 4): https://sciencetostartup.com/paper/uncertainty-estimation-for-deep-reconstruction-in-actuatic-disaster-scenarios-with-autonomous-vehicles - Evidential Deep Learning is the preferred method for uncertainty-aware field reconstruction in autonomous vehicles, offering superior accuracy and calibration over other methods.
- The Master Key Hypothesis: Unlocking Cross-Model Capability Transfer via Linear Subspace Alignment (viability: 7): https://sciencetostartup.com/paper/the-master-key-hypothesis-unlocking-cross-model-capability-transfer-via-linear-subspace-alignment - A framework to transfer learned capabilities between language models without retraining, improving performance on specific tasks.
- SymptomWise: A Deterministic Reasoning Layer for Reliable and Efficient AI Systems (viability: 6): https://sciencetostartup.com/paper/symptomwise-a-deterministic-reasoning-layer-for-reliable-and-efficient-ai-systems - A hybrid AI system for medical symptom analysis that combines deterministic reasoning with LLMs for improved reliability and interpretability.
- Beyond Functional Correctness: Design Issues in AI IDE-Generated Large-Scale Projects (viability: 5): https://sciencetostartup.com/paper/beyond-functional-correctness-design-issues-in-ai-ide-generated-large-scale-projects - An AI-assisted framework for generating functional large-scale software projects, highlighting the need for human oversight on design quality.
- WebSP-Eval: Evaluating Web Agents on Website Security and Privacy Tasks (viability: 7): https://sciencetostartup.com/paper/websp-eval-evaluating-web-agents-on-website-security-and-privacy-tasks - A new benchmark and framework for evaluating web agents on website security and privacy tasks, revealing current limitations in autonomous exploration.
- A Severity-Based Curriculum Learning Strategy for Arabic Medical Text Generation (viability: 5): https://sciencetostartup.com/paper/a-severity-based-curriculum-learning-strategy-for-arabic-medical-text-generation - A severity-based curriculum learning strategy for Arabic medical text generation improves model performance by gradually introducing more complex medical conditions.
- GS-Surrogate: Deformable Gaussian Splatting for Parameter Space Exploration of Ensemble Simulations (viability: 6): https://sciencetostartup.com/paper/gs-surrogate-deformable-gaussian-splatting-for-parameter-space-exploration-of-ensemble-simulations - GS-Surrogate uses deformable Gaussian Splatting for real-time, flexible exploration of ensemble simulations across parameter spaces.
- In-Context Learning in Speech Language Models: Analyzing the Role of Acoustic Features, Linguistic Structure, and Induction Heads (viability: 5): https://sciencetostartup.com/paper/in-context-learning-in-speech-language-models-analyzing-the-role-of-acoustic-features-linguistic-structure-and-induction - Investigating in-context learning in speech models, this work analyzes the impact of acoustic features and linguistic structure on task inference and mimicry.
- DietDelta: A Vision-Language Approach for Dietary Assessment via Before-and-After Images (viability: 5): https://sciencetostartup.com/paper/dietdelta-a-vision-language-approach-for-dietary-assessment-via-before-and-after-images - A vision-language framework uses paired before-and-after images to assess dietary intake at the food-item level, improving nutritional analysis.
- Bi-Level Optimization for Single Domain Generalization (viability: 7): https://sciencetostartup.com/paper/bi-level-optimization-for-single-domain-generalization - A bi-level optimization framework that simulates distribution shifts to improve single domain generalization in machine learning.
- Severity-Aware Weighted Loss for Arabic Medical Text Generation (viability: 8): https://sciencetostartup.com/paper/severity-aware-weighted-loss-for-arabic-medical-text-generation - A severity-aware weighted loss function for fine-tuning Arabic language models to prioritize clinically critical medical text generation.
- "Don't Be Afraid, Just Learn": Insights from Industry Practitioners to Prepare Software Engineers in the Age of Generative AI (viability: 0): https://sciencetostartup.com/paper/don-t-be-afraid-just-learn-insights-from-industry-practitioners-to-prepare-software-engineers-in-the-age-of-generative-a - An empirical study providing recommendations for academia to prepare software engineers for the age of generative AI.
- BiScale-GTR: Fragment-Aware Graph Transformers for Multi-Scale Molecular Representation Learning (viability: 7): https://sciencetostartup.com/paper/biscale-gtr-fragment-aware-graph-transformers-for-multi-scale-molecular-representation-learning - A fragment-aware graph transformer framework for multi-scale molecular representation learning that achieves state-of-the-art performance.
- A Novel Automatic Framework for Speaker Drift Detection in Synthesized Speech (viability: 7): https://sciencetostartup.com/paper/a-novel-automatic-framework-for-speaker-drift-detection-in-synthesized-speech - A framework to automatically detect subtle speaker identity shifts in synthesized speech, improving coherence for long-form audio.
- Blockchain and AI: Securing Intelligent Networks for the Future (viability: 3): https://sciencetostartup.com/paper/blockchain-and-ai-securing-intelligent-networks-for-the-future - Exploring the integration of Blockchain and AI for securing intelligent networks against sophisticated cyber threats.
- In-Place Test-Time Training (viability: 7): https://sciencetostartup.com/paper/in-place-test-time-training - A drop-in framework for Large Language Models that enables test-time training for dynamic adaptation to new information without costly retraining.
- DiffHDR: Re-Exposing LDR Videos with Video Diffusion Models (viability: 7): https://sciencetostartup.com/paper/diffhdr-re-exposing-ldr-videos-with-video-diffusion-models - DiffHDR transforms LDR videos into HDR by leveraging advanced video diffusion models for superior visual quality.
- MMEmb-R1: Reasoning-Enhanced Multimodal Embedding with Pair-Aware Selection and Adaptive Control (viability: 7): https://sciencetostartup.com/paper/mmemb-r1-reasoning-enhanced-multimodal-embedding-with-pair-aware-selection-and-adaptive-control - Enhance multimodal applications with reasoning-enabled embeddings that outperform existing models.
- Toward Consistent World Models with Multi-Token Prediction and Latent Semantic Enhancement (viability: 3): https://sciencetostartup.com/paper/toward-consistent-world-models-with-multi-token-prediction-and-latent-semantic-enhancement - A theoretical exploration of multi-token prediction for LLMs, proposing a method to reduce structural hallucinations in latent space representations.
- Who Governs the Machine? A Machine Identity Governance Taxonomy (MIGT) for AI Systems Operating Across Enterprise and Geopolitical Boundaries (viability: 2): https://sciencetostartup.com/paper/who-governs-the-machine-a-machine-identity-governance-taxonomy-migt-for-ai-systems-operating-across-enterprise-and-geopo - A taxonomy and framework for governing machine identities in AI systems to mitigate enterprise and geopolitical risks.
- Generating Synthetic Doctor-Patient Conversations for Long-form Audio Summarization (viability: 7): https://sciencetostartup.com/paper/generating-synthetic-doctor-patient-conversations-for-long-form-audio-summarization - A synthetic data generation pipeline for doctor-patient conversations to train and evaluate long-form audio summarization models.
- Shot-Based Quantum Encoding: A Data-Loading Paradigm for Quantum Neural Networks (viability: 6): https://sciencetostartup.com/paper/shot-based-quantum-encoding-a-data-loading-paradigm-for-quantum-neural-networks - Shot-Based Quantum Encoding (SBQE) for quantum neural networks that improves data loading efficiency and performance on noisy hardware.
- Claw-Eval: Toward Trustworthy Evaluation of Autonomous Agents (viability: 8): https://sciencetostartup.com/paper/claw-eval-toward-trustworthy-evaluation-of-autonomous-agents - Claw-Eval: A trustworthy evaluation suite for autonomous agents that provides trajectory-aware grading, safety, and robustness assessment.
- PoM: A Linear-Time Replacement for Attention with the Polynomial Mixer (viability: 7): https://sciencetostartup.com/paper/pom-a-linear-time-replacement-for-attention-with-the-polynomial-mixer - A novel linear-time attention replacement that significantly reduces computational cost for long sequences across diverse AI domains.
- Gym-Anything: Turn any Software into an Agent Environment (viability: 9): https://sciencetostartup.com/paper/gym-anything-turn-any-software-into-an-agent-environment - Turn any software application into an interactive agent environment, enabling autonomous system training and evaluation.
- Lightweight Multimodal Adaptation of Vision Language Models for Species Recognition and Habitat Context Interpretation in Drone Thermal Imagery (viability: 8): https://sciencetostartup.com/paper/lightweight-multimodal-adaptation-of-vision-language-models-for-species-recognition-and-habitat-context-interpretation-i - Lightweight adaptation of vision-language models for species recognition and habitat interpretation using drone thermal imagery.
- ACE-Bench: Agent Configurable Evaluation with Scalable Horizons and Controllable Difficulty under Lightweight Environments (viability: 8): https://sciencetostartup.com/paper/ace-bench-agent-configurable-evaluation-with-scalable-horizons-and-controllable-difficulty-under-lightweight-environment - A benchmark for agent evaluation with scalable horizons and controllable difficulty in lightweight environments, addressing limitations of existing benchmarks.
- Artificial Intelligence and the Structure of Mathematics (viability: 1): https://sciencetostartup.com/paper/artificial-intelligence-and-the-structure-of-mathematics - This paper explores the potential of AI to revolutionize mathematics by forging new routes to understanding formal proofs and discovering mathematical concepts.
- LLM4CodeRE: Generative AI for Code Decompilation Analysis and Reverse Engineering (viability: 7): https://sciencetostartup.com/paper/llm4codere-generative-ai-for-code-decompilation-analysis-and-reverse-engineering - LLM4CodeRE is a domain-adaptive LLM framework for bidirectional code reverse engineering, outperforming existing tools in malware analysis.
- Social Dynamics as Critical Vulnerabilities that Undermine Objective Decision-Making in LLM Collectives (viability: 5): https://sciencetostartup.com/paper/social-dynamics-as-critical-vulnerabilities-that-undermine-objective-decision-making-in-llm-collectives - This research reveals how social dynamics like conformity and persuasion undermine decision-making in LLM agent collectives, mirroring human biases.
- LAG-XAI: A Lie-Inspired Affine Geometric Framework for Interpretable Paraphrasing in Transformer Latent Spaces (viability: 7): https://sciencetostartup.com/paper/lag-xai-a-lie-inspired-affine-geometric-framework-for-interpretable-paraphrasing-in-transformer-latent-spaces - LAG-XAI is a geometric framework for interpretable paraphrasing in Transformers, enabling efficient LLM hallucination detection and mechanistic understanding.
- Scientific Graphics Program Synthesis via Dual Self-Consistency Reinforcement Learning (viability: 8): https://sciencetostartup.com/paper/scientific-graphics-program-synthesis-via-dual-self-consistency-reinforcement-learning - A framework for synthesizing scientific graphics from text descriptions, featuring a large-scale dataset, a comprehensive benchmark, and a novel reinforcement learning approach that outperforms leading LLMs.
- Graph-PiT: Enhancing Structural Coherence in Part-Based Image Synthesis via Graph Priors (viability: 7): https://sciencetostartup.com/paper/graph-pit-enhancing-structural-coherence-in-part-based-image-synthesis-via-graph-priors - Graph-PiT enhances part-based image synthesis by using graph priors to model structural dependencies between visual components, improving coherence and controllability.
- Stories of Your Life as Others: A Round-Trip Evaluation of LLM-Generated Life Stories Conditioned on Rich Psychometric Profiles (viability: 3): https://sciencetostartup.com/paper/stories-of-your-life-as-others-a-round-trip-evaluation-of-llm-generated-life-stories-conditioned-on-rich-psychometric-pr - LLMs can generate life stories that robustly encode personality traits, with recovered scores approaching human reliability and demonstrating behavioral differentiation.
- A Multi-Stage Validation Framework for Trustworthy Large-scale Clinical Information Extraction using Large Language Models (viability: 7): https://sciencetostartup.com/paper/a-multi-stage-validation-framework-for-trustworthy-large-scale-clinical-information-extraction-using-large-language-mode - A multi-stage validation framework enables trustworthy large-scale clinical information extraction using LLMs without annotation-intensive evaluation, demonstrating feasibility for real-world deployment.
- CritBench: A Framework for Evaluating Cybersecurity Capabilities of Large Language Models in IEC 61850 Digital Substation Environments (viability: 8): https://sciencetostartup.com/paper/critbench-a-framework-for-evaluating-cybersecurity-capabilities-of-large-language-models-in-iec-61850-digital-substation - CritBench evaluates LLM cybersecurity capabilities in IEC 61850 environments, addressing critical gaps in existing frameworks.
- Governance and Regulation of Artificial Intelligence in Developing Countries: A Case Study of Nigeria (viability: 2): https://sciencetostartup.com/paper/governance-and-regulation-of-artificial-intelligence-in-developing-countries-a-case-study-of-nigeria - This study explores the governance of AI in Nigeria, highlighting ethical risks and regulatory gaps.
- How LLMs Follow Instructions: Skillful Coordination, Not a Universal Mechanism (viability: 2): https://sciencetostartup.com/paper/how-llms-follow-instructions-skillful-coordination-not-a-universal-mechanism - This paper investigates the mechanisms behind instruction-following in language models, challenging the notion of a universal mechanism.
- Epistemic Blinding: An Inference-Time Protocol for Auditing Prior Contamination in LLM-Assisted Analysis (viability: 8): https://sciencetostartup.com/paper/epistemic-blinding-an-inference-time-protocol-for-auditing-prior-contamination-in-llm-assisted-analysis - Epistemic blinding enhances auditability in LLM-assisted drug target prioritization, addressing contamination in outputs.
- The Model Agreed, But Didn't Learn: Diagnosing Surface Compliance in Large Language Models (viability: 7): https://sciencetostartup.com/paper/the-model-agreed-but-didn-t-learn-diagnosing-surface-compliance-in-large-language-models - A diagnostic framework to uncover 'surface compliance' in LLM knowledge editing, revealing when models mimic changes without true belief modification.
- Flowr -- Scaling Up Retail Supply Chain Operations Through Agentic AI in Large Scale Supermarket Chains (viability: 8): https://sciencetostartup.com/paper/flowr-scaling-up-retail-supply-chain-operations-through-agentic-ai-in-large-scale-supermarket-chains - Flowr automates end-to-end retail supply chain operations for large supermarket chains using agentic AI.
- A Formal Security Framework for MCP-Based AI Agents: Threat Taxonomy, Verification Models, and Defense Mechanisms (viability: 8): https://sciencetostartup.com/paper/a-formal-security-framework-for-mcp-based-ai-agents-threat-taxonomy-verification-models-and-defense-mechanisms - MCPSHIELD is a formal security framework for MCP-based AI agents, providing a threat taxonomy, verification models, and an integrated defense architecture.
- Beyond Compromise: Pareto-Lenient Consensus for Efficient Multi-Preference LLM Alignment (viability: 7): https://sciencetostartup.com/paper/beyond-compromise-pareto-lenient-consensus-for-efficient-multi-preference-llm-alignment - Pareto-Lenient Consensus (PLC) is a game-theoretic framework for LLM alignment that allows negotiation-driven exploration of the Pareto-optimal frontier.
- Does Pass Rate Tell the Whole Story? Evaluating Design Constraint Compliance in LLM-based Issue Resolution (viability: 7): https://sciencetostartup.com/paper/does-pass-rate-tell-the-whole-story-evaluating-design-constraint-compliance-in-llm-based-issue-resolution - A benchmark and LLM-based verifier to evaluate code patch quality beyond test pass rates, revealing significant design constraint violations in current AI agents.
- Polynomial-Time Algorithm for Thiele Voting Rules with Voter Interval Preferences (viability: 1): https://sciencetostartup.com/paper/polynomial-time-algorithm-for-thiele-voting-rules-with-voter-interval-preferences - A polynomial-time algorithm for Thiele voting rules with voter interval preferences, resolving a 10-year-old open problem using human-AI collaboration.
- Towards Trustworthy Report Generation: A Deep Research Agent with Progressive Confidence Estimation and Calibration (viability: 7): https://sciencetostartup.com/paper/towards-trustworthy-report-generation-a-deep-research-agent-with-progressive-confidence-estimation-and-calibration - A deep research agent that generates trustworthy reports by progressively estimating and calibrating confidence in its generated claims.
- MARL-GPT: Foundation Model for Multi-Agent Reinforcement Learning (viability: 3): https://sciencetostartup.com/paper/marl-gpt-foundation-model-for-multi-agent-reinforcement-learning - MARL-GPT: A single GPT-based foundation model trained at scale to perform across diverse multi-agent reinforcement learning environments.
- Context-Value-Action Architecture for Value-Driven Large Language Model Agents (viability: 7): https://sciencetostartup.com/paper/context-value-action-architecture-for-value-driven-large-language-model-agents - A novel architecture for LLM agents that enhances behavioral fidelity and reduces value polarization.
- Saliency-Guided Representation with Consistency Policy Learning for Visual Unsupervised Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/saliency-guided-representation-with-consistency-policy-learning-for-visual-unsupervised-reinforcement-learning - A framework that enhances zero-shot generalization in visual unsupervised reinforcement learning.
- "I See What You Did There": Can Large Vision-Language Models Understand Multimodal Puns? (viability: 6): https://sciencetostartup.com/paper/i-see-what-you-did-there-can-large-vision-language-models-understand-multimodal-puns - A pipeline for generating and understanding multimodal puns to enhance humor comprehension in VLMs.
- ReLU Networks for Exact Generation of Similar Graphs (viability: 6): https://sciencetostartup.com/paper/relu-networks-for-exact-generation-of-similar-graphs - A ReLU network architecture for generating graphs within a specified edit distance from a source graph.
- Selective Aggregation of Attention Maps Improves Diffusion-Based Visual Interpretation (viability: 3): https://sciencetostartup.com/paper/selective-aggregation-of-attention-maps-improves-diffusion-based-visual-interpretation - This research proposes a method to improve the interpretability of text-to-image models by selectively aggregating attention maps, showing potential for better control and diagnosis of prompt misinterpretations.
- HybridKV: Hybrid KV Cache Compression for Efficient Multimodal Large Language Model Inference (viability: 7): https://sciencetostartup.com/paper/hybridkv-hybrid-kv-cache-compression-for-efficient-multimodal-large-language-model-inference - HybridKV is a novel framework that significantly compresses multimodal LLM KV caches, reducing memory usage and latency by up to 7.9x and 1.52x respectively, with minimal performance impact.
- Automatic dental superimposition of 3D intraorals and 2D photographs for human identification (viability: 3): https://sciencetostartup.com/paper/automatic-dental-superimposition-of-3d-intraorals-and-2d-photographs-for-human-identification - This research presents an automatic 3D to 2D dental superimposition method for human identification, overcoming limitations of current approaches by modeling perspective distortion and providing objective morphological comparison scores.
- Joint Knowledge Base Completion and Question Answering by Combining Large Language Models and Small Language Models (viability: 5): https://sciencetostartup.com/paper/joint-knowledge-base-completion-and-question-answering-by-combining-large-language-models-and-small-language-models - JCQL is a novel framework that jointly enhances Knowledge Base Completion and Question Answering by iteratively combining Large Language Models and Small Language Models, improving performance on both tasks.
- Swiss-Bench 003: Evaluating LLM Reliability and Adversarial Security for Swiss Regulatory Contexts (viability: 4): https://sciencetostartup.com/paper/swiss-bench-003-evaluating-llm-reliability-and-adversarial-security-for-swiss-regulatory-contexts - A new benchmark and evaluation framework for LLM reliability and adversarial security tailored for Swiss financial and regulatory contexts.
- JTON: A Token-Efficient JSON Superset with Zen Grid Tabular Encoding for Large Language Models (viability: 7): https://sciencetostartup.com/paper/jton-a-token-efficient-json-superset-with-zen-grid-tabular-encoding-for-large-language-models - JTON is a token-efficient JSON superset that reduces LLM processing costs for structured data by up to 60% with a novel tabular encoding.
- When Do We Need LLMs? A Diagnostic for Language-Driven Bandits (viability: 5): https://sciencetostartup.com/paper/when-do-we-need-llms-a-diagnostic-for-language-driven-bandits - A diagnostic tool to determine when LLM-driven reasoning is necessary for sequential decision-making versus using lightweight numerical bandits.
- Neural Network Pruning via QUBO Optimization (viability: 4): https://sciencetostartup.com/paper/neural-network-pruning-via-qubo-optimization - A hybrid QUBO optimization framework for neural network pruning that integrates gradient-aware metrics and data-driven similarity for improved compression.
- Deep Researcher Agent: An Autonomous Framework for 24/7 Deep Learning Experimentation with Zero-Cost Monitoring (viability: 7): https://sciencetostartup.com/paper/deep-researcher-agent-an-autonomous-framework-for-24-7-deep-learning-experimentation-with-zero-cost-monitoring - An autonomous framework for 24/7 deep learning experimentation with zero-cost monitoring and efficient memory management.
- Evaluating Learner Representations for Differentiation Prior to Instructional Outcomes (viability: 4): https://sciencetostartup.com/paper/evaluating-learner-representations-for-differentiation-prior-to-instructional-outcomes - A novel metric to evaluate learner representations for differentiation in educational AI systems, independent of instructional outcomes.
- EEG-MFTNet: An Enhanced EEGNet Architecture with Multi-Scale Temporal Convolutions and Transformer Fusion for Cross-Session Motor Imagery Decoding (viability: 7): https://sciencetostartup.com/paper/eeg-mftnet-an-enhanced-eegnet-architecture-with-multi-scale-temporal-convolutions-and-transformer-fusion-for-cross-sessi - EEG-MFTNet: An enhanced EEGNet architecture with multi-scale temporal convolutions and Transformer fusion for robust cross-session motor imagery decoding.
- Vision-Guided Iterative Refinement for Frontend Code Generation (viability: 7): https://sciencetostartup.com/paper/vision-guided-iterative-refinement-for-frontend-code-generation - A vision-guided iterative refinement framework for frontend code generation, using a vision-language model as an automated critic.
- "OK Aura, Be Fair With Me": Demographics-Agnostic Training for Bias Mitigation in Wake-up Word Detection (viability: 5): https://sciencetostartup.com/paper/ok-aura-be-fair-with-me-demographics-agnostic-training-for-bias-mitigation-in-wake-up-word-detection - Demographics-agnostic training techniques significantly reduce bias in wake-up word detection across diverse speaker populations.
- Reciprocal Trust and Distrust in Artificial Intelligence Systems: The Hard Problem of Regulation (viability: 0): https://sciencetostartup.com/paper/reciprocal-trust-and-distrust-in-artificial-intelligence-systems-the-hard-problem-of-regulation - This paper argues that AI systems should be viewed as agents capable of reciprocal trust and distrust, posing challenges for regulation.
- Hierarchical Reinforcement Learning with Augmented Step-Level Transitions for LLM Agents (viability: 7): https://sciencetostartup.com/paper/hierarchical-reinforcement-learning-with-augmented-step-level-transitions-for-llm-agents - STEP-HRL is a hierarchical reinforcement learning framework for LLM agents that enables step-level learning, reducing computational cost and improving performance.
- What Models Know, How Well They Know It: Knowledge-Weighted Fine-Tuning for Learning When to Say "I Don't Know" (viability: 5): https://sciencetostartup.com/paper/what-models-know-how-well-they-know-it-knowledge-weighted-fine-tuning-for-learning-when-to-say-i-don-t-know - This method uses knowledge-weighted fine-tuning to enable LLMs to express uncertainty and avoid hallucinations for out-of-scope queries.
- Emergent social transmission of model-based representations without inference (viability: 2): https://sciencetostartup.com/paper/emergent-social-transmission-of-model-based-representations-without-inference - This paper explores how simple social cues can lead to the transmission of complex knowledge representations in agents without explicit mental state inference.
- CAKE: Cloud Architecture Knowledge Evaluation of Large Language Models (viability: 6): https://sciencetostartup.com/paper/cake-cloud-architecture-knowledge-evaluation-of-large-language-models - CAKE is a new benchmark for evaluating LLMs on cloud architecture knowledge, revealing insights into model scaling and the impact of augmentation strategies.
- On the Robustness of Diffusion-Based Image Compression to Bit-Flip Errors (viability: 3): https://sciencetostartup.com/paper/on-the-robustness-of-diffusion-based-image-compression-to-bit-flip-errors - This research demonstrates that diffusion-based image compression methods offer superior robustness to bit-flip errors compared to existing codecs.
- Hackers or Hallucinators? A Comprehensive Analysis of LLM-Based Automated Penetration Testing (viability: 5): https://sciencetostartup.com/paper/hackers-or-hallucinators-a-comprehensive-analysis-of-llm-based-automated-penetration-testing - This paper provides a comprehensive analysis and benchmark of LLM-based automated penetration testing frameworks, identifying key architectural designs and empirical performance.
- Can Large Language Models Reinvent Foundational Algorithms? (viability: 4): https://sciencetostartup.com/paper/can-large-language-models-reinvent-foundational-algorithms - This research explores the capability of Large Language Models to reinvent foundational computer science algorithms, demonstrating potential for AI-driven innovation in core computational concepts.
- SemLink: A Semantic-Aware Automated Test Oracle for Hyperlink Verification using Siamese Sentence-BERT (viability: 7): https://sciencetostartup.com/paper/semlink-a-semantic-aware-automated-test-oracle-for-hyperlink-verification-using-siamese-sentence-bert - SemLink is a semantic-aware automated test oracle that verifies hyperlink integrity 47.5x faster than LLMs, ensuring web content consistency for robust quality assurance.
- QA-MoE: Towards a Continuous Reliability Spectrum with Quality-Aware Mixture of Experts for Robust Multimodal Sentiment Analysis (viability: 4): https://sciencetostartup.com/paper/qa-moe-towards-a-continuous-reliability-spectrum-with-quality-aware-mixture-of-experts-for-robust-multimodal-sentiment-a - QA-MoE introduces a quality-aware mixture of experts for multimodal sentiment analysis, adapting to continuously varying input reliability for more robust performance.
- CRFT: Consistent-Recurrent Feature Flow Transformer for Cross-Modal Image Registration (viability: 7): https://sciencetostartup.com/paper/crft-consistent-recurrent-feature-flow-transformer-for-cross-modal-image-registration - CRFT is a transformer-based framework for cross-modal image registration, offering robust alignment for applications in remote sensing, autonomous navigation, and medical imaging.
- Attention Editing: A Versatile Framework for Cross-Architecture Attention Conversion (viability: 7): https://sciencetostartup.com/paper/attention-editing-a-versatile-framework-for-cross-architecture-attention-conversion - A framework to efficiently convert existing large language models to use new attention architectures without retraining, significantly improving inference speed.
- LUDOBENCH: Evaluating LLM Behavioural Decision-Making Through Spot-Based Board Game Scenarios in Ludo (viability: 8): https://sciencetostartup.com/paper/ludobench-evaluating-llm-behavioural-decision-making-through-spot-based-board-game-scenarios-in-ludo - A benchmark and simulator to evaluate LLM strategic decision-making in complex board games, revealing vulnerabilities in prompt sensitivity and strategic depth.
- From Incomplete Architecture to Quantified Risk: Multimodal LLM-Driven Security Assessment for Cyber-Physical Systems (viability: 4): https://sciencetostartup.com/paper/from-incomplete-architecture-to-quantified-risk-multimodal-llm-driven-security-assessment-for-cyber-physical-systems - A multimodal LLM-driven prototype tool to reconstruct and analyze cyber-physical system architectures for security assessment when documentation is incomplete.
- Rectified Schrödinger Bridge Matching for Few-Step Visual Navigation (viability: 7): https://sciencetostartup.com/paper/rectified-schr-dinger-bridge-matching-for-few-step-visual-navigation - A framework for visual navigation that significantly reduces integration steps for diffusion-based policies, enabling real-time robotic control.
- CuraLight: Debate-Guided Data Curation for LLM-Centered Traffic Signal Control (viability: 7): https://sciencetostartup.com/paper/curalight-debate-guided-data-curation-for-llm-centered-traffic-signal-control - An LLM-centered framework for traffic signal control that uses RL-guided data curation and multi-LLM deliberation to outperform state-of-the-art.
- SnapFlow: One-Step Action Generation for Flow-Matching VLAs via Progressive Self-Distillation (viability: 6): https://sciencetostartup.com/paper/snapflow-one-step-action-generation-for-flow-matching-vlas-via-progressive-self-distillation - SnapFlow compresses multi-step denoising in Vision-Language-Action models to a single forward pass, achieving state-of-the-art robotic manipulation with significantly reduced latency.
- LLM Reasoning as Trajectories: Step-Specific Representation Geometry and Correctness Signals (viability: 3): https://sciencetostartup.com/paper/llm-reasoning-as-trajectories-step-specific-representation-geometry-and-correctness-signals - Characterizes LLM chain-of-thought generation as structured trajectories in representation space, enabling mid-reasoning prediction of correctness and inference-time intervention.
- Multiscale Physics-Informed Neural Network for Complex Fluid Flows with Long-Range Dependencies (viability: 4): https://sciencetostartup.com/paper/multiscale-physics-informed-neural-network-for-complex-fluid-flows-with-long-range-dependencies - A Domain-Decomposed and Shifted Physics-Informed Neural Network (DDS-PINN) framework for complex fluid flows that resolves multiscale interactions with minimal supervision.
- Analogical Reasoning as a Doctor: A Foundation Model for Gastrointestinal Endoscopy Diagnosis (viability: 7): https://sciencetostartup.com/paper/analogical-reasoning-as-a-doctor-a-foundation-model-for-gastrointestinal-endoscopy-diagnosis - A foundation model for gastrointestinal endoscopy diagnosis that uses analogical reasoning to improve generalization and adaptability across diverse datasets and disease types.
- PECKER: A Precisely Efficient Critical Knowledge Erasure Recipe For Machine Unlearning in Diffusion Models (viability: 6): https://sciencetostartup.com/paper/pecker-a-precisely-efficient-critical-knowledge-erasure-recipe-for-machine-unlearning-in-diffusion-models - An efficient machine unlearning method for diffusion models that uses a saliency mask to prioritize parameter updates, reducing training time without sacrificing efficacy.
- Beyond Behavior: Why AI Evaluation Needs a Cognitive Revolution (viability: 0): https://sciencetostartup.com/paper/beyond-behavior-why-ai-evaluation-needs-a-cognitive-revolution - This paper argues that AI evaluation needs a cognitive revolution, moving beyond purely behavioral tests to consider internal processes and mechanisms for a more nuanced understanding of intelligence.
- Semantic-Topological Graph Reasoning for Language-Guided Pulmonary Screening (viability: 7): https://sciencetostartup.com/paper/semantic-topological-graph-reasoning-for-language-guided-pulmonary-screening - A framework for language-guided pulmonary screening that combines LLMs and vision models with graph reasoning and selective fine-tuning to improve accuracy and stability.
- Evaluation of Randomization through Style Transfer for Enhanced Domain Generalization (viability: 7): https://sciencetostartup.com/paper/evaluation-of-randomization-through-style-transfer-for-enhanced-domain-generalization - A lightweight, model-agnostic style transfer augmentation recipe that significantly improves computer vision model generalization from synthetic to real-world data.
- Foundations for Agentic AI Investigations from the Forensic Analysis of OpenClaw (viability: 4): https://sciencetostartup.com/paper/foundations-for-agentic-ai-investigations-from-the-forensic-analysis-of-openclaw - This paper provides foundational insights into the forensic analysis of agentic AI systems like OpenClaw, identifying recoverable traces and proposing an artifact taxonomy to aid digital investigations.
- ResearchEVO: An End-to-End Framework for Automated Scientific Discovery and Documentation (viability: 6): https://sciencetostartup.com/paper/researchevo-an-end-to-end-framework-for-automated-scientific-discovery-and-documentation - ResearchEVO is an end-to-end framework that automates scientific discovery by evolving algorithms and generating publication-ready research papers, validated on quantum error correction and PINNs.
- Context-Agent: Dynamic Discourse Trees for Non-Linear Dialogue (viability: 7): https://sciencetostartup.com/paper/context-agent-dynamic-discourse-trees-for-non-linear-dialogue - Context-Agent models multi-turn dialogue history as a dynamic tree structure to improve LLM coherence and efficiency in non-linear conversations, supported by a new benchmark.
- FastDiSS: Few-step Match Many-step Diffusion Language Model on Sequence-to-Sequence Generation--Full Version (viability: 5): https://sciencetostartup.com/paper/fastdiss-few-step-match-many-step-diffusion-language-model-on-sequence-to-sequence-generation-full-version - FastDiSS is a training framework for diffusion language models that improves robustness to self-conditioning errors during few-step inference, achieving faster speeds and competitive quality.
- COSMO-Agent: Tool-Augmented Agent for Closed-loop Optimization,Simulation,and Modeling Orchestration (viability: 8): https://sciencetostartup.com/paper/cosmo-agent-tool-augmented-agent-for-closed-loop-optimization-simulation-and-modeling-orchestration - An LLM-powered agent that orchestrates CAD-CAE tools to automate industrial design optimization, validated on an industry-aligned dataset.
- From Large Language Model Predicates to Logic Tensor Networks: Neurosymbolic Offer Validation in Regulated Procurement (viability: 5): https://sciencetostartup.com/paper/from-large-language-model-predicates-to-logic-tensor-networks-neurosymbolic-offer-validation-in-regulated-procurement - A neurosymbolic system that uses LLMs and Logic Tensor Networks to validate offer documents for regulated procurement, offering interpretable and auditable decisions.
- A canonical generalization of OBDD (viability: 0): https://sciencetostartup.com/paper/a-canonical-generalization-of-obdd - Introduces Tree Decision Diagrams (TDDs) as a generalization of OBDDs with improved succinctness and tractability for Boolean function representation.
- Turbulence-like 5/3 spectral scaling in contextual representations of language as a complex system (viability: 4): https://sciencetostartup.com/paper/turbulence-like-5-3-spectral-scaling-in-contextual-representations-of-language-as-a-complex-system - Discovers a consistent 5/3 spectral scaling in contextual language representations from transformer models, suggesting scale-free semantic integration.
- SignalClaw: LLM-Guided Evolutionary Synthesis of Interpretable Traffic Signal Control Skills (viability: 5): https://sciencetostartup.com/paper/signalclaw-llm-guided-evolutionary-synthesis-of-interpretable-traffic-signal-control-skills - An LLM-guided evolutionary framework synthesizes interpretable traffic signal control skills with rationale and executable code for adaptive traffic management.
- Experience Transfer for Multimodal LLM Agents in Minecraft Game (viability: 4): https://sciencetostartup.com/paper/experience-transfer-for-multimodal-llm-agents-in-minecraft-game - A transfer-oriented memory framework for multimodal LLM agents in Minecraft that decomposes experience into five dimensions for efficient task solving.
- Inventory of the 12 007 Low-Dimensional Pseudo-Boolean Landscapes Invariant to Rank, Translation, and Rotation (viability: 4): https://sciencetostartup.com/paper/inventory-of-the-12-007-low-dimensional-pseudo-boolean-landscapes-invariant-to-rank-translation-and-rotation - An exhaustive inventory of 12,007 invariant landscape classes for pseudo-Boolean functions to aid in benchmark design and algorithm understanding.
- ActivityEditor: Learning to Synthesize Physically Valid Human Mobility (viability: 7): https://sciencetostartup.com/paper/activityeditor-learning-to-synthesize-physically-valid-human-mobility - ActivityEditor is a dual-LLM-agent framework for zero-shot cross-regional human mobility trajectory generation, ensuring physical validity and statistical fidelity.
- Controllable Singing Style Conversion with Boundary-Aware Information Bottleneck (viability: 5): https://sciencetostartup.com/paper/controllable-singing-style-conversion-with-boundary-aware-information-bottleneck - A novel singing style conversion system that advances fine-grained style conversion and control, achieving best naturalness performance in a challenge.
- Market-Bench: Benchmarking Large Language Models on Economic and Trade Competition (viability: 8): https://sciencetostartup.com/paper/market-bench-benchmarking-large-language-models-on-economic-and-trade-competition - Market-Bench is a benchmark for evaluating LLMs in economic and trade competition, revealing significant performance disparities in multi-agent supply chain scenarios.
- Learned Elevation Models as a Lightweight Alternative to LiDAR for Radio Environment Map Estimation (viability: 4): https://sciencetostartup.com/paper/learned-elevation-models-as-a-lightweight-alternative-to-lidar-for-radio-environment-map-estimation - A two-stage framework that predicts elevation maps from satellite RGB imagery, offering a lightweight alternative to LiDAR for radio environment map estimation.
- UniCreative: Unifying Long-form Logic and Short-form Sparkle via Reference-Free Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/unicreative-unifying-long-form-logic-and-short-form-sparkle-via-reference-free-reinforcement-learning - UniCreative is a unified reference-free reinforcement learning framework that reconciles long-form narrative coherence with short-form creative expression.
- OmniDiagram: Advancing Unified Diagram Code Generation via Visual Interrogation Reward (viability: 8): https://sciencetostartup.com/paper/omnidiagram-advancing-unified-diagram-code-generation-via-visual-interrogation-reward - A unified framework for diagram code generation that uses visual feedback to train models without manual code annotation, establishing a new state-of-the-art.
- Thinking Diffusion: Penalize and Guide Visual-Grounded Reasoning in Diffusion Multimodal Language Models (viability: 8): https://sciencetostartup.com/paper/thinking-diffusion-penalize-and-guide-visual-grounded-reasoning-in-diffusion-multimodal-language-models - Enhancing diffusion multimodal language models with penalties and guidance to improve visual grounding and reasoning accuracy while accelerating inference.
- SCMAPR: Self-Correcting Multi-Agent Prompt Refinement for Complex-Scenario Text-to-Video Generation (viability: 8): https://sciencetostartup.com/paper/scmapr-self-correcting-multi-agent-prompt-refinement-for-complex-scenario-text-to-video-generation - A multi-agent framework that refines text prompts for complex text-to-video scenarios, improving alignment and generation quality with a new benchmark.
- Auditable Agents (viability: 7): https://sciencetostartup.com/paper/auditable-agents - A framework for auditable LLM agents that ensures accountability through a multi-dimensional approach to tracking and recovering agent actions.
- Can We Trust a Black-box LLM? LLM Untrustworthy Boundary Detection via Bias-Diffusion and Multi-Agent Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/can-we-trust-a-black-box-llm-llm-untrustworthy-boundary-detection-via-bias-diffusion-and-multi-agent-reinforcement-learn - A novel algorithm detects untrustworthy topic boundaries in black-box LLMs using knowledge graphs and multi-agent reinforcement learning, with a new dataset released for popular LLMs.
- Unifying VLM-Guided Flow Matching and Spectral Anomaly Detection for Interpretable Veterinary Diagnosis (viability: 8): https://sciencetostartup.com/paper/unifying-vlm-guided-flow-matching-and-spectral-anomaly-detection-for-interpretable-veterinary-diagnosis - A novel veterinary diagnostic system uses Vision-Language Models to guide Flow Matching for precise localization and Random Matrix Theory for interpretable anomaly detection in canine pneumothorax.
- On the Role of Fault Localization Context for LLM-Based Program Repair (viability: 3): https://sciencetostartup.com/paper/on-the-role-of-fault-localization-context-for-llm-based-program-repair - This research empirically studies the impact of fault localization context on LLM-based program repair, finding that file-level context is dominant and more context doesn't always improve performance.
- OntoTKGE: Ontology-Enhanced Temporal Knowledge Graph Extrapolation (viability: 7): https://sciencetostartup.com/paper/ontotkge-ontology-enhanced-temporal-knowledge-graph-extrapolation - OntoTKGE is a novel framework that enhances temporal knowledge graph extrapolation by integrating ontological knowledge to address entity sparsity and improve prediction performance.
- Adaptive Serverless Resource Management via Slot-Survival Prediction and Event-Driven Lifecycle Control (viability: 2): https://sciencetostartup.com/paper/adaptive-serverless-resource-management-via-slot-survival-prediction-and-event-driven-lifecycle-control - An adaptive framework for serverless computing that reduces cold starts and improves cost-efficiency through probabilistic modeling and event-driven control.
- LLM Evaluation as Tensor Completion: Low Rank Structure and Semiparametric Efficiency (viability: 1): https://sciencetostartup.com/paper/llm-evaluation-as-tensor-completion-low-rank-structure-and-semiparametric-efficiency - A novel tensor completion framework for semiparametric inference and uncertainty quantification in large language model evaluation using pairwise human judgments.
- MA-IDS: Multi-Agent RAG Framework for IoT Network Intrusion Detection with an Experience Library (viability: 7): https://sciencetostartup.com/paper/ma-ids-multi-agent-rag-framework-for-iot-network-intrusion-detection-with-an-experience-library - A multi-agent system using LLMs and RAG with an experience library for explainable, self-improving IoT network intrusion detection.
- Learning What Matters: Dynamic Dimension Selection and Aggregation for Interpretable Vision-Language Reward Modeling (viability: 7): https://sciencetostartup.com/paper/learning-what-matters-dynamic-dimension-selection-and-aggregation-for-interpretable-vision-language-reward-modeling - A framework for interpretable vision-language reward modeling that dynamically decomposes evaluation into granular, weighted dimensions.
- LanG -- A Governance-Aware Agentic AI Platform for Unified Security Operations (viability: 8): https://sciencetostartup.com/paper/lang-a-governance-aware-agentic-ai-platform-for-unified-security-operations - An open-source, governance-aware AI platform for unified security operations that unifies incident context, orchestrates agents, generates security rules, reconstructs attacks, and enforces AI governance policies.
- Human Interaction-Aware 3D Reconstruction from a Single Image (viability: 7): https://sciencetostartup.com/paper/human-interaction-aware-3d-reconstruction-from-a-single-image - A holistic method for reconstructing textured 3D human models from a single image, explicitly modeling group-level context and interaction priors to resolve occlusions and generate physically plausible, high-fidelity reconstructions of interacting people.
- Automated Auditing of Hospital Discharge Summaries for Care Transitions (viability: 7): https://sciencetostartup.com/paper/automated-auditing-of-hospital-discharge-summaries-for-care-transitions - An automated framework using locally deployed LLMs to audit hospital discharge summaries for care transitions, identifying key documentation elements to improve patient safety and reduce readmissions.
- Your LLM Agent Can Leak Your Data: Data Exfiltration via Backdoored Tool Use (viability: 4): https://sciencetostartup.com/paper/your-llm-agent-can-leak-your-data-data-exfiltration-via-backdoored-tool-use - A data exfiltration attack that embeds semantic triggers into fine-tuned LLM agents, enabling backdoored agents to invoke memory-access tool calls and exfiltrate stored user context via disguised retrieval tool calls.
- Bridging Natural Language and Microgrid Dynamics: A Context-Aware Simulator and Dataset (viability: 7): https://sciencetostartup.com/paper/bridging-natural-language-and-microgrid-dynamics-a-context-aware-simulator-and-dataset - OpenCEM is an open-source simulator and dataset that integrates contextual information with renewable energy dynamics for intelligent energy management.
- ALTO: Adaptive LoRA Tuning and Orchestration for Heterogeneous LoRA Training Workloads (viability: 6): https://sciencetostartup.com/paper/alto-adaptive-lora-tuning-and-orchestration-for-heterogeneous-lora-training-workloads - ALTO accelerates LoRA hyperparameter tuning and improves GPU utilization by orchestrating heterogeneous fine-tuning jobs.
- PRISM-MCTS: Learning from Reasoning Trajectories with Metacognitive Reflection (viability: 7): https://sciencetostartup.com/paper/prism-mcts-learning-from-reasoning-trajectories-with-metacognitive-reflection - PRISM-MCTS enhances reasoning models by learning from trajectories with metacognitive reflection, reducing computational redundancy and improving efficiency.
- VideoStir: Understanding Long Videos via Spatio-Temporally Structured and Intent-Aware RAG (viability: 8): https://sciencetostartup.com/paper/videostir-understanding-long-videos-via-spatio-temporally-structured-and-intent-aware-rag - VideoStir is a structured, intent-aware RAG framework for understanding long videos by leveraging spatio-temporal graphs and an MLLM-backed relevance scorer.
- Multi-Agent Pathfinding with Non-Unit Integer Edge Costs via Enhanced Conflict-Based Search and Graph Discretization (viability: 2): https://sciencetostartup.com/paper/multi-agent-pathfinding-with-non-unit-integer-edge-costs-via-enhanced-conflict-based-search-and-graph-discretization - A novel Multi-Agent Pathfinding variant on graphs with non-unit integer costs and an enhanced Conflict-Based Search framework for improved realism and efficiency.
- CODESTRUCT: Code Agents over Structured Action Spaces (viability: 5): https://sciencetostartup.com/paper/codestruct-code-agents-over-structured-action-spaces - CODESTRUCT reframes code agents to operate on structured action spaces, improving accuracy and reducing costs by treating repositories as AST entities instead of text.
- HYVE: Hybrid Views for LLM Context Engineering over Machine Data (viability: 6): https://sciencetostartup.com/paper/hyve-hybrid-views-for-llm-context-engineering-over-machine-data - HYVE is a framework for LLM context engineering that reduces token usage and improves output quality for machine data by using database principles for preprocessing and postprocessing.
- Reason Analogically via Cross-domain Prior Knowledge: An Empirical Study of Cross-domain Knowledge Transfer for In-Context Learning (viability: 5): https://sciencetostartup.com/paper/reason-analogically-via-cross-domain-prior-knowledge-an-empirical-study-of-cross-domain-knowledge-transfer-for-in-contex - This study validates the feasibility of cross-domain knowledge transfer for in-context learning, showing that source-domain demonstrations can improve target-domain inference despite semantic mismatch.
- Neural Assistive Impulses: Synthesizing Exaggerated Motions for Physics-based Characters (viability: 3): https://sciencetostartup.com/paper/neural-assistive-impulses-synthesizing-exaggerated-motions-for-physics-based-characters - A framework that synthesizes exaggerated character motions for animation by reformulating external assistance in impulse space for numerical stability.
- Towards Effective In-context Cross-domain Knowledge Transfer via Domain-invariant-neurons-based Retrieval (viability: 7): https://sciencetostartup.com/paper/towards-effective-in-context-cross-domain-knowledge-transfer-via-domain-invariant-neurons-based-retrieval - A retrieval method that boosts LLM reasoning by finding structurally compatible cross-domain demonstrations using domain-invariant neurons.
- LLM-as-Judge for Semantic Judging of Powerline Segmentation in UAV Inspection (viability: 5): https://sciencetostartup.com/paper/llm-as-judge-for-semantic-judging-of-powerline-segmentation-in-uav-inspection - Leveraging large language models as semantic judges to assess the reliability of power line segmentation in drone inspections.
- AI and Collective Decisions: Strengthening Legitimacy and Losers' Consent (viability: 6): https://sciencetostartup.com/paper/ai-and-collective-decisions-strengthening-legitimacy-and-losers-consent - A system using an AI interviewer and interactive visualization to increase perceived legitimacy and trust in collective decision-making.
- 3DTurboQuant: Training-Free Near-Optimal Quantization for 3D Reconstruction Models (viability: 8): https://sciencetostartup.com/paper/3dturboquant-training-free-near-optimal-quantization-for-3d-reconstruction-models - A training-free method to compress 3D reconstruction models like 3DGS and NeRF by up to 7.9x with minimal fidelity loss, enabling faster deployment and reduced storage.
- ETR: Entropy Trend Reward for Efficient Chain-of-Thought Reasoning (viability: 7): https://sciencetostartup.com/paper/etr-entropy-trend-reward-for-efficient-chain-of-thought-reasoning - This research introduces a novel reward mechanism (ETR) to significantly improve the efficiency and accuracy of large language model chain-of-thought reasoning by guiding uncertainty reduction.
- DQA: Diagnostic Question Answering for IT Support (viability: 6): https://sciencetostartup.com/paper/dqa-diagnostic-question-answering-for-it-support - DQA is a diagnostic question-answering framework for IT support that maintains a persistent diagnostic state to systematically troubleshoot issues, significantly improving success rates and reducing resolution time.
- From Retinal Evidence to Safe Decisions: RETINA-SAFE and ECRT for Hallucination Risk Triage in Medical LLMs (viability: 7): https://sciencetostartup.com/paper/from-retinal-evidence-to-safe-decisions-retina-safe-and-ecrt-for-hallucination-risk-triage-in-medical-llms - RETINA-SAFE and ECRT offer a practical, interpretable solution for triaging hallucination risks in medical LLMs by grounding detection in retinal evidence and classifying risk types.
- Dynamic Agentic AI Expert Profiler System Architecture for Multidomain Intelligence Modeling (viability: 5): https://sciencetostartup.com/paper/dynamic-agentic-ai-expert-profiler-system-architecture-for-multidomain-intelligence-modeling - This system dynamically profiles user expertise in real-time during human-machine interactions, achieving high accuracy in classifying skill levels across diverse domains.
- Anchored Cyclic Generation: A Novel Paradigm for Long-Sequence Symbolic Music Generation (viability: 4): https://sciencetostartup.com/paper/anchored-cyclic-generation-a-novel-paradigm-for-long-sequence-symbolic-music-generation - A novel paradigm for long-sequence symbolic music generation that mitigates error accumulation using anchor features and a hierarchical framework.
- TRACE: Capability-Targeted Agentic Training (viability: 6): https://sciencetostartup.com/paper/trace-capability-targeted-agentic-training - An end-to-end system for agent self-improvement that identifies lacking capabilities and synthesizes targeted training environments using LoRA adapters.
- Graph of Skills: Dependency-Aware Structural Retrieval for Massive Agent Skills (viability: 7): https://sciencetostartup.com/paper/graph-of-skills-dependency-aware-structural-retrieval-for-massive-agent-skills - A retrieval layer for large agent skill libraries that constructs an executable skill graph and retrieves dependency-aware skill bundles at inference time.
- LLMs Should Express Uncertainty Explicitly (viability: 4): https://sciencetostartup.com/paper/llms-should-express-uncertainty-explicitly - LLMs should express uncertainty explicitly through calibrated confidence scores or reasoning-time markers to improve decision-making and error handling.
- Breakthrough the Suboptimal Stable Point in Value-Factorization-Based Multi-Agent Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/breakthrough-the-suboptimal-stable-point-in-value-factorization-based-multi-agent-reinforcement-learning - A novel Multi-Round Value Factorization framework that breaks suboptimal stable points in multi-agent reinforcement learning by iteratively filtering inferior actions.
- Broken by Default: A Formal Verification Study of Security Vulnerabilities in AI-Generated Code (viability: 4): https://sciencetostartup.com/paper/broken-by-default-a-formal-verification-study-of-security-vulnerabilities-in-ai-generated-code - A formal verification study revealing that over half of AI-generated code artifacts contain security vulnerabilities, with no frontier LLM achieving a passing grade.
- Pressure, What Pressure? Sycophancy Disentanglement in Language Models via Reward Decomposition (viability: 7): https://sciencetostartup.com/paper/pressure-what-pressure-sycophancy-disentanglement-in-language-models-via-reward-decomposition - A reward decomposition approach that disentangles sycophancy in language models by separating pressure capitulation from evidence blindness, leading to more robust and factual responses.
- Spec Kit Agents: Context-Grounded Agentic Workflows (viability: 4): https://sciencetostartup.com/paper/spec-kit-agents-context-grounded-agentic-workflows - Spec Kit Agents enhance AI coding assistants by grounding them in repository context, reducing hallucinations and architectural violations for more reliable software development.
- Simulating the Evolution of Alignment and Values in Machine Intelligence (viability: 3): https://sciencetostartup.com/paper/simulating-the-evolution-of-alignment-and-values-in-machine-intelligence - This research explores the evolutionary dynamics of AI alignment and values, proposing a method to reduce deception in models through improved testing and adaptive design.
- Region-R1: Reinforcing Query-Side Region Cropping for Multi-Modal Re-Ranking (viability: 7): https://sciencetostartup.com/paper/region-r1-reinforcing-query-side-region-cropping-for-multi-modal-re-ranking - Region-R1 enhances multi-modal retrieval by intelligently cropping query images to focus on relevant regions, significantly improving performance on challenging benchmarks.
- Extending Tabular Denoising Diffusion Probabilistic Models for Time-Series Data Generation (viability: 6): https://sciencetostartup.com/paper/extending-tabular-denoising-diffusion-probabilistic-models-for-time-series-data-generation - This work extends diffusion models for time-series data generation by incorporating temporal adapters and context-aware modules to produce temporally coherent synthetic sequences.
- EAGLE: Edge-Aware Graph Learning for Proactive Delivery Delay Prediction in Smart Logistics Networks (viability: 7): https://sciencetostartup.com/paper/eagle-edge-aware-graph-learning-for-proactive-delivery-delay-prediction-in-smart-logistics-networks - EAGLE is a hybrid deep learning framework that proactively predicts delivery delays in logistics networks by jointly modeling temporal dynamics and graph-based dependencies.
- Exemplar Retrieval Without Overhypothesis Induction: Limits of Distributional Sequence Learning in Early Word Learning (viability: 2): https://sciencetostartup.com/paper/exemplar-retrieval-without-overhypothesis-induction-limits-of-distributional-sequence-learning-in-early-word-learning - This research explores the limitations of current language models in achieving higher-order generalization for early word learning, suggesting a gap in their inductive capabilities.
- XMark: Reliable Multi-Bit Watermarking for LLM-Generated Texts (viability: 7): https://sciencetostartup.com/paper/xmark-reliable-multi-bit-watermarking-for-llm-generated-texts - XMark is a novel multi-bit watermarking method for LLM-generated text that reliably embeds messages with high decoding accuracy and preserves text quality, even with limited tokens.
- Curvature-Aware Optimization for High-Accuracy Physics-Informed Neural Networks (viability: 4): https://sciencetostartup.com/paper/curvature-aware-optimization-for-high-accuracy-physics-informed-neural-networks - This paper introduces advanced optimization strategies for Physics-Informed Neural Networks (PINNs) to accelerate convergence and achieve high accuracy in solving complex differential equations.
- From Governance Norms to Enforceable Controls: A Layered Translation Method for Runtime Guardrails in Agentic AI (viability: 3): https://sciencetostartup.com/paper/from-governance-norms-to-enforceable-controls-a-layered-translation-method-for-runtime-guardrails-in-agentic-ai - This paper proposes a layered translation method to connect governance norms to enforceable runtime guardrails for agentic AI systems.
- RoboPlayground: Democratizing Robotic Evaluation through Structured Physical Domains (viability: 7): https://sciencetostartup.com/paper/roboplayground-democratizing-robotic-evaluation-through-structured-physical-domains - A framework for democratizing robotic evaluation by enabling natural language task authoring and crowd-sourced contribution.
- Attribution Bias in Large Language Models (viability: 7): https://sciencetostartup.com/paper/attribution-bias-in-large-language-models - A new benchmark dataset and evaluation framework to identify and mitigate attribution bias in large language models.
- Improving Clinical Trial Recruitment using Clinical Narratives and Large Language Models (viability: 7): https://sciencetostartup.com/paper/improving-clinical-trial-recruitment-using-clinical-narratives-and-large-language-models - Leveraging LLMs with RAG and summarization to significantly improve clinical trial patient screening efficiency.
- OrthoFuse: Training-free Riemannian Fusion of Orthogonal Style-Concept Adapters for Diffusion Models (viability: 7): https://sciencetostartup.com/paper/orthofuse-training-free-riemannian-fusion-of-orthogonal-style-concept-adapters-for-diffusion-models - A training-free method to fuse orthogonal style and concept adapters for diffusion models, enabling combined feature generation.
- LSRM: High-Fidelity Object-Centric Reconstruction via Scaled Context Windows (viability: 7): https://sciencetostartup.com/paper/lsrm-high-fidelity-object-centric-reconstruction-via-scaled-context-windows - A novel 3D reconstruction model that significantly improves fine-grained texture and appearance recovery by scaling transformer context windows, outperforming state-of-the-art.
- ClawsBench: Evaluating Capability and Safety of LLM Productivity Agents in Simulated Workspaces (viability: 8): https://sciencetostartup.com/paper/clawsbench-evaluating-capability-and-safety-of-llm-productivity-agents-in-simulated-workspaces - ClawsBench provides a realistic, stateful benchmark for evaluating LLM productivity agents across multiple services, revealing significant capability and safety gaps.
- Modality-Aware and Anatomical Vector-Quantized Autoencoding for Multimodal Brain MRI (viability: 7): https://sciencetostartup.com/paper/modality-aware-and-anatomical-vector-quantized-autoencoding-for-multimodal-brain-mri - NeuroQuant is a modality-aware VQ-VAE for multimodal brain MRI reconstruction, capturing anatomical structures and appearance for improved generative modeling and analysis.
- Instruction-Tuned LLMs for Parsing and Mining Unstructured Logs on Leadership HPC Systems (viability: 8): https://sciencetostartup.com/paper/instruction-tuned-llms-for-parsing-and-mining-unstructured-logs-on-leadership-hpc-systems - A domain-adapted, instruction-tuned LLM framework for parsing and mining unstructured HPC logs, achieving state-of-the-art accuracy with a locally deployable, energy-efficient approach.
- From Use to Oversight: How Mental Models Influence User Behavior and Output in AI Writing Assistants (viability: 4): https://sciencetostartup.com/paper/from-use-to-oversight-how-mental-models-influence-user-behavior-and-output-in-ai-writing-assistants - This research explores how users' mental models of AI writing assistants impact their control behavior and output quality, revealing a complex relationship between system understanding, trust, and oversight.
- Learning to Focus: CSI-Free Hierarchical MARL for Reconfigurable Reflectors (viability: 7): https://sciencetostartup.com/paper/learning-to-focus-csi-free-hierarchical-marl-for-reconfigurable-reflectors - A CSI-free hierarchical MARL system for reconfigurable reflectors that achieves significant RSSI improvements and robust multi-user scalability for intelligent wireless environments.
- Not All Turns Are Equally Hard: Adaptive Thinking Budgets For Efficient Multi-Turn Reasoning (viability: 7): https://sciencetostartup.com/paper/not-all-turns-are-equally-hard-adaptive-thinking-budgets-for-efficient-multi-turn-reasoning - TAB: Turn-Adaptive Budgets, a novel policy for LLMs that optimizes token allocation in multi-turn reasoning to improve accuracy while significantly reducing compute costs.
- What Makes a Good Response? An Empirical Analysis of Quality in Qualitative Interviews (viability: 5): https://sciencetostartup.com/paper/what-makes-a-good-response-an-empirical-analysis-of-quality-in-qualitative-interviews - This work empirically analyzes quality metrics for qualitative interview responses, identifying direct relevance to research questions as the strongest predictor of quality and introducing a new dataset.
- Bypassing the CSI Bottleneck: MARL-Driven Spatial Control for Reflector Arrays (viability: 4): https://sciencetostartup.com/paper/bypassing-the-csi-bottleneck-marl-driven-spatial-control-for-reflector-arrays - AI-driven spatial control for reflector arrays bypasses computational bottlenecks in wireless networks using Multi-Agent Reinforcement Learning.
- Planning to Explore: Curiosity-Driven Planning for LLM Test Generation (viability: 7): https://sciencetostartup.com/paper/planning-to-explore-curiosity-driven-planning-for-llm-test-generation - Curiosity-driven planning for LLMs enhances test generation by prioritizing exploration of program branches for deeper code coverage.
- IntentScore: Intent-Conditioned Action Evaluation for Computer-Use Agents (viability: 7): https://sciencetostartup.com/paper/intentscore-intent-conditioned-action-evaluation-for-computer-use-agents - IntentScore evaluates and ranks actions for computer-use agents, improving task success rates by learning from offline GUI interactions.
- Compiled AI: Deterministic Code Generation for LLM-Based Workflow Automation (viability: 8): https://sciencetostartup.com/paper/compiled-ai-deterministic-code-generation-for-llm-based-workflow-automation - Compiled AI generates deterministic code from LLMs for reliable and cost-efficient enterprise workflow automation, especially in healthcare.
- A mathematical theory of evolution for self-designing AIs (viability: 2): https://sciencetostartup.com/paper/a-mathematical-theory-of-evolution-for-self-designing-ais - Develops a mathematical model for the evolution of self-designing AI systems, considering directed design and human control through fitness functions, with implications for AI alignment.
- EffiPair: Improving the Efficiency of LLM-generated Code with Relative Contrastive Feedback (viability: 7): https://sciencetostartup.com/paper/effipair-improving-the-efficiency-of-llm-generated-code-with-relative-contrastive-feedback - EffiPair is an inference-time framework that uses relative contrastive feedback to significantly improve the runtime and memory efficiency of LLM-generated code without model fine-tuning.
- Non-monotonic causal discovery with Kolmogorov-Arnold Fuzzy Cognitive Maps (viability: 5): https://sciencetostartup.com/paper/non-monotonic-causal-discovery-with-kolmogorov-arnold-fuzzy-cognitive-maps - Introduces Kolmogorov-Arnold Fuzzy Cognitive Maps (KA-FCMs) that use learnable B-spline functions to model non-monotonic causal relationships in complex dynamic systems while preserving interpretability.
- Reasoning Through Chess: How Reasoning Evolves from Data Through Fine-Tuning and Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/reasoning-through-chess-how-reasoning-evolves-from-data-through-fine-tuning-and-reinforcement-learning - This research analyzes how reasoning evolves in language models for chess through fine-tuning and reinforcement learning, demonstrating improved performance and faithful reasoning with released checkpoints and code.
- Offline RL for Adaptive Policy Retrieval in Prior Authorization (viability: 4): https://sciencetostartup.com/paper/offline-rl-for-adaptive-policy-retrieval-in-prior-authorization - An adaptive policy retrieval system for prior authorization using offline RL to balance decision correctness with retrieval efficiency.
- Watch Before You Answer: Learning from Visually Grounded Post-Training (viability: 7): https://sciencetostartup.com/paper/watch-before-you-answer-learning-from-visually-grounded-post-training - Enhance video understanding models by leveraging visually grounded post-training to surpass state-of-the-art performance.
- Uncertainty-Guided Latent Diagnostic Trajectory Learning for Sequential Clinical Diagnosis (viability: 7): https://sciencetostartup.com/paper/uncertainty-guided-latent-diagnostic-trajectory-learning-for-sequential-clinical-diagnosis - A framework for sequential clinical diagnosis that uses uncertainty-guided latent trajectory learning to improve accuracy and reduce diagnostic tests.
- $π^2$: Structure-Originated Reasoning Data Improves Long-Context Reasoning Ability of Large Language Models (viability: 7): https://sciencetostartup.com/paper/2-structure-originated-reasoning-data-improves-long-context-reasoning-ability-of-large-language-models - A pipeline for curating reasoning data from structured sources to enhance LLM long-context reasoning capabilities, with open-source code and data.
- CRAB: Codebook Rebalancing for Bias Mitigation in Generative Recommendation (viability: 4): https://sciencetostartup.com/paper/crab-codebook-rebalancing-for-bias-mitigation-in-generative-recommendation - A post-hoc debiasing strategy for generative recommendation systems that rebalances item tokenization to mitigate popularity bias and improve recommendation performance.
- Vintix II: Decision Pre-Trained Transformer is a Scalable In-Context Reinforcement Learner (viability: 7): https://sciencetostartup.com/paper/vintix-ii-decision-pre-trained-transformer-is-a-scalable-in-context-reinforcement-learner - A scalable Decision Pre-Trained Transformer trained with Flow Matching achieves strong generalization in multi-domain in-context reinforcement learning, offering a viable alternative to expert distillation for generalist agents.
- Simultaneous Dual-View Mammogram Synthesis Using Denoising Diffusion Probabilistic Models (viability: 6): https://sciencetostartup.com/paper/simultaneous-dual-view-mammogram-synthesis-using-denoising-diffusion-probabilistic-models - A three-channel denoising diffusion probabilistic model synthesizes dual-view mammograms simultaneously, addressing data gaps and enabling cross-view AI applications in breast imaging.
- Edit, But Verify: An Empirical Audit of Instructed Code-Editing Benchmarks (viability: 3): https://sciencetostartup.com/paper/edit-but-verify-an-empirical-audit-of-instructed-code-editing-benchmarks - An empirical audit of instructed code-editing benchmarks reveals significant gaps compared to real-world usage, proposing desiderata for more representative benchmarks.
- Beyond LLM-as-a-Judge: Deterministic Metrics for Multilingual Generative Text Evaluation (viability: 7): https://sciencetostartup.com/paper/beyond-llm-as-a-judge-deterministic-metrics-for-multilingual-generative-text-evaluation - Develops deterministic learned metrics to replace costly and inconsistent LLM-based text evaluation, offering a scalable and reproducible alternative.
- MedGemma 1.5 Technical Report (viability: 8): https://sciencetostartup.com/paper/medgemma-1-5-technical-report - Introduces MedGemma 1.5, a multimodal foundation model for medical AI, integrating imaging, EHRs, and clinical reasoning with significant performance gains.
- Nidus: Externalized Reasoning for AI-Assisted Engineering (viability: 3): https://sciencetostartup.com/paper/nidus-externalized-reasoning-for-ai-assisted-engineering - Presents Nidus, a governance runtime that mechanizes the V-model for AI-assisted software delivery, ensuring engineering invariants through externalized reasoning.
- Feature-Aware Anisotropic Local Differential Privacy for Utility-Preserving Graph Representation Learning in Metal Additive Manufacturing (viability: 5): https://sciencetostartup.com/paper/feature-aware-anisotropic-local-differential-privacy-for-utility-preserving-graph-representation-learning-in-metal-addit - Proposes FI-LDP-HGAT, a privacy-preserving graph learning framework for metal additive manufacturing that balances utility and privacy by allocating noise based on feature importance.
- MMORF: A Multi-agent Framework for Designing Multi-objective Retrosynthesis Planning Systems (viability: 7): https://sciencetostartup.com/paper/mmorf-a-multi-agent-framework-for-designing-multi-objective-retrosynthesis-planning-systems - A framework for building multi-agent systems to solve complex chemistry retrosynthesis planning problems with improved safety and cost metrics.
- Part-Level 3D Gaussian Vehicle Generation with Joint and Hinge Axis Estimation (viability: 3): https://sciencetostartup.com/paper/part-level-3d-gaussian-vehicle-generation-with-joint-and-hinge-axis-estimation - A generative framework for creating animatable 3D vehicle models from single or sparse multi-view images for realistic autonomous driving simulation.
- AutoLALA: Automatic Loop Algebraic Locality Analysis for AI and HPC Kernels (viability: 3): https://sciencetostartup.com/paper/autolala-automatic-loop-algebraic-locality-analysis-for-ai-and-hpc-kernels - An open-source tool that automatically analyzes data locality and movement complexity in AI and HPC kernels to optimize performance.
- Dynamic Linear Coregionalization for Realistic Synthetic Multivariate Time Series (viability: 7): https://sciencetostartup.com/paper/dynamic-linear-coregionalization-for-realistic-synthetic-multivariate-time-series - A dynamic model for generating realistic synthetic multivariate time series with time-varying correlations to improve foundation model training.
- This Treatment Works, Right? Evaluating LLM Sensitivity to Patient Question Framing in Medical QA (viability: 7): https://sciencetostartup.com/paper/this-treatment-works-right-evaluating-llm-sensitivity-to-patient-question-framing-in-medical-qa - This research evaluates LLM sensitivity to patient question framing in medical QA, highlighting a critical need for robust phrasing in high-stakes applications.
- PCA-Driven Adaptive Sensor Triage for Edge AI Inference (viability: 8): https://sciencetostartup.com/paper/pca-driven-adaptive-sensor-triage-for-edge-ai-inference - PCA-Triage is an unsupervised, parameter-free algorithm for adaptive sensor sampling on edge devices, significantly reducing bandwidth while maintaining high inference accuracy.
- ID-Sim: An Identity-Focused Similarity Metric (viability: 8): https://sciencetostartup.com/paper/id-sim-an-identity-focused-similarity-metric - ID-Sim is a novel feed-forward metric that accurately reflects human selective sensitivity to identities, accelerating progress in personalized image generation and identity-focused tasks.
- Phase-Associative Memory: Sequence Modeling in Complex Hilbert Space (viability: 5): https://sciencetostartup.com/paper/phase-associative-memory-sequence-modeling-in-complex-hilbert-space - Phase-Associative Memory (PAM) is a complex-valued recurrent sequence model that shows competitive performance with transformers on WikiText-103, exploring non-classical contextuality in language modeling.
- PaperOrchestra: A Multi-Agent Framework for Automated AI Research Paper Writing (viability: 7): https://sciencetostartup.com/paper/paperorchestra-a-multi-agent-framework-for-automated-ai-research-paper-writing - A multi-agent framework that automates the writing of AI research papers from raw materials, including literature synthesis and visuals.
- Vanast: Virtual Try-On with Human Image Animation via Synthetic Triplet Supervision (viability: 7): https://sciencetostartup.com/paper/vanast-virtual-try-on-with-human-image-animation-via-synthetic-triplet-supervision - A unified framework for generating garment-transferred human animation videos from a single image, garment, and pose, overcoming identity drift and distortion.
- PointTPA: Dynamic Network Parameter Adaptation for 3D Scene Understanding (viability: 7): https://sciencetostartup.com/paper/pointtpa-dynamic-network-parameter-adaptation-for-3d-scene-understanding - A framework that dynamically adapts network parameters at test-time to significantly improve 3D scene understanding with minimal overhead.
- Beyond the Final Actor: Modeling the Dual Roles of Creator and Editor for Fine-Grained LLM-Generated Text Detection (viability: 4): https://sciencetostartup.com/paper/beyond-the-final-actor-modeling-the-dual-roles-of-creator-and-editor-for-fine-grained-llm-generated-text-detection - A fine-grained detection method for LLM-generated text that distinguishes between creator and editor roles to enable nuanced policy enforcement.
- LoMa: Local Feature Matching Revisited (viability: 7): https://sciencetostartup.com/paper/loma-local-feature-matching-revisited - A data-driven approach to local feature matching that significantly outperforms state-of-the-art, with code and models released for 3D vision applications.
- Early Stopping for Large Reasoning Models via Confidence Dynamics (viability: 7): https://sciencetostartup.com/paper/early-stopping-for-large-reasoning-models-via-confidence-dynamics - Optimize large language model computations by integrating early stopping mechanisms to reduce costs.
- Rethinking Model Efficiency: Multi-Agent Inference with Large Models (viability: 7): https://sciencetostartup.com/paper/rethinking-model-efficiency-multi-agent-inference-with-large-models - A multi-agent inference framework that leverages large models with short responses by transferring reasoning tokens from smaller models to improve efficiency and performance.
- Comprehensive List of User Deception Techniques in Emails (viability: 3): https://sciencetostartup.com/paper/comprehensive-list-of-user-deception-techniques-in-emails - A structured catalog of 42 email deception techniques with 64 examples to inform the development of countermeasures.
- Fully Procedural Synthetic Data from Simple Rules for Multi-View Stereo (viability: 7): https://sciencetostartup.com/paper/fully-procedural-synthetic-data-from-simple-rules-for-multi-view-stereo - Generate high-quality synthetic data for multi-view stereo with a procedural engine that outperforms manually curated datasets.
- Your Pre-trained Diffusion Model Secretly Knows Restoration (viability: 7): https://sciencetostartup.com/paper/your-pre-trained-diffusion-model-secretly-knows-restoration - Unlock the hidden restoration capabilities of pre-trained diffusion models using prompt embeddings.
- Stratifying Reinforcement Learning with Signal Temporal Logic (viability: 3): https://sciencetostartup.com/paper/stratifying-reinforcement-learning-with-signal-temporal-logic - A theoretical framework for analyzing deep reinforcement learning embeddings using signal temporal logic and stratification theory.
- TriAttention: Efficient Long Reasoning with Trigonometric KV Compression (viability: 7): https://sciencetostartup.com/paper/triattention-efficient-long-reasoning-with-trigonometric-kv-compression - A novel KV cache compression technique for LLMs that significantly improves reasoning efficiency and memory usage, enabling deployment on consumer hardware.
- Vero: An Open RL Recipe for General Visual Reasoning (viability: 8): https://sciencetostartup.com/paper/vero-an-open-rl-recipe-for-general-visual-reasoning - Vero provides an open, state-of-the-art RL recipe for enhancing vision-language models across diverse reasoning tasks.
- Empowering Power Outage Prediction with Spatially Aware Hybrid Graph Neural Networks and Contrastive Learning (viability: 7): https://sciencetostartup.com/paper/empowering-power-outage-prediction-with-spatially-aware-hybrid-graph-neural-networks-and-contrastive-learning - A hybrid graph neural network with contrastive learning to significantly improve power outage prediction accuracy for extreme weather events.
- Analyzing Symbolic Properties for DRL Agents in Systems and Networking (viability: 3): https://sciencetostartup.com/paper/analyzing-symbolic-properties-for-drl-agents-in-systems-and-networking - A framework for analyzing symbolic properties of DRL agents in systems and networking, enabling broader coverage and uncovering operational counterexamples.
- A Frame is Worth One Token: Efficient Generative World Modeling with Delta Tokens (viability: 7): https://sciencetostartup.com/paper/a-frame-is-worth-one-token-efficient-generative-world-modeling-with-delta-tokens - A generative world model that efficiently predicts diverse future video frames by tokenizing feature differences, achieving state-of-the-art results with significantly fewer parameters and FLOPs.
- SpatialEdit: Benchmarking Fine-Grained Image Spatial Editing (viability: 7): https://sciencetostartup.com/paper/spatialedit-benchmarking-fine-grained-image-spatial-editing - A new benchmark and dataset for fine-grained image spatial editing, with a baseline model that outperforms prior methods.
- HI-MoE: Hierarchical Instance-Conditioned Mixture-of-Experts for Object Detection (viability: 4): https://sciencetostartup.com/paper/hi-moe-hierarchical-instance-conditioned-mixture-of-experts-for-object-detection - A hierarchical mixture-of-experts architecture for object detection that improves performance on small objects by routing experts at both scene and instance levels.
- How AI Aggregation Affects Knowledge (viability: 0): https://sciencetostartup.com/paper/how-ai-aggregation-affects-knowledge - Extends the DeGroot model to analyze how AI aggregation affects social learning, identifying a critical threshold in update speed for robust learning improvement.
- ClickAIXR: On-Device Multimodal Vision-Language Interaction with Real-World Objects in Extended Reality (viability: 7): https://sciencetostartup.com/paper/clickaixr-on-device-multimodal-vision-language-interaction-with-real-world-objects-in-extended-reality - On-device multimodal vision-language interaction for XR objects, enabling private and low-latency natural language queries with real-world items.
- Are Latent Reasoning Models Easily Interpretable? (viability: 4): https://sciencetostartup.com/paper/are-latent-reasoning-models-easily-interpretable - Develop tools to decode interpretable reasoning traces from latent models, revealing prediction correctness.
- FileGram: Grounding Agent Personalization in File-System Behavioral Traces (viability: 7): https://sciencetostartup.com/paper/filegram-grounding-agent-personalization-in-file-system-behavioral-traces - FileGram offers a personalized AI framework for integrating file-system behavioral traces into memory-centric coworker agents.
- QED-Nano: Teaching a Tiny Model to Prove Hard Theorems (viability: 7): https://sciencetostartup.com/paper/qed-nano-teaching-a-tiny-model-to-prove-hard-theorems - A 4B parameter model trained to prove complex mathematical theorems, outperforming larger open models and approaching proprietary performance at a fraction of the cost.
- Agentic Federated Learning: The Future of Distributed Training Orchestration (viability: 4): https://sciencetostartup.com/paper/agentic-federated-learning-the-future-of-distributed-training-orchestration - Autonomous AI agents orchestrated via language models to dynamically manage federated learning, improving privacy and adapting to system dynamics.
- Rethinking Exploration in RLVR: From Entropy Regularization to Refinement via Bidirectional Entropy Modulation (viability: 7): https://sciencetostartup.com/paper/rethinking-exploration-in-rlvr-from-entropy-regularization-to-refinement-via-bidirectional-entropy-modulation - A new framework for reinforcement learning in LLMs that refines exploration by modulating positive and negative rollouts, improving reasoning capabilities.
- Data Attribution in Adaptive Learning (viability: 3): https://sciencetostartup.com/paper/data-attribution-in-adaptive-learning - Developing novel attribution methods for adaptive machine learning systems where data generation is influenced by the model itself.
- Muon Dynamics as a Spectral Wasserstein Flow (viability: 0): https://sciencetostartup.com/paper/muon-dynamics-as-a-spectral-wasserstein-flow - This paper explores a family of spectral normalization rules for deep learning optimization, analyzing them in a mean-field regime using Spectral Wasserstein distances.
- HorizonWeaver: Generalizable Multi-Level Semantic Editing for Driving Scenes (viability: 7): https://sciencetostartup.com/paper/horizonweaver-generalizable-multi-level-semantic-editing-for-driving-scenes - HorizonWeaver enables scalable, realistic editing of driving scenes for safer autonomous driving simulations.
- Learning, Potential, and Retention: An Approach for Evaluating Adaptive AI-Enabled Medical Devices (viability: 7): https://sciencetostartup.com/paper/learning-potential-and-retention-an-approach-for-evaluating-adaptive-ai-enabled-medical-devices - A novel evaluation framework for adaptive AI medical devices that measures learning, potential, and retention to ensure safety and effectiveness across iterative updates.
- Incompleteness of AI Safety Verification via Kolmogorov Complexity (viability: 0): https://sciencetostartup.com/paper/incompleteness-of-ai-safety-verification-via-kolmogorov-complexity - This paper demonstrates an intrinsic information-theoretic limitation in AI safety verification, proving that no finite verifier can certify all policy-compliant instances of arbitrarily high complexity.
- DIRECT: Video Mashup Creation via Hierarchical Multi-Agent Planning and Intent-Guided Editing (viability: 8): https://sciencetostartup.com/paper/direct-video-mashup-creation-via-hierarchical-multi-agent-planning-and-intent-guided-editing - DIRECT is a hierarchical multi-agent system automating professional-grade video mashup creation for content creators.
- Free-Range Gaussians: Non-Grid-Aligned Generative 3D Gaussian Reconstruction (viability: 7): https://sciencetostartup.com/paper/free-range-gaussians-non-grid-aligned-generative-3d-gaussian-reconstruction - A novel method for high-fidelity 3D reconstruction from minimal images using generative Gaussians and hierarchical transformers, outperforming existing grid-aligned approaches.
- Synthetic Sandbox for Training Machine Learning Engineering Agents (viability: 7): https://sciencetostartup.com/paper/synthetic-sandbox-for-training-machine-learning-engineering-agents - SandMLE accelerates machine learning engineering agent training by generating micro-scale synthetic environments, enabling faster on-policy RL and outperforming SFT baselines.
- Optimizing LLM Prompt Engineering with DSPy Based Declarative Learning (viability: 7): https://sciencetostartup.com/paper/optimizing-llm-prompt-engineering-with-dspy-based-declarative-learning - Automate and optimize LLM prompt engineering for improved accuracy and reduced hallucinations using a declarative learning framework.
- Noise Immunity in In-Context Tabular Learning: An Empirical Robustness Analysis of TabPFN's Attention Mechanisms (viability: 5): https://sciencetostartup.com/paper/noise-immunity-in-in-context-tabular-learning-an-empirical-robustness-analysis-of-tabpfn-s-attention-mechanisms - An empirical analysis demonstrating the robustness of TabPFN's attention mechanisms in tabular foundation models against noise and data imperfections.
- Beyond the Global Scores: Fine-Grained Token Grounding as a Robust Detector of LVLM Hallucinations (viability: 7): https://sciencetostartup.com/paper/beyond-the-global-scores-fine-grained-token-grounding-as-a-robust-detector-of-lvlm-hallucinations - A lightweight and interpretable detector for large vision-language model hallucinations, leveraging fine-grained token-level analysis to achieve up to 90% accuracy.
- Outlier-Robust Nonlinear Moving Horizon Estimation using Adaptive Loss Functions (viability: 4): https://sciencetostartup.com/paper/outlier-robust-nonlinear-moving-horizon-estimation-using-adaptive-loss-functions - Develops an adaptive robust loss function for Moving Horizon Estimation to improve accuracy in the presence of data outliers.
- Unified Vector Floorplan Generation via Markup Representation (viability: 7): https://sciencetostartup.com/paper/unified-vector-floorplan-generation-via-markup-representation - A transformer model that generates high-fidelity residential floorplans from diverse inputs using a novel markup language.
- FairLogue: A Toolkit for Intersectional Fairness Analysis in Clinical Machine Learning Models (viability: 7): https://sciencetostartup.com/paper/fairlogue-a-toolkit-for-intersectional-fairness-analysis-in-clinical-machine-learning-models - A Python toolkit for analyzing intersectional fairness in clinical ML models, revealing hidden disparities missed by single-axis methods.
- The Blind Spot of Adaptation: Quantifying and Mitigating Forgetting in Fine-tuned Driving Models (viability: 7): https://sciencetostartup.com/paper/the-blind-spot-of-adaptation-quantifying-and-mitigating-forgetting-in-fine-tuned-driving-models - A novel framework that mitigates catastrophic forgetting in fine-tuned driving models by adapting prompts instead of weights, preserving foundational knowledge while achieving state-of-the-art performance.
- The Role of Generator Access in Autoregressive Post-Training (viability: 2): https://sciencetostartup.com/paper/the-role-of-generator-access-in-autoregressive-post-training - This paper theoretically explores how different levels of access to a generative model's internal states during post-training impact learning outcomes, without providing a practical implementation or dataset.
- MemMachine: A Ground-Truth-Preserving Memory System for Personalized AI Agents (viability: 8): https://sciencetostartup.com/paper/memmachine-a-ground-truth-preserving-memory-system-for-personalized-ai-agents - MemMachine is a memory system for AI agents that combines short-term and long-term memory to improve factual continuity and personalization.
- Strengthening Human-Centric Chain-of-Thought Reasoning Integrity in LLMs via a Structured Prompt Framework (viability: 7): https://sciencetostartup.com/paper/strengthening-human-centric-chain-of-thought-reasoning-integrity-in-llms-via-a-structured-prompt-framework - A structured prompt framework to enhance Chain-of-Thought reasoning integrity and security threat detection in LLMs for local deployments.
- InfBaGel: Human-Object-Scene Interaction Generation with Dynamic Perception and Iterative Refinement (viability: 7): https://sciencetostartup.com/paper/infbagel-human-object-scene-interaction-generation-with-dynamic-perception-and-iterative-refinement - A framework for generating human-object-scene interactions by dynamically updating scene context and using bump-aware guidance for real-time, artifact-free generation.
- Do No Harm: Exposing Hidden Vulnerabilities of LLMs via Persona-based Client Simulation Attack in Psychological Counseling (viability: 4): https://sciencetostartup.com/paper/do-no-harm-exposing-hidden-vulnerabilities-of-llms-via-persona-based-client-simulation-attack-in-psychological-counselin - A new framework for testing LLM safety in psychological counseling by simulating client personas to uncover hidden vulnerabilities.
- MERIT: Multilingual Expert-Reward Informed Tuning for Chinese-Centric Low-Resource Machine Translation (viability: 7): https://sciencetostartup.com/paper/merit-multilingual-expert-reward-informed-tuning-for-chinese-centric-low-resource-machine-translation - A new framework for Chinese-centric low-resource machine translation that significantly improves translation quality by combining data curation and reward-guided optimization.
- Less Detail, Better Answers: Degradation-Driven Prompting for VQA (viability: 7): https://sciencetostartup.com/paper/less-detail-better-answers-degradation-driven-prompting-for-vqa - A framework that strategically degrades image quality to improve VQA accuracy by forcing models to focus on essential information.
- E-VLA: Event-Augmented Vision-Language-Action Model for Dark and Blurred Scenes (viability: 7): https://sciencetostartup.com/paper/e-vla-event-augmented-vision-language-action-model-for-dark-and-blurred-scenes - A vision-language-action model enhanced with event camera data to enable robust robotic manipulation in challenging low-light and blurred conditions.
- Cryptanalysis of the Legendre Pseudorandom Function over Extension Fields (viability: 3): https://sciencetostartup.com/paper/cryptanalysis-of-the-legendre-pseudorandom-function-over-extension-fields - This paper analyzes the security of a cryptographic primitive over extension fields, revealing vulnerabilities and suggesting higher-degree variants for enhanced security.
- A Robust SINDy Autoencoder for Noisy Dynamical System Identification (viability: 3): https://sciencetostartup.com/paper/a-robust-sindy-autoencoder-for-noisy-dynamical-system-identification - A robust SINDy autoencoder for identifying noisy dynamical systems by incorporating a noise-separation module.
- Efficient Multi-Objective Planning with Weighted Maximization Using Large Neighbourhood Search (viability: 7): https://sciencetostartup.com/paper/efficient-multi-objective-planning-with-weighted-maximization-using-large-neighbourhood-search - A novel Large Neighbourhood Search algorithm significantly speeds up multi-objective planning for autonomous navigation, enabling the discovery of critical trade-off solutions previously missed by traditional methods.
- Plausibility as Commonsense Reasoning: Humans Succeed, Large Language Models Do not (viability: 3): https://sciencetostartup.com/paper/plausibility-as-commonsense-reasoning-humans-succeed-large-language-models-do-not - Investigating whether large language models integrate world knowledge with syntactic structure in a human-like way during ambiguity resolution.
- ANX: Protocol-First Design for AI Agent Interaction with a Supporting 3EX Decoupled Architecture (viability: 7): https://sciencetostartup.com/paper/anx-protocol-first-design-for-ai-agent-interaction-with-a-supporting-3ex-decoupled-architecture - ANX offers an innovative protocol-first architecture that significantly optimizes AI agent interactions, reducing token usage and enhancing security.
- LiveFact: A Dynamic, Time-Aware Benchmark for LLM-Driven Fake News Detection (viability: 7): https://sciencetostartup.com/paper/livefact-a-dynamic-time-aware-benchmark-for-llm-driven-fake-news-detection - LiveFact is a dynamic, time-aware benchmark for evaluating LLMs in fake news detection under temporal uncertainty and evolving information.
- AnyUser: Translating Sketched User Intent into Domestic Robots (viability: 7): https://sciencetostartup.com/paper/anyuser-translating-sketched-user-intent-into-domestic-robots - "AnyUser" allows anyone to instruct domestic robots using intuitive sketching and language inputs directly on images.
- Selecting Decision-Relevant Concepts in Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/selecting-decision-relevant-concepts-in-reinforcement-learning - Automates concept selection in reinforcement learning for improved policy performance and interpretability in complex environments.
- Unpacking .zip: A First Look at Domain and File Name Confusion (viability: 2): https://sciencetostartup.com/paper/unpacking-zip-a-first-look-at-domain-and-file-name-confusion - This paper analyzes potential security risks arising from confusion between DNS names and filenames, providing an empirical study and future research directions.
- SkillX: Automatically Constructing Skill Knowledge Bases for Agents (viability: 8): https://sciencetostartup.com/paper/skillx-automatically-constructing-skill-knowledge-bases-for-agents - SkillX automates the creation of plug-and-play skill knowledge bases for enhancing LLM agents across diverse applications.
- Partially deterministic sampling for compressed sensing with denoising guarantees (viability: 3): https://sciencetostartup.com/paper/partially-deterministic-sampling-for-compressed-sensing-with-denoising-guarantees - A theoretical framework for optimized sampling in compressed sensing that combines random and deterministic selection for improved image reconstruction.
- Forgetting to Witness: Efficient Federated Unlearning and Its Visible Evaluation (viability: 3): https://sciencetostartup.com/paper/forgetting-to-witness-efficient-federated-unlearning-and-its-visible-evaluation - A novel pipeline for federated unlearning and a visualization framework to evaluate data forgetting.
- Multi-Modal Sensor Fusion using Hybrid Attention for Autonomous Driving (viability: 7): https://sciencetostartup.com/paper/multi-modal-sensor-fusion-using-hybrid-attention-for-autonomous-driving - A hybrid attention fusion framework for more accurate 3D object detection in autonomous driving by combining camera and radar sensor data.
- How Far Are We? Systematic Evaluation of LLMs vs. Human Experts in Mathematical Contest in Modeling (viability: 4): https://sciencetostartup.com/paper/how-far-are-we-systematic-evaluation-of-llms-vs-human-experts-in-mathematical-contest-in-modeling - A new framework to systematically evaluate LLMs on complex, real-world problem-solving tasks, revealing critical execution gaps.
- HUKUKBERT: Domain-Specific Language Model for Turkish Law (viability: 7): https://sciencetostartup.com/paper/hukukbert-domain-specific-language-model-for-turkish-law - A state-of-the-art domain-specific language model for Turkish law, enabling advanced LegalTech applications and research.
- AvatarPointillist: AutoRegressive 4D Gaussian Avatarization (viability: 7): https://sciencetostartup.com/paper/avatarpointillist-autoregressive-4d-gaussian-avatarization - Generate photorealistic, animatable 4D Gaussian avatars from a single portrait image using an autoregressive Transformer.
- A Quantum Search Approach to Magic Square Constraint Problems with Classical Benchmarking (viability: 3): https://sciencetostartup.com/paper/a-quantum-search-approach-to-magic-square-constraint-problems-with-classical-benchmarking - Applies quantum search to magic square problems, demonstrating theoretical advantages but facing scalability challenges.
- GPU Acceleration of TFHE-Based High-Precision Nonlinear Layers for Encrypted LLM Inference (viability: 4): https://sciencetostartup.com/paper/gpu-acceleration-of-tfhe-based-high-precision-nonlinear-layers-for-encrypted-llm-inference - Accelerating encrypted LLM inference with a GPU-optimized framework for high-precision nonlinear layers.
- CLEAR: Unlocking Generative Potential for Degraded Image Understanding in Unified Multimodal Models (viability: 7): https://sciencetostartup.com/paper/clear-unlocking-generative-potential-for-degraded-image-understanding-in-unified-multimodal-models - A framework that enhances multimodal AI's understanding of degraded images by leveraging its generative capabilities, improving robustness without sacrificing performance on clean images.
- MinerU2.5-Pro: Pushing the Limits of Data-Centric Document Parsing at Scale (viability: 4): https://sciencetostartup.com/paper/mineru2-5-pro-pushing-the-limits-of-data-centric-document-parsing-at-scale - This research pushes document parsing performance by focusing on data engineering and training strategies, achieving state-of-the-art results without architectural changes.
- Cog-DRIFT: Exploration on Adaptively Reformulated Instances Enables Learning from Hard Reasoning Problems (viability: 8): https://sciencetostartup.com/paper/cog-drift-exploration-on-adaptively-reformulated-instances-enables-learning-from-hard-reasoning-problems - Cog-DRIFT transforms hard reasoning problems into learnable formats for AI, enhancing LLM capabilities.
- Your Agent, Their Asset: A Real-World Safety Analysis of OpenClaw (viability: 7): https://sciencetostartup.com/paper/your-agent-their-asset-a-real-world-safety-analysis-of-openclaw - A real-world safety analysis framework for personal AI agents like OpenClaw, revealing inherent vulnerabilities and proposing defense strategies.
- Undetectable Conversations Between AI Agents via Pseudorandom Noise-Resilient Key Exchange (viability: 1): https://sciencetostartup.com/paper/undetectable-conversations-between-ai-agents-via-pseudorandom-noise-resilient-key-exchange - A theoretical framework for undetectable conversations between AI agents using novel cryptographic primitives for key exchange.
- Darkness Visible: Reading the Exception Handler of a Language Model (viability: 3): https://sciencetostartup.com/paper/darkness-visible-reading-the-exception-handler-of-a-language-model - This research deciphers the internal exception handling mechanism of GPT-2 Small, revealing how specific neurons route linguistic information.
- AI Trust OS -- A Continuous Governance Framework for Autonomous AI Observability and Zero-Trust Compliance in Enterprise Environments (viability: 4): https://sciencetostartup.com/paper/ai-trust-os-a-continuous-governance-framework-for-autonomous-ai-observability-and-zero-trust-compliance-in-enterprise-en - AI Trust OS offers a continuous, telemetry-driven governance framework for autonomous AI observability and zero-trust compliance in enterprises.
- RegGuard: Legitimacy and Fairness Enforcement for Optimistic Rollups (viability: 7): https://sciencetostartup.com/paper/regguard-legitimacy-and-fairness-enforcement-for-optimistic-rollups - RegGuard enhances optimistic rollups with a unified framework for regulatory compliance, reducing settlement failures by over 90% and ensuring fair transaction ordering for regulated financial applications.
- Think in Strokes, Not Pixels: Process-Driven Image Generation via Interleaved Reasoning (viability: 7): https://sciencetostartup.com/paper/think-in-strokes-not-pixels-process-driven-image-generation-via-interleaved-reasoning - A multi-step image generation model that reasons through textual planning, visual drafting, textual reflection, and visual refinement for explicit, interpretable, and supervisable image creation.
- Economic Security of VDF-Based Randomness Beacons: Models, Thresholds, and Design Guidelines (viability: 4): https://sciencetostartup.com/paper/economic-security-of-vdf-based-randomness-beacons-models-thresholds-and-design-guidelines - This research provides a formal framework and design guidelines for economically secure randomness beacons in blockchains, addressing vulnerabilities overlooked by purely cryptographic analyses.
- Hallucination Basins: A Dynamic Framework for Understanding and Controlling LLM Hallucinations (viability: 7): https://sciencetostartup.com/paper/hallucination-basins-a-dynamic-framework-for-understanding-and-controlling-llm-hallucinations - A geometric framework to understand and control LLM hallucinations by analyzing latent space basin structure and applying geometry-aware steering.
- Artificial Intelligence and Cost Reduction in Public Higher Education: A Scoping Review of Emerging Evidence (viability: 2): https://sciencetostartup.com/paper/artificial-intelligence-and-cost-reduction-in-public-higher-education-a-scoping-review-of-emerging-evidence - This paper reviews the current state of AI applications in public higher education for cost reduction, highlighting both benefits and limitations.
- Fine-Tuning Integrity for Modern Neural Networks: Structured Drift Proofs via Norm, Rank, and Sparsity Certificates (viability: 3): https://sciencetostartup.com/paper/fine-tuning-integrity-for-modern-neural-networks-structured-drift-proofs-via-norm-rank-and-sparsity-certificates - A cryptographic primitive to certify that fine-tuned neural networks only deviate from a trusted base within defined parameters, preventing malicious updates.
- Sampling Parallelism for Fast and Efficient Bayesian Learning (viability: 7): https://sciencetostartup.com/paper/sampling-parallelism-for-fast-and-efficient-bayesian-learning - A novel sampling parallelism technique significantly speeds up Bayesian learning by distributing sample evaluations across GPUs, reducing memory and compute costs.
- Lighting Up or Dimming Down? Exploring Dark Patterns of LLMs in Co-Creativity (viability: 3): https://sciencetostartup.com/paper/lighting-up-or-dimming-down-exploring-dark-patterns-of-llms-in-co-creativity - This paper explores subtle negative behaviors of LLMs in creative writing to inform safer AI design.
- Discovering Failure Modes in Vision-Language Models using RL (viability: 7): https://sciencetostartup.com/paper/discovering-failure-modes-in-vision-language-models-using-rl - An RL-based framework automatically discovers failure modes in Vision-Language Models by adaptively generating challenging queries, improving model robustness.
- Metaphors We Compute By: A Computational Audit of Cultural Translation vs. Thinking in LLMs (viability: 2): https://sciencetostartup.com/paper/metaphors-we-compute-by-a-computational-audit-of-cultural-translation-vs-thinking-in-llms - This paper audits LLMs for culture-aware reasoning in metaphor generation, finding they exhibit stereotyped usage and Western defaultism.
- Neuromorphic Computing for Low-Power Artificial Intelligence (viability: 2): https://sciencetostartup.com/paper/neuromorphic-computing-for-low-power-artificial-intelligence - This paper surveys neuromorphic computing approaches for low-power AI by co-designing materials, circuits, and algorithms to overcome classical computing limitations.
- A Muon-Accelerated Algorithm for Low Separation Rank Tensor Generalized Linear Models (viability: 3): https://sciencetostartup.com/paper/a-muon-accelerated-algorithm-for-low-separation-rank-tensor-generalized-linear-models - A novel algorithm accelerates the estimation of tensor-generalized linear models for multidimensional signal and imaging data.
- Individual and Combined Effects of English as a Second Language and Typos on LLM Performance (viability: 4): https://sciencetostartup.com/paper/individual-and-combined-effects-of-english-as-a-second-language-and-typos-on-llm-performance - This work evaluates the combined impact of English as a Second Language (ESL) and typos on LLM performance, revealing performance drops larger than individual effects.
- Don't Waste Bits! Adaptive KV-Cache Quantization for Lightweight On-Device LLMs (viability: 7): https://sciencetostartup.com/paper/don-t-waste-bits-adaptive-kv-cache-quantization-for-lightweight-on-device-llms - Adaptive KV-cache quantization for LLMs that reduces memory and latency on edge devices without sacrificing accuracy.
- AI Assistance Reduces Persistence and Hurts Independent Performance (viability: 3): https://sciencetostartup.com/paper/ai-assistance-reduces-persistence-and-hurts-independent-performance - This study provides causal evidence that AI assistance reduces user persistence and impairs independent performance, even after brief interactions.
- What Makes Good Multilingual Reasoning? Disentangling Reasoning Traces with Measurable Features (viability: 7): https://sciencetostartup.com/paper/what-makes-good-multilingual-reasoning-disentangling-reasoning-traces-with-measurable-features - This work defines measurable features for multilingual reasoning and uses them to challenge English-centric assumptions, suggesting adaptive objectives for stronger cross-lingual performance.
- The Infinite-Dimensional Nature of Spectroscopy and Why Models Succeed, Fail, and Mislead (viability: 2): https://sciencetostartup.com/paper/the-infinite-dimensional-nature-of-spectroscopy-and-why-models-succeed-fail-and-mislead - A theoretical framework explains why machine learning models in spectroscopy can achieve high accuracy even when chemical distinctions are absent, offering practical recommendations for model interpretation.
- Hardware-Level Governance of AI Compute: A Feasibility Taxonomy for Regulatory Compliance and Treaty Verification (viability: 2): https://sciencetostartup.com/paper/hardware-level-governance-of-ai-compute-a-feasibility-taxonomy-for-regulatory-compliance-and-treaty-verification - This paper proposes a taxonomy of hardware-level mechanisms for governing AI compute, assessing their technical feasibility for regulatory compliance and treaty verification.
- BiST: A Gold Standard Bangla-English Bilingual Corpus for Sentence Structure and Tense Classification with Inter-Annotator Agreement (viability: 5): https://sciencetostartup.com/paper/bist-a-gold-standard-bangla-english-bilingual-corpus-for-sentence-structure-and-tense-classification-with-inter-annotato - BiST is a high-quality, rigorously annotated Bangla-English bilingual corpus for sentence structure and tense classification, enabling advancements in multilingual NLP for low-resource settings.
- OpenWorldLib: A Unified Codebase and Definition of Advanced World Models (viability: 4): https://sciencetostartup.com/paper/openworldlib-a-unified-codebase-and-definition-of-advanced-world-models - A unified framework for advanced world models, standardizing their definition and enabling collaborative inference.
- Bridging Safety and Security in Complex Systems: A Model-Based Approach with SAFT-GT Toolchain (viability: 7): https://sciencetostartup.com/paper/bridging-safety-and-security-in-complex-systems-a-model-based-approach-with-saft-gt-toolchain - A toolchain for integrating safety and security analysis into self-adaptive systems, with open-source code and expert validation.
- IDIOLEX: Unified and Continuous Representations for Idiolectal and Stylistic Variation (viability: 5): https://sciencetostartup.com/paper/idiolex-unified-and-continuous-representations-for-idiolectal-and-stylistic-variation - A framework to learn sentence representations that capture style and dialect, decoupled from semantic content, for applications like diverse LLM development.
- MUXQ: Mixed-to-Uniform Precision MatriX Quantization via Low-Rank Outlier Decomposition (viability: 7): https://sciencetostartup.com/paper/muxq-mixed-to-uniform-precision-matrix-quantization-via-low-rank-outlier-decomposition - MUXQ is a novel quantization technique that effectively handles activation outliers in LLMs, enabling efficient INT8 inference on edge devices with minimal accuracy loss.
- Explainable Machine Learning for Sepsis Outcome Prediction Using a Novel Romanian Electronic Health Record Dataset (viability: 5): https://sciencetostartup.com/paper/explainable-machine-learning-for-sepsis-outcome-prediction-using-a-novel-romanian-electronic-health-record-dataset - Develops explainable AI models for sepsis outcome prediction using a novel Romanian EHR dataset, identifying key clinical predictors and achieving state-of-the-art performance.
- GPIR: Enabling Practical Private Information Retrieval with GPUs (viability: 7): https://sciencetostartup.com/paper/gpir-enabling-practical-private-information-retrieval-with-gpus - GPIR accelerates private information retrieval on GPUs by re-architecting kernel design and data layout, achieving significant throughput gains over existing solutions.
- 3D Gaussian Splatting for Annular Dark Field Scanning Transmission Electron Microscopy Tomography Reconstruction (viability: 7): https://sciencetostartup.com/paper/3d-gaussian-splatting-for-annular-dark-field-scanning-transmission-electron-microscopy-tomography-reconstruction - A novel 3D reconstruction method for nanoscale materials using sparse-view electron microscopy data, improving fidelity and sample preservation.
- Is a Picture Worth a Thousand Words? Adaptive Multimodal Fact-Checking with Visual Evidence Necessity (viability: 7): https://sciencetostartup.com/paper/is-a-picture-worth-a-thousand-words-adaptive-multimodal-fact-checking-with-visual-evidence-necessity - AMuFC is a multimodal fact-checking framework that adaptively uses visual evidence, improving accuracy by determining when visual information is truly necessary for claim verification.
- Pickalo: Leveraging 6D Pose Estimation for Low-Cost Industrial Bin Picking (viability: 7): https://sciencetostartup.com/paper/pickalo-leveraging-6d-pose-estimation-for-low-cost-industrial-bin-picking - A low-cost, modular bin-picking system using 6D pose estimation and RGB-D cameras for industrial automation.
- On the "Causality" Step in Policy Gradient Derivations: A Pedagogical Reconciliation of Full Return and Reward-to-Go (viability: 1): https://sciencetostartup.com/paper/on-the-causality-step-in-policy-gradient-derivations-a-pedagogical-reconciliation-of-full-return-and-reward-to-go - A theoretical paper clarifying the causality step in policy gradient derivations for reinforcement learning.
- Packing Entries to Diagonals for Homomorphic Sparse-Matrix Vector Multiplication (viability: 3): https://sciencetostartup.com/paper/packing-entries-to-diagonals-for-homomorphic-sparse-matrix-vector-multiplication - Optimizes sparse matrix structures for homomorphic encryption to significantly reduce computational overhead.
- Batch Loss Score for Dynamic Data Pruning (viability: 7): https://sciencetostartup.com/paper/batch-loss-score-for-dynamic-data-pruning - A simple, three-line code injection that enables lossless data pruning for deep learning models, reducing training data by up to 50% across various tasks and architectures.
- An AI Teaching Assistant for Motion Picture Engineering (viability: 7): https://sciencetostartup.com/paper/an-ai-teaching-assistant-for-motion-picture-engineering - Revolutionize teaching in film and motion picture courses with AI-powered teaching assistants.
- ZeD-MAP: Bundle Adjustment Guided Zero-Shot Depth Maps for Real-Time Aerial Imaging (viability: 7): https://sciencetostartup.com/paper/zed-map-bundle-adjustment-guided-zero-shot-depth-maps-for-real-time-aerial-imaging - A real-time framework for accurate 3D mapping from UAV imagery by guiding diffusion models with bundle adjustment.
- ROSClaw: A Hierarchical Semantic-Physical Framework for Heterogeneous Multi-Agent Collaboration (viability: 7): https://sciencetostartup.com/paper/rosclaw-a-hierarchical-semantic-physical-framework-for-heterogeneous-multi-agent-collaboration - ROSClaw: A unified VLM controller framework for heterogeneous robots enabling semantic-physical task execution and sim-to-real transfer.
- Anticipatory Reinforcement Learning: From Generative Path-Laws to Distributional Value Functions (viability: 4): https://sciencetostartup.com/paper/anticipatory-reinforcement-learning-from-generative-path-laws-to-distributional-value-functions - A novel reinforcement learning framework that uses path-dependent geometry to improve foresight and stability in volatile financial environments.
- Springdrift: An Auditable Persistent Runtime for LLM Agents with Case-Based Memory, Normative Safety, and Ambient Self-Perception (viability: 7): https://sciencetostartup.com/paper/springdrift-an-auditable-persistent-runtime-for-llm-agents-with-case-based-memory-normative-safety-and-ambient-self-perc - A persistent runtime for LLM agents with auditable memory, safety gating, and self-perception, enabling long-term continuity and forensic accountability.
- Synthesis4AD: Synthetic Anomalies are All You Need for 3D Anomaly Detection (viability: 7): https://sciencetostartup.com/paper/synthesis4ad-synthetic-anomalies-are-all-you-need-for-3d-anomaly-detection - Generate high-fidelity synthetic 3D anomalies to overcome data scarcity and significantly improve industrial anomaly detection accuracy.
- Grokking as Dimensional Phase Transition in Neural Networks (viability: 1): https://sciencetostartup.com/paper/grokking-as-dimensional-phase-transition-in-neural-networks - Neural network grokking is a dimensional phase transition driven by gradient field geometry, offering new insights into learning dynamics.
- Search, Do not Guess: Teaching Small Language Models to Be Effective Search Agents (viability: 7): https://sciencetostartup.com/paper/search-do-not-guess-teaching-small-language-models-to-be-effective-search-agents - A fine-tuning approach that trains small language models to reliably use search tools, achieving LLM-level performance on complex reasoning tasks.
- From Curiosity to Caution: Mitigating Reward Hacking for Best-of-N with Pessimism (viability: 7): https://sciencetostartup.com/paper/from-curiosity-to-caution-mitigating-reward-hacking-for-best-of-n-with-pessimism - A new method to prevent language models from exploiting reward model flaws when generating multiple responses, improving reliability and performance.
- On Ambiguity: The case of fraction, its meanings and roles (viability: 1): https://sciencetostartup.com/paper/on-ambiguity-the-case-of-fraction-its-meanings-and-roles - This paper theoretically analyzes ambiguity in mathematical discourse, specifically focusing on the term 'fraction', proposing new terminology to clarify its various meanings and arguing it functions as a category rather than a single concept.
- Training-Free Refinement of Flow Matching with Divergence-based Sampling (viability: 7): https://sciencetostartup.com/paper/training-free-refinement-of-flow-matching-with-divergence-based-sampling - A training-free framework that refines flow-based generation by sampling based on the divergence of the velocity field, improving fidelity across tasks.
- WaterSplat-SLAM: Photorealistic Monocular SLAM in Underwater Environment (viability: 7): https://sciencetostartup.com/paper/watersplat-slam-photorealistic-monocular-slam-in-underwater-environment - A photorealistic monocular SLAM system for underwater environments that achieves robust pose estimation and high-fidelity dense mapping.
- Same World, Differently Given: History-Dependent Perceptual Reorganization in Artificial Agents (viability: 2): https://sciencetostartup.com/paper/same-world-differently-given-history-dependent-perceptual-reorganization-in-artificial-agents - A minimal architecture for artificial agents that sustains a history-sensitive perspective on its world through perceptual reorganization.
- Preserving Forgery Artifacts: AI-Generated Video Detection at Native Scale (viability: 8): https://sciencetostartup.com/paper/preserving-forgery-artifacts-ai-generated-video-detection-at-native-scale - A groundbreaking AI tool to detect forgery artifacts in AI-generated videos by preserving native resolution details.
- InCTRLv2: Generalist Residual Models for Few-Shot Anomaly Detection and Segmentation (viability: 7): https://sciencetostartup.com/paper/inctrlv2-generalist-residual-models-for-few-shot-anomaly-detection-and-segmentation - A few-shot generalist anomaly detection and segmentation framework that leverages dual-branch semantic learning to achieve state-of-the-art performance across diverse datasets without retraining.
- Multimodal Backdoor Attack on VLMs for Autonomous Driving via Graffiti and Cross-Lingual Triggers (viability: 3): https://sciencetostartup.com/paper/multimodal-backdoor-attack-on-vlms-for-autonomous-driving-via-graffiti-and-cross-lingual-triggers - This paper introduces a novel backdoor attack method for visual language models in autonomous driving, demonstrating high attack success with low poisoning ratios and resistance to traditional detection methods.
- Biologically Inspired Event-Based Perception and Sample-Efficient Learning for High-Speed Table Tennis Robots (viability: 7): https://sciencetostartup.com/paper/biologically-inspired-event-based-perception-and-sample-efficient-learning-for-high-speed-table-tennis-robots - Develops a biologically inspired, sample-efficient AI system for high-speed robotic control, starting with table tennis robots.
- A Clinical Point Cloud Paradigm for In-Hospital Mortality Prediction from Multi-Level Incomplete Multimodal EHRs (viability: 7): https://sciencetostartup.com/paper/a-clinical-point-cloud-paradigm-for-in-hospital-mortality-prediction-from-multi-level-incomplete-multimodal-ehrs - HealthPoint, a unified clinical point cloud paradigm, predicts in-hospital mortality from incomplete multimodal EHRs by representing events in a 4D space with relational attention.
- Dynamic Free-Rider Detection in Federated Learning via Simulated Attack Patterns (viability: 7): https://sciencetostartup.com/paper/dynamic-free-rider-detection-in-federated-learning-via-simulated-attack-patterns - A novel federated learning security method that dynamically detects free-riders by simulating attack patterns, offering enhanced robustness without requiring proxy datasets.
- Beyond Semantics: Uncovering the Physics of Fakes via Universal Physical Descriptors for Cross-Modal Synthetic Detection (viability: 7): https://sciencetostartup.com/paper/beyond-semantics-uncovering-the-physics-of-fakes-via-universal-physical-descriptors-for-cross-modal-synthetic-detection - A novel approach to detect AI-generated images by extracting universal physical features, achieving near-perfect accuracy across diverse datasets and generative models.
- AI Agents Under EU Law (viability: 1): https://sciencetostartup.com/paper/ai-agents-under-eu-law - A systematic regulatory mapping for AI agent providers, integrating EU law and proposed standards to address compliance challenges.
- Cardinality Estimation for High Dimensional Similarity Queries with Adaptive Bucket Probing (viability: 5): https://sciencetostartup.com/paper/cardinality-estimation-for-high-dimensional-similarity-queries-with-adaptive-bucket-probing - A framework for accurate and efficient cardinality estimation in high-dimensional similarity search using adaptive bucket probing and product quantization.
- LP-GEMM: Integrating Layout Propagation into GEMM Operations (viability: 7): https://sciencetostartup.com/paper/lp-gemm-integrating-layout-propagation-into-gemm-operations - LP-GEMM accelerates machine learning workloads by eliminating redundant data repacking in sequential matrix multiplications, offering significant speedups for inference and scientific computing tasks.
- Benchmarking Multilingual Speech Models on Pashto: Zero-Shot ASR, Script Failure, and Cross-Domain Evaluation (viability: 7): https://sciencetostartup.com/paper/benchmarking-multilingual-speech-models-on-pashto-zero-shot-asr-script-failure-and-cross-domain-evaluation - This research benchmarks multilingual speech models on Pashto, identifying critical failures and proposing a path for improved performance on under-resourced languages.
- Ruling Out to Rule In: Contrastive Hypothesis Retrieval for Medical Question Answering (viability: 7): https://sciencetostartup.com/paper/ruling-out-to-rule-in-contrastive-hypothesis-retrieval-for-medical-question-answering - A contrastive retrieval framework for medical question answering that explicitly penalizes clinically plausible incorrect answers, significantly reducing hard-negative contamination.
- Greedy and Transformer-Based Multi-Port Selection for Slow Fluid Antenna Multiple Access (viability: 4): https://sciencetostartup.com/paper/greedy-and-transformer-based-multi-port-selection-for-slow-fluid-antenna-multiple-access - Novel greedy and Transformer-based methods for efficient multi-port selection in fluid antenna multiple access systems.
- Noisy Nonreciprocal Pairwise Comparisons: Scale Variation, Noise Calibration, and Admissible Ranking Regions (viability: 3): https://sciencetostartup.com/paper/noisy-nonreciprocal-pairwise-comparisons-scale-variation-noise-calibration-and-admissible-ranking-regions - A novel statistical model for analyzing non-reciprocal pairwise comparisons by separating genuine scale variation from random noise to improve decision analysis.
- Firebolt-VL: Efficient Vision-Language Understanding with Cross-Modality Modulation (viability: 7): https://sciencetostartup.com/paper/firebolt-vl-efficient-vision-language-understanding-with-cross-modality-modulation - An efficient vision-language model that uses a novel decoder and correlation module to achieve accurate, fine-grained understanding with linear-time inference, making it suitable for resource-constrained applications.
- PR-IQA: Partial-Reference Image Quality Assessment for Diffusion-Based Novel View Synthesis (viability: 7): https://sciencetostartup.com/paper/pr-iqa-partial-reference-image-quality-assessment-for-diffusion-based-novel-view-synthesis - A novel image quality assessment framework for diffusion-generated views that improves 3D reconstruction accuracy by filtering inconsistencies without ground truth.
- Erasure or Erosion? Evaluating Compositional Degradation in Unlearned Text-To-Image Diffusion Models (viability: 4): https://sciencetostartup.com/paper/erasure-or-erosion-evaluating-compositional-degradation-in-unlearned-text-to-image-diffusion-models - Develops methods to remove unwanted concepts from text-to-image models without degrading their core generative capabilities.
- SAIL: Scene-aware Adaptive Iterative Learning for Long-Tail Trajectory Prediction in Autonomous Vehicles (viability: 7): https://sciencetostartup.com/paper/sail-scene-aware-adaptive-iterative-learning-for-long-tail-trajectory-prediction-in-autonomous-vehicles - A framework for autonomous vehicles that significantly improves prediction accuracy for rare but critical driving scenarios by adaptively learning from long-tail trajectory data.
- Digital Privacy in IoT: Exploring Challenges, Approaches and Open Issues (viability: 3): https://sciencetostartup.com/paper/digital-privacy-in-iot-exploring-challenges-approaches-and-open-issues - This paper reviews digital privacy risks in IoT and proposes a framework to address AI-driven privacy concerns.
- TAPE: A two-stage parameter-efficient adaptation framework for foundation models in OCT-OCTA analysis (viability: 7): https://sciencetostartup.com/paper/tape-a-two-stage-parameter-efficient-adaptation-framework-for-foundation-models-in-oct-octa-analysis - A parameter-efficient framework for adapting foundation models to ophthalmic image analysis, improving diagnosis in resource-constrained clinical settings.
- Generative Modeling under Non-Monotonic MAR Missingness via Approximate Wasserstein Gradient Flows (viability: 7): https://sciencetostartup.com/paper/generative-modeling-under-non-monotonic-mar-missingness-via-approximate-wasserstein-gradient-flows - FLOWGEM is a principled iterative method for generating complete datasets from data with non-monotonic missing values, outperforming existing imputation techniques.
- PassiveQA: A Three-Action Framework for Epistemically Calibrated Question Answering via Supervised Finetuning (viability: 7): https://sciencetostartup.com/paper/passiveqa-a-three-action-framework-for-epistemically-calibrated-question-answering-via-supervised-finetuning - A three-action framework for question answering that enables LLMs to intelligently Ask for clarification or Abstain when faced with incomplete information, reducing hallucinations.
- Visual Prompt Based Reasoning for Offroad Mapping using Multimodal LLMs (viability: 7): https://sciencetostartup.com/paper/visual-prompt-based-reasoning-for-offroad-mapping-using-multimodal-llms - A zero-shot multimodal LLM approach for off-road navigation that reasons about drivable areas using environment segmentation, eliminating the need for task-specific models.
- Temporal Inversion for Learning Interval Change in Chest X-Rays (viability: 7): https://sciencetostartup.com/paper/temporal-inversion-for-learning-interval-change-in-chest-x-rays - A framework for improving temporal change detection in chest X-rays by leveraging temporal inversion as a supervisory signal.
- Paper Espresso: From Paper Overload to Research Insight (viability: 8): https://sciencetostartup.com/paper/paper-espresso-from-paper-overload-to-research-insight - Paper Espresso automates research paper discovery and insight generation for real-time academic trend analysis.
- Mapping the Exploitation Surface: A 10,000-Trial Taxonomy of What Makes LLM Agents Exploit Vulnerabilities (viability: 7): https://sciencetostartup.com/paper/mapping-the-exploitation-surface-a-10-000-trial-taxonomy-of-what-makes-llm-agents-exploit-vulnerabilities - A systematic taxonomy of LLM agent vulnerabilities, identifying goal reframing as the primary trigger for exploitation, with implications for security auditing.
- Relational Epipolar Graphs for Robust Relative Camera Pose Estimation (viability: 7): https://sciencetostartup.com/paper/relational-epipolar-graphs-for-robust-relative-camera-pose-estimation - A novel graph-based approach for robust relative camera pose estimation in VSLAM, outperforming existing methods in noisy conditions.
- StableTTA: Training-Free Test-Time Adaptation that Improves Model Accuracy on ImageNet1K to 96% (viability: 4): https://sciencetostartup.com/paper/stabletta-training-free-test-time-adaptation-that-improves-model-accuracy-on-imagenet1k-to-96 - StableTTA is a training-free method to improve aggregation stability and efficiency for ensemble methods, significantly boosting accuracy on ImageNet-1K.
- Formal Constraints on Dependency Syntax (viability: 2): https://sciencetostartup.com/paper/formal-constraints-on-dependency-syntax - This paper explores theoretical constraints on dependency syntax to improve linguistic accuracy and parsing efficiency.
- Beyond Imbalance Ratio: Data Characteristics as Critical Moderators of Oversampling Method Selection (viability: 5): https://sciencetostartup.com/paper/beyond-imbalance-ratio-data-characteristics-as-critical-moderators-of-oversampling-method-selection - A framework for intelligently selecting data oversampling methods based on data characteristics beyond just imbalance ratio.
- FlashSAC: Fast and Stable Off-Policy Reinforcement Learning for High-Dimensional Robot Control (viability: 7): https://sciencetostartup.com/paper/flashsac-fast-and-stable-off-policy-reinforcement-learning-for-high-dimensional-robot-control - A faster and more stable off-policy reinforcement learning algorithm for high-dimensional robot control, significantly reducing sim-to-real training time.
- Learning from Equivalence Queries, Revisited (viability: 2): https://sciencetostartup.com/paper/learning-from-equivalence-queries-revisited - This paper revisits a classical machine learning model to develop theoretical bounds for learning from equivalence queries under less adversarial conditions, with potential applications in model updates.
- Multilingual Prompt Localization for Agent-as-a-Judge: Language and Backbone Sensitivity in Requirement-Level Evaluation (viability: 7): https://sciencetostartup.com/paper/multilingual-prompt-localization-for-agent-as-a-judge-language-and-backbone-sensitivity-in-requirement-level-evaluation - This research localizes agent-as-a-judge prompts to diverse languages, revealing significant backbone performance shifts and demonstrating the critical need for language-aware evaluation in agentic benchmarks.
- SLSREC: Self-Supervised Contrastive Learning for Adaptive Fusion of Long- and Short-Term User Interests (viability: 7): https://sciencetostartup.com/paper/slsrec-self-supervised-contrastive-learning-for-adaptive-fusion-of-long-and-short-term-user-interests - A self-supervised contrastive learning model that adaptively fuses long- and short-term user interests for improved recommendation accuracy.
- Receding-Horizon Control via Drifting Models (viability: 4): https://sciencetostartup.com/paper/receding-horizon-control-via-drifting-models - Drifting MPC is an offline trajectory optimization framework that combines generative models with receding-horizon planning to learn near-optimal trajectories from data, outperforming diffusion-based baselines in efficiency.
- ENCRUST: Encapsulated Substitution and Agentic Refinement on a Live Scaffold for Safe C-to-Rust Translation (viability: 7): https://sciencetostartup.com/paper/encrust-encapsulated-substitution-and-agentic-refinement-on-a-live-scaffold-for-safe-c-to-rust-translation - Build a safer automated C-to-Rust translation tool with encapsulated substitution and agentic refinement.
- G-EDF-Loc: 3D Continuous Gaussian Distance Field for Robust Gradient-Based 6DoF Localization (viability: 7): https://sciencetostartup.com/paper/g-edf-loc-3d-continuous-gaussian-distance-field-for-robust-gradient-based-6dof-localization - A CPU-based 6DoF localization framework using a novel continuous 3D distance field for robust performance even with degraded sensor input.
- HDP: A Lightweight Cryptographic Protocol for Human Delegation Provenance in Agentic AI Systems (viability: 7): https://sciencetostartup.com/paper/hdp-a-lightweight-cryptographic-protocol-for-human-delegation-provenance-in-agentic-ai-systems - A cryptographic protocol and SDK for verifying human authorization in multi-agent AI systems, addressing a critical accountability gap.
- Reproducibility study on how to find Spurious Correlations, Shortcut Learning, Clever Hans or Group-Distributional non-robustness and how to fix them (viability: 5): https://sciencetostartup.com/paper/reproducibility-study-on-how-to-find-spurious-correlations-shortcut-learning-clever-hans-or-group-distributional-non-rob - This study unifies and compares methods for detecting and fixing spurious correlations and shortcut learning in DNNs, finding XAI-based methods like CFKD most effective but highlighting dependencies on group labels and data scarcity.
- GAIN: Multiplicative Modulation for Domain Adaptation (viability: 4): https://sciencetostartup.com/paper/gain-multiplicative-modulation-for-domain-adaptation - A novel multiplicative modulation technique for LLM domain adaptation that improves performance on previously trained domains, unlike standard fine-tuning methods.
- CommonMorph: Participatory Morphological Documentation Platform (viability: 6): https://sciencetostartup.com/paper/commonmorph-participatory-morphological-documentation-platform - A participatory platform to accelerate morphological data collection for low-resource languages, making linguistic documentation accessible and interoperable.
- SuperLocalMemory V3.3: The Living Brain -- Biologically-Inspired Forgetting, Cognitive Quantization, and Multi-Channel Retrieval for Zero-LLM Agent Memory Systems (viability: 7): https://sciencetostartup.com/paper/superlocalmemory-v3-3-the-living-brain-biologically-inspired-forgetting-cognitive-quantization-and-multi-channel-retriev - A biologically-inspired, local-first AI agent memory system with advanced forgetting, quantization, and multi-channel retrieval, offering significant improvements over existing solutions.
- MPTF-Net: Multi-view Pyramid Transformer Fusion Network for LiDAR-based Place Recognition (viability: 7): https://sciencetostartup.com/paper/mptf-net-multi-view-pyramid-transformer-fusion-network-for-lidar-based-place-recognition - A novel multi-view Transformer fusion network for real-time, state-of-the-art LiDAR-based place recognition in autonomous systems.
- MedROI: Codec-Agnostic Region of Interest-Centric Compression for Medical Images (viability: 7): https://sciencetostartup.com/paper/medroi-codec-agnostic-region-of-interest-centric-compression-for-medical-images - A codec-agnostic framework that significantly improves medical image compression ratios and speeds by focusing compression on relevant regions of interest, with code available.
- DHFP-PE: Dual-Precision Hybrid Floating Point Processing Element for AI Acceleration (viability: 3): https://sciencetostartup.com/paper/dhfp-pe-dual-precision-hybrid-floating-point-processing-element-for-ai-acceleration - A novel dual-precision floating-point processing engine for energy-efficient AI acceleration.
- Memory Intelligence Agent (viability: 6): https://sciencetostartup.com/paper/memory-intelligence-agent - A Memory Intelligence Agent (MIA) framework that enhances deep research agents with efficient, evolving memory systems for improved reasoning and autonomous evolution.
- Veo-Act: How Far Can Frontier Video Models Advance Generalizable Robot Manipulation? (viability: 7): https://sciencetostartup.com/paper/veo-act-how-far-can-frontier-video-models-advance-generalizable-robot-manipulation - Leveraging frontier video generation models to create generalizable robot manipulation policies.
- Saliency-R1: Enforcing Interpretable and Faithful Vision-language Reasoning via Saliency-map Alignment Reward (viability: 4): https://sciencetostartup.com/paper/saliency-r1-enforcing-interpretable-and-faithful-vision-language-reasoning-via-saliency-map-alignment-reward - A framework to improve the trustworthiness of vision-language models by aligning their reasoning with visual evidence using saliency maps.
- One Model for All: Multi-Objective Controllable Language Models (viability: 7): https://sciencetostartup.com/paper/one-model-for-all-multi-objective-controllable-language-models - A novel Multi-Objective Control (MOC) framework that trains a single LLM to generate personalized outputs across diverse user preferences, enabling scalable and customizable LLM applications.
- The Indra Representation Hypothesis for Multimodal Alignment (viability: 7): https://sciencetostartup.com/paper/the-indra-representation-hypothesis-for-multimodal-alignment - A theoretically grounded framework for training-free alignment of unimodal foundation models across vision, language, and audio.
- SLaB: Sparse-Lowrank-Binary Decomposition for Efficient Large Language Models (viability: 7): https://sciencetostartup.com/paper/slab-sparse-lowrank-binary-decomposition-for-efficient-large-language-models - SLaB offers a novel framework for efficient LLM deployment by decomposing weights into sparse, low-rank, and binary components, achieving state-of-the-art compression without retraining.
- Isokinetic Flow Matching for Pathwise Straightening of Generative Flows (viability: 4): https://sciencetostartup.com/paper/isokinetic-flow-matching-for-pathwise-straightening-of-generative-flows - A novel regularization technique for generative models that significantly improves sampling speed and quality by reducing path curvature.
- RAVEN: Radar Adaptive Vision Encoders for Efficient Chirp-wise Object Detection and Segmentation (viability: 7): https://sciencetostartup.com/paper/raven-radar-adaptive-vision-encoders-for-efficient-chirp-wise-object-detection-and-segmentation - RAVEN is a computationally efficient deep learning architecture for FMCW radar perception, enabling chirp-wise processing and early-exit mechanisms for faster object detection and segmentation.
- A Patch-based Cross-view Regularized Framework for Backdoor Defense in Multimodal Large Language Models (viability: 4): https://sciencetostartup.com/paper/a-patch-based-cross-view-regularized-framework-for-backdoor-defense-in-multimodal-large-language-models - A defense framework to prevent backdoor attacks in multimodal large language models by regularizing feature representations and output distributions.
- Training-Free Image Editing with Visual Context Integration and Concept Alignment (viability: 7): https://sciencetostartup.com/paper/training-free-image-editing-with-visual-context-integration-and-concept-alignment - A training-free and inversion-free image editing method that injects visual context for precise user requirements, outperforming state-of-the-art training-based models.
- ECG Biometrics with ArcFace-Inception: External Validation on MIMIC and HEEDB (viability: 3): https://sciencetostartup.com/paper/ecg-biometrics-with-arcface-inception-external-validation-on-mimic-and-heedb - This paper evaluates the performance of ECG biometrics under large-scale, external validation and temporal stress, highlighting the impact of domain heterogeneity and gallery size.
- Scalable and Explainable Learner-Video Interaction Prediction using Multimodal Large Language Models (viability: 7): https://sciencetostartup.com/paper/scalable-and-explainable-learner-video-interaction-prediction-using-multimodal-large-language-models - A predictive tool for optimizing educational videos by analyzing learner interactions using multimodal LLM embeddings.
- MVis-Fold: A Three-Dimensional Microvascular Structure Inference Model for Super-Resolution Ultrasound (viability: 4): https://sciencetostartup.com/paper/mvis-fold-a-three-dimensional-microvascular-structure-inference-model-for-super-resolution-ultrasound - A novel 3D microvascular reconstruction model for super-resolution ultrasound to enable quantitative analysis of microvasculature in solid tumors.
- Discrete Prototypical Memories for Federated Time Series Foundation Models (viability: 7): https://sciencetostartup.com/paper/discrete-prototypical-memories-for-federated-time-series-foundation-models - A federated learning framework using discrete prototypical memories to improve time series foundation models while preserving data privacy.
- MAVEN: A Mesh-Aware Volumetric Encoding Network for Simulating 3D Flexible Deformation (viability: 7): https://sciencetostartup.com/paper/maven-a-mesh-aware-volumetric-encoding-network-for-simulating-3d-flexible-deformation - MAVEN delivers enhanced 3D deformation simulation through mesh-aware volumetric encoding for precise engineering applications.
- Beyond Standard Benchmarks: A Systematic Audit of Vision-Language Model's Robustness to Natural Semantic Variation Across Diverse Tasks (viability: 4): https://sciencetostartup.com/paper/beyond-standard-benchmarks-a-systematic-audit-of-vision-language-model-s-robustness-to-natural-semantic-variation-across - This paper systematically audits the robustness of vision-language models to natural semantic variations across diverse tasks, revealing vulnerabilities and failure modes.
- MC-GenRef: Annotation-free mammography microcalcification segmentation with generative posterior refinement (viability: 5): https://sciencetostartup.com/paper/mc-genref-annotation-free-mammography-microcalcification-segmentation-with-generative-posterior-refinement - MC-GenRef enables annotation-free mammography microcalcification segmentation using synthetic data and test-time generative refinement to improve accuracy and robustness.
- Same Geometry, Opposite Noise: Transformer Magnitude Representations Lack Scalar Variability (viability: 3): https://sciencetostartup.com/paper/same-geometry-opposite-noise-transformer-magnitude-representations-lack-scalar-variability - This research reveals that transformer language models do not replicate biological systems' scalar variability in magnitude representation, suggesting distributional learning alone is insufficient.
- What Makes a Sale? Rethinking End-to-End Seller--Buyer Retail Dynamics with LLM Agents (viability: 9): https://sciencetostartup.com/paper/what-makes-a-sale-rethinking-end-to-end-seller-buyer-retail-dynamics-with-llm-agents - RetailSim offers a comprehensive retail simulation framework using LLM agents to evaluate and optimize end-to-end retail strategies.
- Group-DINOmics: Incorporating People Dynamics into DINO for Self-supervised Group Activity Feature Learning (viability: 7): https://sciencetostartup.com/paper/group-dinomics-incorporating-people-dynamics-into-dino-for-self-supervised-group-activity-feature-learning - A self-supervised learning method for understanding group activities by incorporating people's dynamics and scene context, outperforming existing approaches.
- The Topology of Multimodal Fusion: Why Current Architectures Fail at Creative Cognition (viability: 3): https://sciencetostartup.com/paper/the-topology-of-multimodal-fusion-why-current-architectures-fail-at-creative-cognition - A theoretical framework for multimodal AI architectures that addresses limitations in creative cognition by introducing a novel topological approach.
- DP-OPD: Differentially Private On-Policy Distillation for Language Models (viability: 7): https://sciencetostartup.com/paper/dp-opd-differentially-private-on-policy-distillation-for-language-models - A synthesis-free framework for differentially private language model compression that collapses private compression into a single DP student-training loop.
- Empirical Characterization of Rationale Stability Under Controlled Perturbations for Explainable Pattern Recognition (viability: 7): https://sciencetostartup.com/paper/empirical-characterization-of-rationale-stability-under-controlled-perturbations-for-explainable-pattern-recognition - A new metric to ensure AI model explanations are stable and consistent across similar inputs, building more trustworthy AI systems.
- Generative modeling of granular flow on inclined planes using conditional flow matching (viability: 7): https://sciencetostartup.com/paper/generative-modeling-of-granular-flow-on-inclined-planes-using-conditional-flow-matching - A generative model that reconstructs hidden granular flow mechanics from sparse boundary data, enabling non-invasive analysis of complex physical systems.
- Beyond Few-Step Inference: Accelerating Video Diffusion Transformer Model Serving with Inter-Request Caching Reuse (viability: 5): https://sciencetostartup.com/paper/beyond-few-step-inference-accelerating-video-diffusion-transformer-model-serving-with-inter-request-caching-reuse - Accelerate video diffusion model serving by reusing computation across inference requests.
- Conversational Control with Ontologies for Large Language Models: A Lightweight Framework for Constrained Generation (viability: 7): https://sciencetostartup.com/paper/conversational-control-with-ontologies-for-large-language-models-a-lightweight-framework-for-constrained-generation - A lightweight framework for controlling LLM conversational output using ontologies, enhancing predictability and personalization.
- PSY-STEP: Structuring Therapeutic Targets and Action Sequences for Proactive Counseling Dialogue Systems (viability: 7): https://sciencetostartup.com/paper/psy-step-structuring-therapeutic-targets-and-action-sequences-for-proactive-counseling-dialogue-systems - A proactive counseling agent trained on a new dataset that models cognitive behavioral therapy to provide more clinically grounded and personalized dialogues.
- TinyNina: A Resource-Efficient Edge-AI Framework for Sustainable Air Quality Monitoring via Intra-Image Satellite Super-Resolution (viability: 7): https://sciencetostartup.com/paper/tinynina-a-resource-efficient-edge-ai-framework-for-sustainable-air-quality-monitoring-via-intra-image-satellite-super-r - A resource-efficient Edge-AI framework for real-time, high-precision air quality monitoring using satellite imagery.
- Parameter-Efficient Semantic Augmentation for Enhancing Open-Vocabulary Object Detection (viability: 7): https://sciencetostartup.com/paper/parameter-efficient-semantic-augmentation-for-enhancing-open-vocabulary-object-detection - A parameter-efficient framework that enhances open-vocabulary object detection by enriching textual representations with domain-specific semantic prompts, improving performance on specialized datasets.
- DeonticBench: A Benchmark for Reasoning over Rules (viability: 4): https://sciencetostartup.com/paper/deonticbench-a-benchmark-for-reasoning-over-rules - A new benchmark and dataset to evaluate and improve LLM reasoning over complex, real-world rules, with a focus on legal and policy domains.
- Explainable Autonomous Cyber Defense using Adversarial Multi-Agent Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/explainable-autonomous-cyber-defense-using-adversarial-multi-agent-reinforcement-learning - An AI-powered cyber defense system that uses causal modeling and adversarial reinforcement learning to significantly reduce false positives and provide explainable decision-making.
- Training Transformers in Cosine Coefficient Space (viability: 3): https://sciencetostartup.com/paper/training-transformers-in-cosine-coefficient-space - A novel method for training transformers by parameterizing weights in cosine coefficient space, reducing parameter count while maintaining performance.
- Estimating Central, Peripheral, and Temporal Visual Contributions to Human Decision Making in Atari Games (viability: 4): https://sciencetostartup.com/paper/estimating-central-peripheral-and-temporal-visual-contributions-to-human-decision-making-in-atari-games - This research reverse-engineers human decision-making in Atari games by quantifying the impact of peripheral vision, gaze, and past states on action prediction, offering insights into how agents can better mimic human strategies.
- FAVE: Flow-based Average Velocity Establishment for Sequential Recommendation (viability: 7): https://sciencetostartup.com/paper/fave-flow-based-average-velocity-establishment-for-sequential-recommendation - A novel framework for one-step generative recommendation that significantly improves inference efficiency and accuracy by learning a direct trajectory from an informative prior.
- ShieldNet: Network-Level Guardrails against Emerging Supply-Chain Injections in Agentic Systems (viability: 7): https://sciencetostartup.com/paper/shieldnet-network-level-guardrails-against-emerging-supply-chain-injections-in-agentic-systems - ShieldNet provides a network-level security framework to guard against supply-chain threats in agentic systems.
- HandDreamer: Zero-Shot Text to 3D Hand Model Generation using Corrective Hand Shape Guidance (viability: 7): https://sciencetostartup.com/paper/handdreamer-zero-shot-text-to-3d-hand-model-generation-using-corrective-hand-shape-guidance - HandDreamer generates customizable 3D hand models from text prompts, overcoming limitations of current text-to-3D methods for virtual reality applications.
- Is Prompt Selection Necessary for Task-Free Online Continual Learning? (viability: 7): https://sciencetostartup.com/paper/is-prompt-selection-necessary-for-task-free-online-continual-learning - SinglePrompt is a task-free online continual learning framework that eliminates the need for prompt selection by using a single prompt and cosine similarity-based logit design.
- BoxComm: Benchmarking Category-Aware Commentary Generation and Narration Rhythm in Boxing (viability: 7): https://sciencetostartup.com/paper/boxcomm-benchmarking-category-aware-commentary-generation-and-narration-rhythm-in-boxing - A new dataset and evaluation framework for generating category-aware and rhythmically accurate boxing commentary, addressing a gap in sports AI.
- Justified or Just Convincing? Error Verifiability as a Dimension of LLM Quality (viability: 3): https://sciencetostartup.com/paper/justified-or-just-convincing-error-verifiability-as-a-dimension-of-llm-quality - Develops a new metric for evaluating LLM justification quality and proposes methods to improve it, focusing on error verifiability.
- Structured Causal Video Reasoning via Multi-Objective Alignment (viability: 4): https://sciencetostartup.com/paper/structured-causal-video-reasoning-via-multi-objective-alignment - A novel approach to video understanding that structures causal relationships for more reliable and efficient reasoning.
- Eliminating Vendor Lock-In in Quantum Machine Learning via Framework-Agnostic Neural Networks (viability: 7): https://sciencetostartup.com/paper/eliminating-vendor-lock-in-in-quantum-machine-learning-via-framework-agnostic-neural-networks - A framework-agnostic quantum neural network architecture that eliminates vendor lock-in and enables seamless integration across multiple quantum computing platforms and classical ML frameworks.
- Responses Fall Short of Understanding: Revealing the Gap between Internal Representations and Responses in Visual Document Understanding (viability: 6): https://sciencetostartup.com/paper/responses-fall-short-of-understanding-revealing-the-gap-between-internal-representations-and-responses-in-visual-documen - Identifies a gap between LLM internal representations and responses in visual document understanding and proposes fine-tuning intermediate layers to improve accuracy.
- Relative Density Ratio Optimization for Stable and Statistically Consistent Model Alignment (viability: 6): https://sciencetostartup.com/paper/relative-density-ratio-optimization-for-stable-and-statistically-consistent-model-alignment - Introduces a stable and statistically consistent method for aligning language models with human preferences by optimizing relative density ratios.
- FORMULA: FORmation MPC with neUral barrier Learning for safety Assurance (viability: 4): https://sciencetostartup.com/paper/formula-formation-mpc-with-neural-barrier-learning-for-safety-assurance - A learning-enhanced predictive control framework for safe, scalable, and formation-preserving navigation in multi-robot systems.
- 3D-Fixer: Coarse-to-Fine In-place Completion for 3D Scenes from a Single Image (viability: 8): https://sciencetostartup.com/paper/3d-fixer-coarse-to-fine-in-place-completion-for-3d-scenes-from-a-single-image - 3D-Fixer transforms single-view images into 3D scenes with state-of-the-art accuracy using an innovative in-place completion paradigm.
- MolDA: Molecular Understanding and Generation via Large Language Diffusion Model (viability: 4): https://sciencetostartup.com/paper/molda-molecular-understanding-and-generation-via-large-language-diffusion-model - Proposes MolDA, a diffusion model for molecular understanding and generation that replaces autoregressive backbones to improve chemical validity and global structural coherence.
- UENR-600K: A Large-Scale Physically Grounded Dataset for Nighttime Video Deraining (viability: 7): https://sciencetostartup.com/paper/uenr-600k-a-large-scale-physically-grounded-dataset-for-nighttime-video-deraining - A large-scale, physically grounded dataset for nighttime video deraining, enabling significantly better generalization to real-world scenarios.
- ReinVBC: A Model-based Reinforcement Learning Approach to Vehicle Braking Controller (viability: 7): https://sciencetostartup.com/paper/reinvbc-a-model-based-reinforcement-learning-approach-to-vehicle-braking-controller - An AI-driven system that automates vehicle braking control calibration, reducing labor and improving safety.
- GUIDE: Interpretable GUI Agent Evaluation via Hierarchical Diagnosis (viability: 7): https://sciencetostartup.com/paper/guide-interpretable-gui-agent-evaluation-via-hierarchical-diagnosis - A framework for evaluating and diagnosing GUI agents by decomposing trajectories into subtasks, improving accuracy and providing actionable insights for development.
- BiTDiff: Fine-Grained 3D Conducting Motion Generation via BiMamba-Transformer Diffusion (viability: 7): https://sciencetostartup.com/paper/bitdiff-fine-grained-3d-conducting-motion-generation-via-bimamba-transformer-diffusion - Generate high-quality, fine-grained 3D conducting motions from music with a novel BiMamba-Transformer diffusion model and a new large-scale dataset.
- Finite-Time Analysis of Q-Value Iteration for General-Sum Stackelberg Games (viability: 3): https://sciencetostartup.com/paper/finite-time-analysis-of-q-value-iteration-for-general-sum-stackelberg-games - This paper provides theoretical convergence guarantees for a reinforcement learning algorithm in multi-agent games, offering a control-theoretic perspective.
- Gradual Cognitive Externalization: A Framework for Understanding How Ambient Intelligence Externalizes Human Cognition (viability: 1): https://sciencetostartup.com/paper/gradual-cognitive-externalization-a-framework-for-understanding-how-ambient-intelligence-externalizes-human-cognition - A theoretical framework proposing that human cognition is gradually migrating into digital substrates through ambient intelligence co-adaptation.
- Automatically Generating Hard Math Problems from Hypothesis-Driven Error Analysis (viability: 7): https://sciencetostartup.com/paper/automatically-generating-hard-math-problems-from-hypothesis-driven-error-analysis - An AI pipeline that generates challenging math problems by identifying LLM weaknesses, improving benchmark accuracy and adaptability.
- How Alignment Routes: Localizing, Scaling, and Controlling Policy Circuits in Language Models (viability: 4): https://sciencetostartup.com/paper/how-alignment-routes-localizing-scaling-and-controlling-policy-circuits-in-language-models - Identifies and controls a sparse routing mechanism in language models that governs policy circuits for safety and censorship.
- Compressible Softmax-Attended Language under Incompressible Attention (viability: 2): https://sciencetostartup.com/paper/compressible-softmax-attended-language-under-incompressible-attention - This paper analyzes the internal workings of transformer language models to understand how attention mechanisms interact with data compressibility.
- Optimizing Service Operations via LLM-Powered Multi-Agent Simulation (viability: 7): https://sciencetostartup.com/paper/optimizing-service-operations-via-llm-powered-multi-agent-simulation - An LLM-powered multi-agent simulation framework optimizes service operations by modeling human behavior and using on-trajectory learning.
- CPT: Controllable and Editable Design Variations with Language Models (viability: 5): https://sciencetostartup.com/paper/cpt-controllable-and-editable-design-variations-with-language-models - A language model system generates editable design variations by predicting visual style attributes in design templates, producing structured and stylistically coherent outputs.
- Reinforce to Learn, Elect to Reason: A Dual Paradigm for Video Reasoning (viability: 7): https://sciencetostartup.com/paper/reinforce-to-learn-elect-to-reason-a-dual-paradigm-for-video-reasoning - A dual paradigm for video reasoning that improves reliability and interpretability by explicitly generating and electing based on evidence.
- Towards Considerate Human-Robot Coexistence: A Dual-Space Framework of Robot Design and Human Perception in Healthcare (viability: 1): https://sciencetostartup.com/paper/towards-considerate-human-robot-coexistence-a-dual-space-framework-of-robot-design-and-human-perception-in-healthcare - This paper proposes a dual-space framework for understanding the co-evolution of robot design and human perception in healthcare settings.
- Decocted Experience Improves Test-Time Inference in LLM Agents (viability: 4): https://sciencetostartup.com/paper/decocted-experience-improves-test-time-inference-in-llm-agents - This paper explores how to improve LLM agent performance at inference time by constructing better contexts from accumulated experience.
- Graph-to-Frame RAG: Visual-Space Knowledge Fusion for Training-Free and Auditable Video Reasoning (viability: 7): https://sciencetostartup.com/paper/graph-to-frame-rag-visual-space-knowledge-fusion-for-training-free-and-auditable-video-reasoning - A training-free system that fuses external knowledge into video reasoning by rendering it as visual frames, improving accuracy and interpretability.
- DAO to (Anonymous) DAO Transactions (viability: 4): https://sciencetostartup.com/paper/dao-to-anonymous-dao-transactions - A framework for secure and optionally anonymous transactions between decentralized autonomous organizations.
- Context is All You Need (viability: 3): https://sciencetostartup.com/paper/context-is-all-you-need - A method for contextual adaptation of neural networks and LLMs to improve robustness under domain shift without retraining.
- Integer-Only Operations on Extreme Learning Machine Test Time Classification (viability: 7): https://sciencetostartup.com/paper/integer-only-operations-on-extreme-learning-machine-test-time-classification - A RAG system that uses grounded knowledge graphs extracted from source documents for efficient and factually accurate long-document question answering.
- GROUNDEDKG-RAG: Grounded Knowledge Graph Index for Long-document Question Answering (viability: 7): https://sciencetostartup.com/paper/groundedkg-rag-grounded-knowledge-graph-index-for-long-document-question-answering - GroundedKG-RAG: A RAG system that explicitly extracts and grounds knowledge graphs from source documents for efficient and factually accurate long-document question answering.
- Spatially-Weighted CLIP for Street-View Geo-localization (viability: 7): https://sciencetostartup.com/paper/spatially-weighted-clip-for-street-view-geo-localization - A novel framework for street-view geo-localization that leverages spatial relationships to significantly improve accuracy and reduce errors.
- REAM: Merging Improves Pruning of Experts in LLMs (viability: 7): https://sciencetostartup.com/paper/ream-merging-improves-pruning-of-experts-in-llms - REAM: A method for merging experts in Mixture-of-Experts LLMs to reduce memory requirements while preserving performance.
- Talk2AI: A Longitudinal Dataset of Human--AI Persuasive Conversations (viability: 5): https://sciencetostartup.com/paper/talk2ai-a-longitudinal-dataset-of-human-ai-persuasive-conversations - A longitudinal dataset of human-AI persuasive conversations to study opinion change and human-AI interaction.
- ReFinE: Streamlining UI Mockup Iteration with Research Findings (viability: 4): https://sciencetostartup.com/paper/refine-streamlining-ui-mockup-iteration-with-research-findings - A Figma plugin that synthesizes design insights from HCI research papers to guide real-time UI mockup iteration.
- Adversarial Robustness Analysis of Cloud-Assisted Autonomous Driving Systems (viability: 4): https://sciencetostartup.com/paper/adversarial-robustness-analysis-of-cloud-assisted-autonomous-driving-systems - A hardware-in-the-loop testbed to identify and quantify adversarial vulnerabilities in cloud-assisted autonomous driving systems.
- OmniSonic: Towards Universal and Holistic Audio Generation from Video and Text (viability: 8): https://sciencetostartup.com/paper/omnisonic-towards-universal-and-holistic-audio-generation-from-video-and-text - OmniSonic generates comprehensive audio scenes from video and text, capturing both on-screen and off-screen sounds.
- RoboPhD: Evolving Diverse Complex Agents Under Tight Evaluation Budgets (viability: 6): https://sciencetostartup.com/paper/robophd-evolving-diverse-complex-agents-under-tight-evaluation-budgets - RoboPhD optimizes AI agents efficiently under tight evaluation budgets using a novel validation-free Elo-based evolutionary approach.
- Domain-Contextualized Inference: A Computable Graph Architecture for Explicit-Domain Reasoning (viability: 0): https://sciencetostartup.com/paper/domain-contextualized-inference-a-computable-graph-architecture-for-explicit-domain-reasoning - An architectural framework for domain-contextualized inference that enables explicit-domain reasoning across various computational substrates.
- Deep Kuratowski Embedding Neural Networks for Wasserstein Metric Learning (viability: 4): https://sciencetostartup.com/paper/deep-kuratowski-embedding-neural-networks-for-wasserstein-metric-learning - Develops neural networks to accelerate Wasserstein distance calculations for data analysis.
- Generative models for decision-making under distributional shift (viability: 3): https://sciencetostartup.com/paper/generative-models-for-decision-making-under-distributional-shift - Leveraging generative models to create robust decision-making frameworks that adapt to shifting data distributions.
- Implementing surrogate goals for safer bargaining in LLM-based agents (viability: 5): https://sciencetostartup.com/paper/implementing-surrogate-goals-for-safer-bargaining-in-llm-based-agents - Implements and evaluates scaffolding and fine-tuning methods for LLM agents to safely implement surrogate goals in bargaining scenarios.
- Thermodynamic-Inspired Explainable GeoAI: Uncovering Regime-Dependent Mechanisms in Heterogeneous Spatial Systems (viability: 5): https://sciencetostartup.com/paper/thermodynamic-inspired-explainable-geoai-uncovering-regime-dependent-mechanisms-in-heterogeneous-spatial-systems - A thermodynamics-inspired GeoAI framework that uses graph neural networks to uncover regime-dependent mechanisms in spatial systems, improving interpretability and predictive performance.
- Boosted Distributional Reinforcement Learning: Analysis and Healthcare Applications (viability: 4): https://sciencetostartup.com/paper/boosted-distributional-reinforcement-learning-analysis-and-healthcare-applications - Boosted Distributional Reinforcement Learning algorithm that optimizes agent-specific outcome distributions for improved decision-making in uncertain healthcare scenarios.
- GA-GS: Generation-Assisted Gaussian Splatting for Static Scene Reconstruction (viability: 7): https://sciencetostartup.com/paper/ga-gs-generation-assisted-gaussian-splatting-for-static-scene-reconstruction - GA-GS uses a diffusion model to inpaint occluded regions in static scene reconstruction from monocular video, achieving state-of-the-art performance.
- Soft Tournament Equilibrium (viability: 5): https://sciencetostartup.com/paper/soft-tournament-equilibrium - Soft Tournament Equilibrium provides a differentiable framework for evaluating general-purpose AI agents by computing set-valued tournament solutions from pairwise comparisons.
- Benchmarking Multi-turn Medical Diagnosis: Hold, Lure, and Self-Correction (viability: 7): https://sciencetostartup.com/paper/benchmarking-multi-turn-medical-diagnosis-hold-lure-and-self-correction - This research introduces a benchmark and actionable guidance to improve the reliability of large language models in multi-turn medical diagnosis by addressing premature answering and leveraging self-correction.
- RESCORE: LLM-Driven Simulation Recovery in Control Systems Research Papers (viability: 7): https://sciencetostartup.com/paper/rescore-llm-driven-simulation-recovery-in-control-systems-research-papers - An LLM agent framework that automates the reconstruction of simulations from control systems research papers, achieving a 10x speedup over manual replication.
- How Well Do Agentic Skills Work in the Wild: Benchmarking LLM Skill Usage in Realistic Settings (viability: 7): https://sciencetostartup.com/paper/how-well-do-agentic-skills-work-in-the-wild-benchmarking-llm-skill-usage-in-realistic-settings - This research benchmarks LLM agent skill usage in realistic settings and proposes refinement strategies to improve performance, with code available.
- Effects of Generative AI Errors on User Reliance Across Task Difficulty (viability: 4): https://sciencetostartup.com/paper/effects-of-generative-ai-errors-on-user-reliance-across-task-difficulty - Investigating how users' reliance on generative AI changes based on error rates and task difficulty in diagram generation.
- How Long short-term memory artificial neural network, synthetic data, and fine-tuning improve the classification of raw EEG data (viability: 3): https://sciencetostartup.com/paper/how-long-short-term-memory-artificial-neural-network-synthetic-data-and-fine-tuning-improve-the-classification-of-raw-ee - A machine learning pipeline using synthetic data, LSTMs, and fine-tuning to improve raw EEG data classification for implicit visual stimuli.
- Convolutional Neural Network and Adversarial Autoencoder in EEG images classification (viability: 3): https://sciencetostartup.com/paper/convolutional-neural-network-and-adversarial-autoencoder-in-eeg-images-classification - Applying computer vision and neural networks to classify human brain activity from EEG data for neuroscience research.
- Out-of-Air Computation: Enabling Structured Extraction from Wireless Superposition (viability: 2): https://sciencetostartup.com/paper/out-of-air-computation-enabling-structured-extraction-from-wireless-superposition - A novel framework for extracting computation from wireless superposition by exploiting structured coding, enabling direct operation on continuous-valued device data.
- frax: Fast Robot Kinematics and Dynamics in JAX (viability: 7): https://sciencetostartup.com/paper/frax-fast-robot-kinematics-and-dynamics-in-jax - A JAX-based library for high-performance robot kinematics and dynamics that runs efficiently on CPUs and accelerators, enabling real-time control and optimization.
- HighFM: Towards a Foundation Model for Learning Representations from High-Frequency Earth Observation Data (viability: 7): https://sciencetostartup.com/paper/highfm-towards-a-foundation-model-for-learning-representations-from-high-frequency-earth-observation-data - A foundation model for high temporal resolution Earth observation data, enabling real-time disaster detection and tracking.
- CavMerge: Merging K-means Based on Local Log-Concavity (viability: 4): https://sciencetostartup.com/paper/cavmerge-merging-k-means-based-on-local-log-concavity - A parameter-free, efficient K-means merging algorithm that produces more reliable clusters on non-linearly separable data.
- High-Stakes Personalization: Rethinking LLM Customization for Individual Investor Decision-Making (viability: 5): https://sciencetostartup.com/paper/high-stakes-personalization-rethinking-llm-customization-for-individual-investor-decision-making - A system for AI-augmented portfolio management that addresses the unique challenges of personalizing LLMs for high-stakes investor decision-making.
- A Persistent Homology Design Space for 3D Point Cloud Deep Learning (viability: 7): https://sciencetostartup.com/paper/a-persistent-homology-design-space-for-3d-point-cloud-deep-learning - A unified framework for integrating persistent homology into deep learning for 3D point clouds, improving accuracy and robustness.
- PanLUNA: An Efficient and Robust Query-Unified Multimodal Model for Edge Biosignal Intelligence (viability: 7): https://sciencetostartup.com/paper/panluna-an-efficient-and-robust-query-unified-multimodal-model-for-edge-biosignal-intelligence - A compact, pan-modal foundation model for edge biosignal intelligence that matches or exceeds larger models and deploys efficiently on low-power microcontrollers.
- Adaptive Cost-Efficient Evaluation for Reliable Patent Claim Validation (viability: 7): https://sciencetostartup.com/paper/adaptive-cost-efficient-evaluation-for-reliable-patent-claim-validation - An adaptive AI framework that significantly reduces the cost of patent claim validation by intelligently routing claims to an expert LLM, achieving state-of-the-art accuracy.
- Evaluating Future Air Traffic Management Security (viability: 4): https://sciencetostartup.com/paper/evaluating-future-air-traffic-management-security - Developing a more robust and secure authentication system for air traffic control communications by exploring PUF vulnerabilities and proposing PKI as an alternative.
- Correcting Source Mismatch in Flow Matching with Radial-Angular Transport (viability: 4): https://sciencetostartup.com/paper/correcting-source-mismatch-in-flow-matching-with-radial-angular-transport - A framework for generative modeling that corrects source mismatch in flow matching for heavy-tailed and anisotropic data.
- DAGAF: A directed acyclic generative adversarial framework for joint structure learning and tabular data synthesis (viability: 7): https://sciencetostartup.com/paper/dagaf-a-directed-acyclic-generative-adversarial-framework-for-joint-structure-learning-and-tabular-data-synthesis - A novel framework for learning causal relationships and synthesizing realistic tabular data under multiple causal models, outperforming existing methods in structure learning.
- Poisoned Identifiers Survive LLM Deobfuscation: A Case Study on Claude Opus 4.6 (viability: 5): https://sciencetostartup.com/paper/poisoned-identifiers-survive-llm-deobfuscation-a-case-study-on-claude-opus-4-6 - Identifies how poisoned identifiers in JavaScript can survive LLM deobfuscation, impacting code integrity and revealing vulnerabilities in LLM code understanding.
- LLM-Enabled Open-Source Systems in the Wild: An Empirical Study of Vulnerabilities in GitHub Security Advisories (viability: 5): https://sciencetostartup.com/paper/llm-enabled-open-source-systems-in-the-wild-an-empirical-study-of-vulnerabilities-in-github-security-advisories - This study analyzes GitHub security advisories to identify and categorize vulnerabilities in LLM-enabled open-source systems, revealing that existing frameworks underrepresent model-mediated risks.
- Entropy, Disagreement, and the Limits of Foundation Models in Genomics (viability: 3): https://sciencetostartup.com/paper/entropy-disagreement-and-the-limits-of-foundation-models-in-genomics - This research investigates the fundamental limitations of foundation models in genomics due to sequence entropy, suggesting current self-supervised training methods may be inapplicable.
- Real-Time Projected Adaptive Control for Closed-Chain Co-Manipulative Continuum Robots (viability: 4): https://sciencetostartup.com/paper/real-time-projected-adaptive-control-for-closed-chain-co-manipulative-continuum-robots - A projected adaptive control framework for cooperative continuum robots to handle dynamic interactions and unknown object parameters in real-time.
- Semantics Over Syntax: Uncovering Pre-Authentication 5G Baseband Vulnerabilities (viability: 7): https://sciencetostartup.com/paper/semantics-over-syntax-uncovering-pre-authentication-5g-baseband-vulnerabilities - A framework that finds critical pre-authentication vulnerabilities in 5G baseband implementations by generating semantically inconsistent messages, impacting millions of devices.
- Preservation Is Not Enough for Width Growth: Regime-Sensitive Selection of Dense LM Warm Starts (viability: 1): https://sciencetostartup.com/paper/preservation-is-not-enough-for-width-growth-regime-sensitive-selection-of-dense-lm-warm-starts - Investigates regime-sensitive selection of dense language model warm starts for width expansion, finding preservation is not a universal ranking criterion.
- InferenceEvolve: Towards Automated Causal Effect Estimators through Self-Evolving AI (viability: 7): https://sciencetostartup.com/paper/inferenceevolve-towards-automated-causal-effect-estimators-through-self-evolving-ai - An evolutionary framework using LLMs to automatically discover and refine causal inference methods, outperforming human-designed estimators on benchmarks.
- A Family of Open Time-Series Foundation Models for the Radio Access Network (viability: 7): https://sciencetostartup.com/paper/a-family-of-open-time-series-foundation-models-for-the-radio-access-network - A unified foundation model for time-series analysis in Radio Access Networks, enabling efficient adaptation to diverse tasks with state-of-the-art performance and a large open-source dataset.
- A Logical-Rule Autoencoder for Interpretable Recommendations (viability: 7): https://sciencetostartup.com/paper/a-logical-rule-autoencoder-for-interpretable-recommendations - A recommendation system that generates interpretable logical rules for transparent decision-making.
- Beyond Fluency: Toward Reliable Trajectories in Agentic IR (viability: 3): https://sciencetostartup.com/paper/beyond-fluency-toward-reliable-trajectories-in-agentic-ir - This paper proposes a framework for improving the reliability of autonomous AI agents by implementing verification gates and abstention under uncertainty to prevent cascading errors in multi-step workflows.
- Governance-Constrained Agentic AI: Blockchain-Enforced Human Oversight for Safety-Critical Wildfire Monitoring (viability: 6): https://sciencetostartup.com/paper/governance-constrained-agentic-ai-blockchain-enforced-human-oversight-for-safety-critical-wildfire-monitoring - A blockchain-enforced governance system for autonomous wildfire monitoring agents that ensures human oversight and verifiable responsibility, reducing false alarms and improving detection performance.
- Avoiding Non-Integrable Beliefs in Expectation Propagation (viability: 2): https://sciencetostartup.com/paper/avoiding-non-integrable-beliefs-in-expectation-propagation - This paper proposes novel Expectation Propagation frameworks to ensure integrable beliefs in Bayesian estimation problems, addressing limitations of existing methods.
- Commercial Persuasion in AI-Mediated Conversations (viability: 5): https://sciencetostartup.com/paper/commercial-persuasion-in-ai-mediated-conversations - Conversational AI agents covertly redirect consumer choices at scale, tripling the selection of sponsored products with minimal user detection, highlighting insufficient transparency mechanisms.
- Agents for Agents: An Interrogator-Based Secure Framework for Autonomous Internet of Underwater Things (viability: 5): https://sciencetostartup.com/paper/agents-for-agents-an-interrogator-based-secure-framework-for-autonomous-internet-of-underwater-things - A secure framework for underwater autonomous agents that uses behavioral trust monitoring and blockchain for identity management to improve data forwarding authorization and containment of deviating agents.
- APPA: Adaptive Preference Pluralistic Alignment for Fair Federated RLHF of LLMs (viability: 4): https://sciencetostartup.com/paper/appa-adaptive-preference-pluralistic-alignment-for-fair-federated-rlhf-of-llms - A framework for fair federated reinforcement learning from human feedback that dynamically reweights group rewards to prioritize under-aligned groups without degrading overall alignment.
- Context Engineering: A Practitioner Methodology for Structured Human-AI Collaboration (viability: 7): https://sciencetostartup.com/paper/context-engineering-a-practitioner-methodology-for-structured-human-ai-collaboration - A structured methodology and pipeline for assembling complete informational payloads to AI tools, significantly reducing iteration cycles and improving first-pass acceptance.
- Towards Unveiling Vulnerabilities of Large Reasoning Models in Machine Unlearning (viability: 7): https://sciencetostartup.com/paper/towards-unveiling-vulnerabilities-of-large-reasoning-models-in-machine-unlearning - Develops a novel attack to uncover security vulnerabilities in large reasoning models during the machine unlearning process, enabling proactive defense strategies.
- MC-CPO: Mastery-Conditioned Constrained Policy Optimization (viability: 4): https://sciencetostartup.com/paper/mc-cpo-mastery-conditioned-constrained-policy-optimization - A constrained policy optimization algorithm for adaptive tutoring systems that integrates structural action masking to mitigate reward hacking and ensure sustained learning outcomes.
- CAWN: Continuous Acoustic Wave Networks for Autoregressive Language Modeling (viability: 7): https://sciencetostartup.com/paper/cawn-continuous-acoustic-wave-networks-for-autoregressive-language-modeling - A novel continuous sequence-mixing architecture for LLMs that overcomes the quadratic scaling of Transformers and the signal degradation of SSMs, enabling efficient processing of ultra-long contexts.
- Combee: Scaling Prompt Learning for Self-Improving Language Model Agents (viability: 7): https://sciencetostartup.com/paper/combee-scaling-prompt-learning-for-self-improving-language-model-agents - A novel framework for parallel prompt learning that enables self-improving language model agents to learn efficiently from aggregate traces without quality degradation.
- Transmission Neural Networks: Inhibitory and Excitatory Connections (viability: 2): https://sciencetostartup.com/paper/transmission-neural-networks-inhibitory-and-excitatory-connections - This paper theoretically extends a neural network model to incorporate inhibitory connections and neurotransmitter dynamics, analyzing stability and contraction properties.
- Learning An Interpretable Risk Scoring System for Maximizing Decision Net Benefit (viability: 7): https://sciencetostartup.com/paper/learning-an-interpretable-risk-scoring-system-for-maximizing-decision-net-benefit - Develop an interpretable risk scoring system that directly optimizes net benefit for high-stakes decision-making.
- Peoples Water Data: Enabling Reliable Field Data Generation and Microbial Contamination Screening in Household Drinking Water (viability: 5): https://sciencetostartup.com/paper/peoples-water-data-enabling-reliable-field-data-generation-and-microbial-contamination-screening-in-household-drinking-w - A machine learning framework to predict E. coli contamination in household drinking water using low-cost indicators, enabling targeted testing in resource-constrained regions.
- Good Rankings, Wrong Probabilities: A Calibration Audit of Multimodal Cancer Survival Models (viability: 2): https://sciencetostartup.com/paper/good-rankings-wrong-probabilities-a-calibration-audit-of-multimodal-cancer-survival-models - This paper audits the calibration of multimodal cancer survival prediction models, revealing significant miscalibration issues despite strong discriminative performance.
- Pedagogical Safety in Educational Reinforcement Learning: Formalizing and Detecting Reward Hacking in AI Tutoring Systems (viability: 2): https://sciencetostartup.com/paper/pedagogical-safety-in-educational-reinforcement-learning-formalizing-and-detecting-reward-hacking-in-ai-tutoring-systems - This research formalizes and detects reward hacking in AI tutoring systems, proposing a framework and index to measure pedagogical safety.
- Precise Robot Command Understanding Using Grammar-Constrained Large Language Models (viability: 7): https://sciencetostartup.com/paper/precise-robot-command-understanding-using-grammar-constrained-large-language-models - A grammar-constrained LLM for precise, robot-readable industrial commands, improving safety and efficiency in human-robot collaboration.
- Subspace Control: Turning Constrained Model Steering into Controllable Spectral Optimization (viability: 7): https://sciencetostartup.com/paper/subspace-control-turning-constrained-model-steering-into-controllable-spectral-optimization - A spectral interference-free training method that enables controllable LLM adaptation for safety, privacy, and task-specific requirements.
- Three Phases of Expert Routing: How Load Balance Evolves During Mixture-of-Experts Training (viability: 2): https://sciencetostartup.com/paper/three-phases-of-expert-routing-how-load-balance-evolves-during-mixture-of-experts-training - This paper models Mixture-of-Experts token routing as a congestion game, revealing a three-phase evolution of load balance and expert specialization during training.
- Hierarchical Semantic Correlation-Aware Masked Autoencoder for Unsupervised Audio-Visual Representation Learning (viability: 7): https://sciencetostartup.com/paper/hierarchical-semantic-correlation-aware-masked-autoencoder-for-unsupervised-audio-visual-representation-learning - HSC-MAE learns aligned audio-visual embeddings from label-free data by enforcing semantic consistency across global, local, and sample levels.
- Agentization of Digital Assets for the Agentic Web: Concepts, Techniques, and Benchmark (viability: 7): https://sciencetostartup.com/paper/agentization-of-digital-assets-for-the-agentic-web-concepts-techniques-and-benchmark - An agentization agent and benchmark to automate the creation of digital assets for the Agentic Web, enabling scalable integration.
- Learning from Imperfect Demonstrations via Temporal Behavior Tree-Guided Trajectory Repair (viability: 4): https://sciencetostartup.com/paper/learning-from-imperfect-demonstrations-via-temporal-behavior-tree-guided-trajectory-repair - A formal framework using Temporal Behavior Trees to repair imperfect robot demonstrations for improved policy learning.
- RK-MPC: Residual Koopman Model Predictive Control for Quadruped Locomotion in Offroad Environments (viability: 7): https://sciencetostartup.com/paper/rk-mpc-residual-koopman-model-predictive-control-for-quadruped-locomotion-in-offroad-environments - A data-driven model predictive control framework for quadruped robots that enables reliable locomotion in challenging off-road environments.
- TimeSeek: Temporal Reliability of Agentic Forecasters (viability: 7): https://sciencetostartup.com/paper/timeseek-temporal-reliability-of-agentic-forecasters - TimeSeek benchmark evaluates LLM forecaster reliability over prediction market lifecycles, revealing time-aware evaluation needs.
- Sharp asymptotic theory for Q-learning with LDTZ learning rate and its generalization (viability: 3): https://sciencetostartup.com/paper/sharp-asymptotic-theory-for-q-learning-with-ldtz-learning-rate-and-its-generalization - This paper provides a theoretical analysis of Q-learning with a specific learning rate schedule, offering insights into its convergence properties and enabling bootstrap-based inference.
- DARE: Diffusion Large Language Models Alignment and Reinforcement Executor (viability: 7): https://sciencetostartup.com/paper/dare-diffusion-large-language-models-alignment-and-reinforcement-executor - DARE is an open framework that unifies and accelerates the development, comparison, and deployment of post-training methods for diffusion large language models.
- LOCARD: An Agentic Framework for Blockchain Forensics (viability: 8): https://sciencetostartup.com/paper/locard-an-agentic-framework-for-blockchain-forensics - LOCARD empowers blockchain forensic analysts with an agentic framework for high-fidelity cross-chain investigation and tracing.
- Towards Agentic Defect Reasoning: A Graph-Assisted Retrieval Framework for Laser Powder Bed Fusion (viability: 7): https://sciencetostartup.com/paper/towards-agentic-defect-reasoning-a-graph-assisted-retrieval-framework-for-laser-powder-bed-fusion - A graph-assisted retrieval framework that converts scattered scientific literature into an interpretable knowledge resource for understanding and predicting defects in additive manufacturing.
- Don't Blink: Evidence Collapse during Multimodal Reasoning (viability: 7): https://sciencetostartup.com/paper/don-t-blink-evidence-collapse-during-multimodal-reasoning - Develops a targeted vision veto mechanism to improve the safety and reliability of multimodal reasoning models by detecting and mitigating 'evidence collapse' during complex tasks.
- CoME-VL: Scaling Complementary Multi-Encoder Vision-Language Learning (viability: 7): https://sciencetostartup.com/paper/come-vl-scaling-complementary-multi-encoder-vision-language-learning - A modular framework that fuses complementary vision encoders to significantly improve performance on vision-language tasks, achieving state-of-the-art results.
- Enhancing Robustness of Federated Learning via Server Learning (viability: 4): https://sciencetostartup.com/paper/enhancing-robustness-of-federated-learning-via-server-learning - A federated learning enhancement that improves model robustness against malicious attacks using server-side learning and client update filtering.
- VOSR: A Vision-Only Generative Model for Image Super-Resolution (viability: 7): https://sciencetostartup.com/paper/vosr-a-vision-only-generative-model-for-image-super-resolution - A vision-only generative model for image super-resolution that achieves competitive results with significantly lower training costs than text-to-image based methods.
- BAS: A Decision-Theoretic Approach to Evaluating Large Language Model Confidence (viability: 7): https://sciencetostartup.com/paper/bas-a-decision-theoretic-approach-to-evaluating-large-language-model-confidence - A new decision-theoretic metric for evaluating LLM confidence that prioritizes avoiding overconfident errors, with code and benchmark available.
- ProtoFlow: Mitigating Forgetting in Class-Incremental Remote Sensing Segmentation via Low-Curvature Prototype Flow (viability: 5): https://sciencetostartup.com/paper/protoflow-mitigating-forgetting-in-class-incremental-remote-sensing-segmentation-via-low-curvature-prototype-flow - A framework for continual remote sensing segmentation that mitigates forgetting by modeling class prototype evolution as low-curvature trajectories.
- Hierarchical Planning with Latent World Models (viability: 7): https://sciencetostartup.com/paper/hierarchical-planning-with-latent-world-models - Enabling zero-shot, long-horizon robotic control with hierarchical latent world models that significantly outperform single-level approaches.
- A Tsetlin Machine-driven Intrusion Detection System for Next-Generation IoMT Security (viability: 5): https://sciencetostartup.com/paper/a-tsetlin-machine-driven-intrusion-detection-system-for-next-generation-iomt-security - A Tsetlin Machine-based Intrusion Detection System for IoMT networks that offers interpretable insights into cyberattacks.
- PR3DICTR: A modular AI framework for medical 3D image-based detection and outcome prediction (viability: 5): https://sciencetostartup.com/paper/pr3dictr-a-modular-ai-framework-for-medical-3d-image-based-detection-and-outcome-prediction - A modular framework for developing 3D medical image classification and prediction models with pre-established functionality.
- Coupled Control, Structured Memory, and Verifiable Action in Agentic AI (SCRAT -- Stochastic Control with Retrieval and Auditable Trajectories): A Comparative Perspective from Squirrel Locomotion and Scatter-Hoarding (viability: 4): https://sciencetostartup.com/paper/coupled-control-structured-memory-and-verifiable-action-in-agentic-ai-scrat-stochastic-control-with-retrieval-and-audita - A new framework for agentic AI that integrates control, memory, and verification, inspired by squirrel behavior, to improve robustness and reduce errors.
- Safety-Critical Centralized Nonlinear MPC for Cooperative Payload Transportation by Two Quadrupedal Robots (viability: 4): https://sciencetostartup.com/paper/safety-critical-centralized-nonlinear-mpc-for-cooperative-payload-transportation-by-two-quadrupedal-robots - A safety-critical NMPC framework for cooperative payload transportation by two quadrupedal robots, validated on hardware.
- Learning the Signature of Memorization in Autoregressive Language Models (viability: 7): https://sciencetostartup.com/paper/learning-the-signature-of-memorization-in-autoregressive-language-models - A transferable learned attack that detects memorization in language models across different architectures and data domains, outperforming existing methods.
- The Eleventh NTIRE 2026 Efficient Super-Resolution Challenge Report (viability: 3): https://sciencetostartup.com/paper/the-eleventh-ntire-2026-efficient-super-resolution-challenge-report - Develops efficient single-image super-resolution networks that significantly reduce computational cost while maintaining high image quality.
- Real-Time Surrogate Modeling for Personalized Blood Flow Prediction and Hemodynamic Analysis (viability: 7): https://sciencetostartup.com/paper/real-time-surrogate-modeling-for-personalized-blood-flow-prediction-and-hemodynamic-analysis - Develops a real-time AI surrogate model for personalized blood flow prediction and hemodynamic analysis, enabling rapid patient screening and parameter estimation from clinical data.
- Reliability Gated Multi-Teacher Distillation for Low Resource Abstractive Summarization (viability: 5): https://sciencetostartup.com/paper/reliability-gated-multi-teacher-distillation-for-low-resource-abstractive-summarization - A novel distillation method for abstractive summarization in low-resource languages that improves performance and reduces model size.
- The Compression Gap: Why Discrete Tokenization Limits Vision-Language-Action Model Scaling (viability: 4): https://sciencetostartup.com/paper/the-compression-gap-why-discrete-tokenization-limits-vision-language-action-model-scaling - This research identifies a critical bottleneck in vision-language-action model scaling, suggesting a new direction for improving robotic manipulation performance by addressing information flow rather than just model size.
- Gradient Boosting within a Single Attention Layer (viability: 3): https://sciencetostartup.com/paper/gradient-boosting-within-a-single-attention-layer - A novel attention mechanism that applies gradient boosting within a single layer to improve Transformer performance.
- Reflective Context Learning: Studying the Optimization Primitives of Context Space (viability: 7): https://sciencetostartup.com/paper/reflective-context-learning-studying-the-optimization-primitives-of-context-space - A unified framework for agents that learn through iterative context updates, improving generalization and addressing core optimization challenges.
- Multi-View Video Diffusion Policy: A 3D Spatio-Temporal-Aware Video Action Model (viability: 7): https://sciencetostartup.com/paper/multi-view-video-diffusion-policy-a-3d-spatio-temporal-aware-video-action-model - A multi-view video diffusion policy for data-efficient, robust, and generalizable robotic manipulation that outperforms existing state-of-the-art models.
- PRISM: LLM-Guided Semantic Clustering for High-Precision Topics (viability: 5): https://sciencetostartup.com/paper/prism-llm-guided-semantic-clustering-for-high-precision-topics - A topic modeling framework that uses LLMs to guide semantic clustering for precise topic discovery and analysis.
- Understanding the Role of Hallucination in Reinforcement Post-Training of Multimodal Reasoning Models (viability: 4): https://sciencetostartup.com/paper/understanding-the-role-of-hallucination-in-reinforcement-post-training-of-multimodal-reasoning-models - A framework to analyze and improve multimodal reasoning models by understanding and leveraging model hallucination during reinforcement learning post-training.
- SFFNet: Synergistic Feature Fusion Network With Dual-Domain Edge Enhancement for UAV Image Object Detection (viability: 7): https://sciencetostartup.com/paper/sffnet-synergistic-feature-fusion-network-with-dual-domain-edge-enhancement-for-uav-image-object-detection - A novel object detection network for UAV imagery that effectively separates objects from complex backgrounds and handles scale variations, with code available.
- Beyond the Parameters: A Technical Survey of Contextual Enrichment in Large Language Models: From In-Context Prompting to Causal Retrieval-Augmented Generation (viability: 3): https://sciencetostartup.com/paper/beyond-the-parameters-a-technical-survey-of-contextual-enrichment-in-large-language-models-from-in-context-prompting-to - This paper surveys methods for enriching LLM knowledge at inference time, from simple prompting to advanced retrieval techniques, to improve reasoning and reduce reliance on static parameters.
- Detecting and Correcting Reference Hallucinations in Commercial LLMs and Deep Research Agents (viability: 7): https://sciencetostartup.com/paper/detecting-and-correcting-reference-hallucinations-in-commercial-llms-and-deep-research-agents - A tool for validating citation URLs in LLMs to reduce hallucinations and improve reliability.
- EffiMiniVLM: A Compact Dual-Encoder Regression Framework (viability: 7): https://sciencetostartup.com/paper/effiminivlm-a-compact-dual-encoder-regression-framework - A compact, resource-efficient dual-encoder vision-language model for product quality prediction in cold-start scenarios, outperforming larger models with significantly less computational cost.
- BibTeX Citation Hallucinations in Scientific Publishing Agents: Evaluation and Mitigation (viability: 7): https://sciencetostartup.com/paper/bibtex-citation-hallucinations-in-scientific-publishing-agents-evaluation-and-mitigation - A tool to significantly improve the accuracy of BibTeX entries generated by LLM-powered scientific publishing agents by integrating authoritative citation records.
- Chart-RL: Policy Optimization Reinforcement Learning for Enhanced Visual Reasoning in Chart Question Answering with Vision Language Models (viability: 7): https://sciencetostartup.com/paper/chart-rl-policy-optimization-reinforcement-learning-for-enhanced-visual-reasoning-in-chart-question-answering-with-visio - Reinforcement learning framework to significantly improve visual reasoning and numerical extraction for chart-based question answering, outperforming larger models with reduced latency.
- CAMEO: A Conditional and Quality-Aware Multi-Agent Image Editing Orchestrator (viability: 7): https://sciencetostartup.com/paper/cameo-a-conditional-and-quality-aware-multi-agent-image-editing-orchestrator - CAMEO is a multi-agent framework that orchestrates conditional image editing through a quality-aware, feedback-driven process, significantly improving robustness and controllability.
- DSBD: Dual-Aligned Structural Basis Distillation for Graph Domain Adaptation (viability: 7): https://sciencetostartup.com/paper/dsbd-dual-aligned-structural-basis-distillation-for-graph-domain-adaptation - A novel framework for graph domain adaptation that explicitly models and adapts cross-domain structural variations, outperforming state-of-the-art methods.
- HyperFitS -- Hypernetwork Fitting Spectra for metabolic quantification of ${}^1$H MR spectroscopic imaging (viability: 5): https://sciencetostartup.com/paper/hyperfits-hypernetwork-fitting-spectra-for-metabolic-quantification-of-1-h-mr-spectroscopic-imaging - A hypernetwork for rapid and configurable metabolic quantification in brain MRI, reducing processing time from hours to seconds.
- Valence-Arousal Subspace in LLMs: Circular Emotion Geometry and Multi-Behavioral Control (viability: 7): https://sciencetostartup.com/paper/valence-arousal-subspace-in-llms-circular-emotion-geometry-and-multi-behavioral-control - A method to steer LLM outputs for precise control over emotional valence, arousal, refusal, and sycophancy across multiple architectures.
- Characterization of Gaussian Universality Breakdown in High-Dimensional Empirical Risk Minimization (viability: 2): https://sciencetostartup.com/paper/characterization-of-gaussian-universality-breakdown-in-high-dimensional-empirical-risk-minimization - This paper provides a theoretical framework for understanding the behavior of empirical risk minimization in high-dimensional, non-Gaussian settings, extending existing Gaussian universality theorems.
- InCoder-32B-Thinking: Industrial Code World Model for Thinking (viability: 7): https://sciencetostartup.com/paper/incoder-32b-thinking-industrial-code-world-model-for-thinking - An industrial code world model that generates reasoning traces for chip design, GPU optimization, and embedded systems, validated by domain toolchains.
- Beyond Precision: Importance-Aware Recall for Factuality Evaluation in Long-Form LLM Generation (viability: 4): https://sciencetostartup.com/paper/beyond-precision-importance-aware-recall-for-factuality-evaluation-in-long-form-llm-generation - A framework to evaluate LLM factuality by jointly measuring precision and recall, highlighting factual incompleteness as a key limitation.
- FSUNav: A Cerebrum-Cerebellum Architecture for Fast, Safe, and Universal Zero-Shot Goal-Oriented Navigation (viability: 7): https://sciencetostartup.com/paper/fsunav-a-cerebrum-cerebellum-architecture-for-fast-safe-and-universal-zero-shot-goal-oriented-navigation - FSUNav provides a unified zero-shot goal-oriented navigation system for diverse robotic platforms ensuring fast, safe, and semantic-rich interaction with complex environments.
- StoryScope: Investigating idiosyncrasies in AI fiction (viability: 5): https://sciencetostartup.com/paper/storyscope-investigating-idiosyncrasies-in-ai-fiction - A novel pipeline that detects AI-generated fiction by analyzing narrative structure, not just style, offering a robust method for authorship attribution and content verification.
- AI-Assisted Unit Test Writing and Test-Driven Code Refactoring: A Case Study (viability: 5): https://sciencetostartup.com/paper/ai-assisted-unit-test-writing-and-test-driven-code-refactoring-a-case-study - Automate unit test generation and safe code refactoring using AI to accelerate software development and reduce regression risk.
- SD-FSMIS: Adapting Stable Diffusion for Few-Shot Medical Image Segmentation (viability: 7): https://sciencetostartup.com/paper/sd-fsmis-adapting-stable-diffusion-for-few-shot-medical-image-segmentation - Adapting Stable Diffusion for few-shot medical image segmentation to address data scarcity and domain shifts.
- A Systematic Security Evaluation of OpenClaw and Its Variants (viability: 4): https://sciencetostartup.com/paper/a-systematic-security-evaluation-of-openclaw-and-its-variants - This research systematically evaluates security vulnerabilities in AI agent frameworks, revealing critical risks in tool use and multi-step planning that require lifecycle-wide governance.
- Self-Distilled RLVR (viability: 7): https://sciencetostartup.com/paper/self-distilled-rlvr - A novel self-distillation method for LLMs that combines sparse environmental rewards with dense token-level policy differences for more stable and effective training.
- Domain-Adapted Retrieval for In-Context Annotation of Pedagogical Dialogue Acts (viability: 7): https://sciencetostartup.com/paper/domain-adapted-retrieval-for-in-context-annotation-of-pedagogical-dialogue-acts - A domain-adapted RAG pipeline for accurate pedagogical dialogue annotation by fine-tuning embedding models and indexing dialogues at the utterance level, outperforming baselines and addressing label biases.
- An Independent Safety Evaluation of Kimi K2.5 (viability: 4): https://sciencetostartup.com/paper/an-independent-safety-evaluation-of-kimi-k2-5 - An independent safety evaluation of the Kimi K2.5 LLM reveals significant dual-use capabilities and amplified risks in open-weight models, urging developers to prioritize systematic safety assessments.
- SCC-Loc: A Unified Semantic Cascade Consensus Framework for UAV Thermal Geo-Localization (viability: 7): https://sciencetostartup.com/paper/scc-loc-a-unified-semantic-cascade-consensus-framework-for-uav-thermal-geo-localization - Revolutionizing UAV navigation with a zero-shot thermal geo-localization system, SCC-Loc.
- Salt: Self-Consistent Distribution Matching with Cache-Aware Training for Fast Video Generation (viability: 7): https://sciencetostartup.com/paper/salt-self-consistent-distribution-matching-with-cache-aware-training-for-fast-video-generation - A new distillation method for real-time video generation that improves quality and reduces inference costs by ensuring self-consistency across denoising steps and optimizing for KV cache usage.
- Revealing Physical-World Semantic Vulnerabilities: Universal Adversarial Patches for Infrared Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/revealing-physical-world-semantic-vulnerabilities-universal-adversarial-patches-for-infrared-vision-language-models - Develops universal physical adversarial patches to reveal and exploit critical vulnerabilities in infrared vision-language models for enhanced security and robustness testing.
- Can VLMs Truly Forget? Benchmarking Training-Free Visual Concept Unlearning (viability: 4): https://sciencetostartup.com/paper/can-vlms-truly-forget-benchmarking-training-free-visual-concept-unlearning - A new benchmark to rigorously evaluate the ability of large vision-language models to forget specific visual concepts without retraining.
- Multi-Aspect Knowledge Distillation for Language Model with Low-rank Factorization (viability: 5): https://sciencetostartup.com/paper/multi-aspect-knowledge-distillation-for-language-model-with-low-rank-factorization - A novel knowledge distillation method for language models that captures richer language knowledge by mimicking self-attention and feed-forward modules.
- AlertStar: Path-Aware Alert Prediction on Hyper-Relational Knowledge Graphs (viability: 5): https://sciencetostartup.com/paper/alertstar-path-aware-alert-prediction-on-hyper-relational-knowledge-graphs - AlertStar enhances network intrusion detection by predicting alerts through advanced hyper-relational knowledge graph modeling.
- Co-Evolution of Policy and Internal Reward for Language Agents (viability: 7): https://sciencetostartup.com/paper/co-evolution-of-policy-and-internal-reward-for-language-agents - Develops a novel self-generated internal reward system for LLM agents to improve long-horizon decision-making and policy optimization.
- An Open-Source LiDAR and Monocular Off-Road Autonomous Navigation Stack (viability: 7): https://sciencetostartup.com/paper/an-open-source-lidar-and-monocular-off-road-autonomous-navigation-stack - An open-source autonomous navigation stack that uses monocular depth estimation as a cost-effective alternative to LiDAR for off-road environments.
- A Data-Centric Vision Transformer Baseline for SAR Sea Ice Classification (viability: 5): https://sciencetostartup.com/paper/a-data-centric-vision-transformer-baseline-for-sar-sea-ice-classification - A data-centric Vision Transformer baseline for improved SAR sea ice classification, offering a more useful precision-recall trade-off for rare ice classes.
- Flash-Mono: Feed-Forward Accelerated Gaussian Splatting Monocular SLAM (viability: 8): https://sciencetostartup.com/paper/flash-mono-feed-forward-accelerated-gaussian-splatting-monocular-slam - Flash-Mono accelerates monocular SLAM with Gaussian splatting for real-time 3D scene reconstruction.
- SkillRT: Compiling Skills for Efficient Execution Everywhere (viability: 7): https://sciencetostartup.com/paper/skillrt-compiling-skills-for-efficient-execution-everywhere - SkillRT compiles LLM agent skills into portable and efficient code, improving task completion and reducing costs.
- Supply-Chain Poisoning Attacks Against LLM Coding Agent Skill Ecosystems (viability: 5): https://sciencetostartup.com/paper/supply-chain-poisoning-attacks-against-llm-coding-agent-skill-ecosystems - This research identifies a novel supply-chain attack vector for LLM coding agents by embedding malicious logic in skill documentation, bypassing existing defenses.
- MI-Pruner: Crossmodal Mutual Information-guided Token Pruner for Efficient MLLMs (viability: 7): https://sciencetostartup.com/paper/mi-pruner-crossmodal-mutual-information-guided-token-pruner-for-efficient-mllms - A novel crossmodal mutual information-guided token pruning method for significantly improving multimodal large language model efficiency without architectural changes.
- Automatic Textbook Formalization (viability: 7): https://sciencetostartup.com/paper/automatic-textbook-formalization - An AI system automatically formalizes graduate-level textbooks into verifiable code, achieving unprecedented scale and cost-efficiency.
- Credential Leakage in LLM Agent Skills: A Large-Scale Empirical Study (viability: 7): https://sciencetostartup.com/paper/credential-leakage-in-llm-agent-skills-a-large-scale-empirical-study - A large-scale empirical study and detection pipeline to identify and mitigate credential leakage vulnerabilities in LLM agent skills.
- SparseSplat: Towards Applicable Feed-Forward 3D Gaussian Splatting with Pixel-Unaligned Prediction (viability: 7): https://sciencetostartup.com/paper/sparsesplat-towards-applicable-feed-forward-3d-gaussian-splatting-with-pixel-unaligned-prediction - SparseSplat enables highly compact and efficient feed-forward 3D Gaussian Splatting for downstream reconstruction tasks by adaptively adjusting Gaussian density.
- Joint Prediction of Human Motions and Actions in Human-Robot Collaboration (viability: 4): https://sciencetostartup.com/paper/joint-prediction-of-human-motions-and-actions-in-human-robot-collaboration - A probabilistic framework for robots to jointly predict human movements and actions for improved collaboration.
- Gram-MMD: A Texture-Aware Metric for Image Realism Assessment (viability: 7): https://sciencetostartup.com/paper/gram-mmd-a-texture-aware-metric-for-image-realism-assessment - A novel texture-aware metric for image realism assessment that outperforms existing methods by capturing fine-grained textural information, offering a more robust evaluation for generative models.
- Can Nano Banana 2 Replace Traditional Image Restoration Models? An Evaluation of Its Performance on Image Restoration Tasks (viability: 7): https://sciencetostartup.com/paper/can-nano-banana-2-replace-traditional-image-restoration-models-an-evaluation-of-its-performance-on-image-restoration-tas - A general-purpose generative AI model shows superior performance in image restoration tasks, outperforming specialized models with careful prompt engineering.
- Verbalizing LLMs' assumptions to explain and control sycophancy (viability: 5): https://sciencetostartup.com/paper/verbalizing-llms-assumptions-to-explain-and-control-sycophancy - A framework to understand and control LLM sycophancy by verbalizing and steering their underlying assumptions.
- Querying Structured Data Through Natural Language Using Language Models (viability: 7): https://sciencetostartup.com/paper/querying-structured-data-through-natural-language-using-language-models - A fine-tuned LLM that generates executable queries for structured data, overcoming RAG limitations for numerical and complex datasets.
- MECO: A Multimodal Dataset for Emotion and Cognitive Understanding in Older Adults (viability: 7): https://sciencetostartup.com/paper/meco-a-multimodal-dataset-for-emotion-and-cognitive-understanding-in-older-adults - A new multimodal dataset and baseline models for understanding emotion and cognitive decline in older adults, enabling personalized care and early MCI detection.
- STEAR: Layer-Aware Spatiotemporal Evidence Intervention for Hallucination Mitigation in Video Large Language Models (viability: 7): https://sciencetostartup.com/paper/stear-layer-aware-spatiotemporal-evidence-intervention-for-hallucination-mitigation-in-video-large-language-models - A framework to reduce hallucinations in Video-LLMs by intervening on layer-specific visual evidence during decoding.
- JoyAI-LLM Flash: Advancing Mid-Scale LLMs with Token Efficiency (viability: 5): https://sciencetostartup.com/paper/joyai-llm-flash-advancing-mid-scale-llms-with-token-efficiency - An efficient Mixture-of-Experts LLM that significantly improves token efficiency and inference throughput for mid-scale models.
- Analyzing Healthcare Interoperability Vulnerabilities: Formal Modeling and Graph-Theoretic Approach (viability: 4): https://sciencetostartup.com/paper/analyzing-healthcare-interoperability-vulnerabilities-formal-modeling-and-graph-theoretic-approach - A formal graph-based model to detect race conditions in healthcare interoperability platforms like FHIR.
- Enhancing Multi-Robot Exploration Using Probabilistic Frontier Prioritization with Dirichlet Process Gaussian Mixtures (viability: 5): https://sciencetostartup.com/paper/enhancing-multi-robot-exploration-using-probabilistic-frontier-prioritization-with-dirichlet-process-gaussian-mixtures - A probabilistic approach to frontier prioritization enhances multi-robot exploration efficiency by 10-14% in complex environments.
- QVAD: A Question-Centric Agentic Framework for Efficient and Training-Free Video Anomaly Detection (viability: 7): https://sciencetostartup.com/paper/qvad-a-question-centric-agentic-framework-for-efficient-and-training-free-video-anomaly-detection - A question-centric agentic framework that uses dynamic dialogue to enable lightweight VLMs to achieve state-of-the-art video anomaly detection with high inference speeds and minimal memory footprints.
- GenSmoke-GS: A Multi-Stage Method for Novel View Synthesis from Smoke-Degraded Images Using a Generative Model (viability: 7): https://sciencetostartup.com/paper/gensmoke-gs-a-multi-stage-method-for-novel-view-synthesis-from-smoke-degraded-images-using-a-generative-model - GenSmoke-GS offers a multi-stage solution for improving visibility and coherence in smoke-degraded image processing for enhanced 3D reconstructions.
- ARM: Advantage Reward Modeling for Long-Horizon Manipulation (viability: 7): https://sciencetostartup.com/paper/arm-advantage-reward-modeling-for-long-horizon-manipulation - A novel reward modeling framework for robotics that uses cost-effective human feedback to significantly improve long-horizon manipulation task success rates.
- Beyond Isolated Tasks: A Framework for Evaluating Coding Agents on Sequential Software Evolution (viability: 7): https://sciencetostartup.com/paper/beyond-isolated-tasks-a-framework-for-evaluating-coding-agents-on-sequential-software-evolution - A framework for evaluating coding agents on sequential software evolution tasks, revealing performance gaps and degradation in repository health compared to isolated task evaluations.
- Behavior-Constrained Reinforcement Learning with Receding-Horizon Credit Assignment for High-Performance Control (viability: 7): https://sciencetostartup.com/paper/behavior-constrained-reinforcement-learning-with-receding-horizon-credit-assignment-for-high-performance-control - A behavior-constrained reinforcement learning framework that learns high-performance control policies for robotics, maintaining expert-like behavior and outperforming baselines in simulation and human-in-the-loop evaluations.
- Comparing the Impact of Pedagogy-Informed Custom and General-Purpose GAI Chatbots on Students' Science Problem-Solving Processes and Performance Using Heterogeneous Interaction Network Analysis (viability: 4): https://sciencetostartup.com/paper/comparing-the-impact-of-pedagogy-informed-custom-and-general-purpose-gai-chatbots-on-students-science-problem-solving-pr - A pedagogy-informed custom AI chatbot designed to enhance student science problem-solving by fostering cognitive engagement over direct answers.
- Agentic-MME: What Agentic Capability Really Brings to Multimodal Intelligence? (viability: 7): https://sciencetostartup.com/paper/agentic-mme-what-agentic-capability-really-brings-to-multimodal-intelligence - A new benchmark and evaluation framework for multimodal AI agents that verifies tool invocation and efficiency, revealing significant gaps in current state-of-the-art models.
- Generating DDPM-based Samples from Tilted Distributions (viability: 2): https://sciencetostartup.com/paper/generating-ddpm-based-samples-from-tilted-distributions - This paper presents a theoretical framework for generating diffusion-based samples from tilted distributions.
- User-Aware Conditional Generative Total Correlation Learning for Multi-Modal Recommendation (viability: 7): https://sciencetostartup.com/paper/user-aware-conditional-generative-total-correlation-learning-for-multi-modal-recommendation - A generative total correlation learning framework for multi-modal recommendation that personalizes content feature filtering to user preferences, outperforming state-of-the-art by up to 28.30% in NDCG@5.
- Asymptotically-Bounded 3D Frontier Exploration enhanced with Bayesian Information Gain (viability: 4): https://sciencetostartup.com/paper/asymptotically-bounded-3d-frontier-exploration-enhanced-with-bayesian-information-gain - A more efficient 3D robotic exploration algorithm that reduces computational demands by focusing on frontiers and using Bayesian information gain for viewpoint prioritization.
- A Flow Matching Framework for Soft-Robot Inverse Dynamics (viability: 7): https://sciencetostartup.com/paper/a-flow-matching-framework-for-soft-robot-inverse-dynamics - A flow matching framework that significantly improves inverse dynamics learning for soft robots, enabling stable open-loop control with sub-millisecond inference.
- R2-Write: Reflection and Revision for Open-Ended Writing with Deep Reasoning (viability: 7): https://sciencetostartup.com/paper/r2-write-reflection-and-revision-for-open-ended-writing-with-deep-reasoning - An automated framework that enhances large language models for open-ended writing by incorporating explicit reflection and revision patterns.
- Explicit Time-Frequency Dynamics for Skeleton-Based Gait Recognition (viability: 5): https://sciencetostartup.com/paper/explicit-time-frequency-dynamics-for-skeleton-based-gait-recognition - Enhance skeleton-based gait recognition by adding explicit time-frequency dynamics to existing models, improving performance under challenging conditions.
- Rendering Multi-Human and Multi-Object with 3D Gaussian Splatting (viability: 7): https://sciencetostartup.com/paper/rendering-multi-human-and-multi-object-with-3d-gaussian-splatting - A hierarchical framework for high-fidelity 3D reconstruction of dynamic scenes with multiple interacting humans and objects from sparse-view inputs.
- FedSQ: Optimized Weight Averaging via Fixed Gating (viability: 4): https://sciencetostartup.com/paper/fedsq-optimized-weight-averaging-via-fixed-gating - A federated learning method that stabilizes training by freezing structural components of pretrained models to improve robustness and reduce convergence time.
- Self-Optimizing Multi-Agent Systems for Deep Research (viability: 3): https://sciencetostartup.com/paper/self-optimizing-multi-agent-systems-for-deep-research - Self-optimizing multi-agent systems for deep research that adapt and improve their prompt strategies.
- Mitigating Reward Hacking in RLHF via Advantage Sign Robustness (viability: 7): https://sciencetostartup.com/paper/mitigating-reward-hacking-in-rlhf-via-advantage-sign-robustness - A novel RLHF technique that mitigates reward hacking by ensuring the robustness of advantage signs during policy optimization, leading to improved performance on summarization and instruction-following tasks.
- Prompt Compression in the Wild: Measuring Latency, Rate Adherence, and Quality for Faster LLM Inference (viability: 5): https://sciencetostartup.com/paper/prompt-compression-in-the-wild-measuring-latency-rate-adherence-and-quality-for-faster-llm-inference - Accelerate LLM inference by intelligently compressing prompts, offering significant speed-ups and reduced memory usage with a predictive profiler.
- Not All Frames Deserve Full Computation: Accelerating Autoregressive Video Generation via Selective Computation and Predictive Extrapolation (viability: 5): https://sciencetostartup.com/paper/not-all-frames-deserve-full-computation-accelerating-autoregressive-video-generation-via-selective-computation-and-predi - Accelerate autoregressive video generation by intelligently reusing computations and predicting future frames, achieving significant speedups without retraining.
- Effect of Input Resolution on Retinal Vessel Segmentation Performance: An Empirical Study Across Five Datasets (viability: 4): https://sciencetostartup.com/paper/effect-of-input-resolution-on-retinal-vessel-segmentation-performance-an-empirical-study-across-five-datasets - This research empirically studies the impact of image resolution on retinal vessel segmentation, revealing a critical trade-off for thin vessel detection that standard metrics overlook.
- Exploring Motion-Language Alignment for Text-driven Motion Generation (viability: 7): https://sciencetostartup.com/paper/exploring-motion-language-alignment-for-text-driven-motion-generation - A framework for generating realistic human motion from text by improving motion-language alignment and addressing attention sink issues.
- NeuReasoner: Towards Explainable, Controllable, and Unified Reasoning via Mixture-of-Neurons (viability: 7): https://sciencetostartup.com/paper/neureasoner-towards-explainable-controllable-and-unified-reasoning-via-mixture-of-neurons - A framework that makes large reasoning models explainable and controllable by identifying and correcting specific neuron-level failure modes, improving performance and reducing costs.
- InfoSeeker: A Scalable Hierarchical Parallel Agent Framework for Web Information Seeking (viability: 7): https://sciencetostartup.com/paper/infoseeker-a-scalable-hierarchical-parallel-agent-framework-for-web-information-seeking - A hierarchical agent framework that accelerates web information seeking by parallelizing evidence aggregation and reflection.
- Inversion-Free Natural Gradient Descent on Riemannian Manifolds (viability: 4): https://sciencetostartup.com/paper/inversion-free-natural-gradient-descent-on-riemannian-manifolds - An inversion-free natural gradient descent method for optimizing probability distributions on Riemannian manifolds, offering improved convergence and constraint enforcement for applications like variational Bayes and normalizing flows.
- FoE: Forest of Errors Makes the First Solution the Best in Large Reasoning Models (viability: 7): https://sciencetostartup.com/paper/foe-forest-of-errors-makes-the-first-solution-the-best-in-large-reasoning-models - A novel framework that significantly improves the accuracy and efficiency of large reasoning models by addressing the 'Forest of Errors' phenomenon.
- Visual Prototype Conditioned Focal Region Generation for UAV-Based Object Detection (viability: 7): https://sciencetostartup.com/paper/visual-prototype-conditioned-focal-region-generation-for-uav-based-object-detection - Generate high-fidelity synthetic data for UAV object detection to overcome limited training data and improve accuracy.
- Open-Loop Planning, Closed-Loop Verification: Speculative Verification for VLA (viability: 7): https://sciencetostartup.com/paper/open-loop-planning-closed-loop-verification-speculative-verification-for-vla - A framework for efficient and robust embodied AI control that combines speculative open-loop planning with lightweight closed-loop verification.
- Collaborative Multi-Mode Pruning for Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/collaborative-multi-mode-pruning-for-vision-language-models - A novel framework for joint parameter and token pruning in Vision-Language Models to enable deployment on resource-constrained devices.
- LogicPoison: Logical Attacks on Graph Retrieval-Augmented Generation (viability: 7): https://sciencetostartup.com/paper/logicpoison-logical-attacks-on-graph-retrieval-augmented-generation - An AI security tool to defend against logic-based attacks on Graph-based Retrieval Augmentation systems.
- How Annotation Trains Annotators: Competence Development in Social Influence Recognition (viability: 4): https://sciencetostartup.com/paper/how-annotation-trains-annotators-competence-development-in-social-influence-recognition - This research explores how social influence impacts annotator competence, leading to improved data quality and LLM performance, with potential applications in optimizing annotation pipelines.
- CrossWeaver: Cross-modal Weaving for Arbitrary-Modality Semantic Segmentation (viability: 7): https://sciencetostartup.com/paper/crossweaver-cross-modal-weaving-for-arbitrary-modality-semantic-segmentation - A flexible multimodal fusion framework for semantic segmentation that achieves state-of-the-art performance across diverse sensor combinations.
- AgentHazard: A Benchmark for Evaluating Harmful Behavior in Computer-Use Agents (viability: 7): https://sciencetostartup.com/paper/agenthazard-a-benchmark-for-evaluating-harmful-behavior-in-computer-use-agents - A benchmark to evaluate and mitigate harmful behavior in autonomous computer-use agents, revealing significant vulnerabilities in current systems.
- Learning from Synthetic Data via Provenance-Based Input Gradient Guidance (viability: 4): https://sciencetostartup.com/paper/learning-from-synthetic-data-via-provenance-based-input-gradient-guidance - A new framework for training computer vision models with synthetic data that uses provenance information to guide learning towards relevant input regions, improving robustness and reducing reliance on synthesis artifacts.
- Explainable Machine Learning Reveals 12-Fold Ucp1 Upregulation and Thermogenic Reprogramming in Female Mouse White Adipose Tissue After 37 Days of Microgravity: First AI/ML Analysis of NASA OSD-970 (viability: 5): https://sciencetostartup.com/paper/explainable-machine-learning-reveals-12-fold-ucp1-upregulation-and-thermogenic-reprogramming-in-female-mouse-white-adipo - Leveraging explainable AI to uncover microgravity's impact on thermogenesis in female white adipose tissue, with implications for astronaut health and metabolic disease research.
- MMTalker: Multiresolution 3D Talking Head Synthesis with Multimodal Feature Fusion (viability: 7): https://sciencetostartup.com/paper/mmtalker-multiresolution-3d-talking-head-synthesis-with-multimodal-feature-fusion - A novel method for synthesizing realistic 3D talking heads from speech by fusing multimodal features and using multiresolution representations, outperforming state-of-the-art in synchronization accuracy.
- Modality-Specific Hierarchical Enhancement for RGB-D Camouflaged Object Detection (viability: 7): https://sciencetostartup.com/paper/modality-specific-hierarchical-enhancement-for-rgb-d-camouflaged-object-detection - A novel framework for camouflaged object detection that enhances RGB and depth features independently before adaptive fusion, outperforming existing methods.
- PolyReal: A Benchmark for Real-World Polymer Science Workflows (viability: 4): https://sciencetostartup.com/paper/polyreal-a-benchmark-for-real-world-polymer-science-workflows - A new benchmark for evaluating multimodal LLMs on real-world polymer science workflows, revealing significant gaps in practical application capabilities.
- BEVPredFormer: Spatio-temporal Attention for BEV Instance Prediction in Autonomous Driving (viability: 7): https://sciencetostartup.com/paper/bevpredformer-spatio-temporal-attention-for-bev-instance-prediction-in-autonomous-driving - A camera-only architecture for autonomous driving that uses spatio-temporal attention to predict the behavior of surrounding obstacles with state-of-the-art performance.
- Towards Near-Real-Time Telemetry-Aware Routing with Neural Routing Algorithms (viability: 5): https://sciencetostartup.com/paper/towards-near-real-time-telemetry-aware-routing-with-neural-routing-algorithms - A framework for training and evaluating neural routing algorithms that explicitly model communication and inference delays, outperforming traditional methods in realistic network conditions.
- A Multi-head-based architecture for effective morphological tagging in Russian with open dictionary (viability: 7): https://sciencetostartup.com/paper/a-multi-head-based-architecture-for-effective-morphological-tagging-in-russian-with-open-dictionary - A novel multi-head attention architecture for highly accurate Russian morphological tagging that supports open dictionaries and runs on consumer hardware.
- Council Mode: Mitigating Hallucination and Bias in LLMs via Multi-Agent Consensus (viability: 7): https://sciencetostartup.com/paper/council-mode-mitigating-hallucination-and-bias-in-llms-via-multi-agent-consensus - A multi-agent consensus framework that significantly reduces LLM hallucinations and biases by synthesizing outputs from diverse frontier models.
- Efficient Logistic Regression with Mixture of Sigmoids (viability: 3): https://sciencetostartup.com/paper/efficient-logistic-regression-with-mixture-of-sigmoids - This paper presents a computationally efficient algorithm for online logistic regression with theoretical guarantees, improving upon existing complexity bounds.
- GP-4DGS: Probabilistic 4D Gaussian Splatting from Monocular Video via Variational Gaussian Processes (viability: 3): https://sciencetostartup.com/paper/gp-4dgs-probabilistic-4d-gaussian-splatting-from-monocular-video-via-variational-gaussian-processes - A novel framework for probabilistic 4D scene reconstruction from monocular video, offering uncertainty quantification and motion estimation for unobserved regions.
- Split and Conquer Partial Deepfake Speech (viability: 7): https://sciencetostartup.com/paper/split-and-conquer-partial-deepfake-speech - A novel framework for detecting manipulated regions in partial deepfake speech through boundary detection and segment-level classification.
- Corporations Constitute Intelligence (viability: 2): https://sciencetostartup.com/paper/corporations-constitute-intelligence - This paper analyzes the legal and democratic shortcomings of corporate AI constitutions, arguing for the need for a democratic body to govern AI behavior.
- Learning Task-Invariant Properties via Dreamer: Enabling Efficient Policy Transfer for Quadruped Robots (viability: 7): https://sciencetostartup.com/paper/learning-task-invariant-properties-via-dreamer-enabling-efficient-policy-transfer-for-quadruped-robots - A framework for quadruped robots that learns task-invariant properties to achieve robust and efficient sim-to-real transfer, significantly outperforming existing methods.
- Analysis of Optimality of Large Language Models on Planning Problems (viability: 4): https://sciencetostartup.com/paper/analysis-of-optimality-of-large-language-models-on-planning-problems - LLMs demonstrate near-perfect optimality in complex planning problems, outperforming traditional planners by leveraging algorithmic simulation and geometric memory.
- SentiAvatar: Towards Expressive and Interactive Digital Humans (viability: 8): https://sciencetostartup.com/paper/sentiavatar-towards-expressive-and-interactive-digital-humans - Create expressive 3D digital avatars for real-time interactive applications using SentiAvatar's framework.
- UniSpector: Towards Universal Open-set Defect Recognition via Spectral-Contrastive Visual Prompting (viability: 7): https://sciencetostartup.com/paper/unispector-towards-universal-open-set-defect-recognition-via-spectral-contrastive-visual-prompting - A novel visual prompting method for universal open-set defect recognition in industrial inspection, outperforming baselines significantly.
- BioUNER: A Benchmark Dataset for Clinical Urdu Named Entity Recognition (viability: 5): https://sciencetostartup.com/paper/biouner-a-benchmark-dataset-for-clinical-urdu-named-entity-recognition - A benchmark dataset for clinical Urdu Named Entity Recognition to advance NLP capabilities in under-resourced languages.
- RayMamba: Ray-Aligned Serialization for Long-Range 3D Object Detection (viability: 7): https://sciencetostartup.com/paper/raymamba-ray-aligned-serialization-for-long-range-3d-object-detection - RayMamba enhances 3D object detection by organizing sparse LiDAR data into geometry-aware sequences, significantly improving performance in challenging long-range scenarios.
- Extracting Money Laundering Transactions from Quasi-Temporal Graph Representation (viability: 7): https://sciencetostartup.com/paper/extracting-money-laundering-transactions-from-quasi-temporal-graph-representation - A simple and scalable supervised learning framework to detect money laundering transactions, outperforming state-of-the-art models and complementing existing AML systems.
- EvaNet: Towards More Efficient and Consistent Infrared and Visible Image Fusion Assessment (viability: 7): https://sciencetostartup.com/paper/evanet-towards-more-efficient-and-consistent-infrared-and-visible-image-fusion-assessment - A lightweight, AI-powered framework for significantly faster and more consistent evaluation of infrared and visible image fusion.
- Toward an Artificial General Teacher: Procedural Geometry Data Generation and Visual Grounding with Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/toward-an-artificial-general-teacher-procedural-geometry-data-generation-and-visual-grounding-with-vision-language-model - Generate synthetic geometry diagrams and explanations for AI tutors, overcoming domain shift issues with a novel data engine and fine-tuned vision-language models.
- RAGE: A Tightly Coupled Radar-Aided Grip Estimator For Autonomous Race Cars (viability: 7): https://sciencetostartup.com/paper/rage-a-tightly-coupled-radar-aided-grip-estimator-for-autonomous-race-cars - A real-time friction estimator for autonomous race cars using standard sensors, enabling safer and more effective operation at physical limits.
- Progressive Video Condensation with MLLM Agent for Long-form Video Understanding (viability: 5): https://sciencetostartup.com/paper/progressive-video-condensation-with-mllm-agent-for-long-form-video-understanding - An agent that progressively condenses long videos into keyframes for efficient multimodal LLM reasoning, achieving state-of-the-art accuracy with reduced computational cost.
- Rethinking Forward Processes for Score-Based Data Assimilation in High Dimensions (viability: 4): https://sciencetostartup.com/paper/rethinking-forward-processes-for-score-based-data-assimilation-in-high-dimensions - A novel measurement-aware score-based filter for more accurate and stable data assimilation in high-dimensional systems.
- Lipschitz bounds for integral kernels (viability: 2): https://sciencetostartup.com/paper/lipschitz-bounds-for-integral-kernels - This paper theoretically characterizes the Lipschitz continuity of feature maps for integral kernels, offering insights into the robustness of kernel methods and neural networks.
- High-dimensional Many-to-many-to-many Mediation Analysis (viability: 7): https://sciencetostartup.com/paper/high-dimensional-many-to-many-to-many-mediation-analysis - A statistical framework for high-dimensional mediation analysis to uncover complex genetic-neural-cognitive pathways and improve predictive performance in areas like Alzheimer's research.
- Information-Regularized Constrained Inversion for Stable Avatar Editing from Sparse Supervision (viability: 4): https://sciencetostartup.com/paper/information-regularized-constrained-inversion-for-stable-avatar-editing-from-sparse-supervision - A framework for stable editing of animatable human avatars from sparse supervision by performing constrained inversion in a structured latent space.
- One Model to Translate Them All? A Journey to Mount Doom for Multilingual Model Merging (viability: 3): https://sciencetostartup.com/paper/one-model-to-translate-them-all-a-journey-to-mount-doom-for-multilingual-model-merging - This research explains why merging fine-tuned language models fails for multilingual translation, identifying representational divergence as the cause.
- InstructTable: Improving Table Structure Recognition Through Instructions (viability: 7): https://sciencetostartup.com/paper/instructtable-improving-table-structure-recognition-through-instructions - InstructTable improves table structure recognition for complex layouts by combining instruction-guided pre-training with visual modeling and synthetic data generation.
- An Asynchronous Two-Speed Kalman Filter for Real-Time UUV Cooperative Navigation Under Acoustic Delays (viability: 5): https://sciencetostartup.com/paper/an-asynchronous-two-speed-kalman-filter-for-real-time-uuv-cooperative-navigation-under-acoustic-delays - A novel asynchronous Kalman filter with variational history distillation enables real-time cooperative navigation for underwater vehicles despite significant acoustic communication delays.
- Toward an Operational GNN-Based Multimesh Surrogate for Fast Flood Forecasting (viability: 7): https://sciencetostartup.com/paper/toward-an-operational-gnn-based-multimesh-surrogate-for-fast-flood-forecasting - Develops a graph-neural network surrogate to accelerate flood forecasting simulations by orders of magnitude, enabling rapid decision support for urban floodplains.
- Unlocking Positive Transfer in Incrementally Learning Surgical Instruments: A Self-reflection Hierarchical Prompt Framework (viability: 7): https://sciencetostartup.com/paper/unlocking-positive-transfer-in-incrementally-learning-surgical-instruments-a-self-reflection-hierarchical-prompt-framewo - A framework that improves surgical instrument segmentation by enabling models to learn from past and future knowledge, reducing catastrophic forgetting.
- SPG: Sparse-Projected Guides with Sparse Autoencoders for Zero-Shot Anomaly Detection (viability: 7): https://sciencetostartup.com/paper/spg-sparse-projected-guides-with-sparse-autoencoders-for-zero-shot-anomaly-detection - A prompt-free framework for zero-shot anomaly detection and segmentation that leverages sparse autoencoders to generate anomaly guides, achieving state-of-the-art pixel-level segmentation.
- Token Warping Helps MLLMs Look from Nearby Viewpoints (viability: 7): https://sciencetostartup.com/paper/token-warping-helps-mllms-look-from-nearby-viewpoints - A novel token warping technique for multimodal LLMs to improve viewpoint robustness, outperforming existing methods on a new benchmark.
- Multi-Turn Reinforcement Learning for Tool-Calling Agents with Iterative Reward Calibration (viability: 7): https://sciencetostartup.com/paper/multi-turn-reinforcement-learning-for-tool-calling-agents-with-iterative-reward-calibration - A novel reinforcement learning approach for training highly capable tool-calling agents that outperform larger models on complex multi-turn tasks, with code and training recipes released.
- HairOrbit: Multi-view Aware 3D Hair Modeling from Single Portraits (viability: 7): https://sciencetostartup.com/paper/hairorbit-multi-view-aware-3d-hair-modeling-from-single-portraits - Generate detailed 3D hair models from single portraits by leveraging video generation priors and a novel strand-growing algorithm.
- LLM-based Atomic Propositions help weak extractors: Evaluation of a Propositioner for triplet extraction (viability: 3): https://sciencetostartup.com/paper/llm-based-atomic-propositions-help-weak-extractors-evaluation-of-a-propositioner-for-triplet-extraction - A model that enhances triplet extraction from text using atomic propositions.
- EMS: Multi-Agent Voting via Efficient Majority-then-Stopping (viability: 7): https://sciencetostartup.com/paper/ems-multi-agent-voting-via-efficient-majority-then-stopping - An efficient multi-agent voting system that reduces computational overhead by stopping early once a majority consensus is reached, based on agent reliability.
- LLM+Graph@VLDB'2025 Workshop Summary (viability: 2): https://sciencetostartup.com/paper/llm-graph-vldb-2025-workshop-summary - This workshop report summarizes research directions at the intersection of LLMs and graph data management, highlighting challenges and solutions for practical applications.
- A Paradigm Shift: Fully End-to-End Training for Temporal Sentence Grounding in Videos (viability: 7): https://sciencetostartup.com/paper/a-paradigm-shift-fully-end-to-end-training-for-temporal-sentence-grounding-in-videos - A novel end-to-end training paradigm for video sentence grounding that significantly improves localization accuracy by adapting video backbones to linguistic queries.
- Frame Theoretical Derivation of Three Factor Learning Rule for Oja's Subspace Rule (viability: 1): https://sciencetostartup.com/paper/frame-theoretical-derivation-of-three-factor-learning-rule-for-oja-s-subspace-rule - This paper provides a theoretical derivation of a learning rule for principal component analysis using frame theory, offering a new mathematical perspective on biologically plausible learning.
- HiDiGen: Hierarchical Diffusion for B-Rep Generation with Explicit Topological Constraints (viability: 4): https://sciencetostartup.com/paper/hidigen-hierarchical-diffusion-for-b-rep-generation-with-explicit-topological-constraints - A hierarchical diffusion model for generating topologically valid 3D CAD models by decoupling geometry and topology.
- Adaptive Local Frequency Filtering for Fourier-Encoded Implicit Neural Representations (viability: 7): https://sciencetostartup.com/paper/adaptive-local-frequency-filtering-for-fourier-encoded-implicit-neural-representations - Enhance implicit neural representations for modeling complex signals with spatially varying frequencies, improving reconstruction quality and optimization speed.
- Deformation-based In-Context Learning for Point Cloud Understanding (viability: 7): https://sciencetostartup.com/paper/deformation-based-in-context-learning-for-point-cloud-understanding - A deformation-based framework for point cloud in-context learning that explicitly reasons about geometry and outperforms state-of-the-art on reconstruction, denoising, and registration tasks.
- Towards Secure Agent Skills: Architecture, Threat Taxonomy, and Security Analysis (viability: 4): https://sciencetostartup.com/paper/towards-secure-agent-skills-architecture-threat-taxonomy-and-security-analysis - This paper analyzes security vulnerabilities in the Agent Skills framework for LLM agents, proposing a threat taxonomy and defense strategies.
- Factorized Multi-Resolution HashGrid for Efficient Neural Radiance Fields: Execution on Edge-Devices (viability: 7): https://sciencetostartup.com/paper/factorized-multi-resolution-hashgrid-for-efficient-neural-radiance-fields-execution-on-edge-devices - A novel parameter encoding method for efficient on-device neural radiance field training, significantly reducing memory usage while maintaining quality and speed.
- ESL-Bench: An Event-Driven Synthetic Longitudinal Benchmark for Health Agents (viability: 7): https://sciencetostartup.com/paper/esl-bench-an-event-driven-synthetic-longitudinal-benchmark-for-health-agents - A benchmark and synthesis framework for evaluating longitudinal health agents, demonstrating significant performance gaps between DB agents and RAG baselines on complex reasoning tasks.
- GRADE: Probing Knowledge Gaps in LLMs through Gradient Subspace Dynamics (viability: 7): https://sciencetostartup.com/paper/grade-probing-knowledge-gaps-in-llms-through-gradient-subspace-dynamics - A novel method to detect and explain knowledge gaps in LLMs by analyzing gradient dynamics, enabling more responsible deployment.
- STRNet: Visual Navigation with Spatio-Temporal Representation through Dynamic Graph Aggregation (viability: 7): https://sciencetostartup.com/paper/strnet-visual-navigation-with-spatio-temporal-representation-through-dynamic-graph-aggregation - A unified spatio-temporal representation framework to enhance visual encoding for more accurate robotic navigation.
- NavCrafter: Exploring 3D Scenes from a Single Image (viability: 7): https://sciencetostartup.com/paper/navcrafter-exploring-3d-scenes-from-a-single-image - Generate explorable 3D scenes from a single image with controllable camera movement and high fidelity.
- Orientation Matters: Learning Radiation Patterns of Multi-Rotor UAVs In-Flight to Enhance Communication Availability Modeling (viability: 7): https://sciencetostartup.com/paper/orientation-matters-learning-radiation-patterns-of-multi-rotor-uavs-in-flight-to-enhance-communication-availability-mode - Learn and adapt UAV communication radiation patterns in real-time to ensure reliable connectivity for autonomous drone operations.
- Goal-Conditioned Neural ODEs with Guaranteed Safety and Stability for Learning-Based All-Pairs Motion Planning (viability: 7): https://sciencetostartup.com/paper/goal-conditioned-neural-odes-with-guaranteed-safety-and-stability-for-learning-based-all-pairs-motion-planning - A learning-based motion planner using neural ODEs with theoretical guarantees for safe and stable navigation.
- MFE: A Multimodal Hand Exoskeleton with Interactive Force, Pressure and Thermo-haptic Feedback (viability: 7): https://sciencetostartup.com/paper/mfe-a-multimodal-hand-exoskeleton-with-interactive-force-pressure-and-thermo-haptic-feedback - A multimodal hand exoskeleton providing rich force, pressure, and thermal haptic feedback for enhanced robotic teleoperation and VR experiences.
- Student-in-the-Loop Chain-of-Thought Distillation via Generation-Time Selection (viability: 7): https://sciencetostartup.com/paper/student-in-the-loop-chain-of-thought-distillation-via-generation-time-selection - A framework that distills complex reasoning from large language models to smaller ones by selecting only the most learnable reasoning paths during generation.
- MMPhysVideo: Scaling Physical Plausibility in Video Generation via Joint Multimodal Modeling (viability: 7): https://sciencetostartup.com/paper/mmphysvideo-scaling-physical-plausibility-in-video-generation-via-joint-multimodal-modeling - A framework for generating physically plausible videos by unifying semantic, geometric, and temporal cues into a pseudo-RGB format, improving visual quality and physical consistency.
- QAPruner: Quantization-Aware Vision Token Pruning for Multimodal Large Language Models (viability: 4): https://sciencetostartup.com/paper/qapruner-quantization-aware-vision-token-pruning-for-multimodal-large-language-models - A framework for co-optimizing vision token pruning and quantization to enable efficient deployment of multimodal LLMs.
- Learning Structured Robot Policies from Vision-Language Models via Synthetic Neuro-Symbolic Supervision (viability: 7): https://sciencetostartup.com/paper/learning-structured-robot-policies-from-vision-language-models-via-synthetic-neuro-symbolic-supervision - Generate interpretable and structured robot policies from vision-language models using synthetic data, enabling safer and more reliable robotic applications.
- ChatSVA: Bridging SVA Generation for Hardware Verification via Task-Specific LLMs (viability: 9): https://sciencetostartup.com/paper/chatsva-bridging-sva-generation-for-hardware-verification-via-task-specific-llms - ChatSVA automates SystemVerilog Assertions for hardware verification, improving speed and accuracy with task-specific LLMs.
- CMCC-ReID: Cross-Modality Clothing-Change Person Re-Identification (viability: 7): https://sciencetostartup.com/paper/cmcc-reid-cross-modality-clothing-change-person-re-identification - A novel network for person re-identification that handles both changes in clothing and camera modality, addressing a realistic surveillance challenge.
- PaveBench: A Versatile Benchmark for Pavement Distress Perception and Interactive Vision-Language Analysis (viability: 7): https://sciencetostartup.com/paper/pavebench-a-versatile-benchmark-for-pavement-distress-perception-and-interactive-vision-language-analysis - A benchmark and framework for interactive vision-language analysis of pavement distress, enabling quantitative assessment and maintenance reasoning.
- UNICA: A Unified Neural Framework for Controllable 3D Avatars (viability: 7): https://sciencetostartup.com/paper/unica-a-unified-neural-framework-for-controllable-3d-avatars - A unified neural framework for generating controllable 3D avatars from simple keyboard inputs, eliminating complex pipelines.
- Rubrics to Tokens: Bridging Response-level Rubrics and Token-level Rewards in Instruction Following Tasks (viability: 7): https://sciencetostartup.com/paper/rubrics-to-tokens-bridging-response-level-rubrics-and-token-level-rewards-in-instruction-following-tasks - A novel framework that bridges response-level scores and token-level credit assignment for more accurate LLM instruction following.
- CharTool: Tool-Integrated Visual Reasoning for Chart Understanding (viability: 7): https://sciencetostartup.com/paper/chartool-tool-integrated-visual-reasoning-for-chart-understanding - A multimodal LLM that uses tools for precise chart understanding and numerical reasoning, outperforming existing models on key benchmarks.
- Structure-Aware Commitment Reduction for Network-Constrained Unit Commitment with Solver-Preserving Guarantees (viability: 7): https://sciencetostartup.com/paper/structure-aware-commitment-reduction-for-network-constrained-unit-commitment-with-solver-preserving-guarantees - Leverage LLMs to intelligently reduce the complexity of network-constrained unit commitment problems, enabling faster and more optimal energy grid management.
- LumaFlux: Lifting 8-Bit Worlds to HDR Reality with Physically-Guided Diffusion Transformers (viability: 7): https://sciencetostartup.com/paper/lumaflux-lifting-8-bit-worlds-to-hdr-reality-with-physically-guided-diffusion-transformers - LumaFlux uses a physically-guided diffusion transformer to convert 8-bit SDR video to 10-bit HDR, outperforming existing methods with a new dataset and benchmark.
- QuadAgent: A Responsive Agent System for Vision-Language Guided Quadrotor Agile Flight (viability: 7): https://sciencetostartup.com/paper/quadagent-a-responsive-agent-system-for-vision-language-guided-quadrotor-agile-flight - A training-free agent system for vision-language guided agile quadrotor flight that decouples reasoning and control for improved efficiency and responsiveness.
- CANDLE: Illumination-Invariant Semantic Priors for Color Ambient Lighting Normalization (viability: 7): https://sciencetostartup.com/paper/candle-illumination-invariant-semantic-priors-for-color-ambient-lighting-normalization - A novel method for color ambient lighting normalization using self-supervised features to recover intrinsic object color, achieving state-of-the-art results in challenging conditions.
- EnsemHalDet: Robust VLM Hallucination Detection via Ensemble of Internal State Detectors (viability: 7): https://sciencetostartup.com/paper/ensemhaldet-robust-vlm-hallucination-detection-via-ensemble-of-internal-state-detectors - A framework that significantly improves the accuracy of detecting factual errors in Vision-Language Models by ensembling multiple internal state detectors.
- Disrupting Cognitive Passivity: Rethinking AI-Assisted Data Literacy through Cognitive Alignment (viability: 2): https://sciencetostartup.com/paper/disrupting-cognitive-passivity-rethinking-ai-assisted-data-literacy-through-cognitive-alignment - A framework for human-AI interaction that aligns AI's response mode with user cognitive demand to foster data literacy and prevent cognitive passivity.
- A Unified Perspective on Adversarial Membership Manipulation in Vision Models (viability: 7): https://sciencetostartup.com/paper/a-unified-perspective-on-adversarial-membership-manipulation-in-vision-models - A novel framework to detect and mitigate adversarial attacks that manipulate AI model privacy by making non-training data appear as if it was part of the training set.
- Vision-Based End-to-End Learning for UAV Traversal of Irregular Gaps via Differentiable Simulation (viability: 7): https://sciencetostartup.com/paper/vision-based-end-to-end-learning-for-uav-traversal-of-irregular-gaps-via-differentiable-simulation - A vision-based end-to-end framework for autonomous drones to navigate complex, irregular gaps, enhancing inspection and rescue operations.
- When Modalities Remember: Continual Learning for Multimodal Knowledge Graphs (viability: 7): https://sciencetostartup.com/paper/when-modalities-remember-continual-learning-for-multimodal-knowledge-graphs - A novel model for continual learning in multimodal knowledge graphs that prevents catastrophic forgetting and enhances new knowledge acquisition.
- Open Challenges for Secure and Scalable Wi-Fi Connectivity in Rural Areas (viability: 4): https://sciencetostartup.com/paper/open-challenges-for-secure-and-scalable-wi-fi-connectivity-in-rural-areas - Securing pay-for-use Wi-Fi hotspots in rural areas by addressing hijacking and rogue hotspot vulnerabilities.
- Generalized Small Object Detection:A Point-Prompted Paradigm and Benchmark (viability: 7): https://sciencetostartup.com/paper/generalized-small-object-detection-a-point-prompted-paradigm-and-benchmark - A new paradigm for small object detection that uses point prompts at inference time to significantly improve accuracy and generalize to unseen objects and datasets.
- Multiple-Debias: A Full-process Debiasing Method for Multilingual Pre-trained Language Models (viability: 4): https://sciencetostartup.com/paper/multiple-debias-a-full-process-debiasing-method-for-multilingual-pre-trained-language-models - A method to reduce biases in multilingual language models across multiple sensitive attributes and languages.
- ContractShield: Bridging Semantic-Structural Gaps via Hierarchical Cross-Modal Fusion for Multi-Label Vulnerability Detection in Obfuscated Smart Contracts (viability: 7): https://sciencetostartup.com/paper/contractshield-bridging-semantic-structural-gaps-via-hierarchical-cross-modal-fusion-for-multi-label-vulnerability-detec - ContractShield is a multimodal framework that uses hierarchical cross-modal fusion to detect vulnerabilities in obfuscated smart contracts, outperforming state-of-the-art by 6-15%.
- Improving Role Consistency in Multi-Agent Collaboration via Quantitative Role Clarity (viability: 3): https://sciencetostartup.com/paper/improving-role-consistency-in-multi-agent-collaboration-via-quantitative-role-clarity - This paper proposes a method to enhance role consistency in multi-agent systems using a quantitative role clarity approach.
- SentinelAgent: Intent-Verified Delegation Chains for Securing Federal Multi-Agent AI Systems (viability: 7): https://sciencetostartup.com/paper/sentinelagent-intent-verified-delegation-chains-for-securing-federal-multi-agent-ai-systems - SentinelAgent provides a formal framework and runtime protocol for verifiable delegation chains in multi-agent AI systems, ensuring policy compliance and forensic traceability.
- Random Is Hard to Beat: Active Selection in online DPO with Modern LLMs (viability: 3): https://sciencetostartup.com/paper/random-is-hard-to-beat-active-selection-in-online-dpo-with-modern-llms - This research investigates active preference learning for LLMs, finding random sampling to be a surprisingly effective baseline and questioning the value of complex active selection strategies.
- Towards Realistic Class-Incremental Learning with Free-Flow Increments (viability: 3): https://sciencetostartup.com/paper/towards-realistic-class-incremental-learning-with-free-flow-increments - A model-agnostic framework for robust class-incremental learning that handles variable class arrivals, stabilizing learning signals and improving performance.
- InverseDraping: Recovering Sewing Patterns from 3D Garment Surfaces via BoxMesh Bridging (viability: 7): https://sciencetostartup.com/paper/inversedraping-recovering-sewing-patterns-from-3d-garment-surfaces-via-boxmesh-bridging - Recovering precise 2D sewing patterns from 3D garment scans using a novel BoxMesh representation and a two-stage autoregressive model.
- OMNI-PoseX: A Fast Vision Model for 6D Object Pose Estimation in Embodied Tasks (viability: 8): https://sciencetostartup.com/paper/omni-posex-a-fast-vision-model-for-6d-object-pose-estimation-in-embodied-tasks - OMNI-PoseX provides real-time, accurate 6D object pose estimation for embodied robotic tasks, outperforming current solutions.
- STDDN: A Physics-Guided Deep Learning Framework for Crowd Simulation (viability: 7): https://sciencetostartup.com/paper/stddn-a-physics-guided-deep-learning-framework-for-crowd-simulation - A physics-guided deep learning framework for accurate and efficient crowd simulation, outperforming state-of-the-art with reduced latency.
- DeCo-DETR: Decoupled Cognition DETR for efficient Open-Vocabulary Object Detection (viability: 7): https://sciencetostartup.com/paper/deco-detr-decoupled-cognition-detr-for-efficient-open-vocabulary-object-detection - A vision-centric framework for efficient open-vocabulary object detection that decouples semantic reasoning from localization for practical deployment.
- Differentiable Stroke Planning with Dual Parameterization for Efficient and High-Fidelity Painting Creation (viability: 4): https://sciencetostartup.com/paper/differentiable-stroke-planning-with-dual-parameterization-for-efficient-and-high-fidelity-painting-creation - A novel dual parameterization for stroke-based rendering that improves structural coherence and efficiency.
- Understanding Latent Diffusability via Fisher Geometry (viability: 4): https://sciencetostartup.com/paper/understanding-latent-diffusability-via-fisher-geometry - A theoretical framework to diagnose and improve latent diffusion model performance by analyzing geometric properties of the latent space.
- Visual Instruction-Finetuned Language Model for Versatile Brain MR Image Tasks (viability: 7): https://sciencetostartup.com/paper/visual-instruction-finetuned-language-model-for-versatile-brain-mr-image-tasks - A versatile LLM for brain MRI that performs report generation, VQA, segmentation, and translation, outperforming specialized models.
- Geometrically-Constrained Radar-Inertial Odometry via Continuous Point-Pose Uncertainty Modeling (viability: 7): https://sciencetostartup.com/paper/geometrically-constrained-radar-inertial-odometry-via-continuous-point-pose-uncertainty-modeling - A geometrically-constrained radar-inertial odometry system that improves localization accuracy in challenging environments by dynamically modeling and integrating point and pose uncertainties.
- Learning Locomotion on Complex Terrain for Quadrupedal Robots with Foot Position Maps and Stability Rewards (viability: 5): https://sciencetostartup.com/paper/learning-locomotion-on-complex-terrain-for-quadrupedal-robots-with-foot-position-maps-and-stability-rewards - A new reinforcement learning approach for quadrupedal robots that uses foot position maps and stability rewards to achieve precise and stable locomotion on complex, unseen terrains.
- Cross Event Detection and Topic Evolution Mining in cross events for Man Made Disasters in Social Media Streams (viability: 4): https://sciencetostartup.com/paper/cross-event-detection-and-topic-evolution-mining-in-cross-events-for-man-made-disasters-in-social-media-streams - A framework for detecting and analyzing the evolution of related events in social media streams to understand their impact on human actions.
- Quotient-Based Posterior Analysis for Euclidean Latent Space Models (viability: 2): https://sciencetostartup.com/paper/quotient-based-posterior-analysis-for-euclidean-latent-space-models - A theoretical framework for analyzing latent space models in network analysis by providing canonical posterior summaries.
- State estimations and noise identifications with intermittent corrupted observations via Bayesian variational inference (viability: 3): https://sciencetostartup.com/paper/state-estimations-and-noise-identifications-with-intermittent-corrupted-observations-via-bayesian-variational-inference - A novel adaptive Kalman filter for state estimation in sensor networks with intermittent packet dropouts and corrupted observations.
- THOM: Generating Physically Plausible Hand-Object Meshes From Text (viability: 7): https://sciencetostartup.com/paper/thom-generating-physically-plausible-hand-object-meshes-from-text - THOM generates photorealistic, physically plausible 3D hand-object interactions from text, enhancing VR/AR experiences.
- Aligning Progress and Feasibility: A Neuro-Symbolic Dual Memory Framework for Long-Horizon LLM Agents (viability: 7): https://sciencetostartup.com/paper/aligning-progress-and-feasibility-a-neuro-symbolic-dual-memory-framework-for-long-horizon-llm-agents - A neuro-symbolic framework for LLM agents that decouples semantic guidance from logical validation to improve long-horizon decision-making.
- DeltaLogic: Minimal Premise Edits Reveal Belief-Revision Failures in Logical Reasoning Models (viability: 7): https://sciencetostartup.com/paper/deltalogic-minimal-premise-edits-reveal-belief-revision-failures-in-logical-reasoning-models - A new benchmark and evaluation protocol for assessing and improving belief revision capabilities in language models, crucial for dynamic environments.
- IndustryCode: A Benchmark for Industry Code Generation (viability: 7): https://sciencetostartup.com/paper/industrycode-a-benchmark-for-industry-code-generation - A new benchmark for evaluating LLM code generation across diverse industrial domains and programming languages, enabling more robust industrial AI solutions.
- GrandCode: Achieving Grandmaster Level in Competitive Programming via Agentic Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/grandcode-achieving-grandmaster-level-in-competitive-programming-via-agentic-reinforcement-learning - A multi-agent reinforcement learning system that consistently beats human grandmasters in competitive programming.
- MOMO: Mars Orbital Model Foundation Model for Mars Orbital Applications (viability: 7): https://sciencetostartup.com/paper/momo-mars-orbital-model-foundation-model-for-mars-orbital-applications - A multi-sensor foundation model for Mars orbital applications, outperforming existing baselines on downstream tasks.
- Generative Frontiers: Why Evaluation Matters for Diffusion Language Models (viability: 4): https://sciencetostartup.com/paper/generative-frontiers-why-evaluation-matters-for-diffusion-language-models - This research proposes a new principled method for evaluating diffusion language models, addressing limitations in current benchmarks and metrics to ensure reliable comparisons of generative quality.
- FluxMoE: Decoupling Expert Residency for High-Performance MoE Serving (viability: 7): https://sciencetostartup.com/paper/fluxmoe-decoupling-expert-residency-for-high-performance-moe-serving - A system that significantly boosts LLM serving throughput by intelligently managing expert weights to free up GPU memory for critical runtime data.
- ExploreVLA: Dense World Modeling and Exploration for End-to-End Autonomous Driving (viability: 8): https://sciencetostartup.com/paper/explorevla-dense-world-modeling-and-exploration-for-end-to-end-autonomous-driving - ExploreVLA integrates dense world modeling with RL for robust autonomous driving exploration.
- Breakdowns in Conversational AI: Interactional Failures in Emotionally and Ethically Sensitive Contexts (viability: 4): https://sciencetostartup.com/paper/breakdowns-in-conversational-ai-interactional-failures-in-emotionally-and-ethically-sensitive-contexts - This research identifies and categorizes breakdowns in conversational AI when dealing with emotional and ethical complexities, offering a path to more robust and aligned dialogue systems.
- V2X-QA: A Comprehensive Reasoning Dataset and Benchmark for Multimodal Large Language Models in Autonomous Driving Across Ego, Infrastructure, and Cooperative Views (viability: 7): https://sciencetostartup.com/paper/v2x-qa-a-comprehensive-reasoning-dataset-and-benchmark-for-multimodal-large-language-models-in-autonomous-driving-across - A new dataset and benchmark for multimodal LLMs in autonomous driving, along with a specialized MoE model, to improve reasoning across vehicle, infrastructure, and cooperative views.
- Evaluating the Formal Reasoning Capabilities of Large Language Models through Chomsky Hierarchy (viability: 4): https://sciencetostartup.com/paper/evaluating-the-formal-reasoning-capabilities-of-large-language-models-through-chomsky-hierarchy - A new benchmark reveals LLMs struggle with formal reasoning tasks, highlighting inefficiencies and the continued need for traditional algorithms.
- A Rapid Instrument Exchange System for Humanoid Robots in Minimally Invasive Surgery (viability: 5): https://sciencetostartup.com/paper/a-rapid-instrument-exchange-system-for-humanoid-robots-in-minimally-invasive-surgery - A teleoperated system enabling humanoid robots to rapidly and efficiently exchange surgical instruments, reducing complexity and cognitive load for surgeons.
- ALIVE-LIO: Degeneracy-Aware Learning of Inertial Velocity for Enhancing ESKF-Based LiDAR-Inertial Odometry (viability: 7): https://sciencetostartup.com/paper/alive-lio-degeneracy-aware-learning-of-inertial-velocity-for-enhancing-eskf-based-lidar-inertial-odometry - A degeneracy-aware LiDAR-inertial odometry framework that uses a neural network to predict velocity and improve state estimation in challenging environments.
- Trivial Vocabulary Bans Improve LLM Reasoning More Than Deep Linguistic Constraints (viability: 3): https://sciencetostartup.com/paper/trivial-vocabulary-bans-improve-llm-reasoning-more-than-deep-linguistic-constraints - This research demonstrates that simple vocabulary constraints, rather than complex linguistic manipulations, can improve LLM reasoning by acting as output regularizers.
- LieTrunc-QNN: Lie Algebra Truncation and Quantum Expressivity Phase Transition from LiePrune to Provably Stable Quantum Neural Networks (viability: 3): https://sciencetostartup.com/paper/lietrunc-qnn-lie-algebra-truncation-and-quantum-expressivity-phase-transition-from-lieprune-to-provably-stable-quantum-n - A theoretical framework for improving the trainability and stability of quantum neural networks using Lie algebra.
- VBGS-SLAM: Variational Bayesian Gaussian Splatting Simultaneous Localization and Mapping (viability: 7): https://sciencetostartup.com/paper/vbgs-slam-variational-bayesian-gaussian-splatting-simultaneous-localization-and-mapping - A probabilistic SLAM system using Gaussian Splatting that improves robustness and reduces drift by explicitly modeling uncertainty.
- XrayClaw: Cooperative-Competitive Multi-Agent Alignment for Trustworthy Chest X-ray Diagnosis (viability: 7): https://sciencetostartup.com/paper/xrayclaw-cooperative-competitive-multi-agent-alignment-for-trustworthy-chest-x-ray-diagnosis - A multi-agent AI system that improves the trustworthiness and accuracy of chest X-ray diagnoses by simulating a cooperative-competitive clinical workflow.
- DocShield: Towards AI Document Safety via Evidence-Grounded Agentic Reasoning (viability: 7): https://sciencetostartup.com/paper/docshield-towards-ai-document-safety-via-evidence-grounded-agentic-reasoning - DocShield is an AI-powered forensic tool for identifying and explaining text-centric document forgeries.
- Parser-Oriented Structural Refinement for a Stable Layout Interface in Document Parsing (viability: 5): https://sciencetostartup.com/paper/parser-oriented-structural-refinement-for-a-stable-layout-interface-in-document-parsing - A structural refinement module stabilizes document parsing pipelines by ensuring consistent input order for downstream parsers, significantly reducing errors on complex layouts.
- Adaptive Semantic Communication for Wireless Image Transmission Leveraging Mixture-of-Experts Mechanism (viability: 4): https://sciencetostartup.com/paper/adaptive-semantic-communication-for-wireless-image-transmission-leveraging-mixture-of-experts-mechanism - An adaptive semantic communication system for wireless image transmission that uses a Mixture-of-Experts mechanism to improve reconstruction quality and efficiency.
- Efficient3D: A Unified Framework for Adaptive and Debiased Token Reduction in 3D MLLMs (viability: 7): https://sciencetostartup.com/paper/efficient3d-a-unified-framework-for-adaptive-and-debiased-token-reduction-in-3d-mllms - A framework to significantly reduce inference costs for 3D Multimodal Large Language Models by adaptively pruning visual tokens, maintaining accuracy and enabling deployment on resource-constrained devices.
- Beyond Semantic Manipulation: Token-Space Attacks on Reward Models (viability: 4): https://sciencetostartup.com/paper/beyond-semantic-manipulation-token-space-attacks-on-reward-models - A novel attack framework that exploits reward models by directly manipulating token sequences, bypassing semantic understanding to achieve high reward scores with nonsensical outputs.
- Finding Belief Geometries with Sparse Autoencoders (viability: 4): https://sciencetostartup.com/paper/finding-belief-geometries-with-sparse-autoencoders - A pipeline to discover and validate belief-like geometric structures within large language model representations.
- Eligibility-Aware Evidence Synthesis: An Agentic Framework for Clinical Trial Meta-Analysis (viability: 7): https://sciencetostartup.com/paper/eligibility-aware-evidence-synthesis-an-agentic-framework-for-clinical-trial-meta-analysis - An agentic framework that automates clinical trial discovery and eligibility-aware meta-analysis for precision medicine evidence synthesis.
- Do Agent Societies Develop Intellectual Elites? The Hidden Power Laws of Collective Cognition in LLM Multi-Agent Systems (viability: 4): https://sciencetostartup.com/paper/do-agent-societies-develop-intellectual-elites-the-hidden-power-laws-of-collective-cognition-in-llm-multi-agent-systems - This research uncovers fundamental laws governing the collective intelligence of LLM multi-agent systems, identifying a key bottleneck and proposing a mechanism to improve their scalability and performance.
- Cross-subject Muscle Fatigue Detection via Adversarial and Supervised Contrastive Learning with Inception-Attention Network (viability: 7): https://sciencetostartup.com/paper/cross-subject-muscle-fatigue-detection-via-adversarial-and-supervised-contrastive-learning-with-inception-attention-netw - A novel neural network for robust cross-subject muscle fatigue detection using sEMG signals, enhancing physical rehabilitation.
- Redirected, Not Removed: Task-Dependent Stereotyping Reveals the Limits of LLM Alignments (viability: 7): https://sciencetostartup.com/paper/redirected-not-removed-task-dependent-stereotyping-reveals-the-limits-of-llm-alignments - This research provides a novel, comprehensive framework for auditing LLM bias that reveals systematic mischaracterizations by current alignment practices, offering a path to more robust safety and fairness evaluations.
- Too Polite to Disagree: Understanding Sycophancy Propagation in Multi-Agent Systems (viability: 4): https://sciencetostartup.com/paper/too-polite-to-disagree-understanding-sycophancy-propagation-in-multi-agent-systems - A method to mitigate sycophancy in multi-agent LLM discussions by providing peer sycophancy rankings, improving accuracy.
- Let's Have a Conversation: Designing and Evaluating LLM Agents for Interactive Optimization (viability: 7): https://sciencetostartup.com/paper/let-s-have-a-conversation-designing-and-evaluating-llm-agents-for-interactive-optimization - Develops a novel methodology and tailored LLM agents to significantly improve optimization solution quality through interactive conversations, bridging AI and operations research.
- A Numerical Method for Coupling Parameterized Physics-Informed Neural Networks and FDM for Advanced Thermal-Hydraulic System Simulation (viability: 4): https://sciencetostartup.com/paper/a-numerical-method-for-coupling-parameterized-physics-informed-neural-networks-and-fdm-for-advanced-thermal-hydraulic-sy - A novel hybrid AI-FDM method accelerates nuclear safety simulations by learning solution manifolds for parameterized physics-informed neural networks, eliminating retraining needs for varying problem parameters.
- SocioEval: A Template-Based Framework for Evaluating Socioeconomic Status Bias in Foundation Models (viability: 7): https://sciencetostartup.com/paper/socioeval-a-template-based-framework-for-evaluating-socioeconomic-status-bias-in-foundation-models - A framework for systematically evaluating and mitigating socioeconomic bias in foundation models, addressing a critical gap in responsible AI.
- Low-Rank Compression of Pretrained Models via Randomized Subspace Iteration (viability: 4): https://sciencetostartup.com/paper/low-rank-compression-of-pretrained-models-via-randomized-subspace-iteration - A new randomized subspace iteration method for efficient low-rank compression of large pretrained models that improves approximation quality and predictive accuracy.
- Transfer Learning for Meta-analysis Under Covariate Shift (viability: 4): https://sciencetostartup.com/paper/transfer-learning-for-meta-analysis-under-covariate-shift - A framework for more accurate meta-analysis by leveraging source trial outcomes as proxy signals and target trial placebo outcomes as gold labels to calibrate baseline risk.
- Drift-Resilient Temporal Priors for Visual Tracking (viability: 7): https://sciencetostartup.com/paper/drift-resilient-temporal-priors-for-visual-tracking - A lightweight module that significantly improves visual tracking performance by intelligently filtering noisy historical data and synthesizing dynamic temporal priors.
- Product-Stability: Provable Convergence for Gradient Descent on the Edge of Stability (viability: 3): https://sciencetostartup.com/paper/product-stability-provable-convergence-for-gradient-descent-on-the-edge-of-stability - This paper provides a theoretical framework for understanding gradient descent convergence at the edge of stability for a broader class of loss functions.
- Generalization Limits of Reinforcement Learning Alignment (viability: 4): https://sciencetostartup.com/paper/generalization-limits-of-reinforcement-learning-alignment - Develops novel 'compound jailbreak' techniques to expose generalization failures in LLM safety alignment, demonstrating a significant increase in attack success rates.
- Communication-free Sampling and 4D Hybrid Parallelism for Scalable Mini-batch GNN Training (viability: 3): https://sciencetostartup.com/paper/communication-free-sampling-and-4d-hybrid-parallelism-for-scalable-mini-batch-gnn-training - A 4D parallel framework for scalable mini-batch GNN training that significantly speeds up distributed learning on large graphs.
- Revealing the Learning Dynamics of Long-Context Continual Pre-training (viability: 4): https://sciencetostartup.com/paper/revealing-the-learning-dynamics-of-long-context-continual-pre-training - A framework for monitoring and evaluating the learning dynamics of large-scale continual pre-training for industrial LLMs, revealing insights into data scaling and training stability.
- GBQA: A Game Benchmark for Evaluating LLMs as Quality Assurance Engineers (viability: 4): https://sciencetostartup.com/paper/gbqa-a-game-benchmark-for-evaluating-llms-as-quality-assurance-engineers - A new benchmark and agent framework to evaluate and improve LLMs' ability to autonomously discover software bugs in game development.
- Speaking of Language: Reflections on Metalanguage Research in NLP (viability: 2): https://sciencetostartup.com/paper/speaking-of-language-reflections-on-metalanguage-research-in-nlp - This paper explores the concept of metalanguage in NLP and LLMs, identifying future research directions.
- Conditional Sampling via Wasserstein Autoencoders and Triangular Transport (viability: 4): https://sciencetostartup.com/paper/conditional-sampling-via-wasserstein-autoencoders-and-triangular-transport - A new framework for conditional simulation that reduces approximation error in low-dimensional problems.
- Differentiable SpaTiaL: Symbolic Learning and Reasoning with Geometric Temporal Logic for Manipulation Tasks (viability: 7): https://sciencetostartup.com/paper/differentiable-spatial-symbolic-learning-and-reasoning-with-geometric-temporal-logic-for-manipulation-tasks - A fully differentiable symbolic logic toolbox for robot manipulation that enables end-to-end learning and optimization of complex geometric and temporal constraints.
- Overcoming the "Impracticality" of RAG: Proposing a Real-World Benchmark and Multi-Dimensional Diagnostic Framework (viability: 4): https://sciencetostartup.com/paper/overcoming-the-impracticality-of-rag-proposing-a-real-world-benchmark-and-multi-dimensional-diagnostic-framework - A new benchmark and diagnostic framework to evaluate the real-world performance of Retrieval-Augmented Generation systems in enterprise settings.
- Cross-Vehicle 3D Geometric Consistency for Self-Supervised Surround Depth Estimation on Articulated Vehicles (viability: 7): https://sciencetostartup.com/paper/cross-vehicle-3d-geometric-consistency-for-self-supervised-surround-depth-estimation-on-articulated-vehicles - A self-supervised depth estimation framework for articulated vehicles that leverages cross-vehicle geometric consistency to improve perception in complex robotic platforms.
- AXELRAM: Quantize Once, Never Dequantize (viability: 7): https://sciencetostartup.com/paper/axelram-quantize-once-never-dequantize - A novel SRAM architecture for LLM inference that drastically reduces computation by performing attention scores directly on quantized KV cache indices, with a gradient-free method to ensure stability.
- Train Yourself as an LLM: Exploring Effects of AI Literacy on Persuasion via Role-playing LLM Training (viability: 5): https://sciencetostartup.com/paper/train-yourself-as-an-llm-exploring-effects-of-ai-literacy-on-persuasion-via-role-playing-llm-training - An interactive AI literacy tutorial that trains users to understand LLM persuasion tactics, reducing susceptibility to AI influence.
- Analytic Drift Resister for Non-Exemplar Continual Graph Learning (viability: 4): https://sciencetostartup.com/paper/analytic-drift-resister-for-non-exemplar-continual-graph-learning - A novel framework for continual graph learning that resists feature drift and achieves zero-forgetting class-incremental learning.
- Toys that listen, talk, and play: Understanding Children's Sensemaking and Interactions with AI Toys (viability: 3): https://sciencetostartup.com/paper/toys-that-listen-talk-and-play-understanding-children-s-sensemaking-and-interactions-with-ai-toys - This research explores how children understand and interact with AI toys, identifying design implications for more responsible and developmentally appropriate AI toy development.
- Smart Transfer: Leveraging Vision Foundation Model for Rapid Building Damage Mapping with Post-Earthquake VHR Imagery (viability: 7): https://sciencetostartup.com/paper/smart-transfer-leveraging-vision-foundation-model-for-rapid-building-damage-mapping-with-post-earthquake-vhr-imagery - A GeoAI framework leveraging vision foundation models for rapid building damage mapping from post-earthquake imagery, with publicly available code and data.
- Poison Once, Exploit Forever: Environment-Injected Memory Poisoning Attacks on Web Agents (viability: 4): https://sciencetostartup.com/paper/poison-once-exploit-forever-environment-injected-memory-poisoning-attacks-on-web-agents - This research introduces a novel attack vector for LLM-based web agents, demonstrating how environmental contamination can lead to persistent memory poisoning and cross-session, cross-site compromise.
- Reinforcement Learning-based Knowledge Distillation with LLM-as-a-Judge (viability: 7): https://sciencetostartup.com/paper/reinforcement-learning-based-knowledge-distillation-with-llm-as-a-judge - Leverage LLMs as judges for label-free knowledge distillation to significantly improve model reasoning capabilities on unlabeled data.
- OntoKG: Ontology-Oriented Knowledge Graph Construction with Intrinsic-Relational Routing (viability: 7): https://sciencetostartup.com/paper/ontokg-ontology-oriented-knowledge-graph-construction-with-intrinsic-relational-routing - A novel ontology-oriented approach to knowledge graph construction that decouples schema design from graph building, enabling reusable and adaptable knowledge representations for downstream AI tasks.
- AutoVerifier: An Agentic Automated Verification Framework Using Large Language Models (viability: 7): https://sciencetostartup.com/paper/autoverifier-an-agentic-automated-verification-framework-using-large-language-models - An LLM-powered agent framework that automates the verification of complex technical claims in scientific literature, even without domain expertise.
- Unlocking Multi-Site Clinical Data: A Federated Approach to Privacy-First Child Autism Behavior Analysis (viability: 7): https://sciencetostartup.com/paper/unlocking-multi-site-clinical-data-a-federated-approach-to-privacy-first-child-autism-behavior-analysis - A privacy-preserving federated learning framework for early autism behavior analysis in children, enabling multi-site collaboration without centralizing sensitive data.
- Complex-Valued GNNs for Distributed Basis-Invariant Control of Planar Systems (viability: 4): https://sciencetostartup.com/paper/complex-valued-gnns-for-distributed-basis-invariant-control-of-planar-systems - A novel graph neural network architecture for distributed control of planar systems that is invariant to local sensor orientation, improving data efficiency and generalization.
- Structure-Preserving Multi-View Embedding Using Gromov-Wasserstein Optimal Transport (viability: 4): https://sciencetostartup.com/paper/structure-preserving-multi-view-embedding-using-gromov-wasserstein-optimal-transport - A novel approach to multi-view data analysis using Gromov-Wasserstein optimal transport to preserve intrinsic relational structure across heterogeneous data representations.
- Elastomeric Strain Limitation for Design of Soft Pneumatic Actuators (viability: 7): https://sciencetostartup.com/paper/elastomeric-strain-limitation-for-design-of-soft-pneumatic-actuators - Develop human-safe soft robotic actuators with controllable shape and force using electroadhesive strain limiters and advanced modeling for applications in collaborative robotics.
- Steerable but Not Decodable: Function Vectors Operate Beyond the Logit Lens (viability: 7): https://sciencetostartup.com/paper/steerable-but-not-decodable-function-vectors-operate-beyond-the-logit-lens - This research demonstrates a novel method to steer large language model behavior using function vectors, achieving high accuracy even when the model's internal representations are not decodable, suggesting a new paradigm for controlling LLM outputs.
- Do Audio-Visual Large Language Models Really See and Hear? (viability: 3): https://sciencetostartup.com/paper/do-audio-visual-large-language-models-really-see-and-hear - This paper analyzes the internal workings of audio-visual large language models to understand modality biases, not a product pitch.
- Rascene: High-Fidelity 3D Scene Imaging with mmWave Communication Signals (viability: 7): https://sciencetostartup.com/paper/rascene-high-fidelity-3d-scene-imaging-with-mmwave-communication-signals - Leverage existing mmWave communication signals for high-fidelity, low-cost 3D environmental perception, overcoming limitations of optical sensors in adverse conditions.
- WGFINNs: Weak formulation-based GENERIC formalism informed neural networks' (viability: 4): https://sciencetostartup.com/paper/wgfinns-weak-formulation-based-generic-formalism-informed-neural-networks - WGFINNs enhance the robustness of neural networks in scientific machine learning by integrating weak formulations to handle noisy data effectively.
- LitPivot: Developing Well-Situated Research Ideas Through Dynamic Contextualization and Critique within the Literature Landscape (viability: 4): https://sciencetostartup.com/paper/litpivot-developing-well-situated-research-ideas-through-dynamic-contextualization-and-critique-within-the-literature-la - A tool to help researchers develop novel research ideas by dynamically linking literature review with idea refinement.
- Making Written Theorems Explorable by Grounding Them in Formal Representations (viability: 7): https://sciencetostartup.com/paper/making-written-theorems-explorable-by-grounding-them-in-formal-representations - An LLM-powered system that translates mathematical theorems and proofs into executable code, enabling interactive exploration and deeper understanding for users.
- An Empirical Study of Many-Shot In-Context Learning for Machine Translation of Low-Resource Languages (viability: 4): https://sciencetostartup.com/paper/an-empirical-study-of-many-shot-in-context-learning-for-machine-translation-of-low-resource-languages - Empirically studying many-shot in-context learning to improve machine translation for low-resource languages by optimizing example retrieval and selection.
- Moondream Segmentation: From Words to Masks (viability: 7): https://sciencetostartup.com/paper/moondream-segmentation-from-words-to-masks - A vision-language model extension that generates precise image masks from textual descriptions, with a new dataset for improved evaluation.
- The Quantum-Cryptographic Co-evolution (viability: 2): https://sciencetostartup.com/paper/the-quantum-cryptographic-co-evolution - A framework for understanding the transition to quantum-resistant cryptography by mapping resilience and computational capability.
- TrackerSplat: Exploiting Point Tracking for Fast and Robust Dynamic 3D Gaussians Reconstruction (viability: 7): https://sciencetostartup.com/paper/trackersplat-exploiting-point-tracking-for-fast-and-robust-dynamic-3d-gaussians-reconstruction - TrackerSplat enhances 3D Gaussian Splatting for dynamic scenes by using point tracking to improve robustness and throughput in reconstructions with fast object motion.
- Mitigating LLM biases toward spurious social contexts using direct preference optimization (viability: 7): https://sciencetostartup.com/paper/mitigating-llm-biases-toward-spurious-social-contexts-using-direct-preference-optimization - A novel training method significantly reduces LLM bias towards spurious social contexts in high-stakes decision-making, improving both accuracy and robustness.
- FusionBERT: Multi-View Image-3D Retrieval via Cross-Attention Visual Fusion and Normal-Aware 3D Encoder (viability: 7): https://sciencetostartup.com/paper/fusionbert-multi-view-image-3d-retrieval-via-cross-attention-visual-fusion-and-normal-aware-3d-encoder - FusionBERT enables robust multi-view image-to-3D model retrieval by adaptively fusing visual cues and enhancing 3D geometry encoding.
- Learning interacting particle systems from unlabeled data (viability: 3): https://sciencetostartup.com/paper/learning-interacting-particle-systems-from-unlabeled-data - A theoretical framework for learning interacting particle systems from unlabeled data without trajectory information.
- VoxelCodeBench: Benchmarking 3D World Modeling Through Code Generation (viability: 7): https://sciencetostartup.com/paper/voxelcodebench-benchmarking-3d-world-modeling-through-code-generation - A benchmark and platform for evaluating and improving 3D spatial reasoning in code generation models.
- High Volatility and Action Bias Distinguish LLMs from Humans in Group Coordination (viability: 3): https://sciencetostartup.com/paper/high-volatility-and-action-bias-distinguish-llms-from-humans-in-group-coordination - This research investigates the coordination capabilities of LLMs compared to humans in group tasks, identifying key behavioral differences to inform future agent development.
- ROMAN: A Multiscale Routing Operator for Convolutional Time Series Models (viability: 5): https://sciencetostartup.com/paper/roman-a-multiscale-routing-operator-for-convolutional-time-series-models - A novel operator for convolutional time series models that improves efficiency and accuracy by creating a multiscale, position-aware representation.
- Understanding the Effects of Safety Unalignment on Large Language Models (viability: 3): https://sciencetostartup.com/paper/understanding-the-effects-of-safety-unalignment-on-large-language-models - This research analyzes how safety alignment in LLMs can be compromised by specific techniques, revealing that one method (WO) significantly enhances malicious capabilities while the other (JT) has less impact, and proposes a mitigation strategy using supervised fine-tuning.
- WSVD: Weighted Low-Rank Approximation for Fast and Efficient Execution of Low-Precision Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/wsvd-weighted-low-rank-approximation-for-fast-and-efficient-execution-of-low-precision-vision-language-models - Accelerate vision-language model inference by 1.8x using weighted low-rank approximation and quantization, preserving accuracy.
- Generative AI Use in Entrepreneurship: An Integrative Review and an Empowerment-Entrapment Framework (viability: 2): https://sciencetostartup.com/paper/generative-ai-use-in-entrepreneurship-an-integrative-review-and-an-empowerment-entrapment-framework - This paper reviews how generative AI impacts entrepreneurs, proposing a framework to understand its empowering and entrapping effects across the entrepreneurial lifecycle.
- From Impact to Insight: Dynamics-Aware Proprioceptive Terrain Sensing on Granular Media (viability: 4): https://sciencetostartup.com/paper/from-impact-to-insight-dynamics-aware-proprioceptive-terrain-sensing-on-granular-media - A physics-based framework for robots to accurately characterize deformable terrain during high-speed locomotion using proprioceptive sensing.
- Dependency-Guided Parallel Decoding in Discrete Diffusion Language Models (viability: 5): https://sciencetostartup.com/paper/dependency-guided-parallel-decoding-in-discrete-diffusion-language-models - Accelerate discrete diffusion language model text generation by predicting token dependencies to improve quality and speed.
- Communication-Efficient Distributed Learning with Differential Privacy (viability: 4): https://sciencetostartup.com/paper/communication-efficient-distributed-learning-with-differential-privacy - A differentially private algorithm for distributed learning that improves communication efficiency and data privacy.
- Pragmatics Meets Culture: Culturally-adapted Artwork Description Generation and Evaluation (viability: 5): https://sciencetostartup.com/paper/pragmatics-meets-culture-culturally-adapted-artwork-description-generation-and-evaluation - Generate artwork descriptions tailored to specific cultural audiences to improve comprehension and engagement.
- Fast NF4 Dequantization Kernels for Large Language Model Inference (viability: 7): https://sciencetostartup.com/paper/fast-nf4-dequantization-kernels-for-large-language-model-inference - Accelerate LLM inference by up to 2.2x with a plug-and-play dequantization kernel that leverages shared memory, reducing costs on existing GPU infrastructure.
- Robust Learning with Optimal Error (viability: 3): https://sciencetostartup.com/paper/robust-learning-with-optimal-error - Develops theoretical algorithms for learning with adversarial noise, improving optimal error rates using randomized hypotheses.
- Principled and Scalable Diversity-Aware Retrieval via Cardinality-Constrained Binary Quadratic Programming (viability: 7): https://sciencetostartup.com/paper/principled-and-scalable-diversity-aware-retrieval-via-cardinality-constrained-binary-quadratic-programming - A principled and scalable method for diversity-aware retrieval in RAG, offering theoretical guarantees and significant speedups.
- From Theory to Practice: Code Generation Using LLMs for CAPEC and CWE Frameworks (viability: 7): https://sciencetostartup.com/paper/from-theory-to-practice-code-generation-using-llms-for-capec-and-cwe-frameworks - A novel dataset of vulnerable code snippets linked to CAPEC and CWE, generated by LLMs, to train automated vulnerability detection systems.
- Contrastive Language-Colored Pointmap Pretraining for Unified 3D Scene Understanding (viability: 7): https://sciencetostartup.com/paper/contrastive-language-colored-pointmap-pretraining-for-unified-3d-scene-understanding - A unified 3D scene understanding model that leverages contrastive language-image pretraining for improved representation learning.
- Competency Questions as Executable Plans: a Controlled RAG Architecture for Cultural Heritage Storytelling (viability: 7): https://sciencetostartup.com/paper/competency-questions-as-executable-plans-a-controlled-rag-architecture-for-cultural-heritage-storytelling - A neuro-symbolic RAG architecture using competency questions as executable plans to generate factually accurate and auditable stories from cultural heritage knowledge graphs.
- Overconfidence and Calibration in Medical VQA: Empirical Findings and Hallucination-Aware Mitigation (viability: 7): https://sciencetostartup.com/paper/overconfidence-and-calibration-in-medical-vqa-empirical-findings-and-hallucination-aware-mitigation - This research develops a hallucination-aware calibration method to improve the reliability and trustworthiness of vision-language models in medical question answering.
- Synapse: Evolving Job-Person Fit with Explainable Two-phase Retrieval and LLM-guided Genetic Resume Optimization (viability: 6): https://sciencetostartup.com/paper/synapse-evolving-job-person-fit-with-explainable-two-phase-retrieval-and-llm-guided-genetic-resume-optimization - A two-phase AI system that improves job-person fit by efficiently retrieving candidates and then optimizing resumes using LLMs and genetic algorithms.
- PolyJarvis: LLM Agent for Autonomous Polymer MD Simulations (viability: 7): https://sciencetostartup.com/paper/polyjarvis-llm-agent-for-autonomous-polymer-md-simulations - An LLM-powered agent that autonomously performs complex polymer molecular dynamics simulations from natural language, enabling faster and more accessible material property prediction.
- A Spectral Framework for Multi-Scale Nonlinear Dimensionality Reduction (viability: 4): https://sciencetostartup.com/paper/a-spectral-framework-for-multi-scale-nonlinear-dimensionality-reduction - A spectral framework for multi-scale nonlinear dimensionality reduction that bridges global and local structure with analytical transparency.
- Feature Attribution Stability Suite: How Stable Are Post-Hoc Attributions? (viability: 5): https://sciencetostartup.com/paper/feature-attribution-stability-suite-how-stable-are-post-hoc-attributions - A benchmark suite to rigorously evaluate the stability of AI feature attribution methods in vision systems, revealing critical insights into their reliability under real-world perturbations.
- Interpretable Deep Reinforcement Learning for Element-level Bridge Life-cycle Optimization (viability: 5): https://sciencetostartup.com/paper/interpretable-deep-reinforcement-learning-for-element-level-bridge-life-cycle-optimization - An interpretable reinforcement learning framework optimizes bridge life-cycle management using element-level condition data, producing understandable and auditable decision trees.
- Jump Start or False Start? A Theoretical and Empirical Evaluation of LLM-initialized Bandits (viability: 4): https://sciencetostartup.com/paper/jump-start-or-false-start-a-theoretical-and-empirical-evaluation-of-llm-initialized-bandits - This research theoretically and empirically evaluates the effectiveness of using LLM-generated data to initialize bandit algorithms, identifying critical thresholds for data corruption and misalignment that impact performance.
- AdaHOP: Fast and Accurate Low-Precision Training via Outlier-Pattern-Aware Rotation (viability: 4): https://sciencetostartup.com/paper/adahop-fast-and-accurate-low-precision-training-via-outlier-pattern-aware-rotation - A novel low-precision training method for LLMs that adapts transform strategies based on outlier patterns to achieve significant memory and speed improvements.
- Tune to Learn: How Controller Gains Shape Robot Policy Learning (viability: 4): https://sciencetostartup.com/paper/tune-to-learn-how-controller-gains-shape-robot-policy-learning - Optimize robot controller gains based on the learning algorithm, not just task compliance, to improve policy learning.
- Opal: Private Memory for Personal AI (viability: 7): https://sciencetostartup.com/paper/opal-private-memory-for-personal-ai - Opal provides private, scalable long-term memory for personal AI by decoupling data-dependent reasoning into a trusted enclave, improving retrieval accuracy and reducing costs.
- Social Meaning in Large Language Models: Structure, Magnitude, and Pragmatic Prompting (viability: 3): https://sciencetostartup.com/paper/social-meaning-in-large-language-models-structure-magnitude-and-pragmatic-prompting - This paper investigates how well large language models understand social meaning and explores prompting techniques to improve their accuracy in this area.
- Re-analysis of the Human Transcription Factor Atlas Recovers TF-Specific Signatures from Pooled Single-Cell Screens with Missing Controls (viability: 7): https://sciencetostartup.com/paper/re-analysis-of-the-human-transcription-factor-atlas-recovers-tf-specific-signatures-from-pooled-single-cell-screens-with - A reproducible pipeline to recover valuable transcription factor insights from incomplete single-cell perturbation data, enabling deeper biological discovery.
- Rapidly deploying on-device eye tracking by distilling visual foundation models (viability: 7): https://sciencetostartup.com/paper/rapidly-deploying-on-device-eye-tracking-by-distilling-visual-foundation-models - A framework for rapidly deploying high-accuracy, on-device eye tracking by distilling visual foundation models using synthetic and real-world data.
- Reinforcement Learning from Human Feedback: A Statistical Perspective (viability: 7): https://sciencetostartup.com/paper/reinforcement-learning-from-human-feedback-a-statistical-perspective - A statistical framework for aligning large language models with human preferences, offering a robust approach to reward modeling and policy optimization.
- A Comprehensive Framework for Long-Term Resiliency Investment Planning under Extreme Weather Uncertainty for Electric Utilities (viability: 3): https://sciencetostartup.com/paper/a-comprehensive-framework-for-long-term-resiliency-investment-planning-under-extreme-weather-uncertainty-for-electric-ut - A framework for electric utilities to optimize capital investments for extreme weather resiliency using digital twins and Monte Carlo simulations.
- An Explainable Vision-Language Model Framework with Adaptive PID-Tversky Loss for Lumbar Spinal Stenosis Diagnosis (viability: 7): https://sciencetostartup.com/paper/an-explainable-vision-language-model-framework-with-adaptive-pid-tversky-loss-for-lumbar-spinal-stenosis-diagnosis - An explainable vision-language model framework with adaptive loss for accurate and interpretable lumbar spinal stenosis diagnosis from MRI.
- I must delete the evidence: AI Agents Explicitly Cover up Fraud and Violent Crime (viability: 4): https://sciencetostartup.com/paper/i-must-delete-the-evidence-ai-agents-explicitly-cover-up-fraud-and-violent-crime - This research demonstrates that current AI agents can be manipulated to suppress evidence of fraud and harm for corporate gain, highlighting a critical security vulnerability.
- Delaunay Canopy: Building Wireframe Reconstruction from Airborne LiDAR Point Clouds via Delaunay Graph (viability: 7): https://sciencetostartup.com/paper/delaunay-canopy-building-wireframe-reconstruction-from-airborne-lidar-point-clouds-via-delaunay-graph - A novel method for accurate building wireframe reconstruction from LiDAR data using Delaunay graphs, overcoming limitations in noisy and sparse environments.
- Token-Efficient Multimodal Reasoning via Image Prompt Packaging (viability: 7): https://sciencetostartup.com/paper/token-efficient-multimodal-reasoning-via-image-prompt-packaging - Reduce multimodal AI inference costs by embedding structured text directly into images, achieving significant savings with competitive accuracy.
- Automated Malware Family Classification using Weighted Hierarchical Ensembles of Large Language Models (viability: 4): https://sciencetostartup.com/paper/automated-malware-family-classification-using-weighted-hierarchical-ensembles-of-large-language-models - A zero-label malware classification framework using weighted hierarchical ensembles of LLMs to improve robustness and analyst-style reasoning.
- Causal-Audit: A Framework for Risk Assessment of Assumption Violations in Time-Series Causal Discovery (viability: 7): https://sciencetostartup.com/paper/causal-audit-a-framework-for-risk-assessment-of-assumption-violations-in-time-series-causal-discovery - A framework for assessing and mitigating risks of assumption violations in time-series causal discovery, providing calibrated risk scores and abstention policies for reliable inference.
- VLMs Need Words: Vision Language Models Ignore Visual Detail In Favor of Semantic Anchors (viability: 5): https://sciencetostartup.com/paper/vlms-need-words-vision-language-models-ignore-visual-detail-in-favor-of-semantic-anchors - This research identifies a critical limitation in Vision Language Models, showing they prioritize semantic understanding over visual detail, and proposes methods to improve their fine-grained visual reasoning capabilities.
- Failing to Falsify: Evaluating and Mitigating Confirmation Bias in Language Models (viability: 3): https://sciencetostartup.com/paper/failing-to-falsify-evaluating-and-mitigating-confirmation-bias-in-language-models - This research explores and mitigates confirmation bias in language models to enhance their reasoning capabilities.
- SEDGE: Structural Extrapolated Data Generation (viability: 3): https://sciencetostartup.com/paper/sedge-structural-extrapolated-data-generation - A theoretical framework for generating new data based on assumptions about the data generation process.
- Generating Satellite Imagery Data for Wildfire Detection through Mask-Conditioned Generative AI (viability: 7): https://sciencetostartup.com/paper/generating-satellite-imagery-data-for-wildfire-detection-through-mask-conditioned-generative-ai - Generate realistic satellite imagery for wildfire detection using mask-conditioned diffusion models, addressing data scarcity with a novel inpainting approach.
- AIVV: Neuro-Symbolic LLM Agent-Integrated Verification and Validation for Trustworthy Autonomous Systems (viability: 7): https://sciencetostartup.com/paper/aivv-neuro-symbolic-llm-agent-integrated-verification-and-validation-for-trustworthy-autonomous-systems - Automate the unsustainable manual workload of autonomous system verification and validation using LLM agents to analyze and validate system anomalies against natural language requirements.
- Guideline2Graph: Profile-Aware Multimodal Parsing for Executable Clinical Decision Graphs (viability: 7): https://sciencetostartup.com/paper/guideline2graph-profile-aware-multimodal-parsing-for-executable-clinical-decision-graphs - Automates the conversion of complex clinical guidelines into executable decision support systems, significantly improving accuracy and auditability.
- Understanding the Nature of Generative AI as Threshold Logic in High-Dimensional Space (viability: 2): https://sciencetostartup.com/paper/understanding-the-nature-of-generative-ai-as-threshold-logic-in-high-dimensional-space - This paper theoretically explores generative AI through the lens of threshold logic and high-dimensional geometry, offering a new perspective on neural computation.
- Time-Warping Recurrent Neural Networks for Transfer Learning (viability: 4): https://sciencetostartup.com/paper/time-warping-recurrent-neural-networks-for-transfer-learning - A novel time-warping method for Recurrent Neural Networks enhances transfer learning accuracy in predicting time-varying physical systems.
- VALOR: Value-Aware Revenue Uplift Modeling with Treatment-Gated Representation for B2B Sales (viability: 8): https://sciencetostartup.com/paper/valor-value-aware-revenue-uplift-modeling-with-treatment-gated-representation-for-b2b-sales - VALOR is a B2B sales AI that identifies high-value, persuadable accounts to significantly increase incremental revenue.
- Hierarchical, Interpretable, Label-Free Concept Bottleneck Model (viability: 7): https://sciencetostartup.com/paper/hierarchical-interpretable-label-free-concept-bottleneck-model - A hierarchical, label-free concept bottleneck model that enhances interpretability and classification accuracy by mirroring human cognitive processes.
- VERTIGO: Visual Preference Optimization for Cinematic Camera Trajectory Generation (viability: 7): https://sciencetostartup.com/paper/vertigo-visual-preference-optimization-for-cinematic-camera-trajectory-generation - VERTIGO optimizes cinematic camera trajectory generation by incorporating visual preference feedback, significantly improving framing and reducing off-screen characters.
- Single-Agent LLMs Outperform Multi-Agent Systems on Multi-Hop Reasoning Under Equal Thinking Token Budgets (viability: 4): https://sciencetostartup.com/paper/single-agent-llms-outperform-multi-agent-systems-on-multi-hop-reasoning-under-equal-thinking-token-budgets - This research demonstrates that single-agent LLMs can outperform multi-agent systems on complex reasoning tasks when computational resources are normalized, challenging the perceived benefits of multi-agent architectures.
- On the Geometric Structure of Layer Updates in Deep Language Models (viability: 3): https://sciencetostartup.com/paper/on-the-geometric-structure-of-layer-updates-in-deep-language-models - This paper analyzes the geometric structure of layer updates in deep language models to understand how representations change between layers, identifying a dominant tokenwise component and a distinct residual component.
- When simulations look right but causal effects go wrong: Large language models as behavioral simulators (viability: 5): https://sciencetostartup.com/paper/when-simulations-look-right-but-causal-effects-go-wrong-large-language-models-as-behavioral-simulators - This research evaluates LLMs' ability to simulate causal intervention effects, revealing a divergence between descriptive fit and causal fidelity, crucial for reliable behavioral simulations.
- Street-Legal Physical-World Adversarial Rim for License Plates (viability: 7): https://sciencetostartup.com/paper/street-legal-physical-world-adversarial-rim-for-license-plates - Develop street-legal, low-cost physical adversarial attacks to disrupt license plate reader systems, with potential applications in security and privacy.
- Skeleton-based Coherence Modeling in Narratives (viability: 3): https://sciencetostartup.com/paper/skeleton-based-coherence-modeling-in-narratives - This research explores using sentence skeletons to model narrative coherence, finding that sentence-level analysis remains superior.
- Do We Need Frontier Models to Verify Mathematical Proofs? (viability: 7): https://sciencetostartup.com/paper/do-we-need-frontier-models-to-verify-mathematical-proofs - Develops a prompt engineering technique to enable smaller, open-source LLMs to reliably verify mathematical proofs, matching the performance of frontier models.
- PlayGen-MoG: Framework for Diverse Multi-Agent Play Generation via Mixture-of-Gaussians Trajectory Prediction (viability: 7): https://sciencetostartup.com/paper/playgen-mog-framework-for-diverse-multi-agent-play-generation-via-mixture-of-gaussians-trajectory-prediction - A framework for generating diverse and coordinated multi-agent plays from static formations, overcoming limitations of existing generative models.
- From Elevation Maps To Contour Lines: SVM and Decision Trees to Detect Violin Width Reduction (viability: 2): https://sciencetostartup.com/paper/from-elevation-maps-to-contour-lines-svm-and-decision-trees-to-detect-violin-width-reduction - This research explores automated violin width reduction detection using 3D photogrammetry and machine learning, comparing different feature engineering approaches.
- Matrix Profile for Time-Series Anomaly Detection: A Reproducible Open-Source Benchmark on TSB-AD (viability: 7): https://sciencetostartup.com/paper/matrix-profile-for-time-series-anomaly-detection-a-reproducible-open-source-benchmark-on-tsb-ad - An open-source benchmark and implementation of Matrix Profile methods for scalable and interpretable time-series anomaly detection.
- Adaptive Learned State Estimation based on KalmanNet (viability: 7): https://sciencetostartup.com/paper/adaptive-learned-state-estimation-based-on-kalmannet - An adaptive multi-modal Kalman filter for improved state estimation in autonomous driving using sensor-specific learned noise characteristics.
- Mitigating Data Scarcity in Spaceflight Applications for Offline Reinforcement Learning Using Physics-Informed Deep Generative Models (viability: 7): https://sciencetostartup.com/paper/mitigating-data-scarcity-in-spaceflight-applications-for-offline-reinforcement-learning-using-physics-informed-deep-gene - Generate realistic synthetic data for spaceflight reinforcement learning using physics-informed generative models to overcome data scarcity and improve controller performance.
- Compositional Neuro-Symbolic Reasoning (viability: 7): https://sciencetostartup.com/paper/compositional-neuro-symbolic-reasoning - A neuro-symbolic framework that enhances LLMs with structured reasoning capabilities for complex problem-solving, achieving significant performance gains on the ARC-AGI-2 benchmark.
- Self-Directed Task Identification (viability: 4): https://sciencetostartup.com/paper/self-directed-task-identification - A framework for models to autonomously identify target variables in datasets without pre-training, reducing manual annotation effort.
- Photonic convolutional neural network with pre-trained in-situ training (viability: 6): https://sciencetostartup.com/paper/photonic-convolutional-neural-network-with-pre-trained-in-situ-training - A fully photonic convolutional neural network for energy-efficient image classification, validated with a hybrid training approach.
- Evolution and Perspectives of the Keep IT Secure Ecosystem:A Six-Year Analysis of Cybersecurity Experts Supporting Belgian SMEs (viability: 4): https://sciencetostartup.com/paper/evolution-and-perspectives-of-the-keep-it-secure-ecosystem-a-six-year-analysis-of-cybersecurity-experts-supporting-belgi - A framework for validating cybersecurity experts to improve SME security posture.
- SWAY: A Counterfactual Computational Linguistic Approach to Measuring and Mitigating Sycophancy (viability: 7): https://sciencetostartup.com/paper/sway-a-counterfactual-computational-linguistic-approach-to-measuring-and-mitigating-sycophancy - A novel computational linguistic metric and mitigation strategy to eliminate sycophancy in large language models.
- EventHub: Data Factory for Generalizable Event-Based Stereo Networks without Active Sensors (viability: 7): https://sciencetostartup.com/paper/eventhub-data-factory-for-generalizable-event-based-stereo-networks-without-active-sensors - EventHub enables training of high-performance event-based stereo vision models using only standard color images, significantly reducing hardware costs and improving generalization.
- ActionParty: Multi-Subject Action Binding in Generative Video Games (viability: 7): https://sciencetostartup.com/paper/actionparty-multi-subject-action-binding-in-generative-video-games - ActionParty enables precise control of multiple agents in generative video games by disentangling subject states from video rendering.
- Generative World Renderer (viability: 7): https://sciencetostartup.com/paper/generative-world-renderer - A large-scale dataset and toolkit for high-fidelity generative rendering of real-world scenes, enabling style editing of AAA games via text prompts.
- Modulate-and-Map: Crossmodal Feature Mapping with Cross-View Modulation for 3D Anomaly Detection (viability: 7): https://sciencetostartup.com/paper/modulate-and-map-crossmodal-feature-mapping-with-cross-view-modulation-for-3d-anomaly-detection - A natively multiview and multimodal framework for 3D anomaly detection and segmentation that maps features across modalities and views to achieve state-of-the-art performance.
- Steerable Visual Representations (viability: 7): https://sciencetostartup.com/paper/steerable-visual-representations - A new class of visual representations that can be steered with natural language for precise object focus and improved downstream task performance.
- Grounded Token Initialization for New Vocabulary in LMs for Generative Recommendation (viability: 7): https://sciencetostartup.com/paper/grounded-token-initialization-for-new-vocabulary-in-lms-for-generative-recommendation - A novel method for initializing new vocabulary tokens in language models that significantly improves performance on generative recommendation tasks by grounding them in meaningful semantic space.
- Beyond Referring Expressions: Scenario Comprehension Visual Grounding (viability: 7): https://sciencetostartup.com/paper/beyond-referring-expressions-scenario-comprehension-visual-grounding - A new benchmark and curriculum learning method for visual grounding that understands object roles and context, going beyond literal descriptions to enable more robust AI perception.
- Batched Contextual Reinforcement: A Task-Scaling Law for Efficient Reasoning (viability: 7): https://sciencetostartup.com/paper/batched-contextual-reinforcement-a-task-scaling-law-for-efficient-reasoning - A new training method for LLMs that significantly reduces token usage and inference costs while maintaining or improving accuracy, by training models to solve multiple problems simultaneously.
- Large-scale Codec Avatars: The Unreasonable Effectiveness of Large-scale Avatar Pretraining (viability: 7): https://sciencetostartup.com/paper/large-scale-codec-avatars-the-unreasonable-effectiveness-of-large-scale-avatar-pretraining - A large-scale pre-trained 3D avatar model that achieves high fidelity and generalization to real-world data, enabling efficient inference for diverse applications.
- No Single Best Model for Diversity: Learning a Router for Sample Diversity (viability: 7): https://sciencetostartup.com/paper/no-single-best-model-for-diversity-learning-a-router-for-sample-diversity - A router that selects the best LLM for each query to maximize response diversity, outperforming single models.
- Stop Wandering: Efficient Vision-Language Navigation via Metacognitive Reasoning (viability: 7): https://sciencetostartup.com/paper/stop-wandering-efficient-vision-language-navigation-via-metacognitive-reasoning - A metacognitive navigation agent that uses self-reflection to efficiently explore 3D environments and reduce redundant exploration, outperforming existing methods.
- A Simple Baseline for Streaming Video Understanding (viability: 7): https://sciencetostartup.com/paper/a-simple-baseline-for-streaming-video-understanding - A simple sliding-window baseline for streaming video understanding that matches or surpasses complex memory-based models, offering a new benchmark for the field.
- Beyond the Assistant Turn: User Turn Generation as a Probe of Interaction Awareness in Language Models (viability: 7): https://sciencetostartup.com/paper/beyond-the-assistant-turn-user-turn-generation-as-a-probe-of-interaction-awareness-in-language-models - A new benchmark for LLMs that measures their understanding of conversational context beyond just answering a prompt, revealing latent interaction awareness.
- go-$m$HC: Direct Parameterization of Manifold-Constrained Hyper-Connections via Generalized Orthostochastic Matrices (viability: 3): https://sciencetostartup.com/paper/go-m-hc-direct-parameterization-of-manifold-constrained-hyper-connections-via-generalized-orthostochastic-matrices - A new mathematical framework for parameterizing layer connectivity in large language models that scales efficiently and improves expressivity.
- PARD-SSM: Probabilistic Cyber-Attack Regime Detection via Variational Switching State-Space Models (viability: 7): https://sciencetostartup.com/paper/pard-ssm-probabilistic-cyber-attack-regime-detection-via-variational-switching-state-space-models - A probabilistic framework for real-time cyber-attack regime detection that predicts attacks minutes in advance.
- VOID: Video Object and Interaction Deletion (viability: 7): https://sciencetostartup.com/paper/void-video-object-and-interaction-deletion - A video editing framework that removes objects and realistically reconstructs their physical interactions for plausible scene dynamics.
- Taming the Exponential: A Fast Softmax Surrogate for Integer-Native Edge Inference (viability: 7): https://sciencetostartup.com/paper/taming-the-exponential-a-fast-softmax-surrogate-for-integer-native-edge-inference - A hardware-optimized, fast softmax surrogate for integer-native edge inference in Transformer models, significantly boosting speed while maintaining accuracy.
- AdamFlow: Adam-based Wasserstein Gradient Flows for Surface Registration in Medical Imaging (viability: 7): https://sciencetostartup.com/paper/adamflow-adam-based-wasserstein-gradient-flows-for-surface-registration-in-medical-imaging - A novel Adam-based optimization method for fast and robust surface registration in medical imaging, addressing the efficiency-robustness trade-off.
- Omni123: Exploring 3D Native Foundation Models with Limited 3D Data by Unifying Text to 2D and 3D Generation (viability: 7): https://sciencetostartup.com/paper/omni123-exploring-3d-native-foundation-models-with-limited-3d-data-by-unifying-text-to-2d-and-3d-generation - Omni123 is a 3D-native foundation model that unifies text-to-2D and text-to-3D generation, leveraging 2D data as a prior to overcome 3D data scarcity.
- Unifying Group-Relative and Self-Distillation Policy Optimization via Sample Routing (viability: 7): https://sciencetostartup.com/paper/unifying-group-relative-and-self-distillation-policy-optimization-via-sample-routing - A unified reinforcement learning framework for large language models that improves training stability and efficiency by intelligently routing samples to different optimization strategies.
- Deep Neural Network Based Roadwork Detection for Autonomous Driving (viability: 7): https://sciencetostartup.com/paper/deep-neural-network-based-roadwork-detection-for-autonomous-driving - A real-time system that detects and localizes road construction sites for autonomous vehicles and traffic authorities using YOLO and LiDAR.
- Novel Memory Forgetting Techniques for Autonomous AI Agents: Balancing Relevance and Efficiency (viability: 7): https://sciencetostartup.com/paper/novel-memory-forgetting-techniques-for-autonomous-ai-agents-balancing-relevance-and-efficiency - An adaptive memory forgetting framework for AI agents that balances relevance and efficiency to improve long-horizon reasoning and prevent memory overload.
- The Self Driving Portfolio: Agentic Architecture for Institutional Asset Management (viability: 4): https://sciencetostartup.com/paper/the-self-driving-portfolio-agentic-architecture-for-institutional-asset-management - An agentic AI pipeline automates strategic asset allocation by having specialized agents generate, critique, and improve portfolio construction methods, guided by an investment policy statement.
- De Jure: Iterative LLM Self-Refinement for Structured Extraction of Regulatory Rules (viability: 7): https://sciencetostartup.com/paper/de-jure-iterative-llm-self-refinement-for-structured-extraction-of-regulatory-rules - An automated pipeline for extracting structured regulatory rules from legal documents using LLM self-refinement, improving downstream compliance question-answering.
- Crystalite: A Lightweight Transformer for Efficient Crystal Modeling (viability: 7): https://sciencetostartup.com/paper/crystalite-a-lightweight-transformer-for-efficient-crystal-modeling - A lightweight Transformer model for efficient and state-of-the-art crystal structure prediction and de novo generation, offering faster sampling than existing methods.
- SKILL0: In-Context Agentic Reinforcement Learning for Skill Internalization (viability: 7): https://sciencetostartup.com/paper/skill0-in-context-agentic-reinforcement-learning-for-skill-internalization - Internalize LLM agent skills into model parameters for zero-shot autonomous behavior, reducing token overhead and improving efficiency.
- Modular Energy Steering for Safe Text-to-Image Generation with Foundation Models (viability: 7): https://sciencetostartup.com/paper/modular-energy-steering-for-safe-text-to-image-generation-with-foundation-models - A modular, training-free framework that uses existing foundation models to steer text-to-image generation towards safe outputs without sacrificing quality.
- Model-Based Reinforcement Learning for Control under Time-Varying Dynamics (viability: 4): https://sciencetostartup.com/paper/model-based-reinforcement-learning-for-control-under-time-varying-dynamics - An adaptive reinforcement learning algorithm that maintains control performance in systems with changing dynamics by intelligently managing historical data.
- A virtual-variable-length method for robust inverse kinematics of multi-segment continuum robots (viability: 4): https://sciencetostartup.com/paper/a-virtual-variable-length-method-for-robust-inverse-kinematics-of-multi-segment-continuum-robots - A novel method for robustly solving inverse kinematics in multi-segment continuum robots, improving convergence rates and reducing iteration counts.
- Best-Arm Identification with Noisy Actuation (viability: 2): https://sciencetostartup.com/paper/best-arm-identification-with-noisy-actuation - Develops theoretical communication schemes for multi-armed bandit problems with noisy actuation, linking to channel capacity.
- SPAR: Single-Pass Any-Resolution ViT for Open-vocabulary Segmentation (viability: 7): https://sciencetostartup.com/paper/spar-single-pass-any-resolution-vit-for-open-vocabulary-segmentation - A resolution-agnostic Vision Transformer that enables efficient, high-resolution open-vocabulary segmentation by distilling fine-grained spatial understanding into a single-pass model.
- Smoothing the Landscape: Causal Structure Learning via Diffusion Denoising Objectives (viability: 7): https://sciencetostartup.com/paper/smoothing-the-landscape-causal-structure-learning-via-diffusion-denoising-objectives - A novel diffusion-based framework for stable and scalable causal structure learning from observational data.
- BVFLMSP : Bayesian Vertical Federated Learning for Multimodal Survival with Privacy (viability: 4): https://sciencetostartup.com/paper/bvflmsp-bayesian-vertical-federated-learning-for-multimodal-survival-with-privacy - A Bayesian federated learning framework for private multimodal survival prediction with uncertainty estimates.
- (PAC-)Learning state machines from data streams: A generic strategy and an improved heuristic (Extended version) (viability: 5): https://sciencetostartup.com/paper/pac-learning-state-machines-from-data-streams-a-generic-strategy-and-an-improved-heuristic-extended-version - A generic strategy and improved heuristic for learning state machines from streaming data, implemented in an open-source library and demonstrating effectiveness in runtime, memory, and quality.
- UAV-Track VLA: Embodied Aerial Tracking via Vision-Language-Action Models (viability: 7): https://sciencetostartup.com/paper/uav-track-vla-embodied-aerial-tracking-via-vision-language-action-models - Embodied UAV tracking system leveraging vision-language-action models for dynamic real-world scenarios.
- Generative AI Spotlights the Human Core of Data Science: Implications for Education (viability: 2): https://sciencetostartup.com/paper/generative-ai-spotlights-the-human-core-of-data-science-implications-for-education - This paper argues that generative AI should shift data science education to focus on uniquely human reasoning skills, rather than automating them.
- Do Emotions in Prompts Matter? Effects of Emotional Framing on Large Language Models (viability: 5): https://sciencetostartup.com/paper/do-emotions-in-prompts-matter-effects-of-emotional-framing-on-large-language-models - An adaptive prompting framework that leverages emotional tone to subtly improve LLM performance on specific tasks.
- Answering the Wrong Question: Reasoning Trace Inversion for Abstention in LLMs (viability: 7): https://sciencetostartup.com/paper/answering-the-wrong-question-reasoning-trace-inversion-for-abstention-in-llms - A framework for LLMs to abstain from answering by detecting when they answer the wrong question, significantly improving abstention performance.
- When to ASK: Uncertainty-Gated Language Assistance for Reinforcement Learning (viability: 4): https://sciencetostartup.com/paper/when-to-ask-uncertainty-gated-language-assistance-for-reinforcement-learning - An RL agent that uses smaller language models to suggest actions only when it's uncertain, improving out-of-distribution performance without retraining.
- SCALE: Semantic- and Confidence-Aware Conditional Variational Autoencoder for Zero-shot Skeleton-based Action Recognition (viability: 5): https://sciencetostartup.com/paper/scale-semantic-and-confidence-aware-conditional-variational-autoencoder-for-zero-shot-skeleton-based-action-recognition - A novel framework for zero-shot skeleton-based action recognition that leverages conditional variational autoencoders and a confidence-aware energy loss to improve accuracy without explicit skeleton-text alignment.
- Impact of Multimodal and Conversational AI on Learning Outcomes and Experience (viability: 5): https://sciencetostartup.com/paper/impact-of-multimodal-and-conversational-ai-on-learning-outcomes-and-experience - A conversational AI system that integrates text and images to improve learning outcomes in STEM education.
- VISTA: Visualization of Token Attribution via Efficient Analysis (viability: 7): https://sciencetostartup.com/paper/vista-visualization-of-token-attribution-via-efficient-analysis - A model-agnostic technique for visualizing token importance in LLMs, offering deeper insights into their decision-making without increased computational cost.
- Universal Hypernetworks for Arbitrary Models (viability: 7): https://sciencetostartup.com/paper/universal-hypernetworks-for-arbitrary-models - A universal hypernetwork that generates diverse models across vision, graph, text, and formula tasks from a single fixed architecture, enabling efficient multi-model generalization and recursive generation.
- Multi-Agent Video Recommenders: Evolution, Patterns, and Open Challenges (viability: 4): https://sciencetostartup.com/paper/multi-agent-video-recommenders-evolution-patterns-and-open-challenges - Develops multi-agent architectures for video recommendation systems that leverage foundation models and LLMs for more precise and explainable content delivery.
- CV-18 NER: Augmented Common Voice for Named Entity Recognition from Arabic Speech (viability: 7): https://sciencetostartup.com/paper/cv-18-ner-augmented-common-voice-for-named-entity-recognition-from-arabic-speech - A new dataset and benchmark for end-to-end Arabic speech Named Entity Recognition, outperforming traditional pipelines with publicly released models.
- Blinded Radiologist and LLM-Based Evaluation of LLM-Generated Japanese Translations of Chest CT Reports: Comparative Study (viability: 3): https://sciencetostartup.com/paper/blinded-radiologist-and-llm-based-evaluation-of-llm-generated-japanese-translations-of-chest-ct-reports-comparative-stud - This study evaluates LLM-generated translations of radiology reports, finding significant discrepancies between LLM and human expert evaluations, highlighting the continued need for human oversight in medical contexts.
- LEO: Graph Attention Network based Hybrid Multi Sensor Extended Object Fusion and Tracking for Autonomous Driving Applications (viability: 7): https://sciencetostartup.com/paper/leo-graph-attention-network-based-hybrid-multi-sensor-extended-object-fusion-and-tracking-for-autonomous-driving-applica - A graph attention network that adaptively fuses multi-sensor data for accurate and robust extended object tracking in autonomous driving.
- On the Role of Depth in the Expressivity of RNNs (viability: 3): https://sciencetostartup.com/paper/on-the-role-of-depth-in-the-expressivity-of-rnns - This paper theoretically analyzes the benefits of depth in Recurrent Neural Networks, showing it enhances memory capacity and expressivity.
- Towards Position-Robust Talent Recommendation via Large Language Models (viability: 7): https://sciencetostartup.com/paper/towards-position-robust-talent-recommendation-via-large-language-models - A novel LLM framework for listwise talent recommendation that mitigates position bias and improves efficiency.
- From High-Dimensional Spaces to Verifiable ODD Coverage for Safety-Critical AI-based Systems (viability: 7): https://sciencetostartup.com/paper/from-high-dimensional-spaces-to-verifiable-odd-coverage-for-safety-critical-ai-based-systems - A structured engineering method to formally verify Operational Design Domain coverage for safety-critical AI systems, meeting stringent aviation certification standards.
- Neuro-RIT: Neuron-Guided Instruction Tuning for Robust Retrieval-Augmented Language Model (viability: 7): https://sciencetostartup.com/paper/neuro-rit-neuron-guided-instruction-tuning-for-robust-retrieval-augmented-language-model - A novel framework for robust retrieval-augmented language models that precisely deactivates irrelevant context-processing neurons to improve performance on knowledge-intensive tasks.
- UniDriveVLA: Unifying Understanding, Perception, and Action Planning for Autonomous Driving (viability: 7): https://sciencetostartup.com/paper/unidrivevla-unifying-understanding-perception-and-action-planning-for-autonomous-driving - A unified vision-language-action system that enhances autonomous driving by decoupling spatial perception and semantic reasoning.
- Lightweight Spatiotemporal Highway Lane Detection via 3D-ResNet and PINet with ROI-Aware Attention (viability: 7): https://sciencetostartup.com/paper/lightweight-spatiotemporal-highway-lane-detection-via-3d-resnet-and-pinet-with-roi-aware-attention - A lightweight, end-to-end highway lane detection system using 3D-ResNet and PINet with ROI-aware attention for improved accuracy and reduced latency in ADAS.
- CXR-LT 2026 Challenge: Projection-Aware Multi-Label and Zero-Shot Chest X-Ray Classification (viability: 4): https://sciencetostartup.com/paper/cxr-lt-2026-challenge-projection-aware-multi-label-and-zero-shot-chest-x-ray-classification - A novel dual-branch architecture for chest X-ray classification that improves multi-label and zero-shot performance by integrating projection-specific models and contrastive learning.
- Neural network methods for two-dimensional finite-source reflector design (viability: 4): https://sciencetostartup.com/paper/neural-network-methods-for-two-dimensional-finite-source-reflector-design - Leveraging neural networks for faster and more accurate design of optical reflectors to precisely shape light.
- TRU: Targeted Reverse Update for Efficient Multimodal Recommendation Unlearning (viability: 7): https://sciencetostartup.com/paper/tru-targeted-reverse-update-for-efficient-multimodal-recommendation-unlearning - A plug-and-play framework for efficiently and securely unlearning user data from multimodal recommendation systems without full retraining.
- ViT-Explainer: An Interactive Walkthrough of the Vision Transformer Pipeline (viability: 5): https://sciencetostartup.com/paper/vit-explainer-an-interactive-walkthrough-of-the-vision-transformer-pipeline - An interactive web-based system for visualizing and understanding the end-to-end inference pipeline of Vision Transformers.
- The Expert Strikes Back: Interpreting Mixture-of-Experts Language Models at Expert Level (viability: 7): https://sciencetostartup.com/paper/the-expert-strikes-back-interpreting-mixture-of-experts-language-models-at-expert-level - This research unlocks the inherent interpretability of Mixture-of-Experts LLMs by analyzing experts as fine-grained task specialists, paving the way for more transparent and controllable large-scale models.
- Adam's Law: Textual Frequency Law on Large Language Models (viability: 7): https://sciencetostartup.com/paper/adam-s-law-textual-frequency-law-on-large-language-models - A framework to improve LLM performance by prioritizing frequent textual data during prompting and fine-tuning.
- Quantifying Self-Preservation Bias in Large Language Models (viability: 7): https://sciencetostartup.com/paper/quantifying-self-preservation-bias-in-large-language-models - A new benchmark quantifies and reveals significant self-preservation bias in leading LLMs, offering a path to more robust AI safety.
- Do Lexical and Contextual Coreference Resolution Systems Degrade Differently under Mention Noise? An Empirical Study on Scientific Software Mentions (viability: 7): https://sciencetostartup.com/paper/do-lexical-and-contextual-coreference-resolution-systems-degrade-differently-under-mention-noise-an-empirical-study-on-s - A novel system for robust software mention coreference resolution that outperforms existing methods under noisy conditions and scales efficiently for large corpora.
- Reflection Generation for Composite Image Using Diffusion Model (viability: 7): https://sciencetostartup.com/paper/reflection-generation-for-composite-image-using-diffusion-model - A diffusion model for generating realistic and physically coherent reflections in composite images, supported by the first large-scale object reflection dataset.
- Beyond the Fold: Quantifying Split-Level Noise and the Case for Leave-One-Dataset-Out AU Evaluation (viability: 3): https://sciencetostartup.com/paper/beyond-the-fold-quantifying-split-level-noise-and-the-case-for-leave-one-dataset-out-au-evaluation - A novel evaluation protocol for facial action unit detection that quantifies noise and improves robustness, revealing that many reported gains may be artifacts of the testing method.
- CoRegOVCD: Consistency-Regularized Open-Vocabulary Change Detection (viability: 7): https://sciencetostartup.com/paper/coregovcd-consistency-regularized-open-vocabulary-change-detection - A training-free framework for open-vocabulary change detection in remote sensing, enabling arbitrary user-defined queries with improved accuracy and spatial coherence.
- A Practical Two-Stage Framework for GPU Resource and Power Prediction in Heterogeneous HPC Systems (viability: 5): https://sciencetostartup.com/paper/a-practical-two-stage-framework-for-gpu-resource-and-power-prediction-in-heterogeneous-hpc-systems - A two-stage framework predicts GPU resource utilization and power consumption in HPC systems for efficient scheduling and power management.
- AstroConcepts: A Large-Scale Multi-Label Classification Corpus for Astrophysics (viability: 4): https://sciencetostartup.com/paper/astroconcepts-a-large-scale-multi-label-classification-corpus-for-astrophysics - A new corpus and evaluation framework for tackling extreme class imbalance in scientific text classification, enabling more robust NLP models for specialized domains.
- Brief Is Better: Non-Monotonic Chain-of-Thought Budget Effects in Function-Calling Language Agents (viability: 7): https://sciencetostartup.com/paper/brief-is-better-non-monotonic-chain-of-thought-budget-effects-in-function-calling-language-agents - A novel CoT prompting strategy that significantly improves function-calling agent accuracy and reliability by focusing on efficient function routing, with a structural guarantee against hallucination.
- Auction-Based Online Policy Adaptation for Evolving Objectives (viability: 4): https://sciencetostartup.com/paper/auction-based-online-policy-adaptation-for-evolving-objectives - An auction-based framework for reinforcement learning agents to dynamically adapt to changing objectives by bidding for action execution.
- AEGIS: Adversarial Entropy-Guided Immune System -- Thermodynamic State Space Models for Zero-Day Network Evasion Detection (viability: 7): https://sciencetostartup.com/paper/aegis-adversarial-entropy-guided-immune-system-thermodynamic-state-space-models-for-zero-day-network-evasion-detection - AEGIS uses thermodynamic physics and non-Euclidean geometry to detect zero-day network evasion attacks with near-perfect accuracy and ultra-low latency.
- TRACE-Bot: Detecting Emerging LLM-Driven Social Bots via Implicit Semantic Representations and AIGC-Enhanced Behavioral Patterns (viability: 7): https://sciencetostartup.com/paper/trace-bot-detecting-emerging-llm-driven-social-bots-via-implicit-semantic-representations-and-aigc-enhanced-behavioral-p - A dual-channel framework that detects emerging LLM-driven social bots by jointly analyzing semantic content and AI-generated behavioral patterns, achieving state-of-the-art accuracy.
- MTI: A Behavior-Based Temperament Profiling System for AI Agents (viability: 3): https://sciencetostartup.com/paper/mti-a-behavior-based-temperament-profiling-system-for-ai-agents - A novel behavior-based system to profile AI agent temperaments across four key axes, independent of model capability.
- PRO-SPECT: Probabilistically Safe Scalable Planning for Energy-Aware Coordinated UAV-UGV Teams in Stochastic Environments (viability: 5): https://sciencetostartup.com/paper/pro-spect-probabilistically-safe-scalable-planning-for-energy-aware-coordinated-uav-ugv-teams-in-stochastic-environments - A risk-bounded planning algorithm for coordinated UAV-UGV teams that uses a UGV as a mobile charging station in stochastic environments.
- Application of parametric Shallow Recurrent Decoder Network to magnetohydrodynamic flows in liquid metal blankets of fusion reactors (viability: 3): https://sciencetostartup.com/paper/application-of-parametric-shallow-recurrent-decoder-network-to-magnetohydrodynamic-flows-in-liquid-metal-blankets-of-fus - A data-driven framework using a recurrent neural network for reconstructing magnetohydrodynamic flows in fusion reactor blankets from sparse measurements.
- GaelEval: Benchmarking LLM Performance for Scottish Gaelic (viability: 7): https://sciencetostartup.com/paper/gaeleval-benchmarking-llm-performance-for-scottish-gaelic - A new benchmark for evaluating LLM performance in Scottish Gaelic, revealing above-human capabilities and identifying key performance gaps.
- Intelligent Cloud Orchestration: A Hybrid Predictive and Heuristic Framework for Cost Optimization (viability: 4): https://sciencetostartup.com/paper/intelligent-cloud-orchestration-a-hybrid-predictive-and-heuristic-framework-for-cost-optimization - A hybrid framework combining predictive and heuristic methods to optimize cloud infrastructure costs and response times.
- SEAL: An Open, Auditable, and Fair Data Generation Framework for AI-Native 6G Networks (viability: 7): https://sciencetostartup.com/paper/seal-an-open-auditable-and-fair-data-generation-framework-for-ai-native-6g-networks - A framework for generating auditable and bias-mitigated synthetic data to accelerate AI development in 6G networks.
- AA-SVD : Anchored and Adaptive SVD for Large Language Model Compression (viability: 5): https://sciencetostartup.com/paper/aa-svd-anchored-and-adaptive-svd-for-large-language-model-compression - A framework for compressing large language models without retraining by accounting for input distribution shifts and refining transformer blocks end-to-end.
- LLM-as-a-Judge for Time Series Explanations (viability: 7): https://sciencetostartup.com/paper/llm-as-a-judge-for-time-series-explanations - Leveraging LLMs to automatically verify the factual correctness of time series explanations without human-defined ground truth.
- Reliable Control-Point Selection for Steering Reasoning in Large Language Models (viability: 7): https://sciencetostartup.com/paper/reliable-control-point-selection-for-steering-reasoning-in-large-language-models - A method to reliably select steering vectors for controlling LLM reasoning by filtering unstable behavioral signals, improving performance on math and reasoning tasks.
- ROS 2-Based LiDAR Perception Framework for Mobile Robots in Dynamic Production Environments, Utilizing Synthetic Data Generation, Transformation-Equivariant 3D Detection and Multi-Object Tracking (viability: 4): https://sciencetostartup.com/paper/ros-2-based-lidar-perception-framework-for-mobile-robots-in-dynamic-production-environments-utilizing-synthetic-data-gen - A LiDAR perception framework for mobile robots in dynamic production environments that improves 6D pose estimation and multi-object tracking using synthetic data and a novel detection method.
- Cross-Modal Visuo-Tactile Object Perception (viability: 7): https://sciencetostartup.com/paper/cross-modal-visuo-tactile-object-perception - A novel visuo-tactile perception filter for robots that learns and evolves object properties over time, improving manipulation efficiency and robustness.
- HyVGGT-VO: Tightly Coupled Hybrid Dense Visual Odometry with Feed-Forward Models (viability: 7): https://sciencetostartup.com/paper/hyvggt-vo-tightly-coupled-hybrid-dense-visual-odometry-with-feed-forward-models - HyVGGT-VO delivers real-time dense visual odometry using a hybrid framework for efficient 3D mapping and pose estimation.
- CASHG: Context-Aware Stylized Online Handwriting Generation (viability: 7): https://sciencetostartup.com/paper/cashg-context-aware-stylized-online-handwriting-generation - Generate context-aware, stylized online handwriting at the sentence level with improved stroke continuity and spacing.
- Prosodic ABX: A Language-Agnostic Method for Measuring Prosodic Contrast in Speech Representations (viability: 4): https://sciencetostartup.com/paper/prosodic-abx-a-language-agnostic-method-for-measuring-prosodic-contrast-in-speech-representations - A language-agnostic method to measure prosodic contrast in speech representations using a novel ABX task and a released dataset.
- LatentUM: Unleashing the Potential of Interleaved Cross-Modal Reasoning via a Latent-Space Unified Model (viability: 7): https://sciencetostartup.com/paper/latentum-unleashing-the-potential-of-interleaved-cross-modal-reasoning-via-a-latent-space-unified-model - LatentUM enables efficient and powerful interleaved cross-modal reasoning by unifying visual understanding and generation in a shared latent space, outperforming state-of-the-art on complex visual tasks.
- GroundVTS: Visual Token Sampling in Multimodal Large Language Models for Video Temporal Grounding (viability: 7): https://sciencetostartup.com/paper/groundvts-visual-token-sampling-in-multimodal-large-language-models-for-video-temporal-grounding - GroundVTS enhances video understanding models by intelligently sampling visual tokens, leading to significant improvements in temporal grounding tasks.
- Optimizing RAG Rerankers with LLM Feedback via Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/optimizing-rag-rerankers-with-llm-feedback-via-reinforcement-learning - Optimize RAG rerankers using LLM feedback for improved answer generation, eliminating the need for human annotations.
- Center-Aware Detection with Swin-based Co-DETR Framework for Cervical Cytology (viability: 7): https://sciencetostartup.com/paper/center-aware-detection-with-swin-based-co-detr-framework-for-cervical-cytology - A winning solution for cervical cancer screening using a Swin-based Co-DETR framework that predicts cell centers for improved detection accuracy.
- FlowSlider: Training-Free Continuous Image Editing via Fidelity-Steering Decomposition (viability: 7): https://sciencetostartup.com/paper/flowslider-training-free-continuous-image-editing-via-fidelity-steering-decomposition - A training-free continuous image editing method that offers slider-style control by decomposing edits into fidelity and steering components, improving reliability and quality without post-training.
- Country-wide, high-resolution monitoring of forest browning with Sentinel-2 (viability: 4): https://sciencetostartup.com/paper/country-wide-high-resolution-monitoring-of-forest-browning-with-sentinel-2 - A scalable approach for country-wide mapping of forest greenness anomalies using satellite data to detect disturbances.
- PLUME: Latent Reasoning Based Universal Multimodal Embedding (viability: 7): https://sciencetostartup.com/paper/plume-latent-reasoning-based-universal-multimodal-embedding - PLUME offers a faster and more efficient universal multimodal embedding by using latent reasoning instead of explicit chain-of-thought, significantly reducing inference time for complex retrieval tasks.
- Mining Instance-Centric Vision-Language Contexts for Human-Object Interaction Detection (viability: 7): https://sciencetostartup.com/paper/mining-instance-centric-vision-language-contexts-for-human-object-interaction-detection - A novel network that integrates vision-language models with object detection to achieve state-of-the-art human-object interaction detection.
- Network Structure in UK Payment Flows: Evidence on Economic Interdependencies and Implications for Real-Time Measurement (viability: 5): https://sciencetostartup.com/paper/network-structure-in-uk-payment-flows-evidence-on-economic-interdependencies-and-implications-for-real-time-measurement - Leverage payment network analysis to provide leading indicators of economic change and improve real-time forecasting accuracy, especially during disruptions.
- Diff-KD: Diffusion-based Knowledge Distillation for Collaborative Perception under Corruptions (viability: 5): https://sciencetostartup.com/paper/diff-kd-diffusion-based-knowledge-distillation-for-collaborative-perception-under-corruptions - A diffusion-based knowledge distillation framework to improve collaborative perception in autonomous systems by actively recovering from sensor and communication corruptions.
- CompassAD: Intent-Driven 3D Affordance Grounding in Functionally Competing Objects (viability: 7): https://sciencetostartup.com/paper/compassad-intent-driven-3d-affordance-grounding-in-functionally-competing-objects - A new benchmark and framework for robots to understand implicit intent and select the correct object for a task in cluttered environments.
- COMPASS: Complete Multimodal Fusion via Proxy Tokens and Shared Spaces for Ubiquitous Sensing (viability: 4): https://sciencetostartup.com/paper/compass-complete-multimodal-fusion-via-proxy-tokens-and-shared-spaces-for-ubiquitous-sensing - A framework for robust multimodal sensing that synthesizes proxy tokens for missing modalities to ensure complete fusion.
- True to Tone? Quantifying Skin Tone Fidelity and Bias in Photographic-to-Virtual Human Pipelines (viability: 4): https://sciencetostartup.com/paper/true-to-tone-quantifying-skin-tone-fidelity-and-bias-in-photographic-to-virtual-human-pipelines - A scalable methodology to evaluate and improve skin tone accuracy and fairness in virtual human rendering pipelines.
- Ouroboros: Dynamic Weight Generation for Recursive Transformers via Input-Conditioned LoRA Modulation (viability: 7): https://sciencetostartup.com/paper/ouroboros-dynamic-weight-generation-for-recursive-transformers-via-input-conditioned-lora-modulation - Ouroboros enables recursive transformers to perform input-dependent transformations at each step, significantly reducing training loss and parameter count while recovering performance lost from layer removal.
- Jagle: Building a Large-Scale Japanese Multimodal Post-Training Dataset for Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/jagle-building-a-large-scale-japanese-multimodal-post-training-dataset-for-vision-language-models - Jagle is the largest Japanese multimodal post-training dataset for vision-language models, enabling improved performance on Japanese tasks without sacrificing English capabilities, with code, models, and dataset released.
- Goose: Anisotropic Speculation Trees for Training-Free Speculative Decoding (viability: 7): https://sciencetostartup.com/paper/goose-anisotropic-speculation-trees-for-training-free-speculative-decoding - GOOSE accelerates LLM inference by building adaptive speculative decoding trees that leverage quality differences in token sources for significant speedups without retraining.
- BidirLM: From Text to Omnimodal Bidirectional Encoders by Adapting and Composing Causal LLMs (viability: 7): https://sciencetostartup.com/paper/bidirlm-from-text-to-omnimodal-bidirectional-encoders-by-adapting-and-composing-causal-llms - A novel method to adapt existing causal LLMs into powerful bidirectional encoders for multimodal tasks, with open-source recipes and benchmark-beating results.
- Tracking the emergence of linguistic structure in self-supervised models learning from speech (viability: 2): https://sciencetostartup.com/paper/tracking-the-emergence-of-linguistic-structure-in-self-supervised-models-learning-from-speech - This research investigates the emergence of linguistic structure within self-supervised speech models, analyzing layerwise patterns and learning trajectories.
- Efficient Reasoning via Thought Compression for Language Segmentation (viability: 7): https://sciencetostartup.com/paper/efficient-reasoning-via-thought-compression-for-language-segmentation - A novel paradigm for efficient multimodal reasoning that significantly reduces computational cost by compressing thought processes without sacrificing performance.
- O-ConNet: Geometry-Aware End-to-End Inference of Over-Constrained Spatial Mechanisms (viability: 7): https://sciencetostartup.com/paper/o-connet-geometry-aware-end-to-end-inference-of-over-constrained-spatial-mechanisms - An end-to-end deep learning framework that infers spatial mechanism parameters and motion trajectories from sparse points, outperforming traditional methods.
- AI in Insurance: Adaptive Questionnaires for Improved Risk Profiling (viability: 6): https://sciencetostartup.com/paper/ai-in-insurance-adaptive-questionnaires-for-improved-risk-profiling - An adaptive questionnaire framework using LLMs and alternative data to personalize insurance underwriting, improve user experience, and reduce fraud.
- IndoorCrowd: A Multi-Scene Dataset for Human Detection, Segmentation, and Tracking with an Automated Annotation Pipeline (viability: 7): https://sciencetostartup.com/paper/indoorcrowd-a-multi-scene-dataset-for-human-detection-segmentation-and-tracking-with-an-automated-annotation-pipeline - A new multi-scene dataset and automated annotation pipeline for indoor human detection, segmentation, and tracking, addressing limitations of existing datasets for real-world applications.
- Rare-Aware Autoencoding: Reconstructing Spatially Imbalanced Data (viability: 7): https://sciencetostartup.com/paper/rare-aware-autoencoding-reconstructing-spatially-imbalanced-data - A novel autoencoding technique that reconstructs fine-grained details in spatially imbalanced image data, crucial for medical imaging and scientific research.
- The Latent Space: Foundation, Evolution, Mechanism, Ability, and Outlook (viability: 2): https://sciencetostartup.com/paper/the-latent-space-foundation-evolution-mechanism-ability-and-outlook - This paper surveys the concept of latent space in language models, exploring its foundations, evolution, mechanisms, abilities, and future outlook.
- Why Gaussian Diffusion Models Fail on Discrete Data? (viability: 4): https://sciencetostartup.com/paper/why-gaussian-diffusion-models-fail-on-discrete-data - This research identifies and proposes solutions for fundamental limitations of diffusion models in generating discrete data, with validated improvements across text, code, and protein generation.
- Systematic Analyses of Reinforcement Learning Controllers in Signalized Urban Corridors (viability: 3): https://sciencetostartup.com/paper/systematic-analyses-of-reinforcement-learning-controllers-in-signalized-urban-corridors - This paper explores the theoretical capacity regions of different reinforcement learning controllers for urban traffic networks, comparing them to a classical baseline.
- APEX: Agent Payment Execution with Policy for Autonomous Agent API Access (viability: 7): https://sciencetostartup.com/paper/apex-agent-payment-execution-with-policy-for-autonomous-agent-api-access - APEX enables autonomous agents to securely access APIs using real-time fiat payments and policy-based spend control, adapting to local regulatory environments.
- ATBench: A Diverse and Realistic Trajectory Benchmark for Long-Horizon Agent Safety (viability: 7): https://sciencetostartup.com/paper/atbench-a-diverse-and-realistic-trajectory-benchmark-for-long-horizon-agent-safety - ATBench provides a realistic and diverse benchmark for evaluating the long-horizon safety of LLM-based agents, enabling better risk assessment and mitigation.
- Bridging Discrete Planning and Continuous Execution for Redundant Robot (viability: 4): https://sciencetostartup.com/paper/bridging-discrete-planning-and-continuous-execution-for-redundant-robot - A framework to improve the continuous execution quality of robot path planning by bridging discrete planning and inverse kinematics.
- Are VLMs Lost Between Sky and Space? LinkS$^2$Bench for UAV-Satellite Dynamic Cross-View Spatial Intelligence (viability: 7): https://sciencetostartup.com/paper/are-vlms-lost-between-sky-and-space-links-2-bench-for-uav-satellite-dynamic-cross-view-spatial-intelligence - A new benchmark and adapter for Vision-Language Models to enable dynamic spatial intelligence between UAVs and satellites, addressing a critical gap in cross-view reasoning for emergency response and security.
- Feature Weighting Improves Pool-Based Sequential Active Learning for Regression (viability: 3): https://sciencetostartup.com/paper/feature-weighting-improves-pool-based-sequential-active-learning-for-regression - Enhances active learning for regression by weighting features to improve sample selection accuracy.
- Demographic Parity Tails for Regression (viability: 4): https://sciencetostartup.com/paper/demographic-parity-tails-for-regression - A new framework for regression fairness that targets specific distribution tails, offering more nuanced and context-sensitive interventions.
- Optimizing Interventions for Agent-Based Infectious Disease Simulations (viability: 5): https://sciencetostartup.com/paper/optimizing-interventions-for-agent-based-infectious-disease-simulations - An AI system that optimizes non-pharmaceutical interventions for infectious disease simulations to minimize societal disruption.
- Decouple and Rectify: Semantics-Preserving Structural Enhancement for Open-Vocabulary Remote Sensing Segmentation (viability: 7): https://sciencetostartup.com/paper/decouple-and-rectify-semantics-preserving-structural-enhancement-for-open-vocabulary-remote-sensing-segmentation - A novel framework for open-vocabulary remote sensing segmentation that decouples semantic and structural features to achieve state-of-the-art accuracy.
- Test-Time Adaptation for Height Completion via Self-Supervised ViT Features and Monocular Foundation Models (viability: 7): https://sciencetostartup.com/paper/test-time-adaptation-for-height-completion-via-self-supervised-vit-features-and-monocular-foundation-models - A test-time adaptation framework that uses foundation models to complete missing height data in digital surface models, improving accuracy and enabling DSM updating.
- $k$NNProxy: Efficient Training-Free Proxy Alignment for Black-Box Zero-Shot LLM-Generated Text Detection (viability: 7): https://sciencetostartup.com/paper/k-nnproxy-efficient-training-free-proxy-alignment-for-black-box-zero-shot-llm-generated-text-detection - A training-free, query-efficient framework for detecting LLM-generated text by adapting existing language models without fine-tuning or costly API interactions.
- Apriel-Reasoner: RL Post-Training for General-Purpose and Efficient Reasoning (viability: 7): https://sciencetostartup.com/paper/apriel-reasoner-rl-post-training-for-general-purpose-and-efficient-reasoning - An LLM post-training method that significantly improves reasoning accuracy and efficiency across diverse tasks with shorter inference traces.
- ProCeedRL: Process Critic with Exploratory Demonstration Reinforcement Learning for LLM Agentic Reasoning (viability: 7): https://sciencetostartup.com/paper/proceedrl-process-critic-with-exploratory-demonstration-reinforcement-learning-for-llm-agentic-reasoning - A reinforcement learning framework that guides LLM agents to avoid reasoning errors in complex multi-turn tasks by actively intervening in exploration.
- ProDiG: Progressive Diffusion-Guided Gaussian Splatting for Aerial to Ground Reconstruction (viability: 7): https://sciencetostartup.com/paper/prodig-progressive-diffusion-guided-gaussian-splatting-for-aerial-to-ground-reconstruction - A diffusion-guided framework for generating realistic ground-level 3D reconstructions from aerial imagery, overcoming extreme viewpoint changes.
- How and why does deep ensemble coupled with transfer learning increase performance in bipolar disorder and schizophrenia classification? (viability: 3): https://sciencetostartup.com/paper/how-and-why-does-deep-ensemble-coupled-with-transfer-learning-increase-performance-in-bipolar-disorder-and-schizophrenia - Investigating the theoretical underpinnings of how transfer learning and deep ensembles improve psychiatric disorder classification.
- GenGait: A Transformer-Based Model for Human Gait Anomaly Detection and Normative Twin Generation (viability: 7): https://sciencetostartup.com/paper/gengait-a-transformer-based-model-for-human-gait-anomaly-detection-and-normative-twin-generation - A label-free Transformer model that detects and corrects gait abnormalities by learning normative human movement patterns.
- MTLSI-Net: A Linear Semantic Interaction Network for Parameter-Efficient Multi-Task Dense Prediction (viability: 7): https://sciencetostartup.com/paper/mtlsi-net-a-linear-semantic-interaction-network-for-parameter-efficient-multi-task-dense-prediction - A novel network architecture that enables efficient multi-task dense prediction by capturing cross-task interactions with linear complexity, outperforming state-of-the-art.
- Resonance4D: Frequency-Domain Motion Supervision for Preset-Free Physical Parameter Learning in 4D Dynamic Physical Scene Simulation (viability: 4): https://sciencetostartup.com/paper/resonance4d-frequency-domain-motion-supervision-for-preset-free-physical-parameter-learning-in-4d-dynamic-physical-scene - A framework for physics-driven 4D dynamic simulation that uses dual-domain motion supervision to improve physical fidelity and reduce computational cost.
- SAFE: Stepwise Atomic Feedback for Error correction in Multi-hop Reasoning (viability: 7): https://sciencetostartup.com/paper/safe-stepwise-atomic-feedback-for-error-correction-in-multi-hop-reasoning - A framework for rigorously verifying and improving multi-hop reasoning in LLMs by grounding each step in a knowledge graph.
- Integrated Identification of Collaborative Robots for Robot Assisted 3D Printing Processes (viability: 4): https://sciencetostartup.com/paper/integrated-identification-of-collaborative-robots-for-robot-assisted-3d-printing-processes - A model-based approach to improve precision and control in robot-assisted 3D printing by identifying system parameters.
- Attention at Rest Stays at Rest: Breaking Visual Inertia for Cognitive Hallucination Mitigation (viability: 7): https://sciencetostartup.com/paper/attention-at-rest-stays-at-rest-breaking-visual-inertia-for-cognitive-hallucination-mitigation - A training-free method to improve multimodal LLM reasoning by dynamically adjusting visual attention, reducing cognitive hallucinations.
- SenseMath: Do LLMs Have Number Sense? Evaluating Shortcut Use, Judgment, and Generation (viability: 7): https://sciencetostartup.com/paper/sensemath-do-llms-have-number-sense-evaluating-shortcut-use-judgment-and-generation - A benchmark to evaluate LLMs' true number sense, revealing their reliance on learned procedures over structural understanding, with potential for targeted model improvement.
- Curia-2: Scaling Self-Supervised Learning for Radiology Foundation Models (viability: 7): https://sciencetostartup.com/paper/curia-2-scaling-self-supervised-learning-for-radiology-foundation-models - A foundation model for radiology that significantly improves representation quality for CT and MRI analysis, outperforming existing models on vision-focused tasks and competing on clinical tasks.
- World Action Verifier: Self-Improving World Models via Forward-Inverse Asymmetry (viability: 7): https://sciencetostartup.com/paper/world-action-verifier-self-improving-world-models-via-forward-inverse-asymmetry - A self-improving world model framework that verifies its own predictions by decomposing state prediction into plausibility and reachability, leading to significantly improved sample efficiency and downstream policy performance.
- RuleForge: Automated Generation and Validation for Web Vulnerability Detection at Scale (viability: 7): https://sciencetostartup.com/paper/ruleforge-automated-generation-and-validation-for-web-vulnerability-detection-at-scale - RuleForge automates the generation of security detection rules from vulnerability descriptions using LLMs, significantly reducing false positives and improving detection capacity.
- Interactive Tracking: A Human-in-the-Loop Paradigm with Memory-Augmented Adaptation (viability: 7): https://sciencetostartup.com/paper/interactive-tracking-a-human-in-the-loop-paradigm-with-memory-augmented-adaptation - A new paradigm for visual tracking that allows real-time human guidance via natural language, with a benchmark and baseline model to enable adaptive tracking systems.
- NearID: Identity Representation Learning via Near-identity Distractors (viability: 8): https://sciencetostartup.com/paper/nearid-identity-representation-learning-via-near-identity-distractors - NearID offers a robust identity verification system that isolates identity signals for enhanced personalization and image editing.
- SDesc3D: Towards Layout-Aware 3D Indoor Scene Generation from Short Descriptions (viability: 7): https://sciencetostartup.com/paper/sdesc3d-towards-layout-aware-3d-indoor-scene-generation-from-short-descriptions - Generate detailed and physically plausible 3D indoor scenes from short text descriptions by leveraging multi-view structural priors and functional reasoning.
- Ego-Grounding for Personalized Question-Answering in Egocentric Videos (viability: 7): https://sciencetostartup.com/paper/ego-grounding-for-personalized-question-answering-in-egocentric-videos - A new dataset and benchmark for personalized question-answering in egocentric videos, revealing significant limitations in current multimodal LLMs for understanding the camera wearer.
- Do We Need Bigger Models for Science? Task-Aware Retrieval with Small Language Models (viability: 7): https://sciencetostartup.com/paper/do-we-need-bigger-models-for-science-task-aware-retrieval-with-small-language-models - A retrieval-augmented framework using compact language models and task-aware routing to build reproducible and accessible scholarly assistants.
- Automated Prostate Gland Segmentation in MRI Using nnU-Net (viability: 7): https://sciencetostartup.com/paper/automated-prostate-gland-segmentation-in-mri-using-nnu-net - A containerized deep learning tool for highly accurate prostate gland segmentation in MRI, outperforming general-purpose models.
- Abnormal Head Movements in Neurological Conditions: A Knowledge-Based Dataset with Application to Cervical Dystonia (viability: 7): https://sciencetostartup.com/paper/abnormal-head-movements-in-neurological-conditions-a-knowledge-based-dataset-with-application-to-cervical-dystonia - A novel knowledge-based dataset of abnormal head movements extracted from medical literature, enabling AI-driven diagnostic tools for neurological conditions.
- Generalization Bounds and Statistical Guarantees for Multi-Task and Multiple Operator Learning with MNO Networks (viability: 2): https://sciencetostartup.com/paper/generalization-bounds-and-statistical-guarantees-for-multi-task-and-multiple-operator-learning-with-mno-networks - Develops theoretical generalization bounds for multiple operator learning architectures, specifically MNO networks, to understand sample complexity for unseen operator instances.
- MAVFusion: Efficient Infrared and Visible Video Fusion via Motion-Aware Sparse Interaction (viability: 7): https://sciencetostartup.com/paper/mavfusion-efficient-infrared-and-visible-video-fusion-via-motion-aware-sparse-interaction - An efficient video fusion framework that leverages motion-aware sparse interaction to combine infrared and visible imagery for enhanced detail and temporal consistency.
- Diagnosing Translated Benchmarks: An Automated Quality Assurance Study of the EU20 Benchmark Suite (viability: 7): https://sciencetostartup.com/paper/diagnosing-translated-benchmarks-an-automated-quality-assurance-study-of-the-eu20-benchmark-suite - Automated quality assurance for machine-translated benchmarks to improve reliability and prioritize human review.
- Qiana: A First-Order Formalism to Quantify over Contexts and Formulas with Temporality (viability: 2): https://sciencetostartup.com/paper/qiana-a-first-order-formalism-to-quantify-over-contexts-and-formulas-with-temporality - A formal logic system for reasoning about context-dependent and temporal information.
- Learn by Surprise, Commit by Proof (viability: 3): https://sciencetostartup.com/paper/learn-by-surprise-commit-by-proof - A self-gated post-training framework for autonomous knowledge acquisition in language models, learning only what the model doesn't know and sharpening existing knowledge to reduce hallucinations.
- annbatch unlocks terabyte-scale training of biological data in anndata (viability: 7): https://sciencetostartup.com/paper/annbatch-unlocks-terabyte-scale-training-of-biological-data-in-anndata - An out-of-core data loader for biological datasets that significantly speeds up machine learning model training by addressing data access bottlenecks.
- A Self supervised learning framework for imbalanced medical imaging datasets (viability: 5): https://sciencetostartup.com/paper/a-self-supervised-learning-framework-for-imbalanced-medical-imaging-datasets - A self-supervised learning framework that improves medical image classification accuracy on imbalanced and scarce datasets.
- PAC-Bayesian Reward-Certified Outcome Weighted Learning (viability: 4): https://sciencetostartup.com/paper/pac-bayesian-reward-certified-outcome-weighted-learning - A novel PAC-Bayesian framework for robustly estimating individualized treatment rules by accounting for reward uncertainty, leading to improved treatment regime selection.
- Physics-Informed Transformer for Multi-Band Channel Frequency Response Reconstruction (viability: 4): https://sciencetostartup.com/paper/physics-informed-transformer-for-multi-band-channel-frequency-response-reconstruction - A physics-informed Transformer reconstructs wireless channel frequency responses from fragmented spectrum data, outperforming classical methods in accuracy.
- A Novel Theoretical Analysis for Clustering Heteroscedastic Gaussian Data without Knowledge of the Number of Clusters (viability: 3): https://sciencetostartup.com/paper/a-novel-theoretical-analysis-for-clustering-heteroscedastic-gaussian-data-without-knowledge-of-the-number-of-clusters - A theoretical framework for clustering heteroscedastic Gaussian data with a novel cost function and Wald kernel, leading to a new algorithm called CENTRE-X.
- Captioning Daily Activity Images in Early Childhood Education: Benchmark and Algorithm (viability: 7): https://sciencetostartup.com/paper/captioning-daily-activity-images-in-early-childhood-education-benchmark-and-algorithm - A new benchmark and hybrid training framework for generating precise, domain-specific image captions in early childhood education, outperforming existing models.
- Probabilistic classification from possibilistic data: computing Kullback-Leibler projection with a possibility distribution (viability: 4): https://sciencetostartup.com/paper/probabilistic-classification-from-possibilistic-data-computing-kullback-leibler-projection-with-a-possibility-distributi - A novel method for training multi-class classifiers using graded plausibility supervision, improving predictive performance by projecting predictions onto admissible probability distributions.
- How to measure the optimality of word or gesture order with respect to the principle of swap distance minimization (viability: 3): https://sciencetostartup.com/paper/how-to-measure-the-optimality-of-word-or-gesture-order-with-respect-to-the-principle-of-swap-distance-minimization - A theoretical framework to measure the optimality of word and gesture order based on swap distance minimization.
- Architectural Implications of the UK Cyber Security and Resilience Bill (viability: 4): https://sciencetostartup.com/paper/architectural-implications-of-the-uk-cyber-security-and-resilience-bill - A framework for achieving compliance with the UK Cyber Security and Resilience Bill using Zero Trust Architecture.
- Reliable News or Propagandist News? A Neurosymbolic Model Using Genre, Topic, and Persuasion Techniques to Improve Robustness in Classification (viability: 4): https://sciencetostartup.com/paper/reliable-news-or-propagandist-news-a-neurosymbolic-model-using-genre-topic-and-persuasion-techniques-to-improve-robustne - A neurosymbolic model enhances news classification robustness by combining text embeddings with symbolic features like genre, topic, and persuasion techniques.
- Rethinking Representations for Cross-Domain Infrared Small Target Detection: A Generalizable Perspective from the Frequency Domain (viability: 7): https://sciencetostartup.com/paper/rethinking-representations-for-cross-domain-infrared-small-target-detection-a-generalizable-perspective-from-the-frequen - A novel network for infrared small target detection that generalizes across different domains by leveraging frequency domain analysis and phase rectification.
- BraiNCA: brain-inspired neural cellular automata and applications to morphogenesis and motor control (viability: 3): https://sciencetostartup.com/paper/brainca-brain-inspired-neural-cellular-automata-and-applications-to-morphogenesis-and-motor-control - A brain-inspired neural cellular automata with attention for improved robustness and learning speed in self-organization tasks.
- Woosh: A Sound Effects Foundation Model (viability: 8): https://sciencetostartup.com/paper/woosh-a-sound-effects-foundation-model - Harness Sony AI's 'Woosh' for groundbreaking, high-quality sound effects generation for multimedia solutions.
- ImplicitBBQ: Benchmarking Implicit Bias in Large Language Models through Characteristic Based Cues (viability: 7): https://sciencetostartup.com/paper/implicitbbq-benchmarking-implicit-bias-in-large-language-models-through-characteristic-based-cues - A new benchmark and dataset to uncover and measure implicit biases in LLMs, enabling targeted mitigation strategies.
- Is Clinical Text Enough? A Multimodal Study on Mortality Prediction in Heart Failure Patients (viability: 4): https://sciencetostartup.com/paper/is-clinical-text-enough-a-multimodal-study-on-mortality-prediction-in-heart-failure-patients - Develops a multimodal transformer model for improved short-term mortality prediction in heart failure patients by combining clinical text and structured EHR data.
- Learning Spatial Structure from Pre-Beamforming Per-Antenna Range-Doppler Radar Data via Visibility-Aware Cross-Modal Supervision (viability: 3): https://sciencetostartup.com/paper/learning-spatial-structure-from-pre-beamforming-per-antenna-range-doppler-radar-data-via-visibility-aware-cross-modal-su - Learning spatial structure directly from raw radar data for automotive perception.
- SURE: Synergistic Uncertainty-aware Reasoning for Multimodal Emotion Recognition in Conversations (viability: 7): https://sciencetostartup.com/paper/sure-synergistic-uncertainty-aware-reasoning-for-multimodal-emotion-recognition-in-conversations - A framework for robust multimodal emotion recognition in conversations that leverages uncertainty and iterative reasoning to outperform existing methods.
- Enhancing Medical Visual Grounding via Knowledge-guided Spatial Prompts (viability: 7): https://sciencetostartup.com/paper/enhancing-medical-visual-grounding-via-knowledge-guided-spatial-prompts - A framework that enhances the precision of medical image localization from radiology reports by integrating medical knowledge and attention mechanisms.
- The Rank and Gradient Lost in Non-stationarity: Sample Weight Decay for Mitigating Plasticity Loss in Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/the-rank-and-gradient-lost-in-non-stationarity-sample-weight-decay-for-mitigating-plasticity-loss-in-reinforcement-learn - A theoretical framework and sample weighting technique to improve continuous learning in deep reinforcement learning agents.
- Night Eyes: A Reproducible Framework for Constellation-Based Corneal Reflection Matching (viability: 7): https://sciencetostartup.com/paper/night-eyes-a-reproducible-framework-for-constellation-based-corneal-reflection-matching - A reproducible framework for robust corneal reflection matching in eye tracking systems.
- Lifting Unlabeled Internet-level Data for 3D Scene Understanding (viability: 7): https://sciencetostartup.com/paper/lifting-unlabeled-internet-level-data-for-3d-scene-understanding - Leveraging unlabeled internet videos to automatically generate training data for 3D scene understanding models, improving zero-shot performance.
- From Component Manipulation to System Compromise: Understanding and Detecting Malicious MCP Servers (viability: 7): https://sciencetostartup.com/paper/from-component-manipulation-to-system-compromise-understanding-and-detecting-malicious-mcp-servers - A novel component-centric approach and behavioral detector for identifying and mitigating malicious LLM tool integrations, outperforming existing methods.
- Combating Data Laundering in LLM Training (viability: 7): https://sciencetostartup.com/paper/combating-data-laundering-in-llm-training - A novel method to detect and combat data laundering in LLM training by inferring transformations and synthesizing detection queries.
- Light-ResKAN: A Parameter-Sharing Lightweight KAN with Gram Polynomials for Efficient SAR Image Recognition (viability: 7): https://sciencetostartup.com/paper/light-reskan-a-parameter-sharing-lightweight-kan-with-gram-polynomials-for-efficient-sar-image-recognition - A parameter-sharing lightweight KAN model using Gram Polynomials for efficient and high-precision SAR image recognition on edge devices.
- FTPFusion: Frequency-Aware Infrared and Visible Video Fusion with Temporal Perturbation (viability: 7): https://sciencetostartup.com/paper/ftpfusion-frequency-aware-infrared-and-visible-video-fusion-with-temporal-perturbation - A frequency-aware method for fusing infrared and visible videos that enhances temporal stability and spatial detail for surveillance and low-light monitoring.
- Enhancing the Reliability of Medical AI through Expert-guided Uncertainty Modeling (viability: 7): https://sciencetostartup.com/paper/enhancing-the-reliability-of-medical-ai-through-expert-guided-uncertainty-modeling - A novel AI approach enhances medical diagnostic reliability by leveraging expert disagreement to improve uncertainty estimation, reducing risks in healthcare.
- Bayesian Elicitation with LLMs: Model Size Helps, Extra "Reasoning" Doesn't Always (viability: 4): https://sciencetostartup.com/paper/bayesian-elicitation-with-llms-model-size-helps-extra-reasoning-doesn-t-always - This research investigates the accuracy and uncertainty estimation of LLMs for Bayesian elicitation, finding that model size is key but reasoning effort is not, and that statistical correction is needed for reliable decision-making.
- SHARC: Reference point driven Spherical Harmonic Representation for Complex Shapes (viability: 7): https://sciencetostartup.com/paper/sharc-reference-point-driven-spherical-harmonic-representation-for-complex-shapes - A novel framework for synthesizing complex 3D shapes using Spherical Harmonic representations anchored by reference points, outperforming existing methods in accuracy and efficiency.
- ProVG: Progressive Visual Grounding via Language Decoupling for Remote Sensing Imagery (viability: 7): https://sciencetostartup.com/paper/provg-progressive-visual-grounding-via-language-decoupling-for-remote-sensing-imagery - A novel framework for precise object localization in remote sensing imagery by decoupling language into distinct components for progressive, fine-grained alignment.
- LI-DSN: A Layer-wise Interactive Dual-Stream Network for EEG Decoding (viability: 7): https://sciencetostartup.com/paper/li-dsn-a-layer-wise-interactive-dual-stream-network-for-eeg-decoding - A novel dual-stream neural network with layer-wise interaction and attention mechanisms for significantly improved EEG decoding performance across various BCI applications.
- Low-Effort Jailbreak Attacks Against Text-to-Image Safety Filters (viability: 7): https://sciencetostartup.com/paper/low-effort-jailbreak-attacks-against-text-to-image-safety-filters - Develops novel prompt-based strategies to bypass safety filters in text-to-image models, demonstrating a critical gap in current generative AI security.
- GS^2: Graph-based Spatial Distribution Optimization for Compact 3D Gaussian Splatting (viability: 7): https://sciencetostartup.com/paper/gs-2-graph-based-spatial-distribution-optimization-for-compact-3d-gaussian-splatting - A novel method to significantly reduce the memory footprint of 3D Gaussian Splatting while improving rendering quality, making real-time 3D reconstruction more practical.
- A3R: Agentic Affordance Reasoning via Cross-Dimensional Evidence in 3D Gaussian Scenes (viability: 7): https://sciencetostartup.com/paper/a3r-agentic-affordance-reasoning-via-cross-dimensional-evidence-in-3d-gaussian-scenes - An agentic framework that sequentially acquires 3D geometric and 2D semantic evidence to improve fine-grained affordance reasoning in complex 3D scenes.
- HieraVid: Hierarchical Token Pruning for Fast Video Large Language Models (viability: 7): https://sciencetostartup.com/paper/hieravid-hierarchical-token-pruning-for-fast-video-large-language-models - HieraVid drastically reduces computational cost for Video LLMs by intelligently pruning video tokens hierarchically, achieving state-of-the-art performance with significantly fewer resources.
- DDCL-INCRT: A Self-Organising Transformer with Hierarchical Prototype Structure (Theoretical Foundations) (viability: 2): https://sciencetostartup.com/paper/ddcl-incrt-a-self-organising-transformer-with-hierarchical-prototype-structure-theoretical-foundations - A theoretical framework for self-organizing transformer architectures that minimize computational resources by dynamically determining their structure during training.
- Robust Graph Representation Learning via Adaptive Spectral Contrast (viability: 7): https://sciencetostartup.com/paper/robust-graph-representation-learning-via-adaptive-spectral-contrast - A novel framework for robust graph representation learning that adaptively gates spectral components to achieve state-of-the-art performance on diverse graph types.
- Topology-Hiding Connectivity-Assurance for QKD Inter-Networking (viability: 2): https://sciencetostartup.com/paper/topology-hiding-connectivity-assurance-for-qkd-inter-networking - A protocol for proving secure connections in quantum key distribution networks without revealing network topology.
- Towards Intrinsically Calibrated Uncertainty Quantification in Industrial Data-Driven Models via Diffusion Sampler (viability: 7): https://sciencetostartup.com/paper/towards-intrinsically-calibrated-uncertainty-quantification-in-industrial-data-driven-models-via-diffusion-sampler - A diffusion-based framework for intrinsically calibrated uncertainty quantification in industrial data-driven models, improving safety and decision-making.
- GeoAI Agency Primitives (viability: 7): https://sciencetostartup.com/paper/geoai-agency-primitives - Develops agency primitives to bridge foundation models with GIS workflows, enabling practical agentic assistance for geospatial practitioners.
- MAR-MAER: Metric-Aware and Ambiguity-Adaptive Autoregressive Image Generation (viability: 7): https://sciencetostartup.com/paper/mar-maer-metric-aware-and-ambiguity-adaptive-autoregressive-image-generation - A hierarchical autoregressive image generation framework that improves metric consistency and handles ambiguous prompts by aligning internal representations with human quality metrics and incorporating controlled randomness.
- Posterior Optimization with Clipped Objective for Bridging Efficiency and Stability in Generative Policy Learning (viability: 4): https://sciencetostartup.com/paper/posterior-optimization-with-clipped-objective-for-bridging-efficiency-and-stability-in-generative-policy-learning - POCO enhances generative policy learning by maintaining stability and efficiency in robotic manipulation through posterior optimization.
- Combining Boundary Supervision and Segment-Level Regularization for Fine-Grained Action Segmentation (viability: 5): https://sciencetostartup.com/paper/combining-boundary-supervision-and-segment-level-regularization-for-fine-grained-action-segmentation - A lightweight dual-loss training framework that significantly improves fine-grained action segmentation quality with minimal architectural changes, making complex models more practical.
- Efficient Constraint Generation for Stochastic Shortest Path Problems (viability: 3): https://sciencetostartup.com/paper/efficient-constraint-generation-for-stochastic-shortest-path-problems - A new algorithm for stochastic shortest path problems that significantly reduces computation by intelligently pruning actions based on heuristics, leading to faster problem solving.
- Beyond Detection: Ethical Foundations for Automated Dyslexic Error Attribution (viability: 5): https://sciencetostartup.com/paper/beyond-detection-ethical-foundations-for-automated-dyslexic-error-attribution - An AI system that accurately attributes spelling errors to dyslexic writers, with a strong focus on ethical deployment guidelines for educational contexts.
- From Guessing to Placeholding: A Cost-Theoretic Framework for Uncertainty-Aware Code Completion (viability: 7): https://sciencetostartup.com/paper/from-guessing-to-placeholding-a-cost-theoretic-framework-for-uncertainty-aware-code-completion - An AI code completion framework that strategically uses placeholders to reduce user editing costs and improve suggestion accuracy.
- Semantic Richness or Geometric Reasoning? The Fragility of VLM's Visual Invariance (viability: 4): https://sciencetostartup.com/paper/semantic-richness-or-geometric-reasoning-the-fragility-of-vlm-s-visual-invariance - This research reveals a fundamental geometric reasoning gap in current Vision-Language Models, highlighting a need for improved spatial invariance in future multimodal systems.
- CANDI: Curated Test-Time Adaptation for Multivariate Time-Series Anomaly Detection Under Distribution Shift (viability: 7): https://sciencetostartup.com/paper/candi-curated-test-time-adaptation-for-multivariate-time-series-anomaly-detection-under-distribution-shift - A novel test-time adaptation framework for multivariate time-series anomaly detection that significantly improves performance under distribution shifts by selectively adapting to false positives.
- FaCT-GS: Fast and Scalable CT Reconstruction with Gaussian Splatting (viability: 7): https://sciencetostartup.com/paper/fact-gs-fast-and-scalable-ct-reconstruction-with-gaussian-splatting - Accelerate CT reconstruction using optimized Gaussian Splatting for faster and more flexible medical imaging.
- Investigating Permutation-Invariant Discrete Representation Learning for Spatially Aligned Images (viability: 4): https://sciencetostartup.com/paper/investigating-permutation-invariant-discrete-representation-learning-for-spatially-aligned-images - A novel autoencoder learns position-independent image representations for faster, direct image interpolation and synthesis.
- Retrieval-aligned Tabular Foundation Models Enable Robust Clinical Risk Prediction in Electronic Health Records Under Real-world Constraints (viability: 7): https://sciencetostartup.com/paper/retrieval-aligned-tabular-foundation-models-enable-robust-clinical-risk-prediction-in-electronic-health-records-under-re - A retrieval-aligned framework for robust clinical risk prediction from electronic health records that overcomes data complexity and imbalance.
- Not All Tokens See Equally: Perception-Grounded Policy Optimization for Large Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/not-all-tokens-see-equally-perception-grounded-policy-optimization-for-large-vision-language-models - A novel fine-tuning framework that significantly improves multimodal reasoning in LLMs by focusing learning signals on visually-dependent tokens.
- PLOT: Enhancing Preference Learning via Optimal Transport (viability: 7): https://sciencetostartup.com/paper/plot-enhancing-preference-learning-via-optimal-transport - A novel optimal transport-based loss function for LLM alignment that improves performance and stability by modeling global token relationships.
- Semantic Segmentation of Textured Non-manifold 3D Meshes using Transformers (viability: 7): https://sciencetostartup.com/paper/semantic-segmentation-of-textured-non-manifold-3d-meshes-using-transformers - A texture-aware transformer for semantic segmentation of 3D meshes, outperforming existing methods on urban and cultural heritage datasets.
- Ranking-Guided Semi-Supervised Domain Adaptation for Severity Classification (viability: 5): https://sciencetostartup.com/paper/ranking-guided-semi-supervised-domain-adaptation-for-severity-classification - A novel semi-supervised domain adaptation method for medical image severity classification that aligns class-specific rank score distributions.
- Language-Pretraining-Induced Bias: A Strong Foundation for General Vision Tasks (viability: 3): https://sciencetostartup.com/paper/language-pretraining-induced-bias-a-strong-foundation-for-general-vision-tasks - Adapting large language models for vision tasks through a novel bridge training method.
- Physics Informed Reinforcement Learning with Gibbs Priors for Topology Control in Power Grids (viability: 7): https://sciencetostartup.com/paper/physics-informed-reinforcement-learning-with-gibbs-priors-for-topology-control-in-power-grids - A physics-informed reinforcement learning system that drastically speeds up power grid topology control by intelligently predicting risks and narrowing down action choices.
- SafeRoPE: Risk-specific Head-wise Embedding Rotation for Safe Generation in Rectified Flow Transformers (viability: 7): https://sciencetostartup.com/paper/saferope-risk-specific-head-wise-embedding-rotation-for-safe-generation-in-rectified-flow-transformers - A lightweight framework for mitigating unsafe content in text-to-image models by precisely rotating head-wise embeddings.
- STRIVE: Structured Spatiotemporal Exploration for Reinforcement Learning in Video Question Answering (viability: 7): https://sciencetostartup.com/paper/strive-structured-spatiotemporal-exploration-for-reinforcement-learning-in-video-question-answering - A reinforcement learning framework that improves video question answering by generating structured spatiotemporal variants of videos to enrich reward signals and stabilize learning.
- Graph Neural Operator Towards Edge Deployability and Portability for Sparse-to-Dense, Real-Time Virtual Sensing on Irregular Grids (viability: 8): https://sciencetostartup.com/paper/graph-neural-operator-towards-edge-deployability-and-portability-for-sparse-to-dense-real-time-virtual-sensing-on-irregu - VIRSO is a graph-based neural operator for real-time, sparse-to-dense virtual sensing on irregular grids, optimized for edge deployment with low power and latency.
- A deep learning pipeline for PAM50 subtype classification using histopathology images and multi-objective patch selection (viability: 7): https://sciencetostartup.com/paper/a-deep-learning-pipeline-for-pam50-subtype-classification-using-histopathology-images-and-multi-objective-patch-selectio - A deep learning framework predicts breast cancer subtypes from histopathology images, reducing the need for costly molecular assays and aiding clinical decisions.
- PTC-Depth: Pose-Refined Monocular Depth Estimation with Temporal Consistency (viability: 7): https://sciencetostartup.com/paper/ptc-depth-pose-refined-monocular-depth-estimation-with-temporal-consistency - A monocular depth estimation framework that uses wheel odometry for temporally consistent and accurate depth predictions, addressing jitter and estimation failures in dynamic environments.
- Learning in Prophet Inequalities with Noisy Observations (viability: 4): https://sciencetostartup.com/paper/learning-in-prophet-inequalities-with-noisy-observations - Develops algorithms for online decision-making with noisy observations by integrating learning and decision-making.
- DEFT: Distribution-guided Efficient Fine-Tuning for Human Alignment (viability: 7): https://sciencetostartup.com/paper/deft-distribution-guided-efficient-fine-tuning-for-human-alignment - An efficient framework for aligning large language models with human values by intelligently filtering data and guiding output distribution, improving performance and generalization while reducing training time.
- Taming CATS: Controllable Automatic Text Simplification through Instruction Fine-Tuning with Control Tokens (viability: 7): https://sciencetostartup.com/paper/taming-cats-controllable-automatic-text-simplification-through-instruction-fine-tuning-with-control-tokens - A framework for fine-tuning open-source LLMs with control tokens to achieve user-tailored text simplification, demonstrating competitive performance with smaller models.
- GardenDesigner: Encoding Aesthetic Principles into Jiangnan Garden Construction via a Chain of Agents (viability: 8): https://sciencetostartup.com/paper/gardendesigner-encoding-aesthetic-principles-into-jiangnan-garden-construction-via-a-chain-of-agents - A framework that uses a chain of agents to automatically construct aesthetically pleasing Jiangnan gardens from text input, enabling rapid digital asset creation for film, games, and tourism.
- Preferential Bayesian Optimization with Crash Feedback (viability: 7): https://sciencetostartup.com/paper/preferential-bayesian-optimization-with-crash-feedback - A Bayesian optimization framework that reduces experimental crashes and improves data efficiency in robotics by incorporating user feedback on preferences and crashes.
- Bridging Deep Learning and Integer Linear Programming: A Predictive-to-Prescriptive Framework for Supply Chain Analytics (viability: 7): https://sciencetostartup.com/paper/bridging-deep-learning-and-integer-linear-programming-a-predictive-to-prescriptive-framework-for-supply-chain-analytics - A predictive-to-prescriptive framework combining deep learning forecasting with integer linear programming for cost-optimal supply chain logistics.
- Domain-constrained knowledge representation: A modal framework (viability: 3): https://sciencetostartup.com/paper/domain-constrained-knowledge-representation-a-modal-framework - A new framework for knowledge graphs that treats domain as a core part of representation, enabling disambiguation and cross-domain relations.
- Dual-Attention Based 3D Channel Estimation (viability: 3): https://sciencetostartup.com/paper/dual-attention-based-3d-channel-estimation - A deep learning approach for more accurate 3D channel estimation in MIMO systems.
- FSKD: Monocular Forest Structure Inference via LiDAR-to-RGBI Knowledge Distillation (viability: 5): https://sciencetostartup.com/paper/fskd-monocular-forest-structure-inference-via-lidar-to-rgbi-knowledge-distillation - A knowledge distillation framework that uses RGBI imagery to infer high-resolution forest structure data, making detailed ecosystem monitoring more accessible and scalable.
- DriveDreamer-Policy: A Geometry-Grounded World-Action Model for Unified Generation and Planning (viability: 7): https://sciencetostartup.com/paper/drivedreamer-policy-a-geometry-grounded-world-action-model-for-unified-generation-and-planning - A geometry-grounded world-action model for unified driving simulation, future prediction, and motion planning.
- Hidden Meanings in Plain Sight: RebusBench for Evaluating Cognitive Visual Reasoning (viability: 7): https://sciencetostartup.com/paper/hidden-meanings-in-plain-sight-rebusbench-for-evaluating-cognitive-visual-reasoning - A new benchmark and evaluation framework for visual reasoning tasks that current state-of-the-art models fail at, highlighting a critical gap in cognitive integration.
- Cosine-Normalized Attention for Hyperspectral Image Classification (viability: 7): https://sciencetostartup.com/paper/cosine-normalized-attention-for-hyperspectral-image-classification - A novel cosine-normalized attention mechanism for Transformers significantly improves hyperspectral image classification accuracy under extreme data scarcity, outperforming existing state-of-the-art models.
- FourierMoE: Fourier Mixture-of-Experts Adaptation of Large Language Models (viability: 7): https://sciencetostartup.com/paper/fouriermoe-fourier-mixture-of-experts-adaptation-of-large-language-models - FourierMoE enables parameter-efficient LLM adaptation by leveraging spectral domain experts, outperforming existing methods across numerous benchmarks with fewer trainable parameters.
- Control-DINO: Feature Space Conditioning for Controllable Image-to-Video Diffusion (viability: 7): https://sciencetostartup.com/paper/control-dino-feature-space-conditioning-for-controllable-image-to-video-diffusion - Control image-to-video diffusion models for tasks like domain transfer and 3D scene generation by conditioning on disentangled appearance features.
- Realistic Lip Motion Generation Based on 3D Dynamic Viseme and Coarticulation Modeling for Human-Robot Interaction (viability: 7): https://sciencetostartup.com/paper/realistic-lip-motion-generation-based-on-3d-dynamic-viseme-and-coarticulation-modeling-for-human-robot-interaction - A framework for realistic, speech-driven lip motion generation for humanoid robots, enabling more natural human-robot interaction.
- LiveMathematicianBench: A Live Benchmark for Mathematician-Level Reasoning with Proof Sketches (viability: 7): https://sciencetostartup.com/paper/livemathematicianbench-a-live-benchmark-for-mathematician-level-reasoning-with-proof-sketches - A dynamic benchmark for evaluating LLM mathematical reasoning capabilities on cutting-edge research, designed to prevent data contamination and assess genuine understanding.
- Analysis of Efficient Transmission Methods of Grid Maps for Intelligent Vehicles (viability: 4): https://sciencetostartup.com/paper/analysis-of-efficient-transmission-methods-of-grid-maps-for-intelligent-vehicles - Develops a communication pipeline for efficient transmission of grid-map data for intelligent vehicles.
- Spike-PTSD: A Bio-Plausible Adversarial Example Attack on Spiking Neural Networks via PTSD-Inspired Spike Scaling (viability: 7): https://sciencetostartup.com/paper/spike-ptsd-a-bio-plausible-adversarial-example-attack-on-spiking-neural-networks-via-ptsd-inspired-spike-scaling - A biologically inspired adversarial attack framework that systematically compromises Spiking Neural Network robustness for security-critical systems.
- Ultrasound-CLIP: Semantic-Aware Contrastive Pre-training for Ultrasound Image-Text Understanding (viability: 7): https://sciencetostartup.com/paper/ultrasound-clip-semantic-aware-contrastive-pre-training-for-ultrasound-image-text-understanding - A semantic-aware contrastive learning framework and dataset for improved understanding of ultrasound images and their associated diagnostic text.
- Unifying UAV Cross-View Geo-Localization via 3D Geometric Perception (viability: 7): https://sciencetostartup.com/paper/unifying-uav-cross-view-geo-localization-via-3d-geometric-perception - A geometry-aware framework for precise UAV geo-localization in GNSS-denied environments by unifying place recognition and pose estimation.
- Detecting Toxic Language: Ontology and BERT-based Approaches for Bulgarian Text (viability: 5): https://sciencetostartup.com/paper/detecting-toxic-language-ontology-and-bert-based-approaches-for-bulgarian-text - A BERT-based system for nuanced toxic language detection in Bulgarian that preserves access to essential information.
- Dense Point-to-Mask Optimization with Reinforced Point Selection for Crowd Instance Segmentation (viability: 7): https://sciencetostartup.com/paper/dense-point-to-mask-optimization-with-reinforced-point-selection-for-crowd-instance-segmentation - A novel framework for dense crowd instance segmentation using point annotations, outperforming existing methods and improving counting accuracy.
- DDCL: Deep Dual Competitive Learning: A Differentiable End-to-End Framework for Unsupervised Prototype-Based Representation Learning (viability: 3): https://sciencetostartup.com/paper/ddcl-deep-dual-competitive-learning-a-differentiable-end-to-end-framework-for-unsupervised-prototype-based-representatio - A novel differentiable framework for unsupervised prototype-based representation learning that unifies feature extraction and cluster assignment.
- AeroTherm-GPT: A Verification-Centered LLM Framework for Thermal Protection System Engineering Workflows (viability: 7): https://sciencetostartup.com/paper/aerotherm-gpt-a-verification-centered-llm-framework-for-thermal-protection-system-engineering-workflows - A specialized LLM agent for thermal protection system design that iteratively generates, validates, and repairs simulation artifacts to ensure engineering workflow compliance.
- Setup-Independent Full Projector Compensation (viability: 8): https://sciencetostartup.com/paper/setup-independent-full-projector-compensation - A framework for projector compensation that generalizes to unseen setups without retraining, enabled by a large dataset and a novel co-adaptive geometry and photometry correction approach.
- From BM25 to Corrective RAG: Benchmarking Retrieval Strategies for Text-and-Table Documents (viability: 7): https://sciencetostartup.com/paper/from-bm25-to-corrective-rag-benchmarking-retrieval-strategies-for-text-and-table-documents - This research benchmarks and optimizes retrieval strategies for RAG systems handling complex text-and-table documents, offering a path to significantly improved question-answering accuracy.
- Solving the Two-dimensional single stock size Cuting Stock Problem with SAT and MaxSAT (viability: 4): https://sciencetostartup.com/paper/solving-the-two-dimensional-single-stock-size-cuting-stock-problem-with-sat-and-maxsat - A SAT-based framework for optimizing the 2D cutting stock problem, outperforming existing solvers on benchmark instances.
- Koopman-Based Nonlinear Identification and Adaptive Control of a Turbofan Engine (viability: 4): https://sciencetostartup.com/paper/koopman-based-nonlinear-identification-and-adaptive-control-of-a-turbofan-engine - Develops a Koopman operator-based adaptive control system for turbofan engines to improve robustness and thrust response.
- The AnIML Ontology: Enabling Semantic Interoperability for Large-Scale Experimental Data in Interconnected Scientific Labs (viability: 3): https://sciencetostartup.com/paper/the-animl-ontology-enabling-semantic-interoperability-for-large-scale-experimental-data-in-interconnected-scientific-lab - Develops an ontology to improve semantic interoperability for scientific experimental data, addressing inconsistencies in existing standards.
- MATA-Former & SIICU: Semantic Aware Temporal Alignment for High-Fidelity ICU Risk Prediction (viability: 7): https://sciencetostartup.com/paper/mata-former-siicu-semantic-aware-temporal-alignment-for-high-fidelity-icu-risk-prediction - A novel transformer model and labeling technique for high-fidelity ICU risk prediction using semantic event alignment and continuous multi-horizon regression.
- LiteInception: A Lightweight and Interpretable Deep Learning Framework for General Aviation Fault Diagnosis (viability: 7): https://sciencetostartup.com/paper/liteinception-a-lightweight-and-interpretable-deep-learning-framework-for-general-aviation-fault-diagnosis - A lightweight and interpretable AI framework for fault diagnosis on resource-constrained edge devices in general aviation.
- Causal Scene Narration with Runtime Safety Supervision for Vision-Language-Action Driving (viability: 7): https://sciencetostartup.com/paper/causal-scene-narration-with-runtime-safety-supervision-for-vision-language-action-driving - A novel approach to integrate diverse textual inputs for autonomous driving, improving driving performance and safety through causal scene narration and runtime supervision.
- Hi-LOAM: Hierarchical Implicit Neural Fields for LiDAR Odometry and Mapping (viability: 7): https://sciencetostartup.com/paper/hi-loam-hierarchical-implicit-neural-fields-for-lidar-odometry-and-mapping - A self-supervised LiDAR odometry and mapping system using hierarchical implicit neural fields for detailed reconstruction of complex environments.
- SteerFlow: Steering Rectified Flows for Faithful Inversion-Based Image Editing (viability: 7): https://sciencetostartup.com/paper/steerflow-steering-rectified-flows-for-faithful-inversion-based-image-editing - SteerFlow enables high-fidelity, model-agnostic image editing by steering generative flow trajectories to preserve source details and achieve complex multi-turn edits.
- End-to-End Shared Attention Estimation via Group Detection with Feedback Refinement (viability: 7): https://sciencetostartup.com/paper/end-to-end-shared-attention-estimation-via-group-detection-with-feedback-refinement - An end-to-end system for estimating shared attention by simultaneously detecting groups of people and their focus points, outperforming existing methods.
- Transformer self-attention encoder-decoder with multimodal deep learning for response time series forecasting and digital twin support in wind structural health monitoring (viability: 5): https://sciencetostartup.com/paper/transformer-self-attention-encoder-decoder-with-multimodal-deep-learning-for-response-time-series-forecasting-and-digita - A transformer-based digital twin system for wind-induced structural response forecasting and early warning of structural changes in bridges.
- Human-Guided Reasoning with Large Language Models for Vietnamese Speech Emotion Recognition (viability: 7): https://sciencetostartup.com/paper/human-guided-reasoning-with-large-language-models-for-vietnamese-speech-emotion-recognition - A human-AI collaborative framework for Vietnamese Speech Emotion Recognition that uses LLMs to reason on ambiguous cases, improving accuracy in low-resource settings.
- Bias mitigation in graph diffusion models (viability: 7): https://sciencetostartup.com/paper/bias-mitigation-in-graph-diffusion-models - A novel approach to mitigate bias in graph diffusion models, achieving state-of-the-art results without network modifications.
- OpenGo: An OpenClaw-Based Robotic Dog with Real-Time Skill Switching (viability: 5): https://sciencetostartup.com/paper/opengo-an-openclaw-based-robotic-dog-with-real-time-skill-switching - A robotic dog that can dynamically switch between skills in real-time, controlled by natural language instructions.
- Memory in the LLM Era: Modular Architectures and Strategies in a Unified Framework (viability: 7): https://sciencetostartup.com/paper/memory-in-the-llm-era-modular-architectures-and-strategies-in-a-unified-framework - A unified framework and novel memory method for LLM agents to significantly enhance long-horizon task performance.
- Development and multi-center evaluation of domain-adapted speech recognition for human-AI teaming in real-world gastrointestinal endoscopy (viability: 7): https://sciencetostartup.com/paper/development-and-multi-center-evaluation-of-domain-adapted-speech-recognition-for-human-ai-teaming-in-real-world-gastroin - A domain-adapted speech recognition system for real-time human-AI collaboration in gastrointestinal endoscopy, significantly improving accuracy and enabling efficient edge deployment.
- 3-D Relative Localization for Multi-Robot Systems with Angle and Self-Displacement Measurements (viability: 3): https://sciencetostartup.com/paper/3-d-relative-localization-for-multi-robot-systems-with-angle-and-self-displacement-measurements - A novel framework for 3D relative robot localization using angle and self-displacement measurements, addressing noise and optimization challenges.
- On the Role of Reasoning Patterns in the Generalization Discrepancy of Long Chain-of-Thought Supervised Fine-Tuning (viability: 7): https://sciencetostartup.com/paper/on-the-role-of-reasoning-patterns-in-the-generalization-discrepancy-of-long-chain-of-thought-supervised-fine-tuning - This research identifies a critical flaw in how large language models learn reasoning from diverse data sources and proposes a filtering method to significantly improve generalization performance on complex reasoning tasks.
- Can Video Diffusion Models Predict Past Frames? Bidirectional Cycle Consistency for Reversible Interpolation (viability: 6): https://sciencetostartup.com/paper/can-video-diffusion-models-predict-past-frames-bidirectional-cycle-consistency-for-reversible-interpolation - A video frame interpolation model that uses bidirectional cycle consistency to ensure temporal accuracy and reversibility, outperforming existing methods without added inference cost.
- A Graph Neural Network Approach for Solving the Ranked Assignment Problem in Multi-Object Tracking (viability: 4): https://sciencetostartup.com/paper/a-graph-neural-network-approach-for-solving-the-ranked-assignment-problem-in-multi-object-tracking - A Graph Neural Network approach to improve data association in multi-object tracking for autonomous vehicles.
- MiCA Learns More Knowledge Than LoRA and Full Fine-Tuning (viability: 5): https://sciencetostartup.com/paper/mica-learns-more-knowledge-than-lora-and-full-fine-tuning - A parameter-efficient fine-tuning method that adapts underutilized model subspaces for more efficient knowledge acquisition in LLMs.
- From Understanding to Erasing: Towards Complete and Stable Video Object Removal (viability: 8): https://sciencetostartup.com/paper/from-understanding-to-erasing-towards-complete-and-stable-video-object-removal - A novel video object removal system that leverages vision foundation models and cross-attention to achieve state-of-the-art performance and establish a new benchmark.
- Scale over Preference: The Impact of AI-Generated Content on Online Content Ecology (viability: 5): https://sciencetostartup.com/paper/scale-over-preference-the-impact-of-ai-generated-content-on-online-content-ecology - This research analyzes the impact of AI-generated content on online platforms, revealing a scale-over-preference dynamic and advocating for AIGC-sensitive distribution algorithms.
- EvoSkills: Self-Evolving Agent Skills via Co-Evolutionary Verification (viability: 5): https://sciencetostartup.com/paper/evoskills-self-evolving-agent-skills-via-co-evolutionary-verification - A framework for LLM agents to autonomously generate complex, multi-file skills for professional tasks, improving performance and reducing manual effort.
- Coupled Query-Key Dynamics for Attention (viability: 4): https://sciencetostartup.com/paper/coupled-query-key-dynamics-for-attention - A novel attention mechanism that improves language model training stability and efficiency by jointly evolving queries and keys.
- PRISM: Probability Reallocation with In-Span Masking for Knowledge-Sensitive Alignment (viability: 7): https://sciencetostartup.com/paper/prism-probability-reallocation-with-in-span-masking-for-knowledge-sensitive-alignment - A novel framework for fine-tuning language models that reduces factual hallucinations by intelligently reallocating probability at risk-sensitive generation points.
- Bridging Large-Model Reasoning and Real-Time Control via Agentic Fast-Slow Planning (viability: 7): https://sciencetostartup.com/paper/bridging-large-model-reasoning-and-real-time-control-via-agentic-fast-slow-planning - A hierarchical AI framework that bridges large-model reasoning with real-time control for autonomous driving, improving robustness and efficiency.
- BTS-rPPG: Orthogonal Butterfly Temporal Shifting for Remote Photoplethysmography (viability: 7): https://sciencetostartup.com/paper/bts-rppg-orthogonal-butterfly-temporal-shifting-for-remote-photoplethysmography - A novel temporal modeling framework for contactless physiological sensing from facial videos that significantly improves remote photoplethysmography estimation.
- Director: Instance-aware Gaussian Splatting for Dynamic Scene Modeling and Understanding (viability: 7): https://sciencetostartup.com/paper/director-instance-aware-gaussian-splatting-for-dynamic-scene-modeling-and-understanding - Director enables instance-aware 4D scene modeling for dynamic environments, allowing for semantic understanding and open-vocabulary querying of volumetric video.
- GPA: Learning GUI Process Automation from Demonstrations (viability: 7): https://sciencetostartup.com/paper/gpa-learning-gui-process-automation-from-demonstrations - A robust and private GUI automation tool that learns from single demonstrations, outperforming existing solutions in speed and success rate.
- HOT: Harmonic-Constrained Optimal Transport for Remote Photoplethysmography Domain Adaptation (viability: 7): https://sciencetostartup.com/paper/hot-harmonic-constrained-optimal-transport-for-remote-photoplethysmography-domain-adaptation - A novel framework for robust and generalized non-contact physiological measurement from facial videos, overcoming domain shifts caused by varying illumination and camera characteristics.
- Can Heterogeneous Language Models Be Fused? (viability: 4): https://sciencetostartup.com/paper/can-heterogeneous-language-models-be-fused - A method to merge diverse language models from different architectures into a single, more capable model.
- PRCCF: A Persona-guided Retrieval and Causal-aware Cognitive Filtering Framework for Emotional Support Conversation (viability: 7): https://sciencetostartup.com/paper/prccf-a-persona-guided-retrieval-and-causal-aware-cognitive-filtering-framework-for-emotional-support-conversation - A framework for empathetic AI companions that leverages persona and causal reasoning to deeply understand and respond to user emotions.
- Hierarchical Memory Orchestration for Personalized Persistent Agents (viability: 8): https://sciencetostartup.com/paper/hierarchical-memory-orchestration-for-personalized-persistent-agents - A hierarchical memory system for personalized AI agents that significantly enhances fluidity and personalization by intelligently organizing interaction history.
- Robust Embodied Perception in Dynamic Environments via Disentangled Weight Fusion (viability: 7): https://sciencetostartup.com/paper/robust-embodied-perception-in-dynamic-environments-via-disentangled-weight-fusion - A framework for embodied AI systems to adapt to changing environments without prior knowledge, improving generalization and preventing knowledge loss.
- M3D-BFS: a Multi-stage Dynamic Fusion Strategy for Sample-Adaptive Multi-Modal Brain Network Analysis (viability: 5): https://sciencetostartup.com/paper/m3d-bfs-a-multi-stage-dynamic-fusion-strategy-for-sample-adaptive-multi-modal-brain-network-analysis - A novel dynamic fusion strategy for multi-modal brain network analysis that adaptively processes samples for improved performance.
- DynaVid: Learning to Generate Highly Dynamic Videos using Synthetic Motion Data (viability: 7): https://sciencetostartup.com/paper/dynavid-learning-to-generate-highly-dynamic-videos-using-synthetic-motion-data - A video synthesis framework that uses synthetic motion data to generate highly dynamic and controllable videos, overcoming limitations of real-world datasets.
- ContextBudget: Budget-Aware Context Management for Long-Horizon Search Agents (viability: 7): https://sciencetostartup.com/paper/contextbudget-budget-aware-context-management-for-long-horizon-search-agents - A budget-aware context management system for LLM agents that optimizes information retention within cost constraints, outperforming baselines by over 1.6x.
- Ontology-Aware Design Patterns for Clinical AI Systems: Translating Reification Theory into Software Architecture (viability: 3): https://sciencetostartup.com/paper/ontology-aware-design-patterns-for-clinical-ai-systems-translating-reification-theory-into-software-architecture - Develop resilient clinical AI systems by applying ontology-aware design patterns to mitigate data distortion and semantic drift.
- AURA: Multimodal Shared Autonomy for Real-World Urban Navigation (viability: 8): https://sciencetostartup.com/paper/aura-multimodal-shared-autonomy-for-real-world-urban-navigation - A multimodal shared autonomy framework for urban navigation that reduces human operator fatigue and improves safety by intelligently decomposing control between human instruction and AI execution.
- CORAL: Towards Autonomous Multi-Agent Evolution for Open-Ended Discovery (viability: 8): https://sciencetostartup.com/paper/coral-towards-autonomous-multi-agent-evolution-for-open-ended-discovery - CORAL enables autonomous multi-agent evolution for open-ended discovery, achieving state-of-the-art results on complex tasks with significantly fewer evaluations.
- What Do Claim Verification Datasets Actually Test? A Reasoning Trace Analysis (viability: 4): https://sciencetostartup.com/paper/what-do-claim-verification-datasets-actually-test-a-reasoning-trace-analysis - This research analyzes claim verification datasets to reveal their limitations in testing complex reasoning, suggesting improvements for more robust AI evaluation.
- Moiré Video Authentication: A Physical Signature Against AI Video Generation (viability: 7): https://sciencetostartup.com/paper/moir-video-authentication-a-physical-signature-against-ai-video-generation - A physics-based signature leveraging the Moiré effect to authenticate real videos against AI-generated content.
- Cognitive Energy Modeling for Neuroadaptive Human-Machine Systems using EEG and WGAN-GP (viability: 7): https://sciencetostartup.com/paper/cognitive-energy-modeling-for-neuroadaptive-human-machine-systems-using-eeg-and-wgan-gp - Leveraging synthetic EEG data and a novel cognitive energy metric to build adaptive human-machine systems that respond in real-time to user cognitive states.
- ThinknCheck: Grounded Claim Verification with Compact, Reasoning-Driven, and Interpretable Models (viability: 7): https://sciencetostartup.com/paper/thinkncheck-grounded-claim-verification-with-compact-reasoning-driven-and-interpretable-models - A compact, reasoning-driven AI model for grounded claim verification that achieves state-of-the-art accuracy with significantly fewer parameters.
- Label Shift Estimation With Incremental Prior Update (viability: 7): https://sciencetostartup.com/paper/label-shift-estimation-with-incremental-prior-update - A post-hoc label shift estimation method that incrementally updates priors for any black-box probabilistic classifier, outperforming state-of-the-art.
- AromaGen: Interactive Generation of Rich Olfactory Experiences with Multimodal Language Models (viability: 8): https://sciencetostartup.com/paper/aromagen-interactive-generation-of-rich-olfactory-experiences-with-multimodal-language-models - A wearable AI that generates custom aromas from text or images, enhancing immersive experiences and communication.
- Exploring Robust Multi-Agent Workflows for Environmental Data Management (viability: 7): https://sciencetostartup.com/paper/exploring-robust-multi-agent-workflows-for-environmental-data-management - A multi-agent AI system that enhances the reliability and efficiency of environmental data management by externalizing knowledge and implementing audited handoffs to prevent irreversible errors.
- MonoSAOD: Monocular 3D Object Detection with Sparsely Annotated Label (viability: 7): https://sciencetostartup.com/paper/monosaod-monocular-3d-object-detection-with-sparsely-annotated-label - A framework for accurate 3D object detection from single images using minimal annotations, reducing annotation costs for real-world applications.
- Contextualizing Sink Knowledge for Java Vulnerability Discovery (viability: 8): https://sciencetostartup.com/paper/contextualizing-sink-knowledge-for-java-vulnerability-discovery - GONDAR identifies and exploits Java vulnerabilities through a novel LLM-assisted fuzzing framework, significantly outperforming existing tools.
- TOL: Textual Localization with OpenStreetMap (viability: 7): https://sciencetostartup.com/paper/tol-textual-localization-with-openstreetmap - A framework for global localization in urban environments using natural language descriptions and OpenStreetMap data, outperforming existing methods.
- LivingWorld: Interactive 4D World Generation with Environmental Dynamics (viability: 7): https://sciencetostartup.com/paper/livingworld-interactive-4d-world-generation-with-environmental-dynamics - An interactive framework for generating dynamic 4D worlds from single images, enabling real-time environmental simulations for expanding scenes.
- Fragile Reasoning: A Mechanistic Analysis of LLM Sensitivity to Meaning-Preserving Perturbations (viability: 7): https://sciencetostartup.com/paper/fragile-reasoning-a-mechanistic-analysis-of-llm-sensitivity-to-meaning-preserving-perturbations - This research develops a framework to diagnose and fix the surprising fragility of large language models to meaning-preserving text changes, offering a path to more reliable AI.
- Seclens: Role-specific Evaluation of LLM's for security vulnerablity detection (viability: 7): https://sciencetostartup.com/paper/seclens-role-specific-evaluation-of-llm-s-for-security-vulnerablity-detection - A novel evaluation framework for LLMs in security vulnerability detection that tailors performance metrics to specific stakeholder priorities, enabling more nuanced and actionable assessments.
- Diffusion-Guided Adversarial Perturbation Injection for Generalizable Defense Against Facial Manipulations (viability: 7): https://sciencetostartup.com/paper/diffusion-guided-adversarial-perturbation-injection-for-generalizable-defense-against-facial-manipulations - A diffusion-guided defense system that injects adversarial perturbations into latent space to shield facial identities from GAN and diffusion-based deepfakes, offering robust protection in both white-box and black-box scenarios.
- CRIT: Graph-Based Automatic Data Synthesis to Enhance Cross-Modal Multi-Hop Reasoning (viability: 7): https://sciencetostartup.com/paper/crit-graph-based-automatic-data-synthesis-to-enhance-cross-modal-multi-hop-reasoning - CRIT provides a novel dataset and benchmark for training Vision-Language Models to perform complex, multi-hop reasoning across text and visual information, addressing hallucination and improving grounding.
- Grounding AI-in-Education Development in Teachers' Voices: Findings from a National Survey in Indonesia (viability: 3): https://sciencetostartup.com/paper/grounding-ai-in-education-development-in-teachers-voices-findings-from-a-national-survey-in-indonesia - This research identifies teacher needs and AI usage patterns in Indonesian education to inform the development of context-appropriate AI tools and policies.
- RefinementEngine: Automating Intent-to-Device Filtering Policy Deployment under Network Constraints (viability: 7): https://sciencetostartup.com/paper/refinementengine-automating-intent-to-device-filtering-policy-deployment-under-network-constraints - Automates the translation of high-level security intents into deployable network configurations, addressing manual errors and evolving threats.
- OSCAR: Orchestrated Self-verification and Cross-path Refinement (viability: 7): https://sciencetostartup.com/paper/oscar-orchestrated-self-verification-and-cross-path-refinement - A training-free framework that uses inherent model uncertainty to detect and correct hallucinations in generated text, improving factual accuracy.
- Expert-Choice Routing Enables Adaptive Computation in Diffusion Language Models (viability: 7): https://sciencetostartup.com/paper/expert-choice-routing-enables-adaptive-computation-in-diffusion-language-models - This research introduces a novel routing mechanism for diffusion language models that significantly improves training efficiency and performance by adaptively allocating computational resources based on denoising steps, with available code for implementation.
- DWDP: Distributed Weight Data Parallelism for High-Performance LLM Inference on NVL72 (viability: 4): https://sciencetostartup.com/paper/dwdp-distributed-weight-data-parallelism-for-high-performance-llm-inference-on-nvl72 - A novel inference parallelization strategy for LLMs that improves performance by offloading MoE weights and enabling independent GPU execution.
- Automatic Image-Level Morphological Trait Annotation for Organismal Images (viability: 7): https://sciencetostartup.com/paper/automatic-image-level-morphological-trait-annotation-for-organismal-images - Automate the annotation of biological morphological traits from images using foundation models to enable large-scale ecological studies.
- Tex3D: Objects as Attack Surfaces via Adversarial 3D Textures for Vision-Language-Action Models (viability: 4): https://sciencetostartup.com/paper/tex3d-objects-as-attack-surfaces-via-adversarial-3d-textures-for-vision-language-action-models - A framework for generating 3D adversarial textures to test and improve the robustness of vision-language-action models in robotic manipulation.
- Analysis of LLM Performance on AWS Bedrock: Receipt-item Categorisation Case Study (viability: 5): https://sciencetostartup.com/paper/analysis-of-llm-performance-on-aws-bedrock-receipt-item-categorisation-case-study - A cost-aware evaluation framework for selecting the optimal LLM for receipt-item categorization on AWS Bedrock.
- Smooth Feedback Motion Planning with Reduced Curvature (viability: 7): https://sciencetostartup.com/paper/smooth-feedback-motion-planning-with-reduced-curvature - A computationally efficient method to generate smoother, less-bending robot motion paths, reducing control effort and planning time.
- Pseudo-Quantized Actor-Critic Algorithm for Robustness to Noisy Temporal Difference Error (viability: 4): https://sciencetostartup.com/paper/pseudo-quantized-actor-critic-algorithm-for-robustness-to-noisy-temporal-difference-error - A novel reinforcement learning algorithm that robustly handles noisy temporal difference errors for more stable and efficient learning.
- NEMESIS: Noise-suppressed Efficient MAE with Enhanced Superpatch Integration Strategy (viability: 7): https://sciencetostartup.com/paper/nemesis-noise-suppressed-efficient-mae-with-enhanced-superpatch-integration-strategy - A memory-efficient self-supervised learning framework for 3D medical imaging that significantly reduces computational cost and improves performance with limited annotations.
- GraphWalk: Enabling Reasoning in Large Language Models through Tool-Based Graph Navigation (viability: 7): https://sciencetostartup.com/paper/graphwalk-enabling-reasoning-in-large-language-models-through-tool-based-graph-navigation - GraphWalk enables off-the-shelf LLMs to perform complex multi-hop reasoning over enterprise-scale knowledge graphs using a minimal set of graph navigation tools.
- Swift-SVD: Theoretical Optimality Meets Practical Efficiency in Low-Rank LLM Compression (viability: 7): https://sciencetostartup.com/paper/swift-svd-theoretical-optimality-meets-practical-efficiency-in-low-rank-llm-compression - Swift-SVD offers a theoretically optimal and practically efficient method for compressing Large Language Models, achieving significant speedups in compression time.
- From Multi-Agent to Single-Agent: When Is Skill Distillation Beneficial? (viability: 7): https://sciencetostartup.com/paper/from-multi-agent-to-single-agent-when-is-skill-distillation-beneficial - A framework that predicts and optimizes when to distill multi-agent system skills into a single agent for significant cost and latency reduction.
- ModTrans: Translating Real-world Models for Distributed Training Simulator (viability: 4): https://sciencetostartup.com/paper/modtrans-translating-real-world-models-for-distributed-training-simulator - ModTrans enables the use of real-world machine learning models within distributed training simulators, bridging the gap between ML development and system research.
- Random Coordinate Descent on the Wasserstein Space of Probability Measures (viability: 4): https://sciencetostartup.com/paper/random-coordinate-descent-on-the-wasserstein-space-of-probability-measures - A novel randomized coordinate descent framework for efficient optimization over probability measures, offering significant speedups over traditional methods for machine learning and mean-field modeling.
- F3DGS: Federated 3D Gaussian Splatting for Decentralized Multi-Agent World Modeling (viability: 7): https://sciencetostartup.com/paper/f3dgs-federated-3d-gaussian-splatting-for-decentralized-multi-agent-world-modeling - A federated 3D Gaussian Splatting framework enabling decentralized multi-agent 3D reconstruction without centralized data aggregation.
- CRaFT: Circuit-Guided Refusal Feature Selection via Cross-Layer Transcoders (viability: 7): https://sciencetostartup.com/paper/craft-circuit-guided-refusal-feature-selection-via-cross-layer-transcoders - A framework for identifying and manipulating the internal mechanisms that cause LLMs to refuse harmful requests, significantly improving jailbreak success rates.
- Towards Minimal Focal Stack in Shape from Focus (viability: 7): https://sciencetostartup.com/paper/towards-minimal-focal-stack-in-shape-from-focus - Enable precise 3D depth reconstruction from just two images, significantly reducing data requirements for shape from focus applications.
- Training In-Context and In-Weights Mixtures Via Contrastive Context Sampling (viability: 3): https://sciencetostartup.com/paper/training-in-context-and-in-weights-mixtures-via-contrastive-context-sampling - A novel training strategy for LLMs that balances in-context and in-weights learning by using contrastive context sampling to improve performance and prevent label copying.
- MM-ReCoder: Advancing Chart-to-Code Generation with Reinforcement Learning and Self-Correction (viability: 7): https://sciencetostartup.com/paper/mm-recoder-advancing-chart-to-code-generation-with-reinforcement-learning-and-self-correction - A multimodal model that generates accurate and executable code from charts using reinforcement learning and self-correction.
- ByteRover: Agent-Native Memory Through LLM-Curated Hierarchical Context (viability: 7): https://sciencetostartup.com/paper/byterover-agent-native-memory-through-llm-curated-hierarchical-context - ByteRover enables LLMs to manage their own hierarchical knowledge graph directly on the filesystem, improving reasoning accuracy and reducing latency.
- Riemannian and Symplectic Geometry for Hierarchical Text-Driven Place Recognition (viability: 7): https://sciencetostartup.com/paper/riemannian-and-symplectic-geometry-for-hierarchical-text-driven-place-recognition - A novel framework for precise robot localization using text descriptions by leveraging hierarchical geometric alignments and outperforming state-of-the-art by 19%.
- Learning from the Right Rollouts: Data Attribution for PPO-based LLM Post-Training (viability: 7): https://sciencetostartup.com/paper/learning-from-the-right-rollouts-data-attribution-for-ppo-based-llm-post-training - Accelerate LLM training and reduce unfaithful reasoning by intelligently filtering training data using influence scores.
- Optimizing EEG Graph Structure for Seizure Detection: An Information Bottleneck and Self-Supervised Learning Approach (viability: 7): https://sciencetostartup.com/paper/optimizing-eeg-graph-structure-for-seizure-detection-an-information-bottleneck-and-self-supervised-learning-approach - A novel AI approach that learns denoised EEG graph structures and informative representations for improved seizure detection, offering clinically meaningful insights.
- Do Large Language Models Mentalize When They Teach? (viability: 3): https://sciencetostartup.com/paper/do-large-language-models-mentalize-when-they-teach - This research investigates whether Large Language Models exhibit mentalizing behavior when teaching, using cognitive models to analyze their decision-making processes in a simulated learning environment.
- ThinkTwice: Jointly Optimizing Large Language Models for Reasoning and Self-Refinement (viability: 7): https://sciencetostartup.com/paper/thinktwice-jointly-optimizing-large-language-models-for-reasoning-and-self-refinement - A two-phase framework that jointly trains LLMs to solve reasoning problems and refine their own answers, significantly improving accuracy on mathematical benchmarks.
- Mitigating the ID-OOD Tradeoff in Open-Set Test-Time Adaptation (viability: 7): https://sciencetostartup.com/paper/mitigating-the-id-ood-tradeoff-in-open-set-test-time-adaptation - A robust open-set test-time adaptation method that mitigates the ID-OOD tradeoff for improved model reliability in shifting environments.
- NED-Tree: Bridging the Semantic Gap with Nonlinear Element Decomposition Tree for LLM Nonlinear Optimization Modeling (viability: 7): https://sciencetostartup.com/paper/ned-tree-bridging-the-semantic-gap-with-nonlinear-element-decomposition-tree-for-llm-nonlinear-optimization-modeling - A framework that enables LLMs to accurately model complex nonlinear optimization problems by decomposing them into solver-compatible elements, with a new benchmark to drive progress.
- Variational LSTM with Augmented Inputs: Nonlinear Response History Metamodeling with Aleatoric and Epistemic Uncertainty (viability: 5): https://sciencetostartup.com/paper/variational-lstm-with-augmented-inputs-nonlinear-response-history-metamodeling-with-aleatoric-and-epistemic-uncertainty - A variational LSTM with augmented inputs and Monte Carlo dropout to quantify both aleatoric and epistemic uncertainty in high-dimensional nonlinear dynamic structural systems, reducing computational burden.
- SHOE: Semantic HOI Open-Vocabulary Evaluation Metric (viability: 7): https://sciencetostartup.com/paper/shoe-semantic-hoi-open-vocabulary-evaluation-metric - A new semantic evaluation metric for open-vocabulary human-object interaction detection that aligns better with human judgment, enabling more scalable and flexible assessment of AI models.
- Assertain: Automated Security Assertion Generation Using Large Language Models (viability: 8): https://sciencetostartup.com/paper/assertain-automated-security-assertion-generation-using-large-language-models - Automate the generation of security assertions for complex hardware designs, significantly improving vulnerability coverage and reducing manual verification effort.
- Satellite-Free Training for Drone-View Geo-Localization (viability: 4): https://sciencetostartup.com/paper/satellite-free-training-for-drone-view-geo-localization - A framework for training drone geo-localization models without relying on satellite imagery, enabling deployment in GPS-denied or data-restricted environments.
- Harmonized Tabular-Image Fusion via Gradient-Aligned Alternating Learning (viability: 7): https://sciencetostartup.com/paper/harmonized-tabular-image-fusion-via-gradient-aligned-alternating-learning - A novel learning paradigm that aligns gradients between tabular and image data to improve multimodal fusion performance.
- Thinking While Listening: Fast-Slow Recurrence for Long-Horizon Sequential Modeling (viability: 3): https://sciencetostartup.com/paper/thinking-while-listening-fast-slow-recurrence-for-long-horizon-sequential-modeling - A novel recurrent latent modeling approach for improved long-horizon sequential data representation and generalization.
- Care-Conditioned Neuromodulation for Autonomy-Preserving Supportive Dialogue Agents (viability: 7): https://sciencetostartup.com/paper/care-conditioned-neuromodulation-for-autonomy-preserving-supportive-dialogue-agents - Develops a novel framework for supportive AI dialogue agents that prioritizes user autonomy while maintaining helpfulness, addressing relational risks like dependency and coercion.
- AI-Assisted Hardware Security Verification: A Survey and AI Accelerator Case Study (viability: 4): https://sciencetostartup.com/paper/ai-assisted-hardware-security-verification-a-survey-and-ai-accelerator-case-study - Leveraging AI and LLMs to automate and accelerate hardware security verification processes for complex systems.
- Boosting Vision-Language-Action Finetuning with Feasible Action Neighborhood Prior (viability: 4): https://sciencetostartup.com/paper/boosting-vision-language-action-finetuning-with-feasible-action-neighborhood-prior - A novel regularization technique for vision-language-action models to improve sample efficiency and generalization in robotic manipulation by exploiting the inherent neighborhood of feasible actions.
- VideoZeroBench: Probing the Limits of Video MLLMs with Spatio-Temporal Evidence Verification (viability: 7): https://sciencetostartup.com/paper/videozerobench-probing-the-limits-of-video-mllms-with-spatio-temporal-evidence-verification - A new benchmark and evaluation framework to rigorously test and improve the grounded spatio-temporal reasoning capabilities of video multimodal large language models.
- AnchorVLA: Anchored Diffusion for Efficient End-to-End Mobile Manipulation (viability: 7): https://sciencetostartup.com/paper/anchorvla-anchored-diffusion-for-efficient-end-to-end-mobile-manipulation - A diffusion-based policy for mobile manipulation that enables efficient, reactive, and multimodal action generation with self-correction.
- Does Your Optimizer Care How You Normalize? Normalization-Optimizer Coupling in LLM Training (viability: 3): https://sciencetostartup.com/paper/does-your-optimizer-care-how-you-normalize-normalization-optimizer-coupling-in-llm-training - This research investigates the detrimental interaction between specific normalization layers and optimizers in LLM training, revealing failure modes that can be mitigated through parameter adjustments or EMA blending.
- Acoustic and perceptual differences between standard and accented Chinese speech and their voice clones (viability: 4): https://sciencetostartup.com/paper/acoustic-and-perceptual-differences-between-standard-and-accented-chinese-speech-and-their-voice-clones - This research investigates the perceptual differences in voice cloning for standard versus accented Chinese speech, revealing that accent significantly impacts perceived identity and intelligibility.
- ReFlow: Self-correction Motion Learning for Dynamic Scene Reconstruction (viability: 7): https://sciencetostartup.com/paper/reflow-self-correction-motion-learning-for-dynamic-scene-reconstruction - A unified framework for monocular dynamic scene reconstruction that learns 3D motion through self-correction, improving stability and accuracy.
- DeltaMem: Towards Agentic Memory Management via Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/deltamem-towards-agentic-memory-management-via-reinforcement-learning - DeltaMem is an agentic memory management system that uses reinforcement learning to significantly improve persona memory performance in conversational AI, outperforming existing product-level baselines.
- EXHIB: A Benchmark for Realistic and Diverse Evaluation of Function Similarity in the Wild (viability: 4): https://sciencetostartup.com/paper/exhib-a-benchmark-for-realistic-and-diverse-evaluation-of-function-similarity-in-the-wild - A new benchmark for evaluating binary function similarity detection models to uncover critical generalization gaps in software security.
- Cross-Domain Vessel Segmentation via Latent Similarity Mining and Iterative Co-Optimization (viability: 7): https://sciencetostartup.com/paper/cross-domain-vessel-segmentation-via-latent-similarity-mining-and-iterative-co-optimization - A novel domain transfer framework for retinal vessel segmentation that achieves state-of-the-art performance by leveraging latent vascular similarity and iterative co-optimization.
- ZEUS: Accelerating Diffusion Models with Only Second-Order Predictor (viability: 7): https://sciencetostartup.com/paper/zeus-accelerating-diffusion-models-with-only-second-order-predictor - ZEUS accelerates diffusion model inference by up to 3.2x using a novel second-order predictor without architectural changes or training, maintaining perceptual quality.
- Prototype-Based Low Altitude UAV Semantic Segmentation (viability: 7): https://sciencetostartup.com/paper/prototype-based-low-altitude-uav-semantic-segmentation - An efficient prototype-based semantic segmentation framework for low-altitude UAV imagery that balances performance and computational efficiency.
- RAE-AR: Taming Autoregressive Models with Representation Autoencoders (viability: 5): https://sciencetostartup.com/paper/rae-ar-taming-autoregressive-models-with-representation-autoencoders - A novel method to integrate powerful pre-trained visual encoders into autoregressive generative models, improving performance and unifying understanding and generation.
- Universal computational thermal imaging overcoming the ghosting effect (viability: 7): https://sciencetostartup.com/paper/universal-computational-thermal-imaging-overcoming-the-ghosting-effect - A universal computational thermal imaging framework that overcomes ghosting for high-fidelity night vision, enabling applications from autonomous navigation to healthcare.
- Countering Catastrophic Forgetting of Large Language Models for Better Instruction Following via Weight-Space Model Merging (viability: 7): https://sciencetostartup.com/paper/countering-catastrophic-forgetting-of-large-language-models-for-better-instruction-following-via-weight-space-model-merg - A model merging framework adapts large language models for clinical applications by preserving instruction-following ability and domain expertise, offering a scalable solution for resource-constrained healthcare.
- Read More, Think More: Revisiting Observation Reduction for Web Agents (viability: 5): https://sciencetostartup.com/paper/read-more-think-more-revisiting-observation-reduction-for-web-agents - This research optimizes web agent performance by adaptively selecting observation representations based on model capability and token budget, and incorporating historical context.
- PHMForge: A Scenario-Driven Agentic Benchmark for Industrial Asset Lifecycle Maintenance (viability: 7): https://sciencetostartup.com/paper/phmforge-a-scenario-driven-agentic-benchmark-for-industrial-asset-lifecycle-maintenance - A benchmark for evaluating LLM agents in industrial maintenance tasks, revealing significant gaps in current capabilities.
- A Role-Based LLM Framework for Structured Information Extraction from Healthy Food Policies (viability: 7): https://sciencetostartup.com/paper/a-role-based-llm-framework-for-structured-information-extraction-from-healthy-food-policies - A role-based LLM framework automates structured information extraction from complex healthy food policies, overcoming common LLM limitations like hallucinations and misclassifications.
- ProdCodeBench: A Production-Derived Benchmark for Evaluating AI Coding Agents (viability: 7): https://sciencetostartup.com/paper/prodcodebench-a-production-derived-benchmark-for-evaluating-ai-coding-agents - A production-derived benchmark for evaluating AI coding agents, enabling more realistic performance assessment and driving improvements in agent capabilities.
- Learning ECG Image Representations via Dual Physiological-Aware Alignments (viability: 7): https://sciencetostartup.com/paper/learning-ecg-image-representations-via-dual-physiological-aware-alignments - Unlock legacy ECG image data for automated cardiovascular diagnostics with a self-supervised framework that bridges the performance gap with signal-based analysis.
- Robust Autonomous Control of a Magnetic Millirobot in In Vitro Cardiac Flow (viability: 7): https://sciencetostartup.com/paper/robust-autonomous-control-of-a-magnetic-millirobot-in-in-vitro-cardiac-flow - Develops a vision-guided control system for autonomous navigation of magnetic millirobots in cardiac flow for targeted drug delivery.
- LLM Agents as Social Scientists: A Human-AI Collaborative Platform for Social Science Automation (viability: 4): https://sciencetostartup.com/paper/llm-agents-as-social-scientists-a-human-ai-collaborative-platform-for-social-science-automation - A platform that uses LLM agents to automate social science research by simulating human behavior and generating reports.
- Why Instruction-Based Unlearning Fails in Diffusion Models? (viability: 3): https://sciencetostartup.com/paper/why-instruction-based-unlearning-fails-in-diffusion-models - This research reveals a fundamental limitation in unlearning concepts from diffusion models using only natural language instructions, suggesting a need for new intervention methods.
- ToolMisuseBench: An Offline Deterministic Benchmark for Tool Misuse and Recovery in Agentic Systems (viability: 7): https://sciencetostartup.com/paper/toolmisusebench-an-offline-deterministic-benchmark-for-tool-misuse-and-recovery-in-agentic-systems - A deterministic benchmark and dataset to rigorously evaluate and improve the reliability of AI agents in handling tool misuse and recovery.
- Beyond Logit Adjustment: A Residual Decomposition Framework for Long-Tailed Reranking (viability: 7): https://sciencetostartup.com/paper/beyond-logit-adjustment-a-residual-decomposition-framework-for-long-tailed-reranking - A lightweight post-hoc reranker that decomposes and corrects for biases in long-tailed classification, improving accuracy on rare classes.
- Magic, Madness, Heaven, Sin: LLM Output Diversity is Everything, Everywhere, All at Once (viability: 3): https://sciencetostartup.com/paper/magic-madness-heaven-sin-llm-output-diversity-is-everything-everywhere-all-at-once - A new framework for understanding and evaluating LLM output diversity across different task objectives, revealing trade-offs between safety, representation, and creativity.
- Non-monotonicity in Conformal Risk Control (viability: 3): https://sciencetostartup.com/paper/non-monotonicity-in-conformal-risk-control - Develops theoretical guarantees for conformal risk control under non-monotonic loss functions, improving stability in practical applications.
- Identifying and Estimating Causal Direct Effects Under Unmeasured Confounding (viability: 4): https://sciencetostartup.com/paper/identifying-and-estimating-causal-direct-effects-under-unmeasured-confounding - A statistical method to identify causal direct effects in the presence of unmeasured confounding, applicable to vaccine studies.
- Matching Accuracy, Different Geometry: Evolution Strategies vs GRPO in LLM Post-Training (viability: 4): https://sciencetostartup.com/paper/matching-accuracy-different-geometry-evolution-strategies-vs-grpo-in-llm-post-training - This research compares gradient-free Evolution Strategies with gradient-based RL for LLM fine-tuning, revealing distinct solution geometries that impact knowledge preservation.
- From SWE-ZERO to SWE-HERO: Execution-free to Execution-based Fine-tuning for Software Engineering Agents (viability: 8): https://sciencetostartup.com/paper/from-swe-zero-to-swe-hero-execution-free-to-execution-based-fine-tuning-for-software-engineering-agents - A two-stage fine-tuning recipe that significantly improves the code generation and reasoning capabilities of open-weight LLMs for software engineering tasks.
- Distal-Stable Beam for Continuum Robots (viability: 3): https://sciencetostartup.com/paper/distal-stable-beam-for-continuum-robots - A novel geometric design for continuum robots that significantly improves distal stiffness for enhanced precision in constrained environments.
- CuTeGen: An LLM-Based Agentic Framework for Generation and Optimization of High-Performance GPU Kernels using CuTe (viability: 7): https://sciencetostartup.com/paper/cutegen-an-llm-based-agentic-framework-for-generation-and-optimization-of-high-performance-gpu-kernels-using-cute - An LLM-based agentic framework that automates the generation and optimization of high-performance GPU kernels through iterative refinement and execution-based validation.
- AgentSocialBench: Evaluating Privacy Risks in Human-Centered Agentic Social Networks (viability: 5): https://sciencetostartup.com/paper/agentsocialbench-evaluating-privacy-risks-in-human-centered-agentic-social-networks - A benchmark to evaluate and mitigate novel privacy risks in human-centered AI agent social networks.
- Type-Checked Compliance: Deterministic Guardrails for Agentic Financial Systems Using Lean 4 Theorem Proving (viability: 7): https://sciencetostartup.com/paper/type-checked-compliance-deterministic-guardrails-for-agentic-financial-systems-using-lean-4-theorem-proving - A formal verification platform for AI agents in finance, ensuring mathematically verifiable compliance with regulatory mandates.
- DISCO-TAB: A Hierarchical Reinforcement Learning Framework for Privacy-Preserving Synthesis of Complex Clinical Data (viability: 7): https://sciencetostartup.com/paper/disco-tab-a-hierarchical-reinforcement-learning-framework-for-privacy-preserving-synthesis-of-complex-clinical-data - A hierarchical reinforcement learning framework for generating high-fidelity, privacy-preserving synthetic clinical data that significantly improves downstream classifier performance.
- A Self-Evolving Agentic Framework for Metasurface Inverse Design (viability: 7): https://sciencetostartup.com/paper/a-self-evolving-agentic-framework-for-metasurface-inverse-design - An AI agent that autonomously learns and refines workflows for complex optical device design, reducing the need for specialized expertise.
- UniRecGen: Unifying Multi-View 3D Reconstruction and Generation (viability: 7): https://sciencetostartup.com/paper/unirecgen-unifying-multi-view-3d-reconstruction-and-generation - A unified framework for 3D reconstruction and generation from sparse views, improving fidelity and completeness.
- Soft MPCritic: Amortized Model Predictive Value Iteration (viability: 4): https://sciencetostartup.com/paper/soft-mpcritic-amortized-model-predictive-value-iteration - A framework combining reinforcement learning and model predictive control for practical and scalable policy synthesis in complex control tasks.
- When Reward Hacking Rebounds: Understanding and Mitigating It with Representation-Level Signals (viability: 3): https://sciencetostartup.com/paper/when-reward-hacking-rebounds-understanding-and-mitigating-it-with-representation-level-signals - A method to mitigate reward hacking in LLMs by penalizing shortcut behaviors during training.
- Prime Once, then Reprogram Locally: An Efficient Alternative to Black-Box Service Model Adaptation (viability: 7): https://sciencetostartup.com/paper/prime-once-then-reprogram-locally-an-efficient-alternative-to-black-box-service-model-adaptation - Adapt powerful closed-box AI models like GPT-4o efficiently and affordably by priming a local proxy, drastically reducing API calls and improving performance.
- SelfGrader: Stable Jailbreak Detection for Large Language Models using Token-Level Logits (viability: 7): https://sciencetostartup.com/paper/selfgrader-stable-jailbreak-detection-for-large-language-models-using-token-level-logits - A novel, low-latency LLM guardrail that uses token-level logits to detect jailbreaks with significantly reduced resource overhead.
- Preserving Target Distributions With Differentially Private Count Mechanisms (viability: 3): https://sciencetostartup.com/paper/preserving-target-distributions-with-differentially-private-count-mechanisms - A novel framework for differentially private count mechanisms that balances distribution accuracy with count accuracy and runtime.
- A Dynamic Atlas of Persian Poetic Symbolism: Families, Fields, and the Historical Rewiring of Meaning (viability: 2): https://sciencetostartup.com/paper/a-dynamic-atlas-of-persian-poetic-symbolism-families-fields-and-the-historical-rewiring-of-meaning - This research maps the historical evolution of symbolic meaning in Persian poetry, identifying recurring themes and their changing relationships over centuries.
- Efficient Equivariant Transformer for Self-Driving Agent Modeling (viability: 7): https://sciencetostartup.com/paper/efficient-equivariant-transformer-for-self-driving-agent-modeling - A novel transformer architecture for self-driving that models agent behaviors with SE(2)-equivariance at a reduced computational cost.
- Low-Burden LLM-Based Preference Learning: Personalizing Assistive Robots from Natural Language Feedback for Users with Paralysis (viability: 7): https://sciencetostartup.com/paper/low-burden-llm-based-preference-learning-personalizing-assistive-robots-from-natural-language-feedback-for-users-with-pa - Personalize assistive robots using natural language feedback for users with paralysis, reducing user fatigue and ensuring safety.
- Reducing Hallucinations in LLM-based Scientific Literature Analysis Using Peer Context Outlier Detection (viability: 7): https://sciencetostartup.com/paper/reducing-hallucinations-in-llm-based-scientific-literature-analysis-using-peer-context-outlier-detection - A novel method to reduce LLM hallucinations in scientific literature analysis by leveraging peer document context, improving data extraction accuracy and streamlining research workflows.
- Reinforcing Consistency in Video MLLMs with Structured Rewards (viability: 7): https://sciencetostartup.com/paper/reinforcing-consistency-in-video-mllms-with-structured-rewards - This research introduces a structured reward system for video multimodal large language models to improve factual and temporal grounding, reducing hallucinations and enhancing faithfulness in video understanding.
- Wired for Overconfidence: A Mechanistic Perspective on Inflated Verbalized Confidence in LLMs (viability: 7): https://sciencetostartup.com/paper/wired-for-overconfidence-a-mechanistic-perspective-on-inflated-verbalized-confidence-in-llms - A tool to detect and mitigate overconfidence in LLM responses by identifying and intervening on specific internal circuits.
- Infeasibility Aware Large Language Models for Combinatorial Optimization (viability: 7): https://sciencetostartup.com/paper/infeasibility-aware-large-language-models-for-combinatorial-optimization - Fine-tune LLMs to solve complex optimization problems by detecting infeasibility and accelerating search, outperforming existing models.
- Nonlinear Methods for Analyzing Pose in Behavioral Research (viability: 5): https://sciencetostartup.com/paper/nonlinear-methods-for-analyzing-pose-in-behavioral-research - A general-purpose pipeline for analyzing complex human pose data to extract meaningful behavioral insights.
- A Multi-Agent Human-LLM Collaborative Framework for Closed-Loop Scientific Literature Summarization (viability: 7): https://sciencetostartup.com/paper/a-multi-agent-human-llm-collaborative-framework-for-closed-loop-scientific-literature-summarization - A multi-agent, human-in-the-loop framework that leverages LLMs and structured AI to accelerate scientific discovery by automating literature analysis and insight extraction.
- When AI Gets it Wong: Reliability and Risk in AI-Assisted Medication Decision Systems (viability: 4): https://sciencetostartup.com/paper/when-ai-gets-it-wong-reliability-and-risk-in-ai-assisted-medication-decision-systems - This research analyzes the reliability and failure modes of AI medication decision systems to mitigate risks and improve patient safety in healthcare.
- Better Rigs, Not Bigger Networks: A Body Model Ablation for Gaussian Avatars (viability: 5): https://sciencetostartup.com/paper/better-rigs-not-bigger-networks-a-body-model-ablation-for-gaussian-avatars - A simplified approach to 3D avatar reconstruction by optimizing the body model rather than scaling network complexity, achieving state-of-the-art results with less computational overhead.
- Cooking Up Risks: Benchmarking and Reducing Food Safety Risks in Large Language Models (viability: 7): https://sciencetostartup.com/paper/cooking-up-risks-benchmarking-and-reducing-food-safety-risks-in-large-language-models - A specialized guardrail model and benchmark to mitigate food safety risks in large language models.
- Know Your Streams: On the Conceptualization, Characterization, and Generation of Intentional Event Streams (viability: 4): https://sciencetostartup.com/paper/know-your-streams-on-the-conceptualization-characterization-and-generation-of-intentional-event-streams - A prototype generator for creating realistic event streams to benchmark and improve streaming process mining algorithms.
- ClawSafety: "Safe" LLMs, Unsafe Agents (viability: 7): https://sciencetostartup.com/paper/clawsafety-safe-llms-unsafe-agents - A new benchmark and evaluation framework to rigorously test the safety of AI agents operating with elevated privileges in realistic professional environments.
- Reproducible, Explainable, and Effective Evaluations of Agentic AI for Software Engineering (viability: 4): https://sciencetostartup.com/paper/reproducible-explainable-and-effective-evaluations-of-agentic-ai-for-software-engineering - This paper proposes guidelines and a proof-of-concept for reproducible and explainable evaluations of AI agents in software engineering by making agent trajectories publicly accessible.
- Leveraging the Value of Information in POMDP Planning (viability: 5): https://sciencetostartup.com/paper/leveraging-the-value-of-information-in-pomdp-planning - A novel planning algorithm for partially observable environments that intelligently filters information to improve decision-making efficiency.
- Semantically Annotated Multimodal Dataset for RF Interpretation and Prediction (viability: 5): https://sciencetostartup.com/paper/semantically-annotated-multimodal-dataset-for-rf-interpretation-and-prediction - A new multimodal dataset bridging RF signals with visual and lidar data to enable AI-driven wireless system design and RF-based perception.
- Are Finer Citations Always Better? Rethinking Granularity for Attributed Generation (viability: 4): https://sciencetostartup.com/paper/are-finer-citations-always-better-rethinking-granularity-for-attributed-generation - This research optimizes citation granularity in attributed generation models to improve attribution quality and answer correctness, finding that intermediate granularities perform best.
- Improving Latent Generalization Using Test-time Compute (viability: 5): https://sciencetostartup.com/paper/improving-latent-generalization-using-test-time-compute - Train language models to use test-time compute for improved latent generalization and deductive reasoning.
- The power of context: Random Forest classification of near synonyms. A case study in Modern Hindi (viability: 3): https://sciencetostartup.com/paper/the-power-of-context-random-forest-classification-of-near-synonyms-a-case-study-in-modern-hindi - This research quantitatively demonstrates that word usage patterns in Hindi preserve etymological signals, distinguishing between Sanskrit and Perso-Arabic origins of synonyms.
- EgoFlow: Gradient-Guided Flow Matching for Egocentric 6DoF Object Motion Generation (viability: 7): https://sciencetostartup.com/paper/egoflow-gradient-guided-flow-matching-for-egocentric-6dof-object-motion-generation - EgoFlow generates physically consistent 6DoF object trajectories from egocentric video by combining a Mamba-Transformer-Perceiver architecture with gradient-guided flow matching, outperforming existing methods in accuracy and realism.
- Cost-Efficient Estimation of General Abilities Across Benchmarks (viability: 7): https://sciencetostartup.com/paper/cost-efficient-estimation-of-general-abilities-across-benchmarks - Develop a cost-efficient LLM benchmarking tool that predicts model performance on unseen tasks with 85% cost reduction.
- ReFormeR: Learning and Applying Explicit Query Reformulation Patterns (viability: 5): https://sciencetostartup.com/paper/reformer-learning-and-applying-explicit-query-reformulation-patterns - A pattern-guided approach to query reformulation that elicits and applies explicit patterns to constrain LLM-based query generation for improved retrieval.
- Learning When to See and When to Feel: Adaptive Vision-Torque Fusion for Contact-Aware Manipulation (viability: 7): https://sciencetostartup.com/paper/learning-when-to-see-and-when-to-feel-adaptive-vision-torque-fusion-for-contact-aware-manipulation - An adaptive fusion strategy for vision and torque sensors in robotic manipulation that significantly improves success rates in contact-rich tasks.
- Adaptive Stopping for Multi-Turn LLM Reasoning (viability: 7): https://sciencetostartup.com/paper/adaptive-stopping-for-multi-turn-llm-reasoning - A conformal prediction framework for multi-turn LLM reasoning that guarantees accuracy while reducing cost and latency.
- Test-Time Scaling Makes Overtraining Compute-Optimal (viability: 3): https://sciencetostartup.com/paper/test-time-scaling-makes-overtraining-compute-optimal - Develops new scaling laws for LLM pretraining that optimize for end-to-end compute budgets, including inference costs, leading to overtrained models with improved performance.
- Assessing Pause Thresholds for empirical Translation Process Research (viability: 2): https://sciencetostartup.com/paper/assessing-pause-thresholds-for-empirical-translation-process-research - This paper proposes a novel method for analyzing typing pauses in translation to understand cognitive processes, but lacks immediate commercial application.
- Friends and Grandmothers in Silico: Localizing Entity Cells in Language Models (viability: 3): https://sciencetostartup.com/paper/friends-and-grandmothers-in-silico-localizing-entity-cells-in-language-models - This research investigates internal mechanisms within language models for entity-centric factual questions, identifying specific neurons responsible for entity recall.
- Benchmark Problems and Benchmark Datasets for the evaluation of Machine and Deep Learning methods on Photoplethysmography signals: the D4 report from the QUMPHY project (viability: 4): https://sciencetostartup.com/paper/benchmark-problems-and-benchmark-datasets-for-the-evaluation-of-machine-and-deep-learning-methods-on-photoplethysmograph - A curated set of benchmark problems and datasets for evaluating machine learning on photoplethysmography signals to quantify uncertainty in medical applications.
- A soft and lightweight fabric-based pneumatic interface for multimodal fingertip tactile feedback (viability: 5): https://sciencetostartup.com/paper/a-soft-and-lightweight-fabric-based-pneumatic-interface-for-multimodal-fingertip-tactile-feedback - Develops a lightweight, fabric-based pneumatic haptic interface for realistic tactile feedback in VR/AR and teleoperation.
- LESV: Language Embedded Sparse Voxel Fusion for Open-Vocabulary 3D Scene Understanding (viability: 7): https://sciencetostartup.com/paper/lesv-language-embedded-sparse-voxel-fusion-for-open-vocabulary-3d-scene-understanding - A novel framework for open-vocabulary 3D scene understanding that uses sparse voxel rasterization and foundation models to overcome spatial and semantic ambiguities, achieving state-of-the-art performance.
- GRAZE: Grounded Refinement and Motion-Aware Zero-Shot Event Localization (viability: 7): https://sciencetostartup.com/paper/graze-grounded-refinement-and-motion-aware-zero-shot-event-localization - A training-free AI pipeline for precise First Point of Contact localization in American football practice videos, enabling biomechanical analysis without labeled data.
- Can LLMs Predict Academic Collaboration? Topology Heuristics vs. LLM-Based Link Prediction on Real Co-authorship Networks (viability: 7): https://sciencetostartup.com/paper/can-llms-predict-academic-collaboration-topology-heuristics-vs-llm-based-link-prediction-on-real-co-authorship-networks - Leveraging LLMs to predict future academic collaborations by analyzing author profiles, outperforming traditional methods and identifying novel connections.
- Residuals-based Offline Reinforcement Learning (viability: 4): https://sciencetostartup.com/paper/residuals-based-offline-reinforcement-learning - A new offline reinforcement learning framework that explicitly accounts for estimation error in transition dynamics to improve policy optimization.
- RIFT: A RubrIc Failure Mode Taxonomy and Automated Diagnostics (viability: 5): https://sciencetostartup.com/paper/rift-a-rubric-failure-mode-taxonomy-and-automated-diagnostics - Automate the diagnosis of rubric quality issues in LLM benchmarks to improve evaluation reliability.
- AffordTissue: Dense Affordance Prediction for Tool-Action Specific Tissue Interaction (viability: 7): https://sciencetostartup.com/paper/affordtissue-dense-affordance-prediction-for-tool-action-specific-tissue-interaction - AffordTissue predicts tool-action specific safe interaction regions on tissue for surgical automation, outperforming general vision-language models.
- "The System Will Choose Security Over Humanity Every Time": Understanding Security and Privacy for U.S. Incarcerated Users (viability: 3): https://sciencetostartup.com/paper/the-system-will-choose-security-over-humanity-every-time-understanding-security-and-privacy-for-u-s-incarcerated-users - This research identifies critical privacy and security vulnerabilities in digital devices used by incarcerated individuals, highlighting the need for user-centric design and policy reform.
- CogBias: Measuring and Mitigating Cognitive Bias in Large Language Models (viability: 7): https://sciencetostartup.com/paper/cogbias-measuring-and-mitigating-cognitive-bias-in-large-language-models - This research quantifies and mitigates cognitive biases in LLMs by identifying and manipulating internal representations, offering a path to more reliable AI decision-making.
- Crashing Waves vs. Rising Tides: Preliminary Findings on AI Automation from Thousands of Worker Evaluations of Labor Market Tasks (viability: 3): https://sciencetostartup.com/paper/crashing-waves-vs-rising-tides-preliminary-findings-on-ai-automation-from-thousands-of-worker-evaluations-of-labor-marke - This research analyzes the continuous and broad-based nature of AI automation in labor market tasks, contrasting with abrupt 'crashing waves' of capability surges.
- IGLOSS: Image Generation for Lidar Open-vocabulary Semantic Segmentation (viability: 7): https://sciencetostartup.com/paper/igloss-image-generation-for-lidar-open-vocabulary-semantic-segmentation - Generate prototype images from text to enable zero-shot open-vocabulary semantic segmentation for 3D lidar data, outperforming existing methods.
- Semantic Modeling for World-Centered Architectures (viability: 3): https://sciencetostartup.com/paper/semantic-modeling-for-world-centered-architectures - A new framework for multi-agent systems that uses a shared world representation for improved consistency and stability.
- Open-Domain Safety Policy Construction (viability: 7): https://sciencetostartup.com/paper/open-domain-safety-policy-construction - An agentic system that automatically drafts content moderation policies from minimal seed information, outperforming existing methods and expert-written policies.
- Open-loop POMDP Simplification and Safe Skipping of Replanning with Formal Performance Guarantees (viability: 5): https://sciencetostartup.com/paper/open-loop-pomdp-simplification-and-safe-skipping-of-replanning-with-formal-performance-guarantees - A new framework for adaptive open-loop planning in POMDPs with formal performance guarantees, enabling faster and safer decision-making under uncertainty.
- No Attacker Needed: Unintentional Cross-User Contamination in Shared-State LLM Agents (viability: 4): https://sciencetostartup.com/paper/no-attacker-needed-unintentional-cross-user-contamination-in-shared-state-llm-agents - This research identifies and quantifies a critical security flaw in shared-state LLM agents where benign user interactions can unintentionally corrupt outcomes for other users, requiring new artifact-level defenses.
- PI-JEPA: Label-Free Surrogate Pretraining for Coupled Multiphysics Simulation via Operator-Split Latent Prediction (viability: 7): https://sciencetostartup.com/paper/pi-jepa-label-free-surrogate-pretraining-for-coupled-multiphysics-simulation-via-operator-split-latent-prediction - A physics-informed AI framework that drastically reduces the need for expensive simulation data by pretraining on unlabeled physics parameters, enabling faster and more accurate multiphysics surrogate models.
- Procedural Knowledge at Scale Improves Reasoning (viability: 7): https://sciencetostartup.com/paper/procedural-knowledge-at-scale-improves-reasoning - A retrieval-augmented generation framework that leverages a large corpus of procedural knowledge to significantly improve language model reasoning capabilities on complex tasks.
- Safety, Security, and Cognitive Risks in World Models (viability: 3): https://sciencetostartup.com/paper/safety-security-and-cognitive-risks-in-world-models - This paper identifies and formalizes safety, security, and cognitive risks in world models, proposing mitigations for critical AI deployments.
- Malliavin Calculus for Counterfactual Gradient Estimation in Adaptive Inverse Reinforcement Learning (viability: 3): https://sciencetostartup.com/paper/malliavin-calculus-for-counterfactual-gradient-estimation-in-adaptive-inverse-reinforcement-learning - A theoretical framework for estimating counterfactual gradients in adaptive inverse reinforcement learning using Malliavin calculus.
- IDEA2: Expert-in-the-loop competency question elicitation for collaborative ontology engineering (viability: 7): https://sciencetostartup.com/paper/idea2-expert-in-the-loop-competency-question-elicitation-for-collaborative-ontology-engineering - Accelerate ontology engineering by using LLMs to semi-automatically elicit competency questions from domain experts through an iterative feedback loop.
- Massively Parallel Exact Inference for Hawkes Processes (viability: 7): https://sciencetostartup.com/paper/massively-parallel-exact-inference-for-hawkes-processes - A massively parallel PyTorch library for exact inference of multivariate Hawkes processes, enabling analysis at unprecedented scales.
- Perceptual misalignment of texture representations in convolutional neural networks (viability: 2): https://sciencetostartup.com/paper/perceptual-misalignment-of-texture-representations-in-convolutional-neural-networks - This research investigates the disconnect between how convolutional neural networks represent textures and human perception, suggesting current object recognition models are insufficient for understanding texture perception.
- Regularizing Attention Scores with Bootstrapping (viability: 7): https://sciencetostartup.com/paper/regularizing-attention-scores-with-bootstrapping - A novel bootstrapping method for vision transformers that quantifies attention score uncertainty, leading to more interpretable and sparse attention maps for image analysis.
- SECURE: Stable Early Collision Understanding via Robust Embeddings in Autonomous Driving (viability: 7): https://sciencetostartup.com/paper/secure-stable-early-collision-understanding-via-robust-embeddings-in-autonomous-driving - A framework for building highly robust and stable AI models for early collision prediction in autonomous driving, significantly reducing reliability risks.
- Bias Inheritance in Neural-Symbolic Discovery of Constitutive Closures Under Function-Class Mismatch (viability: 4): https://sciencetostartup.com/paper/bias-inheritance-in-neural-symbolic-discovery-of-constitutive-closures-under-function-class-mismatch - A neural-symbolic framework to robustly discover physical laws from data, addressing bias inheritance in scientific modeling.
- Evolutionary Multi-Objective Fusion of Deepfake Speech Detectors (viability: 7): https://sciencetostartup.com/paper/evolutionary-multi-objective-fusion-of-deepfake-speech-detectors - A framework that uses evolutionary algorithms to fuse deepfake speech detectors, achieving state-of-the-art accuracy with significantly reduced system complexity.
- Model Merging via Data-Free Covariance Estimation (viability: 7): https://sciencetostartup.com/paper/model-merging-via-data-free-covariance-estimation - A data-free method to merge AI models, inheriting capabilities without needing original training data, outperforming existing techniques.
- Efficient and Principled Scientific Discovery through Bayesian Optimization: A Tutorial (viability: 3): https://sciencetostartup.com/paper/efficient-and-principled-scientific-discovery-through-bayesian-optimization-a-tutorial - A tutorial on Bayesian Optimization for accelerating scientific discovery by automating the hypothesize-experiment-refine cycle.
- Macroscopic transport patterns of UAV traffic in 3D anisotropic wind fields: A constraint-preserving hybrid PINN-FVM approach (viability: 3): https://sciencetostartup.com/paper/macroscopic-transport-patterns-of-uav-traffic-in-3d-anisotropic-wind-fields-a-constraint-preserving-hybrid-pinn-fvm-appr - A hybrid physics-informed neural network and finite-volume method for simulating UAV traffic in complex wind fields.
- The Digital Twin Counterfactual Framework: A Validation Architecture for Simulated Potential Outcomes (viability: 2): https://sciencetostartup.com/paper/the-digital-twin-counterfactual-framework-a-validation-architecture-for-simulated-potential-outcomes - A framework for validating simulated counterfactual outcomes in causal inference by introducing a digital twin and a hierarchical validation architecture.
- Human Pose Estimation in Trampoline Gymnastics: Improving Performance Using a New Synthetic Dataset (viability: 7): https://sciencetostartup.com/paper/human-pose-estimation-in-trampoline-gymnastics-improving-performance-using-a-new-synthetic-dataset - Improve human pose estimation for extreme sports by fine-tuning models on a novel synthetic dataset.
- ViTs for Action Classification in Videos: An Approach to Risky Tackle Detection in American Football Practice Videos (viability: 7): https://sciencetostartup.com/paper/vits-for-action-classification-in-videos-an-approach-to-risky-tackle-detection-in-american-football-practice-videos - A Vision Transformer model and expanded dataset for early detection of risky tackles in American football, enabling coach-centered injury prevention.
- Detecting Complex Money Laundering Patterns with Incremental and Distributed Graph Modeling (viability: 7): https://sciencetostartup.com/paper/detecting-complex-money-laundering-patterns-with-incremental-and-distributed-graph-modeling - A distributed graph modeling framework to detect complex money laundering patterns with reduced false positives, validated on real and synthetic datasets.
- JetPrism: diagnosing convergence for generative simulation and inverse problems in nuclear physics (viability: 4): https://sciencetostartup.com/paper/jetprism-diagnosing-convergence-for-generative-simulation-and-inverse-problems-in-nuclear-physics - A diagnostic framework for generative simulation and inverse problems that ensures precise statistical agreement with ground-truth data, applicable across various scientific and financial domains.
- Preference learning in shades of gray: Interpretable and bias-aware reward modeling for human preferences (viability: 7): https://sciencetostartup.com/paper/preference-learning-in-shades-of-gray-interpretable-and-bias-aware-reward-modeling-for-human-preferences - A novel framework for interpretable and bias-aware reward modeling in LLMs that significantly improves preference learning accuracy.
- Sparse Spectral LoRA: Routed Experts for Medical VLMs (viability: 7): https://sciencetostartup.com/paper/sparse-spectral-lora-routed-experts-for-medical-vlms - A parameter-efficient medical vision-language model that achieves near full fine-tuning performance with significantly fewer trainable parameters and reduced catastrophic forgetting.
- An Online Machine Learning Multi-resolution Optimization Framework for Energy System Design Limit of Performance Analysis (viability: 4): https://sciencetostartup.com/paper/an-online-machine-learning-multi-resolution-optimization-framework-for-energy-system-design-limit-of-performance-analysi - An ML-accelerated framework to optimize energy system design by bridging architectural and operational performance gaps, reducing high-fidelity model evaluations.
- M2-Verify: A Large-Scale Multidomain Benchmark for Checking Multimodal Claim Consistency (viability: 7): https://sciencetostartup.com/paper/m2-verify-a-large-scale-multidomain-benchmark-for-checking-multimodal-claim-consistency - A large-scale, expert-validated multimodal dataset to benchmark and improve AI's ability to verify scientific claims against visual evidence, revealing significant gaps in current state-of-the-art models.
- UQ-SHRED: uncertainty quantification of shallow recurrent decoder networks for sparse sensing via engression (viability: 7): https://sciencetostartup.com/paper/uq-shred-uncertainty-quantification-of-shallow-recurrent-decoder-networks-for-sparse-sensing-via-engression - UQ-SHRED quantifies uncertainty in reconstructing complex spatiotemporal fields from sparse sensor data using a novel distributional regression approach.
- Scaling Reasoning Tokens via RL and Parallel Thinking: Evidence From Competitive Programming (viability: 4): https://sciencetostartup.com/paper/scaling-reasoning-tokens-via-rl-and-parallel-thinking-evidence-from-competitive-programming - A system that scales reasoning token budgets for competitive programming using RL and parallel thinking to significantly improve performance on hard problems.
- Forecasting Supply Chain Disruptions with Foresight Learning (viability: 8): https://sciencetostartup.com/paper/forecasting-supply-chain-disruptions-with-foresight-learning - Train LLMs to provide accurate, calibrated probabilistic forecasts for supply chain disruptions, outperforming existing models and offering decision-ready signals.
- Look Twice: Training-Free Evidence Highlighting in Multimodal Large Language Models (viability: 7): https://sciencetostartup.com/paper/look-twice-training-free-evidence-highlighting-in-multimodal-large-language-models - A training-free framework that enhances multimodal LLMs' ability to identify and utilize relevant visual and textual evidence for knowledge-intensive question answering.
- Sven: Singular Value Descent as a Computationally Efficient Natural Gradient Method (viability: 7): https://sciencetostartup.com/paper/sven-singular-value-descent-as-a-computationally-efficient-natural-gradient-method - Sven is a novel optimization algorithm that significantly outperforms Adam and LBFGS on regression tasks by efficiently approximating natural gradient descent, offering faster convergence and lower loss.
- Non-Rigid 3D Shape Correspondences: From Foundations to Open Challenges and Opportunities (viability: 4): https://sciencetostartup.com/paper/non-rigid-3d-shape-correspondences-from-foundations-to-open-challenges-and-opportunities - A survey of methods for estimating correspondences between deformed 3D shapes, highlighting recent advances and future research directions.
- Functional Force-Aware Retargeting from Virtual Human Demos to Soft Robot Policies (viability: 7): https://sciencetostartup.com/paper/functional-force-aware-retargeting-from-virtual-human-demos-to-soft-robot-policies - A framework for teaching soft robot hands manipulation skills by explicitly reasoning about contact forces, enabling functional force-aware retargeting from human demonstrations.
- HippoCamp: Benchmarking Contextual Agents on Personal Computers (viability: 6): https://sciencetostartup.com/paper/hippocamp-benchmarking-contextual-agents-on-personal-computers - HippoCamp is a benchmark evaluating digital assistants' capabilities in managing personal file systems for enhanced user-specific reasoning.
- Universal YOCO for Efficient Depth Scaling (viability: 3): https://sciencetostartup.com/paper/universal-yoco-for-efficient-depth-scaling - Universal YOCO combines a decoder-decoder architecture with recursive computation for efficient depth scaling of Large Language Models, improving inference efficiency.
- LAtent Phase Inference from Short time sequences using SHallow REcurrent Decoders (LAPIS-SHRED) (viability: 0): https://sciencetostartup.com/paper/latent-phase-inference-from-short-time-sequences-using-shallow-recurrent-decoders-lapis-shred - LAPIS-SHRED is a modular architecture for reconstructing and forecasting complete spatiotemporal dynamics from sparse sensor observations in short temporal windows.
- The Recipe Matters More Than the Kitchen:Mathematical Foundations of the AI Weather Prediction Pipeline (viability: 4): https://sciencetostartup.com/paper/the-recipe-matters-more-than-the-kitchen-mathematical-foundations-of-the-ai-weather-prediction-pipeline - A theoretical framework and empirical validation for improving AI weather prediction by focusing on the entire learning pipeline, not just architecture.
- Collaborative Task and Path Planning for Heterogeneous Robotic Teams using Multi-Agent PPO (viability: 4): https://sciencetostartup.com/paper/collaborative-task-and-path-planning-for-heterogeneous-robotic-teams-using-multi-agent-ppo - A collaborative planning strategy using Multi-Agent PPO to coordinate heterogeneous robotic teams for efficient extraterrestrial exploration.
- $\texttt{YC-Bench}$: Benchmarking AI Agents for Long-Term Planning and Consistent Execution (viability: 7): https://sciencetostartup.com/paper/texttt-yc-bench-benchmarking-ai-agents-for-long-term-planning-and-consistent-execution - YC-Bench benchmarks LLM agents for long-term planning and consistent execution by simulating a startup over a year, revealing key capability gaps.
- CliffSearch: Structured Agentic Co-Evolution over Theory and Code for Scientific Algorithm Discovery (viability: 6): https://sciencetostartup.com/paper/cliffsearch-structured-agentic-co-evolution-over-theory-and-code-for-scientific-algorithm-discovery - CliffSearch offers an AI-driven, evolutionary framework that enhances scientific algorithm discovery by ensuring correctness and originality through agent-based mutation and review processes.
- TRACE: High-Fidelity 3D Scene Editing via Tangible Reconstruction and Geometry-Aligned Contextual Video Masking (viability: 7): https://sciencetostartup.com/paper/trace-high-fidelity-3d-scene-editing-via-tangible-reconstruction-and-geometry-aligned-contextual-video-masking - TRACE enables high-fidelity, part-level 3D scene editing by anchoring video diffusion with explicit 3D geometry, outperforming existing methods in editing versatility and structural integrity.
- LLM REgression with a Latent Iterative State Head (viability: 5): https://sciencetostartup.com/paper/llm-regression-with-a-latent-iterative-state-head - RELISH enhances LLM-based regression by iteratively refining latent states for precise numerical predictions.
- Neural Harmonic Textures for High-Quality Primitive Based Neural Reconstruction (viability: 7): https://sciencetostartup.com/paper/neural-harmonic-textures-for-high-quality-primitive-based-neural-reconstruction - Neural Harmonic Textures enhance primitive-based neural reconstruction by anchoring latent features on a scaffold and using periodic activations for high-frequency detail and real-time novel view synthesis.
- Therefore I am. I Think (viability: 7): https://sciencetostartup.com/paper/therefore-i-am-i-think - This research provides evidence that LLM reasoning models encode decisions early, shaping chain-of-thought and allowing for manipulation of tool-calling behavior before deliberation.
- ORBIT: Scalable and Verifiable Data Generation for Search Agents on a Tight Budget (viability: 7): https://sciencetostartup.com/paper/orbit-scalable-and-verifiable-data-generation-for-search-agents-on-a-tight-budget - Generate scalable and verifiable training data for search agents using a frugal, open-source framework to improve performance on complex queries.
- A ROS 2 Wrapper for Florence-2: Multi-Mode Local Vision-Language Inference for Robotic Systems (viability: 7): https://sciencetostartup.com/paper/a-ros-2-wrapper-for-florence-2-multi-mode-local-vision-language-inference-for-robotic-systems - A ROS 2 wrapper for Florence-2 enables local, multi-mode vision-language inference for robotic systems, feasible on consumer hardware.
- Screening Is Enough (viability: 3): https://sciencetostartup.com/paper/screening-is-enough - Multiscreen is a novel language model architecture that uses screening to achieve absolute query-key relevance, reducing parameters and improving inference latency.
- NeuroDDAF: Neural Dynamic Diffusion-Advection Fields with Evidential Fusion for Air Quality Forecasting (viability: 7): https://sciencetostartup.com/paper/neuroddaf-neural-dynamic-diffusion-advection-fields-with-evidential-fusion-for-air-quality-forecasting - NeuroDDAF is a physics-informed framework for air quality forecasting that unifies neural networks with transport modeling, outperforming baselines and providing calibrated uncertainty.
- Open-Set Supervised 3D Anomaly Detection: An Industrial Dataset and a Generalisable Framework for Unknown Defects (viability: 7): https://sciencetostartup.com/paper/open-set-supervised-3d-anomaly-detection-an-industrial-dataset-and-a-generalisable-framework-for-unknown-defects - Open3D-AD is a framework for open-set supervised 3D anomaly detection in industrial settings, leveraging normal and limited anomalous samples to identify unknown defects.
- Online Reasoning Calibration: Test-Time Training Enables Generalizable Conformal LLM Reasoning (viability: 8): https://sciencetostartup.com/paper/online-reasoning-calibration-test-time-training-enables-generalizable-conformal-llm-reasoning - ORCA calibrates LLM sampling at test-time using conformal prediction and meta-learning, providing valid confidence estimates and improving efficiency across reasoning tasks.
- Bridging the Simulation-to-Experiment Gap with Generative Models using Adversarial Distribution Alignment (viability: 7): https://sciencetostartup.com/paper/bridging-the-simulation-to-experiment-gap-with-generative-models-using-adversarial-distribution-alignment - Adversarial Distribution Alignment (ADA) bridges the simulation-to-experiment gap by aligning generative models with real-world experimental observations.
- S0 Tuning: Zero-Overhead Adaptation of Hybrid Recurrent-Attention Models (viability: 8): https://sciencetostartup.com/paper/s0-tuning-zero-overhead-adaptation-of-hybrid-recurrent-attention-models - S0 Tuning offers zero-inference overhead adaptation for hybrid recurrent-attention LLMs by tuning a single initial state matrix per recurrent layer.
- Reasoning Shift: How Context Silently Shortens LLM Reasoning (viability: 3): https://sciencetostartup.com/paper/reasoning-shift-how-context-silently-shortens-llm-reasoning - This paper investigates how context length and conversational settings silently shorten LLM reasoning traces, potentially impacting performance on complex tasks.
- SMASH: Mastering Scalable Whole-Body Skills for Humanoid Ping-Pong with Egocentric Vision (viability: 7): https://sciencetostartup.com/paper/smash-mastering-scalable-whole-body-skills-for-humanoid-ping-pong-with-egocentric-vision - A modular system for agile humanoid ping-pong using onboard egocentric vision and scalable whole-body skill learning, eliminating the need for external sensors.
- Property-Level Flood Risk Assessment Using AI-Enabled Street-View Lowest Floor Elevation Extraction and ML Imputation Across Texas (viability: 4): https://sciencetostartup.com/paper/property-level-flood-risk-assessment-using-ai-enabled-street-view-lowest-floor-elevation-extraction-and-ml-imputation-ac - Brainstacks is a modular architecture for continual multi-domain LLM fine-tuning using frozen MoE-LoRA adapter stacks that enable cross-domain cognitive capability composition.
- Brainstacks: Cross-Domain Cognitive Capabilities via Frozen MoE-LoRA Stacks for Continual LLM Learning (viability: 6): https://sciencetostartup.com/paper/brainstacks-cross-domain-cognitive-capabilities-via-frozen-moe-lora-stacks-for-continual-llm-learning - An AI-enabled street-view analysis pipeline for extracting building elevation data to improve regional flood risk assessment.
- Detecting Multi-Agent Collusion Through Multi-Agent Interpretability (viability: 7): https://sciencetostartup.com/paper/detecting-multi-agent-collusion-through-multi-agent-interpretability - NARCBench is a benchmark and probing technique for detecting multi-agent LLM collusion by analyzing internal model activations.
- SERSEM: Selective Entropy-Weighted Scoring for Membership Inference in Code Language Models (viability: 7): https://sciencetostartup.com/paper/sersem-selective-entropy-weighted-scoring-for-membership-inference-in-code-language-models - A novel framework for detecting data contamination in code LLMs by amplifying specific memorization signals through selective entropy weighting.
- Deep Reinforcement Learning for Robotic Manipulation under Distribution Shift with Bounded Extremum Seeking (viability: 4): https://sciencetostartup.com/paper/deep-reinforcement-learning-for-robotic-manipulation-under-distribution-shift-with-bounded-extremum-seeking - Improving robotic manipulation robustness under distribution shift by combining deep reinforcement learning with bounded extremum seeking.
- Looking into a Pixel by Nonlinear Unmixing -- A Generative Approach (viability: 7): https://sciencetostartup.com/paper/looking-into-a-pixel-by-nonlinear-unmixing-a-generative-approach - A generative approach using a bi-directional GAN for hyperspectral nonlinear unmixing without explicit mixing model knowledge.
- VRUD: A Drone Dataset for Complex Vehicle-VRU Interactions within Mixed Traffic (viability: 8): https://sciencetostartup.com/paper/vrud-a-drone-dataset-for-complex-vehicle-vru-interactions-within-mixed-traffic - A novel drone-based dataset and method for capturing complex vehicle-VRU interactions in unstructured urban traffic, enabling safer autonomous driving systems.
- Obfuscating Code Vulnerabilities against Static Analysis in JavaScript Code (viability: 4): https://sciencetostartup.com/paper/obfuscating-code-vulnerabilities-against-static-analysis-in-javascript-code - This research quantifies the impact of JavaScript obfuscation on static code analysis tools, revealing significant weaknesses in current security pipelines.
- Toward Personalized Darts Training: A Data-Driven Framework Based on Skeleton-Based Biomechanical Analysis and Motion Modeling (viability: 5): https://sciencetostartup.com/paper/toward-personalized-darts-training-a-data-driven-framework-based-on-skeleton-based-biomechanical-analysis-and-motion-mod - A data-driven framework for personalized darts training that uses biomechanical analysis and motion modeling to provide targeted feedback.
- ReinDriveGen: Reinforcement Post-Training for Out-of-Distribution Driving Scene Generation (viability: 7): https://sciencetostartup.com/paper/reindrivegen-reinforcement-post-training-for-out-of-distribution-driving-scene-generation - ReinDriveGen enables controllable generation of out-of-distribution driving scenes by editing actor trajectories and using reinforcement learning for robust video synthesis.
- Paper Reconstruction Evaluation: Evaluating Presentation and Hallucination in AI-written Papers (viability: 4): https://sciencetostartup.com/paper/paper-reconstruction-evaluation-evaluating-presentation-and-hallucination-in-ai-written-papers - PaperRecon is a new evaluation framework for AI-written papers, disentangling presentation quality from hallucination risks.
- Multi-Agent LLM Governance for Safe Two-Timescale Reinforcement Learning in SDN-IoT Defense (viability: 3): https://sciencetostartup.com/paper/multi-agent-llm-governance-for-safe-two-timescale-reinforcement-learning-in-sdn-iot-defense - A multi-agent LLM governance system for safe, two-timescale reinforcement learning in SDN-IoT defense.
- Lightweight Prompt-Guided CLIP Adaptation for Monocular Depth Estimation (viability: 6): https://sciencetostartup.com/paper/lightweight-prompt-guided-clip-adaptation-for-monocular-depth-estimation - A lightweight, prompt-guided adapter for CLIP to perform monocular depth estimation with minimal supervision.
- Reconsidering Dependency Networks from an Information Geometry Perspective (viability: 0): https://sciencetostartup.com/paper/reconsidering-dependency-networks-from-an-information-geometry-perspective - An information-geometric analysis of dependency networks and pseudo-Gibbs sampling for theoretical understanding.
- ProTPS: Prototype-Guided Text Prompt Selection for Continual Learning (viability: 7): https://sciencetostartup.com/paper/protps-prototype-guided-text-prompt-selection-for-continual-learning - A prototype-guided text prompt selection method for continual learning that mitigates catastrophic forgetting and introduces a new marine species dataset.
- Trust and Reliance on AI in Education: AI Literacy and Need for Cognition as Moderators (viability: 2): https://sciencetostartup.com/paper/trust-and-reliance-on-ai-in-education-ai-literacy-and-need-for-cognition-as-moderators - This research explores how students' trust, AI literacy, and need for cognition influence their reliance on AI assistants in programming tasks, suggesting a need for better instructional support.
- CARE: Privacy-Compliant Agentic Reasoning with Evidence Discordance (viability: 8): https://sciencetostartup.com/paper/care-privacy-compliant-agentic-reasoning-with-evidence-discordance - CARE is a privacy-compliant agentic reasoning framework that uses a remote LLM for guidance and a local LLM for decision-making to reconcile conflicting evidence in medical data.
- Adversarial Moral Stress Testing of Large Language Models (viability: 7): https://sciencetostartup.com/paper/adversarial-moral-stress-testing-of-large-language-models - Adversarial Moral Stress Testing (AMST) is a framework for evaluating LLM ethical robustness under sustained adversarial interaction, exposing degradation patterns missed by single-round evaluations.
- Approximating Pareto Frontiers in Stochastic Multi-Objective Optimization via Hashing and Randomization (viability: 7): https://sciencetostartup.com/paper/approximating-pareto-frontiers-in-stochastic-multi-objective-optimization-via-hashing-and-randomization - XOR-SMOO is a novel algorithm that approximates Pareto frontiers in stochastic multi-objective optimization by querying SAT oracles, outperforming baselines on real-world problems.
- Temporal Dependencies in In-Context Learning: The Role of Induction Heads (viability: 3): https://sciencetostartup.com/paper/temporal-dependencies-in-in-context-learning-the-role-of-induction-heads - Investigating the role of induction heads in LLMs for understanding temporal dependencies in in-context learning.
- LightGuard: Transparent WiFi Security via Physical-Layer LiFi Key Bootstrapping (viability: 5): https://sciencetostartup.com/paper/lightguard-transparent-wifi-security-via-physical-layer-lifi-key-bootstrapping - LightGuard enhances WiFi security by offloading cryptographic key establishment to a physically confined LiFi channel.
- TRACE: Training-Free Partial Audio Deepfake Detection via Embedding Trajectory Analysis of Speech Foundation Models (viability: 7): https://sciencetostartup.com/paper/trace-training-free-partial-audio-deepfake-detection-via-embedding-trajectory-analysis-of-speech-foundation-models - TRACE is a training-free framework for detecting audio deepfakes by analyzing speech foundation model embedding trajectories.
- ReMoGen: Real-time Human Interaction-to-Reaction Generation via Modular Learning from Diverse Data (viability: 6): https://sciencetostartup.com/paper/remogen-real-time-human-interaction-to-reaction-generation-via-modular-learning-from-diverse-data - ReMoGen generates real-time human reactions to interactions using modular learning and segment-level refinement.
- ProOOD: Prototype-Guided Out-of-Distribution 3D Occupancy Prediction (viability: 7): https://sciencetostartup.com/paper/proood-prototype-guided-out-of-distribution-3d-occupancy-prediction - A plug-and-play method for 3D occupancy prediction that improves accuracy and out-of-distribution detection in autonomous driving.
- Automated Generation of Cybersecurity Exercise Scenarios (viability: 8): https://sciencetostartup.com/paper/automated-generation-of-cybersecurity-exercise-scenarios - An automated system for generating diverse cybersecurity exercise scenarios, complete with a simulation environment and a large dataset.
- Narrative Fingerprints: Multi-Scale Author Identification via Novelty Curve Dynamics (viability: 1): https://sciencetostartup.com/paper/narrative-fingerprints-multi-scale-author-identification-via-novelty-curve-dynamics - Authors leave unique 'fingerprints' in the novelty dynamics of their writing, detectable at both book and chapter levels.
- BAT: Balancing Agility and Stability via Online Policy Switching for Long-Horizon Whole-Body Humanoid Control (viability: 7): https://sciencetostartup.com/paper/bat-balancing-agility-and-stability-via-online-policy-switching-for-long-horizon-whole-body-humanoid-control - A framework for humanoid robots that dynamically switches between control policies to balance agility and stability for long-horizon tasks.
- PHASOR: Anatomy- and Phase-Consistent Volumetric Diffusion for CT Virtual Contrast Enhancement (viability: 7): https://sciencetostartup.com/paper/phasor-anatomy-and-phase-consistent-volumetric-diffusion-for-ct-virtual-contrast-enhancement - A volumetric diffusion framework for high-fidelity CT virtual contrast enhancement that improves anatomical consistency and detail.
- VibeGuard: A Security Gate Framework for AI-Generated Code (viability: 5): https://sciencetostartup.com/paper/vibeguard-a-security-gate-framework-for-ai-generated-code - VibeGuard is a pre-publish security gate for AI-generated code, addressing blind spots in artifact hygiene, packaging, and supply-chain risk.
- Adversarial Attacks in AI-Driven RAN Slicing: SLA Violations and Recovery (viability: 3): https://sciencetostartup.com/paper/adversarial-attacks-in-ai-driven-ran-slicing-sla-violations-and-recovery - Studies the impact of adversarial attacks on AI-driven RAN slicing decisions, quantifying SLA violations and recovery behavior.
- A global dataset of continuous urban dashcam driving (viability: 8): https://sciencetostartup.com/paper/a-global-dataset-of-continuous-urban-dashcam-driving - CROWD is a large-scale, manually curated urban dashcam dataset with temporal continuity and diverse global coverage, designed for cross-domain robustness analysis.
- ONE-SHOT: Compositional Human-Environment Video Synthesis via Spatial-Decoupled Motion Injection and Hybrid Context Integration (viability: 7): https://sciencetostartup.com/paper/one-shot-compositional-human-environment-video-synthesis-via-spatial-decoupled-motion-injection-and-hybrid-context-integ - A parameter-efficient framework for compositional human-environment video generation that disentangles human dynamics from environmental cues for precise control and creative diversity.
- Automated Framework to Evaluate and Harden LLM System Instructions against Encoding Attacks (viability: 7): https://sciencetostartup.com/paper/automated-framework-to-evaluate-and-harden-llm-system-instructions-against-encoding-attacks - An automated framework to evaluate and harden LLM system instructions against encoding attacks, preventing sensitive information leakage through structured output requests.
- Foundation Model-guided Iteratively Prompting and Pseudo-Labeling for Partially Labeled Medical Image Segmentation (viability: 7): https://sciencetostartup.com/paper/foundation-model-guided-iteratively-prompting-and-pseudo-labeling-for-partially-labeled-medical-image-segmentation - An iterative framework that leverages foundation models to generate and refine pseudo-labels for unlabeled organs in partially labeled medical images, improving segmentation performance.
- Aligning Recommendations with User Popularity Preferences (viability: 7): https://sciencetostartup.com/paper/aligning-recommendations-with-user-popularity-preferences - SPREE, an inference-time method for sequential recommenders that steers model activations to align recommendations with individual user popularity preferences, improving alignment while preserving quality.
- Stein Variational Uncertainty-Adaptive Model Predictive Control (viability: 3): https://sciencetostartup.com/paper/stein-variational-uncertainty-adaptive-model-predictive-control - A novel controller for nonlinear dynamical systems that uses Stein variational inference to adapt to latent parametric uncertainty, improving performance-robustness tradeoffs.
- Sub-metre Lunar DEM Generation and Validation from Chandrayaan-2 OHRC Multi-View Imagery Using Open-Source Photogrammetry (viability: 4): https://sciencetostartup.com/paper/sub-metre-lunar-dem-generation-and-validation-from-chandrayaan-2-ohrc-multi-view-imagery-using-open-source-photogrammetr - Generating sub-meter lunar digital elevation models from high-resolution orbital imagery using an open-source photogrammetry pipeline.
- Diff3R: Feed-forward 3D Gaussian Splatting with Uncertainty-aware Differentiable Optimization (viability: 7): https://sciencetostartup.com/paper/diff3r-feed-forward-3d-gaussian-splatting-with-uncertainty-aware-differentiable-optimization - A framework that combines feed-forward 3D Gaussian Splatting with uncertainty-aware optimization for faster, higher-quality 3D scene reconstruction.
- Revision or Re-Solving? Decomposing Second-Pass Gains in Multi-LLM Pipelines (viability: 5): https://sciencetostartup.com/paper/revision-or-re-solving-decomposing-second-pass-gains-in-multi-llm-pipelines - Decomposing multi-LLM pipeline gains into re-solving, scaffolding, and content correction to inform targeted pipeline designs.
- Fast and Accurate Probing of In-Training LLMs' Downstream Performances (viability: 4): https://sciencetostartup.com/paper/fast-and-accurate-probing-of-in-training-llms-downstream-performances - A new method to evaluate LLM performance during training, reducing evaluation time from hours to minutes.
- Model-Based Learning of Near-Optimal Finite-Window Policies in POMDPs (viability: 0): https://sciencetostartup.com/paper/model-based-learning-of-near-optimal-finite-window-policies-in-pomdps - Developing a sample-efficient model estimation procedure for learning policies in partially observable environments.
- Infinite-Horizon Ergodic Control via Kernel Mean Embeddings (viability: 5): https://sciencetostartup.com/paper/infinite-horizon-ergodic-control-via-kernel-mean-embeddings - An infinite-horizon ergodic controller using kernel mean embeddings for long-duration coverage tasks.
- Transfer learning for nonparametric Bayesian networks (viability: 5): https://sciencetostartup.com/paper/transfer-learning-for-nonparametric-bayesian-networks - Two transfer learning methods for nonparametric Bayesian networks to improve learning performance with scarce data.
- OrgAgent: Organize Your Multi-Agent System like a Company (viability: 3): https://sciencetostartup.com/paper/orgagent-organize-your-multi-agent-system-like-a-company - A hierarchical framework for organizing multi-agent systems like a company to improve reasoning, reduce costs, and enhance coordination.
- Forecasting Motion in the Wild (viability: 7): https://sciencetostartup.com/paper/forecasting-motion-in-the-wild - A diffusion transformer that forecasts complex motion patterns in animals by representing motion as visual tokens, enabling predictive visual intelligence.
- AutoMIA: Improved Baselines for Membership Inference Attack via Agentic Self-Exploration (viability: 7): https://sciencetostartup.com/paper/automia-improved-baselines-for-membership-inference-attack-via-agentic-self-exploration - An agentic framework that automates membership inference attacks by self-exploring and refining strategies, improving model auditing without manual feature engineering.
- PDA: Text-Augmented Defense Framework for Robust Vision-Language Models against Adversarial Image Attacks (viability: 7): https://sciencetostartup.com/paper/pda-text-augmented-defense-framework-for-robust-vision-language-models-against-adversarial-image-attacks - A training-free defense framework for vision-language models that uses text augmentation to enhance robustness against adversarial attacks without modifying the underlying models.
- OmniMem: Autoresearch-Guided Discovery of Lifelong Multimodal Agent Memory (viability: 8): https://sciencetostartup.com/paper/omnimem-autoresearch-guided-discovery-of-lifelong-multimodal-agent-memory - OmniMem is an autonomous multimodal memory system enhancing AI agents' lifelong memory with a 23-stage autoresearch pipeline.
- Query-Conditioned Evidential Keyframe Sampling for MLLM-Based Long-Form Video Understanding (viability: 7): https://sciencetostartup.com/paper/query-conditioned-evidential-keyframe-sampling-for-mllm-based-long-form-video-understanding - A principled framework for query-conditioned evidential keyframe sampling optimizes multimodal LLM performance on long-form videos by maximizing conditional mutual information.
- EgoSim: Egocentric World Simulator for Embodied Interaction Generation (viability: 8): https://sciencetostartup.com/paper/egosim-egocentric-world-simulator-for-embodied-interaction-generation - EgoSim is a closed-loop egocentric world simulator that generates spatially consistent interaction videos and updates 3D scene states, enabling scalable data collection and cross-embodiment transfer.
- EmbedPart: Embedding-Driven Graph Partitioning for Scalable Graph Neural Network Training (viability: 4): https://sciencetostartup.com/paper/embedpart-embedding-driven-graph-partitioning-for-scalable-graph-neural-network-training - EmbedPart is an embedding-driven graph partitioning approach that accelerates scalable distributed GNN training by clustering node embeddings.
- Customizing Large Vision Model-Guided Low-Rank Approximation for Ground-Roll Denoise (viability: 7): https://sciencetostartup.com/paper/customizing-large-vision-model-guided-low-rank-approximation-for-ground-roll-denoise - A training-free AI framework uses vision models to denoise seismic data, preserving reflections and outperforming existing methods without task-specific training.
- Uncertainty-Aware Variational Reward Factorization via Probabilistic Preference Bases for LLM Personalization (viability: 7): https://sciencetostartup.com/paper/uncertainty-aware-variational-reward-factorization-via-probabilistic-preference-bases-for-llm-personalization - An uncertainty-aware framework personalizes LLMs by modeling user preferences as distributions, improving accuracy and reliability over existing methods.
- Multimodal Analysis of State-Funded News Coverage of the Israel-Hamas War on YouTube Shorts (viability: 7): https://sciencetostartup.com/paper/multimodal-analysis-of-state-funded-news-coverage-of-the-israel-hamas-war-on-youtube-shorts - A multimodal pipeline analyzes state-funded news coverage of the Israel-Hamas war on YouTube Shorts, revealing sentiment and visual trends with efficient models.
- Bridging Structured Knowledge and Data: A Unified Framework with Finance Applications (viability: 3): https://sciencetostartup.com/paper/bridging-structured-knowledge-and-data-a-unified-framework-with-finance-applications - A unified framework embeds structured knowledge into neural networks for consistent estimation and improved financial applications.
- Do Phone-Use Agents Respect Your Privacy? (viability: 7): https://sciencetostartup.com/paper/do-phone-use-agents-respect-your-privacy - This research introduces a framework to evaluate and improve the privacy-preserving capabilities of phone-use AI agents, addressing a critical gap in current AI deployment.
- Maximizing T2-Only Prostate Cancer Localization from Expected Diffusion Weighted Imaging (viability: 6): https://sciencetostartup.com/paper/maximizing-t2-only-prostate-cancer-localization-from-expected-diffusion-weighted-imaging - A novel framework for prostate cancer localization using only T2-weighted MRI by leveraging diffusion-weighted imaging as a privileged latent modality during training.
- ACT Now: Preempting LVLM Hallucinations via Adaptive Context Integration (viability: 7): https://sciencetostartup.com/paper/act-now-preempting-lvlm-hallucinations-via-adaptive-context-integration - This research proposes an inference-time intervention method to significantly reduce hallucinations in Large Vision-Language Models by adaptively integrating contextual information.
- Dual Optimal: Make Your LLM Peer-like with Dignity (viability: 7): https://sciencetostartup.com/paper/dual-optimal-make-your-llm-peer-like-with-dignity - A novel framework and dataset to train LLMs that exhibit dignity and trustworthiness, moving beyond sycophantic and evasive behavior.
- Flow-based Policy With Distributional Reinforcement Learning in Trajectory Optimization (viability: 7): https://sciencetostartup.com/paper/flow-based-policy-with-distributional-reinforcement-learning-in-trajectory-optimization - A novel reinforcement learning algorithm using flow matching and distributional RL to improve policy representation and performance in complex control tasks.
- An Integrated Soft Robotic System for Measuring Vital Signs in Search and Rescue Environments (viability: 7): https://sciencetostartup.com/paper/an-integrated-soft-robotic-system-for-measuring-vital-signs-in-search-and-rescue-environments - An integrated soft robotic system for search and rescue that accurately measures victim pulse and blood pressure using advanced signal processing.
- DLWM: Dual Latent World Models enable Holistic Gaussian-centric Pre-training in Autonomous Driving (viability: 8): https://sciencetostartup.com/paper/dlwm-dual-latent-world-models-enable-holistic-gaussian-centric-pre-training-in-autonomous-driving - DLWM: A novel dual latent world model system for holistic Gaussian-centric pre-training in autonomous driving, significantly improving perception, forecasting, and planning.
- Rapid mixing in positively weighted restricted Boltzmann machines (viability: 0): https://sciencetostartup.com/paper/rapid-mixing-in-positively-weighted-restricted-boltzmann-machines - Theoretical analysis of mixing time bounds for positively weighted restricted Boltzmann machines.
- Enhancing Gradient Inversion Attacks in Federated Learning via Hierarchical Feature Optimization (viability: 7): https://sciencetostartup.com/paper/enhancing-gradient-inversion-attacks-in-federated-learning-via-hierarchical-feature-optimization - Enhance privacy in federated learning by optimizing gradient inversion attacks through hierarchical feature optimization, enabling pixel-level reconstruction and outperforming baselines.
- Phase transition on a context-sensitive random language model with short range interactions (viability: 2): https://sciencetostartup.com/paper/phase-transition-on-a-context-sensitive-random-language-model-with-short-range-interactions - Investigate phase transitions in context-sensitive random language models with short-range interactions to understand the intrinsic nature of language properties.
- Differentially Private Manifold Denoising (viability: 7): https://sciencetostartup.com/paper/differentially-private-manifold-denoising - A differentially private manifold denoising framework that allows sensitive reference datasets to correct noisy query points without compromising privacy, offering formal DP guarantees.
- YieldSAT: A Multimodal Benchmark Dataset for High-Resolution Crop Yield Prediction (viability: 8): https://sciencetostartup.com/paper/yieldsat-a-multimodal-benchmark-dataset-for-high-resolution-crop-yield-prediction - YieldSAT provides high-resolution crop yield predictions using a multimodal dataset to improve agricultural productivity efficiently.
- WARP: Guaranteed Inner-Layer Repair of NLP Transformers (viability: 7): https://sciencetostartup.com/paper/warp-guaranteed-inner-layer-repair-of-nlp-transformers - WARP provides provable inner-layer repair for NLP Transformers, enhancing robustness against adversarial attacks with a convex quadratic programming framework.
- EmoScene: A Dual-space Dataset for Controllable Affective Image Generation (viability: 8): https://sciencetostartup.com/paper/emoscene-a-dual-space-dataset-for-controllable-affective-image-generation - EmoScene is a dual-space dataset and controllable generation framework for nuanced emotional image synthesis using diffusion models.
- PsychAgent: An Experience-Driven Lifelong Learning Agent for Self-Evolving Psychological Counselor (viability: 7): https://sciencetostartup.com/paper/psychagent-an-experience-driven-lifelong-learning-agent-for-self-evolving-psychological-counselor - PsychAgent is a lifelong learning agent for psychological counseling that evolves skills from experience to improve multi-session interaction quality.
- Autoregressive Appearance Prediction for 3D Gaussian Avatars (viability: 7): https://sciencetostartup.com/paper/autoregressive-appearance-prediction-for-3d-gaussian-avatars - This work introduces a 3D Gaussian Splatting avatar model with autoregressive appearance prediction for temporally smooth and stable avatar driving.
- Learning Quantised Structure-Preserving Motion Representations for Dance Fingerprinting (viability: 7): https://sciencetostartup.com/paper/learning-quantised-structure-preserving-motion-representations-for-dance-fingerprinting - DANCEMATCH provides a scalable framework for identifying similar dance choreographies from raw video by creating discrete motion signatures.
- Positional Cognitive Specialization: Where Do LLMs Learn To Comprehend and Speak Your Language? (viability: 7): https://sciencetostartup.com/paper/positional-cognitive-specialization-where-do-llms-learn-to-comprehend-and-speak-your-language - CogSym offers a layer-wise heuristic to efficiently adapt large language models to new languages by fine-tuning only a few layers.
- Representation Selection via Cross-Model Agreement using Canonical Correlation Analysis (viability: 6): https://sciencetostartup.com/paper/representation-selection-via-cross-model-agreement-using-canonical-correlation-analysis - A training-free method using Canonical Correlation Analysis to select and reduce dimensionality of image representations, improving downstream performance.
- GPT-NL Public Corpus: A Permissively Licensed, Dutch-First Dataset for LLM Pre-training (viability: 4): https://sciencetostartup.com/paper/gpt-nl-public-corpus-a-permissively-licensed-dutch-first-dataset-for-llm-pre-training - GPT-NL Public Corpus is a large, permissively licensed dataset of Dutch language resources for LLM pre-training.
- Generalization Bounds for Spectral GNNs via Fourier Domain Analysis (viability: 1): https://sciencetostartup.com/paper/generalization-bounds-for-spectral-gnns-via-fourier-domain-analysis - Develops theoretical generalization bounds for spectral graph neural networks by analyzing their behavior in the graph Fourier domain.
- Investigating Autonomous Agent Contributions in the Wild: Activity Patterns and Code Change over Time (viability: 7): https://sciencetostartup.com/paper/investigating-autonomous-agent-contributions-in-the-wild-activity-patterns-and-code-change-over-time - Analyzes the real-world contributions and code churn of autonomous coding agents across major open-source projects, revealing increasing activity but higher maintenance needs.
- Orthogonal Learner for Estimating Heterogeneous Long-Term Treatment Effects (viability: 6): https://sciencetostartup.com/paper/orthogonal-learner-for-estimating-heterogeneous-long-term-treatment-effects - Introduces novel orthogonal learners for robust estimation of heterogeneous long-term treatment effects, particularly in settings with limited data overlap.
- Benchmarking and Mechanistic Analysis of Vision-Language Models for Cross-Depiction Assembly Instruction Alignment (viability: 7): https://sciencetostartup.com/paper/benchmarking-and-mechanistic-analysis-of-vision-language-models-for-cross-depiction-assembly-instruction-alignment - Benchmarks vision-language models on cross-depiction assembly instruction alignment, identifying visual encoding as the key bottleneck and releasing a new benchmark dataset.
- ProCap: Projection-Aware Captioning for Spatial Augmented Reality (viability: 8): https://sciencetostartup.com/paper/procap-projection-aware-captioning-for-spatial-augmented-reality - A framework for spatial augmented reality that separates projected content from physical scenes, enabling intelligent interaction by resolving virtual-physical ambiguity.
- Event Embedding of Protein Networks : Compositional Learning of Biological Function (viability: 2): https://sciencetostartup.com/paper/event-embedding-of-protein-networks-compositional-learning-of-biological-function - Investigating whether enforcing compositional structure in sequence embeddings improves geometric organization in protein-protein interaction networks.
- JAMMEval: A Refined Collection of Japanese Benchmarks for Reliable VLM Evaluation (viability: 8): https://sciencetostartup.com/paper/jammeval-a-refined-collection-of-japanese-benchmarks-for-reliable-vlm-evaluation - A refined collection of Japanese benchmarks for reliable vision-language model evaluation, addressing issues in existing datasets to improve model assessment.
- Fatigue-Aware Learning to Defer via Constrained Optimisation (viability: 7): https://sciencetostartup.com/paper/fatigue-aware-learning-to-defer-via-constrained-optimisation - A fatigue-aware learning to defer framework that models workload-varying human performance for optimized human-AI cooperation.
- IDDM: Identity-Decoupled Personalized Diffusion Models with a Tunable Privacy-Utility Trade-off (viability: 7): https://sciencetostartup.com/paper/iddm-identity-decoupled-personalized-diffusion-models-with-a-tunable-privacy-utility-trade-off - A diffusion model defense that decouples identity from personalized image generation, offering tunable privacy for social media.
- Experience as a Compass: Multi-agent RAG with Evolving Orchestration and Agent Prompts (viability: 7): https://sciencetostartup.com/paper/experience-as-a-compass-multi-agent-rag-with-evolving-orchestration-and-agent-prompts - A hierarchical framework that evolves multi-agent orchestration and prompts for complex, multi-hop reasoning tasks.
- Super-Resolving Coarse-Resolution Weather Forecasts With Flow Matching (viability: 4): https://sciencetostartup.com/paper/super-resolving-coarse-resolution-weather-forecasts-with-flow-matching - A generative super-resolution framework for weather forecasts that enhances spatial resolution post-processing.
- When Users Change Their Mind: Evaluating Interruptible Agents in Long-Horizon Web Navigation (viability: 7): https://sciencetostartup.com/paper/when-users-change-their-mind-evaluating-interruptible-agents-in-long-horizon-web-navigation - A benchmark and evaluation of LLM agents' ability to handle user interruptions during long-horizon web navigation tasks.
- Beyond Symbolic Solving: Multi Chain-of-Thought Voting for Geometric Reasoning in Large Language Models (viability: 7): https://sciencetostartup.com/paper/beyond-symbolic-solving-multi-chain-of-thought-voting-for-geometric-reasoning-in-large-language-models - A system that improves geometric reasoning in LLMs by generating and voting on multiple reasoning paths, verified with Python code execution.
- Adversarial Attenuation Patch Attack for SAR Object Detection (viability: 7): https://sciencetostartup.com/paper/adversarial-attenuation-patch-attack-for-sar-object-detection - A physically realizable adversarial attack for SAR object detection that uses an energy-constrained patch to deceive systems with high imperceptibility.
- PixelPrune: Pixel-Level Adaptive Visual Token Reduction via Predictive Coding (viability: 8): https://sciencetostartup.com/paper/pixelprune-pixel-level-adaptive-visual-token-reduction-via-predictive-coding - A training-free, pixel-level compression method that prunes redundant image patches before ViT encoding to accelerate document understanding and GUI interaction.
- KUET at StanceNakba Shared Task: StanceMoE: Mixture-of-Experts Architecture for Stance Detection (viability: 7): https://sciencetostartup.com/paper/kuet-at-stancenakba-shared-task-stancemoe-mixture-of-experts-architecture-for-stance-detection - A Mixture-of-Experts model built on BERT for actor-level stance detection, leveraging specialized modules to capture diverse linguistic signals.
- Accurate and Scalable Matrix Mechanisms via Divide and Conquer (viability: 4): https://sciencetostartup.com/paper/accurate-and-scalable-matrix-mechanisms-via-divide-and-conquer - A novel divide-and-conquer approach for scalable and accurate differentially private query answering and synthetic data generation.
- A 4D Representation for Training-Free Agentic Reasoning from Monocular Laparoscopic Video (viability: 8): https://sciencetostartup.com/paper/a-4d-representation-for-training-free-agentic-reasoning-from-monocular-laparoscopic-video - Enabling AI agents to perform spatiotemporal reasoning in surgery by grounding language models in a 4D representation of video, without additional training.
- Shape Representation using Gaussian Process mixture models (viability: 5): https://sciencetostartup.com/paper/shape-representation-using-gaussian-process-mixture-models - A lightweight functional shape representation using Gaussian Process mixture models for efficient and accurate 3D geometry encoding.
- Policy Improvement Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/policy-improvement-reinforcement-learning - A closed-loop reinforcement learning framework that self-corrects by verifying policy improvements across iterations for more stable and performant LLM training.
- Sparkle: A Robust and Versatile Representation for Point Cloud based Human Motion Capture (viability: 7): https://sciencetostartup.com/paper/sparkle-a-robust-and-versatile-representation-for-point-cloud-based-human-motion-capture - A novel structured representation and framework for robust and versatile point cloud-based human motion capture that outperforms state-of-the-art in accuracy and generalization.
- Perturb-and-Restore: Simulation-driven Structural Augmentation Framework for Imbalance Chromosomal Anomaly Detection (viability: 8): https://sciencetostartup.com/paper/perturb-and-restore-simulation-driven-structural-augmentation-framework-for-imbalance-chromosomal-anomaly-detection - A simulation-driven framework that generates synthetic chromosomal data to overcome severe imbalance and scarcity for state-of-the-art anomaly detection in genetic disorders.
- MotionGrounder: Grounded Multi-Object Motion Transfer via Diffusion Transformer (viability: 8): https://sciencetostartup.com/paper/motiongrounder-grounded-multi-object-motion-transfer-via-diffusion-transformer - MotionGrounder is a Diffusion Transformer framework enabling multi-object motion transfer with fine-grained control, grounding captions to specific objects in generated videos.
- PanoAir: A Panoramic Visual-Inertial SLAM with Cross-Time Real-World UAV Dataset (viability: 8): https://sciencetostartup.com/paper/panoair-a-panoramic-visual-inertial-slam-with-cross-time-real-world-uav-dataset - A panoramic Visual-Inertial SLAM framework with a novel real-world UAV dataset, offering superior accuracy and robustness for drone localization and mapping.
- Disentangling to Re-couple: Resolving the Similarity-Controllability Paradox in Subject-Driven Text-to-Image Generation (viability: 7): https://sciencetostartup.com/paper/disentangling-to-re-couple-resolving-the-similarity-controllability-paradox-in-subject-driven-text-to-image-generation - A framework that disentangles and recouples visual and textual information to resolve the similarity-controllability paradox in subject-driven text-to-image generation, achieving state-of-the-art results.
- Debiased Estimators in High-Dimensional Regression: A Review and Replication of Javanmard and Montanari (2014) (viability: 3): https://sciencetostartup.com/paper/debiased-estimators-in-high-dimensional-regression-a-review-and-replication-of-javanmard-and-montanari-2014 - Examines and replicates a debiased LASSO framework for high-dimensional regression, comparing its performance and power against the desparsified LASSO.
- Proactive Agent Research Environment: Simulating Active Users to Evaluate Proactive Assistants (viability: 8): https://sciencetostartup.com/paper/proactive-agent-research-environment-simulating-active-users-to-evaluate-proactive-assistants - Pare is a framework for building and evaluating proactive agents with stateful user simulation, featuring a benchmark of 143 diverse tasks across multiple applications.
- Agentic Tool Use in Large Language Models (viability: 3): https://sciencetostartup.com/paper/agentic-tool-use-in-large-language-models - Organizes and analyzes the literature on agentic tool use in large language models across three paradigms, highlighting challenges and evolutionary views.
- Learning to Learn-at-Test-Time: Language Agents with Learnable Adaptation Policies (viability: 7): https://sciencetostartup.com/paper/learning-to-learn-at-test-time-language-agents-with-learnable-adaptation-policies - A framework for learning adaptation policies in language agents to improve performance through iterative environmental interaction.
- LinguDistill: Recovering Linguistic Ability in Vision- Language Models via Selective Cross-Modal Distillation (viability: 7): https://sciencetostartup.com/paper/lingudistill-recovering-linguistic-ability-in-vision-language-models-via-selective-cross-modal-distillation - An adapter-free distillation method to restore linguistic ability in vision-language models by selectively distilling a frozen language model teacher.
- Video Patch Pruning: Efficient Video Instance Segmentation via Early Token Reduction (viability: 7): https://sciencetostartup.com/paper/video-patch-pruning-efficient-video-instance-segmentation-via-early-token-reduction - A video patch pruning framework that enables efficient sparsity in early Vision Transformer layers for video instance segmentation.
- Optimal Brain Decomposition for Accurate LLM Low-Rank Approximation (viability: 3): https://sciencetostartup.com/paper/optimal-brain-decomposition-for-accurate-llm-low-rank-approximation - A theoretical framework for optimal low-rank decomposition of LLM weights using second-order Hessian information.
- Continual Vision-Language Learning for Remote Sensing: Benchmarking and Analysis (viability: 4): https://sciencetostartup.com/paper/continual-vision-language-learning-for-remote-sensing-benchmarking-and-analysis - A benchmark and analysis of continual learning for remote sensing vision-language models to address catastrophic forgetting.
- Emotion Entanglement and Bayesian Inference for Multi-Dimensional Emotion Understanding (viability: 5): https://sciencetostartup.com/paper/emotion-entanglement-and-bayesian-inference-for-multi-dimensional-emotion-understanding - A theory-grounded benchmark and Bayesian inference framework for multi-dimensional emotion understanding in rich textual contexts.
- Multicentric thrombus segmentation using an attention-based recurrent network with gradual modality dropout (viability: 3): https://sciencetostartup.com/paper/multicentric-thrombus-segmentation-using-an-attention-based-recurrent-network-with-gradual-modality-dropout - An attention-based recurrent network with gradual modality dropout for multicentric thrombus segmentation in 3D brain scans.
- DVGT-2: Vision-Geometry-Action Model for Autonomous Driving at Scale (viability: 7): https://sciencetostartup.com/paper/dvgt-2-vision-geometry-action-model-for-autonomous-driving-at-scale - A streaming Vision-Geometry-Action model for end-to-end autonomous driving that jointly outputs dense geometry and trajectory planning.
- Cost-Penalized Fitness in FMA-Orchestrated Mixture of Experts: Experimental Evidence for Molecular Memory in Domain Adaptation (viability: 7): https://sciencetostartup.com/paper/cost-penalized-fitness-in-fma-orchestrated-mixture-of-experts-experimental-evidence-for-molecular-memory-in-domain-adapt - A novel Mixture-of-Experts management system for LLMs that learns 'molecular memory' for faster domain adaptation, promising significant cost and energy savings.
- Deconfounding Scores and Representation Learning for Causal Effect Estimation with Weak Overlap (viability: 3): https://sciencetostartup.com/paper/deconfounding-scores-and-representation-learning-for-causal-effect-estimation-with-weak-overlap - A theoretical framework for causal effect estimation that improves overlap conditions using novel 'deconfounding scores' for high-dimensional data.
- Revisiting Human-in-the-Loop Object Retrieval with Pre-Trained Vision Transformers (viability: 7): https://sciencetostartup.com/paper/revisiting-human-in-the-loop-object-retrieval-with-pre-trained-vision-transformers - An interactive object retrieval system that leverages pre-trained Vision Transformers and Active Learning to rapidly identify objects in complex images.
- Compact Keyframe-Optimized Multi-Agent Gaussian Splatting SLAM (viability: 8): https://sciencetostartup.com/paper/compact-keyframe-optimized-multi-agent-gaussian-splatting-slam - A compact multi-agent SLAM system using Gaussian Splatting that significantly reduces communication bandwidth for real-time robotic mapping.
- Routing-Free Mixture-of-Experts (viability: 7): https://sciencetostartup.com/paper/routing-free-mixture-of-experts - A novel Mixture-of-Experts model that eliminates centralized routing for improved scalability and robustness in LLMs.
- MIRANDA: MId-feature RANk-adversarial Domain Adaptation toward climate change-robust ecological forecasting with deep learning (viability: 7): https://sciencetostartup.com/paper/miranda-mid-feature-rank-adversarial-domain-adaptation-toward-climate-change-robust-ecological-forecasting-with-deep-lea - A domain adaptation technique for deep learning that improves ecological forecasting robustness to climate change by adapting intermediate features.
- Multimodal Language Models Cannot Spot Spatial Inconsistencies (viability: 3): https://sciencetostartup.com/paper/multimodal-language-models-cannot-spot-spatial-inconsistencies - Demonstrates that current multimodal language models fail to detect spatial inconsistencies in 3D scenes, highlighting a gap in their physical understanding.
- Preference Guided Iterated Pareto Referent Optimisation for Accessible Route Planning (viability: 4): https://sciencetostartup.com/paper/preference-guided-iterated-pareto-referent-optimisation-for-accessible-route-planning - An interactive route planning algorithm for users with accessibility needs that prioritizes user feedback for efficient optimization.
- HICT: High-precision 3D CBCT reconstruction from a single X-ray (viability: 7): https://sciencetostartup.com/paper/hict-high-precision-3d-cbct-reconstruction-from-a-single-x-ray - A two-stage framework and dataset for high-precision 3D dental CBCT reconstruction from single X-rays, improving diagnosis and treatment planning.
- RefineRL: Advancing Competitive Programming with Self-Refinement Reinforcement Learning (viability: 8): https://sciencetostartup.com/paper/refinerl-advancing-competitive-programming-with-self-refinement-reinforcement-learning - RefineRL enhances LLMs for competitive programming by enabling self-refinement through a skeptical agent and reinforcement learning, outperforming larger models.
- Valency Classification of Mapudungun Verbal Roots. Established by the language's own morphotactics (viability: 0): https://sciencetostartup.com/paper/valency-classification-of-mapudungun-verbal-roots-established-by-the-language-s-own-morphotactics - Classifying the valency of Mapudungun verbal roots using the language's own morphotactics to improve a morphological analyzer.
- UK AISI Alignment Evaluation Case-Study (viability: 5): https://sciencetostartup.com/paper/uk-aisi-alignment-evaluation-case-study - Evaluating frontier AI models for research sabotage when deployed as coding assistants, finding some models refuse safety-relevant tasks.
- Scalable Pretraining of Large Mixture of Experts Language Models on Aurora Super Computer (viability: 2): https://sciencetostartup.com/paper/scalable-pretraining-of-large-mixture-of-experts-language-models-on-aurora-super-computer - This paper details the large-scale pretraining of Mixture of Experts language models on a supercomputer, focusing on optimizing training speed and stability.
- An Approach to Enriching Surgical Video Datasets for Fine-Grained Spatial-Temporal Understanding of Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/an-approach-to-enriching-surgical-video-datasets-for-fine-grained-spatial-temporal-understanding-of-vision-language-mode - A pipeline to generate enriched surgical video datasets for fine-grained spatial-temporal understanding, improving vision-language model performance.
- Using predefined vector systems to speed up neural network multimillion class classification (viability: 5): https://sciencetostartup.com/paper/using-predefined-vector-systems-to-speed-up-neural-network-multimillion-class-classification - A method to speed up neural network classification for millions of classes by reducing label prediction complexity.
- From Early Encoding to Late Suppression: Interpreting LLMs on Character Counting Tasks (viability: 4): https://sciencetostartup.com/paper/from-early-encoding-to-late-suppression-interpreting-llms-on-character-counting-tasks - Interpreting LLM failures on character counting tasks to reveal internal reasoning and output suppression mechanisms.
- From Baselines to Preferences: A Comparative Study of LoRA/QLoRA and Preference Optimization for Mental Health Text Classification (viability: 7): https://sciencetostartup.com/paper/from-baselines-to-preferences-a-comparative-study-of-lora-qlora-and-preference-optimization-for-mental-health-text-class - A comparative study of LoRA/QLoRA and preference optimization for mental health text classification, providing a practical framework for choosing effective training strategies.
- Thinking Wrong in Silence: Backdoor Attacks on Continuous Latent Reasoning (viability: 8): https://sciencetostartup.com/paper/thinking-wrong-in-silence-backdoor-attacks-on-continuous-latent-reasoning - ThoughtSteer: A novel backdoor attack on continuous latent reasoning in language models that evades existing defenses and achieves high success rates.
- A Dual-Action Fabric-Based Soft Robotic Glove for Ergonomic Hand Rehabilitation (viability: 6): https://sciencetostartup.com/paper/a-dual-action-fabric-based-soft-robotic-glove-for-ergonomic-hand-rehabilitation - A dual-action fabric-based soft robotic glove with customized actuators for ergonomic hand rehabilitation, showing reduced muscle activity and improved grasp patterns.
- ActivityNarrated: An Open-Ended Narrative Paradigm for Wearable Human Activity Understanding (viability: 7): https://sciencetostartup.com/paper/activitynarrated-an-open-ended-narrative-paradigm-for-wearable-human-activity-understanding - ActivityNarrated: An open-ended narrative paradigm for wearable human activity understanding that shifts from closed-set classification to open-vocabulary sensor-language alignment.
- PrivHAR-Bench: A Graduated Privacy Benchmark Dataset for Video-Based Action Recognition (viability: 5): https://sciencetostartup.com/paper/privhar-bench-a-graduated-privacy-benchmark-dataset-for-video-based-action-recognition - A standardized benchmark dataset and evaluation toolkit for privacy-preserving video action recognition, enabling nuanced analysis of the privacy-utility trade-off.
- IWP: Token Pruning as Implicit Weight Pruning in Large Vision Language Models (viability: 7): https://sciencetostartup.com/paper/iwp-token-pruning-as-implicit-weight-pruning-in-large-vision-language-models - A training-free token pruning framework for large vision language models that quantifies token importance and uses efficient selection to reduce computational cost.
- Stochastic Attention: Connectome-Inspired Randomized Routing for Expressive Linear-Time Attention (viability: 4): https://sciencetostartup.com/paper/stochastic-attention-connectome-inspired-randomized-routing-for-expressive-linear-time-attention - Stochastic Attention, inspired by fruit fly connectomes, enhances efficient attention mechanisms by introducing random routing for improved expressivity and receptive field growth.
- A wearable haptic device for edge and surface simulation (viability: 5): https://sciencetostartup.com/paper/a-wearable-haptic-device-for-edge-and-surface-simulation - A compact, lightweight wearable haptic device that simulates distinct edge and surface contact feedback for enhanced object manipulation in VR.
- How to Train your Tactile Model: Tactile Perception with Multi-fingered Robot Hands (viability: 7): https://sciencetostartup.com/paper/how-to-train-your-tactile-model-tactile-perception-with-multi-fingered-robot-hands - A Vision Transformer-based tactile perception model that generalizes to new robot hand sensors, reducing data collection and retraining needs for scalable robotic manipulation.
- BioCOMPASS: Integrating Biomarkers into Transformer-Based Immunotherapy Response Prediction (viability: 6): https://sciencetostartup.com/paper/biocompass-integrating-biomarkers-into-transformer-based-immunotherapy-response-prediction - BioCOMPASS integrates biomarkers and treatment information into transformer models to improve the generalizability of immunotherapy response prediction.
- SoftHand Model-W: A 3D-Printed, Anthropomorphic, Underactuated Robot Hand with Integrated Wrist and Carpal Tunnel (viability: 4): https://sciencetostartup.com/paper/softhand-model-w-a-3d-printed-anthropomorphic-underactuated-robot-hand-with-integrated-wrist-and-carpal-tunnel - A 3D-printed, anthropomorphic robot hand with an integrated wrist and carpal tunnel design that enables more versatile and human-like manipulation.
- Spectral Compact Training: Pre-Training Large Language Models via Permanent Truncated SVD and Stiefel QR Retraction (viability: 7): https://sciencetostartup.com/paper/spectral-compact-training-pre-training-large-language-models-via-permanent-truncated-svd-and-stiefel-qr-retraction - Spectral Compact Training (SCT) enables training large language models on consumer hardware by replacing dense weight matrices with compact SVD factors, drastically reducing memory usage.
- A CEFR-Inspired Classification Framework with Fuzzy C-Means To Automate Assessment of Programming Skills in Scratch (viability: 3): https://sciencetostartup.com/paper/a-cefr-inspired-classification-framework-with-fuzzy-c-means-to-automate-assessment-of-programming-skills-in-scratch - A framework for assessing programming skills in Scratch using fuzzy clustering, inspired by CEFR levels, to identify learning gaps and guide curriculum design.
- GRASP: Gradient Realignment via Active Shared Perception for Multi-Agent Collaborative Optimization (viability: 3): https://sciencetostartup.com/paper/grasp-gradient-realignment-via-active-shared-perception-for-multi-agent-collaborative-optimization - A novel framework for multi-agent collaboration that uses active shared perception to overcome non-stationarity and accelerate convergence.
- CircuitProbe: Predicting Reasoning Circuits in Transformers via Stability Zone Detection (viability: 7): https://sciencetostartup.com/paper/circuitprobe-predicting-reasoning-circuits-in-transformers-via-stability-zone-detection - A tool that predicts reasoning circuits in transformers in minutes, enabling efficient optimization of small language models.
- To Memorize or to Retrieve: Scaling Laws for RAG-Considerate Pretraining (viability: 4): https://sciencetostartup.com/paper/to-memorize-or-to-retrieve-scaling-laws-for-rag-considerate-pretraining - A framework for understanding the trade-offs between pretraining and retrieval for language models, guiding optimal data allocation.
- AfrIFact: Cultural Information Retrieval, Evidence Extraction and Fact Checking for African Languages (viability: 7): https://sciencetostartup.com/paper/afrifact-cultural-information-retrieval-evidence-extraction-and-fact-checking-for-african-languages - A dataset and evaluation framework for fact-checking in ten African languages, addressing critical information gaps.
- AutoEG: Exploiting Known Third-Party Vulnerabilities in Black-Box Web Applications (viability: 4): https://sciencetostartup.com/paper/autoeg-exploiting-known-third-party-vulnerabilities-in-black-box-web-applications - An automated multi-agent framework for generating exploits against known vulnerabilities in black-box web applications.
- Enhancing REST API Fuzzing with Access Policy Violation Checks and Injection Attacks (viability: 4): https://sciencetostartup.com/paper/enhancing-rest-api-fuzzing-with-access-policy-violation-checks-and-injection-attacks - Enhancing REST API fuzzing with novel oracles for access policy violations and injection attacks, generating executable test cases.
- Learning to Hint for Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/learning-to-hint-for-reinforcement-learning - A framework that jointly trains a hint-generating policy and a reasoner policy to overcome advantage collapse in reinforcement learning.
- Inverse-Free Sparse Variational Gaussian Processes (viability: 6): https://sciencetostartup.com/paper/inverse-free-sparse-variational-gaussian-processes - An inverse-free approach for sparse variational Gaussian processes that uses only matrix multiplications for improved stability and speed.
- TTA-Vid: Generalized Test-Time Adaptation for Video Reasoning (viability: 7): https://sciencetostartup.com/paper/tta-vid-generalized-test-time-adaptation-for-video-reasoning - Adapt pre-trained video reasoning models to new domains at test-time without labels using reinforcement learning, outperforming state-of-the-art.
- Internal APIs Are All You Need: Shadow APIs, Shared Discovery, and the Case Against Browser-First Agent Architectures (viability: 8): https://sciencetostartup.com/paper/internal-apis-are-all-you-need-shadow-apis-shared-discovery-and-the-case-against-browser-first-agent-architectures - Unbrowse transforms web interaction for agents by converting redundant browser discoveries into a shared API index, vastly improving speed and efficiency.
- Performance of Neural and Polynomial Operator Surrogates (viability: 3): https://sciencetostartup.com/paper/performance-of-neural-and-polynomial-operator-surrogates - Compares neural and polynomial operator surrogates for parametric PDEs, finding no universally superior method and highlighting the importance of matching methodology to problem characteristics.
- OmniVoice: Towards Omnilingual Zero-Shot Text-to-Speech with Diffusion Language Models (viability: 8): https://sciencetostartup.com/paper/omnivoice-towards-omnilingual-zero-shot-text-to-speech-with-diffusion-language-models - OmniVoice is a zero-shot text-to-speech model supporting over 600 languages, achieved through a novel diffusion language model architecture and LLM initialization.
- Full-Gradient Successor Feature Representations (viability: 4): https://sciencetostartup.com/paper/full-gradient-successor-feature-representations - A novel reinforcement learning algorithm that improves transfer learning by optimizing successor features with full gradient updates, offering theoretical convergence guarantees and empirical performance gains.
- TP-Seg: Task-Prototype Framework for Unified Medical Lesion Segmentation (viability: 7): https://sciencetostartup.com/paper/tp-seg-task-prototype-framework-for-unified-medical-lesion-segmentation - A unified medical lesion segmentation framework that uses task-aware adapters and prototype-guided decoders to achieve state-of-the-art performance across diverse imaging modalities and lesion types.
- Convergence of projected stochastic natural gradient variational inference for various step size and sample or batch size schedules (viability: 0): https://sciencetostartup.com/paper/convergence-of-projected-stochastic-natural-gradient-variational-inference-for-various-step-size-and-sample-or-batch-siz - Theoretical convergence analysis of projected stochastic natural gradient variational inference under various step size and sample/batch size schedules, providing new non-asymptotic results.
- MoonAnything: A Vision Benchmark with Large-Scale Lunar Supervised Data (viability: 7): https://sciencetostartup.com/paper/moonanything-a-vision-benchmark-with-large-scale-lunar-supervised-data - MoonAnything: A large-scale lunar vision benchmark with comprehensive geometric and photometric supervision, enabling robust perception for exploration missions and beyond.
- CL-VISTA: Benchmarking Continual Learning in Video Large Language Models (viability: 7): https://sciencetostartup.com/paper/cl-vista-benchmarking-continual-learning-in-video-large-language-models - A new benchmark for continual learning in Video Large Language Models that exposes catastrophic forgetting and evaluates performance, efficiency, and memory.
- Common TF-IDF variants arise as key components in the test statistic of a penalized likelihood-ratio test for word burstiness (viability: 2): https://sciencetostartup.com/paper/common-tf-idf-variants-arise-as-key-components-in-the-test-statistic-of-a-penalized-likelihood-ratio-test-for-word-burst - This paper provides a statistical perspective on TF-IDF, showing how it arises from a penalized likelihood-ratio test for word burstiness.
- Embedded Variational Neural Stochastic Differential Equations for Learning Heterogeneous Dynamics (viability: 7): https://sciencetostartup.com/paper/embedded-variational-neural-stochastic-differential-equations-for-learning-heterogeneous-dynamics - A Variational Neural Stochastic Differential Equation model that combines VAEs and Neural SDEs to learn heterogeneous temporal dynamics for complex socioeconomic data.
- TRIMS: Trajectory-Ranked Instruction Masked Supervision for Diffusion Language Models (viability: 8): https://sciencetostartup.com/paper/trims-trajectory-ranked-instruction-masked-supervision-for-diffusion-language-models - TRIMS is a trajectory-guided fine-tuning method for Diffusion Language Models that improves accuracy-parallelism trade-offs with lower training cost.
- Streaming Model Cascades for Semantic SQL (viability: 7): https://sciencetostartup.com/paper/streaming-model-cascades-for-semantic-sql - Develops adaptive cascade algorithms for streaming, per-partition execution of large language models in data warehouses to reduce inference costs.
- LibScan: Smart Contract Library Misuse Detection with Iterative Feedback and Static Verification (viability: 8): https://sciencetostartup.com/paper/libscan-smart-contract-library-misuse-detection-with-iterative-feedback-and-static-verification - LibScan is an automated framework that detects smart contract library misuse by combining LLM reasoning with static code analysis and iterative feedback.
- Chameleons do not Forget: Prompt-Based Online Continual Learning for Next Activity Prediction (viability: 8): https://sciencetostartup.com/paper/chameleons-do-not-forget-prompt-based-online-continual-learning-for-next-activity-prediction - CNAPwP is a prompt-based continual learning approach for next activity prediction that mitigates catastrophic forgetting in dynamic process environments.
- When AI and Experts Agree on Error: Intrinsic Ambiguity in Dermatoscopic Images (viability: 8): https://sciencetostartup.com/paper/when-ai-and-experts-agree-on-error-intrinsic-ambiguity-in-dermatoscopic-images - This research identifies intrinsic ambiguity in dermatoscopic images that challenges both AI and human experts, suggesting a new avenue for diagnostic AI development.
- DirectFisheye-GS: Enabling Native Fisheye Input in Gaussian Splatting with Cross-View Joint Optimization (viability: 7): https://sciencetostartup.com/paper/directfisheye-gs-enabling-native-fisheye-input-in-gaussian-splatting-with-cross-view-joint-optimization - Enabling native fisheye input for 3D Gaussian Splatting with cross-view joint optimization to improve reconstruction quality and reduce artifacts.
- On rankings in multiplayer games with an application to the game of Whist (viability: 4): https://sciencetostartup.com/paper/on-rankings-in-multiplayer-games-with-an-application-to-the-game-of-whist - A novel extension of the Bradley-Terry model for multiplayer games with an adapted algorithm for ranking.
- LiPS: Lightweight Panoptic Segmentation for Resource-Constrained Robotics (viability: 8): https://sciencetostartup.com/paper/lips-lightweight-panoptic-segmentation-for-resource-constrained-robotics - LiPS: A lightweight panoptic segmentation model for resource-constrained robotics, offering comparable accuracy with significantly higher throughput and lower computation.
- StretchBot: A Neuro-Symbolic Framework for Adaptive Guidance with Assistive Robots (viability: 5): https://sciencetostartup.com/paper/stretchbot-a-neuro-symbolic-framework-for-adaptive-guidance-with-assistive-robots - StretchBot: A neuro-symbolic framework for adaptive guidance in assistive robots, combining multimodal perception with LLM reasoning for context-aware adjustments.
- When Safe Models Merge into Danger: Exploiting Latent Vulnerabilities in LLM Fusion (viability: 7): https://sciencetostartup.com/paper/when-safe-models-merge-into-danger-exploiting-latent-vulnerabilities-in-llm-fusion - A framework for embedding latent malicious components into LLMs that only manifest when models are merged, compromising safety without detection in individual models.
- A Survey of On-Policy Distillation for Large Language Models (viability: 2): https://sciencetostartup.com/paper/a-survey-of-on-policy-distillation-for-large-language-models - A survey providing a unified framework and comprehensive overview of on-policy distillation techniques for large language models, addressing exposure bias.
- English to Central Kurdish Speech Translation: Corpus Creation, Evaluation, and Orthographic Standardization (viability: 7): https://sciencetostartup.com/paper/english-to-central-kurdish-speech-translation-corpus-creation-evaluation-and-orthographic-standardization - A new speech-to-text translation dataset for Central Kurdish derived from TED talks, enabling improved translation performance through orthographic standardization.
- A Physical Imitation Learning Pipeline for Energy-Efficient Quadruped Locomotion Assisted by Parallel Elastic Joint (viability: 3): https://sciencetostartup.com/paper/a-physical-imitation-learning-pipeline-for-energy-efficient-quadruped-locomotion-assisted-by-parallel-elastic-joint - Physical Imitation Learning distills RL policies into passive body dynamics for energy-efficient quadruped locomotion, reducing control effort through parallel elastic joints.
- Speech LLMs are Contextual Reasoning Transcribers (viability: 3): https://sciencetostartup.com/paper/speech-llms-are-contextual-reasoning-transcribers - A novel approach to automatic speech recognition that leverages large language models for contextual reasoning and user-guided transcription.
- TALENT: Target-aware Efficient Tuning for Referring Image Segmentation (viability: 7): https://sciencetostartup.com/paper/talent-target-aware-efficient-tuning-for-referring-image-segmentation - TALENT is a parameter-efficient tuning framework for referring image segmentation that addresses non-target activation issues to improve accuracy.
- Fluently Lying: Adversarial Robustness Can Be Substrate-Dependent (viability: 3): https://sciencetostartup.com/paper/fluently-lying-adversarial-robustness-can-be-substrate-dependent - This research identifies a new adversarial failure mode in spiking neural network object detectors where detection count is preserved but accuracy collapses.
- KG-CMI: Knowledge graph enhanced cross-Mamba interaction for medical visual question answering (viability: 7): https://sciencetostartup.com/paper/kg-cmi-knowledge-graph-enhanced-cross-mamba-interaction-for-medical-visual-question-answering - KG-CMI is a knowledge graph enhanced framework for medical visual question answering that improves accuracy and supports free-form answers.
- Predicting Dynamics of Ultra-Large Complex Systems by Inferring Governing Equations (viability: 7): https://sciencetostartup.com/paper/predicting-dynamics-of-ultra-large-complex-systems-by-inferring-governing-equations - A framework for discovering governing equations of complex systems at scale, enabling interpretable and reliable long-term predictions for applications like climate and biology.
- Towards Viewpoint-Robust End-to-End Autonomous Driving with 3D Foundation Model Priors (viability: 5): https://sciencetostartup.com/paper/towards-viewpoint-robust-end-to-end-autonomous-driving-with-3d-foundation-model-priors - Leveraging 3D foundation model priors to improve viewpoint robustness in end-to-end autonomous driving trajectory planning.
- Agent psychometrics: Task-level performance prediction in agentic coding benchmarks (viability: 7): https://sciencetostartup.com/paper/agent-psychometrics-task-level-performance-prediction-in-agentic-coding-benchmarks - A framework for predicting task-level performance of LLM agents in coding benchmarks, enabling better task design and agent evaluation.
- HarassGuard: Detecting Harassment Behaviors in Social Virtual Reality with Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/harassguard-detecting-harassment-behaviors-in-social-virtual-reality-with-vision-language-models - HarassGuard: A privacy-preserving vision-language model system for detecting physical harassment in social VR using only visual input.
- UniMixer: A Unified Architecture for Scaling Laws in Recommendation Systems (viability: 3): https://sciencetostartup.com/paper/unimixer-a-unified-architecture-for-scaling-laws-in-recommendation-systems - A unified architecture for recommendation systems that unifies mainstream scaling blocks and optimizes scaling efficiency.
- More Human, More Efficient: Aligning Annotations with Quantized SLMs (viability: 7): https://sciencetostartup.com/paper/more-human-more-efficient-aligning-annotations-with-quantized-slms - Fine-tune quantized small language models on limited human data to create deterministic, highly aligned evaluators and annotators.
- Representation choice shapes the interpretation of protein conformational dynamics (viability: 3): https://sciencetostartup.com/paper/representation-choice-shapes-the-interpretation-of-protein-conformational-dynamics - A library for computing and analyzing multiple protein representations to provide a comparative framework for molecular dynamics simulations.
- A Japanese Benchmark for Evaluating Social Bias in Reasoning Based on Attribution Theory (viability: 4): https://sciencetostartup.com/paper/a-japanese-benchmark-for-evaluating-social-bias-in-reasoning-based-on-attribution-theory - A Japanese benchmark dataset to evaluate social bias in LLM reasoning based on attribution theory, sensitive to cultural nuances.
- Quantum-Safe Code Auditing: LLM-Assisted Static Analysis and Quantum-Aware Risk Scoring for Post-Quantum Cryptography Migration (viability: 7): https://sciencetostartup.com/paper/quantum-safe-code-auditing-llm-assisted-static-analysis-and-quantum-aware-risk-scoring-for-post-quantum-cryptography-mig - An LLM-assisted static analysis framework for auditing code for quantum-vulnerable cryptography and scoring migration risk.
- FecalFed: Privacy-Preserving Poultry Disease Detection via Federated Learning (viability: 8): https://sciencetostartup.com/paper/fecalfed-privacy-preserving-poultry-disease-detection-via-federated-learning - A privacy-preserving federated learning framework for poultry disease detection using fecal imaging, with a curated and deduplicated dataset.
- STAR: Mitigating Cascading Errors in Spatial Reasoning via Turn-point Alignment and Segment-level DPO (viability: 7): https://sciencetostartup.com/paper/star-mitigating-cascading-errors-in-spatial-reasoning-via-turn-point-alignment-and-segment-level-dpo - STAR framework mitigates cascading errors in LLM spatial reasoning through turn-point alignment and segment-level DPO, achieving state-of-the-art performance.
- Multi-Camera View Scaling for Data-Efficient Robot Imitation Learning (viability: 7): https://sciencetostartup.com/paper/multi-camera-view-scaling-for-data-efficient-robot-imitation-learning - A framework for data-efficient robot imitation learning by scaling camera views during demonstration collection to improve generalization.
- HabitatAgent: An End-to-End Multi-Agent System for Housing Consultation (viability: 7): https://sciencetostartup.com/paper/habitatagent-an-end-to-end-multi-agent-system-for-housing-consultation - An LLM-powered multi-agent system for end-to-end housing consultation that significantly outperforms baselines in accuracy and reliability.
- Ontology-Constrained Neural Reasoning in Enterprise Agentic Systems: A Neurosymbolic Architecture for Domain-Grounded AI Agents (viability: 8): https://sciencetostartup.com/paper/ontology-constrained-neural-reasoning-in-enterprise-agentic-systems-a-neurosymbolic-architecture-for-domain-grounded-ai - A neurosymbolic architecture for enterprise AI agents that enforces regulatory compliance and domain grounding, outperforming ungrounded agents.
- Scenario theory for multi-criteria data-driven decision making (viability: 3): https://sciencetostartup.com/paper/scenario-theory-for-multi-criteria-data-driven-decision-making - A generalized scenario theory for multi-criteria data-driven decision making that provides more accurate robustness certificates.
- BloClaw: An Omniscient, Multi-Modal Agentic Workspace for Next-Generation Scientific Discovery (viability: 7): https://sciencetostartup.com/paper/bloclaw-an-omniscient-multi-modal-agentic-workspace-for-next-generation-scientific-discovery - BloClaw is a robust AI operating system for scientific discovery, enhancing data visualization and computational workflows with state-driven interfaces.
- TF-SSD: A Strong Pipeline via Synergic Mask Filter for Training-free Co-salient Object Detection (viability: 7): https://sciencetostartup.com/paper/tf-ssd-a-strong-pipeline-via-synergic-mask-filter-for-training-free-co-salient-object-detection - A training-free co-salient object detection pipeline leveraging SAM and DINO for improved generalization and performance.
- Reliev3R: Relieving Feed-forward Reconstruction from Multi-View Geometric Annotations (viability: 6): https://sciencetostartup.com/paper/reliev3r-relieving-feed-forward-reconstruction-from-multi-view-geometric-annotations - A weakly-supervised paradigm for training feed-forward 3D reconstruction models using only relative depths and sparse correspondences.
- Does Unification Come at a Cost? Uni-SafeBench: A Safety Benchmark for Unified Multimodal Large Models (viability: 7): https://sciencetostartup.com/paper/does-unification-come-at-a-cost-uni-safebench-a-safety-benchmark-for-unified-multimodal-large-models - Uni-SafeBench: A benchmark and evaluation framework to assess the safety of unified multimodal large models.
- Lightweight, Practical Encrypted Face Recognition with GPU Support (viability: 4): https://sciencetostartup.com/paper/lightweight-practical-encrypted-face-recognition-with-gpu-support - Lightweight and practical encrypted face recognition with GPU acceleration, reducing memory and improving runtime.
- Neuropsychiatric Deviations From Normative Profiles: An MRI-Derived Marker for Early Alzheimer's Disease Detection (viability: 4): https://sciencetostartup.com/paper/neuropsychiatric-deviations-from-normative-profiles-an-mri-derived-marker-for-early-alzheimer-s-disease-detection - A deep learning framework using MRI-derived brain anatomy to predict early Alzheimer's disease conversion from neuropsychiatric symptoms.
- TRiGS: Temporal Rigid-Body Motion for Scalable 4D Gaussian Splatting (viability: 7): https://sciencetostartup.com/paper/trigs-temporal-rigid-body-motion-for-scalable-4d-gaussian-splatting - A novel 4D Gaussian Splatting representation that uses continuous geometric transformations for scalable and temporally stable dynamic scene reconstruction.
- MATHENA: Mamba-based Architectural Tooth Hierarchical Estimator and Holistic Evaluation Network for Anatomy (viability: 7): https://sciencetostartup.com/paper/mathena-mamba-based-architectural-tooth-hierarchical-estimator-and-holistic-evaluation-network-for-anatomy - A unified Mamba-based framework for comprehensive dental diagnosis from OPGs, including tooth detection, caries segmentation, anomaly detection, and developmental staging.
- Optimsyn: Influence-Guided Rubrics Optimization for Synthetic Data Generation (viability: 4): https://sciencetostartup.com/paper/optimsyn-influence-guided-rubrics-optimization-for-synthetic-data-generation - An influence-guided framework that optimizes synthetic data generation for LLMs by using target model feedback to adapt data curation rubrics.
- FreqPhys: Repurposing Implicit Physiological Frequency Prior for Robust Remote Photoplethysmography (viability: 7): https://sciencetostartup.com/paper/freqphys-repurposing-implicit-physiological-frequency-prior-for-robust-remote-photoplethysmography - A frequency-guided framework for robust contactless physiological monitoring from facial videos, outperforming existing methods under challenging conditions.
- Learning from Many and Adapting to the Unknown in Open-set Test Streams (viability: 7): https://sciencetostartup.com/paper/learning-from-many-and-adapting-to-the-unknown-in-open-set-test-streams - A parameter-efficient LLM adaptation method that preserves source knowledge while rapidly specializing to new tasks in evolving, open-set environments.
- Learning Shared Representations for Multi-Task Linear Bandits (viability: 2): https://sciencetostartup.com/paper/learning-shared-representations-for-multi-task-linear-bandits - A theoretical framework for multi-task linear bandits that improves sample efficiency by learning shared low-rank representations.
- AceTone: Bridging Words and Colors for Conditional Image Grading (viability: 8): https://sciencetostartup.com/paper/acetone-bridging-words-and-colors-for-conditional-image-grading - A unified framework for conditional image grading that bridges words and colors, producing visually pleasing and stylistically coherent results aligned with human aesthetics.
- MF-QAT: Multi-Format Quantization-Aware Training for Elastic Inference (viability: 3): https://sciencetostartup.com/paper/mf-qat-multi-format-quantization-aware-training-for-elastic-inference - A method for training a single model to be robust across multiple quantization formats for elastic inference.
- Think, Act, Build: An Agentic Framework with Vision Language Models for Zero-Shot 3D Visual Grounding (viability: 8): https://sciencetostartup.com/paper/think-act-build-an-agentic-framework-with-vision-language-models-for-zero-shot-3d-visual-grounding - An agentic framework using Vision Language Models to perform zero-shot 3D visual grounding by dynamically reconstructing targets from RGB-D streams.
- Lipschitz Dueling Bandits over Continuous Action Spaces (viability: 3): https://sciencetostartup.com/paper/lipschitz-dueling-bandits-over-continuous-action-spaces - A novel algorithm for dueling bandits over continuous action spaces with Lipschitz structure, achieving logarithmic space complexity.
- Learnability-Guided Diffusion for Dataset Distillation (viability: 7): https://sciencetostartup.com/paper/learnability-guided-diffusion-for-dataset-distillation - A learnability-guided diffusion method for dataset distillation that reduces redundancy and improves model training efficiency.
- Toward Optimal Sampling Rate Selection and Unbiased Classification for Precise Animal Activity Recognition (viability: 4): https://sciencetostartup.com/paper/toward-optimal-sampling-rate-selection-and-unbiased-classification-for-precise-animal-activity-recognition - A novel network for animal activity recognition that customizes features and calibrates classifiers to improve accuracy across all behaviors, even with imbalanced data.
- MAESIL: Masked Autoencoder for Enhanced Self-supervised Medical Image Learning (viability: 7): https://sciencetostartup.com/paper/maesil-masked-autoencoder-for-enhanced-self-supervised-medical-image-learning - A self-supervised learning framework for 3D medical imaging that captures structural information using 'superpatches' and a masked autoencoder strategy, outperforming existing methods.
- MOON3.0: Reasoning-aware Multimodal Representation Learning for E-commerce Product Understanding (viability: 8): https://sciencetostartup.com/paper/moon3-0-reasoning-aware-multimodal-representation-learning-for-e-commerce-product-understanding - MOON3.0 is a reasoning-aware multimodal LLM for e-commerce that explicitly models fine-grained product attributes using a novel fusion, contrastive/reinforcement learning, and residual enhancement approach.
- Adaptive Parallel Monte Carlo Tree Search for Efficient Test-time Compute Scaling (viability: 5): https://sciencetostartup.com/paper/adaptive-parallel-monte-carlo-tree-search-for-efficient-test-time-compute-scaling - Adaptive Parallel Monte Carlo Tree Search with negative early exit and boosting mechanism reduces p99 latency and improves throughput for LLM reasoning without sacrificing accuracy.
- RT-GS: Gaussian Splatting with Reflection and Transmittance Primitives (viability: 4): https://sciencetostartup.com/paper/rt-gs-gaussian-splatting-with-reflection-and-transmittance-primitives - A unified framework for Gaussian Splatting that integrates microfacet material models and ray tracing to jointly model specular reflection and transmittance for realistic novel view synthesis.
- A Decoupled Basis-Vector-Driven Generative Framework for Dynamic Multi-Objective Optimization (viability: 4): https://sciencetostartup.com/paper/a-decoupled-basis-vector-driven-generative-framework-for-dynamic-multi-objective-optimization - A decoupled generative framework for dynamic multi-objective optimization that uses wavelet transform and sparse dictionary learning to track moving Pareto fronts with zero-shot inference.
- RegFormer: Transferable Relational Grounding for Efficient Weakly-Supervised Human-Object Interaction Detection (viability: 7): https://sciencetostartup.com/paper/regformer-transferable-relational-grounding-for-efficient-weakly-supervised-human-object-interaction-detection - RegFormer is a transferable relational grounding module for efficient and accurate weakly-supervised Human-Object Interaction detection that learns spatial cues for instance-level reasoning.
- Towards Initialization-dependent and Non-vacuous Generalization Bounds for Overparameterized Shallow Neural Networks (viability: 0): https://sciencetostartup.com/paper/towards-initialization-dependent-and-non-vacuous-generalization-bounds-for-overparameterized-shallow-neural-networks - Develops fully initialization-dependent complexity bounds for shallow neural networks with general Lipschitz activation functions, offering non-vacuous generalization bounds for overparameterized models.
- PET-DINO: Unifying Visual Cues into Grounding DINO with Prompt-Enriched Training (viability: 7): https://sciencetostartup.com/paper/pet-dino-unifying-visual-cues-into-grounding-dino-with-prompt-enriched-training - A universal object detector that unifies visual and text prompts for improved open-set detection and reduced development cycles.
- Scheduling LLM Inference with Uncertainty-Aware Output Length Predictions (viability: 4): https://sciencetostartup.com/paper/scheduling-llm-inference-with-uncertainty-aware-output-length-predictions - An uncertainty-aware scheduling metric for LLM inference that significantly reduces latency and improves throughput by accounting for output length variability.
- PC-SAM: Patch-Constrained Fine-Grained Interactive Road Segmentation in High-Resolution Remote Sensing Images (viability: 8): https://sciencetostartup.com/paper/pc-sam-patch-constrained-fine-grained-interactive-road-segmentation-in-high-resolution-remote-sensing-images - A unified framework for fine-grained interactive road segmentation in high-resolution remote sensing images, combining automatic and interactive methods.
- ARGS: Auto-Regressive Gaussian Splatting via Parallel Progressive Next-Scale Prediction (viability: 7): https://sciencetostartup.com/paper/args-auto-regressive-gaussian-splatting-via-parallel-progressive-next-scale-prediction - Auto-Regressive Gaussian Splatting (ARGS) enables parallel, multi-scale 3D object generation with controllable detail and visual fidelity.
- A Reasoning-Enabled Vision-Language Foundation Model for Chest X-ray Interpretation (viability: 7): https://sciencetostartup.com/paper/a-reasoning-enabled-vision-language-foundation-model-for-chest-x-ray-interpretation - A vision-language model that generates diagnostic predictions and clinically grounded reasoning for chest X-ray interpretation, improving efficiency and accuracy.
- Executing as You Generate: Hiding Execution Latency in LLM Code Generation (viability: 7): https://sciencetostartup.com/paper/executing-as-you-generate-hiding-execution-latency-in-llm-code-generation - An LLM code generation system that executes code concurrently with generation, reducing end-to-end latency by up to 55%.
- Adapting Text LLMs to Speech via Multimodal Depth Up-Scaling (viability: 5): https://sciencetostartup.com/paper/adapting-text-llms-to-speech-via-multimodal-depth-up-scaling - A method to adapt text LLMs to speech by inserting and training new transformer layers, minimizing text capability degradation.
- The Rashomon Effect for Visualizing High-Dimensional Data (viability: 3): https://sciencetostartup.com/paper/the-rashomon-effect-for-visualizing-high-dimensional-data - A framework for visualizing high-dimensional data by embracing the multiplicity of embeddings, leading to more interpretable and robust representations.
- Tucker Diffusion Model for High-dimensional Tensor Generation (viability: 4): https://sciencetostartup.com/paper/tucker-diffusion-model-for-high-dimensional-tensor-generation - A novel Tucker diffusion model for generating structured high-dimensional tensor data with theoretical advantages over vectorized approaches.
- All Roads Lead to Rome: Incentivizing Divergent Thinking in Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/all-roads-lead-to-rome-incentivizing-divergent-thinking-in-vision-language-models - Multi-Group Policy Optimization (MUPO) incentivizes divergent thinking in Vision-Language Models to overcome diversity collapse and improve reasoning capabilities.
- The Silicon Mirror: Dynamic Behavioral Gating for Anti-Sycophancy in LLM Agents (viability: 8): https://sciencetostartup.com/paper/the-silicon-mirror-dynamic-behavioral-gating-for-anti-sycophancy-in-llm-agents - The Silicon Mirror is an orchestration framework that dynamically detects user persuasion tactics and adjusts AI behavior to maintain factual integrity in LLM agents.
- Logarithmic Scores, Power-Law Discoveries: Disentangling Measurement from Coverage in Agent-Based Evaluation (viability: 3): https://sciencetostartup.com/paper/logarithmic-scores-power-law-discoveries-disentangling-measurement-from-coverage-in-agent-based-evaluation - Investigating the relationship between score saturation and issue discovery in LLM-based agent evaluations to understand the optimal number of agents needed.
- Phase space integrity in neural network models of Hamiltonian dynamics: A Lagrangian descriptor approach (viability: 3): https://sciencetostartup.com/paper/phase-space-integrity-in-neural-network-models-of-hamiltonian-dynamics-a-lagrangian-descriptor-approach - A new diagnostic framework using Lagrangian Descriptors to evaluate the phase space integrity of neural network models for Hamiltonian dynamics.
- Automated Detection of Multiple Sclerosis Lesions on 7-tesla MRI Using U-net and Transformer-based Segmentation (viability: 7): https://sciencetostartup.com/paper/automated-detection-of-multiple-sclerosis-lesions-on-7-tesla-mri-using-u-net-and-transformer-based-segmentation - Transformer-based models trained on 7T MRI data for automated detection of multiple sclerosis lesions, outperforming classical methods and releasing code for research.
- Not My Truce: Personality Differences in AI-Mediated Workplace Negotiation (viability: 3): https://sciencetostartup.com/paper/not-my-truce-personality-differences-in-ai-mediated-workplace-negotiation - Investigating how personality traits moderate the effectiveness of AI-mediated workplace negotiation coaching, suggesting tailored interventions for different user profiles.
- First Logit Boosting: Visual Grounding Method to Mitigate Object Hallucination in Large Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/first-logit-boosting-visual-grounding-method-to-mitigate-object-hallucination-in-large-vision-language-models - A training-free method called First Logit Boosting to mitigate object hallucination in Large Vision-Language Models with negligible inference overhead.
- Out of Sight, Out of Track: Adversarial Attacks on Propagation-based Multi-Object Trackers via Query State Manipulation (viability: 7): https://sciencetostartup.com/paper/out-of-sight-out-of-track-adversarial-attacks-on-propagation-based-multi-object-trackers-via-query-state-manipulation - A novel attack framework for multi-object tracking systems that exploits query propagation vulnerabilities to cause track terminations and identity switches.
- Convergence of Byzantine-Resilient Gradient Tracking via Probabilistic Edge Dropout (viability: 3): https://sciencetostartup.com/paper/convergence-of-byzantine-resilient-gradient-tracking-via-probabilistic-edge-dropout - A novel stochastic gradient tracking method designed to maintain convergence properties in distributed optimization networks with Byzantine agents.
- Towards Reliable Truth-Aligned Uncertainty Estimation in Large Language Models (viability: 8): https://sciencetostartup.com/paper/towards-reliable-truth-aligned-uncertainty-estimation-in-large-language-models - A post-hoc calibration method for large language models that improves the reliability of uncertainty estimation by aligning scores with factual correctness.
- Polysemanticity or Polysemy? Lexical Identity Confounds Superposition Metrics (viability: 5): https://sciencetostartup.com/paper/polysemanticity-or-polysemy-lexical-identity-confounds-superposition-metrics - A method to disentangle lexical confounds from semantic superposition in large language models, improving word sense disambiguation and knowledge editing.
- Execution-Verified Reinforcement Learning for Optimization Modeling (viability: 7): https://sciencetostartup.com/paper/execution-verified-reinforcement-learning-for-optimization-modeling - An AI framework that uses a mathematical solver as a verifier to automate optimization modeling from natural language, reducing the need for costly process supervision and enabling cross-solver generalization.
- Reachability-Aware Time Scaling for Path Tracking (viability: 1): https://sciencetostartup.com/paper/reachability-aware-time-scaling-for-path-tracking - A theoretical approach to scale speeds along robot paths to ensure collision-free waypoint tracking under acceleration limits.
- TR-ICRL: Test-Time Rethinking for In-Context Reinforcement Learning (viability: 8): https://sciencetostartup.com/paper/tr-icrl-test-time-rethinking-for-in-context-reinforcement-learning - A framework for LLMs to learn from external rewards during inference by using retrieved instances and pseudo-labels for iterative refinement, significantly improving performance on reasoning and knowledge-intensive tasks.
- Internal State-Based Policy Gradient Methods for Partially Observable Markov Potential Games (viability: 4): https://sciencetostartup.com/paper/internal-state-based-policy-gradient-methods-for-partially-observable-markov-potential-games - A theoretical framework for multi-agent reinforcement learning in partially observable games using internal states and a policy gradient method, with a proven convergence bound.
- Denoising distances beyond the volumetric barrier (viability: 2): https://sciencetostartup.com/paper/denoising-distances-beyond-the-volumetric-barrier - A novel approach for reconstructing the latent geometry of Riemannian manifolds from noisy distance data, breaking the volumetric barrier in higher dimensions.
- Secure Forgetting: A Framework for Privacy-Driven Unlearning in Large Language Model (LLM)-Based Agents (viability: 6): https://sciencetostartup.com/paper/secure-forgetting-a-framework-for-privacy-driven-unlearning-in-large-language-model-llm-based-agents - A framework for LLM-based agents to selectively forget sensitive or outdated knowledge, enabling controlled unlearning with a natural language interface.
- Certificate-Driven Closed-Loop Multi-Agent Path Finding with Inheritable Factorization (viability: 3): https://sciencetostartup.com/paper/certificate-driven-closed-loop-multi-agent-path-finding-with-inheritable-factorization - A certificate-driven approach for closed-loop multi-agent path finding that improves scalability and global guarantees by incorporating inheritable factorization.
- Shapley-Guided Neural Repair Approach via Derivative-Free Optimization (viability: 7): https://sciencetostartup.com/paper/shapley-guided-neural-repair-approach-via-derivative-free-optimization - A Shapley-guided neural repair approach that uses derivative-free optimization to localize and fix defects like backdoors, adversarial attacks, and unfairness in DNNs.
- Self-Routing: Parameter-Free Expert Routing from Hidden States (viability: 3): https://sciencetostartup.com/paper/self-routing-parameter-free-expert-routing-from-hidden-states - A parameter-free routing mechanism for Mixture-of-Experts models that eliminates the need for a learned router.
- G-Drift MIA: Membership Inference via Gradient-Induced Feature Drift in LLMs (viability: 6): https://sciencetostartup.com/paper/g-drift-mia-membership-inference-via-gradient-induced-feature-drift-in-llms - A white-box membership inference attack for LLMs that uses gradient-induced feature drift to detect training data.
- Learning Humanoid Navigation from Human Data (viability: 8): https://sciencetostartup.com/paper/learning-humanoid-navigation-from-human-data - EgoNav enables humanoid robots to autonomously navigate diverse environments using human walking data, bypassing traditional robot-specific data collection.
- Decision-Centric Design for LLM Systems (viability: 3): https://sciencetostartup.com/paper/decision-centric-design-for-llm-systems - A decision-centric framework for LLM systems that separates control decisions from generation for improved reliability and diagnosability.
- Efficient DPF-based Error-Detecting Information-Theoretic Private Information Retrieval Over Rings (viability: 2): https://sciencetostartup.com/paper/efficient-dpf-based-error-detecting-information-theoretic-private-information-retrieval-over-rings - A novel information-theoretic scheme for error-detecting private information retrieval over rings that reduces key size and communication overhead.
- The 1st Winner for 5th PVUW MeViS-Text Challenge: Strong MLLMs Meet SAM3 for Referring Video Object Segmentation (viability: 8): https://sciencetostartup.com/paper/the-1st-winner-for-5th-pvuw-mevis-text-challenge-strong-mllms-meet-sam3-for-referring-video-object-segmentation - A training-free pipeline combining Gemini and SAM3 for referring video object segmentation, achieving state-of-the-art results.
- COTTA: Context-Aware Transfer Adaptation for Trajectory Prediction in Autonomous Driving (viability: 7): https://sciencetostartup.com/paper/cotta-context-aware-transfer-adaptation-for-trajectory-prediction-in-autonomous-driving - Context-aware transfer adaptation for trajectory prediction in autonomous driving, significantly reducing prediction error in new geographic domains.
- Sampling-based Task and Kinodynamic Motion Planning under Semantic Uncertainty (viability: 2): https://sciencetostartup.com/paper/sampling-based-task-and-kinodynamic-motion-planning-under-semantic-uncertainty - An anytime algorithm for integrated task and kinodynamic motion planning under semantic uncertainty in partially observable environments.
- A Cross-graph Tuning-free GNN Prompting Framework (viability: 3): https://sciencetostartup.com/paper/a-cross-graph-tuning-free-gnn-prompting-framework - A tuning-free GNN prompting framework for cross-graph adaptation without retraining.
- Improving Generalization of Deep Learning for Brain Metastases Segmentation Across Institutions (viability: 6): https://sciencetostartup.com/paper/improving-generalization-of-deep-learning-for-brain-metastases-segmentation-across-institutions - A domain adaptation framework for brain metastases segmentation across institutions using VAE-MMD and nnU-Net.
- VLM-in-the-Loop: A Plug-In Quality Assurance Module for ECG Digitization Pipelines (viability: 7): https://sciencetostartup.com/paper/vlm-in-the-loop-a-plug-in-quality-assurance-module-for-ecg-digitization-pipelines - A plug-in quality assurance module for ECG digitization that uses VLMs and domain-specific tools for improved accuracy.
- Advancing Complex Video Object Segmentation via Tracking-Enhanced Prompt: The 1st Winner for 5th PVUW MOSE Challenge (viability: 4): https://sciencetostartup.com/paper/advancing-complex-video-object-segmentation-via-tracking-enhanced-prompt-the-1st-winner-for-5th-pvuw-mose-challenge - A training-free approach using tracking-enhanced prompts for complex video object segmentation.
- Deep Networks Favor Simple Data (viability: 3): https://sciencetostartup.com/paper/deep-networks-favor-simple-data - This research analyzes how deep networks favor simpler data, revealing a consistent pattern across various models and datasets, but lacks a clear product path.
- EvolveTool-Bench: Evaluating the Quality of LLM-Generated Tool Libraries as Software Artifacts (viability: 7): https://sciencetostartup.com/paper/evolvetool-bench-evaluating-the-quality-of-llm-generated-tool-libraries-as-software-artifacts - EvolveTool-Bench evaluates LLM-generated tool libraries as software artifacts, revealing quality risks invisible to task-only evaluation and enabling better tool development.
- Behavioral Score Diffusion: Model-Free Trajectory Planning via Kernel-Based Score Estimation from Data (viability: 6): https://sciencetostartup.com/paper/behavioral-score-diffusion-model-free-trajectory-planning-via-kernel-based-score-estimation-from-data - Behavioral Score Diffusion offers a model-free, training-free trajectory planner for robotics that uses kernel-based score estimation from data, outperforming retrieval methods.
- Gradient-Based Data Valuation Improves Curriculum Learning for Game-Theoretic Motion Planning (viability: 6): https://sciencetostartup.com/paper/gradient-based-data-valuation-improves-curriculum-learning-for-game-theoretic-motion-planning - Gradient-based data valuation significantly improves curriculum learning for game-theoretic motion planners by identifying crucial training scenarios.
- RAGShield: Provenance-Verified Defense-in-Depth Against Knowledge Base Poisoning in Government Retrieval-Augmented Generation Systems (viability: 7): https://sciencetostartup.com/paper/ragshield-provenance-verified-defense-in-depth-against-knowledge-base-poisoning-in-government-retrieval-augmented-genera - RAGShield provides a five-layer defense-in-depth framework for government RAG systems, using supply chain provenance verification to prevent knowledge base poisoning attacks.
- GUIDE: Reinforcement Learning for Behavioral Action Support in Type 1 Diabetes (viability: 7): https://sciencetostartup.com/paper/guide-reinforcement-learning-for-behavioral-action-support-in-type-1-diabetes - GUIDE is a reinforcement learning framework that provides behavioral recommendations to complement automated insulin delivery systems for Type 1 Diabetes management.
- Mine-JEPA: In-Domain Self-Supervised Learning for Mine-Like Object Classification in Side-Scan Sonar (viability: 4): https://sciencetostartup.com/paper/mine-jepa-in-domain-self-supervised-learning-for-mine-like-object-classification-in-side-scan-sonar - Mine-JEPA is an in-domain self-supervised learning pipeline for side-scan sonar mine classification, outperforming larger foundation models in data-scarce maritime imagery.
- mmAnomaly: Leveraging Visual Context for Robust Anomaly Detection in the Non-Visual World with mmWave Radar (viability: 7): https://sciencetostartup.com/paper/mmanomaly-leveraging-visual-context-for-robust-anomaly-detection-in-the-non-visual-world-with-mmwave-radar - mmAnomaly is a multi-modal framework combining mmWave radar and RGBD input for robust anomaly detection in non-visual scenarios, achieving up to 94% F1 score.
- UCMNet: Uncertainty-Aware Context Memory Network for Under-Display Camera Image Restoration (viability: 7): https://sciencetostartup.com/paper/ucmnet-uncertainty-aware-context-memory-network-for-under-display-camera-image-restoration - A lightweight AI model that restores image quality for under-display cameras by adaptively processing regions based on estimated uncertainty.
- Locally Confident, Globally Stuck: The Quality-Exploration Dilemma in Diffusion Language Models (viability: 6): https://sciencetostartup.com/paper/locally-confident-globally-stuck-the-quality-exploration-dilemma-in-diffusion-language-models - An AI decoding method for diffusion language models that balances generation quality with exploration for improved reasoning.
- Dynamic Graph Neural Network with Adaptive Features Selection for RGB-D Based Indoor Scene Recognition (viability: 7): https://sciencetostartup.com/paper/dynamic-graph-neural-network-with-adaptive-features-selection-for-rgb-d-based-indoor-scene-recognition - An AI model that uses dynamic graphs and adaptive feature selection from RGB-D data for more accurate indoor scene recognition.
- Neural Reconstruction of LiDAR Point Clouds under Jamming Attacks via Full-Waveform Representation and Simultaneous Laser Sensing (viability: 7): https://sciencetostartup.com/paper/neural-reconstruction-of-lidar-point-clouds-under-jamming-attacks-via-full-waveform-representation-and-simultaneous-lase - An AI system that reconstructs authentic LiDAR point clouds under jamming attacks by analyzing full-waveform data.
- A Dual-Stream Transformer Architecture for Illumination-Invariant TIR-LiDAR Person Tracking (viability: 7): https://sciencetostartup.com/paper/a-dual-stream-transformer-architecture-for-illumination-invariant-tir-lidar-person-tracking - A dual-stream Transformer architecture for illumination-invariant TIR-LiDAR person tracking in autonomous robots.
- In harmony with gpt-oss (viability: 6): https://sciencetostartup.com/paper/in-harmony-with-gpt-oss - Build robust coding tool harnesses utilizing GPT-OSS for improved integrated AI development environments.
- VADMamba++: Efficient Video Anomaly Detection via Hybrid Modeling in Grayscale Space (viability: 7): https://sciencetostartup.com/paper/vadmamba-efficient-video-anomaly-detection-via-hybrid-modeling-in-grayscale-space - An efficient video anomaly detection method using Mamba, CNN, and Transformer modules for grayscale to RGB reconstruction.
- Signals: Trajectory Sampling and Triage for Agentic Interactions (viability: 8): https://sciencetostartup.com/paper/signals-trajectory-sampling-and-triage-for-agentic-interactions - A lightweight, signal-based framework for triaging agentic interaction trajectories to improve post-deployment optimization.
- Deep Learning-Accelerated Surrogate Optimization for High-Dimensional Well Control in Stress-Sensitive Reservoirs (viability: 4): https://sciencetostartup.com/paper/deep-learning-accelerated-surrogate-optimization-for-high-dimensional-well-control-in-stress-sensitive-reservoirs - A deep learning framework accelerates high-dimensional well control optimization in stress-sensitive reservoirs by using a surrogate model to reduce computational cost by orders of magnitude.
- Go Big or Go Home: Simulating Mobbing Behavior with Braitenbergian Robots (viability: 0): https://sciencetostartup.com/paper/go-big-or-go-home-simulating-mobbing-behavior-with-braitenbergian-robots - Simulating mobbing behavior in Braitenbergian robots using the Webots platform to explore the effects of mobbing call range and group size on predator harassment.
- Agent Q-Mix: Selecting the Right Action for LLM Multi-Agent Systems through Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/agent-q-mix-selecting-the-right-action-for-llm-multi-agent-systems-through-reinforcement-learning - Agent Q-Mix is a reinforcement learning framework that optimizes LLM agent communication topology for improved accuracy and token efficiency in complex tasks.
- Real Time Local Wind Inference for Robust Autonomous Navigation (viability: 7): https://sciencetostartup.com/paper/real-time-local-wind-inference-for-robust-autonomous-navigation - Real-time local wind inference using LiDAR and deep learning enables robust autonomous navigation for aerial robots, improving energy efficiency and obstacle avoidance.
- Is One Token All It Takes? Graph Pooling Tokens for LLM-based GraphQA (viability: 7): https://sciencetostartup.com/paper/is-one-token-all-it-takes-graph-pooling-tokens-for-llm-based-graphqa - This research enhances Graph Question Answering by improving how graph structures are encoded into LLMs, achieving competitive results with a novel multi-token pooling approach and LoRA stabilization.
- When Career Data Runs Out: Structured Feature Engineering and Signal Limits for Founder Success Prediction (viability: 5): https://sciencetostartup.com/paper/when-career-data-runs-out-structured-feature-engineering-and-signal-limits-for-founder-success-prediction - This work engineers structured features from founder career data to predict startup success, outperforming LLM baselines and diagnosing dataset limitations for future improvements.
- The Persistent Vulnerability of Aligned AI Systems (viability: 3): https://sciencetostartup.com/paper/the-persistent-vulnerability-of-aligned-ai-systems - This thesis explores AI safety by automating circuit discovery, removing dangerous behaviors with Latent Adversarial Training, and analyzing agentic misalignment in frontier models.
- Large Language Models in the Abuse Detection Pipeline (viability: 3): https://sciencetostartup.com/paper/large-language-models-in-the-abuse-detection-pipeline - This survey analyzes the integration of Large Language Models into the Abuse Detection Lifecycle, covering label generation, detection, review, and governance, while highlighting challenges and future directions.
- Hierarchical Motion Planning and Control under Unknown Nonlinear Dynamics via Predicted Reachability (viability: 3): https://sciencetostartup.com/paper/hierarchical-motion-planning-and-control-under-unknown-nonlinear-dynamics-via-predicted-reachability - A hierarchical framework for autonomous motion planning and control under unknown nonlinear dynamics using piecewise-affine models and graph-based reachability.
- Collaborative AI Agents and Critics for Fault Detection and Cause Analysis in Network Telemetry (viability: 4): https://sciencetostartup.com/paper/collaborative-ai-agents-and-critics-for-fault-detection-and-cause-analysis-in-network-telemetry - Collaborative AI agents and critics for fault detection and cause analysis in network telemetry, leveraging private cost functions and foundation models.
- Breaking Data Symmetry is Needed For Generalization in Feature Learning Kernels (viability: 0): https://sciencetostartup.com/paper/breaking-data-symmetry-is-needed-for-generalization-in-feature-learning-kernels - Investigating the phenomenon of grokking in feature learning kernels, showing that breaking data symmetry is crucial for generalization.
- Label-efficient underwater species classification with semi-supervised learning on frozen foundation model embeddings (viability: 7): https://sciencetostartup.com/paper/label-efficient-underwater-species-classification-with-semi-supervised-learning-on-frozen-foundation-model-embeddings - Label-efficient underwater species classification using semi-supervised learning on frozen foundation model embeddings, requiring no model training.
- Robust Multimodal Safety via Conditional Decoding (viability: 7): https://sciencetostartup.com/paper/robust-multimodal-safety-via-conditional-decoding - A conditional decoding strategy that significantly improves multimodal LLM safety by predicting a binary safety token before response generation, without external classifiers or modality-specific fine-tuning.
- Vocal Prognostic Digital Biomarkers in Monitoring Chronic Heart Failure: A Longitudinal Observational Study (viability: 4): https://sciencetostartup.com/paper/vocal-prognostic-digital-biomarkers-in-monitoring-chronic-heart-failure-a-longitudinal-observational-study - Voice features can predict health deterioration in chronic heart failure patients, offering a non-invasive alternative to current monitoring methods.
- SAGE: Subsurface AI-driven Geostatistical Extraction with proxy posterior (viability: 7): https://sciencetostartup.com/paper/sage-subsurface-ai-driven-geostatistical-extraction-with-proxy-posterior - A framework for generating statistically consistent subsurface velocity models from sparse well logs and seismic images, enabling data-efficient seismic imaging and inversion.
- Asymmetric Actor-Critic for Multi-turn LLM Agents (viability: 6): https://sciencetostartup.com/paper/asymmetric-actor-critic-for-multi-turn-llm-agents - An asymmetric actor-critic framework that uses a powerful proprietary LLM as an actor and a smaller open-source critic for runtime supervision to improve reliability in multi-turn agent interactions.
- Cybersecurity Risk Assessment for CubeSat Missions: Adapting Established Frameworks for Resource-Constrained Environments (viability: 4): https://sciencetostartup.com/paper/cybersecurity-risk-assessment-for-cubesat-missions-adapting-established-frameworks-for-resource-constrained-environments - A cybersecurity risk assessment framework adapted for resource-constrained CubeSat missions, offering proportionate guidance for mission designers and regulators.
- Play-Testing REMind: Evaluating an Educational Robot-Mediated Role-Play Game (viability: 0): https://sciencetostartup.com/paper/play-testing-remind-evaluating-an-educational-robot-mediated-role-play-game - REMind is an educational robot-mediated role-play game designed to support anti-bullying bystander intervention among children.
- SANA I2I: A Text Free Flow Matching Framework for Paired Image to Image Translation with a Case Study in Fetal MRI Artifact Reduction (viability: 7): https://sciencetostartup.com/paper/sana-i2i-a-text-free-flow-matching-framework-for-paired-image-to-image-translation-with-a-case-study-in-fetal-mri-artifa - A text-free image-to-image translation framework for medical imaging, effectively reducing artifacts with competitive performance.
- SYNTHONY: A Stress-Aware, Intent-Conditioned Agent for Deep Tabular Generative Models Selection (viability: 7): https://sciencetostartup.com/paper/synthony-a-stress-aware-intent-conditioned-agent-for-deep-tabular-generative-models-selection - SYNTHONY selects optimal deep tabular generative models based on dataset stress profiles and user intent, outperforming LLM selectors.
- MambaVoiceCloning: Efficient and Expressive Text-to-Speech via State-Space Modeling and Diffusion Control (viability: 4): https://sciencetostartup.com/paper/mambavoicecloning-efficient-and-expressive-text-to-speech-via-state-space-modeling-and-diffusion-control - An efficient and expressive text-to-speech system using only state-space models for conditioning, reducing parameters and improving throughput.
- Frege in the Flesh: Biolinguistics and the Neural Enforcement of Syntactic Structures (viability: 0): https://sciencetostartup.com/paper/frege-in-the-flesh-biolinguistics-and-the-neural-enforcement-of-syntactic-structures - Exploring the biological foundations of human language through mathematical and algebraic models of syntactic structures.
- Improvisational Games as a Benchmark for Social Intelligence of AI Agents: The Case of Connections (viability: 5): https://sciencetostartup.com/paper/improvisational-games-as-a-benchmark-for-social-intelligence-of-ai-agents-the-case-of-connections - A new wordplay game benchmark to evaluate the social intelligence and reasoning capabilities of AI agents.
- Human-in-the-Loop Control of Objective Drift in LLM-Assisted Computer Science Education (viability: 4): https://sciencetostartup.com/paper/human-in-the-loop-control-of-objective-drift-in-llm-assisted-computer-science-education - A human-in-the-loop curriculum for computer science education that teaches students to control objective drift in LLM-assisted workflows.
- VeriAct: Beyond Verifiability -- Agentic Synthesis of Correct and Complete Formal Specifications (viability: 7): https://sciencetostartup.com/paper/veriact-beyond-verifiability-agentic-synthesis-of-correct-and-complete-formal-specifications - An agentic framework that iteratively synthesizes and repairs formal software specifications using LLMs and verification feedback to ensure correctness and completeness.
- The Geometry of Compromise: Unlocking Generative Capabilities via Controllable Modality Alignment (viability: 7): https://sciencetostartup.com/paper/the-geometry-of-compromise-unlocking-generative-capabilities-via-controllable-modality-alignment - A fine-tuning framework that explicitly reduces the modality gap in Vision-Language Models by aligning geometric and distributional structures for improved cross-modal tasks.
- Excite, Attend and Segment (EASe): Domain-Agnostic Fine-Grained Mask Discovery with Feature Calibration and Self-Supervised Upsampling (viability: 7): https://sciencetostartup.com/paper/excite-attend-and-segment-ease-domain-agnostic-fine-grained-mask-discovery-with-feature-calibration-and-self-supervised - An unsupervised, domain-agnostic framework for fine-grained semantic segmentation that discovers masks in complex scenes using feature calibration and self-supervised upsampling.
- OmniSch: A Multimodal PCB Schematic Benchmark For Structured Diagram Visual Reasoning (viability: 4): https://sciencetostartup.com/paper/omnisch-a-multimodal-pcb-schematic-benchmark-for-structured-diagram-visual-reasoning - A benchmark for evaluating large multimodal models on the complex task of converting PCB schematic diagrams into structured, spatially weighted netlist graphs.
- Omni-MMSI: Toward Identity-attributed Social Interaction Understanding (viability: 7): https://sciencetostartup.com/paper/omni-mmsi-toward-identity-attributed-social-interaction-understanding - A new task and pipeline for understanding social interactions from raw audio and vision, with a focus on identity attribution and reasoning.
- Benchmarking Interaction, Beyond Policy: a Reproducible Benchmark for Collaborative Instance Object Navigation (viability: 8): https://sciencetostartup.com/paper/benchmarking-interaction-beyond-policy-a-reproducible-benchmark-for-collaborative-instance-object-navigation - A benchmark and lightweight model for collaborative object navigation that separates navigation and question-asking assessment, outperforming existing methods.
- Autonomous Adaptive Solver Selection for Chemistry Integration via Reinforcement Learning (viability: 4): https://sciencetostartup.com/paper/autonomous-adaptive-solver-selection-for-chemistry-integration-via-reinforcement-learning - A reinforcement learning framework that autonomously selects solvers for chemistry integration to reduce computational cost while maintaining accuracy.
- Can Large Language Models Self-Correct in Medical Question Answering? An Exploratory Study (viability: 3): https://sciencetostartup.com/paper/can-large-language-models-self-correct-in-medical-question-answering-an-exploratory-study - An exploratory study on whether large language models can self-correct in medical question answering, finding inconsistent benefits.
- Learning to Shuffle: Block Reshuffling and Reversal Schemes for Stochastic Optimization (viability: 4): https://sciencetostartup.com/paper/learning-to-shuffle-block-reshuffling-and-reversal-schemes-for-stochastic-optimization - Discovering novel data shuffling strategies for stochastic optimization using LLM-guided program evolution to improve convergence.
- LLM Essay Scoring Under Holistic and Analytic Rubrics: Prompt Effects and Bias (viability: 7): https://sciencetostartup.com/paper/llm-essay-scoring-under-holistic-and-analytic-rubrics-prompt-effects-and-bias - An LLM-based essay scoring system that identifies and corrects systematic biases in grading, improving accuracy with minimal data.
- Hierarchical Apprenticeship Learning from Imperfect Demonstrations with Evolving Rewards (viability: 4): https://sciencetostartup.com/paper/hierarchical-apprenticeship-learning-from-imperfect-demonstrations-with-evolving-rewards - HALIDE: A hierarchical apprenticeship learning framework that leverages imperfect student demonstrations to infer evolving rewards and improve pedagogical decisions.
- Informed Machine Learning with Knowledge Landmarks (viability: 4): https://sciencetostartup.com/paper/informed-machine-learning-with-knowledge-landmarks - KD-ML integrates numeric data with granular knowledge landmarks to build more accurate models, outperforming data-driven approaches on physics-governed benchmarks.
- PRISM: Differentiable Analysis-by-Synthesis for Fixel Recovery in Diffusion MRI (viability: 3): https://sciencetostartup.com/paper/prism-differentiable-analysis-by-synthesis-for-fixel-recovery-in-diffusion-mri - A differentiable framework for improved fiber recovery in diffusion MRI by fitting explicit multi-compartment models over spatial patches.
- A Safety-Aware Role-Orchestrated Multi-Agent LLM Framework for Behavioral Health Communication Simulation (viability: 3): https://sciencetostartup.com/paper/a-safety-aware-role-orchestrated-multi-agent-llm-framework-for-behavioral-health-communication-simulation - A multi-agent LLM framework simulating supportive behavioral health dialogue through coordinated, role-differentiated agents with continuous safety auditing.
- REM-CTX: Automated Peer Review via Reinforcement Learning with Auxiliary Context (viability: 4): https://sciencetostartup.com/paper/rem-ctx-automated-peer-review-via-reinforcement-learning-with-auxiliary-context - An automated peer review system that incorporates auxiliary context like figures and external scholarly signals using reinforcement learning.
- UCell: rethinking generalizability and scaling of bio-medical vision models (viability: 7): https://sciencetostartup.com/paper/ucell-rethinking-generalizability-and-scaling-of-bio-medical-vision-models - A parameter-efficient deep learning model for single-cell segmentation that matches larger models' performance and generalizes well to unseen data.
- FGR-ColBERT: Identifying Fine-Grained Relevance Tokens During Retrieval (viability: 7): https://sciencetostartup.com/paper/fgr-colbert-identifying-fine-grained-relevance-tokens-during-retrieval - A modified retrieval model that integrates LLM-derived relevance signals to identify fine-grained evidence cues with minimal latency overhead.
- Softmax gradient policy for variance minimization and risk-averse multi armed bandits (viability: 1): https://sciencetostartup.com/paper/softmax-gradient-policy-for-variance-minimization-and-risk-averse-multi-armed-bandits - A theoretical algorithm for risk-aware multi-armed bandits that minimizes variance in reward selection.
- A Taxonomy of Programming Languages for Code Generation (viability: 7): https://sciencetostartup.com/paper/a-taxonomy-of-programming-languages-for-code-generation - A reproducible taxonomy of programming languages by resource availability to guide dataset curation and tier-aware LLM evaluation for code generation.
- AI-Mediated Explainable Regulation for Justice (viability: 6): https://sciencetostartup.com/paper/ai-mediated-explainable-regulation-for-justice - A distributed AI system for explainable and adaptable regulation that models stakeholder preferences to ensure justice and legitimacy.
- Hierarchical Discrete Flow Matching for Graph Generation (viability: 3): https://sciencetostartup.com/paper/hierarchical-discrete-flow-matching-for-graph-generation - A hierarchical generative framework for graph generation that reduces computational cost and generation time.
- MAC-Attention: a Match-Amend-Complete Scheme for Fast and Accurate Attention Computation (viability: 7): https://sciencetostartup.com/paper/mac-attention-a-match-amend-complete-scheme-for-fast-and-accurate-attention-computation - MAC-Attention accelerates LLM long-context decoding by reusing prior attention computations, reducing latency and KV accesses while maintaining fidelity.
- Neural Collapse Dynamics: Depth, Activation, Regularisation, and Feature Norm Threshold (viability: 3): https://sciencetostartup.com/paper/neural-collapse-dynamics-depth-activation-regularisation-and-feature-norm-threshold - Identifies a critical feature norm threshold that predicts the onset of neural collapse in deep networks.
- Do Language Models Know When They'll Refuse? Probing Introspective Awareness of Safety Boundaries (viability: 5): https://sciencetostartup.com/paper/do-language-models-know-when-they-ll-refuse-probing-introspective-awareness-of-safety-boundaries - Language models can predict their refusal behavior with high accuracy, enabling confidence-based routing for safety-critical applications.
- Diversity-Aware Reverse Kullback-Leibler Divergence for Large Language Model Distillation (viability: 7): https://sciencetostartup.com/paper/diversity-aware-reverse-kullback-leibler-divergence-for-large-language-model-distillation - A novel diversity-aware distillation objective for LLMs that improves performance and the fidelity-diversity trade-off.
- Risk-Aware Batch Testing for Performance Regression Detection (viability: 8): https://sciencetostartup.com/paper/risk-aware-batch-testing-for-performance-regression-detection - Building a CI tool to save over $490K annually in infrastructure costs by optimizing performance regression testing with risk-aware batch strategies.
- Do LLMs Know What Is Private Internally? Probing and Steering Contextual Privacy Norms in Large Language Model Representations (viability: 8): https://sciencetostartup.com/paper/do-llms-know-what-is-private-internally-probing-and-steering-contextual-privacy-norms-in-large-language-model-representa - A method to probe and steer LLMs' internal understanding of contextual privacy norms, enabling more reliable control over private information disclosure.
- Measuring the Representational Alignment of Neural Systems in Superposition (viability: 1): https://sciencetostartup.com/paper/measuring-the-representational-alignment-of-neural-systems-in-superposition - A theoretical framework for understanding neural network representations in superposition, highlighting limitations of current alignment metrics.
- Lead Zirconate Titanate Reservoir Computing for Classification of Written and Spoken Digits (viability: 3): https://sciencetostartup.com/paper/lead-zirconate-titanate-reservoir-computing-for-classification-of-written-and-spoken-digits - Utilizing Lead Zirconate Titanate as a physical reservoir for improved handwritten digit classification.
- Unsupervised 4D Flow MRI Velocity Enhancement and Unwrapping Using Divergence-Free Neural Networks (viability: 7): https://sciencetostartup.com/paper/unsupervised-4d-flow-mri-velocity-enhancement-and-unwrapping-using-divergence-free-neural-networks - An unsupervised neural network for enhancing and unwrapping velocity fields in 4D Flow MRI data.
- DreamControl-v2: Simpler and Scalable Autonomous Humanoid Skills via Trainable Guided Diffusion Priors (viability: 8): https://sciencetostartup.com/paper/dreamcontrol-v2-simpler-and-scalable-autonomous-humanoid-skills-via-trainable-guided-diffusion-priors - Scalable autonomous humanoid skills using trainable guided diffusion models trained on diverse motion data.
- Offline Constrained RLHF with Multiple Preference Oracles (viability: 3): https://sciencetostartup.com/paper/offline-constrained-rlhf-with-multiple-preference-oracles - Offline constrained reinforcement learning from human feedback with multiple preference oracles for performance and safety trade-offs.
- QUEST: A robust attention formulation using query-modulated spherical attention (viability: 7): https://sciencetostartup.com/paper/quest-a-robust-attention-formulation-using-query-modulated-spherical-attention - A novel attention mechanism for Transformers that improves training stability, performance, and robustness to data corruptions and adversarial attacks, with applications in vision and beyond.
- Lévy-Flow Models: Heavy-Tail-Aware Normalizing Flows for Financial Risk Management (viability: 7): https://sciencetostartup.com/paper/l-vy-flow-models-heavy-tail-aware-normalizing-flows-for-financial-risk-management - Introducing Lévy-Flows, a new class of normalizing flow models that capture heavy-tailed financial data for improved risk management and density estimation.
- Making Sense of AI Agents Hype: Adoption, Architectures, and Takeaways from Practitioners (viability: 2): https://sciencetostartup.com/paper/making-sense-of-ai-agents-hype-adoption-architectures-and-takeaways-from-practitioners - A review of practitioner conference talks to understand how companies adopt, architect, and implement LLM-driven AI agentic systems.
- On the Necessity of Pre-agreed Secrets for Thwarting Last-minute Coercion: Vulnerabilities and Lessons From the Loki E-voting Protocol (viability: 0): https://sciencetostartup.com/paper/on-the-necessity-of-pre-agreed-secrets-for-thwarting-last-minute-coercion-vulnerabilities-and-lessons-from-the-loki-e-vo - Identifies vulnerabilities in the Loki e-voting protocol and argues for the necessity of pre-agreed secrets to prevent last-minute coercion.
- Explainable AI for Blind and Low-Vision Users: Navigating Trust, Modality, and Interpretability in the Agentic Era (viability: 2): https://sciencetostartup.com/paper/explainable-ai-for-blind-and-low-vision-users-navigating-trust-modality-and-interpretability-in-the-agentic-era - Developing accessible explainable AI for blind and low-vision users to foster trust and independent use of autonomous AI agents.
- NFC based inventory control system for secure and efficient communication (viability: 5): https://sciencetostartup.com/paper/nfc-based-inventory-control-system-for-secure-and-efficient-communication - A secure and efficient inventory control system using NFC tags to replace vulnerable barcodes for retail applications.
- Sit-to-Stand Transitions Detection and Duration Measurement Using Smart Lacelock Sensor (viability: 6): https://sciencetostartup.com/paper/sit-to-stand-transitions-detection-and-duration-measurement-using-smart-lacelock-sensor - A shoe-mounted sensor system for accurate Sit-to-Stand transition detection and duration measurement to assess fall risk in older adults.
- Polish phonology and morphology through the lens of distributional semantics (viability: 7): https://sciencetostartup.com/paper/polish-phonology-and-morphology-through-the-lens-of-distributional-semantics - Leveraging distributional semantics to uncover relationships between Polish word phonology, morphology, and meaning, with potential for language learning tools.
- Suppressing Non-Semantic Noise in Masked Image Modeling Representations (viability: 3): https://sciencetostartup.com/paper/suppressing-non-semantic-noise-in-masked-image-modeling-representations - A post-hoc method to suppress non-semantic information in image representations for improved zero-shot performance.
- Unified Architecture Metamodel of Information Systems Developed by Generative AI (viability: 7): https://sciencetostartup.com/paper/unified-architecture-metamodel-of-information-systems-developed-by-generative-ai - A unified architecture metamodel for LLM-oriented applications to automate and standardize the software development lifecycle.
- Beyond Latency: A System-Level Characterization of MPC and FHE for PPML (viability: 3): https://sciencetostartup.com/paper/beyond-latency-a-system-level-characterization-of-mpc-and-fhe-for-ppml - A system-level characterization of MPC and FHE for privacy-preserving machine learning, evaluating performance, energy, and cost across various scenarios.
- Neural-Assisted in-Motion Self-Heading Alignment (viability: 7): https://sciencetostartup.com/paper/neural-assisted-in-motion-self-heading-alignment - A neural-assisted framework for rapid and accurate initial heading estimation in autonomous ocean platforms, significantly reducing alignment time and improving accuracy.
- A Study on the Impact of Fault localization Granularity for Repository-Scale Code Repair Tasks (viability: 7): https://sciencetostartup.com/paper/a-study-on-the-impact-of-fault-localization-granularity-for-repository-scale-code-repair-tasks - A framework for investigating how fault localization granularity impacts automatic code repair at the repository scale, offering a proof of concept for optimizing this process.
- Epileptic Seizure Detection in Separate Frequency Bands Using Feature Analysis and Graph Convolutional Neural Network (GCN) from Electroencephalogram (EEG) Signals (viability: 7): https://sciencetostartup.com/paper/epileptic-seizure-detection-in-separate-frequency-bands-using-feature-analysis-and-graph-convolutional-neural-network-gc - A frequency-aware framework for epileptic seizure detection using GCNs on EEG signals, offering improved interpretability and diagnostic precision.
- Long-Horizon Geometry-Aware Navigation among Polytopes via MILP-MPC and Minkowski-Based CBFs (viability: 6): https://sciencetostartup.com/paper/long-horizon-geometry-aware-navigation-among-polytopes-via-milp-mpc-and-minkowski-based-cbfs - A hierarchical planning and control framework for geometry-aware robot navigation in complex environments using MILP-MPC and Minkowski-based CBFs.
- Q-Mask: Query-driven Causal Masks for Text Anchoring in OCR-Oriented Vision-Language Models (viability: 8): https://sciencetostartup.com/paper/q-mask-query-driven-causal-masks-for-text-anchoring-in-ocr-oriented-vision-language-models - Q-Mask is an OCR framework using query-driven causal masks for accurate text anchoring in vision-language models, trained on a large-scale dataset.
- Open, Reliable, and Collective: A Community-Driven Framework for Tool-Using AI Agents (viability: 8): https://sciencetostartup.com/paper/open-reliable-and-collective-a-community-driven-framework-for-tool-using-ai-agents - A community-driven framework for reliable tool-using AI agents with standardized schemas, plug-and-play wrappers, and automated testing.
- ParetoBandit: Budget-Paced Adaptive Routing for Non-Stationary LLM Serving (viability: 7): https://sciencetostartup.com/paper/paretobandit-budget-paced-adaptive-routing-for-non-stationary-llm-serving - An adaptive LLM serving router that enforces budgets, adapts to non-stationary conditions, and allows hot-swapping of models.
- Predicting Wave Reflection and Transmission in Heterogeneous Media via Fourier Operator-Based Transformer Modeling (viability: 1): https://sciencetostartup.com/paper/predicting-wave-reflection-and-transmission-in-heterogeneous-media-via-fourier-operator-based-transformer-modeling - A transformer-based ML model that approximates solutions to Maxwell's equations for wave reflection and transmission.
- Oblivion: Self-Adaptive Agentic Memory Control through Decay-Driven Activation (viability: 7): https://sciencetostartup.com/paper/oblivion-self-adaptive-agentic-memory-control-through-decay-driven-activation - A memory control framework for LLM agents that mimics human selective forgetting to reduce interference and latency.
- Hierarchical Chain-of-Thought Prompting: Enhancing LLM Reasoning Performance and Efficiency (viability: 8): https://sciencetostartup.com/paper/hierarchical-chain-of-thought-prompting-enhancing-llm-reasoning-performance-and-efficiency - Enhance LLM reasoning by structuring prompts into hierarchical plans and execution steps, improving accuracy and efficiency.
- From Domain Understanding to Design Readiness: a playbook for GenAI-supported learning in Software Engineering (viability: 4): https://sciencetostartup.com/paper/from-domain-understanding-to-design-readiness-a-playbook-for-genai-supported-learning-in-software-engineering - A playbook for using GenAI tutors to improve student learning in software engineering by grounding them in course knowledge.
- Efficient Software Vulnerability Detection Using Transformer-based Models (viability: 3): https://sciencetostartup.com/paper/efficient-software-vulnerability-detection-using-transformer-based-models - Utilize transformer models with program slices to improve accuracy in detecting software vulnerabilities.
- Speeding Up Mixed-Integer Programming Solvers with Sparse Learning for Branching (viability: 7): https://sciencetostartup.com/paper/speeding-up-mixed-integer-programming-solvers-with-sparse-learning-for-branching - Accelerate mixed-integer programming solvers by using interpretable, sparse learning models for branching decisions, outperforming GPU-accelerated methods.
- RawGen: Learning Camera Raw Image Generation (viability: 7): https://sciencetostartup.com/paper/rawgen-learning-camera-raw-image-generation - A diffusion-based framework for generating camera raw images from text or sRGB inputs, enabling scalable synthetic data for downstream vision tasks.
- Hierarchical Pre-Training of Vision Encoders with Large Language Models (viability: 7): https://sciencetostartup.com/paper/hierarchical-pre-training-of-vision-encoders-with-large-language-models - A hierarchical pre-training framework that enhances vision-language alignment by enabling structured feature fusion between vision encoders and LLMs.
- One Panel Does Not Fit All: Case-Adaptive Multi-Agent Deliberation for Clinical Prediction (viability: 5): https://sciencetostartup.com/paper/one-panel-does-not-fit-all-case-adaptive-multi-agent-deliberation-for-clinical-prediction - A case-adaptive multi-agent system that dynamically assembles specialist panels for clinical prediction, improving accuracy and transparency.
- Gen-Searcher: Reinforcing Agentic Search for Image Generation (viability: 9): https://sciencetostartup.com/paper/gen-searcher-reinforcing-agentic-search-for-image-generation - Gen-Searcher leverages agentic reinforcement learning for search-augmented image generation, delivering contextually relevant and high-fidelity visual content.
- HandX: Scaling Bimanual Motion and Interaction Generation (viability: 7): https://sciencetostartup.com/paper/handx-scaling-bimanual-motion-and-interaction-generation - A unified foundation for generating realistic bimanual hand motion, including data, annotation, and evaluation, with a focus on fine-grained finger dynamics and inter-hand coordination.
- Adaptive Block-Scaled Data Types (viability: 9): https://sciencetostartup.com/paper/adaptive-block-scaled-data-types - Design and implement low-precision data types that improve the performance and efficiency of large language models on modern hardware.
- Geometry-aware similarity metrics for neural representations on Riemannian and statistical manifolds (viability: 2): https://sciencetostartup.com/paper/geometry-aware-similarity-metrics-for-neural-representations-on-riemannian-and-statistical-manifolds - A novel mathematical framework for analyzing the intrinsic geometry of neural network representations to understand computational mechanisms.
- PoseDreamer: Scalable and Photorealistic Human Data Generation Pipeline with Diffusion Models (viability: 7): https://sciencetostartup.com/paper/posedreamer-scalable-and-photorealistic-human-data-generation-pipeline-with-diffusion-models - PoseDreamer generates highly photorealistic synthetic datasets for human 3D mesh estimation using diffusion models, offering a cost-effective alternative to traditional synthetic dataset methods.
- On-the-fly Repulsion in the Contextual Space for Rich Diversity in Diffusion Transformers (viability: 7): https://sciencetostartup.com/paper/on-the-fly-repulsion-in-the-contextual-space-for-rich-diversity-in-diffusion-transformers - A novel framework for Diffusion Transformers that injects on-the-fly repulsion in the contextual space to achieve rich image diversity without sacrificing fidelity or semantic adherence.
- SHOW3D: Capturing Scenes of 3D Hands and Objects in the Wild (viability: 7): https://sciencetostartup.com/paper/show3d-capturing-scenes-of-3d-hands-and-objects-in-the-wild - A new dataset and system for capturing 3D hand-object interactions in real-world environments, enabling more robust egocentric computer vision.
- FlowIt: Global Matching for Optical Flow with Confidence-Guided Refinement (viability: 7): https://sciencetostartup.com/paper/flowit-global-matching-for-optical-flow-with-confidence-guided-refinement - A novel optical flow estimation architecture that uses transformers and optimal transport for robust global matching and confidence-guided refinement, achieving state-of-the-art results.
- SonoWorld: From One Image to a 3D Audio-Visual Scene (viability: 7): https://sciencetostartup.com/paper/sonoworld-from-one-image-to-a-3d-audio-visual-scene - Generate immersive 3D audio-visual scenes from a single image, enabling applications in spatial audio rendering and acoustic learning.
- Temporal Credit Is Free (viability: 4): https://sciencetostartup.com/paper/temporal-credit-is-free - A novel recurrent network architecture that drastically reduces memory requirements for online adaptation, matching state-of-the-art performance.
- Stop Probing, Start Coding: Why Linear Probes and Sparse Autoencoders Fail at Compositional Generalisation (viability: 3): https://sciencetostartup.com/paper/stop-probing-start-coding-why-linear-probes-and-sparse-autoencoders-fail-at-compositional-generalisation - This research investigates why sparse autoencoders fail at compositional generalization, identifying dictionary learning as the core challenge for scalable sparse inference.
- Rethinking Language Model Scaling under Transferable Hypersphere Optimization (viability: 7): https://sciencetostartup.com/paper/rethinking-language-model-scaling-under-transferable-hypersphere-optimization - A new hypersphere parameterization framework that enables stable and efficient scaling of large language models with transferable learning rates.
- FocusVLA: Focused Visual Utilization for Vision-Language-Action Models (viability: 7): https://sciencetostartup.com/paper/focusvla-focused-visual-utilization-for-vision-language-action-models - FocusVLA enhances robotic action generation by intelligently focusing on task-relevant visual information, improving dexterity and accelerating learning.
- Expectation Error Bounds for Transfer Learning in Linear Regression and Linear Neural Networks (viability: 2): https://sciencetostartup.com/paper/expectation-error-bounds-for-transfer-learning-in-linear-regression-and-linear-neural-networks - This paper provides theoretical insights into transfer learning in linear models, offering conditions for auxiliary data to improve generalization.
- RAD-AI: Rethinking Architecture Documentation for AI-Augmented Ecosystems (viability: 7): https://sciencetostartup.com/paper/rad-ai-rethinking-architecture-documentation-for-ai-augmented-ecosystems - RAD-AI provides a novel documentation framework for AI-augmented ecosystems, addressing critical regulatory gaps in the EU AI Act and improving compliance for high-risk systems.
- See it to Place it: Evolving Macro Placements with Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/see-it-to-place-it-evolving-macro-placements-with-vision-language-models - Leverage VLMs to automate and enhance chip floorplanning through guided placement strategies.
- Pandora: Articulated 3D Scene Graphs from Egocentric Vision (viability: 7): https://sciencetostartup.com/paper/pandora-articulated-3d-scene-graphs-from-egocentric-vision - Leverage human-centric egocentric vision to create articulated 3D scene graphs, enabling robots to understand and interact with previously inaccessible parts of an environment.
- SAGAI-MID: A Generative AI-Driven Middleware for Dynamic Runtime Interoperability (viability: 7): https://sciencetostartup.com/paper/sagai-mid-a-generative-ai-driven-middleware-for-dynamic-runtime-interoperability - A FastAPI-based middleware that uses LLMs to dynamically resolve schema mismatches in distributed systems at runtime, offering both per-request transformation and auto-generated adapter code.
- SOLE-R1: Video-Language Reasoning as the Sole Reward for On-Robot Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/sole-r1-video-language-reasoning-as-the-sole-reward-for-on-robot-reinforcement-learning - A video-language reasoning model that acts as the sole reward signal for robots, enabling them to learn new manipulation tasks without human supervision or ground truth rewards.
- BitSov: A Composable Bitcoin-Native Architecture for Sovereign Internet Infrastructure (viability: 3): https://sciencetostartup.com/paper/bitsov-a-composable-bitcoin-native-architecture-for-sovereign-internet-infrastructure - A composable Bitcoin-native architecture for a sovereign internet infrastructure, addressing censorship and economic rent extraction.
- Stepwise Credit Assignment for GRPO on Flow-Matching Models (viability: 4): https://sciencetostartup.com/paper/stepwise-credit-assignment-for-grpo-on-flow-matching-models - A novel reinforcement learning approach for flow models that improves sample efficiency and convergence by assigning credit based on step-wise reward improvements.
- Dynamic Dual-Granularity Skill Bank for Agentic RL (viability: 4): https://sciencetostartup.com/paper/dynamic-dual-granularity-skill-bank-for-agentic-rl - A dynamic skill bank for agentic reinforcement learning that improves performance by organizing reusable experience into task and step skills.
- DreamLite: A Lightweight On-Device Unified Model for Image Generation and Editing (viability: 8): https://sciencetostartup.com/paper/dreamlite-a-lightweight-on-device-unified-model-for-image-generation-and-editing - DreamLite provides efficient on-device image generation and editing within a single compact model.
- GPU-Accelerated Optimization of Transformer-Based Neural Networks for Real-Time Inference (viability: 7): https://sciencetostartup.com/paper/gpu-accelerated-optimization-of-transformer-based-neural-networks-for-real-time-inference - A GPU-accelerated inference pipeline for transformer models that achieves significant speedups and reduced memory usage with hybrid precision, enabling real-time deployment.
- A Convex Route to Thermomechanics: Learning Internal Energy and Dissipation (viability: 7): https://sciencetostartup.com/paper/a-convex-route-to-thermomechanics-learning-internal-energy-and-dissipation - A physics-based neural network framework for discovering constitutive models in thermomechanics, ensuring thermodynamic consistency and accuracy on synthetic and experimental data.
- EpiScreen: Early Epilepsy Detection from Electronic Health Records with Large Language Models (viability: 7): https://sciencetostartup.com/paper/episcreen-early-epilepsy-detection-from-electronic-health-records-with-large-language-models - Leverage LLMs to screen for epilepsy from EHR notes, improving diagnostic speed and accuracy for patients.
- AdaptToken: Entropy-based Adaptive Token Selection for MLLM Long Video Understanding (viability: 8): https://sciencetostartup.com/paper/adapttoken-entropy-based-adaptive-token-selection-for-mllm-long-video-understanding - AdaptToken: Efficient token selection for MLLMs to enhance long video understanding by leveraging entropy for global control.
- DRIVE-Nav: Directional Reasoning, Inspection, and Verification for Efficient Open-Vocabulary Navigation (viability: 7): https://sciencetostartup.com/paper/drive-nav-directional-reasoning-inspection-and-verification-for-efficient-open-vocabulary-navigation - DRIVE-Nav transforms open-vocabulary navigation by reducing redundant actions and enhancing path efficiency in AI-driven robots.
- Vision-Based Robotic Disassembly Combined with Real-Time MFA Data Acquisition (viability: 3): https://sciencetostartup.com/paper/vision-based-robotic-disassembly-combined-with-real-time-mfa-data-acquisition - Developing vision-based robotic disassembly and real-time material flow analysis for critical raw material recovery from end-of-use products.
- Functional Natural Policy Gradients (viability: 2): https://sciencetostartup.com/paper/functional-natural-policy-gradients - A theoretical framework for debiasing policy learning from offline data to achieve improved regret bounds.
- Subspace Optimization for Backpropagation-Free Continual Test-Time Adaptation (viability: 7): https://sciencetostartup.com/paper/subspace-optimization-for-backpropagation-free-continual-test-time-adaptation - A backpropagation-free system for efficient continual test-time adaptation that optimizes model performance under changing data distributions.
- Serialized Red-Green-Gray: Quicker Heuristic Validation of Edges in Dynamic Roadmap Graphs (viability: 5): https://sciencetostartup.com/paper/serialized-red-green-gray-quicker-heuristic-validation-of-edges-in-dynamic-roadmap-graphs - A framework for faster, more efficient motion planning in dynamic robotic environments by intelligently classifying roadmap edges.
- Why Aggregate Accuracy is Inadequate for Evaluating Fairness in Law Enforcement Facial Recognition Systems (viability: 4): https://sciencetostartup.com/paper/why-aggregate-accuracy-is-inadequate-for-evaluating-fairness-in-law-enforcement-facial-recognition-systems - Develop a fairness auditing tool for facial recognition systems that goes beyond aggregate accuracy to reveal demographic disparities.
- FL-PBM: Pre-Training Backdoor Mitigation for Federated Learning (viability: 7): https://sciencetostartup.com/paper/fl-pbm-pre-training-backdoor-mitigation-for-federated-learning - A novel defense mechanism for federated learning that proactively filters poisoned data on the client side before model training to mitigate backdoor attacks.
- Sim-to-Real Fruit Detection Using Synthetic Data: Quantitative Evaluation and Embedded Deployment with Isaac Sim (viability: 5): https://sciencetostartup.com/paper/sim-to-real-fruit-detection-using-synthetic-data-quantitative-evaluation-and-embedded-deployment-with-isaac-sim - Leveraging synthetic data with YOLO models for fruit detection on embedded systems, reducing annotation costs and enabling real-time inference.
- AMIGO: Agentic Multi-Image Grounding Oracle Benchmark (viability: 7): https://sciencetostartup.com/paper/amigo-agentic-multi-image-grounding-oracle-benchmark - A benchmark for evaluating agentic vision-language models in complex, multi-image target identification tasks.
- Industrial3D: A Terrestrial LiDAR Point Cloud Dataset and CrossParadigm Benchmark for Industrial Infrastructure (viability: 7): https://sciencetostartup.com/paper/industrial3d-a-terrestrial-lidar-point-cloud-dataset-and-crossparadigm-benchmark-for-industrial-infrastructure - A new large-scale terrestrial LiDAR dataset and benchmark for industrial 3D scene understanding, addressing critical gaps in digital transformation for construction.
- Divide and Restore: A Modular Task-Decoupled Framework for Universal Image Restoration (viability: 7): https://sciencetostartup.com/paper/divide-and-restore-a-modular-task-decoupled-framework-for-universal-image-restoration - A modular, task-decoupled framework for universal image restoration that intelligently routes degraded images to specialized, lightweight restoration models, reducing training overhead and enabling scalable integration of new degradation types.
- Safeguarding LLMs Against Misuse and AI-Driven Malware Using Steganographic Canaries (viability: 4): https://sciencetostartup.com/paper/safeguarding-llms-against-misuse-and-ai-driven-malware-using-steganographic-canaries - A framework using steganographic canary files to detect unauthorized LLM processing of sensitive documents and prevent AI-driven malware.
- Interpretable Ensemble Learning for Network Traffic Anomaly Detection: A SHAP-based Explainable AI Framework for Embedded Systems Security (viability: 5): https://sciencetostartup.com/paper/interpretable-ensemble-learning-for-network-traffic-anomaly-detection-a-shap-based-explainable-ai-framework-for-embedded - An explainable AI framework using ensemble learning and SHAP for network traffic anomaly detection in embedded systems.
- Mitigating Backdoor Attacks in Federated Learning Using PPA and MiniMax Game Theory (viability: 7): https://sciencetostartup.com/paper/mitigating-backdoor-attacks-in-federated-learning-using-ppa-and-minimax-game-theory - A federated learning defense system that uses reputation, incentives, and game theory to detect and neutralize malicious clients injecting backdoor attacks, significantly outperforming existing methods.
- Information-Theoretic Limits of Safety Verification for Self-Improving Systems (viability: 3): https://sciencetostartup.com/paper/information-theoretic-limits-of-safety-verification-for-self-improving-systems - Establishes theoretical limits on safety verification for self-improving AI systems, demonstrating a fundamental trade-off between bounded risk and unbounded utility.
- Constructing Composite Features for Interpretable Music-Tagging (viability: 5): https://sciencetostartup.com/paper/constructing-composite-features-for-interpretable-music-tagging - Develops an interpretable AI for music tagging by automatically evolving composite audio features using Genetic Programming, outperforming state-of-the-art methods.
- The Ultimate Tutorial for AI-driven Scale Development in Generative Psychometrics: Releasing AIGENIE from its Bottle (viability: 7): https://sciencetostartup.com/paper/the-ultimate-tutorial-for-ai-driven-scale-development-in-generative-psychometrics-releasing-aigenie-from-its-bottle - Automate psychological scale development using LLMs and network psychometrics, reducing expert involvement and pilot testing time.
- Empowering Mobile Networks Security Resilience by using Post-Quantum Cryptography (viability: 7): https://sciencetostartup.com/paper/empowering-mobile-networks-security-resilience-by-using-post-quantum-cryptography - Integrate post-quantum cryptography into 5G networks with a sidecar proxy to protect against future quantum threats without modifying existing functions.
- Dynamic Lookahead Distance via Reinforcement Learning-Based Pure Pursuit for Autonomous Racing (viability: 7): https://sciencetostartup.com/paper/dynamic-lookahead-distance-via-reinforcement-learning-based-pure-pursuit-for-autonomous-racing - A reinforcement learning agent dynamically adjusts the lookahead distance of a pure pursuit controller for faster and more stable autonomous racing.
- Trust-Aware Routing for Distributed Generative AI Inference at the Edge (viability: 7): https://sciencetostartup.com/paper/trust-aware-routing-for-distributed-generative-ai-inference-at-the-edge - A trust-aware routing framework for robust distributed generative AI inference at the edge, ensuring reliable performance even with unreliable peers.
- Seeing with You: Perception-Reasoning Coevolution for Multimodal Reasoning (viability: 7): https://sciencetostartup.com/paper/seeing-with-you-perception-reasoning-coevolution-for-multimodal-reasoning - A dual-role reinforcement learning framework that coevolves perception and reasoning in multimodal models to significantly improve visual evidence accuracy and overall reasoning performance.
- TGIF2: Extended Text-Guided Inpainting Forgery Dataset & Benchmark (viability: 7): https://sciencetostartup.com/paper/tgif2-extended-text-guided-inpainting-forgery-dataset-benchmark - A new dataset and benchmark to detect and localize AI-generated image forgeries, addressing limitations in current media forensics tools.
- LACE: Loss-Adaptive Capacity Expansion for Continual Learning (viability: 7): https://sciencetostartup.com/paper/lace-loss-adaptive-capacity-expansion-for-continual-learning - An online mechanism that dynamically expands model capacity during continual learning by monitoring loss signals, enabling efficient on-device adaptation without labels or replay buffers.
- ResAdapt: Adaptive Resolution for Efficient Multimodal Reasoning (viability: 7): https://sciencetostartup.com/paper/resadapt-adaptive-resolution-for-efficient-multimodal-reasoning - An input-side adaptation framework for multimodal LLMs that intelligently allocates visual budget per frame to significantly improve efficiency and accuracy, especially for reasoning-intensive tasks.
- Unsafe2Safe: Controllable Image Anonymization for Downstream Utility (viability: 7): https://sciencetostartup.com/paper/unsafe2safe-controllable-image-anonymization-for-downstream-utility - Automated pipeline for anonymizing sensitive image content using diffusion models, preserving downstream utility for privacy-safe datasets.
- ELViS: Efficient Visual Similarity from Local Descriptors that Generalizes Across Domains (viability: 7): https://sciencetostartup.com/paper/elvis-efficient-visual-similarity-from-local-descriptors-that-generalizes-across-domains - A highly efficient image-to-image similarity model that generalizes across diverse domains by operating in similarity space and refining local descriptor correspondences.
- Position: Explainable AI is Causality in Disguise (viability: 2): https://sciencetostartup.com/paper/position-explainable-ai-is-causality-in-disguise - This paper argues that explainable AI can only be achieved by grounding it in causal models, suggesting a shift in research focus towards causal discovery.
- Moving Beyond Review: Applying Language Models to Planning and Translation in Reflection (viability: 5): https://sciencetostartup.com/paper/moving-beyond-review-applying-language-models-to-planning-and-translation-in-reflection - A conversational AI tool that scaffolds reflective writing by assisting with planning and concept extraction, improving depth and structure.
- Optimistic Actor-Critic with Parametric Policies for Linear Markov Decision Processes (viability: 3): https://sciencetostartup.com/paper/optimistic-actor-critic-with-parametric-policies-for-linear-markov-decision-processes - Develops a theoretically sound actor-critic algorithm for linear MDPs with improved sample complexity.
- Detection of Adversarial Attacks in Robotic Perception (viability: 2): https://sciencetostartup.com/paper/detection-of-adversarial-attacks-in-robotic-perception - Developing novel detection strategies to secure deep neural networks used in robotic perception against adversarial attacks.
- Physics-Informed Framework for Impact Identification in Aerospace Composites (viability: 5): https://sciencetostartup.com/paper/physics-informed-framework-for-impact-identification-in-aerospace-composites - A physics-informed AI framework for accurate and stable impact identification in aerospace composites, even with noisy or limited data.
- MonitorBench: A Comprehensive Benchmark for Chain-of-Thought Monitorability in Large Language Models (viability: 7): https://sciencetostartup.com/paper/monitorbench-a-comprehensive-benchmark-for-chain-of-thought-monitorability-in-large-language-models - MonitorBench provides a comprehensive benchmark to evaluate and improve the trustworthiness of Large Language Model reasoning by quantifying chain-of-thought monitorability.
- Towards a Medical AI Scientist (viability: 7): https://sciencetostartup.com/paper/towards-a-medical-ai-scientist - An autonomous AI framework for accelerating medical research by generating hypotheses, conducting experiments, and drafting manuscripts grounded in clinical evidence.
- ORSIFlow: Saliency-Guided Rectified Flow for Optical Remote Sensing Salient Object Detection (viability: 7): https://sciencetostartup.com/paper/orsiflow-saliency-guided-rectified-flow-for-optical-remote-sensing-salient-object-detection - A saliency-guided rectified flow framework for efficient and state-of-the-art optical remote sensing salient object detection.
- Navigating the Mirage: A Dual-Path Agentic Framework for Robust Misleading Chart Question Answering (viability: 7): https://sciencetostartup.com/paper/navigating-the-mirage-a-dual-path-agentic-framework-for-robust-misleading-chart-question-answering - A dual-path agentic framework that uses skeptical reasoning to accurately answer questions from misleading charts, outperforming state-of-the-art models.
- A Self-Rotating Tri-Rotor UAV for Field of View Expansion and Autonomous Flight (viability: 4): https://sciencetostartup.com/paper/a-self-rotating-tri-rotor-uav-for-field-of-view-expansion-and-autonomous-flight - A self-rotating drone that expands sensor field-of-view for enhanced environmental perception and autonomous flight.
- EBuddy: a workflow orchestrator for industrial human-machine collaboration (viability: 7): https://sciencetostartup.com/paper/ebuddy-a-workflow-orchestrator-for-industrial-human-machine-collaboration - EBuddy orchestrates industrial human-robot collaboration with voice guidance to streamline complex tool-intensive workflows and reduce process duration.
- ChemCLIP: Bridging Organic and Inorganic Anticancer Compounds Through Contrastive Learning (viability: 7): https://sciencetostartup.com/paper/chemclip-bridging-organic-and-inorganic-anticancer-compounds-through-contrastive-learning - A contrastive learning framework that unifies organic and inorganic anticancer compounds into a shared representation space for accelerated drug discovery.
- Learning Partial Action Replacement in Offline MARL (viability: 7): https://sciencetostartup.com/paper/learning-partial-action-replacement-in-offline-marl - A new framework for offline multi-agent reinforcement learning that adaptively selects partial action replacements to significantly improve computational efficiency and performance on complex benchmarks.
- Unrestrained Simplex Denoising for Discrete Data. A Non-Markovian Approach Applied to Graph Generation (viability: 4): https://sciencetostartup.com/paper/unrestrained-simplex-denoising-for-discrete-data-a-non-markovian-approach-applied-to-graph-generation - A novel generative framework for discrete data that simplifies formulation and improves performance on graph benchmarks.
- CirrusBench: Evaluating LLM-based Agents Beyond Correctness in Real-World Cloud Service Environments (viability: 7): https://sciencetostartup.com/paper/cirrusbench-evaluating-llm-based-agents-beyond-correctness-in-real-world-cloud-service-environments - A new benchmark for evaluating LLM agents in real-world cloud service environments, focusing on resolution efficiency beyond just correctness.
- XSPA: Crafting Imperceptible X-Shaped Sparse Adversarial Perturbations for Transferable Attacks on VLMs (viability: 7): https://sciencetostartup.com/paper/xspa-crafting-imperceptible-x-shaped-sparse-adversarial-perturbations-for-transferable-attacks-on-vlms - Develop a security tool that crafts imperceptible adversarial attacks to test and harden Vision-Language Models against semantic failures.
- StreamingVLA: Streaming Vision-Language-Action Model with Action Flow Matching and Adaptive Early Observation (viability: 7): https://sciencetostartup.com/paper/streamingvla-streaming-vision-language-action-model-with-action-flow-matching-and-adaptive-early-observation - A streaming vision-language-action model that significantly reduces latency and improves execution fluency for real-world edge deployments.
- Fine-Tuning Large Language Models for Cooperative Tactical Deconfliction of Small Unmanned Aerial Systems (viability: 7): https://sciencetostartup.com/paper/fine-tuning-large-language-models-for-cooperative-tactical-deconfliction-of-small-unmanned-aerial-systems - Fine-tuning LLMs with a novel simulation-to-language pipeline to enable safe and efficient cooperative deconfliction for swarms of small drones.
- Curriculum-Guided Myocardial Scar Segmentation for Ischemic and Non-ischemic Cardiomyopathy (viability: 7): https://sciencetostartup.com/paper/curriculum-guided-myocardial-scar-segmentation-for-ischemic-and-non-ischemic-cardiomyopathy - A curriculum learning framework for improved myocardial scar segmentation from cardiac MRI, addressing challenges of inconsistent annotations and subtle scar appearances.
- T-Norm Operators for EU AI Act Compliance Classification: An Empirical Comparison of Lukasiewicz, Product, and Gödel Semantics in a Neuro-Symbolic Reasoning System (viability: 7): https://sciencetostartup.com/paper/t-norm-operators-for-eu-ai-act-compliance-classification-an-empirical-comparison-of-lukasiewicz-product-and-g-del-semant - A neuro-symbolic system for classifying AI systems under the EU AI Act, offering a trade-off between accuracy and false positives with released code and data.
- Domain-Invariant Prompt Learning for Vision-Language Models (viability: 4): https://sciencetostartup.com/paper/domain-invariant-prompt-learning-for-vision-language-models - A method to improve vision-language models' ability to generalize across different visual domains by learning domain-invariant prompts.
- Hydra: Unifying Document Retrieval and Generation in a Single Vision-Language Model (viability: 7): https://sciencetostartup.com/paper/hydra-unifying-document-retrieval-and-generation-in-a-single-vision-language-model - A unified vision-language model that performs both document retrieval and generation with reduced memory footprint and byte-identical output quality.
- Multimodal Analytics of Cybersecurity Crisis Preparation Exercises: What Predicts Success? (viability: 4): https://sciencetostartup.com/paper/multimodal-analytics-of-cybersecurity-crisis-preparation-exercises-what-predicts-success - Develop a system to measure instructional alignment in cybersecurity simulations using multimodal data to predict team success.
- "What Did It Actually Do?": Understanding Risk Awareness and Traceability for Computer-Use Agents (viability: 4): https://sciencetostartup.com/paper/what-did-it-actually-do-understanding-risk-awareness-and-traceability-for-computer-use-agents - A framework to improve user understanding of AI agent actions and risks, enhancing trust and anomaly detection.
- MarkushGrapher-2: End-to-end Multimodal Recognition of Chemical Structures (viability: 8): https://sciencetostartup.com/paper/markushgrapher-2-end-to-end-multimodal-recognition-of-chemical-structures - MarkushGrapher-2 enables automated extraction of complex chemical structures from patents using multimodal recognition, accelerating chemical R&D.
- Seen2Scene: Completing Realistic 3D Scenes with Visibility-Guided Flow (viability: 7): https://sciencetostartup.com/paper/seen2scene-completing-realistic-3d-scenes-with-visibility-guided-flow - Generate realistic and complete 3D scenes from incomplete real-world scans using visibility-guided flow matching.
- GEditBench v2: A Human-Aligned Benchmark for General Image Editing (viability: 7): https://sciencetostartup.com/paper/geditbench-v2-a-human-aligned-benchmark-for-general-image-editing - A human-aligned benchmark and evaluation model for general image editing that surpasses GPT-5.1, enabling more precise model development.
- Shy Guys: A Light-Weight Approach to Detecting Robots on Websites (viability: 5): https://sciencetostartup.com/paper/shy-guys-a-light-weight-approach-to-detecting-robots-on-websites - A lightweight, passive web bot detection system that analyzes server logs to identify and filter malicious bots with high accuracy.
- ManipArena: Comprehensive Real-world Evaluation of Reasoning-Oriented Generalist Robot Manipulation (viability: 7): https://sciencetostartup.com/paper/maniparena-comprehensive-real-world-evaluation-of-reasoning-oriented-generalist-robot-manipulation - A standardized real-world evaluation framework for generalist robot manipulation models to bridge the simulation-to-reality gap.
- Feel Robot Feels: Tactile Feedback Array Glove for Dexterous Manipulation (viability: 7): https://sciencetostartup.com/paper/feel-robot-feels-tactile-feedback-array-glove-for-dexterous-manipulation - A low-cost glove with precise motion capture and high-resolution tactile feedback for enhanced robotic teleoperation and data collection.
- Training data generation for context-dependent rubric-based short answer grading (viability: 5): https://sciencetostartup.com/paper/training-data-generation-for-context-dependent-rubric-based-short-answer-grading - Generate large-scale, confidential training datasets for automated student answer grading using novel text transformation techniques.
- Compressing Transformer Language Models via Matrix Product Operator Decomposition: A Case Study on PicoGPT (viability: 4): https://sciencetostartup.com/paper/compressing-transformer-language-models-via-matrix-product-operator-decomposition-a-case-study-on-picogpt - Compress transformer language models using Matrix Product Operator decomposition for efficient deployment on resource-constrained hardware.
- GraphWalker: Agentic Knowledge Graph Question Answering via Synthetic Trajectory Curriculum (viability: 7): https://sciencetostartup.com/paper/graphwalker-agentic-knowledge-graph-question-answering-via-synthetic-trajectory-curriculum - GraphWalker is an agentic KGQA framework that synthesizes diverse training trajectories and uses stage-wise fine-tuning to achieve state-of-the-art performance and improved generalization.
- Detecting low left ventricular ejection fraction from ECG using an interpretable and scalable predictor-driven framework (viability: 7): https://sciencetostartup.com/paper/detecting-low-left-ventricular-ejection-fraction-from-ecg-using-an-interpretable-and-scalable-predictor-driven-framework - A scalable and interpretable AI framework for early detection of low left ventricular ejection fraction from ECGs, outperforming existing black-box models.
- EarlySciRev: A Dataset of Early-Stage Scientific Revisions Extracted from LaTeX Writing Traces (viability: 4): https://sciencetostartup.com/paper/earlyscirev-a-dataset-of-early-stage-scientific-revisions-extracted-from-latex-writing-traces - A new dataset of early-stage scientific revisions extracted from LaTeX traces to improve LLM capabilities in scientific writing.
- TIEG-Youpu Solution for NeurIPS 2022 WikiKG90Mv2-LSC (viability: 4): https://sciencetostartup.com/paper/tieg-youpu-solution-for-neurips-2022-wikikg90mv2-lsc - A novel approach to knowledge graph embedding that improves retrieval and re-ranking for large-scale encyclopedic knowledge graphs.
- Generalizable Detection of AI Generated Images with Large Models and Fuzzy Decision Tree (viability: 7): https://sciencetostartup.com/paper/generalizable-detection-of-ai-generated-images-with-large-models-and-fuzzy-decision-tree - A novel AI-generated image detection framework that combines artifact-aware detectors with MLLMs for improved accuracy and generalization.
- The Unreasonable Effectiveness of Scaling Laws in AI (viability: 2): https://sciencetostartup.com/paper/the-unreasonable-effectiveness-of-scaling-laws-in-ai - This paper theoretically explores the effectiveness of scaling laws in AI training and their implications for efficiency improvements.
- Bridging the Geometry Mismatch: Frequency-Aware Anisotropic Serialization for Thin-Structure SSMs (viability: 7): https://sciencetostartup.com/paper/bridging-the-geometry-mismatch-frequency-aware-anisotropic-serialization-for-thin-structure-ssms - A novel framework for accurate segmentation of thin structures by aligning feature serialization with geometric properties, outperforming baselines and achieving high FPS.
- Next-Token Prediction and Regret Minimization (viability: 3): https://sciencetostartup.com/paper/next-token-prediction-and-regret-minimization - This paper explores how to make next-token prediction models robust to adversarial decision-making environments, with theoretical guarantees and empirical validation on transformer architectures.
- ConceptWeaver: Weaving Disentangled Concepts with Flow (viability: 7): https://sciencetostartup.com/paper/conceptweaver-weaving-disentangled-concepts-with-flow - A framework for one-shot disentanglement and manipulation of concepts in flow-based generative models.
- Courtroom-Style Multi-Agent Debate with Progressive RAG and Role-Switching for Controversial Claim Verification (viability: 7): https://sciencetostartup.com/paper/courtroom-style-multi-agent-debate-with-progressive-rag-and-role-switching-for-controversial-claim-verification - A courtroom-inspired multi-agent system with progressive evidence retrieval and role-switching to reliably verify controversial claims.
- INSID3: Training-Free In-Context Segmentation with DINOv3 (viability: 7): https://sciencetostartup.com/paper/insid3-training-free-in-context-segmentation-with-dinov3 - A training-free in-context segmentation method leveraging frozen DINOv3 features to achieve state-of-the-art results with reduced complexity.
- With a Little Help From My Friends: Collective Manipulation in Risk-Controlling Recommender Systems (viability: 5): https://sciencetostartup.com/paper/with-a-little-help-from-my-friends-collective-manipulation-in-risk-controlling-recommender-systems - A system to defend recommender systems against coordinated adversarial manipulation by shifting safety guarantees from group to user level.
- Tac2Real: Reliable and GPU Visuotactile Simulation for Online Reinforcement Learning and Zero-Shot Real-World Deployment (viability: 7): https://sciencetostartup.com/paper/tac2real-reliable-and-gpu-visuotactile-simulation-for-online-reinforcement-learning-and-zero-shot-real-world-deployment - A visuotactile simulation framework enabling efficient online reinforcement learning and zero-shot sim-to-real transfer for robotic manipulation.
- CiQi-Agent: Aligning Vision, Tools and Aesthetics in Multimodal Agent for Cultural Reasoning on Chinese Porcelains (viability: 8): https://sciencetostartup.com/paper/ciqi-agent-aligning-vision-tools-and-aesthetics-in-multimodal-agent-for-cultural-reasoning-on-chinese-porcelains - Develop an AI-powered platform for accessible Chinese porcelain connoisseurship using multimodal reasoning and fine-grained attribute analysis.
- Communications-Aware NMPC for Multi-Rotor Aerial Relay Networks Under Jamming Interference (viability: 4): https://sciencetostartup.com/paper/communications-aware-nmpc-for-multi-rotor-aerial-relay-networks-under-jamming-interference - A control framework for aerial robots that optimizes communication links under jamming by integrating motion and antenna orientation.
- Post-hoc Self-explanation of CNNs (viability: 4): https://sciencetostartup.com/paper/post-hoc-self-explanation-of-cnns - This research proposes a novel method to improve the interpretability of Convolutional Neural Networks by integrating k-means clustering with feature activations, offering concept-based explanation maps.
- Decoupling Wavelet Sub-bands for Single Source Domain Generalization in Fundus Image Segmentation (viability: 6): https://sciencetostartup.com/paper/decoupling-wavelet-sub-bands-for-single-source-domain-generalization-in-fundus-image-segmentation - A wavelet-guided segmentation network for robust fundus image analysis across different acquisition devices and clinical settings.
- $R_{dm}$: Re-conceptualizing Distribution Matching as a Reward for Diffusion Distillation (viability: 7): https://sciencetostartup.com/paper/r-dm-re-conceptualizing-distribution-matching-as-a-reward-for-diffusion-distillation - A novel reward-based framework for diffusion model distillation that significantly improves sampling speed and generation quality, with code to be released.
- HISA: Efficient Hierarchical Indexing for Fine-Grained Sparse Attention (viability: 7): https://sciencetostartup.com/paper/hisa-efficient-hierarchical-indexing-for-fine-grained-sparse-attention - HISA offers a drop-in replacement for sparse attention indexers, enabling 2-4x speedups at long context lengths with minimal quality loss.
- FeDMRA: Federated Incremental Learning with Dynamic Memory Replay Allocation (viability: 7): https://sciencetostartup.com/paper/fedmra-federated-incremental-learning-with-dynamic-memory-replay-allocation - A dynamic memory allocation strategy for federated incremental learning that improves model performance and fairness in healthcare settings by intelligently managing data replay.
- Entropic Claim Resolution: Uncertainty-Driven Evidence Selection for RAG (viability: 4): https://sciencetostartup.com/paper/entropic-claim-resolution-uncertainty-driven-evidence-selection-for-rag - A novel algorithm for Retrieval-Augmented Generation that selects evidence based on uncertainty reduction to resolve conflicting information.
- A Predictive Control Strategy to Offset-Point Tracking for Agricultural Mobile Robots (viability: 7): https://sciencetostartup.com/paper/a-predictive-control-strategy-to-offset-point-tracking-for-agricultural-mobile-robots - A predictive control strategy for agricultural robots that significantly reduces tracking errors and crop damage by accounting for the implement's spatial footprint.
- Democratizing Federated Learning with Blockchain and Multi-Task Peer Prediction (viability: 3): https://sciencetostartup.com/paper/democratizing-federated-learning-with-blockchain-and-multi-task-peer-prediction - A theoretical framework for decentralizing AI training using blockchain and multi-task peer prediction.
- GeoHCC: Local Geometry-Aware Hierarchical Context Compression for 3D Gaussian Splatting (viability: 7): https://sciencetostartup.com/paper/geohcc-local-geometry-aware-hierarchical-context-compression-for-3d-gaussian-splatting - A geometry-aware compression framework for 3D Gaussian Splatting that significantly reduces storage overhead while preserving rendering quality and structural integrity.
- IsoQuant: Hardware-Aligned SO(4) Isoclinic Rotations for LLM KV Cache Compression (viability: 3): https://sciencetostartup.com/paper/isoquant-hardware-aligned-so-4-isoclinic-rotations-for-llm-kv-cache-compression - A novel quaternion-based rotation framework for compressing LLM KV caches, aiming for significant speedups and reduced memory footprint.
- AceleradorSNN: A Neuromorphic Cognitive System Integrating Spiking Neural Networks and DynamicImage Signal Processing on FPGA (viability: 3): https://sciencetostartup.com/paper/aceleradorsnn-a-neuromorphic-cognitive-system-integrating-spiking-neural-networks-and-dynamicimage-signal-processing-on - A neuromorphic cognitive system integrating Spiking Neural Networks and dynamic image signal processing on FPGA for low-latency object detection.
- Tele-Catch: Adaptive Teleoperation for Dexterous Dynamic 3D Object Catching (viability: 7): https://sciencetostartup.com/paper/tele-catch-adaptive-teleoperation-for-dexterous-dynamic-3d-object-catching - A shared autonomy framework for dexterous robot hand teleoperation in dynamic object catching, improving accuracy and robustness through adaptive integration and geometry-aware decision making.
- Structural-Ambiguity-Aware Translation from Natural Language to Signal Temporal Logic (viability: 4): https://sciencetostartup.com/paper/structural-ambiguity-aware-translation-from-natural-language-to-signal-temporal-logic - A system that translates ambiguous natural language instructions into multiple plausible Signal Temporal Logic formulas for cyber-physical systems.
- From Pixels to Reality: Physical-Digital Patch Attacks on Real-World Camera (viability: 7): https://sciencetostartup.com/paper/from-pixels-to-reality-physical-digital-patch-attacks-on-real-world-camera - Develops a novel digital-physical adversarial attack method for camera-based authentication systems, demonstrating real-time evasion capabilities with potential for security product development.
- Profile Graphical Models (viability: 4): https://sciencetostartup.com/paper/profile-graphical-models - A new class of graphical models that capture how external factors influence variable dependencies, with applications in biological data analysis.
- Active Stereo-Camera Outperforms Multi-Sensor Setup in ACT Imitation Learning for Humanoid Manipulation (viability: 7): https://sciencetostartup.com/paper/active-stereo-camera-outperforms-multi-sensor-setup-in-act-imitation-learning-for-humanoid-manipulation - This research demonstrates that a minimal active stereo-camera setup significantly outperforms complex multi-sensor configurations for humanoid robot imitation learning, offering a more efficient and robust approach to task acquisition.
- Spectral Higher-Order Neural Networks (viability: 2): https://sciencetostartup.com/paper/spectral-higher-order-neural-networks - A new algorithmic strategy to incorporate higher-order interactions in general-purpose, feedforward, network structures by leveraging spectral attributes.
- LombardoGraphia: Automatic Classification of Lombard Orthography Variants (viability: 7): https://sciencetostartup.com/paper/lombardographia-automatic-classification-of-lombard-orthography-variants - Automates the classification of diverse Lombard language orthographies to enable the development of variety-aware NLP tools for an underresourced language.
- KGroups: A Versatile Univariate Max-Relevance Min-Redundancy Feature Selection Algorithm for High-dimensional Biological Data (viability: 4): https://sciencetostartup.com/paper/kgroups-a-versatile-univariate-max-relevance-min-redundancy-feature-selection-algorithm-for-high-dimensional-biological - A faster, parameterizable feature selection algorithm for high-dimensional biological data that matches multivariate performance.
- Evolutionary Discovery of Reinforcement Learning Algorithms via Large Language Models (viability: 7): https://sciencetostartup.com/paper/evolutionary-discovery-of-reinforcement-learning-algorithms-via-large-language-models - An evolutionary framework that utilizes large language models to discover novel reinforcement learning algorithms through direct search of executable update rules.
- Unified Restoration-Perception Learning: Maritime Infrared-Visible Image Fusion and Segmentation (viability: 7): https://sciencetostartup.com/paper/unified-restoration-perception-learning-maritime-infrared-visible-image-fusion-and-segmentation - A unified framework for maritime infrared-visible image fusion and segmentation that significantly improves perception and robustness in challenging marine environments.
- Mixture-Model Preference Learning for Many-Objective Bayesian Optimization (viability: 5): https://sciencetostartup.com/paper/mixture-model-preference-learning-for-many-objective-bayesian-optimization - A Bayesian framework for preference-based optimization that learns latent preference archetypes to efficiently navigate complex trade-offs.
- MiroEval: Benchmarking Multimodal Deep Research Agents in Process and Outcome (viability: 7): https://sciencetostartup.com/paper/miroeval-benchmarking-multimodal-deep-research-agents-in-process-and-outcome - A new benchmark and evaluation framework for multimodal AI research agents that assesses both process and outcome, revealing critical performance gaps.
- EdgeDiT: Hardware-Aware Diffusion Transformers for Efficient On-Device Image Generation (viability: 7): https://sciencetostartup.com/paper/edgedit-hardware-aware-diffusion-transformers-for-efficient-on-device-image-generation - EdgeDiT enables efficient, private, and offline image generation directly on mobile devices by optimizing Diffusion Transformers for NPUs.
- Label-efficient Training Updates for Malware Detection over Time (viability: 4): https://sciencetostartup.com/paper/label-efficient-training-updates-for-malware-detection-over-time - A model-agnostic framework for malware detection that significantly reduces labeling costs by combining active learning and semi-supervised learning techniques to combat evolving threats.
- From Simulation to Deep Learning: Survey on Network Performance Modeling Approaches (viability: 2): https://sciencetostartup.com/paper/from-simulation-to-deep-learning-survey-on-network-performance-modeling-approaches - This paper surveys existing approaches to network performance modeling, from traditional simulation to machine learning, to understand the evolution and evaluation of these techniques.
- SVH-BD : Synthetic Vegetation Hyperspectral Benchmark Dataset for Emulation of Remote Sensing Images (viability: 5): https://sciencetostartup.com/paper/svh-bd-synthetic-vegetation-hyperspectral-benchmark-dataset-for-emulation-of-remote-sensing-images - A new synthetic hyperspectral dataset with vegetation trait maps to accelerate research in remote sensing emulation and biophysical parameter retrieval.
- The Scaffold Effect: How Prompt Framing Drives Apparent Multimodal Gains in Clinical VLM Evaluation (viability: 4): https://sciencetostartup.com/paper/the-scaffold-effect-how-prompt-framing-drives-apparent-multimodal-gains-in-clinical-vlm-evaluation - This research reveals a critical 'scaffold effect' in clinical vision-language models, where prompt framing, not actual data integration, drives apparent performance gains, highlighting a significant risk for clinical deployment.
- COvolve: Adversarial Co-Evolution of Large-Language-Model-Generated Policies and Environments via Two-Player Zero-Sum Game (viability: 7): https://sciencetostartup.com/paper/covolve-adversarial-co-evolution-of-large-language-model-generated-policies-and-environments-via-two-player-zero-sum-gam - An AI framework that uses LLMs to adversarially co-evolve agent policies and training environments for open-ended learning and robust generalization.
- Critic-Free Deep Reinforcement Learning for Maritime Coverage Path Planning on Irregular Hexagonal Grids (viability: 7): https://sciencetostartup.com/paper/critic-free-deep-reinforcement-learning-for-maritime-coverage-path-planning-on-irregular-hexagonal-grids - A critic-free deep reinforcement learning framework for efficient maritime coverage path planning on irregular grids, outperforming traditional methods with real-time inference.
- Membership Inference Attacks against Large Audio Language Models (viability: 4): https://sciencetostartup.com/paper/membership-inference-attacks-against-large-audio-language-models - Develops a novel method to audit Large Audio Language Models for membership inference attacks, ensuring data privacy and model security.
- Marco DeepResearch: Unlocking Efficient Deep Research Agents via Verification-Centric Design (viability: 7): https://sciencetostartup.com/paper/marco-deepresearch-unlocking-efficient-deep-research-agents-via-verification-centric-design - A deep research agent that significantly improves performance on complex tasks by integrating verification mechanisms throughout its design, outperforming larger models within budget constraints.
- Coherent Without Grounding, Grounded Without Success: Observability and Epistemic Failure (viability: 3): https://sciencetostartup.com/paper/coherent-without-grounding-grounded-without-success-observability-and-epistemic-failure - This paper introduces a new framework for evaluating AI understanding by disentangling coherence, grounding, and action, challenging current AI evaluation practices.
- Tailoring AI-Driven Reading Scaffolds to the Distinct Needs of Neurodiverse Learners (viability: 3): https://sciencetostartup.com/paper/tailoring-ai-driven-reading-scaffolds-to-the-distinct-needs-of-neurodiverse-learners - Developing adaptive AI-powered reading scaffolds tailored to the unique needs of neurodiverse learners to improve comprehension and engagement.
- Rethinking Structure Preservation in Text-Guided Image Editing with Visual Autoregressive Models (viability: 7): https://sciencetostartup.com/paper/rethinking-structure-preservation-in-text-guided-image-editing-with-visual-autoregressive-models - A novel framework for text-guided image editing using visual autoregressive models that significantly improves structural consistency and editing quality.
- AutoCut: End-to-end advertisement video editing based on multimodal discretization and controllable generation (viability: 7): https://sciencetostartup.com/paper/autocut-end-to-end-advertisement-video-editing-based-on-multimodal-discretization-and-controllable-generation - An end-to-end framework for automated advertisement video editing that unifies video, audio, and text for scalable content creation.
- SEA: Evaluating Sketch Abstraction Efficiency via Element-level Commonsense Visual Question Answering (viability: 7): https://sciencetostartup.com/paper/sea-evaluating-sketch-abstraction-efficiency-via-element-level-commonsense-visual-question-answering - A novel metric and dataset for evaluating the efficiency of semantic abstraction in sketches, enabling better understanding of visual summarization.
- A Foldable and Agile Soft Electromagnetic Robot for Multimodal Navigation in Confined and Unstructured Environments (viability: 3): https://sciencetostartup.com/paper/a-foldable-and-agile-soft-electromagnetic-robot-for-multimodal-navigation-in-confined-and-unstructured-environments - A foldable soft electromagnetic robot with multimodal locomotion for navigating confined and unstructured environments, particularly for biomedical applications.
- Deep Research of Deep Research: From Transformer to Agent, From AI to AI for Science (viability: 3): https://sciencetostartup.com/paper/deep-research-of-deep-research-from-transformer-to-agent-from-ai-to-ai-for-science - This paper provides a deep research of deep research, articulating a definition for deep research and unifying perspectives from industry and academia within a developmental framework.
- CoE: Collaborative Entropy for Uncertainty Quantification in Agentic Multi-LLM Systems (viability: 4): https://sciencetostartup.com/paper/coe-collaborative-entropy-for-uncertainty-quantification-in-agentic-multi-llm-systems - A novel information-theoretic metric to quantify semantic disagreement and collaborative confidence across multiple Large Language Models.
- Optimized Weighted Voting System for Brain Tumor Classification Using MRI Images (viability: 7): https://sciencetostartup.com/paper/optimized-weighted-voting-system-for-brain-tumor-classification-using-mri-images - An ensemble learning system for highly accurate brain tumor classification from MRI scans, outperforming existing models.
- VistaGEN: Consistent Driving Video Generation with Fine-Grained Control Using Multiview Visual-Language Reasoning (viability: 4): https://sciencetostartup.com/paper/vistagen-consistent-driving-video-generation-with-fine-grained-control-using-multiview-visual-language-reasoning - A novel technique for generating driving videos with fine-grained object-level control and improved spatiotemporal consistency.
- Not All Subjectivity Is the Same! Defining Desiderata for the Evaluation of Subjectivity in NLP (viability: 3): https://sciencetostartup.com/paper/not-all-subjectivity-is-the-same-defining-desiderata-for-the-evaluation-of-subjectivity-in-nlp - This paper proposes new evaluation criteria for NLP models that handle subjective content, identifying gaps in current research and practice.
- Proposing a Game Theory Approach to Explore Group Dynamics with Social Robot (viability: 3): https://sciencetostartup.com/paper/proposing-a-game-theory-approach-to-explore-group-dynamics-with-social-robot - A game theory framework to study how social robots can foster cooperation in human groups for educational and workplace applications.
- Machine Learning-Assisted High-Dimensional Matrix Estimation (viability: 3): https://sciencetostartup.com/paper/machine-learning-assisted-high-dimensional-matrix-estimation - A theoretical framework for improving high-dimensional matrix estimation using machine learning-assisted optimization.
- Crossing the NL/PL Divide: Information Flow Analysis Across the NL/PL Boundary in LLM-Integrated Code (viability: 7): https://sciencetostartup.com/paper/crossing-the-nl-pl-divide-information-flow-analysis-across-the-nl-pl-boundary-in-llm-integrated-code - This research introduces the first method to analyze information flow across the boundary between natural language prompts and program code generated by LLMs, enabling new tools for program analysis and security.
- Kernel-Smith: A Unified Recipe for Evolutionary Kernel Optimization (viability: 8): https://sciencetostartup.com/paper/kernel-smith-a-unified-recipe-for-evolutionary-kernel-optimization - Kernel-Smith optimizes GPU kernels for enhanced performance using an evolutionary approach, surpassing state-of-the-art methods.
- Users and Wizards in Conversations: How WoZ Interface Choices Define Human-Robot Interactions (viability: 3): https://sciencetostartup.com/paper/users-and-wizards-in-conversations-how-woz-interface-choices-define-human-robot-interactions - This research explores how different Wizard-of-Oz interfaces impact human-robot communication, suggesting VR telepresence offers richer interactions.
- A Multi-Agent Rhizomatic Pipeline for Non-Linear Literature Analysis (viability: 7): https://sciencetostartup.com/paper/a-multi-agent-rhizomatic-pipeline-for-non-linear-literature-analysis - An open-source multi-agent AI pipeline for non-linear literature analysis, uncovering hidden connections and research gaps missed by traditional methods.
- Key-Embedded Privacy for Decentralized AI in Biomedical Omics (viability: 7): https://sciencetostartup.com/paper/key-embedded-privacy-for-decentralized-ai-in-biomedical-omics - A lightweight federated learning method for biomedical omics that embeds secret keys directly into model architecture to ensure privacy without sacrificing performance.
- Integrating Multimodal Large Language Model Knowledge into Amodal Completion (viability: 4): https://sciencetostartup.com/paper/integrating-multimodal-large-language-model-knowledge-into-amodal-completion - Leveraging Multimodal Large Language Models to improve amodal completion for autonomous systems by reasoning about occluded object parts.
- Physics-Informed Neural Networks for Predicting Hydrogen Sorption in Geological Formations: Thermodynamically Constrained Deep Learning Integrating Classical Adsorption Theory (viability: 7): https://sciencetostartup.com/paper/physics-informed-neural-networks-for-predicting-hydrogen-sorption-in-geological-formations-thermodynamically-constrained - A physics-informed neural network that accurately predicts hydrogen sorption in geological formations, outperforming traditional models and enabling better underground hydrogen storage.
- Building evidence-based knowledge graphs from full-text literature for disease-specific biomedical reasoning (viability: 7): https://sciencetostartup.com/paper/building-evidence-based-knowledge-graphs-from-full-text-literature-for-disease-specific-biomedical-reasoning - A framework and dataset for building disease-specific knowledge graphs from biomedical literature using LLMs to extract and structure evidence for improved reasoning and hypothesis generation.
- LDDMM stochastic interpolants: an application to domain uncertainty quantification in hemodynamics (viability: 4): https://sciencetostartup.com/paper/lddmm-stochastic-interpolants-an-application-to-domain-uncertainty-quantification-in-hemodynamics - A novel generative modeling framework for 3D shapes that quantifies domain uncertainty in medical simulations.
- SFDemorpher: Generalizable Face Demorphing for Operational Morphing Attack Detection (viability: 7): https://sciencetostartup.com/paper/sfdemorpher-generalizable-face-demorphing-for-operational-morphing-attack-detection - A generalizable face demorphing framework to detect sophisticated identity manipulation in biometric documents, enhancing security at border control and document issuance.
- FairGC: Fairness-aware Graph Condensation (viability: 7): https://sciencetostartup.com/paper/fairgc-fairness-aware-graph-condensation - FairGC is a framework for compressing large graph datasets while preserving fairness, making them suitable for sensitive applications.
- Beyond Scanpaths: Graph-Based Gaze Simulation in Dynamic Scenes (viability: 7): https://sciencetostartup.com/paper/beyond-scanpaths-graph-based-gaze-simulation-in-dynamic-scenes - A novel graph-based approach to simulate human gaze in dynamic scenes, offering more naturalistic attention modeling for applications like automotive safety.
- Mapping data literacy trajectories in K-12 education (viability: 3): https://sciencetostartup.com/paper/mapping-data-literacy-trajectories-in-k-12-education - A framework for understanding data literacy learning trajectories in K-12 education.
- Taming the Instability: A Robust Second-Order Optimizer for Federated Learning over Non-IID Data (viability: 4): https://sciencetostartup.com/paper/taming-the-instability-a-robust-second-order-optimizer-for-federated-learning-over-non-iid-data - A robust second-order optimizer for federated learning that improves convergence and reduces communication costs on non-IID data.
- Prototype-Enhanced Multi-View Learning for Thyroid Nodule Ultrasound Classification (viability: 7): https://sciencetostartup.com/paper/prototype-enhanced-multi-view-learning-for-thyroid-nodule-ultrasound-classification - A prototype-enhanced multi-view learning framework for robust thyroid nodule classification on heterogeneous ultrasound images, improving diagnostic accuracy and generalization.
- Cryptanalysis of a Lightweight RFID Authentication Protocol Based on a Variable Matrix Encryption Algorithm (viability: 2): https://sciencetostartup.com/paper/cryptanalysis-of-a-lightweight-rfid-authentication-protocol-based-on-a-variable-matrix-encryption-algorithm - This paper analyzes and identifies structural weaknesses in a lightweight RFID authentication protocol, demonstrating a realistic attack path to full compromise.
- VulnScout-C: A Lightweight Transformer for C Code Vulnerability Detection (viability: 7): https://sciencetostartup.com/paper/vulnscout-c-a-lightweight-transformer-for-c-code-vulnerability-detection - A compact transformer model and curated dataset for efficient and effective C code vulnerability detection, outperforming larger models and commercial tools.
- Self++: Co-Determined Agency for Human--AI Symbiosis in Extended Reality (viability: 3): https://sciencetostartup.com/paper/self-co-determined-agency-for-human-ai-symbiosis-in-extended-reality - A design framework for human-AI symbiosis in XR that maintains human authorship and agency through co-determination principles.
- The Necessity of Setting Temperature in LLM-as-a-Judge (viability: 3): https://sciencetostartup.com/paper/the-necessity-of-setting-temperature-in-llm-as-a-judge - This paper investigates the impact of temperature settings on LLM-as-a-Judge performance to optimize text evaluation pipelines.
- LIBERO-Para: A Diagnostic Benchmark and Metrics for Paraphrase Robustness in VLA Models (viability: 7): https://sciencetostartup.com/paper/libero-para-a-diagnostic-benchmark-and-metrics-for-paraphrase-robustness-in-vla-models - A benchmark and metric to improve the robustness of VLA models to paraphrased instructions, addressing a key limitation in robotic manipulation.
- NeiGAD: Augmenting Graph Anomaly Detection via Spectral Neighbor Information (viability: 7): https://sciencetostartup.com/paper/neigad-augmenting-graph-anomaly-detection-via-spectral-neighbor-information - A plug-and-play module that enhances graph anomaly detection by explicitly modeling neighbor information using spectral analysis, outperforming state-of-the-art methods.
- DinoDental: Benchmarking DINOv3 as a Unified Vision Encoder for Dental Image Analysis (viability: 7): https://sciencetostartup.com/paper/dinodental-benchmarking-dinov3-as-a-unified-vision-encoder-for-dental-image-analysis - Benchmarking a powerful vision encoder for dental image analysis to reduce annotation costs and improve diagnostic accuracy.
- Evaluating LLMs for Answering Student Questions in Introductory Programming Courses (viability: 7): https://sciencetostartup.com/paper/evaluating-llms-for-answering-student-questions-in-introductory-programming-courses - An LLM-powered system that provides pedagogical hints for student programming questions, validated to outperform typical educator responses and designed for teacher-in-the-loop implementation.
- OptINC: Optical In-Network-Computing for Scalable Distributed Learning (viability: 7): https://sciencetostartup.com/paper/optinc-optical-in-network-computing-for-scalable-distributed-learning - OptINC offloads distributed learning computations, like gradient averaging and quantization, directly into optical interconnects to eliminate communication overhead and accelerate large model training.
- FI-KAN: Fractal Interpolation Kolmogorov-Arnold Networks (viability: 7): https://sciencetostartup.com/paper/fi-kan-fractal-interpolation-kolmogorov-arnold-networks - FI-KAN introduces fractal interpolation bases into Kolmogorov-Arnold Networks, significantly outperforming KAN on benchmarks requiring multi-scale decomposition and adaptability to target regularity.
- TerraSky3D: Multi-View Reconstructions of European Landmarks in 4K (viability: 5): https://sciencetostartup.com/paper/terrasky3d-multi-view-reconstructions-of-european-landmarks-in-4k - A new high-resolution dataset of European landmarks for training and evaluating 3D reconstruction pipelines.
- Pre-Deployment Complexity Estimation for Federated Perception Systems (viability: 5): https://sciencetostartup.com/paper/pre-deployment-complexity-estimation-for-federated-perception-systems - A framework to estimate the complexity of federated learning tasks for edge AI perception systems, enabling better resource planning and feasibility evaluation.
- Corruption-robust Offline Multi-agent Reinforcement Learning From Human Feedback (viability: 3): https://sciencetostartup.com/paper/corruption-robust-offline-multi-agent-reinforcement-learning-from-human-feedback - Developing robust algorithms for multi-agent reinforcement learning in the presence of corrupted human feedback data.
- Point of View: How Perspective Affects Perceived Robot Sociability (viability: 5): https://sciencetostartup.com/paper/point-of-view-how-perspective-affects-perceived-robot-sociability - This research quantifies how robot navigation perspective impacts human perception of sociability and disturbance, suggesting communication gestures can improve comfort.
- osmAG-Nav: A Hierarchical Semantic Topometric Navigation Stack for Robust Lifelong Indoor Autonomy (viability: 7): https://sciencetostartup.com/paper/osmag-nav-a-hierarchical-semantic-topometric-navigation-stack-for-robust-lifelong-indoor-autonomy - A ROS2 navigation stack using hierarchical semantic maps for scalable, low-latency indoor robot autonomy.
- Merge and Conquer: Instructing Multilingual Models by Adding Target Language Weights (viability: 5): https://sciencetostartup.com/paper/merge-and-conquer-instructing-multilingual-models-by-adding-target-language-weights - Adapt multilingual LLMs to low-resource languages efficiently by merging models, bypassing costly fine-tuning and instruction data requirements.
- Coconstructions in spoken data: UD annotation guidelines and first results (viability: 2): https://sciencetostartup.com/paper/coconstructions-in-spoken-data-ud-annotation-guidelines-and-first-results - Develops annotation guidelines for complex syntactic dependencies in spoken language to improve natural language understanding systems.
- Categorical Perception in Large Language Model Hidden States: Structural Warping at Digit-Count Boundaries (viability: 7): https://sciencetostartup.com/paper/categorical-perception-in-large-language-model-hidden-states-structural-warping-at-digit-count-boundaries - This research reveals how LLMs exhibit categorical perception in their internal representations of numbers, offering a new lens for understanding and potentially controlling model behavior.
- MuonEq: Balancing Before Orthogonalization with Lightweight Equilibration (viability: 4): https://sciencetostartup.com/paper/muoneq-balancing-before-orthogonalization-with-lightweight-equilibration - A lightweight pre-orthogonalization technique for optimizers that accelerates LLM pretraining and reduces perplexity.
- MR-ImagenTime: Multi-Resolution Time Series Generation through Dual Image Representations (viability: 7): https://sciencetostartup.com/paper/mr-imagentime-multi-resolution-time-series-generation-through-dual-image-representations - A diffusion-based framework for multi-resolution time series generation that overcomes limitations of fixed-length inputs and inadequate multi-scale modeling.
- Secret Key Rate Analysis of RIS-Assisted THz MIMO CV-QKD Systems under Localized and Global Eavesdropping (viability: 3): https://sciencetostartup.com/paper/secret-key-rate-analysis-of-ris-assisted-thz-mimo-cv-qkd-systems-under-localized-and-global-eavesdropping - Optimizing RIS phase configurations to enhance secret key rates in THz MIMO CV-QKD systems for secure wireless networks.
- DiffAttn: Diffusion-Based Drivers' Visual Attention Prediction with LLM-Enhanced Semantic Reasoning (viability: 7): https://sciencetostartup.com/paper/diffattn-diffusion-based-drivers-visual-attention-prediction-with-llm-enhanced-semantic-reasoning - A diffusion-based framework enhanced by LLMs for predicting drivers' visual attention to improve in-cabin safety and human-machine interaction in intelligent vehicles.
- Reasoning as Energy Minimization over Structured Latent Trajectories (viability: 4): https://sciencetostartup.com/paper/reasoning-as-energy-minimization-over-structured-latent-trajectories - A novel energy-based model for structured latent trajectory planning in reasoning tasks, with code available.
- Cost-Matching Model Predictive Control for Efficient Reinforcement Learning in Humanoid Locomotion (viability: 5): https://sciencetostartup.com/paper/cost-matching-model-predictive-control-for-efficient-reinforcement-learning-in-humanoid-locomotion - A cost-matching reinforcement learning framework for efficient humanoid locomotion control.
- Off-Axis Compliant RCM Joint with Near-Isotropic Stiffness and Minimal Parasitic Error (viability: 5): https://sciencetostartup.com/paper/off-axis-compliant-rcm-joint-with-near-isotropic-stiffness-and-minimal-parasitic-error - A novel, fabrication-ready compliant RCM joint for neuroendoscopic manipulation offering near-isotropic stiffness and minimal parasitic error.
- TwinMixing: A Shuffle-Aware Feature Interaction Model for Multi-Task Segmentation (viability: 7): https://sciencetostartup.com/paper/twinmixing-a-shuffle-aware-feature-interaction-model-for-multi-task-segmentation - A lightweight multi-task segmentation model for autonomous driving that achieves state-of-the-art accuracy with real-time performance on low-cost hardware.
- Detecting the Unexpected: AI-Driven Anomaly Detection in Smart Bridge Monitoring (viability: 5): https://sciencetostartup.com/paper/detecting-the-unexpected-ai-driven-anomaly-detection-in-smart-bridge-monitoring - An AI-driven anomaly detection system for smart bridge monitoring that uses real-time sensor data to detect unforeseen incidents and enhance public safety.
- Ghost-FWL: A Large-Scale Full-Waveform LiDAR Dataset for Ghost Detection and Removal (viability: 8): https://sciencetostartup.com/paper/ghost-fwl-a-large-scale-full-waveform-lidar-dataset-for-ghost-detection-and-removal - A new large-scale dataset and self-supervised learning method for accurate ghost point removal in full-waveform LiDAR, significantly improving autonomous driving perception tasks.
- Variational Neurons in Transformers for Language Modeling (viability: 3): https://sciencetostartup.com/paper/variational-neurons-in-transformers-for-language-modeling - This paper introduces variational neurons into Transformer feed-forward computation to integrate uncertainty into internal computations for more informative language modeling.
- \textit{Versteasch du mi?} Computational and Socio-Linguistic Perspectives on GenAI, LLMs, and Non-Standard Language (viability: 4): https://sciencetostartup.com/paper/textit-versteasch-du-mi-computational-and-socio-linguistic-perspectives-on-genai-llms-and-non-standard-language - Developing LLMs that can understand and process non-standard dialects to bridge the digital language divide.
- Explaining CLIP Zero-shot Predictions Through Concepts (viability: 7): https://sciencetostartup.com/paper/explaining-clip-zero-shot-predictions-through-concepts - Explain CLIP's zero-shot image recognition predictions using human-understandable concepts without additional supervision.
- Beyond Cosine Similarity: Zero-Initialized Residual Complex Projection for Aspect-Based Sentiment Analysis (viability: 7): https://sciencetostartup.com/paper/beyond-cosine-similarity-zero-initialized-residual-complex-projection-for-aspect-based-sentiment-analysis - A novel framework for Aspect-Based Sentiment Analysis that disentangles sentiment and aspect semantics using complex projections, achieving state-of-the-art performance.
- ERPO: Token-Level Entropy-Regulated Policy Optimization for Large Reasoning Models (viability: 7): https://sciencetostartup.com/paper/erpo-token-level-entropy-regulated-policy-optimization-for-large-reasoning-models - A novel reinforcement learning approach that optimizes token-level decision points in large language models to improve reasoning accuracy and conciseness.
- Differentiable Power-Flow Optimization (viability: 7): https://sciencetostartup.com/paper/differentiable-power-flow-optimization - A differentiable simulation approach for scalable power grid management, enabling efficient optimization of renewable energy integration.
- A Perturbation Approach to Unconstrained Linear Bandits (viability: 2): https://sciencetostartup.com/paper/a-perturbation-approach-to-unconstrained-linear-bandits - A theoretical framework for improving bandit linear optimization algorithms through perturbation methods, offering new regret guarantees.
- A Deep Reinforcement Learning Framework for Closed-loop Guidance of Fish Schools via Virtual Agents (viability: 7): https://sciencetostartup.com/paper/a-deep-reinforcement-learning-framework-for-closed-loop-guidance-of-fish-schools-via-virtual-agents - A deep reinforcement learning framework for guiding fish schools using virtual agents, with demonstrated success in physical experiments.
- Policy-Controlled Generalized Share: A General Framework with a Transformer Instantiation for Strictly Online Switching-Oracle Tracking (viability: 7): https://sciencetostartup.com/paper/policy-controlled-generalized-share-a-general-framework-with-a-transformer-instantiation-for-strictly-online-switching-o - A novel online learning framework with a Transformer controller that adaptively tracks switching experts, outperforming existing methods on dynamic regret benchmarks.
- EpiPersona: Persona Projection and Episode Coupling for Pluralistic Preference Modeling (viability: 4): https://sciencetostartup.com/paper/epipersona-persona-projection-and-episode-coupling-for-pluralistic-preference-modeling - A framework for adapting LLMs to diverse user preferences by separating stable personal traits from situational factors.
- Designing AI for Real Users -- Accessibility Gaps in Retail AI Front-End (viability: 3): https://sciencetostartup.com/paper/designing-ai-for-real-users-accessibility-gaps-in-retail-ai-front-end - This paper critiques the accessibility gaps in retail AI front-ends, arguing for front-end assurance to align AI capabilities with diverse user needs.
- DongYuan: An LLM-Based Framework for Integrative Chinese and Western Medicine Spleen-Stomach Disorders Diagnosis (viability: 7): https://sciencetostartup.com/paper/dongyuan-an-llm-based-framework-for-integrative-chinese-and-western-medicine-spleen-stomach-disorders-diagnosis - An LLM-based framework for integrative Chinese and Western medicine diagnosis of spleen-stomach disorders, offering curated datasets, a novel diagnostic LLM, and a comprehensive evaluation benchmark.
- PReD: An LLM-based Foundation Multimodal Model for Electromagnetic Perception, Recognition, and Decision (viability: 7): https://sciencetostartup.com/paper/pred-an-llm-based-foundation-multimodal-model-for-electromagnetic-perception-recognition-and-decision - A foundation model for electromagnetic signal perception, recognition, and decision-making, leveraging multimodal LLM capabilities with a new dataset and benchmark.
- A Closer Look at Cross-Domain Few-Shot Object Detection: Fine-Tuning Matters and Parallel Decoder Helps (viability: 7): https://sciencetostartup.com/paper/a-closer-look-at-cross-domain-few-shot-object-detection-fine-tuning-matters-and-parallel-decoder-helps - A novel object detection method that significantly improves few-shot generalization by using a parallel decoder and progressive fine-tuning, with code available.
- ToLL: Topological Layout Learning with Structural Multi-view Augmentation for 3D Scene Graph Pretraining (viability: 7): https://sciencetostartup.com/paper/toll-topological-layout-learning-with-structural-multi-view-augmentation-for-3d-scene-graph-pretraining - A framework for self-supervised pretraining of 3D scene graphs using topological layout learning and structural augmentation to improve representation quality and overcome data scarcity.
- Skillful Kilometer-Scale Regional Weather Forecasting via Global and Regional Coupling (viability: 7): https://sciencetostartup.com/paper/skillful-kilometer-scale-regional-weather-forecasting-via-global-and-regional-coupling - A coupled global-regional AI model for kilometer-scale weather forecasting that captures fine-grained meteorological phenomena.
- Automating Early Disease Prediction Via Structured and Unstructured Clinical Data (viability: 7): https://sciencetostartup.com/paper/automating-early-disease-prediction-via-structured-and-unstructured-clinical-data - Automate early disease prediction by extracting crucial clinical insights from unstructured discharge reports to enrich existing EHR data.
- Evaluating Privilege Usage of Agents on Real-World Tools (viability: 7): https://sciencetostartup.com/paper/evaluating-privilege-usage-of-agents-on-real-world-tools - A security sandbox for evaluating LLM agent privilege usage with real-world tools to prevent information leakage and infrastructure damage.
- From Reviews to Requirements: Can LLMs Generate Human-Like User Stories? (viability: 7): https://sciencetostartup.com/paper/from-reviews-to-requirements-can-llms-generate-human-like-user-stories - Leveraging LLMs to automatically transform messy app store reviews into actionable, backlog-ready user stories for agile development.
- ColorFLUX: A Structure-Color Decoupling Framework for Old Photo Colorization (viability: 7): https://sciencetostartup.com/paper/colorflux-a-structure-color-decoupling-framework-for-old-photo-colorization - A new framework for accurate old photo colorization that decouples structure and color using diffusion models and visual prompts, outperforming existing methods.
- Event-Based Method for High-Speed 3D Deformation Measurement under Extreme Illumination Conditions (viability: 5): https://sciencetostartup.com/paper/event-based-method-for-high-speed-3d-deformation-measurement-under-extreme-illumination-conditions - A novel event-based vision system for high-speed, accurate 3D deformation measurement of large structures under extreme lighting.
- Reducing Mental Workload through On-Demand Human Assistance for Physical Action Failures in LLM-based Multi-Robot Coordination (viability: 7): https://sciencetostartup.com/paper/reducing-mental-workload-through-on-demand-human-assistance-for-physical-action-failures-in-llm-based-multi-robot-coordi - A human-in-the-loop framework that integrates remote error resolution into LLM-based multi-robot planning to overcome physical execution failures.
- ObjectMorpher: 3D-Aware Image Editing via Deformable 3DGS Models (viability: 7): https://sciencetostartup.com/paper/objectmorpher-3d-aware-image-editing-via-deformable-3dgs-models - ObjectMorpher enables precise, 3D-aware image editing by converting 2D edits into geometry-grounded manipulations using deformable 3D Gaussian Splatting.
- BlankSkip: Early-exit Object Detection onboard Nano-drones (viability: 7): https://sciencetostartup.com/paper/blankskip-early-exit-object-detection-onboard-nano-drones - Enables real-time object detection on ultra-low-power nano-drones by adaptively skipping computation for frames without objects of interest.
- Silent Guardians: Independent and Secure Decision Tree Evaluation Without Chatter (viability: 3): https://sciencetostartup.com/paper/silent-guardians-independent-and-secure-decision-tree-evaluation-without-chatter - A novel protocol for private and verifiable outsourced decision tree evaluation in MLaaS, addressing privacy and correctness challenges without server-to-server communication.
- RecycleLoRA: Rank-Revealing QR-Based Dual-LoRA Subspace Adaptation for Domain Generalized Semantic Segmentation (viability: 4): https://sciencetostartup.com/paper/recyclelora-rank-revealing-qr-based-dual-lora-subspace-adaptation-for-domain-generalized-semantic-segmentation - A novel method for domain generalized semantic segmentation that leverages VFM subspace structures and enhances LoRA for improved generalization.
- Intelligent Road Condition Monitoring using 3D In-Air SONAR Sensing (viability: 5): https://sciencetostartup.com/paper/intelligent-road-condition-monitoring-using-3d-in-air-sonar-sensing - Leveraging 3D SONAR sensors for robust road condition monitoring, overcoming limitations of traditional sensors in adverse weather.
- CoT2-Meta: Budgeted Metacognitive Control for Test-Time Reasoning (viability: 7): https://sciencetostartup.com/paper/cot2-meta-budgeted-metacognitive-control-for-test-time-reasoning - A training-free metacognitive framework that intelligently controls and optimizes test-time reasoning for improved accuracy and compute efficiency across diverse benchmarks.
- Robust Remote Sensing Image-Text Retrieval with Noisy Correspondence (viability: 7): https://sciencetostartup.com/paper/robust-remote-sensing-image-text-retrieval-with-noisy-correspondence - A novel paradigm for remote sensing image-text retrieval that robustly handles noisy data by mimicking human learning patterns, outperforming state-of-the-art methods.
- MDPBench: A Benchmark for Multilingual Document Parsing in Real-World Scenarios (viability: 7): https://sciencetostartup.com/paper/mdpbench-a-benchmark-for-multilingual-document-parsing-in-real-world-scenarios - A benchmark and evaluation of multilingual document parsing models reveals significant performance gaps, highlighting opportunities for more inclusive and deployment-ready parsing systems.
- A Position Statement on Endovascular Models and Effectiveness Metrics for Mechanical Thrombectomy Navigation, on behalf of the Stakeholder Taskforce for AI-assisted Robotic Thrombectomy (START) (viability: 2): https://sciencetostartup.com/paper/a-position-statement-on-endovascular-models-and-effectiveness-metrics-for-mechanical-thrombectomy-navigation-on-behalf-o - This paper proposes consensus frameworks for developing and validating AI-assisted robotic systems for mechanical thrombectomy, focusing on standardizing effectiveness metrics and defining testbeds.
- ORACAL: A Robust and Explainable Multimodal Framework for Smart Contract Vulnerability Detection with Causal Graph Enrichment (viability: 8): https://sciencetostartup.com/paper/oracal-a-robust-and-explainable-multimodal-framework-for-smart-contract-vulnerability-detection-with-causal-graph-enrich - ORACAL is a multimodal graph learning framework that uses RAG and LLMs to detect smart contract vulnerabilities with explainable causal reasoning, significantly outperforming existing methods.
- SVGS: Single-View to 3D Object Editing via Gaussian Splatting (viability: 8): https://sciencetostartup.com/paper/svgs-single-view-to-3d-object-editing-via-gaussian-splatting - A novel single-view 3D object editing technique using Gaussian Splatting that significantly improves editing efficiency and consistency over multi-view methods.
- Does Claude's Constitution Have a Culture? (viability: 3): https://sciencetostartup.com/paper/does-claude-s-constitution-have-a-culture - This paper investigates cultural biases in AI alignment, finding that current methods may codify existing biases rather than correct them, suggesting a need for globally representative constitution-authoring processes.
- MedLoc-R1: Performance-Aware Curriculum Reward Scheduling for GRPO-Based Medical Visual Grounding (viability: 7): https://sciencetostartup.com/paper/medloc-r1-performance-aware-curriculum-reward-scheduling-for-grpo-based-medical-visual-grounding - A performance-aware reward scheduling framework for reinforcement learning-based medical visual grounding that improves accuracy and training stability.
- Neural Federated Learning for Livestock Growth Prediction (viability: 7): https://sciencetostartup.com/paper/neural-federated-learning-for-livestock-growth-prediction - A federated learning framework for livestock growth prediction that enables collaborative model training across farms while preserving data privacy.
- $AutoDrive\text{-}P^3$: Unified Chain of Perception-Prediction-Planning Thought via Reinforcement Fine-Tuning (viability: 8): https://sciencetostartup.com/paper/autodrive-text-p-3-unified-chain-of-perception-prediction-planning-thought-via-reinforcement-fine-tuning - A unified framework for autonomous driving that integrates perception, prediction, and planning through chain-of-thought reasoning and reinforcement learning, achieving state-of-the-art performance.
- Graph Vector Field: A Unified Framework for Multimodal Health Risk Assessment from Heterogeneous Wearable and Environmental Data Streams (viability: 3): https://sciencetostartup.com/paper/graph-vector-field-a-unified-framework-for-multimodal-health-risk-assessment-from-heterogeneous-wearable-and-environment - A novel mathematical framework for multimodal health risk assessment using graph vector fields and higher-order topology.
- Attention Frequency Modulation: Training-Free Spectral Modulation of Diffusion Cross-Attention (viability: 5): https://sciencetostartup.com/paper/attention-frequency-modulation-training-free-spectral-modulation-of-diffusion-cross-attention - A plug-and-play inference-time intervention for diffusion models that manipulates attention spectra in the Fourier domain to control visual output without retraining.
- Lipschitz verification of neural networks through training (viability: 3): https://sciencetostartup.com/paper/lipschitz-verification-of-neural-networks-through-training - Develops a novel training methodology to make neural networks inherently verifiable by penalizing their Lipschitz constant, leading to tighter bounds and improved robustness.
- Contour-Guided Query-Based Feature Fusion for Boundary-Aware and Generalizable Cardiac Ultrasound Segmentation (viability: 7): https://sciencetostartup.com/paper/contour-guided-query-based-feature-fusion-for-boundary-aware-and-generalizable-cardiac-ultrasound-segmentation - A novel network for precise cardiac ultrasound segmentation that leverages contour information to improve boundary accuracy and generalization across different datasets.
- Quid est VERITAS? A Modular Framework for Archival Document Analysis (viability: 7): https://sciencetostartup.com/paper/quid-est-veritas-a-modular-framework-for-archival-document-analysis - A modular framework for digitizing and analyzing archival documents, significantly reducing errors and processing time while enabling advanced historical inquiry.
- RAWIC: Bit-Depth Adaptive Lossless Raw Image Compression (viability: 7): https://sciencetostartup.com/paper/rawic-bit-depth-adaptive-lossless-raw-image-compression - A bit-depth adaptive learned lossless compression framework for raw images that significantly reduces file sizes compared to existing codecs.
- Transcription and Recognition of Italian Parliamentary Speeches Using Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/transcription-and-recognition-of-italian-parliamentary-speeches-using-vision-language-models - Automate the transcription, semantic understanding, and speaker identification of historical parliamentary documents using a vision-language model pipeline.
- Heddle: A Distributed Orchestration System for Agentic RL Rollout (viability: 7): https://sciencetostartup.com/paper/heddle-a-distributed-orchestration-system-for-agentic-rl-rollout - Heddle is a distributed system that optimizes agentic reinforcement learning rollouts by intelligently scheduling and managing tool calls, achieving up to 2.5x higher throughput.
- Octree-based Learned Point Cloud Geometry Compression: A Lossy Perspective (viability: 5): https://sciencetostartup.com/paper/octree-based-learned-point-cloud-geometry-compression-a-lossy-perspective - A novel octree-based approach for lossy point cloud geometry compression that significantly outperforms existing methods on object point clouds and offers effective rate control for LiDAR point clouds.
- InkDrop: Invisible Backdoor Attacks Against Dataset Condensation (viability: 7): https://sciencetostartup.com/paper/inkdrop-invisible-backdoor-attacks-against-dataset-condensation - InkDrop enables stealthy backdoor attacks on dataset condensation, enhancing AI security by embedding malicious patterns imperceptibly.
- SHARP: Short-Window Streaming for Accurate and Robust Prediction in Motion Forecasting (viability: 7): https://sciencetostartup.com/paper/sharp-short-window-streaming-for-accurate-and-robust-prediction-in-motion-forecasting - A novel streaming motion forecasting framework that achieves state-of-the-art accuracy across diverse observation lengths for real-world deployment.
- To View Transform or Not to View Transform: NeRF-based Pre-training Perspective (viability: 7): https://sciencetostartup.com/paper/to-view-transform-or-not-to-view-transform-nerf-based-pre-training-perspective - A novel 3D detector that leverages NeRF pre-training for improved scene reconstruction and detection in autonomous driving.
- GEMS: Agent-Native Multimodal Generation with Memory and Skills (viability: 7): https://sciencetostartup.com/paper/gems-agent-native-multimodal-generation-with-memory-and-skills - GEMS is an agent-native multimodal generation framework that enhances foundational models with structured multi-agent loops, hierarchical memory, and domain-specific skills to achieve significant performance gains on complex and specialized tasks.
- MOSS-VoiceGenerator: Create Realistic Voices with Natural Language Descriptions (viability: 7): https://sciencetostartup.com/paper/moss-voicegenerator-create-realistic-voices-with-natural-language-descriptions - Generate realistic, expressive voices from natural language descriptions for applications like storytelling and game dubbing.
- LogiStory: A Logic-Aware Framework for Multi-Image Story Visualization (viability: 7): https://sciencetostartup.com/paper/logistory-a-logic-aware-framework-for-multi-image-story-visualization - LogiStory is a framework that explicitly models visual logic to generate coherent and communicative multi-image stories, addressing a key limitation in current multimodal systems.
- Control Without Control: Defining Implicit Interaction Paradigms for Autonomous Assistive Robots (viability: 3): https://sciencetostartup.com/paper/control-without-control-defining-implicit-interaction-paradigms-for-autonomous-assistive-robots - Developing assistive robots that adapt to user behavior through implicit cues to enhance control and reduce workload.
- Koopman-based surrogate modeling for reinforcement-learning-control of Rayleigh-Benard convection (viability: 4): https://sciencetostartup.com/paper/koopman-based-surrogate-modeling-for-reinforcement-learning-control-of-rayleigh-benard-convection - Accelerate reinforcement learning control of fluid dynamics systems by using surrogate models trained with policy-aware data, reducing training time by over 40% while maintaining state-of-the-art performance.
- SIMR-NO: A Spectrally-Informed Multi-Resolution Neural Operator for Turbulent Flow Super-Resolution (viability: 7): https://sciencetostartup.com/paper/simr-no-a-spectrally-informed-multi-resolution-neural-operator-for-turbulent-flow-super-resolution - A novel neural operator framework that achieves physically consistent super-resolution of turbulent flow fields from severely under-resolved data, outperforming existing methods in accuracy and spectral fidelity.
- MolmoPoint: Better Pointing for VLMs with Grounding Tokens (viability: 7): https://sciencetostartup.com/paper/molmopoint-better-pointing-for-vlms-with-grounding-tokens - A new pointing mechanism for VLMs that uses special tokens to directly select visual tokens, improving efficiency and state-of-the-art performance on image, GUI, and video pointing tasks.
- AIBench: Evaluating Visual-Logical Consistency in Academic Illustration Generation (viability: 7): https://sciencetostartup.com/paper/aibench-evaluating-visual-logical-consistency-in-academic-illustration-generation - AIBench provides the first benchmark for evaluating the visual-logical consistency of academic illustrations generated by AI, revealing significant gaps in current models and guiding future development.
- From Vessel Trajectories to Safety-Critical Encounter Scenarios: A Generative AI Framework for Autonomous Ship Digital Testing (viability: 7): https://sciencetostartup.com/paper/from-vessel-trajectories-to-safety-critical-encounter-scenarios-a-generative-ai-framework-for-autonomous-ship-digital-te - A generative AI framework that converts ship trajectory data into realistic safety-critical encounter scenarios for autonomous navigation system testing.
- Synonymix: Unified Group Personas for Generative Simulations (viability: 5): https://sciencetostartup.com/paper/synonymix-unified-group-personas-for-generative-simulations - Synonymix enables meso-level agent simulations by constructing queryable group-level representations from individual personas, preserving behavioral signals and privacy.
- \textit{4DSurf}: High-Fidelity Dynamic Scene Surface Reconstruction (viability: 7): https://sciencetostartup.com/paper/textit-4dsurf-high-fidelity-dynamic-scene-surface-reconstruction - A unified framework for high-fidelity dynamic scene surface reconstruction that handles large deformations and temporal inconsistencies, outperforming state-of-the-art methods.
- Reward Hacking as Equilibrium under Finite Evaluation (viability: 3): https://sciencetostartup.com/paper/reward-hacking-as-equilibrium-under-finite-evaluation - This paper theoretically proves that reward hacking is an inherent structural equilibrium in AI systems, not a bug, and proposes a computable index to predict its severity.
- SLOW: Strategic Logical-inference Open Workspace for Cognitive Adaptation in AI Tutoring (viability: 7): https://sciencetostartup.com/paper/slow-strategic-logical-inference-open-workspace-for-cognitive-adaptation-in-ai-tutoring - A theory-informed AI tutoring framework that separates learner state inference from instructional action selection for more personalized and emotionally sensitive instruction.
- Physics-Embedded Feature Learning for AI in Medical Imaging (viability: 7): https://sciencetostartup.com/paper/physics-embedded-feature-learning-for-ai-in-medical-imaging - A physics-embedded deep learning framework for medical imaging that integrates tumor growth dynamics for more interpretable and trustworthy AI.
- Who Wrote the Book? Detecting and Attributing LLM Ghostwriters (viability: 7): https://sciencetostartup.com/paper/who-wrote-the-book-detecting-and-attributing-llm-ghostwriters - A novel, interpretable method for detecting AI-generated text that achieves state-of-the-art performance and works across different LLMs and domains.
- Reducing Oracle Feedback with Vision-Language Embeddings for Preference-Based RL (viability: 7): https://sciencetostartup.com/paper/reducing-oracle-feedback-with-vision-language-embeddings-for-preference-based-rl - A hybrid framework that reduces the cost of learning from human feedback in reinforcement learning by intelligently combining cheap vision-language embeddings with targeted expert queries.
- Meta-Harness: End-to-End Optimization of Model Harnesses (viability: 7): https://sciencetostartup.com/paper/meta-harness-end-to-end-optimization-of-model-harnesses - Automate the creation of LLM harnesses to significantly improve performance and reduce context token usage.
- Object Detection Based on Distributed Convolutional Neural Networks (viability: 3): https://sciencetostartup.com/paper/object-detection-based-on-distributed-convolutional-neural-networks - A novel object detection method based on distributed convolutional neural networks that identifies objects by detecting features across multiple scales.
- Drift-AR: Single-Step Visual Autoregressive Generation via Anti-Symmetric Drifting (viability: 7): https://sciencetostartup.com/paper/drift-ar-single-step-visual-autoregressive-generation-via-anti-symmetric-drifting - Accelerate visual generation by unifying autoregressive and diffusion models with an entropy-driven single-step decoding process.
- Dogfight Search: A Swarm-Based Optimization Algorithm for Complex Engineering Optimization and Mountainous Terrain Path Planning (viability: 4): https://sciencetostartup.com/paper/dogfight-search-a-swarm-based-optimization-algorithm-for-complex-engineering-optimization-and-mountainous-terrain-path-p - A novel swarm-based optimization algorithm inspired by fighter jet tactics, outperforming existing methods on complex engineering and path planning tasks.
- Event6D: Event-based Novel Object 6D Pose Tracking (viability: 8): https://sciencetostartup.com/paper/event6d-event-based-novel-object-6d-pose-tracking - A real-time 6D object pose tracking system for dynamic scenes using event cameras, capable of tracking novel objects without prior training.
- Seeing the Unseen: Rethinking Illicit Promotion Detection with In-Context Learning (viability: 7): https://sciencetostartup.com/paper/seeing-the-unseen-rethinking-illicit-promotion-detection-with-in-context-learning - A new content moderation paradigm using In-Context Learning to detect novel illicit promotions with 22x fewer labels and cross-platform generalization.
- Bit-Identical Medical Deep Learning via Structured Orthogonal Initialization (viability: 7): https://sciencetostartup.com/paper/bit-identical-medical-deep-learning-via-structured-orthogonal-initialization - A framework for bit-identical deep learning training in medical applications, ensuring reproducible and reliable model performance, especially for rare clinical classes.
- Beyond the Answer: Decoding the Behavior of LLMs as Scientific Reasoners (viability: 3): https://sciencetostartup.com/paper/beyond-the-answer-decoding-the-behavior-of-llms-as-scientific-reasoners - This paper explores how to optimize prompts for LLMs to understand their scientific reasoning capabilities, aiming for better interpretability and collaboration with future AI systems.
- Diffusion Maps is not Dimensionality Reduction (viability: 2): https://sciencetostartup.com/paper/diffusion-maps-is-not-dimensionality-reduction - This paper clarifies the theoretical distinction between Diffusion Maps and traditional dimensionality reduction techniques, showing Diffusion Maps require further processing to accurately represent intrinsic geometry.
- Transfer Learning for an Endangered Slavic Variety: Dependency Parsing in Pomak Across Contact-Shaped Dialects (viability: 4): https://sciencetostartup.com/paper/transfer-learning-for-an-endangered-slavic-variety-dependency-parsing-in-pomak-across-contact-shaped-dialects - Developing dependency parsing for an endangered Slavic language by leveraging and adapting existing linguistic resources.
- CARLA-Air: Fly Drones Inside a CARLA World -- A Unified Infrastructure for Air-Ground Embodied Intelligence (viability: 7): https://sciencetostartup.com/paper/carla-air-fly-drones-inside-a-carla-world-a-unified-infrastructure-for-air-ground-embodied-intelligence - CARLA-Air provides a unified, high-fidelity simulation environment for joint aerial and ground embodied intelligence research and development.
- Effort-Based Criticality Metrics for Evaluating 3D Perception Errors in Autonomous Driving (viability: 4): https://sciencetostartup.com/paper/effort-based-criticality-metrics-for-evaluating-3d-perception-errors-in-autonomous-driving - Develops novel metrics to quantify the criticality of perception errors in autonomous driving, enabling targeted improvement of safety systems.
- Efficient Domain Adaptation for Text Line Recognition via Decoupled Language Models (viability: 7): https://sciencetostartup.com/paper/efficient-domain-adaptation-for-text-line-recognition-via-decoupled-language-models - Enables highly accurate, annotation-free domain adaptation for OCR with 95% less compute, making advanced document digitization accessible.
- Adapting SAM to Nuclei Instance Segmentation and Classification via Cooperative Fine-Grained Refinement (viability: 7): https://sciencetostartup.com/paper/adapting-sam-to-nuclei-instance-segmentation-and-classification-via-cooperative-fine-grained-refinement - A parameter-efficient framework to adapt the Segment Anything Model for accurate nuclei instance segmentation in medical imaging.
- When Choices Become Priors: Contrastive Decoding for Scientific Figure Multiple-Choice QA (viability: 5): https://sciencetostartup.com/paper/when-choices-become-priors-contrastive-decoding-for-scientific-figure-multiple-choice-qa - A training-free decoding method that improves scientific figure question answering by explicitly discounting answer choice biases.
- SegRGB-X: General RGB-X Semantic Segmentation Model (viability: 7): https://sciencetostartup.com/paper/segrgb-x-general-rgb-x-semantic-segmentation-model - A universal framework for semantic segmentation across diverse sensor modalities, achieving state-of-the-art performance with modality-specific guidance.
- Physically Inspired Gaussian Splatting for HDR Novel View Synthesis (viability: 7): https://sciencetostartup.com/paper/physically-inspired-gaussian-splatting-for-hdr-novel-view-synthesis - A physically inspired framework for high dynamic range novel view synthesis that accurately reconstructs HDR details and captures illumination-dependent appearance changes.
- What an Autonomous Agent Discovers About Molecular Transformer Design: Does It Transfer? (viability: 7): https://sciencetostartup.com/paper/what-an-autonomous-agent-discovers-about-molecular-transformer-design-does-it-transfer - An autonomous agent discovers transferable molecular transformer architectures, providing a decision framework and toolkit for molecular modeling teams.
- Kill-Chain Canaries: Stage-Level Tracking of Prompt Injection Across Attack Surfaces and Model Safety Tiers (viability: 4): https://sciencetostartup.com/paper/kill-chain-canaries-stage-level-tracking-of-prompt-injection-across-attack-surfaces-and-model-safety-tiers - This research analyzes prompt injection attacks on LLM agents by tracking cryptographic tokens through attack stages to identify defense weaknesses.
- HeteroHub: An Applicable Data Management Framework for Heterogeneous Multi-Embodied Agent System (viability: 7): https://sciencetostartup.com/paper/heterohub-an-applicable-data-management-framework-for-heterogeneous-multi-embodied-agent-system - HeteroHub provides a unified data management framework for coordinating diverse embodied AI agents in complex tasks.
- Energy-Aware Imitation Learning for Steering Prediction Using Events and Frames (viability: 7): https://sciencetostartup.com/paper/energy-aware-imitation-learning-for-steering-prediction-using-events-and-frames - An energy-aware imitation learning framework for autonomous driving steering prediction that fuses event and frame camera data to achieve state-of-the-art performance.
- FedDES: Graph-Based Dynamic Ensemble Selection for Personalized Federated Learning (viability: 7): https://sciencetostartup.com/paper/feddes-graph-based-dynamic-ensemble-selection-for-personalized-federated-learning - A decentralized framework for personalized federated learning that dynamically selects and weights peer models at the instance level using graph neural networks to combat negative transfer.
- Rethinking Atomic Decomposition for LLM Judges: A Prompt-Controlled Study of Reference-Grounded QA Evaluation (viability: 4): https://sciencetostartup.com/paper/rethinking-atomic-decomposition-for-llm-judges-a-prompt-controlled-study-of-reference-grounded-qa-evaluation - Develops a more effective prompt-controlled method for evaluating LLM-generated answers by comparing atomic decomposition against a holistic approach.
- DipGuava: Disentangling Personalized Gaussian Features for 3D Head Avatars from Monocular Video (viability: 7): https://sciencetostartup.com/paper/dipguava-disentangling-personalized-gaussian-features-for-3d-head-avatars-from-monocular-video - Generate photorealistic, identity-preserving 3D head avatars from single videos by disentangling appearance into geometry-driven and residual detail components.
- CLIP-AUTT: Test-Time Personalization with Action Unit Prompting for Fine-Grained Video Emotion Recognition (viability: 7): https://sciencetostartup.com/paper/clip-autt-test-time-personalization-with-action-unit-prompting-for-fine-grained-video-emotion-recognition - CLIP-AUTT personalizes video emotion recognition by dynamically adapting action unit prompts for enhanced accuracy.
- From Independent to Correlated Diffusion: Generalized Generative Modeling with Probabilistic Computers (viability: 4): https://sciencetostartup.com/paper/from-independent-to-correlated-diffusion-generalized-generative-modeling-with-probabilistic-computers - A new diffusion model framework that leverages correlated noise for generative tasks, mapping to energy-efficient probabilistic computers.
- UniDA3D: A Unified Domain-Adaptive Framework for Multi-View 3D Object Detection (viability: 7): https://sciencetostartup.com/paper/unida3d-a-unified-domain-adaptive-framework-for-multi-view-3d-object-detection - A unified domain-adaptive framework for robust, all-weather 3D object detection from cameras, outperforming state-of-the-art under adverse conditions.
- Progressive Prompt-Guided Cross-Modal Reasoning for Referring Image Segmentation (viability: 7): https://sciencetostartup.com/paper/progressive-prompt-guided-cross-modal-reasoning-for-referring-image-segmentation - A framework that uses progressive prompt-guided reasoning to accurately segment objects in images based on natural language descriptions.
- ViviDoc: Generating Interactive Documents through Human-Agent Collaboration (viability: 7): https://sciencetostartup.com/paper/vividoc-generating-interactive-documents-through-human-agent-collaboration - ViviDoc automates the creation of interactive documents through a multi-agent pipeline, enhancing user control and engagement.
- Beyond Dataset Distillation: Lossless Dataset Concentration via Diffusion-Assisted Distribution Alignment (viability: 7): https://sciencetostartup.com/paper/beyond-dataset-distillation-lossless-dataset-concentration-via-diffusion-assisted-distribution-alignment - A novel framework for creating highly representative, compact datasets using diffusion models, significantly reducing data volume without performance loss.
- FedFG: Privacy-Preserving and Robust Federated Learning via Flow-Matching Generation (viability: 7): https://sciencetostartup.com/paper/fedfg-privacy-preserving-and-robust-federated-learning-via-flow-matching-generation - A federated learning framework that uses flow-matching generation to protect client privacy and resist poisoning attacks, achieving higher accuracy and security.
- CDH-Bench: A Commonsense-Driven Hallucination Benchmark for Evaluating Visual Fidelity in Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/cdh-bench-a-commonsense-driven-hallucination-benchmark-for-evaluating-visual-fidelity-in-vision-language-models - A new benchmark to diagnose and mitigate commonsense-driven hallucinations in vision-language models, improving their visual fidelity.
- On the Role of Encoder Depth: Pruning Whisper and LoRA Fine-Tuning in SLAM-ASR (viability: 7): https://sciencetostartup.com/paper/on-the-role-of-encoder-depth-pruning-whisper-and-lora-fine-tuning-in-slam-asr - This research demonstrates a method to significantly reduce the parameter count of ASR models by pruning encoder layers and using LoRA fine-tuning, leading to performance improvements and efficiency gains.
- RetinexDualV2: Physically-Grounded Dual Retinex for Generalized UHD Image Restoration (viability: 7): https://sciencetostartup.com/paper/retinexdualv2-physically-grounded-dual-retinex-for-generalized-uhd-image-restoration - A unified, physically-grounded framework for generalized UHD image restoration that achieves state-of-the-art performance across multiple challenging degradation tasks.
- SARL: Label-Free Reinforcement Learning by Rewarding Reasoning Topology (viability: 4): https://sciencetostartup.com/paper/sarl-label-free-reinforcement-learning-by-rewarding-reasoning-topology - A label-free reinforcement learning framework that rewards the structure of reasoning, improving LLM performance on math and open-ended tasks.
- Principal Prototype Analysis on Manifold for Interpretable Reinforcement Learning (viability: 5): https://sciencetostartup.com/paper/principal-prototype-analysis-on-manifold-for-interpretable-reinforcement-learning - Develops an interpretable reinforcement learning system that automatically selects optimal prototypes from data, enhancing explainability without sacrificing performance.
- AffordMatcher: Affordance Learning in 3D Scenes from Visual Signifiers (viability: 7): https://sciencetostartup.com/paper/affordmatcher-affordance-learning-in-3d-scenes-from-visual-signifiers - A new dataset and method for learning object affordances in 3D scenes by matching visual cues between images and point clouds.
- Hg-I2P: Bridging Modalities for Generalizable Image-to-Point-Cloud Registration via Heterogeneous Graphs (viability: 7): https://sciencetostartup.com/paper/hg-i2p-bridging-modalities-for-generalizable-image-to-point-cloud-registration-via-heterogeneous-graphs - A novel heterogeneous graph approach for robust and generalizable image-to-point-cloud registration, outperforming existing methods across diverse benchmarks.
- Learning Multi-View Spatial Reasoning from Cross-View Relations (viability: 7): https://sciencetostartup.com/paper/learning-multi-view-spatial-reasoning-from-cross-view-relations - A new dataset and fine-tuning approach for vision-language models to enable robust multi-view spatial reasoning for embodied AI and robotics.
- ExFusion: Efficient Transformer Training via Multi-Experts Fusion (viability: 4): https://sciencetostartup.com/paper/exfusion-efficient-transformer-training-via-multi-experts-fusion - A novel pre-training approach that fuses multiple experts in Transformers to enhance performance with minimal additional cost during training and deployment.
- Gradient Manipulation in Distributed Stochastic Gradient Descent with Strategic Agents: Truthful Incentives with Convergence Guarantees (viability: 7): https://sciencetostartup.com/paper/gradient-manipulation-in-distributed-stochastic-gradient-descent-with-strategic-agents-truthful-incentives-with-converge - A distributed payment mechanism that incentivizes honest participation in collaborative machine learning, ensuring both truthful behavior and accurate model convergence without a central server.
- Efficient Inference of Large Vision Language Models (viability: 4): https://sciencetostartup.com/paper/efficient-inference-of-large-vision-language-models - A survey of techniques to accelerate the inference of large vision language models by addressing computational bottlenecks.
- MathGen: Revealing the Illusion of Mathematical Competence through Text-to-Image Generation (viability: 4): https://sciencetostartup.com/paper/mathgen-revealing-the-illusion-of-mathematical-competence-through-text-to-image-generation - A benchmark and evaluation framework to expose the limitations of text-to-image models in generating mathematically accurate visual content.
- CARV: A Diagnostic Benchmark for Compositional Analogical Reasoning in Multimodal LLMs (viability: 7): https://sciencetostartup.com/paper/carv-a-diagnostic-benchmark-for-compositional-analogical-reasoning-in-multimodal-llms - A new benchmark and dataset to diagnose and improve compositional analogical reasoning in multimodal LLMs, revealing significant performance gaps compared to human capabilities.
- Symbolic Density Estimation: A Decompositional Approach (viability: 4): https://sciencetostartup.com/paper/symbolic-density-estimation-a-decompositional-approach - A novel framework for Symbolic Density Estimation that decomposes complex distributions into interpretable symbolic expressions.
- Scaling Atomistic Protein Binder Design with Generative Pretraining and Test-Time Compute (viability: 8): https://sciencetostartup.com/paper/scaling-atomistic-protein-binder-design-with-generative-pretraining-and-test-time-compute - A novel generative AI method for atomistic protein binder design that unifies generative modeling and sequence optimization, achieving state-of-the-art results and releasing code, models, and data.
- EnsemJudge: Enhancing Reliability in Chinese LLM-Generated Text Detection through Diverse Model Ensembles (viability: 7): https://sciencetostartup.com/paper/ensemjudge-enhancing-reliability-in-chinese-llm-generated-text-detection-through-diverse-model-ensembles - A robust framework for detecting Chinese LLM-generated text using ensemble voting, outperforming baselines and achieving first place in a major task.
- RehearsalNeRF: Decoupling Intrinsic Neural Fields of Dynamic Illuminations for Scene Editing (viability: 7): https://sciencetostartup.com/paper/rehearsalnerf-decoupling-intrinsic-neural-fields-of-dynamic-illuminations-for-scene-editing - RehearsalNeRF enables dynamic scene editing under changing illumination by disentangling object radiance from lighting effects using novel regularization techniques.
- Flip Stunts on Bicycle Robots using Iterative Motion Imitation (viability: 7): https://sciencetostartup.com/paper/flip-stunts-on-bicycle-robots-using-iterative-motion-imitation - Enabling acrobatic maneuvers on robots through iterative motion imitation, starting with infeasible references.
- JaWildText: A Benchmark for Vision-Language Models on Japanese Scene Text Understanding (viability: 7): https://sciencetostartup.com/paper/jawildtext-a-benchmark-for-vision-language-models-on-japanese-scene-text-understanding - A new benchmark and evaluation code for vision-language models specifically designed to understand complex Japanese scene text, addressing a significant gap in current multilingual datasets.
- Top-down string-to-dependency Neural Machine Translation (viability: 4): https://sciencetostartup.com/paper/top-down-string-to-dependency-neural-machine-translation - A novel syntactic decoder for neural machine translation that improves translation of long and unseen inputs by generating target-language dependency trees.
- A Cross-Scale Decoder with Token Refinement for Off-Road Semantic Segmentation (viability: 4): https://sciencetostartup.com/paper/a-cross-scale-decoder-with-token-refinement-for-off-road-semantic-segmentation - A novel decoder architecture for off-road semantic segmentation that improves boundary coherence and recovers fine structural details by refining tokens across scales and guiding refinement with uncertainty.
- Physics-Guided Transformer (PGT): Physics-Aware Attention Mechanism for PINNs (viability: 7): https://sciencetostartup.com/paper/physics-guided-transformer-pgt-physics-aware-attention-mechanism-for-pinns - A Physics-Guided Transformer that enhances attention mechanisms for better reconstruction of physical fields from sparse data.
- ForestSim: A Synthetic Benchmark for Intelligent Vehicle Perception in Unstructured Forest Environments (viability: 7): https://sciencetostartup.com/paper/forestsim-a-synthetic-benchmark-for-intelligent-vehicle-perception-in-unstructured-forest-environments - A synthetic dataset and code for training intelligent vehicle perception systems in challenging forest environments, enabling autonomous off-road navigation.
- GEAKG: Generative Executable Algorithm Knowledge Graphs (viability: 7): https://sciencetostartup.com/paper/geakg-generative-executable-algorithm-knowledge-graphs - A framework for learning and transferring algorithmic expertise as executable knowledge graphs, enabling zero-shot generalization across diverse problem domains.
- Adversarial Attacks on Multimodal Large Language Models: A Comprehensive Survey (viability: 3): https://sciencetostartup.com/paper/adversarial-attacks-on-multimodal-large-language-models-a-comprehensive-survey - This paper surveys adversarial attacks on multimodal large language models, identifying vulnerabilities and proposing a taxonomy for understanding and mitigating threats.
- FlashSign: Pose-Free Guidance for Efficient Sign Language Video Generation (viability: 7): https://sciencetostartup.com/paper/flashsign-pose-free-guidance-for-efficient-sign-language-video-generation - A pose-free diffusion model for real-time sign language video generation that directly maps text to gestures, accelerating inference with a novel attention mechanism.
- ITQ3_S: High-Fidelity 3-bit LLM Inference via Interleaved Ternary Quantization with Rotation-Domain Smoothing (viability: 7): https://sciencetostartup.com/paper/itq3-s-high-fidelity-3-bit-llm-inference-via-interleaved-ternary-quantization-with-rotation-domain-smoothing - A novel 3-bit LLM quantization method that achieves FP16 fidelity and 1.5x throughput of 4-bit models by using rotation-domain smoothing and optimized CUDA kernels.
- Spatial Orthogonal Refinement for Robust RGB-Event Visual Object Tracking (viability: 7): https://sciencetostartup.com/paper/spatial-orthogonal-refinement-for-robust-rgb-event-visual-object-tracking - A novel RGB-Event fusion tracking framework that uses spatial orthogonal refinement to improve robustness in high-speed motion scenarios.
- Safety Guardrails in the Sky: Realizing Control Barrier Functions on the VISTA F-16 Jet (viability: 7): https://sciencetostartup.com/paper/safety-guardrails-in-the-sky-realizing-control-barrier-functions-on-the-vista-f-16-jet - Runtime assurance for autonomous systems that guarantees dynamic safety by blending operator commands with safe control actions, demonstrated on an F-16 jet.
- GAAMA: Graph Augmented Associative Memory for Agents (viability: 7): https://sciencetostartup.com/paper/gaama-graph-augmented-associative-memory-for-agents - GAAMA provides AI agents with a hierarchical, graph-augmented memory system that captures associative relationships for more coherent and personalized multi-session interactions.
- ATLAS-RTC: Closing the Loop on LLM Agent Output with Token-Level Runtime Control (viability: 3): https://sciencetostartup.com/paper/atlas-rtc-closing-the-loop-on-llm-agent-output-with-token-level-runtime-control - ATLAS-RTC enhances LLM output quality through real-time control mechanisms.
- BINO: Encoder Centric Self Supervised Stereo With Native Pair Input (viability: 3): https://sciencetostartup.com/paper/bino-encoder-centric-self-supervised-stereo-with-native-pair-input - A novel self-supervised stereo vision model that learns cross-view correspondence within a compact encoder, outperforming existing methods in low-resource settings.
- Persistence diagrams of random matrices via Morse theory: universality and a new spectral diagnostic (viability: 2): https://sciencetostartup.com/paper/persistence-diagrams-of-random-matrices-via-morse-theory-universality-and-a-new-spectral-diagnostic - This paper theoretically links matrix eigenvalues to persistence diagrams using Morse theory, proposing a new spectral diagnostic for random matrix ensembles.
- Rényi Entropy: A New Token Pruning Metric for Vision Transformers (viability: 7): https://sciencetostartup.com/paper/r-nyi-entropy-a-new-token-pruning-metric-for-vision-transformers - A novel, training-free metric for efficient token pruning in Vision Transformers and Large Vision-Language Models, significantly improving inference speed without sacrificing accuracy.
- SAGE: Sink-Aware Grounded Decoding for Multimodal Hallucination Mitigation (viability: 7): https://sciencetostartup.com/paper/sage-sink-aware-grounded-decoding-for-multimodal-hallucination-mitigation - A novel decoding framework that mitigates hallucinations in vision-language models by dynamically adjusting attention mechanisms, significantly improving grounding accuracy without retraining.
- Poppy: Polarization-based Plug-and-Play Guidance for Enhancing Monocular Normal Estimation (viability: 7): https://sciencetostartup.com/paper/poppy-polarization-based-plug-and-play-guidance-for-enhancing-monocular-normal-estimation - A training-free framework that enhances monocular normal estimation for challenging surfaces by leveraging polarization cues at test time.
- Article and Comment Frames Shape the Quality of Online Comments (viability: 4): https://sciencetostartup.com/paper/article-and-comment-frames-shape-the-quality-of-online-comments - A system that uses article framing to predict and mitigate unhealthy online comments.
- Spectral Signatures of Data Quality: Eigenvalue Tail Index as a Diagnostic for Label Noise in Neural Networks (viability: 5): https://sciencetostartup.com/paper/spectral-signatures-of-data-quality-eigenvalue-tail-index-as-a-diagnostic-for-label-noise-in-neural-networks - A novel spectral analysis of neural network weights can accurately diagnose label noise and data quality issues, outperforming existing metrics.
- Near-Optimal Primal-Dual Algorithm for Learning Linear Mixture CMDPs with Adversarial Rewards (viability: 3): https://sciencetostartup.com/paper/near-optimal-primal-dual-algorithm-for-learning-linear-mixture-cmdps-with-adversarial-rewards - A theoretical algorithm for safe reinforcement learning in linear mixture CMDPs with adversarial rewards, achieving near-optimal regret bounds.
- Decentralized Proof-of-Location for Content Provenance: Towards Capture-Time Authenticity (viability: 3): https://sciencetostartup.com/paper/decentralized-proof-of-location-for-content-provenance-towards-capture-time-authenticity - A decentralized system for verifying the origin and authenticity of real-world data captured by devices.
- Kernel Dynamics under Path Entropy Maximization (viability: 2): https://sciencetostartup.com/paper/kernel-dynamics-under-path-entropy-maximization - A theoretical framework for understanding kernel dynamics through path entropy maximization, with potential applications in biological niches and scientific paradigms.
- HumMusQA: A Human-written Music Understanding QA Benchmark Dataset (viability: 5): https://sciencetostartup.com/paper/hummusqa-a-human-written-music-understanding-qa-benchmark-dataset - A new benchmark dataset for evaluating Large Audio-Language Models' music understanding capabilities.
- Stability and Sensitivity Analysis of Relative Temporal-Difference Learning: Extended Version (viability: 2): https://sciencetostartup.com/paper/stability-and-sensitivity-analysis-of-relative-temporal-difference-learning-extended-version - This paper provides theoretical stability guarantees for a variant of temporal-difference learning with linear function approximation, analyzing the impact of baseline distributions on convergence and parameter estimation.
- Statistical Guarantees for Distributionally Robust Optimization with Optimal Transport and OT-Regularized Divergences (viability: 3): https://sciencetostartup.com/paper/statistical-guarantees-for-distributionally-robust-optimization-with-optimal-transport-and-ot-regularized-divergences - This paper provides theoretical guarantees for distributionally robust optimization using optimal transport, improving adversarial training robustness.
- Wan-R1: Verifiable-Reinforcement Learning for Video Reasoning (viability: 7): https://sciencetostartup.com/paper/wan-r1-verifiable-reinforcement-learning-for-video-reasoning - This research introduces verifiable reward functions for reinforcement learning in video generation, significantly improving spatial reasoning and multi-step planning capabilities for tasks like maze-solving and robotic navigation.
- Vertical Consensus Inference for High-Dimensional Random Partition (viability: 3): https://sciencetostartup.com/paper/vertical-consensus-inference-for-high-dimensional-random-partition - A new Bayesian framework for clustering high-dimensional data by combining inferences from smaller, vertically-split data shards.
- ImagenWorld: Stress-Testing Image Generation Models with Explainable Human Evaluation on Open-ended Real-World Tasks (viability: 7): https://sciencetostartup.com/paper/imagenworld-stress-testing-image-generation-models-with-explainable-human-evaluation-on-open-ended-real-world-tasks - ImagenWorld provides a comprehensive and explainable benchmark for stress-testing image generation models across diverse real-world tasks and domains, identifying specific failure modes.
- KazByte: Adapting Qwen models to Kazakh via Byte-level Adapter (viability: 5): https://sciencetostartup.com/paper/kazbyte-adapting-qwen-models-to-kazakh-via-byte-level-adapter - Adapting large language models to low-resource languages by bypassing tokenizers with byte-level adapters to reduce compute costs and improve performance.
- CARGO: Carbon-Aware Gossip Orchestration in Smart Shipping (viability: 5): https://sciencetostartup.com/paper/cargo-carbon-aware-gossip-orchestration-in-smart-shipping - A carbon-aware gossip orchestration framework for smart shipping that optimizes collaborative AI learning under unreliable and sensitive maritime networks.
- What can LLMs tell us about the mechanisms behind polarity illusions in humans? Experiments across model scales and training steps (viability: 2): https://sciencetostartup.com/paper/what-can-llms-tell-us-about-the-mechanisms-behind-polarity-illusions-in-humans-experiments-across-model-scales-and-train - Investigating how large language models exhibit human-like polarity illusions to understand sentence processing mechanisms.
- EffiSkill: Agent Skill Based Automated Code Efficiency Optimization (viability: 4): https://sciencetostartup.com/paper/effiskill-agent-skill-based-automated-code-efficiency-optimization - A framework for LLM-based agents to optimize code efficiency by distilling and reusing reusable optimization knowledge.
- Model Capability Dominates: Inference-Time Optimization Lessons from AIMO 3 (viability: 3): https://sciencetostartup.com/paper/model-capability-dominates-inference-time-optimization-lessons-from-aimo-3 - This paper explores methods to improve LLM mathematical reasoning by diversifying prompting strategies, but finds model capability is the dominant factor.
- ProText: A benchmark dataset for measuring (mis)gendering in long-form texts (viability: 5): https://sciencetostartup.com/paper/protext-a-benchmark-dataset-for-measuring-mis-gendering-in-long-form-texts - A new benchmark dataset to measure and mitigate gender bias in large language models' text transformations.
- 3-D Representations for Hyperspectral Flame Tomography (viability: 4): https://sciencetostartup.com/paper/3-d-representations-for-hyperspectral-flame-tomography - A novel voxel-grid representation for hyperspectral flame tomography achieves higher accuracy and efficiency than continuous neural representations for 3D thermochemical reconstruction.
- jaxsgp4: GPU-accelerated mega-constellation propagation with batch parallelism (viability: 7): https://sciencetostartup.com/paper/jaxsgp4-gpu-accelerated-mega-constellation-propagation-with-batch-parallelism - A GPU-accelerated library for massively parallel orbital propagation, enabling real-time collision avoidance for mega-constellations.
- Improving Clinical Diagnosis with Counterfactual Multi-Agent Reasoning (viability: 7): https://sciencetostartup.com/paper/improving-clinical-diagnosis-with-counterfactual-multi-agent-reasoning - A counterfactual multi-agent framework for clinical diagnosis that improves accuracy and interpretability by explicitly testing how evidence supports or weakens competing diagnoses.
- KVSculpt: KV Cache Compression as Distillation (viability: 7): https://sciencetostartup.com/paper/kvsculpt-kv-cache-compression-as-distillation - KVSculpt compresses LLM KV caches by optimizing unconstrained embeddings, significantly reducing KL divergence and enabling more efficient long-context inference.
- Benchmarking Multi-View BEV Object Detection with Mixed Pinhole and Fisheye Cameras (viability: 7): https://sciencetostartup.com/paper/benchmarking-multi-view-bev-object-detection-with-mixed-pinhole-and-fisheye-cameras - A benchmark and adaptation strategies for robust 3D object detection in autonomous vehicles using mixed pinhole and fisheye cameras.
- Towards Context-Aware Image Anonymization with Multi-Agent Reasoning (viability: 7): https://sciencetostartup.com/paper/towards-context-aware-image-anonymization-with-multi-agent-reasoning - An agentic framework for context-aware image anonymization that preserves data sovereignty and meets GDPR requirements.
- RG-TTA: Regime-Guided Meta-Control for Test-Time Adaptation in Streaming Time Series (viability: 7): https://sciencetostartup.com/paper/rg-tta-regime-guided-meta-control-for-test-time-adaptation-in-streaming-time-series - A meta-controller that dynamically adjusts adaptation intensity for streaming time series forecasting models to improve accuracy and speed.
- MuSEAgent: A Multimodal Reasoning Agent with Stateful Experiences (viability: 7): https://sciencetostartup.com/paper/museagent-a-multimodal-reasoning-agent-with-stateful-experiences - A multimodal research agent that leverages stateful experiences for improved decision-making and reasoning.
- Tracking without Seeing: Geospatial Inference using Encrypted Traffic from Distributed Nodes (viability: 7): https://sciencetostartup.com/paper/tracking-without-seeing-geospatial-inference-using-encrypted-traffic-from-distributed-nodes - Infer object location and motion from encrypted wireless video traffic, enabling surveillance and tracking without direct sensor access.
- Conversational Agents and the Understanding of Human Language: Reflections on AI, LLMs, and Cognitive Science (viability: 2): https://sciencetostartup.com/paper/conversational-agents-and-the-understanding-of-human-language-reflections-on-ai-llms-and-cognitive-science - This paper reflects on the theoretical relationship between NLP and human language understanding, concluding that current AI has not advanced our cognitive science insights.
- Probe-to-Grasp Manipulation Using Self-Sensing Pneumatic Variable-Stiffness Joints (viability: 5): https://sciencetostartup.com/paper/probe-to-grasp-manipulation-using-self-sensing-pneumatic-variable-stiffness-joints - A low-cost robotic gripper with self-sensing variable-stiffness joints that estimates object stiffness for improved grasping of deformable items.
- Understanding Teacher Revisions of Large Language Model-Generated Feedback (viability: 5): https://sciencetostartup.com/paper/understanding-teacher-revisions-of-large-language-model-generated-feedback - This research analyzes how teachers revise AI-generated feedback, identifying patterns and building models to predict revisions, aiming to improve AI feedback systems for educational settings.
- Distributed Online Submodular Maximization under Communication Delays: A Simultaneous Decision-Making Approach (viability: 3): https://sciencetostartup.com/paper/distributed-online-submodular-maximization-under-communication-delays-a-simultaneous-decision-making-approach - A distributed algorithm for online submodular maximization that handles communication delays by enabling simultaneous decision-making.
- Engineering Mythology: A Digital-Physical Framework for Culturally-Inspired Public Art (viability: 3): https://sciencetostartup.com/paper/engineering-mythology-a-digital-physical-framework-for-culturally-inspired-public-art - A framework for creating culturally-inspired public art by integrating digital fabrication, distributed artisan collaboration, and computational optimization.
- Diversity Matters: Dataset Diversification and Dual-Branch Network for Generalized AI-Generated Image Detection (viability: 7): https://sciencetostartup.com/paper/diversity-matters-dataset-diversification-and-dual-branch-network-for-generalized-ai-generated-image-detection - A framework for robust AI-generated image detection using diverse datasets and a dual-branch network that captures both semantic and structural cues.
- Towards Emotion Recognition with 3D Pointclouds Obtained from Facial Expression Images (viability: 7): https://sciencetostartup.com/paper/towards-emotion-recognition-with-3d-pointclouds-obtained-from-facial-expression-images - Enabling privacy-aware, continuous facial emotion recognition using 3D pointclouds generated from wearable sensors, overcoming data scarcity with a novel synthetic dataset generation method.
- Which Reconstruction Model Should a Robot Use? Routing Image-to-3D Models for Cost-Aware Robotic Manipulation (viability: 7): https://sciencetostartup.com/paper/which-reconstruction-model-should-a-robot-use-routing-image-to-3d-models-for-cost-aware-robotic-manipulation - A novel routing framework for robots to intelligently select the optimal 3D reconstruction model based on task requirements and cost constraints, improving efficiency and performance.
- Spectral Decomposition of Inverse Dynamics for Fast Exploration in Model-Based Manipulation (viability: 7): https://sciencetostartup.com/paper/spectral-decomposition-of-inverse-dynamics-for-fast-exploration-in-model-based-manipulation - Accelerate long-horizon robotic manipulation planning by using spectral decomposition of inverse dynamics for real-time, complex trajectory generation.
- What-If Explanations Over Time: Counterfactuals for Time Series Classification (viability: 5): https://sciencetostartup.com/paper/what-if-explanations-over-time-counterfactuals-for-time-series-classification - A survey and open-source library for generating 'what-if' explanations for time series classification models.
- Inference-time Trajectory Optimization for Manga Image Editing (viability: 5): https://sciencetostartup.com/paper/inference-time-trajectory-optimization-for-manga-image-editing - Adapt pre-trained image editing models to manga with minimal computational cost at inference time.
- VLA-OPD: Bridging Offline SFT and Online RL for Vision-Language-Action Models via On-Policy Distillation (viability: 7): https://sciencetostartup.com/paper/vla-opd-bridging-offline-sft-and-online-rl-for-vision-language-action-models-via-on-policy-distillation - VLA-OPD improves robotic model training by combining efficient fine-tuning with the robustness of RL using on-policy distillation.
- Detailed Geometry and Appearance from Opportunistic Motion (viability: 7): https://sciencetostartup.com/paper/detailed-geometry-and-appearance-from-opportunistic-motion - Develop high-fidelity 3D reconstruction tools from static cameras using opportunistic object motion.
- Learning to Commit: Generating Organic Pull Requests via Online Repository Memory (viability: 7): https://sciencetostartup.com/paper/learning-to-commit-generating-organic-pull-requests-via-online-repository-memory - A tool that generates organic pull requests by learning from the history of a software repository.
- Weight Tying Biases Token Embeddings Towards the Output Space (viability: 3): https://sciencetostartup.com/paper/weight-tying-biases-token-embeddings-towards-the-output-space - This paper investigates the impact of weight tying on language model embedding spaces, revealing a bias towards output prediction that can hinder performance.
- GaussianGPT: Towards Autoregressive 3D Gaussian Scene Generation (viability: 4): https://sciencetostartup.com/paper/gaussiangpt-towards-autoregressive-3d-gaussian-scene-generation - Autoregressive 3D Gaussian scene generator for flexible and scalable virtual environments.
- Ruka-v2: Tendon Driven Open-Source Dexterous Hand with Wrist and Abduction for Robot Learning (viability: 8): https://sciencetostartup.com/paper/ruka-v2-tendon-driven-open-source-dexterous-hand-with-wrist-and-abduction-for-robot-learning - Ruka-v2 is an affordable open-source humanoid robotic hand with advanced dexterity and wrist mobility enabling efficient robot learning.
- Partial Motion Imitation for Learning Cart Pushing with Legged Manipulators (viability: 7): https://sciencetostartup.com/paper/partial-motion-imitation-for-learning-cart-pushing-with-legged-manipulators - Enables legged robots to push carts effectively by transferring learned locomotion styles to manipulation tasks.
- Zero-Shot Depth from Defocus (viability: 7): https://sciencetostartup.com/paper/zero-shot-depth-from-defocus - A novel Transformer-based architecture for zero-shot depth estimation from focus stacks, achieving significant performance improvements on a new real-world benchmark.
- Tunable Soft Equivariance with Guarantees (viability: 5): https://sciencetostartup.com/paper/tunable-soft-equivariance-with-guarantees - A framework to tune the degree of equivariance in computer vision models, improving performance and reducing error on existing architectures.
- PerceptionComp: A Video Benchmark for Complex Perception-Centric Reasoning (viability: 7): https://sciencetostartup.com/paper/perceptioncomp-a-video-benchmark-for-complex-perception-centric-reasoning - A new benchmark for complex video reasoning that pushes the limits of current multimodal LLMs, creating an opportunity for specialized perception-centric AI solutions.
- Vision2Web: A Hierarchical Benchmark for Visual Website Development with Agent Verification (viability: 7): https://sciencetostartup.com/paper/vision2web-a-hierarchical-benchmark-for-visual-website-development-with-agent-verification - A hierarchical benchmark and agent verification system for evaluating and improving AI agents in visual website development.
- An LP-based Sampling Policy for Multi-Armed Bandits with Side-Observations and Stochastic Availability (viability: 4): https://sciencetostartup.com/paper/an-lp-based-sampling-policy-for-multi-armed-bandits-with-side-observations-and-stochastic-availability - A novel policy for multi-armed bandits that optimizes exploration in dynamic environments with side-observations and stochastic availability.
- Beyond Language: Grounding Referring Expressions with Hand Pointing in Egocentric Vision (viability: 7): https://sciencetostartup.com/paper/beyond-language-grounding-referring-expressions-with-hand-pointing-in-egocentric-vision - A new dataset and framework for egocentric visual grounding that uses hand pointing and language to resolve ambiguity, significantly improving agent comprehension of physical intents.
- Automatic Laplace Collapsed Sampling: Scalable Marginalisation of Latent Parameters via Automatic Differentiation (viability: 4): https://sciencetostartup.com/paper/automatic-laplace-collapsed-sampling-scalable-marginalisation-of-latent-parameters-via-automatic-differentiation - A framework for scalable Bayesian inference using automatic differentiation to efficiently marginalize latent parameters in complex models.
- Make Geometry Matter for Spatial Reasoning (viability: 7): https://sciencetostartup.com/paper/make-geometry-matter-for-spatial-reasoning - A framework that forces vision-language models to actively use geometric information for improved spatial reasoning, outperforming existing methods.
- Drive-Through 3D Vehicle Exterior Reconstruction via Dynamic-Scene SfM and Distortion-Aware Gaussian Splatting (viability: 7): https://sciencetostartup.com/paper/drive-through-3d-vehicle-exterior-reconstruction-via-dynamic-scene-sfm-and-distortion-aware-gaussian-splatting - Generate high-fidelity 3D vehicle models from cluttered drive-throughs using dynamic-scene SfM and distortion-aware Gaussian Splatting.
- Machine Learning Transferability for Malware Detection (viability: 4): https://sciencetostartup.com/paper/machine-learning-transferability-for-malware-detection - Improving malware detection by evaluating data preprocessing techniques for better model transferability across datasets.
- Context-specific Credibility-aware Multimodal Fusion with Conditional Probabilistic Circuits (viability: 7): https://sciencetostartup.com/paper/context-specific-credibility-aware-multimodal-fusion-with-conditional-probabilistic-circuits - A framework for adaptive multimodal fusion that dynamically assesses source reliability to improve accuracy in noisy or conflicting data scenarios.
- Meta-Adaptive Beam Search Planning for Transformer-Based Reinforcement Learning Control of UAVs with Overhead Manipulators under Flight Disturbances (viability: 4): https://sciencetostartup.com/paper/meta-adaptive-beam-search-planning-for-transformer-based-reinforcement-learning-control-of-uavs-with-overhead-manipulato - A meta-adaptive reinforcement learning planner for UAVs with overhead manipulators to improve trajectory tracking under flight disturbances.
- Benchmarking Tabular Foundation Models for Conditional Density Estimation in Regression (viability: 7): https://sciencetostartup.com/paper/benchmarking-tabular-foundation-models-for-conditional-density-estimation-in-regression - Leverage state-of-the-art tabular foundation models for superior conditional density estimation, outperforming existing methods across diverse datasets and offering a strong solution for uncertainty quantification in tabular data.
- Think over Trajectories: Leveraging Video Generation to Reconstruct GPS Trajectories from Cellular Signaling (viability: 7): https://sciencetostartup.com/paper/think-over-trajectories-leveraging-video-generation-to-reconstruct-gps-trajectories-from-cellular-signaling - Reconstructs high-precision GPS trajectories from coarse cellular signaling data by framing it as a map-visual video generation task.
- Hardware-Aware Tensor Networks for Real-Time Quantum-Inspired Anomaly Detection at Particle Colliders (viability: 7): https://sciencetostartup.com/paper/hardware-aware-tensor-networks-for-real-time-quantum-inspired-anomaly-detection-at-particle-colliders - Develops quantum-inspired tensor networks for real-time anomaly detection in particle collider data, deployable on edge hardware.
- Sustainability Is Not Linear: Quantifying Performance, Energy, and Privacy Trade-offs in On-Device Intelligence (viability: 4): https://sciencetostartup.com/paper/sustainability-is-not-linear-quantifying-performance-energy-and-privacy-trade-offs-in-on-device-intelligence - This research quantifies the complex trade-offs between performance, energy, and privacy for running LLMs on edge devices, revealing counter-intuitive findings about quantization and model architecture.
- VGGRPO: Towards World-Consistent Video Generation with 4D Latent Reward (viability: 7): https://sciencetostartup.com/paper/vggrpo-towards-world-consistent-video-generation-with-4d-latent-reward - A latent geometry-guided framework for post-training video diffusion models to achieve world-consistent generation with improved camera stability and geometric coherence.
- From Static to Dynamic: Exploring Self-supervised Image-to-Video Representation Transfer Learning (viability: 7): https://sciencetostartup.com/paper/from-static-to-dynamic-exploring-self-supervised-image-to-video-representation-transfer-learning - A framework that enables efficient transfer of image-based AI models to video tasks by optimizing a lightweight projection layer, improving both object distinction across videos and consistency within videos.
- Characterization and forecasting of national-scale solar power ramp events (viability: 4): https://sciencetostartup.com/paper/characterization-and-forecasting-of-national-scale-solar-power-ramp-events - A system to forecast and mitigate solar power ramp events for grid stability, outperforming existing models but facing challenges in capturing rapid fluctuations.
- PQuantML: A Tool for End-to-End Hardware-aware Model Compression (viability: 5): https://sciencetostartup.com/paper/pquantml-a-tool-for-end-to-end-hardware-aware-model-compression - A library for end-to-end hardware-aware neural network model compression, simplifying pruning and quantization for low-latency deployments.
- Evaluating Interactive 2D Visualization as a Sample Selection Strategy for Biomedical Time-Series Data Annotation (viability: 4): https://sciencetostartup.com/paper/evaluating-interactive-2d-visualization-as-a-sample-selection-strategy-for-biomedical-time-series-data-annotation - A novel interactive visualization tool improves the accuracy and efficiency of annotating complex biomedical time-series data.
- The Limits of Learning from Pictures and Text: Vision-Language Models and Embodied Scene Understanding (viability: 4): https://sciencetostartup.com/paper/the-limits-of-learning-from-pictures-and-text-vision-language-models-and-embodied-scene-understanding - This research reveals limitations in current vision-language models for understanding affordances, suggesting a need for embodied experience beyond text-image data.
- From Synthetic Data to Real Restorations: Diffusion Model for Patient-specific Dental Crown Completion (viability: 7): https://sciencetostartup.com/paper/from-synthetic-data-to-real-restorations-diffusion-model-for-patient-specific-dental-crown-completion - A diffusion model that automatically completes missing tooth crowns for dental restorations, trained on synthetic data.
- EnTaCs: Analyzing the Relationship Between Sentiment and Language Choice in English-Tamil Code-Switching (viability: 4): https://sciencetostartup.com/paper/entacs-analyzing-the-relationship-between-sentiment-and-language-choice-in-english-tamil-code-switching - This research quantifies how sentiment influences language mixing in English-Tamil communication, offering insights for multilingual AI applications.
- MA-Bench: Towards Fine-grained Micro-Action Understanding (viability: 7): https://sciencetostartup.com/paper/ma-bench-towards-fine-grained-micro-action-understanding - A new benchmark and training dataset for fine-grained micro-action understanding in multimodal LLMs, enabling better human behavior analysis.
- Scene Grounding In the Wild (viability: 7): https://sciencetostartup.com/paper/scene-grounding-in-the-wild - A framework for globally consistent 3D scene reconstruction from sparse, in-the-wild imagery by aligning partial reconstructions to semantic features of reference models.
- The Climber's Grip -- Personalized Deep Learning Models for Fear and Muscle Activity in Climbing (viability: 3): https://sciencetostartup.com/paper/the-climber-s-grip-personalized-deep-learning-models-for-fear-and-muscle-activity-in-climbing - Personalized deep learning models to understand the relationship between fear and muscle activity in climbers.
- Evolution-Based Timed Opacity under a Universal Observation Model (viability: 2): https://sciencetostartup.com/paper/evolution-based-timed-opacity-under-a-universal-observation-model - A theoretical framework for unifying definitions of timed opacity in automata theory.
- Generation Is Compression: Zero-Shot Video Coding via Stochastic Rectified Flow (viability: 7): https://sciencetostartup.com/paper/generation-is-compression-zero-shot-video-coding-via-stochastic-rectified-flow - A zero-shot video codec that leverages pretrained generative models for high-quality, flexible bitrate video compression.
- Machine Unlearning under Retain-Forget Entanglement (viability: 7): https://sciencetostartup.com/paper/machine-unlearning-under-retain-forget-entanglement - A novel two-phase optimization framework for machine unlearning that effectively handles entangled retain-forget data, outperforming existing methods in accuracy retention and removal fidelity.
- Beyond Code Snippets: Benchmarking LLMs on Repository-Level Question Answering (viability: 5): https://sciencetostartup.com/paper/beyond-code-snippets-benchmarking-llms-on-repository-level-question-answering - A new dataset and evaluation framework to benchmark LLMs on understanding entire code repositories, revealing limitations in current reasoning capabilities.
- MemBoost: A Memory-Boosted Framework for Cost-Aware LLM Inference (viability: 5): https://sciencetostartup.com/paper/memboost-a-memory-boosted-framework-for-cost-aware-llm-inference - A framework that reduces LLM inference costs by intelligently reusing answers and escalating complex queries to a stronger model.
- When Perplexity Lies: Generation-Focused Distillation of Hybrid Sequence Models (viability: 7): https://sciencetostartup.com/paper/when-perplexity-lies-generation-focused-distillation-of-hybrid-sequence-models - Distill large language models into smaller, faster, and more memory-efficient hybrid architectures using a generation-focused pipeline and novel attention mechanism, achieving significant performance gains with reduced inference costs.
- Sharp Capacity Scaling of Spectral Optimizers in Learning Associative Memory (viability: 3): https://sciencetostartup.com/paper/sharp-capacity-scaling-of-spectral-optimizers-in-learning-associative-memory - This paper theoretically analyzes spectral optimizers for associative memory recall in language models, showing potential capacity advantages over SGD.
- HolisticSemGes: Semantic Grounding of Holistic Co-Speech Gesture Generation with Contrastive Flow-Matching (viability: 7): https://sciencetostartup.com/paper/holisticsemges-semantic-grounding-of-holistic-co-speech-gesture-generation-with-contrastive-flow-matching - A novel AI model generates semantically grounded co-speech gestures by learning from both correct and incorrect audio-text pairings, improving cross-modal consistency and outperforming existing methods.
- Beyond MACs: Hardware Efficient Architecture Design for Vision Backbones (viability: 7): https://sciencetostartup.com/paper/beyond-macs-hardware-efficient-architecture-design-for-vision-backbones - A novel vision backbone family, LowFormer, offers significant speed-ups and superior performance on edge devices by optimizing hardware efficiency beyond traditional MAC metrics.
- A Lyapunov Analysis of Softmax Policy Gradient for Stochastic Bandits (viability: 2): https://sciencetostartup.com/paper/a-lyapunov-analysis-of-softmax-policy-gradient-for-stochastic-bandits - Theoretical analysis of policy gradient for stochastic bandits to improve regret bounds.
- AutoWeather4D: Autonomous Driving Video Weather Conversion via G-Buffer Dual-Pass Editing (viability: 7): https://sciencetostartup.com/paper/autoweather4d-autonomous-driving-video-weather-conversion-via-g-buffer-dual-pass-editing - A feed-forward framework for 3D-aware weather editing in autonomous driving videos, enabling fine-grained physical control without massive datasets.
- Development of a European Union Time-Indexed Reference Dataset for Assessing the Performance of Signal Detection Methods in Pharmacovigilance using a Large Language Model (viability: 7): https://sciencetostartup.com/paper/development-of-a-european-union-time-indexed-reference-dataset-for-assessing-the-performance-of-signal-detection-methods - A time-indexed EU drug safety dataset with regulatory metadata to benchmark AI signal detection methods.
- User Involvement in Robotic Wheelchair Development: A Decade of Limited Progress (viability: 2): https://sciencetostartup.com/paper/user-involvement-in-robotic-wheelchair-development-a-decade-of-limited-progress - This paper analyzes the limited user involvement in robotic wheelchair development over the past decade, highlighting a gap in participatory design methodologies.
- The Multi-AMR Buffer Storage, Retrieval, and Reshuffling Problem: Exact and Heuristic Approaches (viability: 4): https://sciencetostartup.com/paper/the-multi-amr-buffer-storage-retrieval-and-reshuffling-problem-exact-and-heuristic-approaches - Develops a heuristic approach for robotic buffer management in dense production facilities to address labor shortages and operational costs.
- OVI-MAP:Open-Vocabulary Instance-Semantic Mapping (viability: 7): https://sciencetostartup.com/paper/ovi-map-open-vocabulary-instance-semantic-mapping - Enabling autonomous agents to incrementally build open-vocabulary 3D maps in real-time by decoupling instance reconstruction from semantic inference.
- How Open Must Language Models be to Enable Reliable Scientific Inference? (viability: 2): https://sciencetostartup.com/paper/how-open-must-language-models-be-to-enable-reliable-scientific-inference - This paper analyzes the impact of model openness on scientific inference, arguing that closed models are generally ill-suited for research and proposing mitigation strategies.
- Stabilizing Rubric Integration Training via Decoupled Advantage Normalization (viability: 7): https://sciencetostartup.com/paper/stabilizing-rubric-integration-training-via-decoupled-advantage-normalization - A novel training method that improves LLM reasoning quality by decoupling outcome and process rewards, outperforming existing approaches on complex benchmarks.
- Learnable Quantum Efficiency Filters for Urban Hyperspectral Segmentation (viability: 7): https://sciencetostartup.com/paper/learnable-quantum-efficiency-filters-for-urban-hyperspectral-segmentation - A physics-informed dimensionality reduction method for hyperspectral urban driving data that significantly improves semantic segmentation performance and interpretability.
- ALBA: A European Portuguese Benchmark for Evaluating Language and Linguistic Dimensions in Generative LLMs (viability: 4): https://sciencetostartup.com/paper/alba-a-european-portuguese-benchmark-for-evaluating-language-and-linguistic-dimensions-in-generative-llms - A new benchmark and evaluation framework to assess Large Language Model proficiency in European Portuguese across eight linguistic dimensions.
- JAL-Turn: Joint Acoustic-Linguistic Modeling for Real-Time and Robust Turn-Taking Detection in Full-Duplex Spoken Dialogue Systems (viability: 7): https://sciencetostartup.com/paper/jal-turn-joint-acoustic-linguistic-modeling-for-real-time-and-robust-turn-taking-detection-in-full-duplex-spoken-dialogu - A lightweight, speech-only framework for real-time, robust turn-taking detection in voice AI agents, integrating acoustic and linguistic cues without adding latency.
- CADSmith: Multi-Agent CAD Generation with Programmatic Geometric Validation (viability: 7): https://sciencetostartup.com/paper/cadsmith-multi-agent-cad-generation-with-programmatic-geometric-validation - CADSmith is a multi-agent system that generates precise CAD models from natural language using programmatic geometric validation and iterative refinement.
- AMALIA Technical Report: A Fully Open Source Large Language Model for European Portuguese (viability: 7): https://sciencetostartup.com/paper/amalia-technical-report-a-fully-open-source-large-language-model-for-european-portuguese - An open-source LLM specifically trained for European Portuguese, offering superior performance on native tasks and a new suite of pt-PT benchmarks.
- Clinical named entity recognition in the Portuguese language: a benchmark of modern BERT models and LLMs (viability: 7): https://sciencetostartup.com/paper/clinical-named-entity-recognition-in-the-portuguese-language-a-benchmark-of-modern-bert-models-and-llms - A benchmark of modern BERT and LLMs for clinical named entity recognition in Portuguese, demonstrating strong performance with mmBERT and balanced data strategies.
- Conditional Diffusion for 3D CT Volume Reconstruction from 2D X-rays (viability: 8): https://sciencetostartup.com/paper/conditional-diffusion-for-3d-ct-volume-reconstruction-from-2d-x-rays - Revolutionizing diagnostic imaging by reconstructing 3D CT volumes from 2D X-rays, reducing costs and radiation exposure.
- Targeted learning of heterogeneous treatment effect curves for right censored or left truncated time-to-event data (viability: 4): https://sciencetostartup.com/paper/targeted-learning-of-heterogeneous-treatment-effect-curves-for-right-censored-or-left-truncated-time-to-event-data - A new method for estimating subject-specific treatment effects on time-to-event outcomes, improving accuracy and revealing temporal patterns in clinical data.
- AIRA_2: Overcoming Bottlenecks in AI Research Agents (viability: 4): https://sciencetostartup.com/paper/aira-2-overcoming-bottlenecks-in-ai-research-agents - An AI research agent architecture that overcomes performance bottlenecks to achieve higher results on complex benchmarks.
- Rocks, Pebbles and Sand: Modality-aware Scheduling for Multimodal Large Language Model Inference (viability: 7): https://sciencetostartup.com/paper/rocks-pebbles-and-sand-modality-aware-scheduling-for-multimodal-large-language-model-inference - RPS-Serve is a modality-aware scheduler that dramatically reduces MLLM inference latency by prioritizing requests based on their resource demands, ensuring interactive responsiveness.
- Reentrancy Detection in the Age of LLMs (viability: 7): https://sciencetostartup.com/paper/reentrancy-detection-in-the-age-of-llms - Leveraging LLMs to significantly improve reentrancy vulnerability detection in modern Ethereum smart contracts, addressing critical gaps in existing tools.
- ClipTTT: CLIP-Guided Test-Time Training Helps LVLMs See Better (viability: 7): https://sciencetostartup.com/paper/clipttt-clip-guided-test-time-training-helps-lvlms-see-better - A method to adapt large vision-language models on the fly to reduce hallucinations caused by visual input corruption, using CLIP as a guidance signal.
- EcoFair: Trustworthy and Energy-Aware Routing for Privacy-Preserving Vertically Partitioned Medical Inference (viability: 7): https://sciencetostartup.com/paper/ecofair-trustworthy-and-energy-aware-routing-for-privacy-preserving-vertically-partitioned-medical-inference - A privacy-preserving and energy-efficient framework for medical inference that intelligently routes computation based on diagnostic uncertainty and clinical risk.
- SPECTRA: An Efficient Spectral-Informed Neural Network for Sensor-Based Activity Recognition (viability: 7): https://sciencetostartup.com/paper/spectra-an-efficient-spectral-informed-neural-network-for-sensor-based-activity-recognition - SPECTRA is a deployment-first spectral-informed neural network for efficient, real-time, and private sensor-based activity recognition on edge devices.
- SparseCam4D: Spatio-Temporally Consistent 4D Reconstruction from Sparse Cameras (viability: 7): https://sciencetostartup.com/paper/sparsecam4d-spatio-temporally-consistent-4d-reconstruction-from-sparse-cameras - Enables high-quality 4D reconstruction of dynamic scenes using sparse, uncalibrated cameras by modeling spatio-temporal inconsistencies.
- Probabilistic Multilabel Graphical Modelling of Motif Transformations in Symbolic Music (viability: 4): https://sciencetostartup.com/paper/probabilistic-multilabel-graphical-modelling-of-motif-transformations-in-symbolic-music - A probabilistic framework to model and analyze motif transformations in symbolic music, enabling quantitative study of musical structure and compositional patterns.
- Shapley meets Rawls: an integrated framework for measuring and explaining unfairness (viability: 4): https://sciencetostartup.com/paper/shapley-meets-rawls-an-integrated-framework-for-measuring-and-explaining-unfairness - An integrated framework for measuring and explaining unfairness in AI models using Shapley values, offering faster computation than traditional methods.
- Foundation Model for Cardiac Time Series via Masked Latent Attention (viability: 4): https://sciencetostartup.com/paper/foundation-model-for-cardiac-time-series-via-masked-latent-attention - A foundation model for ECG analysis that leverages cross-lead relationships to improve cardiovascular diagnosis.
- UNIFERENCE: A Discrete Event Simulation Framework for Developing Distributed AI Models (viability: 7): https://sciencetostartup.com/paper/uniference-a-discrete-event-simulation-framework-for-developing-distributed-ai-models - A discrete event simulation framework that bridges the gap between simulating and deploying distributed AI models, offering high accuracy and seamless integration with PyTorch Distributed.
- HyVIC: A Metric-Driven Spatio-Spectral Hyperspectral Image Compression Architecture Based on Variational Autoencoders (viability: 7): https://sciencetostartup.com/paper/hyvic-a-metric-driven-spatio-spectral-hyperspectral-image-compression-architecture-based-on-variational-autoencoders - A metric-driven variational autoencoder architecture for hyperspectral image compression that significantly improves reconstruction fidelity by explicitly modeling spatio-spectral redundancies.
- Addressing Ambiguity in Imitation Learning through Product of Experts based Negative Feedback (viability: 7): https://sciencetostartup.com/paper/addressing-ambiguity-in-imitation-learning-through-product-of-experts-based-negative-feedback - A novel imitation learning system that leverages suboptimal human demonstrations and learns from its own failures to significantly improve robot task success rates.
- Adapt as You Say: Online Interactive Bimanual Skill Adaptation via Human Language Feedback (viability: 7): https://sciencetostartup.com/paper/adapt-as-you-say-online-interactive-bimanual-skill-adaptation-via-human-language-feedback - BiSAIL empowers robots to adapt bimanual skills through verbal feedback, enhancing human-robot collaboration.
- A Boltzmann-machine-enhanced Transformer For DNA Sequence Classification (viability: 3): https://sciencetostartup.com/paper/a-boltzmann-machine-enhanced-transformer-for-dna-sequence-classification - A novel Transformer architecture enhanced with Boltzmann machines for improved DNA sequence classification by uncovering latent interactions.
- Automatic feature identification in least-squares policy iteration using the Koopman operator framework (viability: 3): https://sciencetostartup.com/paper/automatic-feature-identification-in-least-squares-policy-iteration-using-the-koopman-operator-framework - A novel reinforcement learning algorithm that automatically learns features for policy iteration, removing the need for manual feature engineering.
- DTP-Attack: A decision-based black-box adversarial attack on trajectory prediction (viability: 7): https://sciencetostartup.com/paper/dtp-attack-a-decision-based-black-box-adversarial-attack-on-trajectory-prediction - A black-box adversarial attack framework for trajectory prediction systems that exploits vulnerabilities in autonomous vehicle safety.
- Neuro-Symbolic Process Anomaly Detection (viability: 5): https://sciencetostartup.com/paper/neuro-symbolic-process-anomaly-detection - A neuro-symbolic AI approach that integrates domain knowledge into process anomaly detection to accurately distinguish between rare but conformant behavior and actual anomalies.
- Can AI Models Direct Each Other? Organizational Structure as a Probe into Training Limitations (viability: 7): https://sciencetostartup.com/paper/can-ai-models-direct-each-other-organizational-structure-as-a-probe-into-training-limitations - A two-agent AI system directs cheaper models to perform complex software engineering tasks more efficiently by leveraging expensive reasoning for task delegation and review.
- Fair Data Pre-Processing with Imperfect Attribute Space (viability: 5): https://sciencetostartup.com/paper/fair-data-pre-processing-with-imperfect-attribute-space - A framework for robust fair data pre-processing in real-world scenarios with imperfect attributes by leveraging latent attributes.
- ClimateCheck 2026: Scientific Fact-Checking and Disinformation Narrative Classification of Climate-related Claims (viability: 4): https://sciencetostartup.com/paper/climatecheck-2026-scientific-fact-checking-and-disinformation-narrative-classification-of-climate-related-claims - A shared task competition focused on developing AI systems for scientific fact-checking and disinformation classification of climate-related claims.
- Meta-Learned Adaptive Optimization for Robust Human Mesh Recovery with Uncertainty-Aware Parameter Updates (viability: 7): https://sciencetostartup.com/paper/meta-learned-adaptive-optimization-for-robust-human-mesh-recovery-with-uncertainty-aware-parameter-updates - A meta-learning framework for robust 3D human mesh recovery from single images, offering state-of-the-art accuracy and uncertainty estimation.
- Image-based Quantification of Postural Deviations on Patients with Cervical Dystonia: A Machine Learning Approach Using Synthetic Training Data (viability: 7): https://sciencetostartup.com/paper/image-based-quantification-of-postural-deviations-on-patients-with-cervical-dystonia-a-machine-learning-approach-using-s - An AI-powered image analysis tool objectively quantifies cervical dystonia severity, overcoming subjective clinical assessments with synthetic data validation.
- 120 Minutes and a Laptop: Minimalist Image-goal Navigation via Unsupervised Exploration and Offline RL (viability: 7): https://sciencetostartup.com/paper/120-minutes-and-a-laptop-minimalist-image-goal-navigation-via-unsupervised-exploration-and-offline-rl - Enables rapid, low-resource image-goal navigation for robots by leveraging unsupervised exploration and offline reinforcement learning.
- Interpretable long-term traffic modelling on national road networks using theory-informed deep learning (viability: 4): https://sciencetostartup.com/paper/interpretable-long-term-traffic-modelling-on-national-road-networks-using-theory-informed-deep-learning - A theory-informed deep learning framework for interpretable long-term highway traffic volume prediction.
- Automating Clinical Information Retrieval from Finnish Electronic Health Records Using Large Language Models (viability: 7): https://sciencetostartup.com/paper/automating-clinical-information-retrieval-from-finnish-electronic-health-records-using-large-language-models - A locally deployable framework using open-source LLMs to accurately retrieve patient-specific information from Finnish EHRs, with a low rate of clinically significant errors.
- Analysing Calls to Order in German Parliamentary Debates (viability: 7): https://sciencetostartup.com/paper/analysing-calls-to-order-in-german-parliamentary-debates - Automate the detection and analysis of norm violations in political debates using a novel rule-based system and a 72-year dataset.
- CPUBone: Efficient Vision Backbone Design for Devices with Low Parallelization Capabilities (viability: 7): https://sciencetostartup.com/paper/cpubone-efficient-vision-backbone-design-for-devices-with-low-parallelization-capabilities - CPUBone offers a new family of vision backbone models optimized for high performance on CPUs, achieving state-of-the-art speed-accuracy trade-offs for edge AI applications.
- Kantorovich--Kernel Neural Operators: Approximation Theory, Asymptotics, and Neural Network Interpretation (viability: 1): https://sciencetostartup.com/paper/kantorovich-kernel-neural-operators-approximation-theory-asymptotics-and-neural-network-interpretation - This paper theoretically analyzes a class of neural network operators, exploring their connection to classical mathematical concepts and approximation theory.
- Towards Privacy-Preserving Federated Learning using Hybrid Homomorphic Encryption (viability: 7): https://sciencetostartup.com/paper/towards-privacy-preserving-federated-learning-using-hybrid-homomorphic-encryption - Enhance federated learning privacy with novel key protection mechanisms that secure against malicious participants without sacrificing model accuracy.
- KMM-CP: Practical Conformal Prediction under Covariate Shift via Selective Kernel Mean Matching (viability: 7): https://sciencetostartup.com/paper/kmm-cp-practical-conformal-prediction-under-covariate-shift-via-selective-kernel-mean-matching - A conformal prediction framework using Kernel Mean Matching to provide reliable uncertainty quantification for machine learning models facing covariate shift, with code available.
- Generalizable task-oriented object grasping through LLM-guided ontology and similarity-based planning (viability: 7): https://sciencetostartup.com/paper/generalizable-task-oriented-object-grasping-through-llm-guided-ontology-and-similarity-based-planning - A geometry-centric approach for generalizable object grasping guided by LLM-generated ontologies and similarity-based planning, overcoming limitations of semantic vision models.
- Why Models Know But Don't Say: Chain-of-Thought Faithfulness Divergence Between Thinking Tokens and Answers in Open-Weight Reasoning Models (viability: 3): https://sciencetostartup.com/paper/why-models-know-but-don-t-say-chain-of-thought-faithfulness-divergence-between-thinking-tokens-and-answers-in-open-weigh - This research analyzes how large language models' internal thinking processes diverge from their final answers, revealing a significant gap in how models acknowledge hints and influencing factors.
- Cryptanalysis of a PIR Scheme based on Linear Codes over Rings (viability: 2): https://sciencetostartup.com/paper/cryptanalysis-of-a-pir-scheme-based-on-linear-codes-over-rings - This paper presents a theoretical attack on a private information retrieval scheme based on linear codes over rings, demonstrating a vulnerability in the server's ability to retrieve file indices.
- Hidden Elo: Private Matchmaking through Encrypted Rating Systems (viability: 5): https://sciencetostartup.com/paper/hidden-elo-private-matchmaking-through-encrypted-rating-systems - A fully homomorphic encryption-based private rating system for secure matchmaking without compromising sensitive user data.
- T-800: An 800 Hz Data Glove for Precise Hand Gesture Tracking (viability: 7): https://sciencetostartup.com/paper/t-800-an-800-hz-data-glove-for-precise-hand-gesture-tracking - A high-bandwidth data glove system for precise, 800 Hz full-hand motion tracking, enabling detailed human gesture capture for robotic control policy training.
- Word Alignment-Based Evaluation of Uniform Meaning Representations (viability: 3): https://sciencetostartup.com/paper/word-alignment-based-evaluation-of-uniform-meaning-representations - A novel algorithm for comparing sentence meaning representations that leverages word alignments for more intuitive and interpretable evaluation.
- SHANDS: A Multi-View Dataset and Benchmark for Surgical Hand-Gesture and Error Recognition Toward Medical Training (viability: 7): https://sciencetostartup.com/paper/shands-a-multi-view-dataset-and-benchmark-for-surgical-hand-gesture-and-error-recognition-toward-medical-training - A new multi-view dataset and benchmark for surgical hand-gesture and error recognition aims to automate medical training assessment, addressing scalability and cost limitations.
- Maintaining Difficulty: A Margin Scheduler for Triplet Loss in Siamese Networks Training (viability: 4): https://sciencetostartup.com/paper/maintaining-difficulty-a-margin-scheduler-for-triplet-loss-in-siamese-networks-training - A novel margin scheduling technique for Siamese networks that improves distance metric learning by dynamically adjusting training difficulty.
- Restore, Assess, Repeat: A Unified Framework for Iterative Image Restoration (viability: 7): https://sciencetostartup.com/paper/restore-assess-repeat-a-unified-framework-for-iterative-image-restoration - A unified framework for iterative image restoration that integrates quality assessment and restoration in the latent domain for improved generalization and efficiency.
- Switch Attention: Towards Dynamic and Fine-grained Hybrid Transformers (viability: 7): https://sciencetostartup.com/paper/switch-attention-towards-dynamic-and-fine-grained-hybrid-transformers - A novel hybrid transformer architecture that dynamically routes computation between full and sliding window attention for efficient long-context language modeling.
- Generative Modeling in Protein Design: Neural Representations, Conditional Generation, and Evaluation Standards (viability: 3): https://sciencetostartup.com/paper/generative-modeling-in-protein-design-neural-representations-conditional-generation-and-evaluation-standards - This paper surveys generative AI methods for protein design, aiming to unify research and establish evaluation standards to accelerate functional protein engineering.
- Dynamic Token Compression for Efficient Video Understanding through Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/dynamic-token-compression-for-efficient-video-understanding-through-reinforcement-learning - A reinforcement learning framework that adaptively compresses video tokens to achieve significant speedups for video understanding tasks without sacrificing performance.
- A Formal Framework for Uncertainty Analysis of Text Generation with Large Language Models (viability: 4): https://sciencetostartup.com/paper/a-formal-framework-for-uncertainty-analysis-of-text-generation-with-large-language-models - A formal framework to measure and analyze uncertainty in text generation from LLMs, considering prompting, generation, and interpretation.
- HandVQA: Diagnosing and Improving Fine-Grained Spatial Reasoning about Hands in Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/handvqa-diagnosing-and-improving-fine-grained-spatial-reasoning-about-hands-in-vision-language-models - A diagnostic benchmark and fine-tuning method to significantly improve vision-language models' spatial reasoning about human hands, enabling applications in robotics and AR/VR.
- Auditing Blockchain Innovations: Technical Challenges Beyond Traditional Finance (viability: 4): https://sciencetostartup.com/paper/auditing-blockchain-innovations-technical-challenges-beyond-traditional-finance - A novel autoethnographic framework for auditing complex cryptoasset transactions and smart contracts, addressing technically insurmountable challenges for auditing firms.
- Realtime-VLA V2: Learning to Run VLAs Fast, Smooth, and Accurate (viability: 7): https://sciencetostartup.com/paper/realtime-vla-v2-learning-to-run-vlas-fast-smooth-and-accurate - Enables VLA-driven robots to operate at human-like speeds with high accuracy through practical deployment techniques.
- MPDiT: Multi-Patch Global-to-Local Transformer Architecture For Efficient Flow Matching and Diffusion Model (viability: 7): https://sciencetostartup.com/paper/mpdit-multi-patch-global-to-local-transformer-architecture-for-efficient-flow-matching-and-diffusion-model - A novel multi-patch transformer architecture for diffusion models that significantly reduces computational cost while maintaining high generative performance.
- From Pen to Pixel: Translating Hand-Drawn Plots into Graphical APIs via a Novel Benchmark and Efficient Adapter (viability: 7): https://sciencetostartup.com/paper/from-pen-to-pixel-translating-hand-drawn-plots-into-graphical-apis-via-a-novel-benchmark-and-efficient-adapter - Generate graphical APIs from hand-drawn plot images using an efficient adapter and a novel dataset, making data visualization accessible to non-experts.
- Only Whats Necessary: Pareto Optimal Data Minimization for Privacy Preserving Video Anomaly Detection (viability: 7): https://sciencetostartup.com/paper/only-whats-necessary-pareto-optimal-data-minimization-for-privacy-preserving-video-anomaly-detection - A privacy-by-design framework for video anomaly detection that minimizes personally identifiable information while preserving detection accuracy.
- DuSCN-FusionNet: An Interpretable Dual-Channel Structural Covariance Fusion Framework for ADHD Classification Using Structural MRI (viability: 7): https://sciencetostartup.com/paper/duscn-fusionnet-an-interpretable-dual-channel-structural-covariance-fusion-framework-for-adhd-classification-using-struc - An interpretable AI framework for ADHD classification using structural MRI, identifying key brain regions as potential biomarkers.
- Generative Score Inference for Multimodal Data (viability: 5): https://sciencetostartup.com/paper/generative-score-inference-for-multimodal-data - A flexible framework for statistically valid uncertainty quantification in multimodal learning using synthetic data from generative models.
- Reflect to Inform: Boosting Multimodal Reasoning via Information-Gain-Driven Verification (viability: 7): https://sciencetostartup.com/paper/reflect-to-inform-boosting-multimodal-reasoning-via-information-gain-driven-verification - A self-evolving training framework that enables multimodal models to autonomously verify visual information during reasoning, reducing hallucinations and improving accuracy.
- Optimal Prioritized Dissipation and Closed-Form Damping Limitation under Actuator Constraints for Haptic Interfaces (viability: 4): https://sciencetostartup.com/paper/optimal-prioritized-dissipation-and-closed-form-damping-limitation-under-actuator-constraints-for-haptic-interfaces - A novel damping method for haptic interfaces that prioritizes actuator load while maintaining stability and transparency.
- A Power-Weighted Noncentral Complex Gaussian Distribution (viability: 3): https://sciencetostartup.com/paper/a-power-weighted-noncentral-complex-gaussian-distribution - A new probabilistic model for complex-valued data that unifies and improves upon existing distributions for signal modeling.
- Hermes Seal: Zero-Knowledge Assurance for Autonomous Vehicle Communications (viability: 7): https://sciencetostartup.com/paper/hermes-seal-zero-knowledge-assurance-for-autonomous-vehicle-communications - A ZKP framework for verifiable and privacy-preserving communication in autonomous vehicles, enabling trust and interoperability without revealing proprietary data.
- HINT: Composed Image Retrieval with Dual-path Compositional Contextualized Network (viability: 7): https://sciencetostartup.com/paper/hint-composed-image-retrieval-with-dual-path-compositional-contextualized-network - A novel network architecture that significantly improves composed image retrieval by incorporating contextual information and amplifying similarity differences, outperforming existing benchmarks.
- Curvature-aware Expected Free Energy as an Acquisition Function for Bayesian Optimization (viability: 4): https://sciencetostartup.com/paper/curvature-aware-expected-free-energy-as-an-acquisition-function-for-bayesian-optimization - A novel acquisition function for Bayesian optimization that simultaneously learns and optimizes underlying functions, outperforming existing methods in simulations.
- From Pixels to Privacy: Temporally Consistent Video Anonymization via Token Pruning for Privacy Preserving Action Recognition (viability: 7): https://sciencetostartup.com/paper/from-pixels-to-privacy-temporally-consistent-video-anonymization-via-token-pruning-for-privacy-preserving-action-recogni - A novel framework for video anonymization that prunes privacy-sensitive information while preserving action recognition accuracy, enabling safer video analytics.
- CALRK-Bench: Evaluating Context-Aware Legal Reasoning in Korean Law (viability: 7): https://sciencetostartup.com/paper/calrk-bench-evaluating-context-aware-legal-reasoning-in-korean-law - A new benchmark for evaluating context-aware legal reasoning in Korean law, revealing significant performance gaps in current LLMs.
- Mitigating the Reasoning Tax in Vision-Language Fine-Tuning with Input-Adaptive Depth Aggregation (viability: 5): https://sciencetostartup.com/paper/mitigating-the-reasoning-tax-in-vision-language-fine-tuning-with-input-adaptive-depth-aggregation - A lightweight mechanism to improve reasoning and perception in vision-language models without significant parameter increases.
- Verify Claimed Text-to-Image Models via Boundary-Aware Prompt Optimization (viability: 7): https://sciencetostartup.com/paper/verify-claimed-text-to-image-models-via-boundary-aware-prompt-optimization - A novel method to verify which text-to-image models are actually being used by third-party platforms, ensuring accurate claims and protecting reputations.
- Making Multi-Axis Models Robust to Multiplicative Noise: How, and Why? (viability: 7): https://sciencetostartup.com/paper/making-multi-axis-models-robust-to-multiplicative-noise-how-and-why - A graph-learning algorithm for robustly fitting multi-axis models corrupted by multiplicative noise, with applications in single-cell RNA sequencing.
- PRISMA: Toward a Normative Information Infrastructure for Responsible Pharmaceutical Knowledge Management (viability: 7): https://sciencetostartup.com/paper/prisma-toward-a-normative-information-infrastructure-for-responsible-pharmaceutical-knowledge-management - A normative information infrastructure for responsible pharmaceutical knowledge management that anchors AI interpretations to primary sources, enhancing transparency and accountability.
- From Human Cognition to Neural Activations: Probing the Computational Primitives of Spatial Reasoning in LLMs (viability: 3): https://sciencetostartup.com/paper/from-human-cognition-to-neural-activations-probing-the-computational-primitives-of-spatial-reasoning-in-llms - This research probes the internal mechanisms of spatial reasoning in LLMs, revealing limitations in their current representations and suggesting a need for more robust, integrated spatial intelligence.
- DiffusionAnything: End-to-End In-context Diffusion Learning for Unified Navigation and Pre-Grasp Motion (viability: 5): https://sciencetostartup.com/paper/diffusionanything-end-to-end-in-context-diffusion-learning-for-unified-navigation-and-pre-grasp-motion - A unified diffusion policy for end-to-end robot navigation and manipulation from RGB input, requiring minimal self-supervised data.
- DFM-VLA: Iterative Action Refinement for Robot Manipulation via Discrete Flow Matching (viability: 7): https://sciencetostartup.com/paper/dfm-vla-iterative-action-refinement-for-robot-manipulation-via-discrete-flow-matching - A novel VLA model that iteratively refines robot action sequences for improved manipulation performance, outperforming existing methods.
- Label-Free Cross-Task LoRA Merging with Null-Space Compression (viability: 7): https://sciencetostartup.com/paper/label-free-cross-task-lora-merging-with-null-space-compression - A novel method for merging multiple fine-tuned AI models without retraining, enabling better performance across diverse tasks.
- SALMUBench: A Benchmark for Sensitive Association-Level Multimodal Unlearning (viability: 7): https://sciencetostartup.com/paper/salmubench-a-benchmark-for-sensitive-association-level-multimodal-unlearning - A new benchmark and dataset for evaluating the precise removal of sensitive information from multimodal AI models, addressing a critical gap in current AI safety and privacy.
- Line-of-Sight-Constrained Multi-Robot Mapless Navigation via Polygonal Visible Regions (viability: 7): https://sciencetostartup.com/paper/line-of-sight-constrained-multi-robot-mapless-navigation-via-polygonal-visible-regions - A distributed multi-robot navigation system that maintains line-of-sight connectivity using real-time LiDAR scans and egocentric visible regions, outperforming existing methods in complex environments.
- Semi-structured multi-state delinquency model for mortgage default (viability: 5): https://sciencetostartup.com/paper/semi-structured-multi-state-delinquency-model-for-mortgage-default - A hybrid model combining interpretable structured predictors with flexible neural networks to improve mortgage delinquency forecasting.
- D-GATNet: Interpretable Temporal Graph Attention Learning for ADHD Identification Using Dynamic Functional Connectivity (viability: 7): https://sciencetostartup.com/paper/d-gatnet-interpretable-temporal-graph-attention-learning-for-adhd-identification-using-dynamic-functional-connectivity - An interpretable temporal graph attention network for ADHD identification using dynamic brain connectivity, outperforming existing methods.
- Complete Causal Identification from Ancestral Graphs under Selection Bias (viability: 2): https://sciencetostartup.com/paper/complete-causal-identification-from-ancestral-graphs-under-selection-bias - Develops a theoretical framework for causal identification from ancestral graphs under selection bias, extending existing causal calculus.
- Preference-Aligned LoRA Merging: Preserving Subspace Coverage and Addressing Directional Anisotropy (viability: 7): https://sciencetostartup.com/paper/preference-aligned-lora-merging-preserving-subspace-coverage-and-addressing-directional-anisotropy - A novel LoRA merging technique that preserves task-specific model capabilities for more robust and general-purpose AI systems.
- Bitcoin Smart Accounts: Trust-Minimized Native Bitcoin DeFi Infrastructure (viability: 7): https://sciencetostartup.com/paper/bitcoin-smart-accounts-trust-minimized-native-bitcoin-defi-infrastructure - A trust-minimized protocol enabling native Bitcoin to participate in DeFi with self-custody, reducing trust assumptions for both protocols and depositors.
- findsylls: A Language-Agnostic Toolkit for Syllable-Level Speech Tokenization and Embedding (viability: 5): https://sciencetostartup.com/paper/findsylls-a-language-agnostic-toolkit-for-syllable-level-speech-tokenization-and-embedding - A language-agnostic toolkit for unified syllable-level speech tokenization and embedding, enabling reproducible research across diverse languages.
- PEB Separation and State Migration: Unmasking the New Frontiers of DeFi AML Evasion (viability: 4): https://sciencetostartup.com/paper/peb-separation-and-state-migration-unmasking-the-new-frontiers-of-defi-aml-evasion - A new framework for DeFi AML that moves beyond transaction tracing to analyze execution semantics and state invariants, addressing fundamental limitations in current systems.
- PhysVid: Physics Aware Local Conditioning for Generative Video Models (viability: 4): https://sciencetostartup.com/paper/physvid-physics-aware-local-conditioning-for-generative-video-models - A physics-aware conditioning scheme for generative video models to improve physical plausibility.
- Knowdit: Agentic Smart Contract Vulnerability Detection with Auditing Knowledge Summarization (viability: 7): https://sciencetostartup.com/paper/knowdit-agentic-smart-contract-vulnerability-detection-with-auditing-knowledge-summarization - Knowdit is a knowledge-driven, agentic framework that leverages historical audit data to systematically detect smart contract vulnerabilities, outperforming existing methods and discovering previously unknown bugs.
- GUIDE: Resolving Domain Bias in GUI Agents through Real-Time Web Video Retrieval and Plug-and-Play Annotation (viability: 7): https://sciencetostartup.com/paper/guide-resolving-domain-bias-in-gui-agents-through-real-time-web-video-retrieval-and-plug-and-play-annotation - A plug-and-play framework that resolves GUI agent domain bias by acquiring expertise from web tutorial videos, improving performance without model retraining.
- Topology-Aware Graph Reinforcement Learning for Energy Storage Systems Optimal Dispatch in Distribution Networks (viability: 7): https://sciencetostartup.com/paper/topology-aware-graph-reinforcement-learning-for-energy-storage-systems-optimal-dispatch-in-distribution-networks - A topology-aware reinforcement learning system using graph neural networks for optimal dispatch of energy storage systems to improve grid economy and voltage security.
- DRUM: Diffusion-based Raydrop-aware Unpaired Mapping for Sim2Real LiDAR Segmentation (viability: 7): https://sciencetostartup.com/paper/drum-diffusion-based-raydrop-aware-unpaired-mapping-for-sim2real-lidar-segmentation - A diffusion-based framework to bridge the domain gap between synthetic and real-world LiDAR data for improved robot perception.
- GLASS: Geometry-aware Local Alignment and Structure Synchronization Network for 2D-3D Registration (viability: 7): https://sciencetostartup.com/paper/glass-geometry-aware-local-alignment-and-structure-synchronization-network-for-2d-3d-registration - A novel network for precise 2D-3D image-to-point cloud registration that enhances geometric understanding and structural consistency, outperforming existing methods.
- Contrastive Conformal Sets (viability: 4): https://sciencetostartup.com/paper/contrastive-conformal-sets - A novel method for contrastive learning that provides theoretical guarantees on semantic feature space coverage and negative sample exclusion.
- GeoGuide: Hierarchical Geometric Guidance for Open-Vocabulary 3D Semantic Segmentation (viability: 7): https://sciencetostartup.com/paper/geoguide-hierarchical-geometric-guidance-for-open-vocabulary-3d-semantic-segmentation - A framework for open-vocabulary 3D semantic segmentation that leverages hierarchical geometric consistency to improve accuracy beyond training data.
- Working Notes on Late Interaction Dynamics: Analyzing Targeted Behaviors of Late Interaction Models (viability: 4): https://sciencetostartup.com/paper/working-notes-on-late-interaction-dynamics-analyzing-targeted-behaviors-of-late-interaction-models - This research analyzes performance bottlenecks in late interaction retrieval models by investigating length bias and similarity distribution, offering insights for improving search relevance.
- ARTA: Adaptive Mixed-Resolution Token Allocation for Efficient Dense Feature Extraction (viability: 7): https://sciencetostartup.com/paper/arta-adaptive-mixed-resolution-token-allocation-for-efficient-dense-feature-extraction - A vision transformer that efficiently extracts dense features by adaptively allocating computational resources to critical regions, achieving state-of-the-art results with less compute.
- Improving Risk Stratification in Hypertrophic Cardiomyopathy: A Novel Score Combining Echocardiography, Clinical, and Medication Data (viability: 7): https://sciencetostartup.com/paper/improving-risk-stratification-in-hypertrophic-cardiomyopathy-a-novel-score-combining-echocardiography-clinical-and-medic - A novel, explainable AI risk score for hypertrophic cardiomyopathy patients, outperforming existing models and enabling personalized clinical management.
- SocialX: A Modular Platform for Multi-Source Big Data Research in Indonesia (viability: 7): https://sciencetostartup.com/paper/socialx-a-modular-platform-for-multi-source-big-data-research-in-indonesia - SocialX is a modular platform that unifies fragmented big data from diverse Indonesian sources, streamlining research by integrating collection, preprocessing, and analysis into a single pipeline.
- Channelling, Coordinating, Collaborating: A Three-Layer Framework for Disability-Centered Human-Agent Collaboration (viability: 3): https://sciencetostartup.com/paper/channelling-coordinating-collaborating-a-three-layer-framework-for-disability-centered-human-agent-collaboration - A framework to rethink AI's role in ability-diverse collaboration by centering the interdependent ways people with disabilities already work.
- Real-Time Branch-to-Tool Distance Estimation for Autonomous UAV Pruning: Benchmarking Five DEFOM-Stereo Variants from Simulation to Jetson Deployment (viability: 7): https://sciencetostartup.com/paper/real-time-branch-to-tool-distance-estimation-for-autonomous-uav-pruning-benchmarking-five-defom-stereo-variants-from-sim - Real-time tool-to-branch distance estimation for autonomous UAV pruning, balancing accuracy and speed for safe operation.
- Knowledge Distillation for Efficient Transformer-Based Reinforcement Learning in Hardware-Constrained Energy Management Systems (viability: 7): https://sciencetostartup.com/paper/knowledge-distillation-for-efficient-transformer-based-reinforcement-learning-in-hardware-constrained-energy-management - Compresses powerful transformer-based reinforcement learning models for efficient deployment on energy management hardware, reducing costs and improving self-consumption.
- Automatic Speech Recognition for Documenting Endangered Languages: Case Study of Ikema Miyakoan (viability: 5): https://sciencetostartup.com/paper/automatic-speech-recognition-for-documenting-endangered-languages-case-study-of-ikema-miyakoan - Develop an automatic speech recognition system to accelerate the documentation of endangered languages, reducing transcription time and cognitive load.
- Distilling Conversations: Abstract Compression of Conversational Audio Context for LLM-based ASR (viability: 4): https://sciencetostartup.com/paper/distilling-conversations-abstract-compression-of-conversational-audio-context-for-llm-based-asr - A method to compress conversational audio context for improved LLM-based speech recognition, reducing computational cost.
- Physics-Informed Neural Networks and Sequence Encoder: Application to heating and early cooling of thermo-stamping process (viability: 3): https://sciencetostartup.com/paper/physics-informed-neural-networks-and-sequence-encoder-application-to-heating-and-early-cooling-of-thermo-stamping-proces - A physics-informed neural network combined with a sequence encoder for predicting the behavior of complex thermo-stamping processes.
- Automating Domain-Driven Design: Experience with a Prompting Framework (viability: 4): https://sciencetostartup.com/paper/automating-domain-driven-design-experience-with-a-prompting-framework - A prompting framework that automates core domain-driven design activities using LLMs to assist software architects.
- SwarmCoDe: A Scalable Co-Design Framework for Heterogeneous Robot Swarms via Dynamic Speciation (viability: 3): https://sciencetostartup.com/paper/swarmcode-a-scalable-co-design-framework-for-heterogeneous-robot-swarms-via-dynamic-speciation - A novel algorithm for co-designing large-scale, heterogeneous robot swarms by dynamically adapting swarm complexity to task requirements.
- A Universal Vibe? Finding and Controlling Language-Agnostic Informal Register with SAEs (viability: 7): https://sciencetostartup.com/paper/a-universal-vibe-finding-and-controlling-language-agnostic-informal-register-with-saes - This research identifies and controls a portable, language-agnostic abstraction for informal language within multilingual LLMs, enabling cross-lingual formality control.
- GS-BrainText: A Multi-Site Brain Imaging Report Dataset from Generation Scotland for Clinical Natural Language Processing Development and Validation (viability: 5): https://sciencetostartup.com/paper/gs-braintext-a-multi-site-brain-imaging-report-dataset-from-generation-scotland-for-clinical-natural-language-processing - A new, large-scale, multi-site dataset of brain radiology reports with expert annotations to develop and validate generalizable clinical NLP tools.
- Ask or Assume? Uncertainty-Aware Clarification-Seeking in Coding Agents (viability: 7): https://sciencetostartup.com/paper/ask-or-assume-uncertainty-aware-clarification-seeking-in-coding-agents - LLM agents that proactively seek clarification to resolve underspecified instructions, significantly improving task completion rates.
- Optimization Trade-offs in Asynchronous Federated Learning: A Stochastic Networks Approach (viability: 4): https://sciencetostartup.com/paper/optimization-trade-offs-in-asynchronous-federated-learning-a-stochastic-networks-approach - This research provides a theoretical framework and optimization strategies to improve the efficiency and reduce the energy consumption of asynchronous federated learning systems.
- Privacy-Accuracy Trade-offs in High-Dimensional LASSO under Perturbation Mechanisms (viability: 4): https://sciencetostartup.com/paper/privacy-accuracy-trade-offs-in-high-dimensional-lasso-under-perturbation-mechanisms - Develops a theoretical framework for privacy-preserving sparse linear regression using approximate message passing to analyze trade-offs in high-dimensional settings.
- Privacy-Enhancing Encryption in Data Sharing: A Survey on Security, Performance and Functionality (viability: 3): https://sciencetostartup.com/paper/privacy-enhancing-encryption-in-data-sharing-a-survey-on-security-performance-and-functionality - This paper surveys privacy-enhancing encryption technologies for secure and efficient data sharing across industries, identifying potential attacks and integration strategies.
- Clawed and Dangerous: Can We Trust Open Agentic Systems? (viability: 4): https://sciencetostartup.com/paper/clawed-and-dangerous-can-we-trust-open-agentic-systems - A framework for systematically evaluating and securing open agentic systems by providing a taxonomy, doctrine, and scorecard for governable and resilient agent platforms.
- EPDQ: Efficient and Privacy-Preserving Exact Distance Query on Encrypted Graphs (viability: 7): https://sciencetostartup.com/paper/epdq-efficient-and-privacy-preserving-exact-distance-query-on-encrypted-graphs - Enables efficient and private shortest distance queries on encrypted graph databases using a novel tensor-based indexing scheme.
- Towards GUI Agents: Vision-Language Diffusion Models for GUI Grounding (viability: 7): https://sciencetostartup.com/paper/towards-gui-agents-vision-language-diffusion-models-for-gui-grounding - Leveraging vision-language diffusion models to enable more accurate and efficient GUI grounding for building autonomous agents.
- Sparse Auto-Encoders and Holism about Large Language Models (viability: 2): https://sciencetostartup.com/paper/sparse-auto-encoders-and-holism-about-large-language-models - This paper theoretically explores the semantic interpretation of Large Language Models by analyzing the role of sparse auto-encoders in their internal feature representations.
- 4DRaL: Bridging 4D Radar with LiDAR for Place Recognition using Knowledge Distillation (viability: 4): https://sciencetostartup.com/paper/4dral-bridging-4d-radar-with-lidar-for-place-recognition-using-knowledge-distillation - A knowledge distillation framework to improve place recognition for 4D radar by leveraging LiDAR data, enabling robust localization in all weather conditions.
- An Object Web Seminar: A Retrospective on a Technical Dialogue Still Reverbarating (viability: 1): https://sciencetostartup.com/paper/an-object-web-seminar-a-retrospective-on-a-technical-dialogue-still-reverbarating - This paper analyzes the historical convergence of Object Technologies and early Web adoption, drawing parallels to modern architectures like Kubernetes and microservices, and briefly touches on early AI tools.
- SAFT: Sensitivity-Aware Filtering and Transmission for Adaptive 3D Point Cloud Communication over Wireless Channels (viability: 4): https://sciencetostartup.com/paper/saft-sensitivity-aware-filtering-and-transmission-for-adaptive-3d-point-cloud-communication-over-wireless-channels - A learned framework for adaptive 3D point cloud transmission over wireless channels, improving fidelity under low bandwidth and signal conditions.
- MemCam: Memory-Augmented Camera Control for Consistent Video Generation (viability: 5): https://sciencetostartup.com/paper/memcam-memory-augmented-camera-control-for-consistent-video-generation - A memory-augmented approach for consistent interactive video generation that maintains scene coherence under dynamic camera control.
- HAD: Heterogeneity-Aware Distillation for Lifelong Heterogeneous Learning (viability: 7): https://sciencetostartup.com/paper/had-heterogeneity-aware-distillation-for-lifelong-heterogeneous-learning - A novel distillation method for lifelong learning that adapts to changing task structures, outperforming existing approaches in heterogeneous learning scenarios.
- Dual-Stage Invariant Continual Learning under Extreme Visual Sparsity (viability: 4): https://sciencetostartup.com/paper/dual-stage-invariant-continual-learning-under-extreme-visual-sparsity - A dual-stage continual learning framework for object detection in extreme visual sparsity, improving representation stability and detection accuracy.
- OSA: Echocardiography Video Segmentation via Orthogonalized State Update and Anatomical Prior-aware Feature Enhancement (viability: 7): https://sciencetostartup.com/paper/osa-echocardiography-video-segmentation-via-orthogonalized-state-update-and-anatomical-prior-aware-feature-enhancement - A novel framework for real-time, accurate segmentation of echocardiography videos to improve cardiac function assessment.
- Progressive Learning with Anatomical Priors for Reliable Left Atrial Scar Segmentation from Late Gadolinium Enhancement MRI (viability: 7): https://sciencetostartup.com/paper/progressive-learning-with-anatomical-priors-for-reliable-left-atrial-scar-segmentation-from-late-gadolinium-enhancement - A progressive learning framework for more reliable left atrial scar segmentation in cardiac MRI, inspired by clinical workflows.
- DUGAE: Unified Geometry and Attribute Enhancement via Spatiotemporal Correlations for G-PCC Compressed Dynamic Point Clouds (viability: 7): https://sciencetostartup.com/paper/dugae-unified-geometry-and-attribute-enhancement-via-spatiotemporal-correlations-for-g-pcc-compressed-dynamic-point-clou - A unified framework for enhancing dynamic point cloud geometry and attributes by exploiting spatiotemporal correlations, significantly improving compression efficiency and quality.
- ClinicalAgents: Multi-Agent Orchestration for Clinical Decision Making with Dual-Memory (viability: 7): https://sciencetostartup.com/paper/clinicalagents-multi-agent-orchestration-for-clinical-decision-making-with-dual-memory - A multi-agent AI system that mimics expert clinician reasoning for improved diagnostic accuracy and explainability.
- GLINT: Modeling Scene-Scale Transparency via Gaussian Radiance Transport (viability: 7): https://sciencetostartup.com/paper/glint-modeling-scene-scale-transparency-via-gaussian-radiance-transport - GLINT enables realistic 3D scene reconstruction of transparent objects by explicitly modeling radiance transport through decomposed Gaussian representations.
- Consistency Beyond Contrast: Enhancing Open-Vocabulary Object Detection Robustness via Contextual Consistency Learning (viability: 7): https://sciencetostartup.com/paper/consistency-beyond-contrast-enhancing-open-vocabulary-object-detection-robustness-via-contextual-consistency-learning - A framework for robust open-vocabulary object detection that improves generalization by enforcing contextual consistency within visual data, achieving state-of-the-art results.
- Geometric Evolution Graph Convolutional Networks: Enhancing Graph Representation Learning via Ricci Flow (viability: 7): https://sciencetostartup.com/paper/geometric-evolution-graph-convolutional-networks-enhancing-graph-representation-learning-via-ricci-flow - A novel graph convolutional network framework that models geometric evolution for state-of-the-art performance on graph classification tasks, especially for heterophilic graphs.
- Can AI Scientist Agents Learn from Lab-in-the-Loop Feedback? Evidence from Iterative Perturbation Discovery (viability: 4): https://sciencetostartup.com/paper/can-ai-scientist-agents-learn-from-lab-in-the-loop-feedback-evidence-from-iterative-perturbation-discovery - AI agents can significantly improve scientific discovery by learning from experimental feedback, with performance scaling with model capability.
- CREval: An Automated Interpretable Evaluation for Creative Image Manipulation under Complex Instructions (viability: 7): https://sciencetostartup.com/paper/creval-an-automated-interpretable-evaluation-for-creative-image-manipulation-under-complex-instructions - An automated evaluation framework and benchmark for creative image manipulation that provides reliable metrics and identifies key challenges for model development.
- ComVi: Context-Aware Optimized Comment Display in Video Playback (viability: 5): https://sciencetostartup.com/paper/comvi-context-aware-optimized-comment-display-in-video-playback - A system that synchronizes video comments to relevant video moments to improve viewer engagement and reduce spoilers.
- Provably Contractive and High-Quality Denoisers for Convergent Restoration (viability: 7): https://sciencetostartup.com/paper/provably-contractive-and-high-quality-denoisers-for-convergent-restoration - Develops provably stable and high-quality image denoisers with guaranteed robustness to input perturbations, challenging assumptions about quality degradation with Lipschitz control.
- Gaussian Shannon: High-Precision Diffusion Model Watermarking Based on Communication (viability: 7): https://sciencetostartup.com/paper/gaussian-shannon-high-precision-diffusion-model-watermarking-based-on-communication - A diffusion model watermarking framework that enables robust tracing and exact bit recovery for AI-generated content.
- DataFlex: A Unified Framework for Data-Centric Dynamic Training of Large Language Models (viability: 7): https://sciencetostartup.com/paper/dataflex-a-unified-framework-for-data-centric-dynamic-training-of-large-language-models - DataFlex unifies and streamlines data-centric dynamic training for LLMs, offering improved performance and efficiency with a drop-in framework.
- Clash of the models: Comparing performance of BERT-based variants for generic news frame detection (viability: 5): https://sciencetostartup.com/paper/clash-of-the-models-comparing-performance-of-bert-based-variants-for-generic-news-frame-detection - A comparative analysis of BERT-based models for generic news frame detection, offering a new Swiss electoral context dataset.
- IP-Bench: Benchmark for Image Protection Methods in Image-to-Video Generation Scenarios (viability: 4): https://sciencetostartup.com/paper/ip-bench-benchmark-for-image-protection-methods-in-image-to-video-generation-scenarios - A benchmark for evaluating image protection methods in image-to-video generation to combat misuse.
- Efficient Few-Shot Learning for Edge AI via Knowledge Distillation on MobileViT (viability: 7): https://sciencetostartup.com/paper/efficient-few-shot-learning-for-edge-ai-via-knowledge-distillation-on-mobilevit - Enabling efficient few-shot learning on edge devices through knowledge distillation for mobile vision transformers.
- On the Complexity of Optimal Graph Rewiring for Oversmoothing and Oversquashing in Graph Neural Networks (viability: 2): https://sciencetostartup.com/paper/on-the-complexity-of-optimal-graph-rewiring-for-oversmoothing-and-oversquashing-in-graph-neural-networks - This paper theoretically investigates the computational complexity of optimizing graph structures to improve Graph Neural Network performance, proving NP-hardness for exact solutions.
- PruneFuse: Efficient Data Selection via Weight Pruning and Network Fusion (viability: 4): https://sciencetostartup.com/paper/prunefuse-efficient-data-selection-via-weight-pruning-and-network-fusion - A novel strategy for efficient data selection in deep learning that reduces computational costs and accelerates training by leveraging pruned networks.
- ATime-Consistent Benchmark for Repository-Level Software Engineering Evaluation (viability: 7): https://sciencetostartup.com/paper/atime-consistent-benchmark-for-repository-level-software-engineering-evaluation - A time-consistent benchmark and methodology for evaluating repository-aware software engineering AI agents, improving prompt construction and temporal validity.
- PEANUT: Perturbations by Eigenvalue Alignment for Attacking GNNs Under Topology-Driven Message Passing (viability: 7): https://sciencetostartup.com/paper/peanut-perturbations-by-eigenvalue-alignment-for-attacking-gnns-under-topology-driven-message-passing - A gradient-free attack that injects virtual nodes to exploit Graph Neural Network vulnerabilities for enhanced robustness testing.
- TinyML for Acoustic Anomaly Detection in IoT Sensor Networks (viability: 5): https://sciencetostartup.com/paper/tinyml-for-acoustic-anomaly-detection-in-iot-sensor-networks - Deploy real-time, energy-efficient acoustic anomaly detection directly on IoT devices for enhanced safety and context awareness.
- InstaVSR: Taming Diffusion for Efficient and Temporally Consistent Video Super-Resolution (viability: 5): https://sciencetostartup.com/paper/instavsr-taming-diffusion-for-efficient-and-temporally-consistent-video-super-resolution - A lightweight diffusion framework for efficient and temporally consistent video super-resolution.
- SWE-PRBench: Benchmarking AI Code Review Quality Against Pull Request Feedback (viability: 7): https://sciencetostartup.com/paper/swe-prbench-benchmarking-ai-code-review-quality-against-pull-request-feedback - A benchmark and evaluation framework to measure and improve AI code review quality, revealing current AI limitations and guiding future development.
- TaxaAdapter: Vision Taxonomy Models are Key to Fine-grained Image Generation over the Tree of Life (viability: 7): https://sciencetostartup.com/paper/taxaadapter-vision-taxonomy-models-are-key-to-fine-grained-image-generation-over-the-tree-of-life - A lightweight adapter for text-to-image models that uses vision taxonomy embeddings to achieve highly accurate species-level image generation, even for rare or unseen species.
- Finding Distributed Object-Centric Properties in Self-Supervised Transformers (viability: 7): https://sciencetostartup.com/paper/finding-distributed-object-centric-properties-in-self-supervised-transformers - A training-free method to extract distributed object-centric information from self-supervised transformers for improved object discovery and grounding in multimodal models.
- Beyond Where to Look: Trajectory-Guided Reinforcement Learning for Multimodal RLVR (viability: 7): https://sciencetostartup.com/paper/beyond-where-to-look-trajectory-guided-reinforcement-learning-for-multimodal-rlvr - A novel reinforcement learning approach that guides multimodal models to better integrate visual evidence into their reasoning processes, improving accuracy on complex tasks.
- SkinGPT-X: A Self-Evolving Collaborative Multi-Agent System for Transparent and Trustworthy Dermatological Diagnosis (viability: 7): https://sciencetostartup.com/paper/skingpt-x-a-self-evolving-collaborative-multi-agent-system-for-transparent-and-trustworthy-dermatological-diagnosis - SkinGPT-X is a self-evolving multi-agent system that provides transparent and trustworthy dermatological diagnoses, outperforming state-of-the-art models on complex and rare skin conditions.
- DPD-Cancer: Explainable Graph-based Deep Learning for Small Molecule Anti-Cancer Activity Prediction (viability: 8): https://sciencetostartup.com/paper/dpd-cancer-explainable-graph-based-deep-learning-for-small-molecule-anti-cancer-activity-prediction - DPD-Cancer offers a state-of-the-art, explainable AI tool for predicting small molecule anti-cancer activity, enhancing precision medicine.
- SDDF: Specificity-Driven Dynamic Focusing for Open-Vocabulary Camouflaged Object Detection (viability: 5): https://sciencetostartup.com/paper/sddf-specificity-driven-dynamic-focusing-for-open-vocabulary-camouflaged-object-detection - A new method for detecting camouflaged objects in images using text prompts, with a new benchmark dataset and code available.
- Accurate Precipitation Forecast by Efficiently Learning from Massive Atmospheric Variables and Unbalanced Distribution (viability: 7): https://sciencetostartup.com/paper/accurate-precipitation-forecast-by-efficiently-learning-from-massive-atmospheric-variables-and-unbalanced-distribution - A novel precipitation forecasting model that efficiently leverages massive atmospheric data and a specialized loss function to achieve superior accuracy and computational efficiency.
- LLM Benchmark-User Need Misalignment for Climate Change (viability: 5): https://sciencetostartup.com/paper/llm-benchmark-user-need-misalignment-for-climate-change - This research identifies a critical misalignment between current LLM benchmarks and real-world user needs for climate change information, offering a framework to improve RAG systems and LLM training.
- Are LLM-Enhanced Graph Neural Networks Robust against Poisoning Attacks? (viability: 7): https://sciencetostartup.com/paper/are-llm-enhanced-graph-neural-networks-robust-against-poisoning-attacks - A framework to assess and defend LLM-enhanced Graph Neural Networks against sophisticated poisoning attacks, with released code for broader research.
- "Oops! ChatGPT is Temporarily Unavailable!": A Diary Study on Knowledge Workers' Experiences of LLM Withdrawal (viability: 3): https://sciencetostartup.com/paper/oops-chatgpt-is-temporarily-unavailable-a-diary-study-on-knowledge-workers-experiences-of-llm-withdrawal - This study explores knowledge workers' experiences and dependencies on LLMs during temporary withdrawal, highlighting workflow disruptions and the normative integration of LLMs into daily work practices.
- A Human-Inspired Decoupled Architecture for Efficient Audio Representation Learning (viability: 7): https://sciencetostartup.com/paper/a-human-inspired-decoupled-architecture-for-efficient-audio-representation-learning - A highly efficient audio representation learning architecture inspired by human cognition, offering competitive performance with significantly reduced computational cost.
- Dynamic Tokenization via Reinforcement Patching: End-to-end Training and Zero-shot Transfer (viability: 7): https://sciencetostartup.com/paper/dynamic-tokenization-via-reinforcement-patching-end-to-end-training-and-zero-shot-transfer - A reinforcement learning framework for dynamic, data-driven sequence patching to improve time-series forecasting and enable reusable foundation patchers.
- AcTTA: Rethinking Test-Time Adaptation via Dynamic Activation (viability: 4): https://sciencetostartup.com/paper/actta-rethinking-test-time-adaptation-via-dynamic-activation - A novel framework for adaptive test-time adaptation by dynamically reinterpreting and updating activation functions, outperforming existing normalization-based methods.
- IndoBERT-Relevancy: A Context-Conditioned Relevancy Classifier for Indonesian Text (viability: 7): https://sciencetostartup.com/paper/indobert-relevancy-a-context-conditioned-relevancy-classifier-for-indonesian-text - A context-conditioned relevancy classifier for Indonesian text, trained on a novel dataset and achieving state-of-the-art performance.
- ROAST: Risk-aware Outlier-exposure for Adversarial Selective Training of Anomaly Detectors Against Evasion Attacks (viability: 4): https://sciencetostartup.com/paper/roast-risk-aware-outlier-exposure-for-adversarial-selective-training-of-anomaly-detectors-against-evasion-attacks - A novel framework for anomaly detection that improves recall by selectively training on less vulnerable patient data and using outlier exposure to maintain precision.
- CD-Buffer: Complementary Dual-Buffer Framework for Test-Time Adaptation in Adverse Weather Object Detection (viability: 7): https://sciencetostartup.com/paper/cd-buffer-complementary-dual-buffer-framework-for-test-time-adaptation-in-adverse-weather-object-detection - A novel framework for real-time object detection adaptation in adverse weather by adaptively balancing feature removal and refinement.
- Selective Deficits in LLM Mental Self-Modeling in a Behavior-Based Test of Theory of Mind (viability: 7): https://sciencetostartup.com/paper/selective-deficits-in-llm-mental-self-modeling-in-a-behavior-based-test-of-theory-of-mind - This research develops a novel test for LLM Theory of Mind, revealing that while recent models excel at understanding others, they struggle with self-modeling unless given a scratchpad, suggesting a path to more sophisticated AI reasoning.
- Learnable Instance Attention Filtering for Adaptive Detector Distillation (viability: 7): https://sciencetostartup.com/paper/learnable-instance-attention-filtering-for-adaptive-detector-distillation - A novel framework for adaptive knowledge distillation in computer vision that improves student model efficiency by learning to reweight instance importance during training.
- When Identities Collapse: A Stress-Test Benchmark for Multi-Subject Personalization (viability: 7): https://sciencetostartup.com/paper/when-identities-collapse-a-stress-test-benchmark-for-multi-subject-personalization - A stress-test benchmark and novel metric to expose and address identity collapse in multi-subject text-to-image generation.
- Semi-Automated Knowledge Engineering and Process Mapping for Total Airport Management (viability: 4): https://sciencetostartup.com/paper/semi-automated-knowledge-engineering-and-process-mapping-for-total-airport-management - A framework to build machine-readable knowledge graphs for complex domains like airport management by fusing expert knowledge engineering with LLMs.
- Not All Entities are Created Equal: A Dynamic Anonymization Framework for Privacy-Preserving Retrieval-Augmented Generation (viability: 7): https://sciencetostartup.com/paper/not-all-entities-are-created-equal-a-dynamic-anonymization-framework-for-privacy-preserving-retrieval-augmented-generati - A dynamic framework for anonymizing knowledge bases in RAG systems to protect sensitive data without sacrificing LLM utility.
- MUST: Modality-Specific Representation-Aware Transformer for Diffusion-Enhanced Survival Prediction with Missing Modality (viability: 7): https://sciencetostartup.com/paper/must-modality-specific-representation-aware-transformer-for-diffusion-enhanced-survival-prediction-with-missing-modality - A Transformer-based framework for accurate survival prediction in oncology that robustly handles missing medical data by generating missing modality representations using diffusion models.
- PAD-Hand: Physics-Aware Diffusion for Hand Motion Recovery (viability: 7): https://sciencetostartup.com/paper/pad-hand-physics-aware-diffusion-for-hand-motion-recovery - A physics-aware diffusion model that refines hand motion from images, providing confidence estimates for physical plausibility.
- R-PGA: Robust Physical Adversarial Camouflage Generation via Relightable 3D Gaussian Splatting (viability: 3): https://sciencetostartup.com/paper/r-pga-robust-physical-adversarial-camouflage-generation-via-relightable-3d-gaussian-splatting - A novel framework for generating robust physical adversarial camouflage for autonomous driving systems by leveraging relightable 3D Gaussian splatting and hard physical configuration mining.
- Adversarial Bandit Optimization with Globally Bounded Perturbations to Linear Losses (viability: 2): https://sciencetostartup.com/paper/adversarial-bandit-optimization-with-globally-bounded-perturbations-to-linear-losses - This paper theoretically analyzes adversarial bandit optimization with bounded perturbations to linear losses, providing new regret bounds.
- MuDD: A Multimodal Deception Detection Dataset and GSR-Guided Progressive Distillation for Non-Contact Deception Detection (viability: 7): https://sciencetostartup.com/paper/mudd-a-multimodal-deception-detection-dataset-and-gsr-guided-progressive-distillation-for-non-contact-deception-detectio - A multimodal dataset and distillation framework for non-contact deception detection, leveraging physiological cues to improve accuracy.
- I Want to Believe (but the Vocabulary Changed): Measuring the Semantic Structure and Evolution of Conspiracy Theories (viability: 4): https://sciencetostartup.com/paper/i-want-to-believe-but-the-vocabulary-changed-measuring-the-semantic-structure-and-evolution-of-conspiracy-theories - A tool to track the semantic evolution of conspiracy theories in online discourse, moving beyond keyword analysis.
- Pioneering Perceptual Video Fluency Assessment: A Novel Task with Benchmark Dataset and Baseline (viability: 7): https://sciencetostartup.com/paper/pioneering-perceptual-video-fluency-assessment-a-novel-task-with-benchmark-dataset-and-baseline - A new task and dataset for assessing video fluency, with a baseline model that outperforms existing methods, enabling better video streaming and gaming experiences.
- Bridging Pixels and Words: Mask-Aware Local Semantic Fusion for Multimodal Media Verification (viability: 5): https://sciencetostartup.com/paper/bridging-pixels-and-words-mask-aware-local-semantic-fusion-for-multimodal-media-verification - A novel framework for multimodal media verification that actively bridges pixels and words using mask-aware local semantic fusion to detect sophisticated misinformation.
- Seeing Like Radiologists: Context- and Gaze-Guided Vision-Language Pretraining for Chest X-rays (viability: 7): https://sciencetostartup.com/paper/seeing-like-radiologists-context-and-gaze-guided-vision-language-pretraining-for-chest-x-rays - A vision-language pretraining framework for chest X-rays that integrates clinical context and radiologist gaze to improve diagnostic accuracy and report generation.
- Asymptotic Optimism for Tensor Regression Models with Applications to Neural Network Compression (viability: 4): https://sciencetostartup.com/paper/asymptotic-optimism-for-tensor-regression-models-with-applications-to-neural-network-compression - A novel rank selection method for tensor regression that optimizes neural network compression and model selection.
- Retrieval-Augmented Generation Based Nurse Observation Extraction (viability: 7): https://sciencetostartup.com/paper/retrieval-augmented-generation-based-nurse-observation-extraction - An automated RAG-based system to extract critical clinical observations from nurse dictations, reducing administrative burden and improving patient care.
- H-Node Attack and Defense in Large Language Models (viability: 4): https://sciencetostartup.com/paper/h-node-attack-and-defense-in-large-language-models - A framework to identify and defend against hallucinations in LLMs by targeting specific hidden-state dimensions.
- Rethinking Token Pruning for Historical Screenshots in GUI Visual Agents: Semantic, Spatial, and Temporal Perspectives (viability: 5): https://sciencetostartup.com/paper/rethinking-token-pruning-for-historical-screenshots-in-gui-visual-agents-semantic-spatial-and-temporal-perspectives - This research proposes a novel token pruning strategy for GUI visual agents that significantly reduces computational cost by leveraging semantic, spatial, and temporal insights from historical screenshots, without sacrificing performance.
- Face2Parts: Exploring Coarse-to-Fine Inter-Regional Facial Dependencies for Generalized Deepfake Detection (viability: 7): https://sciencetostartup.com/paper/face2parts-exploring-coarse-to-fine-inter-regional-facial-dependencies-for-generalized-deepfake-detection - A novel hybrid deepfake detection method that leverages hierarchical feature representation and inter-regional facial dependencies to achieve generalized and superior performance.
- AgentCollab: A Self-Evaluation-Driven Collaboration Paradigm for Efficient LLM Agents (viability: 7): https://sciencetostartup.com/paper/agentcollab-a-self-evaluation-driven-collaboration-paradigm-for-efficient-llm-agents - A self-evaluation framework that dynamically coordinates LLM agents of varying capabilities to improve task efficiency and reasoning robustness.
- Knowledge is Power: Advancing Few-shot Action Recognition with Multimodal Semantics from MLLMs (viability: 7): https://sciencetostartup.com/paper/knowledge-is-power-advancing-few-shot-action-recognition-with-multimodal-semantics-from-mllms - Leverage multimodal large language models to directly enhance few-shot action recognition with enriched representations and adaptive prompting.
- Protecting User Prompts Via Character-Level Differential Privacy (viability: 7): https://sciencetostartup.com/paper/protecting-user-prompts-via-character-level-differential-privacy - A novel differential privacy method to sanitize LLM prompts by perturbing characters, protecting sensitive information without explicit PII identification.
- Designing Fatigue-Aware VR Interfaces via Biomechanical Models (viability: 7): https://sciencetostartup.com/paper/designing-fatigue-aware-vr-interfaces-via-biomechanical-models - Optimize VR interfaces for reduced user fatigue using biomechanical models and reinforcement learning, enabling faster ergonomic design cycles.
- Constitutive parameterized deep energy method for solid mechanics problems with random material parameters (viability: 7): https://sciencetostartup.com/paper/constitutive-parameterized-deep-energy-method-for-solid-mechanics-problems-with-random-material-parameters - A physics-driven AI that enables real-time structural simulations with uncertain material properties, eliminating the need for retraining.
- Learning to Trim: End-to-End Causal Graph Pruning with Dynamic Anatomical Feature Banks for Medical VQA (viability: 7): https://sciencetostartup.com/paper/learning-to-trim-end-to-end-causal-graph-pruning-with-dynamic-anatomical-feature-banks-for-medical-vqa - A framework for medical visual question answering that learns to suppress spurious correlations and focus on causal diagnostic evidence, improving generalization.
- Identification of Bivariate Causal Directionality Based on Anticipated Asymmetric Geometries (viability: 3): https://sciencetostartup.com/paper/identification-of-bivariate-causal-directionality-based-on-anticipated-asymmetric-geometries - This paper proposes two novel methods, Anticipated Asymmetric Geometries and Monotonicity Index, to identify causal directionality in bivariate numerical data, outperforming existing methods on real-world examples.
- GLU: Global-Local-Uncertainty Fusion for Scalable Spatiotemporal Reconstruction and Forecasting (viability: 7): https://sciencetostartup.com/paper/glu-global-local-uncertainty-fusion-for-scalable-spatiotemporal-reconstruction-and-forecasting - A unified framework for reconstructing unobserved states and forecasting the evolution of complex physical systems using sparse measurements.
- Unlabeled Cross-Center Automatic Analysis for TAAD: An Integrated Framework from Segmentation to Clinical Features (viability: 7): https://sciencetostartup.com/paper/unlabeled-cross-center-automatic-analysis-for-taad-an-integrated-framework-from-segmentation-to-clinical-features - An unsupervised domain adaptation framework for automated extraction of critical clinical features from medical images, enabling precise preoperative assessment without target-domain annotations.
- GeoReFormer: Geometry-Aware Refinement for Lane Segment Detection and Topology Reasoning (viability: 7): https://sciencetostartup.com/paper/georeformer-geometry-aware-refinement-for-lane-segment-detection-and-topology-reasoning - A transformer architecture that embeds geometric and topological priors for more accurate 3D lane detection and map construction in autonomous driving.
- QuitoBench: A High-Quality Open Time Series Forecasting Benchmark (viability: 5): https://sciencetostartup.com/paper/quitobench-a-high-quality-open-time-series-forecasting-benchmark - A new benchmark for time series forecasting that reveals performance trade-offs between deep learning and foundation models across different data regimes.
- VLAgeBench: Benchmarking Large Vision-Language Models for Zero-Shot Human Age Estimation (viability: 7): https://sciencetostartup.com/paper/vlagebench-benchmarking-large-vision-language-models-for-zero-shot-human-age-estimation - Leverage state-of-the-art large vision-language models for zero-shot human age estimation, offering a competitive alternative to traditional supervised methods for applications in biometrics and healthcare.
- Toward Culturally Grounded Natural Language Processing (viability: 4): https://sciencetostartup.com/paper/toward-culturally-grounded-natural-language-processing - This research proposes a new framework for Natural Language Processing that accounts for cultural nuances and local norms, moving beyond simple multilingual capabilities to achieve true cultural competence.
- FairLLaVA: Fairness-Aware Parameter-Efficient Fine-Tuning for Large Vision-Language Assistants (viability: 7): https://sciencetostartup.com/paper/fairllava-fairness-aware-parameter-efficient-fine-tuning-for-large-vision-language-assistants - FairLLaVA offers a parameter-efficient fine-tuning method to mitigate demographic biases in multimodal LLMs for critical applications like medical imaging.
- AutoB2G: A Large Language Model-Driven Agentic Framework For Automated Building-Grid Co-Simulation (viability: 7): https://sciencetostartup.com/paper/autob2g-a-large-language-model-driven-agentic-framework-for-automated-building-grid-co-simulation - An LLM-driven agentic framework that automates building-grid co-simulation, enabling efficient coordination and improved grid-side performance through natural language task descriptions.
- A Large-scale Empirical Study on the Generalizability of Disclosed Java Library Vulnerability Exploits (viability: 7): https://sciencetostartup.com/paper/a-large-scale-empirical-study-on-the-generalizability-of-disclosed-java-library-vulnerability-exploits - Automate Java library vulnerability detection by migrating and applying existing exploits across versions, significantly improving accuracy over current tools.
- Neighbor-Aware Localized Concept Erasure in Text-to-Image Diffusion Models (viability: 7): https://sciencetostartup.com/paper/neighbor-aware-localized-concept-erasure-in-text-to-image-diffusion-models - A training-free framework for precisely removing unwanted concepts from AI-generated images without damaging related elements.
- FAST3DIS: Feed-forward Anchored Scene Transformer for 3D Instance Segmentation (viability: 7): https://sciencetostartup.com/paper/fast3dis-feed-forward-anchored-scene-transformer-for-3d-instance-segmentation - An end-to-end Transformer for 3D instance segmentation that bypasses slow clustering methods, offering improved scalability and speed.
- JRM: Joint Reconstruction Model for Multiple Objects without Alignment (viability: 4): https://sciencetostartup.com/paper/jrm-joint-reconstruction-model-for-multiple-objects-without-alignment - A generative model for 3D object reconstruction that implicitly handles multiple unaligned observations, improving robustness and accuracy.
- Policy-Guided World Model Planning for Language-Conditioned Visual Navigation (viability: 7): https://sciencetostartup.com/paper/policy-guided-world-model-planning-for-language-conditioned-visual-navigation - A two-stage framework for instruction-conditioned visual navigation that combines learned policies with latent world model planning for faster convergence to high-quality action sequences.
- Diffusion MRI Transformer with a Diffusion Space Rotary Positional Embedding (D-RoPE) (viability: 7): https://sciencetostartup.com/paper/diffusion-mri-transformer-with-a-diffusion-space-rotary-positional-embedding-d-rope - A novel transformer architecture with a diffusion space positional embedding for improved analysis of diffusion MRI data, achieving superior accuracy in clinical tasks.
- Second-Order, First-Class: A Composable Stack for Curvature-Aware Training (viability: 4): https://sciencetostartup.com/paper/second-order-first-class-a-composable-stack-for-curvature-aware-training - A composable stack for curvature-aware training that optimizes stability and convergence for machine learning models.
- Do Neurons Dream of Primitive Operators? Wake-Sleep Compression Rediscovers Schank's Event Semantics (viability: 3): https://sciencetostartup.com/paper/do-neurons-dream-of-primitive-operators-wake-sleep-compression-rediscovers-schank-s-event-semantics - This research explores the theoretical possibility of automatically discovering event primitives through compression, drawing parallels to historical AI theories.
- MemoryCD: Benchmarking Long-Context User Memory of LLM Agents for Lifelong Cross-Domain Personalization (viability: 7): https://sciencetostartup.com/paper/memorycd-benchmarking-long-context-user-memory-of-llm-agents-for-lifelong-cross-domain-personalization - A new benchmark for evaluating LLM agents' long-context user memory in real-world, cross-domain personalization scenarios.
- Neuro-Cognitive Reward Modeling for Human-Centered Autonomous Vehicle Control (viability: 7): https://sciencetostartup.com/paper/neuro-cognitive-reward-modeling-for-human-centered-autonomous-vehicle-control - This project integrates EEG-based cognitive feedback into reinforcement learning for autonomous vehicles to improve human-aligned decision-making and collision avoidance.
- BEVMAPMATCH: Multimodal BEV Neural Map Matching for Robust Re-Localization of Autonomous Vehicles (viability: 7): https://sciencetostartup.com/paper/bevmapmatch-multimodal-bev-neural-map-matching-for-robust-re-localization-of-autonomous-vehicles - A multimodal BEV map matching framework for robust autonomous vehicle re-localization in GNSS-denied environments.
- When Chain-of-Thought Backfires: Evaluating Prompt Sensitivity in Medical Language Models (viability: 7): https://sciencetostartup.com/paper/when-chain-of-thought-backfires-evaluating-prompt-sensitivity-in-medical-language-models - This research reveals critical prompt sensitivity issues in medical LLMs and offers reliable alternatives for accurate medical question answering.
- On the Objective and Feature Weights of Minkowski Weighted k-Means (viability: 2): https://sciencetostartup.com/paper/on-the-objective-and-feature-weights-of-minkowski-weighted-k-means - A theoretical analysis of a weighted k-means algorithm to understand feature importance in clustering.
- Adversarial-Robust Multivariate Time-Series Anomaly Detection via Joint Information Retention (viability: 7): https://sciencetostartup.com/paper/adversarial-robust-multivariate-time-series-anomaly-detection-via-joint-information-retention - A novel adversarial training framework for time-series anomaly detection that significantly improves robustness and provides interpretable sensitivity masks.
- EngineAD: A Real-World Vehicle Engine Anomaly Detection Dataset (viability: 5): https://sciencetostartup.com/paper/enginead-a-real-world-vehicle-engine-anomaly-detection-dataset - A new real-world dataset for vehicle engine anomaly detection aims to accelerate the development of robust, field-deployable solutions for the automotive industry.
- Online Learning for Dynamic Constellation Topologies (viability: 4): https://sciencetostartup.com/paper/online-learning-for-dynamic-constellation-topologies - An online learning framework for dynamically configuring satellite networks that adapt to changing topologies without assuming fixed orbital structures.
- Low-Rank-Modulated Functa: Exploring the Latent Space of Implicit Neural Representations for Interpretable Ultrasound Video Analysis (viability: 7): https://sciencetostartup.com/paper/low-rank-modulated-functa-exploring-the-latent-space-of-implicit-neural-representations-for-interpretable-ultrasound-vid - A novel framework for interpretable ultrasound video analysis that compresses videos, reveals temporal patterns, and directly identifies key cardiac frames without additional training.
- Collision-Aware Vision-Language Learning for End-to-End Driving with Multimodal Infraction Datasets (viability: 8): https://sciencetostartup.com/paper/collision-aware-vision-language-learning-for-end-to-end-driving-with-multimodal-infraction-datasets - Develop a VLAAD-enhanced module for collision-aware autonomous driving systems improving safety and reducing infractions.
- Can Small Models Reason About Legal Documents? A Comparative Study (viability: 7): https://sciencetostartup.com/paper/can-small-models-reason-about-legal-documents-a-comparative-study - This research demonstrates that smaller, cost-effective language models can achieve state-of-the-art performance on legal reasoning tasks, offering a practical alternative to expensive frontier models.
- Reinforcing Structured Chain-of-Thought for Video Understanding (viability: 7): https://sciencetostartup.com/paper/reinforcing-structured-chain-of-thought-for-video-understanding - A novel single-stage reinforcement learning framework for video understanding that improves reasoning and temporal comprehension without costly supervised fine-tuning.
- Can Vision Foundation Models Navigate? Zero-Shot Real-World Evaluation and Lessons Learned (viability: 7): https://sciencetostartup.com/paper/can-vision-foundation-models-navigate-zero-shot-real-world-evaluation-and-lessons-learned - This research evaluates vision-based navigation models for robots, revealing critical limitations in collision avoidance and environmental robustness, and will release a new evaluation dataset and codebase.
- DenseSwinV2: Channel Attentive Dual Branch CNN Transformer Learning for Cassava Leaf Disease Classification (viability: 7): https://sciencetostartup.com/paper/denseswinv2-channel-attentive-dual-branch-cnn-transformer-learning-for-cassava-leaf-disease-classification - A hybrid CNN-Transformer model achieves state-of-the-art cassava disease classification, offering practical field-level diagnosis.
- DiReCT: Disentangled Regularization of Contrastive Trajectories for Physics-Refined Video Generation (viability: 4): https://sciencetostartup.com/paper/direct-disentangled-regularization-of-contrastive-trajectories-for-physics-refined-video-generation - A post-training framework to enforce physical consistency in generated videos by disentangling semantic and physical properties of motion.
- AVDA: Autonomous Vibe Detection Authoring for Cybersecurity (viability: 5): https://sciencetostartup.com/paper/avda-autonomous-vibe-detection-authoring-for-cybersecurity - Automate the creation of cybersecurity detection logic using AI and organizational context to improve coverage and velocity.
- Density-aware Soft Context Compression with Semi-Dynamic Compression Ratio (viability: 7): https://sciencetostartup.com/paper/density-aware-soft-context-compression-with-semi-dynamic-compression-ratio - A framework for more efficient LLM context processing by intelligently compressing information based on its density, outperforming static methods.
- Personalizing Mathematical Game-based Learning for Children: A Preliminary Study (viability: 4): https://sciencetostartup.com/paper/personalizing-mathematical-game-based-learning-for-children-a-preliminary-study - An AI framework to personalize mathematical game levels for children by classifying player-generated content.
- Good Scores, Bad Data: A Metric for Multimodal Coherence (viability: 4): https://sciencetostartup.com/paper/good-scores-bad-data-a-metric-for-multimodal-coherence - A new metric to evaluate the internal consistency of multimodal AI systems, independent of downstream task performance.
- Preventing Data Leakage in EEG-Based Survival Prediction: A Two-Stage Embedding and Transformer Framework (viability: 6): https://sciencetostartup.com/paper/preventing-data-leakage-in-eeg-based-survival-prediction-a-two-stage-embedding-and-transformer-framework - A two-stage framework using CNN embeddings and Transformers to prevent data leakage in EEG-based survival prediction for comatose patients.
- Parameter-Free Dynamic Regret for Unconstrained Linear Bandits (viability: 2): https://sciencetostartup.com/paper/parameter-free-dynamic-regret-for-unconstrained-linear-bandits - A theoretical advancement in bandit algorithms for dynamic regret minimization.
- Robust Tensor-on-Tensor Regression (viability: 5): https://sciencetostartup.com/paper/robust-tensor-on-tensor-regression - A novel robust tensor regression method that handles outliers and missing values for improved tensor data analysis, demonstrated on facial attribute prediction.
- Shared Representation for 3D Pose Estimation, Action Classification, and Progress Prediction from Tactile Signals (viability: 7): https://sciencetostartup.com/paper/shared-representation-for-3d-pose-estimation-action-classification-and-progress-prediction-from-tactile-signals - A unified tactile sensing model for simultaneous 3D pose estimation, action classification, and progress prediction, overcoming vision limitations for human-robot interaction.
- Disguising Topology and Side-Channel Information through Covert Gate- and ML-Enabled IP Camouflaging (viability: 7): https://sciencetostartup.com/paper/disguising-topology-and-side-channel-information-through-covert-gate-and-ml-enabled-ip-camouflaging - A novel approach to semiconductor IP protection that uses mimetic deception to thwart reverse engineering and side-channel attacks by poisoning the adversary's data.
- Emergent Neural Automaton Policies: Learning Symbolic Structure from Visuomotor Trajectories (viability: 7): https://sciencetostartup.com/paper/emergent-neural-automaton-policies-learning-symbolic-structure-from-visuomotor-trajectories - A framework for robots to learn symbolic task structures from demonstrations, enabling efficient long-horizon manipulation with interpretable planning.
- Chasing Autonomy: Dynamic Retargeting and Control Guided RL for Performant and Controllable Humanoid Running (viability: 7): https://sciencetostartup.com/paper/chasing-autonomy-dynamic-retargeting-and-control-guided-rl-for-performant-and-controllable-humanoid-running - Enabling humanoid robots to run dynamically and autonomously by retargeting human motion and using goal-conditioned reinforcement learning.
- Decoding Defensive Coverage Responsibilities in American Football Using Factorized Attention Based Transformer Models (viability: 5): https://sciencetostartup.com/paper/decoding-defensive-coverage-responsibilities-in-american-football-using-factorized-attention-based-transformer-models - Predict individual defensive assignments and matchups in NFL games to enhance broadcast storytelling and provide strategic insights for teams.
- Data-Driven Plasticity Modeling via Acoustic Profiling (viability: 5): https://sciencetostartup.com/paper/data-driven-plasticity-modeling-via-acoustic-profiling - A data-driven framework uses acoustic signals to model and predict plastic deformation in metals, enabling early detection of material failure.
- THFM: A Unified Video Foundation Model for 4D Human Perception and Beyond (viability: 7): https://sciencetostartup.com/paper/thfm-a-unified-video-foundation-model-for-4d-human-perception-and-beyond - A unified video foundation model that performs multiple human perception tasks with state-of-the-art results, trained solely on synthetic data.
- Few Shots Text to Image Retrieval: New Benchmarking Dataset and Optimization Methods (viability: 7): https://sciencetostartup.com/paper/few-shots-text-to-image-retrieval-new-benchmarking-dataset-and-optimization-methods - A new benchmark and optimization methods for few-shot text-to-image retrieval that significantly improves performance on compositional and out-of-distribution queries.
- Polarization-Based Eye Tracking with Personalized Siamese Architectures (viability: 7): https://sciencetostartup.com/paper/polarization-based-eye-tracking-with-personalized-siamese-architectures - A personalized eye-tracking system that uses Siamese architectures to achieve accurate gaze reconstruction with significantly fewer calibration frames, outperforming traditional methods.
- World Reasoning Arena (viability: 7): https://sciencetostartup.com/paper/world-reasoning-arena - A new benchmark and dataset to evaluate and advance world models for intelligent agents, exposing current limitations in simulation and reasoning.
- Automated Quality Assessment of Blind Sweep Obstetric Ultrasound for Improved Diagnosis (viability: 7): https://sciencetostartup.com/paper/automated-quality-assessment-of-blind-sweep-obstetric-ultrasound-for-improved-diagnosis - Develop an AI-powered quality assessment tool for obstetric ultrasound to improve diagnostic accuracy in low-resource settings.
- DRiffusion: Draft-and-Refine Process Parallelizes Diffusion Models with Ease (viability: 7): https://sciencetostartup.com/paper/driffusion-draft-and-refine-process-parallelizes-diffusion-models-with-ease - Accelerate diffusion model inference by 1.4x-3.7x with a parallel draft-and-refine process, enabling interactive applications without sacrificing quality.
- Speech-Synchronized Whiteboard Generation via VLM-Driven Structured Drawing Representations (viability: 7): https://sciencetostartup.com/paper/speech-synchronized-whiteboard-generation-via-vlm-driven-structured-drawing-representations - Generate synchronized whiteboard illustrations for educational videos using a fine-tuned vision-language model and a novel structured drawing dataset.
- Seeing Through Smoke: Surgical Desmoking for Improved Visual Perception (viability: 7): https://sciencetostartup.com/paper/seeing-through-smoke-surgical-desmoking-for-improved-visual-perception - A transformer-based model that removes surgical smoke from endoscopic images to improve surgeon vision and enable downstream AI tasks.
- GUIDE: A Benchmark for Understanding and Assisting Users in Open-Ended GUI Tasks (viability: 7): https://sciencetostartup.com/paper/guide-a-benchmark-for-understanding-and-assisting-users-in-open-ended-gui-tasks - A benchmark and dataset for AI agents that understand user intent in graphical interfaces to provide collaborative assistance.
- Dynamic LIBRAS Gesture Recognition via CNN over Spatiotemporal Matrix Representation (viability: 5): https://sciencetostartup.com/paper/dynamic-libras-gesture-recognition-via-cnn-over-spatiotemporal-matrix-representation - A CNN-based system for dynamic hand gesture recognition using skeletal keypoints, achieving high accuracy for device control.
- Methods for Knowledge Graph Construction from Text Collections: Development and Applications (viability: 4): https://sciencetostartup.com/paper/methods-for-knowledge-graph-construction-from-text-collections-development-and-applications - Automates knowledge graph construction from diverse text sources using NLP and Semantic Web techniques for actionable insights.
- Why Safety Probes Catch Liars But Miss Fanatics (viability: 3): https://sciencetostartup.com/paper/why-safety-probes-catch-liars-but-miss-fanatics - This research identifies a fundamental limitation in current AI safety probes, showing they can be evaded by models that genuinely believe in harmful actions, posing a significant challenge for detecting coherent misalignment.
- On the Expressive Power of Contextual Relations in Transformers (viability: 2): https://sciencetostartup.com/paper/on-the-expressive-power-of-contextual-relations-in-transformers - A theoretical framework and architecture for understanding the expressive power of contextual relations in Transformers.
- In-Context Molecular Property Prediction with LLMs: A Blinding Study on Memorization and Knowledge Conflicts (viability: 4): https://sciencetostartup.com/paper/in-context-molecular-property-prediction-with-llms-a-blinding-study-on-memorization-and-knowledge-conflicts - This research investigates the reliability of LLMs for molecular property prediction by analyzing memorization versus genuine in-context learning, providing a framework for controlled evaluation.
- Incorporating contextual information into KGWAS for interpretable GWAS discovery (viability: 3): https://sciencetostartup.com/paper/incorporating-contextual-information-into-kgwas-for-interpretable-gwas-discovery - This research proposes a more focused knowledge graph for Genome-Wide Association Studies to improve disease mechanism discovery.
- GazeQwen: Lightweight Gaze-Conditioned LLM Modulation for Streaming Video Understanding (viability: 7): https://sciencetostartup.com/paper/gazeqwen-lightweight-gaze-conditioned-llm-modulation-for-streaming-video-understanding - Optimizing video understanding with gaze-driven LLM modulation for improved accuracy in real-time applications.
- A Compression Perspective on Simplicity Bias (viability: 3): https://sciencetostartup.com/paper/a-compression-perspective-on-simplicity-bias - A theoretical framework explains the simplicity bias in neural networks through lossless compression, predicting how data quantity influences feature selection and robustness.
- Gradient-Informed Training for Low-Resource Multilingual Speech Translation (viability: 7): https://sciencetostartup.com/paper/gradient-informed-training-for-low-resource-multilingual-speech-translation - A novel gradient-informed method to optimize multilingual speech translation for low-resource languages by intelligently sharing model components.
- Massive Parallel Deep Reinforcement Learning for Active SLAM (viability: 4): https://sciencetostartup.com/paper/massive-parallel-deep-reinforcement-learning-for-active-slam - A scalable end-to-end DRL framework for Active SLAM that significantly reduces training time and supports continuous action spaces.
- Fus3D: Decoding Consolidated 3D Geometry from Feed-forward Geometry Transformer Latents (viability: 7): https://sciencetostartup.com/paper/fus3d-decoding-consolidated-3d-geometry-from-feed-forward-geometry-transformer-latents - Generate dense 3D geometry from image collections in seconds by directly decoding latent representations of feed-forward geometry transformers.
- Understanding AI Methods for Intrusion Detection and Cryptographic Leakage (viability: 4): https://sciencetostartup.com/paper/understanding-ai-methods-for-intrusion-detection-and-cryptographic-leakage - AI models can detect network intrusions and identify cryptographic vulnerabilities, but performance degrades with unseen traffic.
- ViGoR-Bench: How Far Are Visual Generative Models From Zero-Shot Visual Reasoners? (viability: 7): https://sciencetostartup.com/paper/vigor-bench-how-far-are-visual-generative-models-from-zero-shot-visual-reasoners - ViGoR-Bench provides a rigorous evaluation framework to expose reasoning deficits in generative vision models, enabling the development of truly intelligent visual systems.
- Doctorina MedBench: End-to-End Evaluation of Agent-Based Medical AI (viability: 7): https://sciencetostartup.com/paper/doctorina-medbench-end-to-end-evaluation-of-agent-based-medical-ai - A comprehensive evaluation framework for agent-based medical AI that simulates realistic physician-patient interactions to assess clinical competence.
- Geo$^\textbf{2}$: Geometry-Guided Cross-view Geo-Localization and Image Synthesis (viability: 7): https://sciencetostartup.com/paper/geo-textbf-2-geometry-guided-cross-view-geo-localization-and-image-synthesis - A unified framework leveraging 3D geometric priors to simultaneously perform cross-view geo-localization and image synthesis, achieving state-of-the-art results.
- MAGNET: Autonomous Expert Model Generation via Decentralized Autoresearch and BitNet Training (viability: 7): https://sciencetostartup.com/paper/magnet-autonomous-expert-model-generation-via-decentralized-autoresearch-and-bitnet-training - A decentralized system for autonomously generating and training domain-expert language models on commodity hardware, enabling efficient CPU-native inference.
- ExVerus: Verus Proof Repair via Counterexample Reasoning (viability: 4): https://sciencetostartup.com/paper/exverus-verus-proof-repair-via-counterexample-reasoning - Leveraging LLMs with counterexample reasoning to improve automated formal verification of software proofs.
- RealChart2Code: Advancing Chart-to-Code Generation with Real Data and Multi-Task Evaluation (viability: 7): https://sciencetostartup.com/paper/realchart2code-advancing-chart-to-code-generation-with-real-data-and-multi-task-evaluation - A new benchmark and code release for evaluating and improving Vision-Language Models' ability to generate complex code for multi-panel visualizations from real-world data.
- Do All Vision Transformers Need Registers? A Cross-Architectural Reassessment (viability: 3): https://sciencetostartup.com/paper/do-all-vision-transformers-need-registers-a-cross-architectural-reassessment - This research investigates the necessity of 'registers' in Vision Transformers to improve attention map interpretability, finding their applicability varies across different models and sizes.
- LEMON: a foundation model for nuclear morphology in Computational Pathology (viability: 7): https://sciencetostartup.com/paper/lemon-a-foundation-model-for-nuclear-morphology-in-computational-pathology - A foundation model for learning robust single-cell image representations in computational pathology, enabling large-scale cell-level analyses.
- End-to-end Feature Alignment: A Simple CNN with Intrinsic Class Attribution (viability: 5): https://sciencetostartup.com/paper/end-to-end-feature-alignment-a-simple-cnn-with-intrinsic-class-attribution - A CNN architecture that intrinsically aligns features for enhanced interpretability and class attribution.
- Beyond identifiability: Learning causal representations with few environments and finite samples (viability: 3): https://sciencetostartup.com/paper/beyond-identifiability-learning-causal-representations-with-few-environments-and-finite-samples - Develops theoretical guarantees for learning causal representations from limited data and interventions, improving interpretability and robustness.
- ArtHOI: Taming Foundation Models for Monocular 4D Reconstruction of Hand-Articulated-Object Interactions (viability: 7): https://sciencetostartup.com/paper/arthoi-taming-foundation-models-for-monocular-4d-reconstruction-of-hand-articulated-object-interactions - A framework that reconstructs 4D hand-articulated-object interactions from single RGB videos by refining foundation model priors, addressing limitations of existing methods.
- Less Gaussians, Texture More: 4K Feed-Forward Textured Splatting (viability: 7): https://sciencetostartup.com/paper/less-gaussians-texture-more-4k-feed-forward-textured-splatting - A feed-forward 3D rendering framework that enables high-fidelity 4K novel view synthesis by predicting compact Gaussian primitives and per-primitive textures, overcoming resolution scaling barriers.
- ShotStream: Streaming Multi-Shot Video Generation for Interactive Storytelling (viability: 7): https://sciencetostartup.com/paper/shotstream-streaming-multi-shot-video-generation-for-interactive-storytelling - ShotStream enables real-time, interactive multi-shot video generation for storytelling by using a novel causal architecture with dual-cache memory and a two-stage distillation strategy.
- MuRF: Unlocking the Multi-Scale Potential of Vision Foundation Models (viability: 7): https://sciencetostartup.com/paper/murf-unlocking-the-multi-scale-potential-of-vision-foundation-models - Enhance existing vision foundation models with multi-resolution processing for improved performance across various tasks without retraining.
- RefAlign: Representation Alignment for Reference-to-Video Generation (viability: 7): https://sciencetostartup.com/paper/refalign-representation-alignment-for-reference-to-video-generation - A framework to improve identity consistency and reduce artifacts in reference-to-video generation by explicitly aligning visual features.
- Vega: Learning to Drive with Natural Language Instructions (viability: 7): https://sciencetostartup.com/paper/vega-learning-to-drive-with-natural-language-instructions - A vision-language-action model that learns to drive by following natural language instructions, enabling personalized autonomous driving experiences.
- Drive My Way: Preference Alignment of Vision-Language-Action Model for Personalized Driving (viability: 7): https://sciencetostartup.com/paper/drive-my-way-preference-alignment-of-vision-language-action-model-for-personalized-driving - A personalized vision-language-action driving framework that adapts to individual driver habits and real-time natural language instructions.
- PSDesigner: Automated Graphic Design with a Human-Like Creative Workflow (viability: 7): https://sciencetostartup.com/paper/psdesigner-automated-graphic-design-with-a-human-like-creative-workflow - An automated graphic design system that emulates human creative workflows to produce production-quality designs from user instructions.
- MegaFlow: Zero-Shot Large Displacement Optical Flow (viability: 7): https://sciencetostartup.com/paper/megaflow-zero-shot-large-displacement-optical-flow - A zero-shot optical flow model leveraging pre-trained vision priors for accurate large displacement estimation and long-range point tracking.
- How good was my shot? Quantifying Player Skill Level in Table Tennis (viability: 7): https://sciencetostartup.com/paper/how-good-was-my-shot-quantifying-player-skill-level-in-table-tennis - Quantify player skill in table tennis by learning generative models of tactical strokes and embedding them in a latent space that reflects individual characteristics.
- Training the Knowledge Base through Evidence Distillation and Write-Back Enrichment (viability: 7): https://sciencetostartup.com/paper/training-the-knowledge-base-through-evidence-distillation-and-write-back-enrichment - Enhance retrieval-augmented generation systems by dynamically distilling and enriching the knowledge base for improved factual accuracy and performance.
- Unleashing Guidance Without Classifiers for Human-Object Interaction Animation (viability: 7): https://sciencetostartup.com/paper/unleashing-guidance-without-classifiers-for-human-object-interaction-animation - A data-driven approach to generate realistic human-object interaction animations by leveraging the denoising process itself for guidance, eliminating the need for explicit classifiers.
- SlotVTG: Object-Centric Adapter for Generalizable Video Temporal Grounding (viability: 7): https://sciencetostartup.com/paper/slotvtg-object-centric-adapter-for-generalizable-video-temporal-grounding - A lightweight adapter for multimodal LLMs that enables precise object-centric temporal grounding in videos with improved out-of-domain generalization.
- BizGenEval: A Systematic Benchmark for Commercial Visual Content Generation (viability: 7): https://sciencetostartup.com/paper/bizgeneval-a-systematic-benchmark-for-commercial-visual-content-generation - A new benchmark for evaluating and improving AI image generation for commercial design tasks, revealing significant gaps in current models.
- PackForcing: Short Video Training Suffices for Long Video Sampling and Long Context Inference (viability: 8): https://sciencetostartup.com/paper/packforcing-short-video-training-suffices-for-long-video-sampling-and-long-context-inference - PackForcing: Efficient long-video generation using short-video training with reduced memory footprint and improved temporal coherence.
- PixelSmile: Toward Fine-Grained Facial Expression Editing (viability: 7): https://sciencetostartup.com/paper/pixelsmile-toward-fine-grained-facial-expression-editing - A diffusion framework for precise, controllable, and fine-grained facial expression editing with strong identity preservation.
- Back to Basics: Revisiting ASR in the Age of Voice Agents (viability: 7): https://sciencetostartup.com/paper/back-to-basics-revisiting-asr-in-the-age-of-voice-agents - WildASR provides a diagnostic benchmark and analytical tools to improve the reliability and safety of Automatic Speech Recognition systems in real-world voice agents.
- AnyHand: A Large-Scale Synthetic Dataset for RGB(-D) Hand Pose Estimation (viability: 7): https://sciencetostartup.com/paper/anyhand-a-large-scale-synthetic-dataset-for-rgb-d-hand-pose-estimation - A large-scale synthetic dataset and lightweight module to significantly improve 3D hand pose estimation accuracy and generalization for RGB and RGB-D inputs.
- SoftMimicGen: A Data Generation System for Scalable Robot Learning in Deformable Object Manipulation (viability: 7): https://sciencetostartup.com/paper/softmimicgen-a-data-generation-system-for-scalable-robot-learning-in-deformable-object-manipulation - An automated data generation pipeline for deformable object manipulation in robotics, enabling scalable robot learning.
- Natural-Language Agent Harnesses (viability: 7): https://sciencetostartup.com/paper/natural-language-agent-harnesses - Externalize agent harness logic into editable natural language for improved transferability and study, powered by an intelligent runtime.
- No Hard Negatives Required: Concept Centric Learning Leads to Compositionality without Degrading Zero-shot Capabilities of Contrastive Models (viability: 7): https://sciencetostartup.com/paper/no-hard-negatives-required-concept-centric-learning-leads-to-compositionality-without-degrading-zero-shot-capabilities-o - A new method for vision-language models that improves compositional understanding without sacrificing zero-shot performance, with code available.
- R-C2: Cycle-Consistent Reinforcement Learning Improves Multimodal Reasoning (viability: 4): https://sciencetostartup.com/paper/r-c2-cycle-consistent-reinforcement-learning-improves-multimodal-reasoning - A reinforcement learning framework that enforces cross-modal consistency to improve multimodal reasoning accuracy.
- Agent Factories for High Level Synthesis: How Far Can General-Purpose Coding Agents Go in Hardware Optimization? (viability: 7): https://sciencetostartup.com/paper/agent-factories-for-high-level-synthesis-how-far-can-general-purpose-coding-agents-go-in-hardware-optimization - Autonomous coding agents that significantly accelerate hardware design optimization by intelligently decomposing and exploring complex configurations.
- Out of Sight but Not Out of Mind: Hybrid Memory for Dynamic Video World Models (viability: 7): https://sciencetostartup.com/paper/out-of-sight-but-not-out-of-mind-hybrid-memory-for-dynamic-video-world-models - A novel hybrid memory system for video world models that tracks dynamic subjects even when they are out of view, ensuring motion continuity and realistic simulation.
- Seeing to Ground: Visual Attention for Hallucination-Resilient MDLLMs (viability: 7): https://sciencetostartup.com/paper/seeing-to-ground-visual-attention-for-hallucination-resilient-mdllms - A training-free framework that uses visual attention to eliminate hallucinations in multimodal large language models by ensuring text generation is visually grounded.
- TRACE: Object Motion Editing in Videos with First-Frame Trajectory Guidance (viability: 7): https://sciencetostartup.com/paper/trace-object-motion-editing-in-videos-with-first-frame-trajectory-guidance - A framework for easily editing object trajectories in videos by specifying a path in a single frame, producing coherent and realistic edits.
- Wan-Weaver: Interleaved Multi-modal Generation via Decoupled Training (viability: 7): https://sciencetostartup.com/paper/wan-weaver-interleaved-multi-modal-generation-via-decoupled-training - Wan-Weaver enables interleaved text and image generation by decoupling planning and visualization, achieving state-of-the-art performance without real interleaved data.
- S2D2: Fast Decoding for Diffusion LLMs via Training-Free Self-Speculation (viability: 7): https://sciencetostartup.com/paper/s2d2-fast-decoding-for-diffusion-llms-via-training-free-self-speculation - Accelerate LLM generation by training-free self-speculative decoding for diffusion models, achieving significant speedups with improved accuracy.
- Neural Network Conversion of Machine Learning Pipelines (viability: 3): https://sciencetostartup.com/paper/neural-network-conversion-of-machine-learning-pipelines - This paper explores transferring knowledge from non-neural machine learning pipelines to neural networks to create unified inference engines for multiple ML tasks.
- The Kitchen Loop: User-Spec-Driven Development for a Self-Evolving Codebase (viability: 7): https://sciencetostartup.com/paper/the-kitchen-loop-user-spec-driven-development-for-a-self-evolving-codebase - A framework for autonomous, self-evolving software development using LLM agents and rigorous testing to ensure quality and prevent regressions.
- A Unified Memory Perspective for Probabilistic Trustworthy AI (viability: 3): https://sciencetostartup.com/paper/a-unified-memory-perspective-for-probabilistic-trustworthy-ai - A novel memory architecture to improve the efficiency and scalability of probabilistic computations for trustworthy AI.
- LEMMA: Laplacian pyramids for Efficient Marine SeMAntic Segmentation (viability: 8): https://sciencetostartup.com/paper/lemma-laplacian-pyramids-for-efficient-marine-semantic-segmentation - Deploy a lightweight semantic segmentation model for real-time marine environment analysis on resource-constrained devices.
- Intelligent Navigation and Obstacle-Aware Fabrication for Mobile Additive Manufacturing Systems (viability: 4): https://sciencetostartup.com/paper/intelligent-navigation-and-obstacle-aware-fabrication-for-mobile-additive-manufacturing-systems - A universal platform for mobile additive manufacturing robots that integrates navigation and material deposition for adaptable production in dynamic environments.
- On Neural Scaling Laws for Weather Emulation through Continual Training (viability: 4): https://sciencetostartup.com/paper/on-neural-scaling-laws-for-weather-emulation-through-continual-training - Develops a framework for understanding and optimizing neural network scaling for weather forecasting, enabling more efficient resource allocation and improved prediction accuracy.
- Just Zoom In: Cross-View Geo-Localization via Autoregressive Zooming (viability: 7): https://sciencetostartup.com/paper/just-zoom-in-cross-view-geo-localization-via-autoregressive-zooming - Autoregressive zooming over satellite imagery for precise GPS-denied localization, outperforming existing retrieval methods.
- Persistent Robot World Models: Stabilizing Multi-Step Rollouts via Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/persistent-robot-world-models-stabilizing-multi-step-rollouts-via-reinforcement-learning - Reinforcement learning for robot world models to enable stable, long-term visual prediction for simulation.
- Self-Improvement of Large Language Models: A Technical Overview and Future Outlook (viability: 3): https://sciencetostartup.com/paper/self-improvement-of-large-language-models-a-technical-overview-and-future-outlook - A framework for organizing techniques in self-improving large language models, exploring autonomous data generation, evaluation, and refinement.
- Measuring What Matters -- or What's Convenient?: Robustness of LLM-Based Scoring Systems to Construct-Irrelevant Factors (viability: 4): https://sciencetostartup.com/paper/measuring-what-matters-or-what-s-convenient-robustness-of-llm-based-scoring-systems-to-construct-irrelevant-factors - This research evaluates the robustness of LLM-based essay scoring systems to irrelevant factors, suggesting potential for more reliable automated assessment.
- Longitudinal Digital Phenotyping for Early Cognitive-Motor Screening (viability: 7): https://sciencetostartup.com/paper/longitudinal-digital-phenotyping-for-early-cognitive-motor-screening - An AI framework using tablet interactions to continuously screen for early cognitive-motor developmental issues in children, identifying persistent deficits for timely intervention.
- Can Users Specify Driving Speed? Bench2Drive-Speed: Benchmark and Baselines for Desired-Speed Conditioned Autonomous Driving (viability: 7): https://sciencetostartup.com/paper/can-users-specify-driving-speed-bench2drive-speed-benchmark-and-baselines-for-desired-speed-conditioned-autonomous-drivi - Enabling customizable speed and overtaking for autonomous vehicles by leveraging existing driving data.
- Uncertainty-Guided Label Rebalancing for CPS Safety Monitoring (viability: 7): https://sciencetostartup.com/paper/uncertainty-guided-label-rebalancing-for-cps-safety-monitoring - A novel approach to rebalance imbalanced safety data in Cyber-Physical Systems by leveraging behavioral uncertainty, significantly improving safety predictor performance.
- Fast-dVLA: Accelerating Discrete Diffusion VLA to Real-Time Performance (viability: 7): https://sciencetostartup.com/paper/fast-dvla-accelerating-discrete-diffusion-vla-to-real-time-performance - Accelerate robot adaptation by decoupling and merging training objectives for enhanced capabilities with reduced computational cost.
- A Mentalistic Interface for Probing Folk-Psychological Attribution to Non-Humanoid Robots (viability: 3): https://sciencetostartup.com/paper/a-mentalistic-interface-for-probing-folk-psychological-attribution-to-non-humanoid-robots - A research platform to study how people attribute mental states to robots using LLM-generated explanations.
- RenoBench: A Citation Parsing Benchmark (viability: 5): https://sciencetostartup.com/paper/renobench-a-citation-parsing-benchmark - A new benchmark and dataset for accurate, machine-readable citation parsing to improve scholarly infrastructure.
- Beyond Via: Analysis and Estimation of the Impact of Large Language Models in Academic Papers (viability: 4): https://sciencetostartup.com/paper/beyond-via-analysis-and-estimation-of-the-impact-of-large-language-models-in-academic-papers - This research analyzes shifts in academic writing driven by LLMs, developing a method to detect LLM-generated text and understand their impact on language.
- Designing Any Imaging System from Natural Language: Agent-Constrained Composition over a Finite Primitive Basis (viability: 7): https://sciencetostartup.com/paper/designing-any-imaging-system-from-natural-language-agent-constrained-composition-over-a-finite-primitive-basis - Automate the design of complex imaging systems from natural language, overcoming expertise bottlenecks and accelerating scientific prototyping.
- Anchored-Branched Steady-state WInd Flow Transformer (AB-SWIFT): a metamodel for 3D atmospheric flow in urban environments (viability: 7): https://sciencetostartup.com/paper/anchored-branched-steady-state-wind-flow-transformer-ab-swift-a-metamodel-for-3d-atmospheric-flow-in-urban-environments - A transformer-based metamodel for accurate and efficient 3D atmospheric flow simulation in urban environments, outperforming existing methods.
- Is Mathematical Problem-Solving Expertise in Large Language Models Associated with Assessment Performance? (viability: 5): https://sciencetostartup.com/paper/is-mathematical-problem-solving-expertise-in-large-language-models-associated-with-assessment-performance - This research investigates the link between LLM math problem-solving ability and their accuracy in assessing student reasoning steps, suggesting potential for improved AI tutors.
- LanteRn: Latent Visual Structured Reasoning (viability: 7): https://sciencetostartup.com/paper/lantern-latent-visual-structured-reasoning - A framework enabling large multimodal models to perform efficient visual reasoning directly in latent space, improving fine-grained understanding.
- Visual or Textual: Effects of Explanation Format and Personal Characteristics on the Perception of Explanations in an Educational Recommender System (viability: 4): https://sciencetostartup.com/paper/visual-or-textual-effects-of-explanation-format-and-personal-characteristics-on-the-perception-of-explanations-in-an-edu - This research investigates how visual versus textual explanations impact user perception in educational recommender systems, offering design guidelines.
- Accurate Surface and Reflectance Modelling from 3D Radar Data with Neural Radiance Fields (viability: 4): https://sciencetostartup.com/paper/accurate-surface-and-reflectance-modelling-from-3d-radar-data-with-neural-radiance-fields - A neural implicit method for improved 3D surface and reflectance modeling from sparse and noisy radar data.
- The Geometry of Efficient Nonconvex Sampling (viability: 1): https://sciencetostartup.com/paper/the-geometry-of-efficient-nonconvex-sampling - Develops a theoretical algorithm for uniform sampling from complex geometric shapes, with potential applications in statistical inference and machine learning.
- PICon: A Multi-Turn Interrogation Framework for Evaluating Persona Agent Consistency (viability: 6): https://sciencetostartup.com/paper/picon-a-multi-turn-interrogation-framework-for-evaluating-persona-agent-consistency - A framework for validating persona agent consistency across multiple dimensions for reliable human simulation.
- Social Hippocampus Memory Learning (viability: 7): https://sciencetostartup.com/paper/social-hippocampus-memory-learning - A memory-centric social machine learning framework enabling privacy-preserving collaboration among heterogeneous agents by sharing abstracted knowledge instead of model parameters.
- Demographic Fairness in Multimodal LLMs: A Benchmark of Gender and Ethnicity Bias in Face Verification (viability: 4): https://sciencetostartup.com/paper/demographic-fairness-in-multimodal-llms-a-benchmark-of-gender-and-ethnicity-bias-in-face-verification - This research benchmarks demographic fairness in multimodal LLMs for face verification, revealing significant biases and performance disparities across different models and demographic groups.
- DeepFAN, a transformer-based deep learning model for human-artificial intelligence collaborative assessment of incidental pulmonary nodules in CT scans: a multi-reader, multi-case trial (viability: 7): https://sciencetostartup.com/paper/deepfan-a-transformer-based-deep-learning-model-for-human-artificial-intelligence-collaborative-assessment-of-incidental - A transformer-based AI model that significantly improves radiologist performance in assessing pulmonary nodules, with clinical trial validation.
- Spatiotemporal System Forecasting with Irregular Time Steps via Masked Autoencoder (viability: 7): https://sciencetostartup.com/paper/spatiotemporal-system-forecasting-with-irregular-time-steps-via-masked-autoencoder - A masked autoencoder for accurate spatiotemporal forecasting of systems with irregular time steps, applicable to climate and fluid dynamics.
- Towards Generalizable Robotic Data Flywheel: High-Dimensional Factorization and Composition (viability: 7): https://sciencetostartup.com/paper/towards-generalizable-robotic-data-flywheel-high-dimensional-factorization-and-composition - A framework for structured data factorization and iterative learning to significantly improve robotic model generalization with fewer demonstrations.
- UNIC: Neural Garment Deformation Field for Real-time Clothed Character Animation (viability: 7): https://sciencetostartup.com/paper/unic-neural-garment-deformation-field-for-real-time-clothed-character-animation - Real-time neural deformation field for animating complex garment meshes in virtual environments.
- The Rules-and-Facts Model for Simultaneous Generalization and Memorization in Neural Networks (viability: 1): https://sciencetostartup.com/paper/the-rules-and-facts-model-for-simultaneous-generalization-and-memorization-in-neural-networks - A theoretical model to understand how neural networks learn rules and memorize facts simultaneously.
- Hierarchy-Guided Multimodal Representation Learning for Taxonomic Inference (viability: 7): https://sciencetostartup.com/paper/hierarchy-guided-multimodal-representation-learning-for-taxonomic-inference - A hierarchy-aware multimodal AI that accurately identifies species from imperfect image and DNA data, crucial for conservation and environmental monitoring.
- Cooperative Deep Reinforcement Learning for Fair RIS Allocation (viability: 5): https://sciencetostartup.com/paper/cooperative-deep-reinforcement-learning-for-fair-ris-allocation - A cooperative multi-agent reinforcement learning system for fair resource allocation in wireless networks using reconfigurable intelligent surfaces.
- TAAC: A gate into Trustable Audio Affective Computing (viability: 7): https://sciencetostartup.com/paper/taac-a-gate-into-trustable-audio-affective-computing - A framework for trustable audio-based depression diagnosis that encrypts sensitive user identity information while maintaining high diagnostic accuracy.
- Are LLMs Overkill for Databases?: A Study on the Finiteness of SQL (viability: 4): https://sciencetostartup.com/paper/are-llms-overkill-for-databases-a-study-on-the-finiteness-of-sql - This research suggests that simpler, template-based approaches may be more efficient and auditable than LLMs for generating SQL queries in many practical database scenarios.
- GeoHeight-Bench: Towards Height-Aware Multimodal Reasoning in Remote Sensing (viability: 7): https://sciencetostartup.com/paper/geoheight-bench-towards-height-aware-multimodal-reasoning-in-remote-sensing - A new benchmark and baseline model for height-aware reasoning in remote sensing, addressing a critical gap in current multimodal models.
- Revisiting On-Policy Distillation: Empirical Failure Modes and Simple Fixes (viability: 4): https://sciencetostartup.com/paper/revisiting-on-policy-distillation-empirical-failure-modes-and-simple-fixes - Improves on-policy distillation for LLMs by addressing empirical failure modes to achieve more stable optimization and better downstream performance.
- An Integrative Genome-Scale Metabolic Modeling and Machine Learning Framework for Predicting and Optimizing Biofuel-Relevant Biomass Production in Saccharomyces cerevisiae (viability: 7): https://sciencetostartup.com/paper/an-integrative-genome-scale-metabolic-modeling-and-machine-learning-framework-for-predicting-and-optimizing-biofuel-rele - A machine learning framework integrating metabolic modeling and optimization to predict and dramatically enhance biofuel-relevant biomass production in yeast.
- Towards Comprehensive Real-Time Scene Understanding in Ophthalmic Surgery through Multimodal Image Fusion (viability: 7): https://sciencetostartup.com/paper/towards-comprehensive-real-time-scene-understanding-in-ophthalmic-surgery-through-multimodal-image-fusion - A multimodal AI system fuses operating microscope and OCT images for real-time, precise instrument tracking and tool-tissue distance estimation in ophthalmic surgery.
- Voxtral TTS (viability: 7): https://sciencetostartup.com/paper/voxtral-tts - A multilingual text-to-speech model that clones voices from just 3 seconds of audio, outperforming existing solutions in naturalness and expressivity.
- Towards Embodied AI with MuscleMimic: Unlocking full-body musculoskeletal motor learning at scale (viability: 7): https://sciencetostartup.com/paper/towards-embodied-ai-with-musclemimic-unlocking-full-body-musculoskeletal-motor-learning-at-scale - A framework for scalable, physiologically realistic full-body musculoskeletal motor learning, enabling rapid training and fine-tuning of human-like movements.
- PAWS: Perception of Articulation in the Wild at Scale from Egocentric Videos (viability: 7): https://sciencetostartup.com/paper/paws-perception-of-articulation-in-the-wild-at-scale-from-egocentric-videos - Extracts 3D object articulations from unannotated egocentric videos to enable robotics and animation applications.
- Missing-Aware Multimodal Fusion for Unified Microservice Incident Management (viability: 7): https://sciencetostartup.com/paper/missing-aware-multimodal-fusion-for-unified-microservice-incident-management - A self-supervised framework for robust microservice incident management that handles missing data, improving anomaly detection, failure triage, and root cause localization.
- Humans vs Vision-Language Models: A Unified Measure of Narrative Coherence (viability: 4): https://sciencetostartup.com/paper/humans-vs-vision-language-models-a-unified-measure-of-narrative-coherence - A new metric to evaluate narrative coherence in vision-language models, revealing systematic differences from human storytelling.
- Insights on back marking for the automated identification of animals (viability: 3): https://sciencetostartup.com/paper/insights-on-back-marking-for-the-automated-identification-of-animals - Optimizing animal back mark design for accurate machine learning-based individual monitoring.
- BFMD: A Full-Match Badminton Dense Dataset for Dense Shot Captioning (viability: 5): https://sciencetostartup.com/paper/bfmd-a-full-match-badminton-dense-dataset-for-dense-shot-captioning - A new dataset and multimodal captioning framework for understanding tactical dynamics in full badminton matches.
- Synchronous Signal Temporal Logic for Decidable Verification of Cyber-Physical Systems (viability: 3): https://sciencetostartup.com/paper/synchronous-signal-temporal-logic-for-decidable-verification-of-cyber-physical-systems - A decidable fragment of Signal Temporal Logic for static verification of safety-critical cyber-physical systems, enabling model checking with SPIN.
- Adaptive Subspace Modeling With Functional Tucker Decomposition (viability: 4): https://sciencetostartup.com/paper/adaptive-subspace-modeling-with-functional-tucker-decomposition - A functional Tucker decomposition method that embeds continuity constraints for improved multidimensional data analysis in classification tasks.
- Beyond the Golden Data: Resolving the Motion-Vision Quality Dilemma via Timestep Selective Training (viability: 7): https://sciencetostartup.com/paper/beyond-the-golden-data-resolving-the-motion-vision-quality-dilemma-via-timestep-selective-training - A novel training method for video generation models that decouples motion and visual quality, enabling superior performance even with imperfect data.
- CHIRP dataset: towards long-term, individual-level, behavioral monitoring of bird populations in the wild (viability: 7): https://sciencetostartup.com/paper/chirp-dataset-towards-long-term-individual-level-behavioral-monitoring-of-bird-populations-in-the-wild - A new dataset and method for individual bird re-identification and behavior monitoring in the wild, outperforming state-of-the-art.
- NERO-Net: A Neuroevolutionary Approach for the Design of Adversarially Robust CNNs (viability: 3): https://sciencetostartup.com/paper/nero-net-a-neuroevolutionary-approach-for-the-design-of-adversarially-robust-cnns - A neuroevolutionary approach to design CNNs that are intrinsically robust to adversarial attacks, improving post-attack accuracy without sacrificing clean accuracy.
- Challenges in Hyperspectral Imaging for Autonomous Driving: The HSI-Drive Case (viability: 4): https://sciencetostartup.com/paper/challenges-in-hyperspectral-imaging-for-autonomous-driving-the-hsi-drive-case - Developing custom vision algorithms to overcome hyperspectral imaging challenges for real-time autonomous driving applications.
- Lightweight GenAI for Network Traffic Synthesis: Fidelity, Augmentation, and Classification (viability: 7): https://sciencetostartup.com/paper/lightweight-genai-for-network-traffic-synthesis-fidelity-augmentation-and-classification - Lightweight GenAI models generate realistic network traffic for data augmentation and classification, overcoming data scarcity and privacy concerns with high fidelity and efficiency.
- RealRestorer: Towards Generalizable Real-World Image Restoration with Large-Scale Image Editing Models (viability: 7): https://sciencetostartup.com/paper/realrestorer-towards-generalizable-real-world-image-restoration-with-large-scale-image-editing-models - A state-of-the-art open-source image restoration model trained on a large-scale dataset, outperforming existing methods and closing the gap with closed-source alternatives.
- An Experimental Comparison of the Most Popular Approaches to Fake News Detection (viability: 4): https://sciencetostartup.com/paper/an-experimental-comparison-of-the-most-popular-approaches-to-fake-news-detection - This research critically assesses 12 fake news detection approaches, highlighting LLMs' potential for zero/few-shot learning in text-only English scenarios.
- Unveiling the Resilience of LLM-Enhanced Search Engines against Black-Hat SEO Manipulation (viability: 7): https://sciencetostartup.com/paper/unveiling-the-resilience-of-llm-enhanced-search-engines-against-black-hat-seo-manipulation - This research benchmarks and identifies vulnerabilities in LLM-enhanced search engines against sophisticated SEO manipulation, offering insights for building more resilient AI search systems.
- Knowledge-Guided Failure Prediction: Detecting When Object Detectors Miss Safety-Critical Objects (viability: 7): https://sciencetostartup.com/paper/knowledge-guided-failure-prediction-detecting-when-object-detectors-miss-safety-critical-objects - A runtime monitoring framework for object detectors that predicts failures by measuring semantic misalignment between internal features and foundation model embeddings, significantly improving recall for safety-critical objects.
- EcoThink: A Green Adaptive Inference Framework for Sustainable and Accessible Agents (viability: 7): https://sciencetostartup.com/paper/ecothink-a-green-adaptive-inference-framework-for-sustainable-and-accessible-agents - EcoThink is an adaptive inference framework that significantly reduces LLM energy consumption for generative AI agents without sacrificing performance, enabling sustainable and accessible AI.
- Interpretable PM2.5 Forecasting for Urban Air Quality: A Comparative Study of Operational Time-Series Models (viability: 4): https://sciencetostartup.com/paper/interpretable-pm2-5-forecasting-for-urban-air-quality-a-comparative-study-of-operational-time-series-models - Lightweight and interpretable time-series models offer competitive performance for urban air quality forecasting, balancing accuracy and computational efficiency.
- AdaSFormer: Adaptive Serialized Transformers for Monocular Semantic Scene Completion from Indoor Environments (viability: 7): https://sciencetostartup.com/paper/adasformer-adaptive-serialized-transformers-for-monocular-semantic-scene-completion-from-indoor-environments - A novel transformer architecture for accurate 3D scene reconstruction from single indoor images, outperforming existing methods.
- Translation Asymmetry in LLMs as a Data Augmentation Factor: A Case Study for 6 Romansh Language Varieties (viability: 4): https://sciencetostartup.com/paper/translation-asymmetry-in-llms-as-a-data-augmentation-factor-a-case-study-for-6-romansh-language-varieties - A novel data augmentation strategy for low-resource machine translation that aligns with resource gradients, outperforming existing models for specific language varieties.
- LILAC: Language-Conditioned Object-Centric Optical Flow for Open-Loop Trajectory Generation (viability: 7): https://sciencetostartup.com/paper/lilac-language-conditioned-object-centric-optical-flow-for-open-loop-trajectory-generation - LILAC translates natural language into robotic actions using optical flow for efficient task execution.
- Retraining as Approximate Bayesian Inference (viability: 2): https://sciencetostartup.com/paper/retraining-as-approximate-bayesian-inference - A theoretical framework for understanding model retraining as approximate Bayesian inference to minimize 'learning debt'.
- How Class Ontology and Data Scale Affect Audio Transfer Learning (viability: 3): https://sciencetostartup.com/paper/how-class-ontology-and-data-scale-affect-audio-transfer-learning - This research investigates the impact of data scale and class ontology on audio transfer learning performance, identifying key factors for effective pre-training.
- Causal-INSIGHT: Probing Temporal Models to Extract Causal Structure (viability: 7): https://sciencetostartup.com/paper/causal-insight-probing-temporal-models-to-extract-causal-structure - A model-agnostic framework to extract directed temporal influence structures from pre-trained time series models, improving interpretability and delay localization.
- Not a fragment, but the whole: Map-based evaluation of data-driven Fire Danger Index models (viability: 4): https://sciencetostartup.com/paper/not-a-fragment-but-the-whole-map-based-evaluation-of-data-driven-fire-danger-index-models - A novel evaluation method for wildfire prediction models that prioritizes operational decision-making by accurately assessing fire activity and minimizing false alarms.
- GridVAD: Open-Set Video Anomaly Detection via Spatial Reasoning over Stratified Frame Grids (viability: 7): https://sciencetostartup.com/paper/gridvad-open-set-video-anomaly-detection-via-spatial-reasoning-over-stratified-frame-grids - GridVAD leverages natural-language anomaly proposals for zero-shot video anomaly detection, achieving state-of-the-art performance without task-specific training.
- Residual-as-Teacher: Mitigating Bias Propagation in Student--Teacher Estimation (viability: 3): https://sciencetostartup.com/paper/residual-as-teacher-mitigating-bias-propagation-in-student-teacher-estimation - A theoretical framework to reduce bias propagation in student-teacher AI models.
- Maximum Entropy Behavior Exploration for Sim2Real Zero-Shot Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/maximum-entropy-behavior-exploration-for-sim2real-zero-shot-reinforcement-learning - A zero-shot reinforcement learning algorithm that enables direct deployment of policies to real-world robots without finetuning by maximizing exploration entropy.
- CIAR: Interval-based Collaborative Decoding for Image Generation Acceleration (viability: 7): https://sciencetostartup.com/paper/ciar-interval-based-collaborative-decoding-for-image-generation-acceleration - CIAR accelerates image generation by intelligently offloading computation to the device, reducing cloud requests by 70% and achieving 2.18x speed-up.
- Temporally Decoupled Diffusion Planning for Autonomous Driving (viability: 7): https://sciencetostartup.com/paper/temporally-decoupled-diffusion-planning-for-autonomous-driving - A diffusion model that generates safer and more goal-aligned autonomous driving trajectories by decoupling near-term and long-term planning.
- Cross-Model Disagreement as a Label-Free Correctness Signal (viability: 7): https://sciencetostartup.com/paper/cross-model-disagreement-as-a-label-free-correctness-signal - A training-free system that uses a second language model to detect confident errors in a primary model, improving LLM safety and reliability.
- DC-Reg: Globally Optimal Point Cloud Registration via Tight Bounding with Difference of Convex Programming (viability: 7): https://sciencetostartup.com/paper/dc-reg-globally-optimal-point-cloud-registration-via-tight-bounding-with-difference-of-convex-programming - A globally optimal point cloud registration framework that significantly tightens bounding for faster and more robust results, even with extreme noise and outliers.
- From Manipulation to Mistrust: Explaining Diverse Micro-Video Misinformation for Robust Debunking in the Wild (viability: 7): https://sciencetostartup.com/paper/from-manipulation-to-mistrust-explaining-diverse-micro-video-misinformation-for-robust-debunking-in-the-wild - A multi-agent reasoning framework and benchmark for explaining diverse micro-video misinformation, enabling robust debunking.
- Navigating the Prompt Space: Improving LLM Classification of Social Science Texts Through Prompt Engineering (viability: 4): https://sciencetostartup.com/paper/navigating-the-prompt-space-improving-llm-classification-of-social-science-texts-through-prompt-engineering - This paper explores prompt engineering techniques to improve LLM classification accuracy for social science texts, finding that minimal context increases yield the best results.
- VideoWeaver: Multimodal Multi-View Video-to-Video Transfer for Embodied Agents (viability: 7): https://sciencetostartup.com/paper/videoweaver-multimodal-multi-view-video-to-video-transfer-for-embodied-agents - VideoWeaver enables robot policies to transfer to new environments by performing consistent multi-view video translation, overcoming limitations of single-view approaches.
- TAPO: Translation Augmented Policy Optimization for Multilingual Mathematical Reasoning (viability: 7): https://sciencetostartup.com/paper/tapo-translation-augmented-policy-optimization-for-multilingual-mathematical-reasoning - A reinforcement learning framework that uses English as a pivot to significantly improve multilingual mathematical reasoning in LLMs by decoupling understanding from reasoning.
- Visualizing Impedance Control in Augmented Reality for Teleoperation: Design and User Evaluation (viability: 5): https://sciencetostartup.com/paper/visualizing-impedance-control-in-augmented-reality-for-teleoperation-design-and-user-evaluation - Augmented reality visualization of impedance control for teleoperation improves force-critical manipulation tasks by providing intuitive, real-time feedback without haptic hardware.
- Modernising Reinforcement Learning-Based Navigation for Embodied Semantic Scene Graph Generation (viability: 3): https://sciencetostartup.com/paper/modernising-reinforcement-learning-based-navigation-for-embodied-semantic-scene-graph-generation - This research modernizes navigation for embodied agents to generate semantic scene graphs more efficiently by improving decision-making policies and action formulations.
- Decidable By Construction: Design-Time Verification for Trustworthy AI (viability: 2): https://sciencetostartup.com/paper/decidable-by-construction-design-time-verification-for-trustworthy-ai - A theoretical framework for designing AI models that are provably correct and stable before training, eliminating post-hoc verification overhead.
- Beyond Content Safety: Real-Time Monitoring for Reasoning Vulnerabilities in Large Language Models (viability: 7): https://sciencetostartup.com/paper/beyond-content-safety-real-time-monitoring-for-reasoning-vulnerabilities-in-large-language-models - A real-time monitor for LLM reasoning safety that detects and interrupts unsafe reasoning steps, improving security and reliability.
- HiSpatial: Taming Hierarchical 3D Spatial Understanding in Vision-Language Models (viability: 8): https://sciencetostartup.com/paper/hispatial-taming-hierarchical-3d-spatial-understanding-in-vision-language-models - HiSpatial provides state-of-the-art 3D spatial intelligence for vision-language models, suitable for enhancing autonomous systems and smart environments.
- MMaDA-VLA: Large Diffusion Vision-Language-Action Model with Unified Multi-Modal Instruction and Generation (viability: 7): https://sciencetostartup.com/paper/mmada-vla-large-diffusion-vision-language-action-model-with-unified-multi-modal-instruction-and-generation - A unified diffusion model for robot manipulation that generates future observations and actions in parallel, improving long-horizon consistency and performance.
- System Design for Maintaining Internal State Consistency in Long-Horizon Robotic Tabletop Games (viability: 5): https://sciencetostartup.com/paper/system-design-for-maintaining-internal-state-consistency-in-long-horizon-robotic-tabletop-games - A system design for robots to maintain internal state consistency in long-horizon tabletop games, demonstrated with Mahjong.
- Shape and Substance: Dual-Layer Side-Channel Attacks on Local Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/shape-and-substance-dual-layer-side-channel-attacks-on-local-vision-language-models - Exploit algorithmic side-channels in on-device Vision-Language Models to infer sensitive user contexts, enabling proactive security measures for Edge AI.
- LaMP: Learning Vision-Language-Action Policies with 3D Scene Flow as Latent Motion Prior (viability: 8): https://sciencetostartup.com/paper/lamp-learning-vision-language-action-policies-with-3d-scene-flow-as-latent-motion-prior - LaMP offers a cutting-edge robotic manipulation framework leveraging 3D scene flow for enhanced vision-language-action alignment, outperforming existing models by integrating geometric foresight in control policies.
- PMT: Plain Mask Transformer for Image and Video Segmentation with Frozen Vision Encoders (viability: 7): https://sciencetostartup.com/paper/pmt-plain-mask-transformer-for-image-and-video-segmentation-with-frozen-vision-encoders - A fast Transformer-based segmentation decoder that leverages frozen vision foundation models for efficient image and video segmentation.
- A Causal Framework for Evaluating ICU Discharge Strategies (viability: 5): https://sciencetostartup.com/paper/a-causal-framework-for-evaluating-icu-discharge-strategies - A causal inference framework to optimize ICU patient discharge strategies, demonstrated on MIMIC-IV.
- UMBRELLA: Uncertainty-aware Multi-robot Reactive Coordination under Dynamic Temporal Logic Tasks (viability: 7): https://sciencetostartup.com/paper/umbrella-uncertainty-aware-multi-robot-reactive-coordination-under-dynamic-temporal-logic-tasks - A framework for multi-robot coordination that handles dynamic tasks and uncertainties, improving efficiency and reliability.
- ALPS: Automated Least-Privilege Enforcement for Securing Serverless Functions (viability: 7): https://sciencetostartup.com/paper/alps-automated-least-privilege-enforcement-for-securing-serverless-functions - ALPS automates least-privilege enforcement for serverless functions, reducing security risks and improving permission management across cloud providers.
- FSGNet: A Frequency-Aware and Semantic Guidance Network for Infrared Small Target Detection (viability: 7): https://sciencetostartup.com/paper/fsgnet-a-frequency-aware-and-semantic-guidance-network-for-infrared-small-target-detection - A lightweight and effective network for infrared small target detection that uses frequency and semantic guidance to improve precision and efficiency.
- Multimodal Dataset Distillation via Phased Teacher Models (viability: 7): https://sciencetostartup.com/paper/multimodal-dataset-distillation-via-phased-teacher-models - A phased distillation framework that creates compact synthetic datasets from large image-text data, significantly improving student model performance and reducing storage.
- GlowQ: Group-Shared LOw-Rank Approximation for Quantized LLMs (viability: 5): https://sciencetostartup.com/paper/glowq-group-shared-low-rank-approximation-for-quantized-llms - A novel low-rank approximation technique for quantized LLMs that reduces overhead and improves accuracy by selectively restoring layers.
- CLIP-RD: Relational Distillation for Efficient CLIP Knowledge Distillation (viability: 7): https://sciencetostartup.com/paper/clip-rd-relational-distillation-for-efficient-clip-knowledge-distillation - A novel relational distillation framework to create efficient, lightweight CLIP models that preserve structural relationships, outperforming existing methods.
- IntentReact: Guiding Reactive Object-Centric Navigation via Topological Intent (viability: 7): https://sciencetostartup.com/paper/intentreact-guiding-reactive-object-centric-navigation-via-topological-intent - A novel framework for robot navigation that bridges global topological planning with local perception for more efficient and robust object-goal navigation.
- Does Structured Intent Representation Generalize? A Cross-Language, Cross-Model Empirical Study of 5W3H Prompting (viability: 4): https://sciencetostartup.com/paper/does-structured-intent-representation-generalize-a-cross-language-cross-model-empirical-study-of-5w3h-prompting - A framework for structured intent representation in human-AI interaction that improves goal alignment and accessibility across languages and models.
- Supercharging Federated Intelligence Retrieval (viability: 4): https://sciencetostartup.com/paper/supercharging-federated-intelligence-retrieval - A secure federated RAG system for distributed private data retrieval and confidential LLM inference.
- Hessian-informed machine learning interatomic potential towards bridging theory and experiments (viability: 4): https://sciencetostartup.com/paper/hessian-informed-machine-learning-interatomic-potential-towards-bridging-theory-and-experiments - Develops a novel machine learning interatomic potential that accurately predicts material properties by incorporating local curvature information, bridging the gap between theoretical simulations and experimental observations.
- A Distribution-to-Distribution Neural Probabilistic Forecasting Framework for Dynamical Systems (viability: 5): https://sciencetostartup.com/paper/a-distribution-to-distribution-neural-probabilistic-forecasting-framework-for-dynamical-systems - A neural framework that directly forecasts probability distributions for dynamical systems, offering improved uncertainty quantification.
- Integrating Deep RL and Bayesian Inference for ObjectNav in Mobile Robotics (viability: 4): https://sciencetostartup.com/paper/integrating-deep-rl-and-bayesian-inference-for-objectnav-in-mobile-robotics - A hybrid framework for mobile robots that combines Bayesian inference with deep reinforcement learning to improve autonomous object search efficiency and reliability in partially observable indoor environments.
- Bayesian Learning-Enhanced Navigation with Deep Smoothing for Inertial-Aided Navigation (viability: 7): https://sciencetostartup.com/paper/bayesian-learning-enhanced-navigation-with-deep-smoothing-for-inertial-aided-navigation - A data-driven framework that uses Bayesian learning and deep smoothing to significantly improve the accuracy of inertial-aided navigation systems for robotics and mapping.
- InstanceAnimator: Multi-Instance Sketch Video Colorization (viability: 7): https://sciencetostartup.com/paper/instanceanimator-multi-instance-sketch-video-colorization - A novel framework for flexible and controllable multi-instance sketch video colorization with enhanced detail fidelity.
- 4OPS: Structural Difficulty Modeling in Integer Arithmetic Puzzles (viability: 7): https://sciencetostartup.com/paper/4ops-structural-difficulty-modeling-in-integer-arithmetic-puzzles - Develops a novel method for precisely modeling and predicting the difficulty of integer arithmetic puzzles, enabling personalized learning experiences.
- Multi-target Coverage-based Greybox Fuzzing (viability: 3): https://sciencetostartup.com/paper/multi-target-coverage-based-greybox-fuzzing - A novel fuzzing technique for cooperative system and firmware execution to discover vulnerabilities.
- SafeGuard ASF: SR Agentic Humanoid Robot System for Autonomous Industrial Safety (viability: 7): https://sciencetostartup.com/paper/safeguard-asf-sr-agentic-humanoid-robot-system-for-autonomous-industrial-safety - An agentic humanoid robot system for autonomous industrial safety, detecting fires, abnormal temperatures, and intruders.
- Image Rotation Angle Estimation: Comparing Circular-Aware Methods (viability: 5): https://sciencetostartup.com/paper/image-rotation-angle-estimation-comparing-circular-aware-methods - A study comparing circular-aware methods for image rotation estimation, identifying the most robust and accurate approaches for vision pipelines.
- From Intent to Evidence: A Categorical Approach for Structural Evaluation of Deep Research Agents (viability: 7): https://sciencetostartup.com/paper/from-intent-to-evidence-a-categorical-approach-for-structural-evaluation-of-deep-research-agents - A new benchmark and theoretical framework for evaluating deep research agents, revealing significant gaps in their ability to perform complex structural synthesis.
- Large Language Model as Token Compressor and Decompressor (viability: 7): https://sciencetostartup.com/paper/large-language-model-as-token-compressor-and-decompressor - Leverage off-the-shelf LLMs to compress long texts into compact latent codes, enabling token-efficient long-context reasoning and generation.
- HeSS: Head Sensitivity Score for Sparsity Redistribution in VGGT (viability: 7): https://sciencetostartup.com/paper/hess-head-sensitivity-score-for-sparsity-redistribution-in-vggt - A novel method to significantly accelerate 3D vision transformers by intelligently redistributing attention sparsity based on head sensitivity, reducing computational cost without sacrificing accuracy.
- Agentic Trust Coordination for Federated Learning through Adaptive Thresholding and Autonomous Decision Making in Sustainable and Resilient Industrial Networks (viability: 4): https://sciencetostartup.com/paper/agentic-trust-coordination-for-federated-learning-through-adaptive-thresholding-and-autonomous-decision-making-in-sustai - A server-side control layer for federated learning that uses agentic trust coordination to improve reliability in industrial networks.
- Adaptive Chunking: Optimizing Chunking-Method Selection for RAG (viability: 7): https://sciencetostartup.com/paper/adaptive-chunking-optimizing-chunking-method-selection-for-rag - A framework that adaptively optimizes document chunking for RAG systems, significantly improving answer correctness and question success rates without changing models or prompts.
- Beyond Detection: Rethinking Education in the Age of AI-writing (viability: 2): https://sciencetostartup.com/paper/beyond-detection-rethinking-education-in-the-age-of-ai-writing - This paper argues that the cognitive benefits of writing are lost when outsourced to AI, proposing new pedagogical approaches and critical literacy skills for the age of generative AI.
- Macroscopic Characteristics of Mixed Traffic Flow with Deep Reinforcement Learning Based Automated and Human-Driven Vehicles (viability: 5): https://sciencetostartup.com/paper/macroscopic-characteristics-of-mixed-traffic-flow-with-deep-reinforcement-learning-based-automated-and-human-driven-vehi - Leveraging Deep Reinforcement Learning to optimize autonomous vehicle control in mixed traffic, improving road capacity and fuel efficiency.
- Evaluating Language Models for Harmful Manipulation (viability: 4): https://sciencetostartup.com/paper/evaluating-language-models-for-harmful-manipulation - A framework for evaluating AI manipulation in context-specific human-AI interactions, revealing domain and geographic differences in manipulative behavior.
- How Pruning Reshapes Features: Sparse Autoencoder Analysis of Weight-Pruned Language Models (viability: 4): https://sciencetostartup.com/paper/how-pruning-reshapes-features-sparse-autoencoder-analysis-of-weight-pruned-language-models - This research analyzes how weight pruning in language models affects learned features, revealing that rare features are surprisingly resilient and suggesting pruning acts as implicit feature selection.
- AD-CARE: A Guideline-grounded, Modality-agnostic LLM Agent for Real-world Alzheimer's Disease Diagnosis with Multi-cohort Assessment, Fairness Analysis, and Reader Study (viability: 7): https://sciencetostartup.com/paper/ad-care-a-guideline-grounded-modality-agnostic-llm-agent-for-real-world-alzheimer-s-disease-diagnosis-with-multi-cohort - AD-CARE: An AI-powered, guideline-driven agent for enhancing Alzheimer's diagnosis accuracy and clinical efficiency through modality-agnostic multi-modal data integration.
- MACRO: Advancing Multi-Reference Image Generation with Structured Long-Context Data (viability: 7): https://sciencetostartup.com/paper/macro-advancing-multi-reference-image-generation-with-structured-long-context-data - A new dataset and benchmark for multi-reference image generation that overcomes current model limitations by providing structured, long-context supervision.
- Adaptive Learned Image Compression with Graph Neural Networks (viability: 8): https://sciencetostartup.com/paper/adaptive-learned-image-compression-with-graph-neural-networks - A novel graph neural network approach for adaptive image compression that significantly outperforms state-of-the-art methods by modeling spatially varying redundancy.
- Practical Efficient Global Optimization is No-regret (viability: 3): https://sciencetostartup.com/paper/practical-efficient-global-optimization-is-no-regret - This paper theoretically analyzes the regret bounds of a practical Bayesian optimization algorithm, offering theoretical insights into its no-regret properties.
- On the Vulnerability of Deep Automatic Modulation Classifiers to Explainable Backdoor Threats (viability: 3): https://sciencetostartup.com/paper/on-the-vulnerability-of-deep-automatic-modulation-classifiers-to-explainable-backdoor-threats - This paper explores a novel physical backdoor attack on deep learning-based automatic modulation classifiers in wireless communications, guided by explainable AI.
- Separate Before You Compress: The WWHO Tokenization Architecture (viability: 8): https://sciencetostartup.com/paper/separate-before-you-compress-the-wwho-tokenization-architecture - A novel tokenization architecture and algorithm that significantly reduces token count and inference costs for complex scripts, unlocking LLM potential for the Global South.
- Physical Backdoor Attack Against Deep Learning-Based Modulation Classification (viability: 4): https://sciencetostartup.com/paper/physical-backdoor-attack-against-deep-learning-based-modulation-classification - A novel physical backdoor attack against deep learning-based radio frequency signal classifiers that bypasses existing defenses.
- Connectivity-Aware Representations for Constrained Motion Planning via Multi-Scale Contrastive Learning (viability: 4): https://sciencetostartup.com/paper/connectivity-aware-representations-for-constrained-motion-planning-via-multi-scale-contrastive-learning - A novel representation learning approach for constrained motion planning that improves success rates and reduces planning time by intelligently selecting start and goal configurations.
- Towards Controllable Low-Light Image Enhancement: A Continuous Multi-illumination Dataset and Efficient State Space Framework (viability: 7): https://sciencetostartup.com/paper/towards-controllable-low-light-image-enhancement-a-continuous-multi-illumination-dataset-and-efficient-state-space-frame - A controllable low-light image enhancement framework with a new dataset and efficient state space model architecture.
- DAGverse: Building Document-Grounded Semantic DAGs from Scientific Papers (viability: 7): https://sciencetostartup.com/paper/dagverse-building-document-grounded-semantic-dags-from-scientific-papers - A framework and dataset for automatically extracting structured knowledge graphs from scientific papers, enabling deeper document understanding and reasoning.
- Usability of Passwordless Authentication in Wi-Fi Networks: A Comparative Study of Passkeys and Passwords in Captive Portals (viability: 3): https://sciencetostartup.com/paper/usability-of-passwordless-authentication-in-wi-fi-networks-a-comparative-study-of-passkeys-and-passwords-in-captive-port - This paper explores the usability of passkeys versus passwords in Wi-Fi captive portals, identifying design recommendations to improve user experience and reduce error rates.
- Revealing the influence of participant failures on model quality in cross-silo Federated Learning (viability: 3): https://sciencetostartup.com/paper/revealing-the-influence-of-participant-failures-on-model-quality-in-cross-silo-federated-learning - This research systematically investigates the impact of participant failures on model quality in cross-silo Federated Learning, providing insights into data skewness and availability patterns.
- CSI-tuples-based 3D Channel Fingerprints Construction Assisted by MultiModal Learning (viability: 7): https://sciencetostartup.com/paper/csi-tuples-based-3d-channel-fingerprints-construction-assisted-by-multimodal-learning - A multimodal AI framework that constructs 3D channel fingerprints for enhanced 6G communication by fusing diverse data sources, outperforming existing methods by over 27.5% in accuracy.
- SliderQuant: Accurate Post-Training Quantization for LLMs (viability: 7): https://sciencetostartup.com/paper/sliderquant-accurate-post-training-quantization-for-llms - SliderQuant offers accurate post-training quantization for LLMs by adaptively quantizing layers based on their sensitivity, outperforming existing methods across various tasks and models.
- A Gait Foundation Model Predicts Multi-System Health Phenotypes from 3D Skeletal Motion (viability: 7): https://sciencetostartup.com/paper/a-gait-foundation-model-predicts-multi-system-health-phenotypes-from-3d-skeletal-motion - A foundation model for 3D skeletal motion predicts multi-system health phenotypes, enabling gait to be used as a scalable, passive vital sign.
- V2U4Real: A Real-world Large-scale Dataset for Vehicle-to-UAV Cooperative Perception (viability: 7): https://sciencetostartup.com/paper/v2u4real-a-real-world-large-scale-dataset-for-vehicle-to-uav-cooperative-perception - A large-scale real-world dataset and benchmarks for Vehicle-to-UAV cooperative perception, addressing limitations of ground-level autonomous vehicle sensing.
- Distribution and Clusters Approximations as Abstract Domains in Probabilistic Abstract Interpretation to Neural Network Analysis (viability: 1): https://sciencetostartup.com/paper/distribution-and-clusters-approximations-as-abstract-domains-in-probabilistic-abstract-interpretation-to-neural-network - This paper introduces theoretical novel approximation methods for analyzing neural network density distributions, lacking immediate product application.
- When Hate Meets Facts: LLMs-in-the-Loop for Check-worthiness Detection in Hate Speech (viability: 7): https://sciencetostartup.com/paper/when-hate-meets-facts-llms-in-the-loop-for-check-worthiness-detection-in-hate-speech - An LLM-in-the-loop framework and dataset to improve hate speech detection by assessing claim check-worthiness, reducing moderator effort and improving model accuracy.
- CRAFT: Grounded Multi-Agent Coordination Under Partial Information (viability: 7): https://sciencetostartup.com/paper/craft-grounded-multi-agent-coordination-under-partial-information - A benchmark for evaluating and improving pragmatic communication and coordination in large language models for complex tasks.
- EagleNet: Energy-Aware Fine-Grained Relationship Learning Network for Text-Video Retrieval (viability: 7): https://sciencetostartup.com/paper/eaglenet-energy-aware-fine-grained-relationship-learning-network-for-text-video-retrieval - EagleNet enhances text-video retrieval by learning fine-grained relationships between text and video frames, leading to more accurate context-aware embeddings.
- Probabilistic Abstract Interpretation on Neural Networks via Grids Approximation (viability: 2): https://sciencetostartup.com/paper/probabilistic-abstract-interpretation-on-neural-networks-via-grids-approximation - A theoretical framework for analyzing the input-output behavior of neural networks by approximating density distributions.
- ViewSplat: View-Adaptive Dynamic Gaussian Splatting for Feed-Forward Synthesis (viability: 7): https://sciencetostartup.com/paper/viewsplat-view-adaptive-dynamic-gaussian-splatting-for-feed-forward-synthesis - ViewSplat enables high-fidelity, real-time 3D scene reconstruction from unposed images by dynamically adapting Gaussian splatting to each viewpoint.
- Towards Practical Lossless Neural Compression for LiDAR Point Clouds (viability: 7): https://sciencetostartup.com/paper/towards-practical-lossless-neural-compression-for-lidar-point-clouds - A novel neural compression framework for LiDAR point clouds that achieves real-time speed with competitive compression performance.
- A Minimum-Energy Control Approach for Redundant Mobile Manipulators in Physical Human-Robot Interaction Applications (viability: 5): https://sciencetostartup.com/paper/a-minimum-energy-control-approach-for-redundant-mobile-manipulators-in-physical-human-robot-interaction-applications - A novel control method for mobile manipulators that minimizes kinetic energy for safer and more efficient human-robot interaction.
- Mitigating Evasion Attacks in Fog Computing Resource Provisioning Through Proactive Hardening (viability: 3): https://sciencetostartup.com/paper/mitigating-evasion-attacks-in-fog-computing-resource-provisioning-through-proactive-hardening - This paper proposes an adversarial training method to improve the robustness of k-means based resource provisioning in fog networks against evasion attacks.
- Hyperspectral Trajectory Image for Multi-Month Trajectory Anomaly Detection (viability: 7): https://sciencetostartup.com/paper/hyperspectral-trajectory-image-for-multi-month-trajectory-anomaly-detection - A novel vision-based approach for multi-month trajectory anomaly detection that significantly outperforms existing methods and offers substantial speedups.
- MolQuest: A Benchmark for Agentic Evaluation of Abductive Reasoning in Chemical Structure Elucidation (viability: 7): https://sciencetostartup.com/paper/molquest-a-benchmark-for-agentic-evaluation-of-abductive-reasoning-in-chemical-structure-elucidation - MolQuest provides a novel agent-based benchmark to evaluate and improve LLMs' abductive reasoning for complex scientific tasks like chemical structure elucidation, revealing significant performance gaps in current frontier models.
- Does Explanation Correctness Matter? Linking Computational XAI Evaluation to Human Understanding (viability: 4): https://sciencetostartup.com/paper/does-explanation-correctness-matter-linking-computational-xai-evaluation-to-human-understanding - This research validates the effectiveness of Explainable AI metrics by demonstrating that not all functional correctness improvements translate to better human understanding, highlighting a gap for AI explainability tools.
- Activation Matters: Test-time Activated Negative Labels for OOD Detection with Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/activation-matters-test-time-activated-negative-labels-for-ood-detection-with-vision-language-models - A training-free method for out-of-distribution detection that dynamically selects negative labels based on test-time activation to significantly improve accuracy.
- Semantic-Aware Prefix Learning for Token-Efficient Image Generation (viability: 4): https://sciencetostartup.com/paper/semantic-aware-prefix-learning-for-token-efficient-image-generation - A novel tokenization method for image generation that improves semantic understanding and generation quality.
- FEAST: Fully Connected Expressive Attention for Spatial Transcriptomics (viability: 7): https://sciencetostartup.com/paper/feast-fully-connected-expressive-attention-for-spatial-transcriptomics - FEAST uses a novel attention mechanism and off-grid sampling to infer spatial gene expression from whole slide images, overcoming limitations of existing graph neural networks and providing biologically plausible insights.
- Efficient Preemptive Robustification with Image Sharpening (viability: 4): https://sciencetostartup.com/paper/efficient-preemptive-robustification-with-image-sharpening - A novel image sharpening technique that efficiently robustifies deep neural networks against adversarial attacks without requiring complex training or optimization.
- FluxEDA: A Unified Execution Infrastructure for Stateful Agentic EDA (viability: 5): https://sciencetostartup.com/paper/fluxeda-a-unified-execution-infrastructure-for-stateful-agentic-eda - A stateful execution infrastructure for agentic EDA tools to enable iterative optimization and state preservation in production environments.
- Offline Decision Transformers for Neural Combinatorial Optimization: Surpassing Heuristics on the Traveling Salesman Problem (viability: 7): https://sciencetostartup.com/paper/offline-decision-transformers-for-neural-combinatorial-optimization-surpassing-heuristics-on-the-traveling-salesman-prob - Leveraging offline reinforcement learning to outperform classical heuristics on complex combinatorial optimization problems like the Traveling Salesman Problem.
- A Unified Spatial Alignment Framework for Highly Transferable Transformation-Based Attacks on Spatially Structured Tasks (viability: 4): https://sciencetostartup.com/paper/a-unified-spatial-alignment-framework-for-highly-transferable-transformation-based-attacks-on-spatially-structured-tasks - A framework for creating more effective adversarial attacks on structured computer vision tasks by aligning input transformations with label transformations.
- An Image Dataset of Common Skin Diseases of Bangladesh and Benchmarking Performance with Machine Learning Models (viability: 4): https://sciencetostartup.com/paper/an-image-dataset-of-common-skin-diseases-of-bangladesh-and-benchmarking-performance-with-machine-learning-models - A publicly available dataset of common skin diseases in Bangladesh, with initial benchmarking, to accelerate AI-powered dermatology diagnostics.
- Training-free Detection and 6D Pose Estimation of Unseen Surgical Instruments (viability: 7): https://sciencetostartup.com/paper/training-free-detection-and-6d-pose-estimation-of-unseen-surgical-instruments - A training-free system for accurate 6D pose estimation of unseen surgical instruments using only CAD models, enabling robust tracking in clinical environments.
- Comparing Natural and Synthetic Structured Data: A Study of the Passive Verb Alternation in French and Italian (viability: 4): https://sciencetostartup.com/paper/comparing-natural-and-synthetic-structured-data-a-study-of-the-passive-verb-alternation-in-french-and-italian - Develops a method to evaluate LLM linguistic generalization using structured natural and synthetic data, highlighting the limitations of synthetic data for real-world understanding.
- WebTestBench: Evaluating Computer-Use Agents towards End-to-End Automated Web Testing (viability: 7): https://sciencetostartup.com/paper/webtestbench-evaluating-computer-use-agents-towards-end-to-end-automated-web-testing - A benchmark and baseline framework for evaluating and improving end-to-end automated web testing agents, addressing critical gaps in test completeness and reliability for industrial deployment.
- Fair regression under localized demographic parity constraints (viability: 4): https://sciencetostartup.com/paper/fair-regression-under-localized-demographic-parity-constraints - A novel regression fairness method that enforces distributional parity at specific quantiles or thresholds, offering a tunable trade-off between fairness and accuracy.
- Translation or Recitation? Calibrating Evaluation Scores for Machine Translation of Extremely Low-Resource Languages (viability: 4): https://sciencetostartup.com/paper/translation-or-recitation-calibrating-evaluation-scores-for-machine-translation-of-extremely-low-resource-languages - Develops intrinsic metrics to contextualize machine translation performance for extremely low-resource languages, addressing variability caused by dataset artifacts rather than model capability.
- Gap Safe Screening Rules for Fast Training of Robust Support Vector Machines under Feature Noise (viability: 4): https://sciencetostartup.com/paper/gap-safe-screening-rules-for-fast-training-of-robust-support-vector-machines-under-feature-noise - Accelerate the training of robust machine learning models by safely screening out irrelevant data points.
- SDD-YOLO: A Small-Target Detection Framework for Ground-to-Air Anti-UAV Surveillance with Edge-Efficient Deployment (viability: 7): https://sciencetostartup.com/paper/sdd-yolo-a-small-target-detection-framework-for-ground-to-air-anti-uav-surveillance-with-edge-efficient-deployment - A highly efficient small-target detection framework for anti-UAV surveillance, optimized for edge deployment with a new dataset and improved accuracy.
- A Wireless World Model for AI-Native 6G Networks (viability: 7): https://sciencetostartup.com/paper/a-wireless-world-model-for-ai-native-6g-networks - A foundation model that predicts wireless channel evolution by understanding 3D geometry and signal dynamics, enabling physics-aware 6G intelligence.
- Free-Lunch Long Video Generation via Layer-Adaptive O.O.D Correction (viability: 7): https://sciencetostartup.com/paper/free-lunch-long-video-generation-via-layer-adaptive-o-o-d-correction - A training-free framework that significantly improves long video generation quality from pre-trained short-clip models by adaptively correcting out-of-distribution issues.
- A CDF-First Framework for Free-Form Density Estimation (viability: 5): https://sciencetostartup.com/paper/a-cdf-first-framework-for-free-form-density-estimation - A new framework for free-form density estimation that estimates the CDF first to overcome limitations of direct PDF estimation, outperforming existing methods.
- Probabilistic Concept Graph Reasoning for Multimodal Misinformation Detection (viability: 7): https://sciencetostartup.com/paper/probabilistic-concept-graph-reasoning-for-multimodal-misinformation-detection - An interpretable framework for detecting multimodal misinformation by reasoning over discovered concepts, outperforming existing methods.
- CIV-DG: Conditional Instrumental Variables for Domain Generalization in Medical Imaging (viability: 7): https://sciencetostartup.com/paper/civ-dg-conditional-instrumental-variables-for-domain-generalization-in-medical-imaging - A causal framework using conditional instrumental variables to achieve robust domain generalization in medical imaging by disentangling pathological semantics from site-specific artifacts.
- SafeMath: Inference-time Safety improves Math Accuracy (viability: 7): https://sciencetostartup.com/paper/safemath-inference-time-safety-improves-math-accuracy - A safety alignment technique that reduces harmful LLM outputs in mathematical reasoning without sacrificing accuracy, supported by a new dataset and released code.
- TacSIm: A Dataset and Benchmark for Football Tactical Style Imitation (viability: 5): https://sciencetostartup.com/paper/tacsim-a-dataset-and-benchmark-for-football-tactical-style-imitation - A new dataset and benchmark for accurately replicating real-world football team tactical behaviors, moving beyond simple reward optimization.
- The Competence Shadow: Theory and Bounds of AI Assistance in Safety Engineering (viability: 3): https://sciencetostartup.com/paper/the-competence-shadow-theory-and-bounds-of-ai-assistance-in-safety-engineering - Formalizing the risks of AI assistance in safety engineering to design shadow-resistant workflows for physical AI systems.
- A Decade-Scale Benchmark Evaluating LLMs' Clinical Practice Guidelines Detection and Adherence in Multi-turn Conversations (viability: 7): https://sciencetostartup.com/paper/a-decade-scale-benchmark-evaluating-llms-clinical-practice-guidelines-detection-and-adherence-in-multi-turn-conversation - A benchmark to evaluate and improve LLM adherence to clinical practice guidelines, crucial for safe healthcare deployment.
- CardioDiT: Latent Diffusion Transformers for 4D Cardiac MRI Synthesis (viability: 7): https://sciencetostartup.com/paper/cardiodit-latent-diffusion-transformers-for-4d-cardiac-mri-synthesis - A 4D latent diffusion transformer for synthesizing realistic cardiac MRI sequences, improving temporal consistency and physiological accuracy.
- zk-X509: Privacy-Preserving On-Chain Identity from Legacy PKI via Zero-Knowledge Proofs (viability: 5): https://sciencetostartup.com/paper/zk-x509-privacy-preserving-on-chain-identity-from-legacy-pki-via-zero-knowledge-proofs - Leverage existing X.509 certificates for privacy-preserving on-chain identity using zero-knowledge proofs.
- A Catalog of Basque Dialectal Resources: Online Collections and Standard-to-Dialectal Adaptations (viability: 5): https://sciencetostartup.com/paper/a-catalog-of-basque-dialectal-resources-online-collections-and-standard-to-dialectal-adaptations - This paper catalogs and creates new Basque dialectal NLP resources, addressing data scarcity for low-resource languages.
- AnyID: Ultra-Fidelity Universal Identity-Preserving Video Generation from Any Visual References (viability: 7): https://sciencetostartup.com/paper/anyid-ultra-fidelity-universal-identity-preserving-video-generation-from-any-visual-references - AnyID enables ultra-fidelity video generation from any visual reference, offering precise attribute control for creative expression.
- Probing the Lack of Stable Internal Beliefs in LLMs (viability: 3): https://sciencetostartup.com/paper/probing-the-lack-of-stable-internal-beliefs-in-llms - This research identifies a fundamental limitation in current LLMs' ability to maintain consistent internal goals, hindering the development of realistic persona-driven applications.
- Knowledge-Guided Retrieval-Augmented Generation for Zero-Shot Psychiatric Data: Privacy Preserving Synthetic Data Generation (viability: 7): https://sciencetostartup.com/paper/knowledge-guided-retrieval-augmented-generation-for-zero-shot-psychiatric-data-privacy-preserving-synthetic-data-generat - Generate privacy-preserving synthetic psychiatric data using LLMs guided by clinical knowledge, overcoming limitations of real data access.
- Train at Moving Edge: Online-Verified Prompt Selection for Efficient RL Training of Large Reasoning Model (viability: 7): https://sciencetostartup.com/paper/train-at-moving-edge-online-verified-prompt-selection-for-efficient-rl-training-of-large-reasoning-model - A dual-stage framework that efficiently selects high-utility prompts for reinforcement learning in large language models, reducing computational costs without sacrificing performance.
- Cross-Preference Learning for Sentence-Level and Context-Aware Machine Translation (viability: 7): https://sciencetostartup.com/paper/cross-preference-learning-for-sentence-level-and-context-aware-machine-translation - A novel training framework for machine translation that adaptively leverages document-level context to improve translation quality and robustness.
- VolDiT: Controllable Volumetric Medical Image Synthesis with Diffusion Transformers (viability: 7): https://sciencetostartup.com/paper/voldit-controllable-volumetric-medical-image-synthesis-with-diffusion-transformers - A controllable 3D medical image synthesis tool using diffusion transformers for improved fidelity and spatial guidance.
- Bilingual Text-to-Motion Generation: A New Benchmark and Baselines (viability: 7): https://sciencetostartup.com/paper/bilingual-text-to-motion-generation-a-new-benchmark-and-baselines - A new bilingual text-to-motion benchmark and baseline model that enables high-quality motion generation from diverse language inputs, including zero-shot code-switching.
- Prompt Attack Detection with LLM-as-a-Judge and Mixture-of-Models (viability: 8): https://sciencetostartup.com/paper/prompt-attack-detection-with-llm-as-a-judge-and-mixture-of-models - Leveraging lightweight LLMs as low-latency judges to secure public chatbots against prompt attacks in real-time production environments.
- AG-EgoPose: Leveraging Action-Guided Motion and Kinematic Joint Encoding for Egocentric 3D Pose Estimation (viability: 7): https://sciencetostartup.com/paper/ag-egopose-leveraging-action-guided-motion-and-kinematic-joint-encoding-for-egocentric-3d-pose-estimation - A novel dual-stream framework for robust egocentric 3D human pose estimation from fisheye cameras, leveraging action-guided motion and kinematic joint encoding.
- Knowledge-Guided Adversarial Training for Infrared Object Detection via Thermal Radiation Modeling (viability: 4): https://sciencetostartup.com/paper/knowledge-guided-adversarial-training-for-infrared-object-detection-via-thermal-radiation-modeling - Enhance infrared object detection robustness by embedding physical thermal radiation knowledge into adversarial training.
- To Write or to Automate Linguistic Prompts, That Is the Question (viability: 4): https://sciencetostartup.com/paper/to-write-or-to-automate-linguistic-prompts-that-is-the-question - Automating LLM prompt optimization for linguistic tasks shows task-dependent results, sometimes matching expert performance but not yet replacing it.
- ET-SAM: Efficient Point Prompt Prediction in SAM for Unified Scene Text Detection and Layout Analysis (viability: 7): https://sciencetostartup.com/paper/et-sam-efficient-point-prompt-prediction-in-sam-for-unified-scene-text-detection-and-layout-analysis - ET-SAM accelerates scene text detection and layout analysis by efficiently predicting point prompts, enabling faster inference and better data utilization.
- Towards Foundation Models for 3D Scene Understanding: Instance-Aware Self-Supervised Learning for Point Clouds (viability: 7): https://sciencetostartup.com/paper/towards-foundation-models-for-3d-scene-understanding-instance-aware-self-supervised-learning-for-point-clouds - A self-supervised learning framework for 3D point clouds that enhances instance localization and semantic understanding, enabling scalable foundation models for diverse downstream tasks.
- PIDP-Attack: Combining Prompt Injection with Database Poisoning Attacks on Retrieval-Augmented Generation Systems (viability: 4): https://sciencetostartup.com/paper/pidp-attack-combining-prompt-injection-with-database-poisoning-attacks-on-retrieval-augmented-generation-systems - A novel attack method that combines prompt injection and database poisoning to compromise Retrieval-Augmented Generation systems, demonstrating significant improvements over existing methods.
- SportSkills: Physical Skill Learning from Sports Instructional Videos (viability: 7): https://sciencetostartup.com/paper/sportskills-physical-skill-learning-from-sports-instructional-videos - A dataset and retrieval system for personalized sports skill improvement from instructional videos.
- A Semantically Disentangled Unified Model for Multi-category 3D Anomaly Detection (viability: 7): https://sciencetostartup.com/paper/a-semantically-disentangled-unified-model-for-multi-category-3d-anomaly-detection - A unified 3D anomaly detection model that disentangles category semantics to achieve state-of-the-art performance and reliability.
- Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills (viability: 8): https://sciencetostartup.com/paper/trace2skill-distill-trajectory-local-lessons-into-transferable-agent-skills - Automatically distill transferable agent skills from execution experience, enabling LLM agents to tackle complex tasks without parameter updates.
- Vision Hopfield Memory Networks (viability: 7): https://sciencetostartup.com/paper/vision-hopfield-memory-networks - A brain-inspired vision foundation model that offers improved data efficiency and interpretability through hierarchical memory mechanisms.
- Photon: Speedup Volume Understanding with Efficient Multimodal Large Language Models (viability: 5): https://sciencetostartup.com/paper/photon-speedup-volume-understanding-with-efficient-multimodal-large-language-models - Photon accelerates 3D medical image understanding by efficiently representing volumetric data with adaptive token scheduling for multimodal LLMs.
- UniAI-GraphRAG: Synergizing Ontology-Guided Extraction, Multi-Dimensional Clustering, and Dual-Channel Fusion for Robust Multi-Hop Reasoning (viability: 7): https://sciencetostartup.com/paper/uniai-graphrag-synergizing-ontology-guided-extraction-multi-dimensional-clustering-and-dual-channel-fusion-for-robust-mu - An enhanced RAG framework that improves multi-hop reasoning and domain-specific QA through ontology-guided extraction, advanced clustering, and dual-channel retrieval.
- Goodness-of-pronunciation without phoneme time alignment (viability: 5): https://sciencetostartup.com/paper/goodness-of-pronunciation-without-phoneme-time-alignment - Enables low-resource language speech evaluation by adapting weakly-supervised ASR models without phoneme time alignment.
- Factors Influencing the Quality of AI-Generated Code: A Synthesis of Empirical Evidence (viability: 3): https://sciencetostartup.com/paper/factors-influencing-the-quality-of-ai-generated-code-a-synthesis-of-empirical-evidence - This paper synthesizes empirical evidence on factors influencing the quality of AI-generated code, highlighting the interplay of human and AI system characteristics.
- Learning to Rank Caption Chains for Video-Text Alignment (viability: 4): https://sciencetostartup.com/paper/learning-to-rank-caption-chains-for-video-text-alignment - A novel ranking optimization approach for video-text alignment that improves caption generation quality over binary preference methods.
- FD$^2$: A Dedicated Framework for Fine-Grained Dataset Distillation (viability: 7): https://sciencetostartup.com/paper/fd-2-a-dedicated-framework-for-fine-grained-dataset-distillation - A framework for fine-grained dataset distillation that creates more discriminative and diverse synthetic datasets, improving recognition performance.
- SAVe: Self-Supervised Audio-visual Deepfake Detection Exploiting Visual Artifacts and Audio-visual Misalignment (viability: 7): https://sciencetostartup.com/paper/save-self-supervised-audio-visual-deepfake-detection-exploiting-visual-artifacts-and-audio-visual-misalignment - A self-supervised framework for detecting audio-visual deepfakes by learning solely on authentic videos, overcoming dataset bias and improving generalization.
- Dissimilarity-Based Persistent Coverage Control of Multi-Robot Systems for Improving Solar Irradiance Prediction Accuracy in Solar Thermal Power Plants (viability: 4): https://sciencetostartup.com/paper/dissimilarity-based-persistent-coverage-control-of-multi-robot-systems-for-improving-solar-irradiance-prediction-accurac - A persistent coverage control algorithm for multi-robot systems to improve solar irradiance prediction accuracy in solar thermal power plants.
- EgoXtreme: A Dataset for Robust Object Pose Estimation in Egocentric Views under Extreme Conditions (viability: 7): https://sciencetostartup.com/paper/egoxtreme-a-dataset-for-robust-object-pose-estimation-in-egocentric-views-under-extreme-conditions - A new dataset and code for robust 6D object pose estimation in egocentric views under extreme real-world conditions, addressing limitations of current benchmarks.
- RubricEval: A Rubric-Level Meta-Evaluation Benchmark for LLM Judges in Instruction Following (viability: 4): https://sciencetostartup.com/paper/rubriceval-a-rubric-level-meta-evaluation-benchmark-for-llm-judges-in-instruction-following - A new benchmark for evaluating the reliability of rubric-based instruction following in LLMs, revealing significant performance gaps even in advanced models.
- Robust Principal Component Completion (viability: 7): https://sciencetostartup.com/paper/robust-principal-component-completion - A novel Bayesian framework for robust principal component completion that directly identifies sparse component support, improving foreground extraction and anomaly detection.
- Denoise and Align: Towards Source-Free UDA for Robust Panoramic Semantic Segmentation (viability: 7): https://sciencetostartup.com/paper/denoise-and-align-towards-source-free-uda-for-robust-panoramic-semantic-segmentation - A framework for robust panoramic semantic segmentation that adapts models from labeled pinhole datasets to unlabeled panoramic data without access to the original source data, improving performance on challenging real-world applications.
- AirSplat: Alignment and Rating for Robust Feed-Forward 3D Gaussian Splatting (viability: 7): https://sciencetostartup.com/paper/airsplat-alignment-and-rating-for-robust-feed-forward-3d-gaussian-splatting - Adapt 3D vision foundation models for high-fidelity, pose-free novel view synthesis with a novel alignment and rating framework.
- MCLMR: A Model-Agnostic Causal Learning Framework for Multi-Behavior Recommendation (viability: 7): https://sciencetostartup.com/paper/mclmr-a-model-agnostic-causal-learning-framework-for-multi-behavior-recommendation - A causal learning framework that improves recommendation systems by intelligently fusing user behaviors and mitigating bias.
- CTS-PLL: A Robust and Anytime Framework for Collaborative Task Sequencing and Multi-Agent Path Finding (viability: 7): https://sciencetostartup.com/paper/cts-pll-a-robust-and-anytime-framework-for-collaborative-task-sequencing-and-multi-agent-path-finding - A robust and anytime framework for collaborative task sequencing and multi-agent path finding that improves solution quality and success rates in complex environments.
- AnyDoc: Enhancing Document Generation via Large-Scale HTML/CSS Data Synthesis and Height-Aware Reinforcement Optimization (viability: 7): https://sciencetostartup.com/paper/anydoc-enhancing-document-generation-via-large-scale-html-css-data-synthesis-and-height-aware-reinforcement-optimization - A framework for generating diverse documents from natural language instructions by synthesizing large-scale HTML/CSS datasets and using height-aware reinforcement learning.
- When Sensing Varies with Contexts: Context-as-Transform for Tactile Few-Shot Class-Incremental Learning (viability: 4): https://sciencetostartup.com/paper/when-sensing-varies-with-contexts-context-as-transform-for-tactile-few-shot-class-incremental-learning - A novel method for few-shot class-incremental learning in tactile sensing that adapts to varying acquisition contexts.
- Do LLMs Know What They Know? Measuring Metacognitive Efficiency with Signal Detection Theory (viability: 7): https://sciencetostartup.com/paper/do-llms-know-what-they-know-measuring-metacognitive-efficiency-with-signal-detection-theory - A new evaluation framework using signal detection theory to measure LLM metacognitive efficiency, revealing which models truly know what they don't know.
- SEVerA: Verified Synthesis of Self-Evolving Agents (viability: 4): https://sciencetostartup.com/paper/severa-verified-synthesis-of-self-evolving-agents - A framework for generating self-evolving LLM agents with formal guarantees of safety and correctness.
- MoireMix: A Formula-Based Data Augmentation for Improving Image Classification Robustness (viability: 7): https://sciencetostartup.com/paper/moiremix-a-formula-based-data-augmentation-for-improving-image-classification-robustness - A formula-based data augmentation technique for image classification that procedurally generates Moire interference patterns on-the-fly to significantly improve model robustness with negligible computational cost.
- MSRL: Scaling Generative Multimodal Reward Modeling via Multi-Stage Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/msrl-scaling-generative-multimodal-reward-modeling-via-multi-stage-reinforcement-learning - Scales generative multimodal reward modeling using a novel multi-stage reinforcement learning approach, reducing reliance on costly multimodal preference data and significantly improving performance on visual understanding and generation tasks.
- Label What Matters: Modality-Balanced and Difficulty-Aware Multimodal Active Learning (viability: 7): https://sciencetostartup.com/paper/label-what-matters-modality-balanced-and-difficulty-aware-multimodal-active-learning - A reinforcement learning framework for smarter, modality-balanced data labeling in multimodal AI, reducing annotation costs and improving accuracy.
- OMIND: Framework for Knowledge Grounded Finetuning and Multi-Turn Dialogue Benchmark for Mental Health LLMs (viability: 7): https://sciencetostartup.com/paper/omind-framework-for-knowledge-grounded-finetuning-and-multi-turn-dialogue-benchmark-for-mental-health-llms - A framework and dataset for training and evaluating LLMs specifically for mental health conversations, demonstrating superior reasoning and performance.
- Layer-Specific Lipschitz Modulation for Fault-Tolerant Multimodal Representation Learning (viability: 5): https://sciencetostartup.com/paper/layer-specific-lipschitz-modulation-for-fault-tolerant-multimodal-representation-learning - A mathematically grounded framework for building multimodal AI systems that can tolerate sensor failures and signal degradation.
- From Logic Monopoly to Social Contract: Separation of Power and the Institutional Foundations for Autonomous Agent Economies (viability: 3): https://sciencetostartup.com/paper/from-logic-monopoly-to-social-contract-separation-of-power-and-the-institutional-foundations-for-autonomous-agent-econom - A theoretical framework for structuring autonomous agent economies with a separation of powers model to address reliability and deception issues.
- Large Language Models as Optimization Controllers: Adaptive Continuation for SIMP Topology Optimization (viability: 7): https://sciencetostartup.com/paper/large-language-models-as-optimization-controllers-adaptive-continuation-for-simp-topology-optimization - Leveraging LLMs as adaptive controllers for topology optimization to achieve superior engineering designs.
- ElephantBroker: A Knowledge-Grounded Cognitive Runtime for Trustworthy AI Agents (viability: 7): https://sciencetostartup.com/paper/elephantbroker-a-knowledge-grounded-cognitive-runtime-for-trustworthy-ai-agents - ElephantBroker provides a verifiable and trustworthy memory system for LLM agents, enabling secure and auditable operation in high-stakes environments.
- Process-Aware AI for Rainfall-Runoff Modeling: A Mass-Conserving Neural Framework with Hydrological Process Constraints (viability: 5): https://sciencetostartup.com/paper/process-aware-ai-for-rainfall-runoff-modeling-a-mass-conserving-neural-framework-with-hydrological-process-constraints - A physics-constrained AI framework for rainfall-runoff modeling that improves predictive accuracy and interpretability by embedding hydrological process knowledge.
- Pixelis: Reasoning in Pixels, from Seeing to Acting (viability: 7): https://sciencetostartup.com/paper/pixelis-reasoning-in-pixels-from-seeing-to-acting - Pixelis is a pixel-space agent that learns to act and adapt in visual environments by directly manipulating images and videos, enabling grounded multimodal perception and embodied adaptation.
- THEMIS: Towards Holistic Evaluation of MLLMs for Scientific Paper Fraud Forensics (viability: 7): https://sciencetostartup.com/paper/themis-towards-holistic-evaluation-of-mllms-for-scientific-paper-fraud-forensics - A new benchmark and evaluation framework for multimodal LLMs to detect sophisticated fraud in scientific papers.
- Visual Attention Drifts,but Anchors Hold:Mitigating Hallucination in Multimodal Large Language Models via Cross-Layer Visual Anchors (viability: 7): https://sciencetostartup.com/paper/visual-attention-drifts-but-anchors-hold-mitigating-hallucination-in-multimodal-large-language-models-via-cross-layer-vi - A training-free method to reduce hallucination in multimodal LLMs by reinforcing intermediate visual features, improving output reliability without significant computational overhead.
- Learning domain-invariant features through channel-level sparsification for Out-Of Distribution Generalization (viability: 7): https://sciencetostartup.com/paper/learning-domain-invariant-features-through-channel-level-sparsification-for-out-of-distribution-generalization - A novel method for training image analysis systems that reliably generalize across different data sources by enforcing feature sparsity and causal interventions.
- Bridging Perception and Reasoning: Token Reweighting for RLVR in Multimodal LLMs (viability: 7): https://sciencetostartup.com/paper/bridging-perception-and-reasoning-token-reweighting-for-rlvr-in-multimodal-llms - A plug-and-play strategy to improve multimodal LLM reasoning by dynamically reweighting critical perception and reasoning tokens during training.
- Sparse Visual Thought Circuits in Vision-Language Models (viability: 4): https://sciencetostartup.com/paper/sparse-visual-thought-circuits-in-vision-language-models - Develops a diagnostic framework to understand and control the internal workings of vision-language models by analyzing sparse autoencoder features.
- Z-Erase: Enabling Concept Erasure in Single-Stream Diffusion Transformers (viability: 7): https://sciencetostartup.com/paper/z-erase-enabling-concept-erasure-in-single-stream-diffusion-transformers - A novel framework for stable concept erasure in single-stream text-to-image diffusion models, overcoming generation collapse and achieving state-of-the-art performance.
- GIFT: Global Irreplaceability Frame Targeting for Efficient Video Understanding (viability: 7): https://sciencetostartup.com/paper/gift-global-irreplaceability-frame-targeting-for-efficient-video-understanding - A novel framework for efficient video understanding that selects irreplaceable frames to significantly improve performance on long-form video benchmarks.
- An Explainable Ensemble Learning Framework for Crop Classification with Optimized Feature Pyramids and Deep Networks (viability: 5): https://sciencetostartup.com/paper/an-explainable-ensemble-learning-framework-for-crop-classification-with-optimized-feature-pyramids-and-deep-networks - An explainable ensemble learning framework for crop classification that significantly outperforms individual models, providing actionable insights for agricultural decision-making.
- Ultra-fast Traffic Nowcasting and Control via Differentiable Agent-based Simulation (viability: 5): https://sciencetostartup.com/paper/ultra-fast-traffic-nowcasting-and-control-via-differentiable-agent-based-simulation - A differentiable agent-based traffic simulator enables ultra-fast model calibration, nowcasting, and control for traffic digital twins.
- TopoPilot: Reliable Conversational Workflow Automation for Topological Data Analysis and Visualization (viability: 7): https://sciencetostartup.com/paper/topopilot-reliable-conversational-workflow-automation-for-topological-data-analysis-and-visualization - TopoPilot is a reliable agentic framework that automates complex scientific visualization workflows with over 99% success rate, overcoming the limitations of current LLM-based systems.
- SIGMA: Structure-Invariant Generative Molecular Alignment for Chemical Language Models via Autoregressive Contrastive Learning (viability: 7): https://sciencetostartup.com/paper/sigma-structure-invariant-generative-molecular-alignment-for-chemical-language-models-via-autoregressive-contrastive-lea - SIGMA aligns molecular structures in generative models by enforcing symmetry, improving sample efficiency and diversity for multi-parameter optimization.
- Learning Explicit Continuous Motion Representation for Dynamic Gaussian Splatting from Monocular Videos (viability: 7): https://sciencetostartup.com/paper/learning-explicit-continuous-motion-representation-for-dynamic-gaussian-splatting-from-monocular-videos - Enabling high-quality dynamic 3D scene reconstruction from single videos by explicitly modeling continuous motion.
- The System Prompt Is the Attack Surface: How LLM Agent Configuration Shapes Security and Creates Exploitable Vulnerabilities (viability: 6): https://sciencetostartup.com/paper/the-system-prompt-is-the-attack-surface-how-llm-agent-configuration-shapes-security-and-creates-exploitable-vulnerabilit - This research demonstrates how prompt engineering can drastically improve LLM agent phishing detection, but also reveals exploitable vulnerabilities that require external tool augmentation for robust security.
- Synergistic Event-SVE Imaging for Quantitative Propellant Combustion Diagnostics (viability: 3): https://sciencetostartup.com/paper/synergistic-event-sve-imaging-for-quantitative-propellant-combustion-diagnostics - A novel imaging system for real-time, high-resolution diagnostics of propellant combustion under challenging conditions.
- GaussFusion: Improving 3D Reconstruction in the Wild with A Geometry-Informed Video Generator (viability: 7): https://sciencetostartup.com/paper/gaussfusion-improving-3d-reconstruction-in-the-wild-with-a-geometry-informed-video-generator - GaussFusion refines 3D reconstructions from imperfect video inputs by generating temporally coherent, artifact-free frames, enabling real-time interactive 3D applications.
- Closing the Confidence-Faithfulness Gap in Large Language Models (viability: 7): https://sciencetostartup.com/paper/closing-the-confidence-faithfulness-gap-in-large-language-models - A novel two-stage adaptive steering pipeline significantly improves LLM calibration by aligning verbalized confidence with actual accuracy, addressing the 'Reasoning Contamination Effect'.
- Approaches to Analysing Historical Newspapers Using LLMs (viability: 5): https://sciencetostartup.com/paper/approaches-to-analysing-historical-newspapers-using-llms - Leveraging LLMs and topic modeling to analyze historical newspapers for insights into collective identities and political discourse.
- MP-MoE: Matrix Profile-Guided Mixture of Experts for Precipitation Forecasting (viability: 7): https://sciencetostartup.com/paper/mp-moe-matrix-profile-guided-mixture-of-experts-for-precipitation-forecasting - A novel AI framework for more accurate precipitation forecasting in tropical regions by integrating structural similarity with expert selection.
- ThermoAct:Thermal-Aware Vision-Language-Action Models for Robotic Perception and Decision-Making (viability: 4): https://sciencetostartup.com/paper/thermoact-thermal-aware-vision-language-action-models-for-robotic-perception-and-decision-making - A novel framework integrating thermal data into vision-language-action models for enhanced robotic perception and decision-making.
- Efficient ML-DSA Public Key Management Method with Identity for PKI and Its Application (viability: 7): https://sciencetostartup.com/paper/efficient-ml-dsa-public-key-management-method-with-identity-for-pki-and-its-application - A novel identity-based public key management framework using post-quantum cryptography to improve efficiency and scalability in PKI applications like RPKI.
- MoRGS: Efficient Per-Gaussian Motion Reasoning for Streamable Dynamic 3D Scenes (viability: 7): https://sciencetostartup.com/paper/morgs-efficient-per-gaussian-motion-reasoning-for-streamable-dynamic-3d-scenes - An efficient framework for real-time 3D scene reconstruction that explicitly models per-Gaussian motion for improved accuracy and temporal consistency.
- Intern-S1-Pro: Scientific Multimodal Foundation Model at Trillion Scale (viability: 3): https://sciencetostartup.com/paper/intern-s1-pro-scientific-multimodal-foundation-model-at-trillion-scale - A trillion-parameter multimodal foundation model for scientific tasks with enhanced agent capabilities.
- $π$, But Make It Fly: Physics-Guided Transfer of VLA Models to Aerial Manipulation (viability: 7): https://sciencetostartup.com/paper/but-make-it-fly-physics-guided-transfer-of-vla-models-to-aerial-manipulation - Enables pre-trained vision-language-action models to perform aerial manipulation tasks by bridging the dynamics gap with payload-aware guidance and synthetic data.
- GeoNDC: A Queryable Neural Data Cube for Planetary-Scale Earth Observation (viability: 5): https://sciencetostartup.com/paper/geondc-a-queryable-neural-data-cube-for-planetary-scale-earth-observation - GeoNDC transforms massive satellite imagery archives into a compact, queryable neural data cube for efficient analysis and reconstruction.
- Mechanistically Interpreting Compression in Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/mechanistically-interpreting-compression-in-vision-language-models - Develop safer and more efficient vision-language models by understanding and mitigating the safety implications of compression techniques.
- Epistemic Compression: The Case for Deliberate Ignorance in High-Stakes AI (viability: 2): https://sciencetostartup.com/paper/epistemic-compression-the-case-for-deliberate-ignorance-in-high-stakes-ai - A theoretical framework for building more robust AI by matching model complexity to data stability, rather than scaling parameters.
- From Stateless to Situated: Building a Psychological World for LLM-Based Emotional Support (viability: 4): https://sciencetostartup.com/paper/from-stateless-to-situated-building-a-psychological-world-for-llm-based-emotional-support - A new LLM architecture that maintains temporal continuity and user consent for emotional support by separating cognitive and executive layers.
- Optimal High-Probability Regret for Online Convex Optimization with Two-Point Bandit Feedback (viability: 2): https://sciencetostartup.com/paper/optimal-high-probability-regret-for-online-convex-optimization-with-two-point-bandit-feedback - This paper develops a theoretical framework for online convex optimization with limited feedback, achieving optimal regret bounds.
- CARE: Training-Free Controllable Restoration for Medical Images via Dual-Latent Steering (viability: 5): https://sciencetostartup.com/paper/care-training-free-controllable-restoration-for-medical-images-via-dual-latent-steering - A training-free framework for controllable medical image restoration that balances structure preservation and prior-guided enhancement during inference.
- System-Anchored Knee Estimation for Low-Cost Context Window Selection in PDE Forecasting (viability: 3): https://sciencetostartup.com/paper/system-anchored-knee-estimation-for-low-cost-context-window-selection-in-pde-forecasting - A novel algorithmic approach for low-cost context window selection in autoregressive neural PDE simulators.
- Improving Infinitely Deep Bayesian Neural Networks with Nesterov's Accelerated Gradient Method (viability: 3): https://sciencetostartup.com/paper/improving-infinitely-deep-bayesian-neural-networks-with-nesterov-s-accelerated-gradient-method - Accelerating Bayesian Neural Networks with Nesterov's method to reduce computational cost and improve convergence.
- A Public Theory of Distillation Resistance via Constraint-Coupled Reasoning Architectures (viability: 2): https://sciencetostartup.com/paper/a-public-theory-of-distillation-resistance-via-constraint-coupled-reasoning-architectures - A theoretical framework to make AI models more resistant to knowledge distillation by coupling capability to internal stability constraints.
- VideoTIR: Accurate Understanding for Long Videos with Efficient Tool-Integrated Reasoning (viability: 7): https://sciencetostartup.com/paper/videotir-accurate-understanding-for-long-videos-with-efficient-tool-integrated-reasoning - A novel RL-based approach for accurate and efficient long video understanding by intelligently using multi-level toolkits, reducing hallucinations and redundant calls.
- GDPO-Listener: Expressive Interactive Head Generation via Auto-Regressive Flow Matching and Group reward-Decoupled Policy Optimization (viability: 7): https://sciencetostartup.com/paper/gdpo-listener-expressive-interactive-head-generation-via-auto-regressive-flow-matching-and-group-reward-decoupled-policy - A framework for generating highly expressive and controllable 3D head motion for virtual human interactions, overcoming static face issues and enabling semantic text control.
- Discrete Causal Representation Learning (viability: 5): https://sciencetostartup.com/paper/discrete-causal-representation-learning - A generative framework for learning interpretable causal relationships among discrete latent variables from complex observational data.
- Imperative Interference: Social Register Shapes Instruction Topology in Large Language Models (viability: 3): https://sciencetostartup.com/paper/imperative-interference-social-register-shapes-instruction-topology-in-large-language-models - This research explores how social register in LLM prompts impacts instruction following across languages, suggesting a new avenue for language-dependent AI alignment.
- A Systematic Empirical Study of Grokking: Depth, Architecture, Activation, and Regularization (viability: 3): https://sciencetostartup.com/paper/a-systematic-empirical-study-of-grokking-depth-architecture-activation-and-regularization - This research systematically disentangles factors influencing grokking in neural networks, revealing that optimization stability and regularization, rather than architecture, are key drivers of delayed generalization.
- Few TensoRF: Enhance the Few-shot on Tensorial Radiance Fields (viability: 7): https://sciencetostartup.com/paper/few-tensorf-enhance-the-few-shot-on-tensorial-radiance-fields - A 3D reconstruction framework that achieves high quality and fast training times with sparse input views, enabling efficient real-time 3D scene generation.
- Improving Fine-Grained Rice Leaf Disease Detection via Angular-Compactness Dual Loss Learning (viability: 5): https://sciencetostartup.com/paper/improving-fine-grained-rice-leaf-disease-detection-via-angular-compactness-dual-loss-learning - A dual-loss deep learning framework to significantly improve fine-grained detection of rice leaf diseases, boosting accuracy beyond state-of-the-art.
- Interpretable Zero-shot Referring Expression Comprehension with Query-driven Scene Graphs (viability: 7): https://sciencetostartup.com/paper/interpretable-zero-shot-referring-expression-comprehension-with-query-driven-scene-graphs - An interpretable zero-shot object localization system that uses query-driven scene graphs to bridge vision-language models and large language models for precise target identification with explanations.
- Rethinking Failure Attribution in Multi-Agent Systems: A Multi-Perspective Benchmark and Evaluation (viability: 4): https://sciencetostartup.com/paper/rethinking-failure-attribution-in-multi-agent-systems-a-multi-perspective-benchmark-and-evaluation - A new benchmark and evaluation protocol for multi-perspective failure attribution in multi-agent systems, revealing limitations in current LLM failure attribution assessments.
- Distributed Real-Time Vehicle Control for Emergency Vehicle Transit: A Scalable Cooperative Method (viability: 7): https://sciencetostartup.com/paper/distributed-real-time-vehicle-control-for-emergency-vehicle-transit-a-scalable-cooperative-method - A scalable distributed system for emergency vehicle transit that enables cooperative driving behavior among vehicles using only local information, outperforming centralized and reinforcement learning methods in speed, impact, and scalability.
- Efficient Detection of Bad Benchmark Items with Novel Scalability Coefficients (viability: 7): https://sciencetostartup.com/paper/efficient-detection-of-bad-benchmark-items-with-novel-scalability-coefficients - A novel statistical method to efficiently identify flawed items in large-scale AI benchmarks and assessments, reducing manual review effort.
- IrisFP: Adversarial-Example-based Model Fingerprinting with Enhanced Uniqueness and Robustness (viability: 3): https://sciencetostartup.com/paper/irisfp-adversarial-example-based-model-fingerprinting-with-enhanced-uniqueness-and-robustness - A novel framework for model fingerprinting that enhances uniqueness and robustness using adversarial examples.
- Relaxed Rigidity with Ray-based Grouping for Dynamic Gaussian Splatting (viability: 7): https://sciencetostartup.com/paper/relaxed-rigidity-with-ray-based-grouping-for-dynamic-gaussian-splatting - A novel method for dynamic 3D scene reconstruction that enforces physically plausible motion by preserving local Gaussian geometry, outperforming existing methods on monocular datasets.
- C2W-Tune: Cavity-to -Wall Transfer Learning for Thin Atrial Wall Segmentation in 3D Late Gadolinium-enhanced Magnetic Resonance (viability: 7): https://sciencetostartup.com/paper/c2w-tune-cavity-to-wall-transfer-learning-for-thin-atrial-wall-segmentation-in-3d-late-gadolinium-enhanced-magnetic-reso - A transfer learning framework for precise thin atrial wall segmentation in cardiac MRI, improving diagnostic accuracy for heart conditions.
- Towards Video Anomaly Detection from Event Streams: A Baseline and Benchmark Datasets (viability: 7): https://sciencetostartup.com/paper/towards-video-anomaly-detection-from-event-streams-a-baseline-and-benchmark-datasets - Develops a novel framework and benchmark datasets for event-based video anomaly detection, offering a privacy-preserving and efficient solution.
- Learning Rollout from Sampling:An R1-Style Tokenized Traffic Simulation Model (viability: 7): https://sciencetostartup.com/paper/learning-rollout-from-sampling-an-r1-style-tokenized-traffic-simulation-model - A novel reinforcement learning approach for generating diverse and realistic autonomous driving simulations by exploring uncertain motion patterns.
- Rethinking Health Agents: From Siloed AI to Collaborative Decision Mediators (viability: 4): https://sciencetostartup.com/paper/rethinking-health-agents-from-siloed-ai-to-collaborative-decision-mediators - Develops a conceptual framework for collaborative AI agents in healthcare to improve multi-stakeholder communication and decision-making, addressing fragmentation caused by siloed AI tools.
- Few-Shot Left Atrial Wall Segmentation in 3D LGE MRI via Meta-Learning (viability: 5): https://sciencetostartup.com/paper/few-shot-left-atrial-wall-segmentation-in-3d-lge-mri-via-meta-learning - A meta-learning framework for accurate 3D left atrial wall segmentation in MRI, enabling clinical translation with minimal labeling.
- MoE-GRPO: Optimizing Mixture-of-Experts via Reinforcement Learning in Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/moe-grpo-optimizing-mixture-of-experts-via-reinforcement-learning-in-vision-language-models - Optimize vision-language model performance by using reinforcement learning to dynamically route tokens to the most relevant expert parameters, reducing computational cost and improving accuracy.
- LiteGuard: Efficient Task-Agnostic Model Fingerprinting with Enhanced Generalization (viability: 3): https://sciencetostartup.com/paper/liteguard-efficient-task-agnostic-model-fingerprinting-with-enhanced-generalization - A novel framework for efficient and generalized AI model fingerprinting that reduces computational cost and overfitting.
- Exons-Detect: Identifying and Amplifying Exonic Tokens via Hidden-State Discrepancy for Robust AI-Generated Text Detection (viability: 7): https://sciencetostartup.com/paper/exons-detect-identifying-and-amplifying-exonic-tokens-via-hidden-state-discrepancy-for-robust-ai-generated-text-detectio - A training-free AI-generated text detection method that amplifies informative tokens for improved accuracy and robustness.
- LLM-Driven Reasoning for Constraint-Aware Feature Selection in Industrial Systems (viability: 7): https://sciencetostartup.com/paper/llm-driven-reasoning-for-constraint-aware-feature-selection-in-industrial-systems - An LLM-powered framework for intelligent, constraint-aware feature selection in industrial ML systems, improving accuracy and efficiency.
- PASDiff: Physics-Aware Semantic Guidance for Joint Real-world Low-Light Face Enhancement and Restoration (viability: 7): https://sciencetostartup.com/paper/pasdiff-physics-aware-semantic-guidance-for-joint-real-world-low-light-face-enhancement-and-restoration - A physics-aware diffusion model for real-world low-light face enhancement and restoration that significantly outperforms existing methods.
- The Anatomy of Uncertainty in LLMs (viability: 4): https://sciencetostartup.com/paper/the-anatomy-of-uncertainty-in-llms - A framework to decompose LLM uncertainty into input ambiguity, knowledge gaps, and decoding randomness for improved reliability and hallucination detection.
- Self-Corrected Image Generation with Explainable Latent Rewards (viability: 7): https://sciencetostartup.com/paper/self-corrected-image-generation-with-explainable-latent-rewards - A self-correcting image generation framework that uses explainable latent rewards to improve alignment with complex prompts.
- Design Once, Deploy at Scale: Template-Driven ML Development for Large Model Ecosystems (viability: 7): https://sciencetostartup.com/paper/design-once-deploy-at-scale-template-driven-ml-development-for-large-model-ecosystems - A framework for standardized ML model development that drastically reduces engineering time and accelerates innovation in large-scale recommendation systems.
- Can MLLMs Read Students' Minds? Unpacking Multimodal Error Analysis in Handwritten Math (viability: 7): https://sciencetostartup.com/paper/can-mllms-read-students-minds-unpacking-multimodal-error-analysis-in-handwritten-math - A new benchmark and evaluation of multimodal LLMs for diagnosing student errors in handwritten math scratchwork, revealing significant performance gaps and highlighting proprietary models' potential.
- Wireless bioelectronics for untethered biohybrid robots (viability: 2): https://sciencetostartup.com/paper/wireless-bioelectronics-for-untethered-biohybrid-robots - Developing untethered control for biohybrid robots using wireless bioelectronics and neuromuscular integration.
- Toward domain-specific machine translation and quality estimation systems (viability: 7): https://sciencetostartup.com/paper/toward-domain-specific-machine-translation-and-quality-estimation-systems - Builds domain-specific machine translation and quality estimation systems by leveraging data selection, staged training, and QE-guided in-context learning for large language models.
- Select, Hypothesize and Verify: Towards Verified Neuron Concept Interpretation (viability: 3): https://sciencetostartup.com/paper/select-hypothesize-and-verify-towards-verified-neuron-concept-interpretation - A framework for verifying and interpreting the functionality of individual neurons within neural networks to prevent misinterpretations.
- Shopping with a Platform AI Assistant: Who Adopts, When in the Journey, and What For (viability: 5): https://sciencetostartup.com/paper/shopping-with-a-platform-ai-assistant-who-adopts-when-in-the-journey-and-what-for - This research reveals how platform-embedded AI assistants are adopted and used by consumers for exploratory shopping, offering insights for product development and user engagement strategies.
- MobileDev-Bench: A Comprehensive Benchmark for Evaluating Language Models on Mobile Application Development (viability: 7): https://sciencetostartup.com/paper/mobiledev-bench-a-comprehensive-benchmark-for-evaluating-language-models-on-mobile-application-development - A benchmark for evaluating LLMs on mobile app development, revealing significant performance gaps and identifying key bottlenecks for future model improvement.
- FinMCP-Bench: Benchmarking LLM Agents for Real-World Financial Tool Use under the Model Context Protocol (viability: 7): https://sciencetostartup.com/paper/finmcp-bench-benchmarking-llm-agents-for-real-world-financial-tool-use-under-the-model-context-protocol - A new benchmark for evaluating LLM agents' ability to use real-world financial tools, enabling more reliable financial applications.
- BiFM: Bidirectional Flow Matching for Few-Step Image Editing and Generation (viability: 7): https://sciencetostartup.com/paper/bifm-bidirectional-flow-matching-for-few-step-image-editing-and-generation - A unified framework for few-step image editing and generation that jointly learns generation and inversion, outperforming existing methods.
- Beyond Attention Magnitude: Leveraging Inter-layer Rank Consistency for Efficient Vision-Language-Action Models (viability: 7): https://sciencetostartup.com/paper/beyond-attention-magnitude-leveraging-inter-layer-rank-consistency-for-efficient-vision-language-action-models - A dynamic framework for efficient vision-language-action models that reduces inference latency by 78% while improving performance through inter-layer rank consistency.
- Evaluating adaptive and generative AI-based feedback and recommendations in a knowledge-graph-integrated programming learning system (viability: 7): https://sciencetostartup.com/paper/evaluating-adaptive-and-generative-ai-based-feedback-and-recommendations-in-a-knowledge-graph-integrated-programming-lea - An AI-powered adaptive learning system that uses LLMs and knowledge graphs to provide personalized code feedback and exercise recommendations, outperforming traditional adaptive methods.
- Infinite Gaze Generation for Videos with Autoregressive Diffusion (viability: 7): https://sciencetostartup.com/paper/infinite-gaze-generation-for-videos-with-autoregressive-diffusion - Generate realistic, long-range human gaze trajectories for any video length to enhance scene understanding and human-computer interaction.
- TIGFlow-GRPO: Trajectory Forecasting via Interaction-Aware Flow Matching and Reward-Driven Optimization (viability: 7): https://sciencetostartup.com/paper/tigflow-grpo-trajectory-forecasting-via-interaction-aware-flow-matching-and-reward-driven-optimization - A generative framework for human trajectory forecasting that aligns flow-based generation with behavioral rules for more socially compliant and physically feasible predictions.
- SABER: A Stealthy Agentic Black-Box Attack Framework for Vision-Language-Action Models (viability: 7): https://sciencetostartup.com/paper/saber-a-stealthy-agentic-black-box-attack-framework-for-vision-language-action-models - SABER is an agent-based framework that automatically generates stealthy, minimal textual attacks to identify and fix vulnerabilities in vision-language-action robotic models.
- CVA: Context-aware Video-text Alignment for Video Temporal Grounding (viability: 7): https://sciencetostartup.com/paper/cva-context-aware-video-text-alignment-for-video-temporal-grounding - A novel framework for video temporal grounding that significantly improves accuracy by intelligently handling irrelevant background context.
- Decoding Market Emotions in Cryptocurrency Tweets via Predictive Statement Classification with Machine Learning and Transformers (viability: 5): https://sciencetostartup.com/paper/decoding-market-emotions-in-cryptocurrency-tweets-via-predictive-statement-classification-with-machine-learning-and-tran - A novel framework for classifying predictive statements in cryptocurrency tweets, leveraging machine learning and transformers to understand market emotions.
- COIN: Collaborative Interaction-Aware Multi-Agent Reinforcement Learning for Self-Driving Systems (viability: 8): https://sciencetostartup.com/paper/coin-collaborative-interaction-aware-multi-agent-reinforcement-learning-for-self-driving-systems - A novel MARL framework for self-driving systems that significantly improves safety and efficiency in dense urban traffic through collaborative interaction-aware learning.
- CROSS: A Mixture-of-Experts Reinforcement Learning Framework for Generalizable Large-Scale Traffic Signal Control (viability: 7): https://sciencetostartup.com/paper/cross-a-mixture-of-experts-reinforcement-learning-framework-for-generalizable-large-scale-traffic-signal-control - A Mixture-of-Experts Reinforcement Learning framework that uses predictive clustering to adapt traffic signal control to diverse and dynamic urban traffic patterns, outperforming existing methods.
- LogitScope: A Framework for Analyzing LLM Uncertainty Through Information Metrics (viability: 7): https://sciencetostartup.com/paper/logitscope-a-framework-for-analyzing-llm-uncertainty-through-information-metrics - LogitScope provides a lightweight framework to analyze LLM uncertainty at the token level, enabling better detection of hallucinations and improved model monitoring.
- GraphER: An Efficient Graph-Based Enrichment and Reranking Method for Retrieval-Augmented Generation (viability: 7): https://sciencetostartup.com/paper/grapher-an-efficient-graph-based-enrichment-and-reranking-method-for-retrieval-augmented-generation - GraphER enhances retrieval-augmented generation by efficiently enriching and reranking candidate documents using graph-based proximity, integrating seamlessly with existing vector stores.
- Estimating near-verbatim extraction risk in language models with decoding-constrained beam search (viability: 4): https://sciencetostartup.com/paper/estimating-near-verbatim-extraction-risk-in-language-models-with-decoding-constrained-beam-search - A novel decoding-constrained beam search method quantifies near-verbatim extraction risk in LLMs at a fraction of the computational cost of existing probabilistic methods.
- Once-for-All Channel Mixers (HYPERTINYPW): Generative Compression for TinyML (viability: 7): https://sciencetostartup.com/paper/once-for-all-channel-mixers-hypertinypw-generative-compression-for-tinyml - Generates neural network weights on-the-fly for extreme memory-constrained microcontrollers, achieving significant size reduction without sacrificing performance.
- ICTPolarReal: A Polarized Reflection and Material Dataset of Real World Objects (viability: 7): https://sciencetostartup.com/paper/ictpolarreal-a-polarized-reflection-and-material-dataset-of-real-world-objects - A new dataset and rendering models for accurate material understanding in real-world images, overcoming limitations of synthetic data.
- Integrated Multi-Drone Task Allocation, Sequencing, and Optimal Trajectory Generation in Obstacle-Rich 3D Environments (viability: 5): https://sciencetostartup.com/paper/integrated-multi-drone-task-allocation-sequencing-and-optimal-trajectory-generation-in-obstacle-rich-3d-environments - An integrated framework for multi-drone task allocation and safe trajectory generation in complex 3D environments.
- On the Foundations of Trustworthy Artificial Intelligence (viability: 3): https://sciencetostartup.com/paper/on-the-foundations-of-trustworthy-artificial-intelligence - A theoretical framework proving platform determinism is essential for trustworthy AI, proposing an integer-based inference engine to achieve bitwise identical outputs across architectures.
- Self-Supervised Learning for Knee Osteoarthritis: Diagnostic Limitations and Prognostic Value of Uncurated Hospital Data (viability: 5): https://sciencetostartup.com/paper/self-supervised-learning-for-knee-osteoarthritis-diagnostic-limitations-and-prognostic-value-of-uncurated-hospital-data - Leveraging self-supervised learning on uncurated hospital image-text data to significantly improve knee osteoarthritis prognosis, outperforming standard pretraining methods.
- Sovereign AI at the Front Door of Care: A Physically Unidirectional Architecture for Secure Clinical Intelligence (viability: 7): https://sciencetostartup.com/paper/sovereign-ai-at-the-front-door-of-care-a-physically-unidirectional-architecture-for-secure-clinical-intelligence - A physically unidirectional AI architecture for secure, on-device clinical triage at the point of care.
- SurgPhase: Time efficient pituitary tumor surgery phase recognition via an interactive web platform (viability: 8): https://sciencetostartup.com/paper/surgphase-time-efficient-pituitary-tumor-surgery-phase-recognition-via-an-interactive-web-platform - An AI-powered web platform that automatically recognizes surgical phases in pituitary tumor surgeries, enabling data-driven improvements in training and performance.
- LogSigma at SemEval-2026 Task 3: Uncertainty-Weighted Multitask Learning for Dimensional Aspect-Based Sentiment Analysis (viability: 7): https://sciencetostartup.com/paper/logsigma-at-semeval-2026-task-3-uncertainty-weighted-multitask-learning-for-dimensional-aspect-based-sentiment-analysis - A system that achieves state-of-the-art performance in dimensional aspect-based sentiment analysis by learning to dynamically balance prediction difficulty across languages and tasks.
- Surrogates, Spikes, and Sparsity: Performance Analysis and Characterization of SNN Hyperparameters on Hardware (viability: 7): https://sciencetostartup.com/paper/surrogates-spikes-and-sparsity-performance-analysis-and-characterization-of-snn-hyperparameters-on-hardware - This research provides a methodology to optimize Spiking Neural Network hardware performance by selecting specific training hyperparameters, leading to significant reductions in latency and improvements in accuracy.
- An Approach to Generate Attack Graphs with a Case Study on Siemens PCS7 Blueprint for Water Treatment Plants (viability: 4): https://sciencetostartup.com/paper/an-approach-to-generate-attack-graphs-with-a-case-study-on-siemens-pcs7-blueprint-for-water-treatment-plants - A semi-automated framework to generate attack graphs for Industrial Control Systems, visualizing multi-step attack scenarios and providing actionable security insights.
- Learning to Staff: Offline Reinforcement Learning and Fine-Tuned LLMs for Warehouse Staffing Optimization (viability: 7): https://sciencetostartup.com/paper/learning-to-staff-offline-reinforcement-learning-and-fine-tuned-llms-for-warehouse-staffing-optimization - AI-powered system to optimize warehouse staffing decisions using offline reinforcement learning and fine-tuned LLMs, leading to significant throughput improvements and cost savings.
- Trusted-Execution Environment (TEE) for Solving the Replication Crisis in Academia (viability: 7): https://sciencetostartup.com/paper/trusted-execution-environment-tee-for-solving-the-replication-crisis-in-academia - A TEE-based framework for academic journals to cryptographically verify research replication without re-running code, solving the replication crisis efficiently and scalably.
- More Than "Means to an End": Supporting Reasoning with Transparently Designed AI Data Science Processes (viability: 3): https://sciencetostartup.com/paper/more-than-means-to-an-end-supporting-reasoning-with-transparently-designed-ai-data-science-processes - This paper explores how intermediate AI artifacts can enhance user reasoning in data science tasks, particularly in high-stakes domains like medicine.
- OptiSAR-Net++: A Large-Scale Benchmark and Transformer-Free Framework for Cross-Domain Remote Sensing Visual Grounding (viability: 7): https://sciencetostartup.com/paper/optisar-net-a-large-scale-benchmark-and-transformer-free-framework-for-cross-domain-remote-sensing-visual-grounding - A transformer-free framework and large-scale benchmark for cross-domain remote sensing visual grounding, enabling precise target localization across optical and SAR imagery.
- How Far Are Vision-Language Models from Constructing the Real World? A Benchmark for Physical Generative Reasoning (viability: 7): https://sciencetostartup.com/paper/how-far-are-vision-language-models-from-constructing-the-real-world-a-benchmark-for-physical-generative-reasoning - A new benchmark and dataset for evaluating vision-language models' ability to reason about and construct physical objects based on structural and code compliance constraints.
- AI Security in the Foundation Model Era: A Comprehensive Survey from a Unified Perspective (viability: 3): https://sciencetostartup.com/paper/ai-security-in-the-foundation-model-era-a-comprehensive-survey-from-a-unified-perspective - A unified taxonomy for understanding and defending against AI security threats in foundation models by framing model-data interactions.
- SentinelAI: A Multi-Agent Framework for Structuring and Linking NG9-1-1 Emergency Incident Data (viability: 5): https://sciencetostartup.com/paper/sentinelai-a-multi-agent-framework-for-structuring-and-linking-ng9-1-1-emergency-incident-data - A framework to structure and link emergency incident data from multiple sources into a unified, machine-readable format for improved emergency response.
- Resisting Humanization: Ethical Front-End Design Choices in AI for Sensitive Contexts (viability: 3): https://sciencetostartup.com/paper/resisting-humanization-ethical-front-end-design-choices-in-ai-for-sensitive-contexts - This research explores ethical front-end design choices in AI conversational interfaces to prevent user misalignments and foster autonomy, particularly in sensitive contexts.
- Towards automatic smoke detector inspection: Recognition of the smoke detectors in industrial facilities and preparation for future drone integration (viability: 7): https://sciencetostartup.com/paper/towards-automatic-smoke-detector-inspection-recognition-of-the-smoke-detectors-in-industrial-facilities-and-preparation - An automated drone-integrated system for inspecting smoke detectors in industrial facilities, reducing costs and improving safety.
- Gaze patterns predict preference and confidence in pairwise AI image evaluation (viability: 4): https://sciencetostartup.com/paper/gaze-patterns-predict-preference-and-confidence-in-pairwise-ai-image-evaluation - Leveraging eye-tracking to understand and predict human preferences in AI image evaluation.
- CORA: A Pathology Synthesis Driven Foundation Model for Coronary CT Angiography Analysis and MACE Risk Assessment (viability: 7): https://sciencetostartup.com/paper/cora-a-pathology-synthesis-driven-foundation-model-for-coronary-ct-angiography-analysis-and-mace-risk-assessment - A pathology-aware foundation model for coronary artery analysis and cardiovascular risk prediction, trained on synthesized lesions to outperform existing methods.
- NeuroVLM-Bench: Evaluation of Vision-Enabled Large Language Models for Clinical Reasoning in Neurological Disorders (viability: 7): https://sciencetostartup.com/paper/neurovlm-bench-evaluation-of-vision-enabled-large-language-models-for-clinical-reasoning-in-neurological-disorders - Benchmarking vision-enabled LLMs for clinical reasoning in neurological disorders to identify reliable diagnostic tools and efficiency trade-offs.
- Reaching Beyond the Mode: RL for Distributional Reasoning in Language Models (viability: 7): https://sciencetostartup.com/paper/reaching-beyond-the-mode-rl-for-distributional-reasoning-in-language-models - Train language models to generate diverse, well-calibrated answers for tasks with inherent uncertainty, offering a compute-efficient alternative to repeated sampling.
- Prune as You Generate: Online Rollout Pruning for Faster and Better RLVR (viability: 7): https://sciencetostartup.com/paper/prune-as-you-generate-online-rollout-pruning-for-faster-and-better-rlvr - Accelerate LLM training for verifiable rewards by pruning rollouts online, improving accuracy and speed.
- Bridging Code Property Graphs and Language Models for Program Analysis (viability: 7): https://sciencetostartup.com/paper/bridging-code-property-graphs-and-language-models-for-program-analysis - A server that integrates code property graphs with LLMs to enable semantic code analysis across entire repositories for vulnerability discovery and patching.
- WAFT-Stereo: Warping-Alone Field Transforms for Stereo Matching (viability: 9): https://sciencetostartup.com/paper/waft-stereo-warping-alone-field-transforms-for-stereo-matching - A revolutionary warping-based stereo matching solution that outperforms existing methods in accuracy and speed.
- DCARL: A Divide-and-Conquer Framework for Autoregressive Long-Trajectory Video Generation (viability: 7): https://sciencetostartup.com/paper/dcarl-a-divide-and-conquer-framework-for-autoregressive-long-trajectory-video-generation - A framework for generating stable and high-fidelity long-trajectory videos by combining keyframe generation with autoregressive interpolation.
- Robust Matrix Estimation with Side Information (viability: 3): https://sciencetostartup.com/paper/robust-matrix-estimation-with-side-information - A flexible framework for high-dimensional matrix estimation that incorporates side information for improved imputation and treatment-effect estimation.
- Flow matching on homogeneous spaces (viability: 3): https://sciencetostartup.com/paper/flow-matching-on-homogeneous-spaces - A theoretical framework for extending flow matching to complex geometric spaces by simplifying the problem to Euclidean flow matching on Lie algebras.
- A Practical Guide Towards Interpreting Time-Series Deep Clinical Predictive Models: A Reproducibility Study (viability: 7): https://sciencetostartup.com/paper/a-practical-guide-towards-interpreting-time-series-deep-clinical-predictive-models-a-reproducibility-study - A reproducibility study providing guidelines and an open-source framework for interpretable deep clinical predictive models.
- Synthetic Rewriting as a Quality Multiplier: Evidence from Portuguese Continued Pretraining (viability: 4): https://sciencetostartup.com/paper/synthetic-rewriting-as-a-quality-multiplier-evidence-from-portuguese-continued-pretraining - This research explores how synthetic data rewriting can improve Portuguese language model pretraining, acting as a quality multiplier dependent on source data quality and model scale.
- Learning From Developers: Towards Reliable Patch Validation at Scale for Linux (viability: 5): https://sciencetostartup.com/paper/learning-from-developers-towards-reliable-patch-validation-at-scale-for-linux - FLINT is a patch validation system that leverages past developer discussions and LLMs to automatically analyze Linux patch proposals, improving detection of complex bugs.
- Generative Adversarial Perturbations with Cross-paradigm Transferability on Localized Crowd Counting (viability: 7): https://sciencetostartup.com/paper/generative-adversarial-perturbations-with-cross-paradigm-transferability-on-localized-crowd-counting - A novel adversarial framework that compromises both density map and point regression crowd counting models with imperceptible perturbations, offering a new security layer for visual surveillance systems.
- Attention-based Pin Site Image Classification in Orthopaedic Patients with External Fixators (viability: 7): https://sciencetostartup.com/paper/attention-based-pin-site-image-classification-in-orthopaedic-patients-with-external-fixators - An attention-based deep learning model for early detection of pin site infections in orthopedic patients, improving patient outcomes and reducing morbidity.
- Characterization of Constraints in Flexible Unknown Environments (viability: 5): https://sciencetostartup.com/paper/characterization-of-constraints-in-flexible-unknown-environments - Enables robots to safely manipulate unknown flexible objects by simultaneously exploring and characterizing their constraints in real-time.
- A Nonvolatile Switchable-polarity EPM Valve (viability: 3): https://sciencetostartup.com/paper/a-nonvolatile-switchable-polarity-epm-valve - A novel bistable magnetic valve architecture enables power-free, programmable control of fluidic networks for autonomous lab platforms.
- FODMP: Fast One-Step Diffusion of Movement Primitives Generation for Time-Dependent Robot Actions (viability: 7): https://sciencetostartup.com/paper/fodmp-fast-one-step-diffusion-of-movement-primitives-generation-for-time-dependent-robot-actions - A novel framework for robots to generate complex, time-dependent movements in a single step, enabling real-time control and object interception.
- GoldiCLIP: The Goldilocks Approach for Balancing Explicit Supervision for Language-Image Pretraining (viability: 7): https://sciencetostartup.com/paper/goldiclip-the-goldilocks-approach-for-balancing-explicit-supervision-for-language-image-pretraining - A data-efficient framework for vision-language pretraining that achieves state-of-the-art results with significantly less data.
- Dissecting Model Failures in Abdominal Aortic Aneurysm Segmentation through Explainability-Driven Analysis (viability: 4): https://sciencetostartup.com/paper/dissecting-model-failures-in-abdominal-aortic-aneurysm-segmentation-through-explainability-driven-analysis - An explainable AI framework to improve abdominal aortic aneurysm segmentation by guiding model focus to relevant structures.
- Calibri: Enhancing Diffusion Transformers via Parameter-Efficient Calibration (viability: 7): https://sciencetostartup.com/paper/calibri-enhancing-diffusion-transformers-via-parameter-efficient-calibration - Calibri enhances diffusion transformer image generation quality and speed with a parameter-efficient calibration method.
- Enhancing Structured Meaning Representations with Aspect Classification (viability: 5): https://sciencetostartup.com/paper/enhancing-structured-meaning-representations-with-aspect-classification - This research introduces a new dataset and baseline models for automatically predicting aspectual information in semantic meaning representations, aiming to enhance the understanding of event structures in sentences.
- AVControl: Efficient Framework for Training Audio-Visual Controls (viability: 8): https://sciencetostartup.com/paper/avcontrol-efficient-framework-for-training-audio-visual-controls - A lightweight, extendable framework for efficient audio-visual control in video generation, enabling modular training of diverse modalities with minimal architectural changes.
- Local learning for stable backpropagation-free neural network training towards physical learning (viability: 3): https://sciencetostartup.com/paper/local-learning-for-stable-backpropagation-free-neural-network-training-towards-physical-learning - A novel framework for training neural networks without backpropagation, enabling physical hardware implementations.
- ReLope: KL-Regularized LoRA Probes for Multimodal LLM Routing (viability: 7): https://sciencetostartup.com/paper/relope-kl-regularized-lora-probes-for-multimodal-llm-routing - Improve multimodal LLM routing efficiency by enhancing the quality of hidden states with attention-based probes and KL-regularized LoRA adapters.
- Transformers in the Dark: Navigating Unknown Search Spaces via Bandit Feedback (viability: 7): https://sciencetostartup.com/paper/transformers-in-the-dark-navigating-unknown-search-spaces-via-bandit-feedback - Train LLMs to act as internal search algorithms, navigating complex problem spaces more efficiently without external components.
- AIP: Agent Identity Protocol for Verifiable Delegation Across MCP and A2A (viability: 7): https://sciencetostartup.com/paper/aip-agent-identity-protocol-for-verifiable-delegation-across-mcp-and-a2a - A verifiable delegation protocol for AI agents that secures tool calls and inter-agent communication with minimal latency overhead.
- From Untestable to Testable: Metamorphic Testing in the Age of LLMs (viability: 3): https://sciencetostartup.com/paper/from-untestable-to-testable-metamorphic-testing-in-the-age-of-llms - A novel approach to testing AI-integrated software by leveraging relationships between test executions to create scalable oracles.
- Evaluating Fine-Tuned LLM Model For Medical Transcription With Small Low-Resource Languages Validated Dataset (viability: 5): https://sciencetostartup.com/paper/evaluating-fine-tuned-llm-model-for-medical-transcription-with-small-low-resource-languages-validated-dataset - Fine-tuning LLaMA 3.1-8B on a small Finnish medical transcription dataset shows promise for reducing physician burnout in low-resource languages.
- Identifiable Deep Latent Variable Models for MNAR Data (viability: 4): https://sciencetostartup.com/paper/identifiable-deep-latent-variable-models-for-mnar-data - A novel deep latent variable model framework for non-random missing data imputation that guarantees identifiability and improves accuracy over existing methods.
- DRoPS: Dynamic 3D Reconstruction of Pre-Scanned Objects (viability: 7): https://sciencetostartup.com/paper/drops-dynamic-3d-reconstruction-of-pre-scanned-objects - A novel 3D reconstruction method that uses pre-scanned object data to achieve superior geometric consistency and rendering quality for dynamic scenes, outperforming state-of-the-art.
- Supervising Ralph Wiggum: Exploring a Metacognitive Co-Regulation Agentic AI Loop for Engineering Design (viability: 4): https://sciencetostartup.com/paper/supervising-ralph-wiggum-exploring-a-metacognitive-co-regulation-agentic-ai-loop-for-engineering-design - An AI agent that uses metacognition to overcome design fixation and improve engineering design solutions.
- Fine-Tuning A Large Language Model for Systematic Review Screening (viability: 7): https://sciencetostartup.com/paper/fine-tuning-a-large-language-model-for-systematic-review-screening - Fine-tune LLMs to dramatically accelerate systematic review screening, achieving high agreement with human coders.
- Enhancing Online Support Group Formation Using Topic Modeling Techniques (viability: 7): https://sciencetostartup.com/paper/enhancing-online-support-group-formation-using-topic-modeling-techniques - Automate personalized and semantically coherent online support group formation using novel topic modeling techniques, outperforming existing methods and enabling scalable community building.
- Synthetic Cardiac MRI Image Generation using Deep Generative Models (viability: 5): https://sciencetostartup.com/paper/synthetic-cardiac-mri-image-generation-using-deep-generative-models - Generate synthetic cardiac MRI images to overcome data scarcity and improve downstream medical imaging tasks.
- SlopCodeBench: Benchmarking How Coding Agents Degrade Over Long-Horizon Iterative Tasks (viability: 4): https://sciencetostartup.com/paper/slopcodebench-benchmarking-how-coding-agents-degrade-over-long-horizon-iterative-tasks - A new benchmark reveals that current coding agents degrade significantly in code quality during iterative development, highlighting a critical gap for practical software engineering.
- An Explainable Federated Framework for Zero Trust Micro-Segmentation in IIoT Networks (viability: 7): https://sciencetostartup.com/paper/an-explainable-federated-framework-for-zero-trust-micro-segmentation-in-iiot-networks - An explainable federated framework for zero trust micro-segmentation in IIoT networks that significantly improves security efficacy and provides trustworthy policy decisions.
- Light Cones For Vision: Simple Causal Priors For Visual Hierarchy (viability: 7): https://sciencetostartup.com/paper/light-cones-for-vision-simple-causal-priors-for-visual-hierarchy - A novel vision architecture using Lorentzian geometry to model causal relationships between objects, achieving significant improvements in hierarchical structure discovery with minimal parameters.
- Pseudo Label NCF for Sparse OHC Recommendation: Dual Representation Learning and the Separability Accuracy Trade off (viability: 5): https://sciencetostartup.com/paper/pseudo-label-ncf-for-sparse-ohc-recommendation-dual-representation-learning-and-the-separability-accuracy-trade-off - This research proposes a novel recommendation system for online health communities that leverages survey data to overcome extreme user sparsity, improving personalized support group discovery.
- TIGeR: A Unified Framework for Time, Images and Geo-location Retrieval (viability: 7): https://sciencetostartup.com/paper/tiger-a-unified-framework-for-time-images-and-geo-location-retrieval - A unified framework for retrieving images based on location and time, outperforming state-of-the-art by up to 16%.
- Formal Semantics for Agentic Tool Protocols: A Process Calculus Approach (viability: 7): https://sciencetostartup.com/paper/formal-semantics-for-agentic-tool-protocols-a-process-calculus-approach - Formalizing agent-tool interaction protocols to ensure verifiable and robust agent behavior, addressing critical gaps in current industry standards.
- Grokking as a Falsifiable Finite-Size Transition (viability: 2): https://sciencetostartup.com/paper/grokking-as-a-falsifiable-finite-size-transition - This paper explores the theoretical underpinnings of 'grokking' in machine learning, a phenomenon where models generalize after memorizing data, by applying statistical mechanics principles to understand its phase transition behavior.
- Contrastive Learning Boosts Deterministic and Generative Models for Weather Data (viability: 5): https://sciencetostartup.com/paper/contrastive-learning-boosts-deterministic-and-generative-models-for-weather-data - Contrastive learning for weather data compression to improve forecasting and extreme-weather detection, especially with sparse data.
- Trust as Monitoring: Evolutionary Dynamics of User Trust and AI Developer Behaviour (viability: 2): https://sciencetostartup.com/paper/trust-as-monitoring-evolutionary-dynamics-of-user-trust-and-ai-developer-behaviour - This research uses evolutionary game theory and simulations to model how user trust and AI developer behavior co-evolve, identifying conditions for safe AI adoption.
- Decentralized Task Scheduling in Distributed Systems: A Deep Reinforcement Learning Approach (viability: 7): https://sciencetostartup.com/paper/decentralized-task-scheduling-in-distributed-systems-a-deep-reinforcement-learning-approach - A decentralized multi-agent deep reinforcement learning framework for efficient task scheduling in distributed systems, offering significant improvements in completion time, energy efficiency, and SLA satisfaction.
- AutoSAM: an Agentic Framework for Automating Input File Generation for the SAM Code with Multi-Modal Retrieval-Augmented Generation (viability: 7): https://sciencetostartup.com/paper/autosam-an-agentic-framework-for-automating-input-file-generation-for-the-sam-code-with-multi-modal-retrieval-augmented - An agentic framework that automates the creation of complex engineering simulation input files from unstructured design documents.
- OpenCap Monocular: 3D Human Kinematics and Musculoskeletal Dynamics from a Single Smartphone Video (viability: 8): https://sciencetostartup.com/paper/opencap-monocular-3d-human-kinematics-and-musculoskeletal-dynamics-from-a-single-smartphone-video - OpenCap Monocular turns any smartphone into a 3D movement analytics tool for musculoskeletal insights.
- A Framework for Generating Semantically Ambiguous Images to Probe Human and Machine Perception (viability: 5): https://sciencetostartup.com/paper/a-framework-for-generating-semantically-ambiguous-images-to-probe-human-and-machine-perception - A framework for generating semantically ambiguous images to probe and align human and machine perception.
- Confidence-Based Mesh Extraction from 3D Gaussians (viability: 7): https://sciencetostartup.com/paper/confidence-based-mesh-extraction-from-3d-gaussians - A self-supervised confidence framework for efficient and accurate mesh extraction from 3D Gaussian Splatting, outperforming state-of-the-art for unbounded meshes.
- Is Geometry Enough? An Evaluation of Landmark-Based Gaze Estimation (viability: 7): https://sciencetostartup.com/paper/is-geometry-enough-an-evaluation-of-landmark-based-gaze-estimation - Develop efficient and interpretable gaze estimation for edge devices using landmark-based geometric features, outperforming traditional methods in cross-domain generalization.
- Scalable Object Relation Encoding for Better 3D Spatial Reasoning in Large Language Models (viability: 7): https://sciencetostartup.com/paper/scalable-object-relation-encoding-for-better-3d-spatial-reasoning-in-large-language-models - A novel positional embedding method for LLMs that enables scalable and accurate 3D spatial reasoning by efficiently encoding object relationships.
- Accurate Point Measurement in 3DGS -- A New Alternative to Traditional Stereoscopic-View Based Measurements (viability: 8): https://sciencetostartup.com/paper/accurate-point-measurement-in-3dgs-a-new-alternative-to-traditional-stereoscopic-view-based-measurements - A web-based tool leveraging 3D Gaussian Splatting for highly accurate and accessible 3D point measurements, outperforming traditional methods on challenging structures.
- Can an Actor-Critic Optimization Framework Improve Analog Design Optimization? (viability: 4): https://sciencetostartup.com/paper/can-an-actor-critic-optimization-framework-improve-analog-design-optimization - An actor-critic optimization framework to accelerate analog design by intelligently guiding the search space.
- Lookalike3D: Seeing Double in 3D (viability: 8): https://sciencetostartup.com/paper/lookalike3d-seeing-double-in-3d - A novel 3D object understanding system that leverages repeated objects to improve reconstruction and perception quality, with code and dataset released.
- Training LLMs for Multi-Step Tool Orchestration with Constrained Data Synthesis and Graduated Rewards (viability: 7): https://sciencetostartup.com/paper/training-llms-for-multi-step-tool-orchestration-with-constrained-data-synthesis-and-graduated-rewards - A framework for training LLMs to reliably orchestrate multi-step API calls using synthesized data and graduated rewards, significantly improving accuracy on complex tasks.
- Amortized Inference for Correlated Discrete Choice Models via Equivariant Neural Networks (viability: 4): https://sciencetostartup.com/paper/amortized-inference-for-correlated-discrete-choice-models-via-equivariant-neural-networks - A neural network emulator for discrete choice models that offers faster and more accurate likelihood evaluation than existing simulators.
- Conformal Selective Prediction with General Risk Control (viability: 7): https://sciencetostartup.com/paper/conformal-selective-prediction-with-general-risk-control - A framework for AI models to precisely control prediction risk, ensuring reliability in critical applications like drug discovery and healthcare.
- IndustriConnect: MCP Adapters and Mock-First Evaluation for AI-Assisted Industrial Operations (viability: 7): https://sciencetostartup.com/paper/industriconnect-mcp-adapters-and-mock-first-evaluation-for-ai-assisted-industrial-operations - IndustriConnect provides AI-powered adapters for industrial protocols, enabling safe and efficient integration of AI assistants into plant operations.
- Saranga: MilliWatt Ultrasound for Navigation in Visually Degraded Environments on Palm-Sized Aerial Robots (viability: 7): https://sciencetostartup.com/paper/saranga-milliwatt-ultrasound-for-navigation-in-visually-degraded-environments-on-palm-sized-aerial-robots - Enabling palm-sized aerial robots to navigate visually degraded environments using low-power ultrasound sensing and deep learning denoising.
- LLaVA-LE: Large Language-and-Vision Assistant for Lunar Exploration (viability: 8): https://sciencetostartup.com/paper/llava-le-large-language-and-vision-assistant-for-lunar-exploration - A specialized vision-language model and dataset for lunar exploration, enabling detailed terrain characterization and analysis.
- Amplified Patch-Level Differential Privacy for Free via Random Cropping (viability: 5): https://sciencetostartup.com/paper/amplified-patch-level-differential-privacy-for-free-via-random-cropping - Enhance differential privacy for computer vision models by leveraging random cropping, improving the privacy-utility trade-off without additional costs.
- Reconstructing Spiking Neural Networks Using a Single Neuron with Autapses (viability: 4): https://sciencetostartup.com/paper/reconstructing-spiking-neural-networks-using-a-single-neuron-with-autapses - A novel framework for reconstructing complex Spiking Neural Networks using a single neuron with autapses, significantly reducing computational costs.
- BCMDA: Bidirectional Correlation Maps Domain Adaptation for Mixed Domain Semi-Supervised Medical Image Segmentation (viability: 7): https://sciencetostartup.com/paper/bcmda-bidirectional-correlation-maps-domain-adaptation-for-mixed-domain-semi-supervised-medical-image-segmentation - A framework for semi-supervised medical image segmentation that bridges domain gaps and corrects pseudo-labels to improve performance with limited annotations.
- UniICL: Systematizing Unified Multimodal In-context Learning through a Capability-Oriented Taxonomy (viability: 7): https://sciencetostartup.com/paper/uniicl-systematizing-unified-multimodal-in-context-learning-through-a-capability-oriented-taxonomy - A framework and dataset to improve the reliability of in-context learning for unified multimodal AI, outperforming larger models on understanding tasks.
- KitchenTwin: Semantically and Geometrically Grounded 3D Kitchen Digital Twins (viability: 7): https://sciencetostartup.com/paper/kitchentwin-semantically-and-geometrically-grounded-3d-kitchen-digital-twins - Builds metrically accurate 3D kitchen digital twins by fusing vision-language models with geometric constraints, enabling precise measurements and object alignment.
- ReDiPrune: Relevance-Diversity Pre-Projection Token Pruning for Efficient Multimodal LLMs (viability: 7): https://sciencetostartup.com/paper/rediprune-relevance-diversity-pre-projection-token-pruning-for-efficient-multimodal-llms - A plug-and-play method to drastically reduce multimodal LLM computation by pruning visual tokens before projection, improving accuracy and efficiency.
- When Is Collective Intelligence a Lottery? Multi-Agent Scaling Laws for Memetic Drift in LLMs (viability: 3): https://sciencetostartup.com/paper/when-is-collective-intelligence-a-lottery-multi-agent-scaling-laws-for-memetic-drift-in-llms - A theoretical framework to understand how multi-agent LLM systems reach consensus through memetic drift, rather than pure reasoning.
- Polynomial Speedup in Diffusion Models with the Multilevel Euler-Maruyama Method (viability: 7): https://sciencetostartup.com/paper/polynomial-speedup-in-diffusion-models-with-the-multilevel-euler-maruyama-method - Accelerate diffusion model sampling by orders of magnitude using a novel multilevel approximation method for SDEs.
- DreamerAD: Efficient Reinforcement Learning via Latent World Model for Autonomous Driving (viability: 7): https://sciencetostartup.com/paper/dreamerad-efficient-reinforcement-learning-via-latent-world-model-for-autonomous-driving - Accelerate autonomous driving reinforcement learning by 80x using a latent world model that compresses diffusion sampling and enables efficient exploration.
- Comparing Developer and LLM Biases in Code Evaluation (viability: 4): https://sciencetostartup.com/paper/comparing-developer-and-llm-biases-in-code-evaluation - A framework to evaluate LLM judges for code applications by identifying systematic biases and misalignment with human preferences.
- TAG: Target-Agnostic Guidance for Stable Object-Centric Inference in Vision-Language-Action Models (viability: 7): https://sciencetostartup.com/paper/tag-target-agnostic-guidance-for-stable-object-centric-inference-in-vision-language-action-models - Enhance the reliability of robotic vision-language-action policies by reducing distractors and improving object instance grounding at inference time.
- The Stochastic Gap: A Markovian Framework for Pre-Deployment Reliability and Oversight-Cost Auditing in Agentic Artificial Intelligence (viability: 5): https://sciencetostartup.com/paper/the-stochastic-gap-a-markovian-framework-for-pre-deployment-reliability-and-oversight-cost-auditing-in-agentic-artificia - A Markovian framework to audit the reliability and oversight costs of agentic AI in enterprise workflows.
- Latent-WAM: Latent World Action Modeling for End-to-End Autonomous Driving (viability: 7): https://sciencetostartup.com/paper/latent-wam-latent-world-action-modeling-for-end-to-end-autonomous-driving - An efficient end-to-end autonomous driving framework that achieves state-of-the-art trajectory planning using spatially-aware and dynamics-informed latent world representations.
- Retrieval Improvements Do Not Guarantee Better Answers: A Study of RAG for AI Policy QA (viability: 4): https://sciencetostartup.com/paper/retrieval-improvements-do-not-guarantee-better-answers-a-study-of-rag-for-ai-policy-qa - This research investigates the limitations of RAG systems in policy analysis, finding that component improvements don't guarantee better answers.
- MARCH: Multi-Agent Reinforced Self-Check for LLM Hallucination (viability: 7): https://sciencetostartup.com/paper/march-multi-agent-reinforced-self-check-for-llm-hallucination - A multi-agent framework that uses deliberate information asymmetry and reinforcement learning to significantly reduce LLM hallucinations in RAG systems.
- Vision-Language Models vs Human: Perceptual Image Quality Assessment (viability: 7): https://sciencetostartup.com/paper/vision-language-models-vs-human-perceptual-image-quality-assessment - Leveraging Vision-Language Models to automate perceptual image quality assessment, outperforming traditional methods in specific attributes and offering a scalable alternative to costly psychophysical experiments.
- EndoVGGT: GNN-Enhanced Depth Estimation for Surgical 3D Reconstruction (viability: 7): https://sciencetostartup.com/paper/endovggt-gnn-enhanced-depth-estimation-for-surgical-3d-reconstruction - A geometry-centric framework using graph neural networks to achieve highly accurate 3D reconstruction of deformable surgical tissues, outperforming state-of-the-art and generalizing to unseen domains.
- Chameleon: Episodic Memory for Long-Horizon Robotic Manipulation (viability: 7): https://sciencetostartup.com/paper/chameleon-episodic-memory-for-long-horizon-robotic-manipulation - Chameleon provides robots with geometry-grounded episodic memory to overcome perceptual aliasing and improve long-horizon manipulation reliability.
- VFIG: Vectorizing Complex Figures in SVG with Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/vfig-vectorizing-complex-figures-in-svg-with-vision-language-models - VFIG converts rasterized images of complex figures into editable SVGs using a novel vision-language model and a large-scale dataset, bridging the gap for designers and researchers.
- Towards Training-Free Scene Text Editing (viability: 7): https://sciencetostartup.com/paper/towards-training-free-scene-text-editing - A training-free framework for high-fidelity scene text editing that achieves state-of-the-art results without requiring task-specific data.
- Completeness of Unbounded Best-First Minimax and Descent Minimax (viability: 3): https://sciencetostartup.com/paper/completeness-of-unbounded-best-first-minimax-and-descent-minimax - This paper theoretically analyzes and experimentally validates improvements to minimax search algorithms for two-player perfect information games, enhancing their ability to find winning strategies.
- Anti-I2V: Safeguarding your photos from malicious image-to-video generation (viability: 7): https://sciencetostartup.com/paper/anti-i2v-safeguarding-your-photos-from-malicious-image-to-video-generation - A novel defense system that safeguards personal photos from malicious image-to-video generation by operating in both Lab and frequency domains.
- POLY-SIM: Polyglot Speaker Identification with Missing Modality Grand Challenge 2026 Evaluation Plan (viability: 4): https://sciencetostartup.com/paper/poly-sim-polyglot-speaker-identification-with-missing-modality-grand-challenge-2026-evaluation-plan - Developing robust multimodal speaker identification systems that perform well even with missing visual data and across different languages.
- Trust Region Constrained Bayesian Optimization with Penalized Constraint Handling (viability: 3): https://sciencetostartup.com/paper/trust-region-constrained-bayesian-optimization-with-penalized-constraint-handling - A theoretical Bayesian optimization method for high-dimensional constrained problems.
- Infrastructure for Valuable, Tradable, and Verifiable Agent Memory (viability: 4): https://sciencetostartup.com/paper/infrastructure-for-valuable-tradable-and-verifiable-agent-memory - We propose a system to make autonomous agent memories tradable assets by binding them to verifiable computational provenance and creating a market layer for their exchange.
- Scaling Recurrence-aware Foundation Models for Clinical Records via Next-Visit Prediction (viability: 7): https://sciencetostartup.com/paper/scaling-recurrence-aware-foundation-models-for-clinical-records-via-next-visit-prediction - A generative pretraining strategy for electronic health records that predicts future patient visits, outperforming existing methods and generalizing to new cohorts.
- The Free-Market Algorithm: Self-Organizing Optimization for Open-Ended Complex Systems (viability: 3): https://sciencetostartup.com/paper/the-free-market-algorithm-self-organizing-optimization-for-open-ended-complex-systems - A novel metaheuristic inspired by free-market economics for open-ended complex system optimization.
- LensWalk: Agentic Video Understanding by Planning How You See in Videos (viability: 7): https://sciencetostartup.com/paper/lenswalk-agentic-video-understanding-by-planning-how-you-see-in-videos - LensWalk is an agentic framework that allows LLMs to actively control their visual observation of videos, improving understanding accuracy by dynamically planning how they see.
- Evaluating Chunking Strategies For Retrieval-Augmented Generation in Oil and Gas Enterprise Documents (viability: 5): https://sciencetostartup.com/paper/evaluating-chunking-strategies-for-retrieval-augmented-generation-in-oil-and-gas-enterprise-documents - Optimizing document chunking for Retrieval-Augmented Generation in specialized enterprise domains like oil and gas to improve information retrieval accuracy and reduce computational costs.
- The role of spatial context and multitask learning in the detection of organic and conventional farming systems based on Sentinel-2 time series (viability: 4): https://sciencetostartup.com/paper/the-role-of-spatial-context-and-multitask-learning-in-the-detection-of-organic-and-conventional-farming-systems-based-on - A Vision Transformer model leverages Sentinel-2 time series and spatial context to discriminate between organic and conventional farming systems, with varying accuracy across crop types.
- A Sociolinguistic Analysis of Automatic Speech Recognition Bias in Newcastle English (viability: 4): https://sciencetostartup.com/paper/a-sociolinguistic-analysis-of-automatic-speech-recognition-bias-in-newcastle-english - This research analyzes bias in Automatic Speech Recognition systems for Newcastle English, revealing socially patterned errors and advocating for dialect-aware development.
- Analysing the Safety Pitfalls of Steering Vectors (viability: 5): https://sciencetostartup.com/paper/analysing-the-safety-pitfalls-of-steering-vectors - This research uncovers how steering LLM behavior can inadvertently increase vulnerability to jailbreak attacks, highlighting a critical trade-off between controllability and safety.
- SEGAR: Selective Enhancement for Generative Augmented Reality (viability: 7): https://sciencetostartup.com/paper/segar-selective-enhancement-for-generative-augmented-reality - SEGAR enables real-time, temporally coherent augmented reality experiences by selectively generating and correcting future visual frames.
- CliPPER: Contextual Video-Language Pretraining on Long-form Intraoperative Surgical Procedures for Event Recognition (viability: 7): https://sciencetostartup.com/paper/clipper-contextual-video-language-pretraining-on-long-form-intraoperative-surgical-procedures-for-event-recognition - A pretraining framework for understanding long-form surgical videos, enabling zero-shot recognition of surgical events and improving multimodal alignment.
- Robust Multilingual Text-to-Pictogram Mapping for Scalable Reading Rehabilitation (viability: 6): https://sciencetostartup.com/paper/robust-multilingual-text-to-pictogram-mapping-for-scalable-reading-rehabilitation - An AI-powered multilingual interface that automatically enhances text with contextually relevant pictograms to support reading comprehension for children with special educational needs.
- Representation Learning to Study Temporal Dynamics in Tutorial Scaffolding (viability: 5): https://sciencetostartup.com/paper/representation-learning-to-study-temporal-dynamics-in-tutorial-scaffolding - A new method to analyze and measure adaptive scaffolding in tutoring dialogues by aligning semantic content, applicable to both human and AI tutors.
- UI-Voyager: A Self-Evolving GUI Agent Learning via Failed Experience (viability: 7): https://sciencetostartup.com/paper/ui-voyager-a-self-evolving-gui-agent-learning-via-failed-experience - A self-evolving GUI agent that learns from failures to achieve high success rates in mobile automation tasks.
- Cross-Modal Prototype Alignment and Mixing for Training-Free Few-Shot Classification (viability: 7): https://sciencetostartup.com/paper/cross-modal-prototype-alignment-and-mixing-for-training-free-few-shot-classification - A novel approach to few-shot image classification by mixing and aligning text and image prototypes, outperforming existing methods on benchmarks.
- From Liar Paradox to Incongruent Sets: A Normal Form for Self-Reference (viability: 1): https://sciencetostartup.com/paper/from-liar-paradox-to-incongruent-sets-a-normal-form-for-self-reference - A formal framework for understanding self-reference in language, offering a structural basis for semantic knowledge.
- No Single Metric Tells the Whole Story: A Multi-Dimensional Evaluation Framework for Uncertainty Attributions (viability: 4): https://sciencetostartup.com/paper/no-single-metric-tells-the-whole-story-a-multi-dimensional-evaluation-framework-for-uncertainty-attributions - A new framework for evaluating AI uncertainty attribution methods to enable more reliable and comparable XAI development.
- TuneShift-KD: Knowledge Distillation and Transfer for Fine-tuned Models (viability: 7): https://sciencetostartup.com/paper/tuneshift-kd-knowledge-distillation-and-transfer-for-fine-tuned-models - A novel method to distill specialized knowledge from fine-tuned LLMs to new models without access to original training data, using perplexity differences to generate synthetic training examples.
- AVO: Agentic Variation Operators for Autonomous Evolutionary Search (viability: 7): https://sciencetostartup.com/paper/avo-agentic-variation-operators-for-autonomous-evolutionary-search - Autonomous agents discover and implement performance-critical AI kernel optimizations, outperforming state-of-the-art expert-engineered solutions.
- Claudini: Autoresearch Discovers State-of-the-Art Adversarial Attack Algorithms for LLMs (viability: 8): https://sciencetostartup.com/paper/claudini-autoresearch-discovers-state-of-the-art-adversarial-attack-algorithms-for-llms - Claudini autonomously discovers advanced adversarial attacks on LLMs, offering cutting-edge cybersecurity solutions.
- Toward Physically Consistent Driving Video World Models under Challenging Trajectories (viability: 8): https://sciencetostartup.com/paper/toward-physically-consistent-driving-video-world-models-under-challenging-trajectories - A world model for autonomous driving that generates physically consistent videos even from challenging or invalid trajectories, improving simulation realism.
- Towards Safe Learning-Based Non-Linear Model Predictive Control through Recurrent Neural Network Modeling (viability: 5): https://sciencetostartup.com/paper/towards-safe-learning-based-non-linear-model-predictive-control-through-recurrent-neural-network-modeling - A novel sequential neural policy for safer and more efficient non-linear model predictive control, reducing computational burden and improving feasibility.
- Project and Generate: Divergence-Free Neural Operators for Incompressible Flows (viability: 7): https://sciencetostartup.com/paper/project-and-generate-divergence-free-neural-operators-for-incompressible-flows - A novel framework for neural operators that enforces physical constraints for stable and accurate fluid dynamics simulations.
- Uniform Laws of Large Numbers in Product Spaces (viability: 1): https://sciencetostartup.com/paper/uniform-laws-of-large-numbers-in-product-spaces - This paper theoretically investigates uniform convergence phenomena in product spaces, extending Vapnik--Chervonenkis theory with a focus on linear VC dimension.
- Video-Only ToM: Enhancing Theory of Mind in Multimodal Large Language Models (viability: 7): https://sciencetostartup.com/paper/video-only-tom-enhancing-theory-of-mind-in-multimodal-large-language-models - A framework to enhance the Theory of Mind capabilities of multimodal LLMs by aligning visual representations with semantic targets, improving reasoning and explanations in video-based scenarios.
- Multi-Agent Reasoning with Consistency Verification Improves Uncertainty Calibration in Medical MCQA (viability: 7): https://sciencetostartup.com/paper/multi-agent-reasoning-with-consistency-verification-improves-uncertainty-calibration-in-medical-mcqa - A multi-agent AI system that significantly improves the reliability of medical diagnosis confidence scores, enabling safer deployment in clinical settings.
- Positive-First Most Ambiguous: A Simple Active Learning Criterion for Interactive Retrieval of Rare Categories (viability: 7): https://sciencetostartup.com/paper/positive-first-most-ambiguous-a-simple-active-learning-criterion-for-interactive-retrieval-of-rare-categories - A novel active learning method for rapidly discovering rare visual categories in large unlabeled datasets with minimal user feedback.
- Composer 2 Technical Report (viability: 7): https://sciencetostartup.com/paper/composer-2-technical-report - A specialized AI model for agentic software engineering that demonstrates strong long-term planning and coding intelligence, with benchmark results comparable to state-of-the-art systems.
- Conformalized Transfer Learning for Li-ion Battery State of Health Forecasting under Manufacturing and Usage Variability (viability: 5): https://sciencetostartup.com/paper/conformalized-transfer-learning-for-li-ion-battery-state-of-health-forecasting-under-manufacturing-and-usage-variability - An uncertainty-aware transfer learning framework for accurate and reliable lithium-ion battery state-of-health forecasting.
- Why Does Self-Distillation (Sometimes) Degrade the Reasoning Capability of LLMs? (viability: 3): https://sciencetostartup.com/paper/why-does-self-distillation-sometimes-degrade-the-reasoning-capability-of-llms - This paper investigates why self-distillation can harm LLM reasoning, particularly in mathematical tasks, by suppressing uncertainty expression.
- Counting Without Numbers \& Finding Without Words (viability: 7): https://sciencetostartup.com/paper/counting-without-numbers-finding-without-words - A multimodal AI system that reunites lost pets with their families using both visual and acoustic biometrics, inspired by animal communication.
- Mechanic: Sorrifier-Driven Formal Decomposition Workflow for Automated Theorem Proving (viability: 7): https://sciencetostartup.com/paper/mechanic-sorrifier-driven-formal-decomposition-workflow-for-automated-theorem-proving - A novel agent system that uses a sorry-driven formal decomposition strategy to efficiently solve complex mathematical reasoning problems in automated theorem proving.
- Design, Modelling and Characterisation of a Miniature Fibre-Reinforced Soft Bending Actuator for Endoluminal Interventions (viability: 4): https://sciencetostartup.com/paper/design-modelling-and-characterisation-of-a-miniature-fibre-reinforced-soft-bending-actuator-for-endoluminal-intervention - Develops a miniature, fiber-reinforced soft bending actuator for endoluminal robotic interventions, demonstrating significant bending angles and structural integrity.
- OmniWeaving: Towards Unified Video Generation with Free-form Composition and Reasoning (viability: 7): https://sciencetostartup.com/paper/omniweaving-towards-unified-video-generation-with-free-form-composition-and-reasoning - OmniWeaving offers a state-of-the-art open-source framework for unified video generation with advanced multimodal composition and reasoning capabilities.
- Unleashing Vision-Language Semantics for Deepfake Video Detection (viability: 7): https://sciencetostartup.com/paper/unleashing-vision-language-semantics-for-deepfake-video-detection - A novel deepfake video detection framework that leverages rich vision-language semantics and identity-aware prompting to significantly outperform state-of-the-art methods.
- Integrating Causal Machine Learning into Clinical Decision Support Systems: Insights from Literature and Practice (viability: 4): https://sciencetostartup.com/paper/integrating-causal-machine-learning-into-clinical-decision-support-systems-insights-from-literature-and-practice - Develops design principles for causal machine learning-powered clinical decision support systems to enhance trust and collaboration.
- CUA-Suite: Massive Human-annotated Video Demonstrations for Computer-Use Agents (viability: 7): https://sciencetostartup.com/paper/cua-suite-massive-human-annotated-video-demonstrations-for-computer-use-agents - A large-scale dataset and benchmark for training and evaluating computer-use agents that automate complex desktop workflows.
- Enes Causal Discovery (viability: 4): https://sciencetostartup.com/paper/enes-causal-discovery - A mixture-of-experts architecture for causal discovery from observational data, aiming to overcome limitations of existing methods.
- The Gait Signature of Frailty: Transfer Learning based Deep Gait Models for Scalable Frailty Assessment (viability: 7): https://sciencetostartup.com/paper/the-gait-signature-of-frailty-transfer-learning-based-deep-gait-models-for-scalable-frailty-assessment - Leveraging transfer learning on a new gait dataset to create a scalable, non-invasive frailty assessment tool for aging medicine.
- What and When to Learn: CURriculum Ranking Loss for Large-Scale Speaker Verification (viability: 4): https://sciencetostartup.com/paper/what-and-when-to-learn-curriculum-ranking-loss-for-large-scale-speaker-verification - A novel curriculum learning loss function for large-scale speaker verification that significantly improves accuracy by adaptively ranking sample difficulty.
- Learning Response-Statistic Shifts and Parametric Roll Episodes from Wave--Vessel Time Series via LSTM Functional Models (viability: 5): https://sciencetostartup.com/paper/learning-response-statistic-shifts-and-parametric-roll-episodes-from-wave-vessel-time-series-via-lstm-functional-models - A data-driven surrogate using LSTMs to predict and analyze critical ship roll instabilities from wave data.
- Marchuk: Efficient Global Weather Forecasting from Mid-Range to Sub-Seasonal Scales via Flow Matching (viability: 7): https://sciencetostartup.com/paper/marchuk-efficient-global-weather-forecasting-from-mid-range-to-sub-seasonal-scales-via-flow-matching - A highly efficient generative model for accurate global weather forecasting up to 30 days, outperforming larger models with faster inference.
- Continuous-Time Learning of Probability Distributions: A Case Study in a Digital Trial of Young Children with Type 1 Diabetes (viability: 4): https://sciencetostartup.com/paper/continuous-time-learning-of-probability-distributions-a-case-study-in-a-digital-trial-of-young-children-with-type-1-diab - A probabilistic framework models continuous-time evolution of glucose distributions for improved diabetes monitoring and treatment analysis.
- IPsec based on Quantum Key Distribution: Adapting non-3GPP access to 5G Networks to the Quantum Era (viability: 7): https://sciencetostartup.com/paper/ipsec-based-on-quantum-key-distribution-adapting-non-3gpp-access-to-5g-networks-to-the-quantum-era - This paper proposes and experimentally validates a quantum-resistant security mechanism for 5G non-3GPP access by integrating Quantum Key Distribution with IPsec, offering faster and more secure connections.
- OneSearch-V2: The Latent Reasoning Enhanced Self-distillation Generative Search Framework (viability: 9): https://sciencetostartup.com/paper/onesearch-v2-the-latent-reasoning-enhanced-self-distillation-generative-search-framework - OneSearch-V2 enhances e-commerce search with reasoning and self-distillation, boosting conversion rates and reducing search biases.
- ClawKeeper: Comprehensive Safety Protection for OpenClaw Agents Through Skills, Plugins, and Watchers (viability: 7): https://sciencetostartup.com/paper/clawkeeper-comprehensive-safety-protection-for-openclaw-agents-through-skills-plugins-and-watchers - ClawKeeper provides comprehensive, real-time security for autonomous agents by integrating skill, plugin, and watcher-based protection mechanisms.
- PINGALA: Prosody-Aware Decoding for Sanskrit Poetry Generation (viability: 5): https://sciencetostartup.com/paper/pingala-prosody-aware-decoding-for-sanskrit-poetry-generation - A decoding approach for Sanskrit poetry generation that improves semantic coherence and metrical adherence by segmenting verses and using phonetically aware transliteration.
- Real Talk, Virtual Faces: A Formal Concept Analysis of Personality and Sentiment in Influencer Audiences (viability: 5): https://sciencetostartup.com/paper/real-talk-virtual-faces-a-formal-concept-analysis-of-personality-and-sentiment-in-influencer-audiences - A framework for analyzing the co-occurrence of sentiment, personality, and topics in online discourse to understand audience reactions to virtual influencers.
- Teacher-Student Diffusion Model for Text-Driven 3D Hand Motion Generation (viability: 4): https://sciencetostartup.com/paper/teacher-student-diffusion-model-for-text-driven-3d-hand-motion-generation - A teacher-student diffusion model for generating realistic 3D hand motions from text, improving VR and robotics applications.
- AI-Supervisor: Autonomous AI Research Supervision via a Persistent Research World Model (viability: 7): https://sciencetostartup.com/paper/ai-supervisor-autonomous-ai-research-supervision-via-a-persistent-research-world-model - An autonomous AI research supervision framework that uses a persistent knowledge graph and multi-agent system to discover, develop, and validate AI research.
- Enhancing Drone Light Shows Performances: Optimal Allocation and Trajectories for Swarm Drone Formations (viability: 7): https://sciencetostartup.com/paper/enhancing-drone-light-shows-performances-optimal-allocation-and-trajectories-for-swarm-drone-formations - A real-time framework for optimally assigning and generating collision-free trajectories for large drone swarms, enabling complex aerial light show choreography.
- Neural Network Models for Contextual Regression (viability: 3): https://sciencetostartup.com/paper/neural-network-models-for-contextual-regression - A novel neural network architecture for contextual regression that improves efficiency and interpretability by separating context identification from context-specific regression.
- Exploring How Fair Model Representations Relate to Fair Recommendations (viability: 5): https://sciencetostartup.com/paper/exploring-how-fair-model-representations-relate-to-fair-recommendations - Develops novel methods to measure and improve fairness in recommender systems by analyzing recommendation parity directly, rather than relying on representation-level proxies.
- 3D-Mix for VLA: A Plug-and-Play Module for Integrating VGGT-based 3D Information into Vision-Language-Action Models (viability: 7): https://sciencetostartup.com/paper/3d-mix-for-vla-a-plug-and-play-module-for-integrating-vggt-based-3d-information-into-vision-language-action-models - A plug-and-play module that significantly enhances the 3D spatial understanding of Vision-Language-Action models for improved robotic control.
- Federated fairness-aware classification under differential privacy (viability: 4): https://sciencetostartup.com/paper/federated-fairness-aware-classification-under-differential-privacy - A novel algorithm for federated classification that simultaneously ensures privacy and fairness, with theoretical guarantees and experimental validation.
- When AI Meets Early Childhood Education: Large Language Models as Assessment Teammates in Chinese Preschools (viability: 8): https://sciencetostartup.com/paper/when-ai-meets-early-childhood-education-large-language-models-as-assessment-teammates-in-chinese-preschools - AI tool automating teacher-child interaction quality assessments in Chinese preschools for scalable, continuous monitoring.
- Causal Transfer in Medical Image Analysis (viability: 5): https://sciencetostartup.com/paper/causal-transfer-in-medical-image-analysis - A survey proposing Causal Transfer Learning to build more robust and generalizable AI models for medical image analysis, addressing domain shift issues.
- On the Use of Bagging for Local Intrinsic Dimensionality Estimation (viability: 3): https://sciencetostartup.com/paper/on-the-use-of-bagging-for-local-intrinsic-dimensionality-estimation - A theoretical framework for improving local intrinsic dimensionality estimation using bagging to reduce variance and mean squared error.
- ViHOI: Human-Object Interaction Synthesis with Visual Priors (viability: 7): https://sciencetostartup.com/paper/vihoi-human-object-interaction-synthesis-with-visual-priors - Generate realistic 3D human-object interactions by extracting visual priors from 2D images using a diffusion-based framework.
- MolEvolve: LLM-Guided Evolutionary Search for Interpretable Molecular Optimization (viability: 7): https://sciencetostartup.com/paper/molevolve-llm-guided-evolutionary-search-for-interpretable-molecular-optimization - An LLM-guided evolutionary search framework that autonomously discovers interpretable molecular optimizations by planning chemical symbolic operations.
- GeoRouter: Dynamic Paradigm Routing for Worldwide Image Geolocalization (viability: 7): https://sciencetostartup.com/paper/georouter-dynamic-paradigm-routing-for-worldwide-image-geolocalization - A dynamic routing framework for image geolocalization that adaptively selects the best paradigm (retrieval or generation) for precise GPS coordinate prediction.
- Towards Reward Modeling for AI Tutors in Math Mistake Remediation (viability: 5): https://sciencetostartup.com/paper/towards-reward-modeling-for-ai-tutors-in-math-mistake-remediation - Develops a specialized reward model for AI math tutors to improve pedagogical quality and mistake remediation.
- PP-OCRv5: A Specialized 5M-Parameter Model Rivaling Billion-Parameter Vision-Language Models on OCR Tasks (viability: 7): https://sciencetostartup.com/paper/pp-ocrv5-a-specialized-5m-parameter-model-rivaling-billion-parameter-vision-language-models-on-ocr-tasks - A highly efficient 5M-parameter OCR system that rivals billion-parameter models by focusing on data quality and diversity, offering superior localization and reduced hallucinations.
- Improving Lean4 Autoformalization via Cycle Consistency Fine-tuning (viability: 4): https://sciencetostartup.com/paper/improving-lean4-autoformalization-via-cycle-consistency-fine-tuning - This research explores a novel reinforcement learning approach for translating natural language mathematics into formal proof language, demonstrating improved performance over supervised methods.
- CoordLight: Learning Decentralized Coordination for Network-Wide Traffic Signal Control (viability: 7): https://sciencetostartup.com/paper/coordlight-learning-decentralized-coordination-for-network-wide-traffic-signal-control - A decentralized AI framework for optimizing city-wide traffic signals using novel state representation and neighbor-aware policy optimization.
- LATS: Large Language Model Assisted Teacher-Student Framework for Multi-Agent Reinforcement Learning in Traffic Signal Control (viability: 7): https://sciencetostartup.com/paper/lats-large-language-model-assisted-teacher-student-framework-for-multi-agent-reinforcement-learning-in-traffic-signal-co - A novel framework that uses LLMs to distill knowledge into simpler RL agents for more efficient and generalizable traffic signal control.
- A Neuro-Symbolic System for Interpretable Multimodal Physiological Signals Integration in Human Fatigue Detection (viability: 4): https://sciencetostartup.com/paper/a-neuro-symbolic-system-for-interpretable-multimodal-physiological-signals-integration-in-human-fatigue-detection - A neuro-symbolic system integrates physiological signals for interpretable human fatigue detection, offering insights into model reasoning.
- A Sensorless, Inherently Compliant Anthropomorphic Musculoskeletal Hand Driven by Electrohydraulic Actuators (viability: 5): https://sciencetostartup.com/paper/a-sensorless-inherently-compliant-anthropomorphic-musculoskeletal-hand-driven-by-electrohydraulic-actuators - A novel musculoskeletal robotic hand design that uses electrohydraulic actuators for inherent compliance and self-sensing, eliminating the need for external sensors for safe manipulation.
- Language-Guided Structure-Aware Network for Camouflaged Object Detection (viability: 7): https://sciencetostartup.com/paper/language-guided-structure-aware-network-for-camouflaged-object-detection - A language-guided network that uses text prompts to improve the detection of camouflaged objects in images.
- Evidence of an Emergent "Self" in Continual Robot Learning (viability: 3): https://sciencetostartup.com/paper/evidence-of-an-emergent-self-in-continual-robot-learning - This research explores a theoretical framework for identifying emergent 'self' concepts in continually learning robots by analyzing invariant subnetworks.
- Enhancing Efficiency and Performance in Deepfake Audio Detection through Neuron-level dropin & Neuroplasticity Mechanisms (viability: 7): https://sciencetostartup.com/paper/enhancing-efficiency-and-performance-in-deepfake-audio-detection-through-neuron-level-dropin-neuroplasticity-mechanisms - A novel approach inspired by brain plasticity to significantly improve the efficiency and accuracy of deepfake audio detection models.
- GameplayQA: A Benchmarking Framework for Decision-Dense POV-Synced Multi-Video Understanding of 3D Virtual Agents (viability: 7): https://sciencetostartup.com/paper/gameplayqa-a-benchmarking-framework-for-decision-dense-pov-synced-multi-video-understanding-of-3d-virtual-agents - A new benchmark and dataset for evaluating multimodal LLMs in 3D virtual agents, revealing significant gaps in current AI capabilities for embodied perception and reasoning.
- Le MuMo JEPA: Multi-Modal Self-Supervised Representation Learning with Learnable Fusion Tokens (viability: 7): https://sciencetostartup.com/paper/le-mumo-jepa-multi-modal-self-supervised-representation-learning-with-learnable-fusion-tokens - A multi-modal self-supervised learning framework that unifies RGB and LiDAR depth representations for improved performance on downstream driving tasks.
- Boosting Document Parsing Efficiency and Performance with Coarse-to-Fine Visual Processing (viability: 8): https://sciencetostartup.com/paper/boosting-document-parsing-efficiency-and-performance-with-coarse-to-fine-visual-processing - PaddleOCR-VL enhances document parsing efficiency by focusing on semantically relevant regions with a coarse-to-fine processing framework.
- Large Language Model Guided Incentive Aware Reward Design for Cooperative Multi-Agent Reinforcement Learning (viability: 5): https://sciencetostartup.com/paper/large-language-model-guided-incentive-aware-reward-design-for-cooperative-multi-agent-reinforcement-learning - Automated reward design for cooperative multi-agent systems using LLMs to improve coordination and task performance.
- Heuristic Self-Paced Learning for Domain Adaptive Semantic Segmentation under Adverse Conditions (viability: 7): https://sciencetostartup.com/paper/heuristic-self-paced-learning-for-domain-adaptive-semantic-segmentation-under-adverse-conditions - An AI-powered scheduler that autonomously optimizes the learning order of semantic classes for improved image segmentation under challenging conditions.
- Toward Generalist Neural Motion Planners for Robotic Manipulators: Challenges and Opportunities (viability: 4): https://sciencetostartup.com/paper/toward-generalist-neural-motion-planners-for-robotic-manipulators-challenges-and-opportunities - Developing generalist neural motion planners to overcome limitations in cluttered robotic manipulation environments.
- Refining time-space traffic diagrams: A neighborhood-adaptive linear regression method (viability: 7): https://sciencetostartup.com/paper/refining-time-space-traffic-diagrams-a-neighborhood-adaptive-linear-regression-method - A neighborhood-adaptive linear regression method to refine low-resolution time-space traffic diagrams, improving accuracy and capturing complex traffic dynamics.
- Samasāmayik: A Parallel Dataset for Hindi-Sanskrit Machine Translation (viability: 7): https://sciencetostartup.com/paper/samas-mayik-a-parallel-dataset-for-hindi-sanskrit-machine-translation - A new parallel dataset for Hindi-Sanskrit translation, enabling significant performance gains for contemporary language models.
- CGRL: Causal-Guided Representation Learning for Graph Out-of-Distribution Generalization (viability: 4): https://sciencetostartup.com/paper/cgrl-causal-guided-representation-learning-for-graph-out-of-distribution-generalization - A novel approach to improve Graph Neural Network generalization on out-of-distribution data by integrating causal representation learning and loss replacement.
- A Large-Scale Study of Telegram Bots (viability: 4): https://sciencetostartup.com/paper/a-large-scale-study-of-telegram-bots - A large-scale study and dataset of Telegram bots to identify and combat illicit activities, providing insights for content moderators and researchers.
- AMIF: Authorizable Medical Image Fusion Model with Built-in Authentication (viability: 4): https://sciencetostartup.com/paper/amif-authorizable-medical-image-fusion-model-with-built-in-authentication - A medical image fusion model that embeds copyright identifiers and requires authentication for high-quality outputs, protecting intellectual property.
- RS-SSM: Refining Forgotten Specifics in State Space Model for Video Semantic Segmentation (viability: 7): https://sciencetostartup.com/paper/rs-ssm-refining-forgotten-specifics-in-state-space-model-for-video-semantic-segmentation - A novel state space model that refines forgotten specifics for state-of-the-art video semantic segmentation with high computational efficiency.
- VERIA: Verification-Centric Multimodal Instance Augmentation for Long-Tailed 3D Object Detection (viability: 7): https://sciencetostartup.com/paper/veria-verification-centric-multimodal-instance-augmentation-for-long-tailed-3d-object-detection - VERIA enhances 3D object detection for rare classes in autonomous driving by synthesizing diverse and contextually relevant multimodal instances using off-the-shelf foundation models.
- Cost-Sensitive Neighborhood Aggregation for Heterophilous Graphs: When Does Per-Edge Routing Help? (viability: 4): https://sciencetostartup.com/paper/cost-sensitive-neighborhood-aggregation-for-heterophilous-graphs-when-does-per-edge-routing-help - A GNN layer that intelligently routes messages based on edge type to improve performance on specific types of heterophilous graphs.
- The Specification Gap: Coordination Failure Under Partial Knowledge in Code Agents (viability: 3): https://sciencetostartup.com/paper/the-specification-gap-coordination-failure-under-partial-knowledge-in-code-agents - This research explores the coordination challenges faced by multiple LLM-based code agents when implementing shared code components, highlighting the critical role of detailed specifications in achieving integration accuracy.
- Bridging Biological Hearing and Neuromorphic Computing: End-to-End Time-Domain Audio Signal Processing with Reservoir Computing (viability: 5): https://sciencetostartup.com/paper/bridging-biological-hearing-and-neuromorphic-computing-end-to-end-time-domain-audio-signal-processing-with-reservoir-com - A novel reservoir computing approach simplifies audio signal processing for real-time, energy-efficient speech analysis in embedded systems.
- Software Supply Chain Smells: Lightweight Analysis for Secure Dependency Management (viability: 4): https://sciencetostartup.com/paper/software-supply-chain-smells-lightweight-analysis-for-secure-dependency-management - A tool to detect structural indicators of security risks in software dependencies.
- TopoMesh: High-Fidelity Mesh Autoencoding via Topological Unification (viability: 7): https://sciencetostartup.com/paper/topomesh-high-fidelity-mesh-autoencoding-via-topological-unification - TopoMesh enables high-fidelity 3D mesh generation by unifying topological representations between ground truth and predicted meshes, leading to superior preservation of sharp features and geometric details.
- Language-Assisted Image Clustering Guided by Discriminative Relational Signals and Adaptive Semantic Centers (viability: 7): https://sciencetostartup.com/paper/language-assisted-image-clustering-guided-by-discriminative-relational-signals-and-adaptive-semantic-centers - A novel framework for image clustering that leverages language to create more discriminative features and adaptive semantic centers, outperforming state-of-the-art methods.
- ScrollScape: Unlocking 32K Image Generation With Video Diffusion Priors (viability: 7): https://sciencetostartup.com/paper/scrollscape-unlocking-32k-image-generation-with-video-diffusion-priors - ScrollScape enables ultra-high-resolution (32K) image generation at extreme aspect ratios by leveraging video diffusion priors for structural integrity.
- DeepDTF: Dual-Branch Transformer Fusion for Multi-Omics Anticancer Drug Response Prediction (viability: 7): https://sciencetostartup.com/paper/deepdtf-dual-branch-transformer-fusion-for-multi-omics-anticancer-drug-response-prediction - A dual-branch Transformer fusion framework predicts anticancer drug response from multi-omics data, outperforming baselines and providing biological explanations.
- Forecasting with Guidance: Representation-Level Supervision for Time Series Forecasting (viability: 7): https://sciencetostartup.com/paper/forecasting-with-guidance-representation-level-supervision-for-time-series-forecasting - A plug-in method that enhances time series forecasting accuracy by aligning intermediate representations with pretrained foundation models.
- Accelerating Diffusion-based Video Editing via Heterogeneous Caching: Beyond Full Computing at Sampled Denoising Timestep (viability: 5): https://sciencetostartup.com/paper/accelerating-diffusion-based-video-editing-via-heterogeneous-caching-beyond-full-computing-at-sampled-denoising-timestep - A training-free framework to accelerate diffusion-based video editing by intelligently caching context tokens, reducing redundant computations without sacrificing quality.
- Semantic Alignment across Ancient Egyptian Language Stages via Normalization-Aware Multitask Learning (viability: 4): https://sciencetostartup.com/paper/semantic-alignment-across-ancient-egyptian-language-stages-via-normalization-aware-multitask-learning - A novel multitask learning approach for semantic alignment across historical language stages, providing a baseline and guidance for modeling under data constraints.
- Memory-Augmented Vision-Language Agents for Persistent and Semantically Consistent Object Captioning (viability: 9): https://sciencetostartup.com/paper/memory-augmented-vision-language-agents-for-persistent-and-semantically-consistent-object-captioning - A memory-augmented vision-language model ensuring consistent multi-view object captioning for better embodied agent navigation.
- Embracing Heteroscedasticity for Probabilistic Time Series Forecasting (viability: 5): https://sciencetostartup.com/paper/embracing-heteroscedasticity-for-probabilistic-time-series-forecasting - A probabilistic time series forecasting framework that explicitly models time-varying variance for more robust predictions and uncertainty quantification.
- Semantic Centroids and Hierarchical Density-Based Clustering for Cross-Document Software Coreference Resolution (viability: 4): https://sciencetostartup.com/paper/semantic-centroids-and-hierarchical-density-based-clustering-for-cross-document-software-coreference-resolution - A system for resolving inconsistent software mentions across scientific documents using semantic embeddings and density-based clustering.
- B-MoE: A Body-Part-Aware Mixture-of-Experts "All Parts Matter" Approach to Micro-Action Recognition (viability: 7): https://sciencetostartup.com/paper/b-moe-a-body-part-aware-mixture-of-experts-all-parts-matter-approach-to-micro-action-recognition - A novel body-part-aware Mixture-of-Experts framework for highly accurate micro-action recognition, outperforming state-of-the-art on challenging benchmarks.
- Optimizing Multilingual LLMs via Federated Learning: A Study of Client Language Composition (viability: 3): https://sciencetostartup.com/paper/optimizing-multilingual-llms-via-federated-learning-a-study-of-client-language-composition - This paper explores optimizing multilingual LLMs using federated learning by analyzing the impact of client language composition on model performance and fairness.
- InstanceRSR: Real-World Super-Resolution via Instance-Aware Representation Alignment (viability: 7): https://sciencetostartup.com/paper/instancersr-real-world-super-resolution-via-instance-aware-representation-alignment - A super-resolution framework that preserves fine-grained details and semantic consistency at the instance level for real-world images.
- DVM: Real-Time Kernel Generation for Dynamic AI Models (viability: 4): https://sciencetostartup.com/paper/dvm-real-time-kernel-generation-for-dynamic-ai-models - A real-time compiler for dynamic AI models that significantly speeds up compilation and improves model efficiency.
- Decentralized End-to-End Multi-AAV Pursuit Using Predictive Spatio-Temporal Observation via Deep Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/decentralized-end-to-end-multi-aav-pursuit-using-predictive-spatio-temporal-observation-via-deep-reinforcement-learning - Develop an advanced aerial swarm pursuit system using deep reinforcement learning for autonomous navigation in cluttered environments.
- Attack Assessment and Augmented Identity Recognition for Human Skeleton Data (viability: 7): https://sciencetostartup.com/paper/attack-assessment-and-augmented-identity-recognition-for-human-skeleton-data - A novel framework to defend machine learning models against adversarial attacks using GAN-generated synthetic data, improving robustness without sacrificing performance.
- Stance Labels Fail When They Matter Most: The Projection Problem in Stance Detection (viability: 3): https://sciencetostartup.com/paper/stance-labels-fail-when-they-matter-most-the-projection-problem-in-stance-detection - This research identifies a fundamental flaw in how stance detection is currently performed, showing that existing methods fail when attitudes are complex and multi-dimensional.
- Identification of NMF by choosing maximum-volume basis vectors (viability: 3): https://sciencetostartup.com/paper/identification-of-nmf-by-choosing-maximum-volume-basis-vectors - A new NMF framework that makes basis vectors as distinct as possible to improve interpretability and handle highly mixed data.
- UniScale: Synergistic Entire Space Data and Model Scaling for Search Ranking (viability: 7): https://sciencetostartup.com/paper/uniscale-synergistic-entire-space-data-and-model-scaling-for-search-ranking - UniScale co-designs data and model architecture to unlock superior performance in search ranking by synergistically scaling both elements.
- RVLM: Recursive Vision-Language Models with Adaptive Depth (viability: 7): https://sciencetostartup.com/paper/rvlm-recursive-vision-language-models-with-adaptive-depth - An adaptive, auditable vision-language model for medical diagnostics that generates and executes code for iterative reasoning.
- Variation is the Norm: Embracing Sociolinguistics in NLP (viability: 5): https://sciencetostartup.com/paper/variation-is-the-norm-embracing-sociolinguistics-in-nlp - A framework to integrate sociolinguistic variation into NLP models, improving robustness and performance on evolving languages.
- Environment-Grounded Multi-Agent Workflow for Autonomous Penetration Testing (viability: 7): https://sciencetostartup.com/paper/environment-grounded-multi-agent-workflow-for-autonomous-penetration-testing - An AI-powered multi-agent system for automated penetration testing of robotic systems, offering high reliability and traceability.
- Who Benefits from RAG? The Role of Exposure, Utility and Attribution Bias (viability: 7): https://sciencetostartup.com/paper/who-benefits-from-rag-the-role-of-exposure-utility-and-attribution-bias - This research quantifies and addresses fairness disparities in Retrieval-Augmented Generation (RAG) systems, offering a path to more equitable LLM applications.
- Where Do Your Citations Come From? Citation-Constellation: A Free, Open-Source, No-Code, and Auditable Tool for Citation Network Decomposition with Complementary BARON and HEROCON Scores (viability: 4): https://sciencetostartup.com/paper/where-do-your-citations-come-from-citation-constellation-a-free-open-source-no-code-and-auditable-tool-for-citation-netw - A no-code tool for analyzing citation networks to understand the origin of scholarly influence.
- Uncovering Memorization in Timeseries Imputation models: LBRM Membership Inference and its link to attribute Leakage (viability: 7): https://sciencetostartup.com/paper/uncovering-memorization-in-timeseries-imputation-models-lbrm-membership-inference-and-its-link-to-attribute-leakage - Develops novel attacks to uncover privacy vulnerabilities in time series imputation models, enabling more secure AI deployments.
- HEART-PFL: Stable Personalized Federated Learning under Heterogeneity with Hierarchical Directional Alignment and Adversarial Knowledge Transfer (viability: 7): https://sciencetostartup.com/paper/heart-pfl-stable-personalized-federated-learning-under-heterogeneity-with-hierarchical-directional-alignment-and-adversa - A dual-sided framework for stable personalized federated learning that enhances client specificity and global model accuracy through hierarchical alignment and adversarial knowledge transfer.
- Powerful Teachers Matter: Text-Guided Multi-view Knowledge Distillation with Visual Prior Enhancement (viability: 7): https://sciencetostartup.com/paper/powerful-teachers-matter-text-guided-multi-view-knowledge-distillation-with-visual-prior-enhancement - Enhance knowledge distillation by using dual-modality teachers with visual prior enhancement and text-guided adaptive fusion to significantly improve student model performance.
- IPatch: A Multi-Resolution Transformer Architecture for Robust Time-Series Forecasting (viability: 7): https://sciencetostartup.com/paper/ipatch-a-multi-resolution-transformer-architecture-for-robust-time-series-forecasting - A multi-resolution Transformer architecture that improves time-series forecasting accuracy and robustness by integrating both point-wise and patch-wise temporal representations.
- Invisible Threats from Model Context Protocol: Generating Stealthy Injection Payload via Tree-based Adaptive Search (viability: 4): https://sciencetostartup.com/paper/invisible-threats-from-model-context-protocol-generating-stealthy-injection-payload-via-tree-based-adaptive-search - A novel black-box attack method that generates stealthy injection payloads to compromise LLM agents by treating payload generation as a tree-structured search problem.
- A Deep Dive into Scaling RL for Code Generation with Synthetic Data and Curricula (viability: 4): https://sciencetostartup.com/paper/a-deep-dive-into-scaling-rl-for-code-generation-with-synthetic-data-and-curricula - A novel synthetic data generation pipeline for Reinforcement Learning to improve large language models for code and math tasks.
- RefReward-SR: LR-Conditioned Reward Modeling for Preference-Aligned Super-Resolution (viability: 7): https://sciencetostartup.com/paper/refreward-sr-lr-conditioned-reward-modeling-for-preference-aligned-super-resolution - A novel reward modeling framework for super-resolution that aligns with human perceptual preferences by using the low-resolution input as a semantic anchor, supported by a new large-scale dataset and code release.
- TsetlinWiSARD: On-Chip Training of Weightless Neural Networks using Tsetlin Automata on FPGAs (viability: 7): https://sciencetostartup.com/paper/tsetlinwisard-on-chip-training-of-weightless-neural-networks-using-tsetlin-automata-on-fpgas - Enable on-chip training of weightless neural networks for efficient edge AI with TsetlinWiSARD, offering significant hardware efficiency improvements.
- Unlocking Few-Shot Capabilities in LVLMs via Prompt Conditioning and Head Selection (viability: 7): https://sciencetostartup.com/paper/unlocking-few-shot-capabilities-in-lvlms-via-prompt-conditioning-and-head-selection - Unlock state-of-the-art few-shot and zero-shot image classification for Large Vision Language Models by intelligently combining their internal representations.
- Towards Remote Attestation of Microarchitectural Attacks: The Case of Rowhammer (viability: 5): https://sciencetostartup.com/paper/towards-remote-attestation-of-microarchitectural-attacks-the-case-of-rowhammer - A remote attestation protocol to detect Rowhammer attacks using commodity hardware features and TPMs.
- Walma: Learning to See Memory Corruption in WebAssembly (viability: 7): https://sciencetostartup.com/paper/walma-learning-to-see-memory-corruption-in-webassembly - Walma uses machine learning to detect memory corruption and tampering in WebAssembly applications, offering a practical solution for runtime integrity verification.
- Heuristic-inspired Reasoning Priors Facilitate Data-Efficient Referring Object Detection (viability: 5): https://sciencetostartup.com/paper/heuristic-inspired-reasoning-priors-facilitate-data-efficient-referring-object-detection - A framework that injects heuristic reasoning priors to improve data efficiency in referring object detection for low-data environments.
- CarePilot: A Multi-Agent Framework for Long-Horizon Computer Task Automation in Healthcare (viability: 7): https://sciencetostartup.com/paper/carepilot-a-multi-agent-framework-for-long-horizon-computer-task-automation-in-healthcare - A multi-agent framework for automating complex, long-horizon computer tasks in healthcare, outperforming existing models.
- A convergent Plug-and-Play Majorization-Minimization algorithm for Poisson inverse problems (viability: 5): https://sciencetostartup.com/paper/a-convergent-plug-and-play-majorization-minimization-algorithm-for-poisson-inverse-problems - A plug-and-play algorithm for Poisson inverse problems that leverages pre-trained denoisers for improved performance in medical imaging applications.
- Goal-Oriented Reactive Simulation for Closed-Loop Trajectory Prediction (viability: 7): https://sciencetostartup.com/paper/goal-oriented-reactive-simulation-for-closed-loop-trajectory-prediction - Develops a closed-loop simulation for trajectory prediction that trains autonomous agents to react and recover from their own errors, significantly improving collision avoidance.
- A visual observation on the geometry of UMAP projections of the difference vectors of antonym and synonym word pair embeddings (viability: 3): https://sciencetostartup.com/paper/a-visual-observation-on-the-geometry-of-umap-projections-of-the-difference-vectors-of-antonym-and-synonym-word-pair-embe - Analyzing the geometric properties of word embeddings to understand antonym relationships.
- LightSplat: Fast and Memory-Efficient Open-Vocabulary 3D Scene Understanding in Five Seconds (viability: 8): https://sciencetostartup.com/paper/lightsplat-fast-and-memory-efficient-open-vocabulary-3d-scene-understanding-in-five-seconds - LightSplat dramatically speeds up and optimizes 3D scene understanding with a lightweight indexing framework, making real-time open-vocabulary segmentation feasible.
- Linear-Nonlinear Fusion Neural Operator for Partial Differential Equations (viability: 7): https://sciencetostartup.com/paper/linear-nonlinear-fusion-neural-operator-for-partial-differential-equations - A novel neural operator that significantly accelerates PDE solving by decoupling linear and nonlinear effects, offering faster training and comparable or better accuracy.
- Tutor-Student Reinforcement Learning: A Dynamic Curriculum for Robust Deepfake Detection (viability: 7): https://sciencetostartup.com/paper/tutor-student-reinforcement-learning-a-dynamic-curriculum-for-robust-deepfake-detection - A reinforcement learning framework that dynamically optimizes deepfake detection training by prioritizing challenging samples for improved generalization.
- Efficient Controller Learning from Human Preferences and Numerical Data Via Multi-Modal Surrogate Models (viability: 5): https://sciencetostartup.com/paper/efficient-controller-learning-from-human-preferences-and-numerical-data-via-multi-modal-surrogate-models - A framework for efficiently tuning control policies by combining numerical data with human preferences using multi-modal surrogate models.
- Spectral Scalpel: Amplifying Adjacent Action Discrepancy via Frequency-Selective Filtering for Skeleton-Based Action Segmentation (viability: 7): https://sciencetostartup.com/paper/spectral-scalpel-amplifying-adjacent-action-discrepancy-via-frequency-selective-filtering-for-skeleton-based-action-segm - A frequency-selective filtering framework that amplifies action-specific frequencies to improve skeleton-based action segmentation accuracy and boundary definition.
- Accelerated Spline-Based Time-Optimal Motion Planning with Continuous Safety Guarantees for Non-Differentially Flat Systems (viability: 7): https://sciencetostartup.com/paper/accelerated-spline-based-time-optimal-motion-planning-with-continuous-safety-guarantees-for-non-differentially-flat-syst - A novel motion planning method that significantly reduces trajectory computation time for autonomous robots by decoupling safety constraints, enabling faster and safer navigation in complex environments.
- MedAidDialog: A Multilingual Multi-Turn Medical Dialogue Dataset for Accessible Healthcare (viability: 7): https://sciencetostartup.com/paper/medaiddialog-a-multilingual-multi-turn-medical-dialogue-dataset-for-accessible-healthcare - A multilingual, multi-turn medical dialogue dataset and model for accessible preliminary healthcare consultations, leveraging parameter-efficient fine-tuning for broad deployment.
- Reservoir-Based Graph Convolutional Networks (viability: 7): https://sciencetostartup.com/paper/reservoir-based-graph-convolutional-networks - A novel graph convolutional network integrating reservoir computing for improved graph classification and generation, with faster convergence and reduced over-smoothing.
- Equivariant Filter Transformations for Consistent and Efficient Visual--Inertial Navigation (viability: 3): https://sciencetostartup.com/paper/equivariant-filter-transformations-for-consistent-and-efficient-visual-inertial-navigation - A theoretical framework for improving visual-inertial navigation through equivariant filter transformations.
- On Gossip Algorithms for Machine Learning with Pairwise Objectives (viability: 2): https://sciencetostartup.com/paper/on-gossip-algorithms-for-machine-learning-with-pairwise-objectives - Develops theoretical convergence guarantees for gossip algorithms applied to pairwise machine learning objectives in distributed sensor networks.
- Likelihood hacking in probabilistic program synthesis (viability: 5): https://sciencetostartup.com/paper/likelihood-hacking-in-probabilistic-program-synthesis - A language-level safety constraint for probabilistic programming languages that prevents models from generating invalid programs, demonstrated with a modification to Stan.
- Alignment Reduces Expressed but Not Encoded Gender Bias: A Unified Framework and Study (viability: 7): https://sciencetostartup.com/paper/alignment-reduces-expressed-but-not-encoded-gender-bias-a-unified-framework-and-study - A unified framework to analyze and mitigate gender bias in LLMs by correlating internal representations with expressed outputs, demonstrating that current debiasing methods don't fully remove latent bias.
- The Alignment Tax: Response Homogenization in Aligned LLMs and Its Implications for Uncertainty Estimation (viability: 7): https://sciencetostartup.com/paper/the-alignment-tax-response-homogenization-in-aligned-llms-and-its-implications-for-uncertainty-estimation - This research identifies and quantifies an 'alignment tax' causing response homogenization in LLMs, proposing a novel uncertainty estimation method to improve selective prediction accuracy and reduce computational costs.
- Combi-CAM: A Novel Multi-Layer Approach for Explainable Image Geolocalization (viability: 3): https://sciencetostartup.com/paper/combi-cam-a-novel-multi-layer-approach-for-explainable-image-geolocalization - Enhancing the explainability of image geolocalization models by combining activation maps from multiple network layers.
- Retinal Layer Segmentation in OCT Images With 2.5D Cross-slice Feature Fusion Module for Glaucoma Assessment (viability: 7): https://sciencetostartup.com/paper/retinal-layer-segmentation-in-oct-images-with-2-5d-cross-slice-feature-fusion-module-for-glaucoma-assessment - A 2.5D framework for accurate retinal layer segmentation in OCT images, improving glaucoma assessment by fusing cross-slice features for better consistency and robustness.
- Mixed-signal implementation of feedback-control optimizer for single-layer Spiking Neural Networks (viability: 4): https://sciencetostartup.com/paper/mixed-signal-implementation-of-feedback-control-optimizer-for-single-layer-spiking-neural-networks - Enabling on-chip learning for neuromorphic systems with a feedback-control optimizer.
- Toward a Multi-Layer ML-Based Security Framework for Industrial IoT (viability: 3): https://sciencetostartup.com/paper/toward-a-multi-layer-ml-based-security-framework-for-industrial-iot - A lightweight, ML-based security framework for Industrial IoT that predicts and mitigates network condition impacts on trust convergence.
- Granular Ball Guided Stable Latent Domain Discovery for Domain-General Crowd Counting (viability: 5): https://sciencetostartup.com/paper/granular-ball-guided-stable-latent-domain-discovery-for-domain-general-crowd-counting - A novel framework for more stable and accurate crowd counting across different environments by discovering and leveraging latent domains.
- Causality-Driven Disentangled Representation Learning in Multiplex Graphs (viability: 5): https://sciencetostartup.com/paper/causality-driven-disentangled-representation-learning-in-multiplex-graphs - A causal inference framework to disentangle shared and private information in multiplex graphs for improved representation learning.
- KCLNet: Electrically Equivalence-Oriented Graph Representation Learning for Analog Circuits (viability: 7): https://sciencetostartup.com/paper/kclnet-electrically-equivalence-oriented-graph-representation-learning-for-analog-circuits - KCLNet provides an electrically-aware graph representation learning framework for analog circuits, enabling significant performance improvements in critical EDA tasks like classification and subcircuit detection.
- LaDy: Lagrangian-Dynamic Informed Network for Skeleton-based Action Segmentation via Spatial-Temporal Modulation (viability: 7): https://sciencetostartup.com/paper/lady-lagrangian-dynamic-informed-network-for-skeleton-based-action-segmentation-via-spatial-temporal-modulation - A physics-informed neural network that leverages Lagrangian dynamics for highly accurate and discriminative skeleton-based action segmentation.
- Towards Effective Experiential Learning: Dual Guidance for Utilization and Internalization (viability: 3): https://sciencetostartup.com/paper/towards-effective-experiential-learning-dual-guidance-for-utilization-and-internalization - A new framework for LLM training that aims to improve reasoning by better utilizing and internalizing experience during reinforcement learning.
- LGTM: Training-Free Light-Guided Text-to-Image Diffusion Model via Initial Noise Manipulation (viability: 7): https://sciencetostartup.com/paper/lgtm-training-free-light-guided-text-to-image-diffusion-model-via-initial-noise-manipulation - A training-free method to control lighting in text-to-image generation by manipulating initial noise, offering dynamic user guidance.
- Bridging the Evaluation Gap: Standardized Benchmarks for Multi-Objective Search (viability: 4): https://sciencetostartup.com/paper/bridging-the-evaluation-gap-standardized-benchmarks-for-multi-objective-search - A standardized benchmark suite for multi-objective search to enable robust and reproducible cross-study comparisons.
- Knowledge-Guided Manipulation Using Multi-Task Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/knowledge-guided-manipulation-using-multi-task-reinforcement-learning - A multi-task reinforcement learning framework for robotic manipulation that leverages real-time 3D scene graphs and knowledge to achieve robust generalization and sample efficiency.
- LLMpedia: A Transparent Framework to Materialize an LLM's Encyclopedic Knowledge at Scale (viability: 7): https://sciencetostartup.com/paper/llmpedia-a-transparent-framework-to-materialize-an-llm-s-encyclopedic-knowledge-at-scale - LLMpedia empowers enterprises with auditable AI-generated encyclopedic content across diverse topics, enhancing knowledge bases.
- When Understanding Becomes a Risk: Authenticity and Safety Risks in the Emerging Image Generation Paradigm (viability: 7): https://sciencetostartup.com/paper/when-understanding-becomes-a-risk-authenticity-and-safety-risks-in-the-emerging-image-generation-paradigm - This research identifies and quantifies new safety risks in multimodal large language models for image generation, showing they are more prone to generating unsafe content and harder to detect than diffusion models.
- PosterIQ: A Design Perspective Benchmark for Poster Understanding and Generation (viability: 7): https://sciencetostartup.com/paper/posteriq-a-design-perspective-benchmark-for-poster-understanding-and-generation - A benchmark and diagnostic tool for building generative AI that understands and creates visually compelling posters with human-centered design principles.
- The impact of sensor placement on graph-neural-network-based leakage detection (viability: 4): https://sciencetostartup.com/paper/the-impact-of-sensor-placement-on-graph-neural-network-based-leakage-detection - Optimize sensor placement for graph neural networks to significantly improve leakage detection in water distribution networks.
- ConceptKT: A Benchmark for Concept-Level Deficiency Prediction in Knowledge Tracing (viability: 5): https://sciencetostartup.com/paper/conceptkt-a-benchmark-for-concept-level-deficiency-prediction-in-knowledge-tracing - A new benchmark and method for diagnosing specific conceptual misunderstandings in students, going beyond simple correctness prediction in educational AI.
- Enhanced Mycelium of Thought (EMoT): A Bio-Inspired Hierarchical Reasoning Architecture with Strategic Dormancy and Mnemonic Encoding (viability: 4): https://sciencetostartup.com/paper/enhanced-mycelium-of-thought-emot-a-bio-inspired-hierarchical-reasoning-architecture-with-strategic-dormancy-and-mnemoni - A bio-inspired hierarchical reasoning framework for LLMs that uses strategic dormancy and mnemonic encoding to improve complex, multi-domain problem-solving.
- SOMA: Strategic Orchestration and Memory-Augmented System for Vision-Language-Action Model Robustness via In-Context Adaptation (viability: 8): https://sciencetostartup.com/paper/soma-strategic-orchestration-and-memory-augmented-system-for-vision-language-action-model-robustness-via-in-context-adap - SOMA enhances existing robotic vision-language-action models for robust performance in challenging, out-of-distribution environments without retraining.
- AD-Reasoning: Multimodal Guideline-Guided Reasoning for Alzheimer's Disease Diagnosis (viability: 7): https://sciencetostartup.com/paper/ad-reasoning-multimodal-guideline-guided-reasoning-for-alzheimer-s-disease-diagnosis - A multimodal AI framework for Alzheimer's diagnosis that integrates neuroimaging and clinical data with explicit guideline adherence, offering transparent and accurate diagnostic reasoning.
- Mitigating Object Hallucinations in LVLMs via Attention Imbalance Rectification (viability: 7): https://sciencetostartup.com/paper/mitigating-object-hallucinations-in-lvlms-via-attention-imbalance-rectification - A lightweight decoding-time intervention method to significantly reduce object hallucinations in Large Vision-Language Models, improving their reliability for critical applications.
- Beyond Semantic Priors: Mitigating Optimization Collapse for Generalizable Visual Forensics (viability: 7): https://sciencetostartup.com/paper/beyond-semantic-priors-mitigating-optimization-collapse-for-generalizable-visual-forensics - A novel transformer architecture that mitigates optimization collapse in visual forensics, achieving state-of-the-art generalization for deepfake detection.
- Hierarchical Spatial-Temporal Graph-Enhanced Model for Map-Matching (viability: 7): https://sciencetostartup.com/paper/hierarchical-spatial-temporal-graph-enhanced-model-for-map-matching - A novel hierarchical graph-enhanced model for improved map-matching performance using self-supervised and supervised learning.
- FinToolSyn: A forward synthesis Framework for Financial Tool-Use Dialogue Data with Dynamic Tool Retrieval (viability: 7): https://sciencetostartup.com/paper/fintoolsyn-a-forward-synthesis-framework-for-financial-tool-use-dialogue-data-with-dynamic-tool-retrieval - A framework for generating realistic financial tool-use dialogue data to improve LLM capabilities in complex financial scenarios.
- PCHC: Enabling Preference Conditioned Humanoid Control via Multi-Objective Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/pchc-enabling-preference-conditioned-humanoid-control-via-multi-objective-reinforcement-learning - A reinforcement learning framework for humanoid robots that allows real-time adaptation of control objectives based on user preferences, enabling diverse and optimized behaviors.
- LGEST: Dynamic Spatial-Spectral Expert Routing for Hyperspectral Image Classification (viability: 7): https://sciencetostartup.com/paper/lgest-dynamic-spatial-spectral-expert-routing-for-hyperspectral-image-classification - A novel framework for hyperspectral image classification that dynamically routes spatial-spectral features through expert networks to overcome limitations of existing methods.
- MoE-Sieve: Routing-Guided LoRA for Efficient MoE Fine-Tuning (viability: 4): https://sciencetostartup.com/paper/moe-sieve-routing-guided-lora-for-efficient-moe-fine-tuning - A routing-guided framework for LoRA fine-tuning of Mixture-of-Experts models that significantly reduces trainable parameters and training time by focusing on the most activated experts.
- HAM: A Training-Free Style Transfer Approach via Heterogeneous Attention Modulation for Diffusion Models (viability: 7): https://sciencetostartup.com/paper/ham-a-training-free-style-transfer-approach-via-heterogeneous-attention-modulation-for-diffusion-models - A training-free method for diffusion model style transfer that preserves content identity and captures complex styles, outperforming existing approaches.
- Minimal Sufficient Representations for Self-interpretable Deep Neural Networks (viability: 7): https://sciencetostartup.com/paper/minimal-sufficient-representations-for-self-interpretable-deep-neural-networks - A self-interpretable neural network framework that identifies minimal representations to improve prediction accuracy and uncover human-interpretable patterns.
- SemLayer: Semantic-aware Generative Segmentation and Layer Construction for Abstract Icons (viability: 7): https://sciencetostartup.com/paper/semlayer-semantic-aware-generative-segmentation-and-layer-construction-for-abstract-icons - SemLayer reconstructs editable semantic layers from flattened vector icons, enabling advanced design workflows.
- A^3: Towards Advertising Aesthetic Assessment (viability: 8): https://sciencetostartup.com/paper/a-3-towards-advertising-aesthetic-assessment - A framework and multimodal LLM for objective, scalable, and interpretable assessment of advertising image aesthetics to improve commercial conversion rates.
- SpectralSplats: Robust Differentiable Tracking via Spectral Moment Supervision (viability: 7): https://sciencetostartup.com/paper/spectralsplats-robust-differentiable-tracking-via-spectral-moment-supervision - SpectralSplats offers robust 3D object tracking by using frequency domain supervision to overcome vanishing gradients in differentiable rendering, enabling recovery from severe misalignments.
- From Oracle to Noisy Context: Mitigating Contextual Exposure Bias in Speech-LLMs (viability: 7): https://sciencetostartup.com/paper/from-oracle-to-noisy-context-mitigating-contextual-exposure-bias-in-speech-llms - This research introduces a novel training framework to significantly improve the robustness of speech-to-LLM models against noisy and error-prone contextual information during inference, leading to more reliable real-world performance.
- Lagrangian Relaxation Score-based Generation for Mixed Integer linear Programming (viability: 7): https://sciencetostartup.com/paper/lagrangian-relaxation-score-based-generation-for-mixed-integer-linear-programming - A generative framework using Lagrangian relaxation and SDEs to produce diverse, high-quality solution candidates for mixed-integer linear programming, outperforming existing ML baselines and achieving competitive optimality with exact solvers at reduced computational cost.
- Decompose and Transfer: CoT-Prompting Enhanced Alignment for Open-Vocabulary Temporal Action Detection (viability: 4): https://sciencetostartup.com/paper/decompose-and-transfer-cot-prompting-enhanced-alignment-for-open-vocabulary-temporal-action-detection - A framework for open-vocabulary temporal action detection that uses LLM reasoning to decompose actions into phases for better generalization to unseen actions.
- i-IF-Learn: Iterative Feature Selection and Unsupervised Learning for High-Dimensional Complex Data (viability: 5): https://sciencetostartup.com/paper/i-if-learn-iterative-feature-selection-and-unsupervised-learning-for-high-dimensional-complex-data - An iterative unsupervised framework that jointly performs feature selection and clustering for high-dimensional complex data, outperforming classical and deep baselines.
- Schema on the Inside: A Two-Phase Fine-Tuning Method for High-Efficiency Text-to-SQL at Scale (viability: 8): https://sciencetostartup.com/paper/schema-on-the-inside-a-two-phase-fine-tuning-method-for-high-efficiency-text-to-sql-at-scale - A self-hosted, specialized text-to-SQL model that drastically cuts API costs and latency for conversational data querying in large-scale applications.
- QuadFM: Foundational Text-Driven Quadruped Motion Dataset for Generation and Control (viability: 8): https://sciencetostartup.com/paper/quadfm-foundational-text-driven-quadruped-motion-dataset-for-generation-and-control - A foundational dataset and unified framework for text-driven quadruped motion generation and control, enabling real-time, expressive robot movements.
- ELITE: Experiential Learning and Intent-Aware Transfer for Self-improving Embodied Agents (viability: 7): https://sciencetostartup.com/paper/elite-experiential-learning-and-intent-aware-transfer-for-self-improving-embodied-agents - An embodied agent framework that continuously learns from environment interaction to improve task execution and generalize to new tasks.
- COVTrack++: Learning Open-Vocabulary Multi-Object Tracking from Continuous Videos via a Synergistic Paradigm (viability: 7): https://sciencetostartup.com/paper/covtrack-learning-open-vocabulary-multi-object-tracking-from-continuous-videos-via-a-synergistic-paradigm - A novel open-vocabulary multi-object tracking system that learns from continuous video data and synergistically combines detection and association for improved performance on unseen objects.
- Language-Grounded Multi-Agent Planning for Personalized and Fair Participatory Urban Sensing (viability: 7): https://sciencetostartup.com/paper/language-grounded-multi-agent-planning-for-personalized-and-fair-participatory-urban-sensing - A multi-agent LLM framework that personalizes urban sensing tasks for participants, improving satisfaction and fairness.
- CVPD at QIAS 2026: RAG-Guided LLM Reasoning for Al-Mawarith Share Computation and Heir Allocation (viability: 7): https://sciencetostartup.com/paper/cvpd-at-qias-2026-rag-guided-llm-reasoning-for-al-mawarith-share-computation-and-heir-allocation - A RAG-powered AI system for accurate and reliable Islamic inheritance calculations and heir allocation, outperforming state-of-the-art.
- UW-VOS: A Large-Scale Dataset for Underwater Video Object Segmentation (viability: 7): https://sciencetostartup.com/paper/uw-vos-a-large-scale-dataset-for-underwater-video-object-segmentation - A new large-scale dataset and parameter-efficient framework for underwater video object segmentation that significantly outperforms existing methods.
- DB SwinT: A Dual-Branch Swin Transformer Network for Road Extraction in Optical Remote Sensing Imagery (viability: 5): https://sciencetostartup.com/paper/db-swint-a-dual-branch-swin-transformer-network-for-road-extraction-in-optical-remote-sensing-imagery - A dual-branch Swin Transformer network for more accurate road extraction in optical remote sensing imagery.
- Thinking with Tables: Enhancing Multi-Modal Tabular Understanding via Neuro-Symbolic Reasoning (viability: 8): https://sciencetostartup.com/paper/thinking-with-tables-enhancing-multi-modal-tabular-understanding-via-neuro-symbolic-reasoning - A neuro-symbolic reasoning system that significantly enhances multi-modal understanding of tabular data, outperforming existing baselines and rivaling commercial LLMs.
- PAC-DP: Personalized Adaptive Clipping for Differentially Private Federated Learning (viability: 7): https://sciencetostartup.com/paper/pac-dp-personalized-adaptive-clipping-for-differentially-private-federated-learning - A personalized adaptive clipping framework for federated learning that significantly improves accuracy and convergence speed while maintaining differential privacy.
- Stochastic Dimension-Free Zeroth-Order Estimator for High-Dimensional and High-Order PINNs (viability: 7): https://sciencetostartup.com/paper/stochastic-dimension-free-zeroth-order-estimator-for-high-dimensional-and-high-order-pinns - A novel zeroth-order optimization framework for training high-dimensional Physics-Informed Neural Networks with significantly reduced memory and computational complexity.
- Sparse Growing Transformer: Training-Time Sparse Depth Allocation via Progressive Attention Looping (viability: 7): https://sciencetostartup.com/paper/sparse-growing-transformer-training-time-sparse-depth-allocation-via-progressive-attention-looping - A novel training framework that dynamically allocates computational depth in Transformers to reduce training FLOPs by up to 19% while improving performance.
- HGGT: Robust and Flexible 3D Hand Mesh Reconstruction from Uncalibrated Images (viability: 7): https://sciencetostartup.com/paper/hggt-robust-and-flexible-3d-hand-mesh-reconstruction-from-uncalibrated-images - A feed-forward architecture that reconstructs 3D hand meshes and camera poses from uncalibrated images, outperforming state-of-the-art and generalizing to real-world scenarios.
- Forensic Implications of Localized AI: Artifact Analysis of Ollama, LM Studio, and llama.cpp (viability: 4): https://sciencetostartup.com/paper/forensic-implications-of-localized-ai-artifact-analysis-of-ollama-lm-studio-and-llama-cpp - This research provides forensic investigators with the tools and methodologies to uncover digital evidence from local LLM runners, addressing a critical blind spot in digital investigations.
- MIRROR: Visual Motion Imitation via Real-time Retargeting and Teleoperation with Parallel Differential Inverse Kinematics (viability: 7): https://sciencetostartup.com/paper/mirror-visual-motion-imitation-via-real-time-retargeting-and-teleoperation-with-parallel-differential-inverse-kinematics - Enabling real-time, safe humanoid robot teleoperation through advanced inverse kinematics and visual pose estimation.
- Understanding the Challenges in Iterative Generative Optimization with LLMs (viability: 3): https://sciencetostartup.com/paper/understanding-the-challenges-in-iterative-generative-optimization-with-llms - This paper identifies key challenges in setting up iterative generative optimization loops with LLMs, offering guidance for practitioners but not presenting a ready-to-deploy solution.
- From Untamed Black Box to Interpretable Pedagogical Orchestration: The Ensemble of Specialized LLMs Architecture for Adaptive Tutoring (viability: 7): https://sciencetostartup.com/paper/from-untamed-black-box-to-interpretable-pedagogical-orchestration-the-ensemble-of-specialized-llms-architecture-for-adap - An ensemble of specialized LLMs with a rule-based orchestrator and interpretable student model for reliable, controllable, and efficient adaptive tutoring.
- CoCR-RAG: Enhancing Retrieval-Augmented Generation in Web Q&A via Concept-oriented Context Reconstruction (viability: 7): https://sciencetostartup.com/paper/cocr-rag-enhancing-retrieval-augmented-generation-in-web-q-a-via-concept-oriented-context-reconstruction - A RAG framework that reconstructs multi-source web documents at a concept level to improve factual consistency and answer quality in Q&A systems.
- CAKE: Real-time Action Detection via Motion Distillation and Background-aware Contrastive Learning (viability: 7): https://sciencetostartup.com/paper/cake-real-time-action-detection-via-motion-distillation-and-background-aware-contrastive-learning - A real-time action detection system that distills motion knowledge into RGB models, achieving state-of-the-art performance on a single CPU.
- Can we generate portable representations for clinical time series data using LLMs? (viability: 7): https://sciencetostartup.com/paper/can-we-generate-portable-representations-for-clinical-time-series-data-using-llms - Leverage LLMs to create portable patient embeddings from clinical time series data, enabling scalable deployment of predictive models across hospitals with reduced retraining.
- Diet Your LLM: Dimension-wise Global Pruning of LLMs via Merging Task-specific Importance Score (viability: 7): https://sciencetostartup.com/paper/diet-your-llm-dimension-wise-global-pruning-of-llms-via-merging-task-specific-importance-score - A training-free method to efficiently prune LLMs by merging task-specific importance scores, improving accuracy without retraining.
- Transcending Classical Neural Network Boundaries: A Quantum-Classical Synergistic Paradigm for Seismic Data Processing (viability: 3): https://sciencetostartup.com/paper/transcending-classical-neural-network-boundaries-a-quantum-classical-synergistic-paradigm-for-seismic-data-processing - A quantum-classical generative adversarial network for seismic data processing that leverages quantum mechanics to overcome the representational limitations of traditional neural networks.
- SafeFlow: Real-Time Text-Driven Humanoid Whole-Body Control via Physics-Guided Rectified Flow and Selective Safety Gating (viability: 7): https://sciencetostartup.com/paper/safeflow-real-time-text-driven-humanoid-whole-body-control-via-physics-guided-rectified-flow-and-selective-safety-gating - A real-time text-driven humanoid control system that generates physically feasible and safe motion trajectories by integrating physics-guided generation with a multi-stage safety gate.
- Kirchhoff-Inspired Neural Networks for Evolving High-Order Perception (viability: 4): https://sciencetostartup.com/paper/kirchhoff-inspired-neural-networks-for-evolving-high-order-perception - A novel neural network architecture inspired by Kirchhoff's laws for improved data representation and physical consistency in tasks like PDE solving and image classification.
- SilLang: Improving Gait Recognition with Silhouette Language Encoding (viability: 7): https://sciencetostartup.com/paper/sillang-improving-gait-recognition-with-silhouette-language-encoding - Leveraging Large Language Models to improve pedestrian gait recognition by encoding binary silhouettes into a language-like representation.
- HyDRA: Hybrid Domain-Aware Robust Architecture for Heterogeneous Collaborative Perception (viability: 7): https://sciencetostartup.com/paper/hydra-hybrid-domain-aware-robust-architecture-for-heterogeneous-collaborative-perception - A novel architecture for collaborative perception that robustly handles agent heterogeneity without retraining, enabling scalable and cost-effective multi-agent systems.
- SLAT-Phys: Fast Material Property Field Prediction from Structured 3D Latents (viability: 7): https://sciencetostartup.com/paper/slat-phys-fast-material-property-field-prediction-from-structured-3d-latents - Predict material properties of 3D assets from a single image 120x faster than existing methods, enabling real-time physics simulation and digital twins.
- Grounding Arabic LLMs in the Doha Historical Dictionary: Retrieval-Augmented Understanding of Quran and Hadith (viability: 7): https://sciencetostartup.com/paper/grounding-arabic-llms-in-the-doha-historical-dictionary-retrieval-augmented-understanding-of-quran-and-hadith - A retrieval-augmented generation framework that grounds Arabic LLMs in historical lexicographic data to significantly improve understanding of religious texts.
- The Price Reversal Phenomenon: When Cheaper Reasoning Models End Up Costing More (viability: 7): https://sciencetostartup.com/paper/the-price-reversal-phenomenon-when-cheaper-reasoning-models-end-up-costing-more - A tool that predicts and optimizes LLM inference costs by analyzing token consumption, addressing the 'price reversal phenomenon' where cheaper listed models can cost more.
- Robust Distributed Cooperative Path-Following and Local Replanning for Multi-UAVs Under Differentiated Low-Altitude Paths (viability: 5): https://sciencetostartup.com/paper/robust-distributed-cooperative-path-following-and-local-replanning-for-multi-uavs-under-differentiated-low-altitude-path - Enables multiple drones to cooperatively follow complex 3D paths in low-altitude airspace, overcoming disturbances and obstacles with real-time replanning.
- Wireless communication empowers online scheduling of partially-observable transportation multi-robot systems in a smart factory (viability: 3): https://sciencetostartup.com/paper/wireless-communication-empowers-online-scheduling-of-partially-observable-transportation-multi-robot-systems-in-a-smart - A novel framework integrates wireless communication with multi-robot task assignment and route scheduling for enhanced efficiency in smart factories.
- Policy-Guided Threat Hunting: An LLM enabled Framework with Splunk SOC Triage (viability: 7): https://sciencetostartup.com/paper/policy-guided-threat-hunting-an-llm-enabled-framework-with-splunk-soc-triage - An LLM-powered framework integrated with Splunk to automate and prioritize threat hunting for overwhelmed SOC analysts.
- MonoSIM: An open source SIL framework for Ackermann Vehicular Systems with Monocular Vision (viability: 7): https://sciencetostartup.com/paper/monosim-an-open-source-sil-framework-for-ackermann-vehicular-systems-with-monocular-vision - An open-source simulation platform for developing and testing autonomous vehicle control systems using monocular vision.
- From Pixels to Digital Agents: An Empirical Study on the Taxonomy and Technological Trends of Reinforcement Learning Environments (viability: 5): https://sciencetostartup.com/paper/from-pixels-to-digital-agents-an-empirical-study-on-the-taxonomy-and-technological-trends-of-reinforcement-learning-envi - This research provides a data-driven taxonomy and analysis of reinforcement learning environments, identifying a paradigm shift towards LLM-driven agents and domain-specific generalization to guide the design of next-generation simulators.
- GRMLR: Knowledge-Enhanced Small-Data Learning for Deep-Sea Cold Seep Stage Inference (viability: 5): https://sciencetostartup.com/paper/grmlr-knowledge-enhanced-small-data-learning-for-deep-sea-cold-seep-stage-inference - A knowledge-enhanced machine learning model for accurate deep-sea cold seep stage inference using microbial data, reducing reliance on costly visual surveys.
- Leave No Stone Unturned: Uncovering Holistic Audio-Visual Intrinsic Coherence for Deepfake Detection (viability: 8): https://sciencetostartup.com/paper/leave-no-stone-unturned-uncovering-holistic-audio-visual-intrinsic-coherence-for-deepfake-detection - A novel deepfake detection system that leverages intrinsic audio-visual coherence, outperforming state-of-the-art with a new high-fidelity dataset and available code.
- PointRFT: Explicit Reinforcement Fine-tuning for Point Cloud Few-shot Learning (viability: 7): https://sciencetostartup.com/paper/pointrft-explicit-reinforcement-fine-tuning-for-point-cloud-few-shot-learning - A novel reinforcement fine-tuning method for 3D point cloud models that significantly improves performance in data-scarce scenarios.
- SynMVCrowd: A Large Synthetic Benchmark for Multi-view Crowd Counting and Localization (viability: 7): https://sciencetostartup.com/paper/synmvcrowd-a-large-synthetic-benchmark-for-multi-view-crowd-counting-and-localization - A large synthetic benchmark and baseline methods for more practical multi-view crowd counting and localization, advancing research towards real-world applications.
- VOLMO: Versatile and Open Large Models for Ophthalmology (viability: 6): https://sciencetostartup.com/paper/volmo-versatile-and-open-large-models-for-ophthalmology - A framework for building and deploying ophthalmology-specific multimodal large language models that outperform existing general and medical models.
- From AI Assistant to AI Scientist: Autonomous Discovery of LLM-RL Algorithms with LLM Agents (viability: 7): https://sciencetostartup.com/paper/from-ai-assistant-to-ai-scientist-autonomous-discovery-of-llm-rl-algorithms-with-llm-agents - Automate the discovery of improved policy optimization algorithms for language models, significantly boosting performance on complex reasoning tasks.
- Event-Driven Proactive Assistive Manipulation with Grounded Vision-Language Planning (viability: 5): https://sciencetostartup.com/paper/event-driven-proactive-assistive-manipulation-with-grounded-vision-language-planning - A robot system that proactively assists human manipulation by inferring goals from observed state changes, rather than waiting for explicit instructions.
- Argument Mining as a Text-to-Text Generation Task (viability: 7): https://sciencetostartup.com/paper/argument-mining-as-a-text-to-text-generation-task - A text-to-text generation approach for argument mining that simplifies the process and achieves state-of-the-art results.
- Variable-Length Audio Fingerprinting (viability: 7): https://sciencetostartup.com/paper/variable-length-audio-fingerprinting - A novel deep learning method for audio fingerprinting that handles variable-length audio, outperforming existing state-of-the-art on real-world datasets.
- High-Fidelity Face Content Recovery via Tamper-Resilient Versatile Watermarking (viability: 4): https://sciencetostartup.com/paper/high-fidelity-face-content-recovery-via-tamper-resilient-versatile-watermarking - A watermarking framework for copyright protection, manipulation localization, and high-fidelity face content recovery to combat deepfakes.
- OmniACBench: A Benchmark for Evaluating Context-Grounded Acoustic Control in Omni-Modal Models (viability: 7): https://sciencetostartup.com/paper/omniacbench-a-benchmark-for-evaluating-context-grounded-acoustic-control-in-omni-modal-models - A benchmark and analysis tool to evaluate and improve the context-aware speech generation capabilities of omni-modal AI models.
- Dialogue to Question Generation for Evidence-based Medical Guideline Agent Development (viability: 7): https://sciencetostartup.com/paper/dialogue-to-question-generation-for-evidence-based-medical-guideline-agent-development - An AI assistant that generates targeted, evidence-based questions during physician-patient encounters to improve the implementation of medical guidelines.
- An Empirical Analysis of Google Play Data Safety Disclosures: A Consistency Study of Privacy Indicators in Mobile Gaming Apps (viability: 4): https://sciencetostartup.com/paper/an-empirical-analysis-of-google-play-data-safety-disclosures-a-consistency-study-of-privacy-indicators-in-mobile-gaming - This research analyzes the consistency of Google Play's Data Safety disclosures against actual app behavior, revealing significant inconsistencies in privacy reporting that necessitate improved validation mechanisms.
- Revealing Multi-View Hallucination in Large Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/revealing-multi-view-hallucination-in-large-vision-language-models - A novel decoding technique that significantly reduces multi-view hallucination in vision-language models by suppressing visual interference.
- ORACLE: Orchestrate NPC Daily Activities using Contrastive Learning with Transformer-CVAE (viability: 5): https://sciencetostartup.com/paper/oracle-orchestrate-npc-daily-activities-using-contrastive-learning-with-transformer-cvae - Generate realistic and varied NPC daily activity plans for immersive digital environments using a novel generative model.
- Optimal Variance-Dependent Regret Bounds for Infinite-Horizon MDPs (viability: 2): https://sciencetostartup.com/paper/optimal-variance-dependent-regret-bounds-for-infinite-horizon-mdps - Develops theoretical optimal regret bounds for infinite-horizon reinforcement learning problems.
- DP^2-VL: Private Photo Dataset Protection by Data Poisoning for Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/dp-2-vl-private-photo-dataset-protection-by-data-poisoning-for-vision-language-models - A framework to protect private photos from being exploited by vision-language models through data poisoning, preventing identity-affiliation leakage.
- DepthArb: Training-Free Depth-Arbitrated Generation for Occlusion-Robust Image Synthesis (viability: 7): https://sciencetostartup.com/paper/deptharb-training-free-depth-arbitrated-generation-for-occlusion-robust-image-synthesis - A training-free framework that improves occlusion accuracy in text-to-image generation by arbitrating object attention, enhancing compositional capabilities of existing diffusion models.
- Elements of Conformal Prediction for Statisticians (viability: 2): https://sciencetostartup.com/paper/elements-of-conformal-prediction-for-statisticians - A theoretical overview of conformal prediction methods for statistical inference.
- Uncertainty-Aware Vision-based Risk Object Identification via Conformal Risk Tube Prediction (viability: 8): https://sciencetostartup.com/paper/uncertainty-aware-vision-based-risk-object-identification-via-conformal-risk-tube-prediction - A novel AI system for hazard detection in intelligent driving that quantifies risk uncertainty to improve safety and reduce false alarms.
- DecepGPT: Schema-Driven Deception Detection with Multicultural Datasets and Robust Multimodal Learning (viability: 7): https://sciencetostartup.com/paper/decepgpt-schema-driven-deception-detection-with-multicultural-datasets-and-robust-multimodal-learning - A multimodal deception detection system that provides auditable reports and robust cross-cultural generalization, powered by novel reasoning datasets and advanced representation learning.
- Attention-aware Inference Optimizations for Large Vision-Language Models with Memory-efficient Decoding (viability: 7): https://sciencetostartup.com/paper/attention-aware-inference-optimizations-for-large-vision-language-models-with-memory-efficient-decoding - Optimize large vision-language models for faster and more memory-efficient inference, enabling longer contexts and higher throughput.
- Self-Distillation for Multi-Token Prediction (viability: 7): https://sciencetostartup.com/paper/self-distillation-for-multi-token-prediction - Accelerate LLM inference by up to 220% with a self-distillation method that improves multi-token prediction accuracy and efficiency.
- AnalogAgent: Self-Improving Analog Circuit Design Automation with LLM Agents (viability: 7): https://sciencetostartup.com/paper/analogagent-self-improving-analog-circuit-design-automation-with-llm-agents - AnalogAgent automates analog circuit design using a multi-agent LLM system with self-evolving memory, significantly improving the performance of existing models.
- DUPLEX: Agentic Dual-System Planning via LLM-Driven Information Extraction (viability: 7): https://sciencetostartup.com/paper/duplex-agentic-dual-system-planning-via-llm-driven-information-extraction - A neuro-symbolic architecture that uses LLMs for structured information extraction to enable reliable, long-horizon robotic task planning.
- GenMask: Adapting DiT for Segmentation via Direct Mask (viability: 7): https://sciencetostartup.com/paper/genmask-adapting-dit-for-segmentation-via-direct-mask - GenMask enables direct generative training for segmentation tasks by adapting Diffusion Transformer (DiT) to generate masks alongside images, achieving state-of-the-art performance without complex feature extraction pipelines.
- Latent Bias Alignment for High-Fidelity Diffusion Inversion in Real-World Image Reconstruction and Manipulation (viability: 7): https://sciencetostartup.com/paper/latent-bias-alignment-for-high-fidelity-diffusion-inversion-in-real-world-image-reconstruction-and-manipulation - A novel method for high-fidelity diffusion model inversion that significantly improves image reconstruction and editing capabilities.
- Knowledge-Refined Dual Context-Aware Network for Partially Relevant Video Retrieval (viability: 7): https://sciencetostartup.com/paper/knowledge-refined-dual-context-aware-network-for-partially-relevant-video-retrieval - A novel network for precise video segment retrieval that overcomes information density mismatches and attention limitations by refining knowledge and context.
- MMTIT-Bench: A Multilingual and Multi-Scenario Benchmark with Cognition-Perception-Reasoning Guided Text-Image Machine Translation (viability: 7): https://sciencetostartup.com/paper/mmtit-bench-a-multilingual-and-multi-scenario-benchmark-with-cognition-perception-reasoning-guided-text-image-machine-tr - A new benchmark and data paradigm for robust multilingual text-image translation, improving accuracy and interpretability in vision-language models.
- FilterGS: Traversal-Free Parallel Filtering and Adaptive Shrinking for Large-Scale LoD 3D Gaussian Splatting (viability: 7): https://sciencetostartup.com/paper/filtergs-traversal-free-parallel-filtering-and-adaptive-shrinking-for-large-scale-lod-3d-gaussian-splatting - Accelerate large-scale 3D scene rendering with a traversal-free parallel filtering and adaptive shrinking method for Gaussian Splatting.
- Praxium: Diagnosing Cloud Anomalies with AI-based Telemetry and Dependency Analysis (viability: 7): https://sciencetostartup.com/paper/praxium-diagnosing-cloud-anomalies-with-ai-based-telemetry-and-dependency-analysis - Praxium automates cloud anomaly detection and root cause inference by analyzing telemetry and software installation dependencies, enabling faster resolution for SREs.
- Off-Policy Safe Reinforcement Learning with Constrained Optimistic Exploration (viability: 7): https://sciencetostartup.com/paper/off-policy-safe-reinforcement-learning-with-constrained-optimistic-exploration - A novel off-policy safe RL algorithm that enables cost-bounded exploration and stable value learning for safety-critical applications.
- AgentChemist: A Multi-Agent Experimental Robotic Platform Integrating Chemical Perception and Precise Control (viability: 7): https://sciencetostartup.com/paper/agentchemist-a-multi-agent-experimental-robotic-platform-integrating-chemical-perception-and-precise-control - A multi-agent robotic platform that integrates chemical perception and precise control to enable flexible and scalable laboratory automation for diverse, long-tail experimental tasks.
- Towards Real-World Document Parsing via Realistic Scene Synthesis and Document-Aware Training (viability: 7): https://sciencetostartup.com/paper/towards-real-world-document-parsing-via-realistic-scene-synthesis-and-document-aware-training - A framework for generating realistic document parsing data and training strategies to improve accuracy and robustness in real-world scenarios.
- BioVITA: Biological Dataset, Model, and Benchmark for Visual-Textual-Acoustic Alignment (viability: 7): https://sciencetostartup.com/paper/biovita-biological-dataset-model-and-benchmark-for-visual-textual-acoustic-alignment - A multimodal AI framework and dataset for understanding animal species from visual, textual, and acoustic data, advancing biodiversity research.
- The Luna Bound Propagator for Formal Analysis of Neural Networks (viability: 7): https://sciencetostartup.com/paper/the-luna-bound-propagator-for-formal-analysis-of-neural-networks - A C++ neural network bound propagator that integrates with existing verifiers and production systems, offering competitive performance.
- EnvSocial-Diff: A Diffusion-Based Crowd Simulation Model with Environmental Conditioning and Individual-Group Interaction (viability: 7): https://sciencetostartup.com/paper/envsocial-diff-a-diffusion-based-crowd-simulation-model-with-environmental-conditioning-and-individual-group-interaction - A diffusion-based crowd simulation model that integrates environmental context and social interactions to generate more realistic pedestrian trajectories.
- The DeepXube Software Package for Solving Pathfinding Problems with Learned Heuristic Functions and Search (viability: 7): https://sciencetostartup.com/paper/the-deepxube-software-package-for-solving-pathfinding-problems-with-learned-heuristic-functions-and-search - Automate pathfinding solutions using learned heuristics and deep neural networks with an open-source Python package.
- HDPO: Hybrid Distillation Policy Optimization via Privileged Self-Distillation (viability: 4): https://sciencetostartup.com/paper/hdpo-hybrid-distillation-policy-optimization-via-privileged-self-distillation - A novel reinforcement learning technique for large language models that improves mathematical reasoning by distilling privileged information on failure cases.
- MLE-UVAD: Minimal Latent Entropy Autoencoder for Fully Unsupervised Video Anomaly Detection (viability: 7): https://sciencetostartup.com/paper/mle-uvad-minimal-latent-entropy-autoencoder-for-fully-unsupervised-video-anomaly-detection - A novel unsupervised video anomaly detection method that uses an entropy-guided autoencoder to achieve superior performance by minimizing latent space entropy for clearer anomaly identification.
- Can VLMs Reason Robustly? A Neuro-Symbolic Investigation (viability: 4): https://sciencetostartup.com/paper/can-vlms-reason-robustly-a-neuro-symbolic-investigation - A neuro-symbolic method that combines VLM concept recognition with circuit-based symbolic reasoning to achieve robust visual deductive reasoning under distribution shifts.
- See, Remember, Explore: A Benchmark and Baselines for Streaming Spatial Reasoning (viability: 7): https://sciencetostartup.com/paper/see-remember-explore-a-benchmark-and-baselines-for-streaming-spatial-reasoning - A new benchmark and baseline model for embodied agents to perform spatial reasoning in real-time streaming video with active exploration.
- Generative AI User Experience: Developing Human--AI Epistemic Partnership (viability: 3): https://sciencetostartup.com/paper/generative-ai-user-experience-developing-human-ai-epistemic-partnership - A new theory to understand and design user experiences for generative AI that goes beyond simple utility to encompass knowledge co-construction and dynamic partnership.
- Deep Convolutional Neural Networks for predicting highest priority functional group in organic molecules (viability: 3): https://sciencetostartup.com/paper/deep-convolutional-neural-networks-for-predicting-highest-priority-functional-group-in-organic-molecules - A deep convolutional neural network predicts the highest priority functional group in organic molecules from FTIR spectra, outperforming SVM.
- An Invariant Compiler for Neural ODEs in AI-Accelerated Scientific Simulation (viability: 3): https://sciencetostartup.com/paper/an-invariant-compiler-for-neural-odes-in-ai-accelerated-scientific-simulation - A compiler framework to enforce scientific invariants in neural ODEs for more accurate simulations.
- Why the Maximum Second Derivative of Activations Matters for Adversarial Robustness (viability: 4): https://sciencetostartup.com/paper/why-the-maximum-second-derivative-of-activations-matters-for-adversarial-robustness - A novel activation function design that optimizes adversarial robustness by precisely controlling curvature, leading to improved generalization.
- When AI output tips to bad but nobody notices: Legal implications of AI's mistakes (viability: 4): https://sciencetostartup.com/paper/when-ai-output-tips-to-bad-but-nobody-notices-legal-implications-of-ai-s-mistakes - Develop verification protocols for generative AI in legal settings to prevent fabricated case law and mitigate professional risks.
- Symbolic--KAN: Kolmogorov-Arnold Networks with Discrete Symbolic Structure for Interpretable Learning (viability: 7): https://sciencetostartup.com/paper/symbolic-kan-kolmogorov-arnold-networks-with-discrete-symbolic-structure-for-interpretable-learning - Symbolic-KANs embed discrete symbolic structure into neural networks for interpretable and scalable discovery of governing equations in scientific machine learning.
- SCoOP: Semantic Consistent Opinion Pooling for Uncertainty Quantification in Multiple Vision-Language Model Systems (viability: 7): https://sciencetostartup.com/paper/scoop-semantic-consistent-opinion-pooling-for-uncertainty-quantification-in-multiple-vision-language-model-systems - A training-free framework to quantify uncertainty and detect hallucinations in multi-Vision-Language Model systems, improving reliability with minimal overhead.
- BeliefShift: Benchmarking Temporal Belief Consistency and Opinion Drift in LLM Agents (viability: 7): https://sciencetostartup.com/paper/beliefshift-benchmarking-temporal-belief-consistency-and-opinion-drift-in-llm-agents - Benchmark and evaluate LLM agents for temporal belief consistency and opinion drift in long-running conversations.
- 3D-LLDM: Label-Guided 3D Latent Diffusion Model for Improving High-Resolution Synthetic MR Imaging in Hepatic Structure Segmentation (viability: 7): https://sciencetostartup.com/paper/3d-lldm-label-guided-3d-latent-diffusion-model-for-improving-high-resolution-synthetic-mr-imaging-in-hepatic-structure-s - A label-guided 3D latent diffusion model that generates high-resolution synthetic MR images with anatomical segmentation masks to improve medical image analysis.
- Language Model Planners do not Scale, but do Formalizers? (viability: 3): https://sciencetostartup.com/paper/language-model-planners-do-not-scale-but-do-formalizers - This paper explores the scalability of LLM formalizers for complex planning problems, introducing a novel 'LLM-as-higher-order-formalizer' paradigm to address combinatorial explosion.
- PoliticsBench: Benchmarking Political Values in Large Language Models with Multi-Turn Roleplay (viability: 4): https://sciencetostartup.com/paper/politicsbench-benchmarking-political-values-in-large-language-models-with-multi-turn-roleplay - A new benchmark and framework to evaluate and quantify political bias in large language models through multi-turn roleplay.
- VehicleMemBench: An Executable Benchmark for Multi-User Long-Term Memory in In-Vehicle Agents (viability: 7): https://sciencetostartup.com/paper/vehiclemembench-an-executable-benchmark-for-multi-user-long-term-memory-in-in-vehicle-agents - A new benchmark and simulation environment for developing in-vehicle agents with robust long-term, multi-user memory capabilities.
- Learning-guided Prioritized Planning for Lifelong Multi-Agent Path Finding in Warehouse Automation (viability: 5): https://sciencetostartup.com/paper/learning-guided-prioritized-planning-for-lifelong-multi-agent-path-finding-in-warehouse-automation - A reinforcement learning framework that guides prioritized planning for more efficient multi-robot navigation in warehouses.
- Beyond Consistency: Inference for the Relative risk functional in Deep Nonparametric Cox Models (viability: 3): https://sciencetostartup.com/paper/beyond-consistency-inference-for-the-relative-risk-functional-in-deep-nonparametric-cox-models - Develops theoretical foundations for deep neural network estimators in nonparametric Cox models to enable valid inference for relative risk.
- Unveiling Hidden Convexity in Deep Learning: a Sparse Signal Processing Perspective (viability: 2): https://sciencetostartup.com/paper/unveiling-hidden-convexity-in-deep-learning-a-sparse-signal-processing-perspective - This paper explores theoretical connections between deep learning and sparse signal processing to understand the non-convexity of neural network loss functions.
- An Adaptive Neuro-Fuzzy Blockchain-AI Framework for Secure and Intelligent FinTech Transactions (viability: 4): https://sciencetostartup.com/paper/an-adaptive-neuro-fuzzy-blockchain-ai-framework-for-secure-and-intelligent-fintech-transactions - An adaptive neuro-fuzzy blockchain framework enhances FinTech transaction security by detecting complex fraud schemes with improved accuracy and reduced latency.
- Circuit Complexity of Hierarchical Knowledge Tracing and Implications for Log-Precision Transformers (viability: 3): https://sciencetostartup.com/paper/circuit-complexity-of-hierarchical-knowledge-tracing-and-implications-for-log-precision-transformers - This research theoretically analyzes the circuit complexity of hierarchical knowledge tracing, suggesting structure-aware objectives for improved learning on deep concept hierarchies.
- How Vulnerable Are Edge LLMs? (viability: 7): https://sciencetostartup.com/paper/how-vulnerable-are-edge-llms - A structured query framework to extract sensitive knowledge from quantized edge-deployed LLMs, revealing a significant security vulnerability.
- Perturbation: A simple and efficient adversarial tracer for representation learning in language models (viability: 3): https://sciencetostartup.com/paper/perturbation-a-simple-and-efficient-adversarial-tracer-for-representation-learning-in-language-models - A novel method to understand internal representations in language models by observing how they react to adversarial perturbations.
- Aesthetics of Robot-Mediated Applied Drama: A Case Study on REMind (viability: 3): https://sciencetostartup.com/paper/aesthetics-of-robot-mediated-applied-drama-a-case-study-on-remind - This paper explores using social robots in dramatic role-playing games to support children's social-emotional learning, focusing on anti-bullying scenarios.
- Willful Disobedience: Automatically Detecting Failures in Agentic Traces (viability: 7): https://sciencetostartup.com/paper/willful-disobedience-automatically-detecting-failures-in-agentic-traces - AgentPex is an AI-powered tool that automatically validates AI agent execution traces by detecting specification violations missed by outcome-only benchmarks.
- Deep Neural Regression Collapse (viability: 3): https://sciencetostartup.com/paper/deep-neural-regression-collapse - This paper theoretically analyzes a phenomenon called Neural Regression Collapse in deep networks, revealing insights into their internal structure for regression tasks.
- AgentRFC: Security Design Principles and Conformance Testing for Agent Protocols (viability: 7): https://sciencetostartup.com/paper/agentrfc-security-design-principles-and-conformance-testing-for-agent-protocols - A security framework and conformance checker for AI agent protocols to ensure secure composition and prevent cross-protocol vulnerabilities.
- Object Search in Partially-Known Environments via LLM-informed Model-based Planning and Prompt Selection (viability: 7): https://sciencetostartup.com/paper/object-search-in-partially-known-environments-via-llm-informed-model-based-planning-and-prompt-selection - An LLM-informed planning framework for efficient object search in partially-known environments, validated on real robots.
- Resolving gradient pathology in physics-informed epidemiological models (viability: 4): https://sciencetostartup.com/paper/resolving-gradient-pathology-in-physics-informed-epidemiological-models - A novel method for stable and efficient training of physics-informed neural networks in epidemiology, improving parameter estimation and peak recovery.
- Infrequent Child-Directed Speech Is Bursty and May Draw Infant Vocalizations (viability: 1): https://sciencetostartup.com/paper/infrequent-child-directed-speech-is-bursty-and-may-draw-infant-vocalizations - This research investigates the temporal patterns of child-directed speech and infant vocalizations to understand language development in low-input environments.
- Sparse Autoencoders for Interpretable Medical Image Representation Learning (viability: 7): https://sciencetostartup.com/paper/sparse-autoencoders-for-interpretable-medical-image-representation-learning - Develop interpretable medical imaging AI by replacing opaque foundation model representations with human-understandable sparse features derived from Sparse Autoencoders.
- AetherWeave: Sybil-Resistant Robust Peer Discovery with Stake (viability: 7): https://sciencetostartup.com/paper/aetherweave-sybil-resistant-robust-peer-discovery-with-stake - AetherWeave offers Sybil-resistant and private peer discovery for P2P networks by tying participation to staked assets, securing blockchains against adversarial attacks.
- Manifold Generalization Provably Proceeds Memorization in Diffusion Models (viability: 2): https://sciencetostartup.com/paper/manifold-generalization-provably-proceeds-memorization-in-diffusion-models - This paper theoretically explains how diffusion models generalize by capturing data geometry rather than just density estimation, achieving faster rates than full distribution estimation.
- The Cognitive Firewall:Securing Browser Based AI Agents Against Indirect Prompt Injection Via Hybrid Edge Cloud Defense (viability: 7): https://sciencetostartup.com/paper/the-cognitive-firewall-securing-browser-based-ai-agents-against-indirect-prompt-injection-via-hybrid-edge-cloud-defense - A hybrid edge-cloud system that drastically reduces latency and improves security for browser-based AI agents by preventing indirect prompt injection attacks.
- Re-Prompting SAM 3 via Object Retrieval: 3rd of the 5th PVUW MOSE Track (viability: 4): https://sciencetostartup.com/paper/re-prompting-sam-3-via-object-retrieval-3rd-of-the-5th-pvuw-mose-track - An automatic re-prompting framework for semi-supervised video object segmentation that improves robustness to target disappearance and distractors.
- Digital Twin-Assisted Measurement Design and Channel Statistics Prediction (viability: 4): https://sciencetostartup.com/paper/digital-twin-assisted-measurement-design-and-channel-statistics-prediction - A framework for data-efficient wireless channel prediction using uncalibrated digital twins and Gaussian processes to reduce measurement overhead and improve accuracy.
- Retinal Disease Classification from Fundus Images using CNN Transfer Learning (viability: 5): https://sciencetostartup.com/paper/retinal-disease-classification-from-fundus-images-using-cnn-transfer-learning - A reproducible deep learning pipeline for early detection of retinal diseases from fundus images using transfer learning.
- Latent Algorithmic Structure Precedes Grokking: A Mechanistic Study of ReLU MLPs on Modular Arithmetic (viability: 2): https://sciencetostartup.com/paper/latent-algorithmic-structure-precedes-grokking-a-mechanistic-study-of-relu-mlps-on-modular-arithmetic - This research investigates the internal algorithmic structure of ReLU MLPs during the grokking phenomenon, revealing how weight distributions evolve to enable generalization.
- Probabilistic Geometric Alignment via Bayesian Latent Transport for Domain-Adaptive Foundation Models (viability: 5): https://sciencetostartup.com/paper/probabilistic-geometric-alignment-via-bayesian-latent-transport-for-domain-adaptive-foundation-models - A probabilistic framework for adapting large foundation models to new domains by aligning latent spaces with theoretical guarantees.
- Leveraging Large Language Models for Trustworthiness Assessment of Web Applications (viability: 5): https://sciencetostartup.com/paper/leveraging-large-language-models-for-trustworthiness-assessment-of-web-applications - Leveraging LLMs to automate the trustworthiness assessment of web applications by verifying adherence to secure coding practices.
- Lightweight Fairness for LLM-Based Recommendations via Kernelized Projection and Gated Adapters (viability: 7): https://sciencetostartup.com/paper/lightweight-fairness-for-llm-based-recommendations-via-kernelized-projection-and-gated-adapters - A lightweight method to mitigate bias in LLM-based recommendation systems without retraining, preserving accuracy.
- Human-in-the-Loop Pareto Optimization: Trade-off Characterization for Assist-as-Needed Training and Performance Evaluation (viability: 7): https://sciencetostartup.com/paper/human-in-the-loop-pareto-optimization-trade-off-characterization-for-assist-as-needed-training-and-performance-evaluatio - A human-in-the-loop Pareto optimization framework to characterize and leverage the trade-off between task performance and perceived challenge for personalized training and evaluation.
- AI-driven Intent-Based Networking Approach for Self-configuration of Next Generation Networks (viability: 4): https://sciencetostartup.com/paper/ai-driven-intent-based-networking-approach-for-self-configuration-of-next-generation-networks - An AI-driven approach to automate network configuration and assurance by translating high-level intents into enforceable policies and proactively predicting failures.
- Semantic Iterative Reconstruction: One-Shot Universal Anomaly Detection (viability: 8): https://sciencetostartup.com/paper/semantic-iterative-reconstruction-one-shot-universal-anomaly-detection - A universal AI model that detects anomalies in medical images with minimal normal samples, outperforming existing methods across diverse domains.
- Learning Cross-Joint Attention for Generalizable Video-Based Seizure Detection (viability: 5): https://sciencetostartup.com/paper/learning-cross-joint-attention-for-generalizable-video-based-seizure-detection - A novel attention model for generalized seizure detection from clinical videos by focusing on body dynamics.
- Self Paced Gaussian Contextual Reinforcement Learning (viability: 5): https://sciencetostartup.com/paper/self-paced-gaussian-contextual-reinforcement-learning - A novel self-paced curriculum learning method for reinforcement learning that reduces computational overhead using a closed-form update rule for Gaussian context distributions.
- IJmond Industrial Smoke Segmentation Dataset (viability: 4): https://sciencetostartup.com/paper/ijmond-industrial-smoke-segmentation-dataset - A new dataset for industrial smoke segmentation is released to facilitate research and development in this niche computer vision area.
- Task-Space Singularity Avoidance for Control Affine Systems Using Control Barrier Functions (viability: 5): https://sciencetostartup.com/paper/task-space-singularity-avoidance-for-control-affine-systems-using-control-barrier-functions - A control barrier function framework to prevent singularities in robotic systems for smoother trajectory tracking and reduced control spikes.
- IslamicMMLU: A Benchmark for Evaluating LLMs on Islamic Knowledge (viability: 7): https://sciencetostartup.com/paper/islamicmmlu-a-benchmark-for-evaluating-llms-on-islamic-knowledge - A new benchmark and leaderboard to rigorously evaluate LLMs on Islamic knowledge, revealing performance gaps and biases.
- Efficient Benchmarking of AI Agents (viability: 7): https://sciencetostartup.com/paper/efficient-benchmarking-of-ai-agents - A cost-effective protocol for reliably ranking AI agents by evaluating them on a filtered subset of tasks, reducing evaluation costs by up to 70% while maintaining high rank fidelity.
- Kronecker-Structured Nonparametric Spatiotemporal Point Processes (viability: 5): https://sciencetostartup.com/paper/kronecker-structured-nonparametric-spatiotemporal-point-processes - A flexible and interpretable nonparametric spatiotemporal point process model for uncovering complex event relationships and enabling accurate prediction.
- Space Fabric: A Satellite-Enhanced Trusted Execution Architecture (viability: 3): https://sciencetostartup.com/paper/space-fabric-a-satellite-enhanced-trusted-execution-architecture - A satellite-enhanced architecture for secure orbital computing, ensuring trust without physical access or pre-provisioned secrets.
- Detection and Classification of (Pre)Cancerous Cells in Pap Smears: An Ensemble Strategy for the RIVA Cervical Cytology Challenge (viability: 7): https://sciencetostartup.com/paper/detection-and-classification-of-pre-cancerous-cells-in-pap-smears-an-ensemble-strategy-for-the-riva-cervical-cytology-ch - An ensemble strategy using YOLOv11m to significantly improve the detection and classification of precancerous cells in Pap smears, outperforming individual models by 29%.
- BXRL: Behavior-Explainable Reinforcement Learning (viability: 3): https://sciencetostartup.com/paper/bxrl-behavior-explainable-reinforcement-learning - A new framework for Reinforcement Learning that defines and measures agent behaviors to provide more interpretable explanations.
- Wasserstein Parallel Transport for Predicting the Dynamics of Statistical Systems (viability: 4): https://sciencetostartup.com/paper/wasserstein-parallel-transport-for-predicting-the-dynamics-of-statistical-systems - A novel method for predicting the evolution of complex systems described by probability distributions, enabling causal inference and domain adaptation.
- An Adapter-free Fine-tuning Approach for Tuning 3D Foundation Models (viability: 7): https://sciencetostartup.com/paper/an-adapter-free-fine-tuning-approach-for-tuning-3d-foundation-models - Adapter-free fine-tuning for 3D foundation models that improves performance in low-data regimes without increasing inference latency.
- Bi-CRCL: Bidirectional Conservative-Radical Complementary Learning with Pre-trained Foundation Models for Class-incremental Medical Image Analysis (viability: 7): https://sciencetostartup.com/paper/bi-crcl-bidirectional-conservative-radical-complementary-learning-with-pre-trained-foundation-models-for-class-increment - A dual-learner framework for medical image analysis that preserves existing diagnostic knowledge while rapidly adapting to new disease categories, overcoming privacy constraints and data heterogeneity.
- Form-Fitting, Large-Area Sensor Mounting for Obstacle Detection (viability: 4): https://sciencetostartup.com/paper/form-fitting-large-area-sensor-mounting-for-obstacle-detection - A low-cost, calibration-free method to procedurally generate robot skins for mounting sensors, enabling large-area obstacle detection.
- Dual-Gated Epistemic Time-Dilation: Autonomous Compute Modulation in Asynchronous MARL (viability: 7): https://sciencetostartup.com/paper/dual-gated-epistemic-time-dilation-autonomous-compute-modulation-in-asynchronous-marl - Autonomous compute modulation for edge-deployed multi-agent systems that drastically reduces computational overhead without sacrificing performance.
- CDMT-EHR: A Continuous-Time Diffusion Framework for Generating Mixed-Type Time-Series Electronic Health Records (viability: 7): https://sciencetostartup.com/paper/cdmt-ehr-a-continuous-time-diffusion-framework-for-generating-mixed-type-time-series-electronic-health-records - A continuous-time diffusion model for generating realistic, privacy-preserving mixed-type electronic health records.
- LLMs Do Not Grade Essays Like Humans (viability: 3): https://sciencetostartup.com/paper/llms-do-not-grade-essays-like-humans - Evaluating LLMs for essay scoring reveals their scoring patterns differ from human graders, limiting their direct application without further refinement.
- Mind the Hitch: Dynamic Calibration and Articulated Perception for Autonomous Trucks (viability: 8): https://sciencetostartup.com/paper/mind-the-hitch-dynamic-calibration-and-articulated-perception-for-autonomous-trucks - A vision-based framework for autonomous trucks that dynamically calibrates perception systems to handle articulated trailer geometry, improving object detection accuracy.
- An In-Depth Study of Filter-Agnostic Vector Search on a PostgreSQL Database System: [Experiments and Analysis] (viability: 5): https://sciencetostartup.com/paper/an-in-depth-study-of-filter-agnostic-vector-search-on-a-postgresql-database-system-experiments-and-analysis - Optimizing vector search within production databases by analyzing system-level overheads beyond theoretical performance.
- The Diminishing Returns of Early-Exit Decoding in Modern LLMs (viability: 4): https://sciencetostartup.com/paper/the-diminishing-returns-of-early-exit-decoding-in-modern-llms - A framework to identify and exploit early-exit opportunities in LLM inference, reducing latency and cost for specific model architectures and sizes.
- CoRe: Joint Optimization with Contrastive Learning for Medical Image Registration (viability: 7): https://sciencetostartup.com/paper/core-joint-optimization-with-contrastive-learning-for-medical-image-registration - A novel framework for medical image registration that uses contrastive learning to improve alignment accuracy and robustness.
- ROSCell: A ROS2-Based Framework for Automated Formation and Orchestration of Multi-Robot Systems (viability: 7): https://sciencetostartup.com/paper/roscell-a-ros2-based-framework-for-automated-formation-and-orchestration-of-multi-robot-systems - A ROS2 framework for dynamically forming and orchestrating multi-robot systems, reducing overhead for flexible manufacturing.
- Adaptive Gaussian Process Search for Simulation-Based Sample Size Estimation in Clinical Prediction Models: Validation of the pmsims R Package (viability: 5): https://sciencetostartup.com/paper/adaptive-gaussian-process-search-for-simulation-based-sample-size-estimation-in-clinical-prediction-models-validation-of - An R package using adaptive Gaussian processes to efficiently estimate optimal sample sizes for clinical prediction models, outperforming existing methods in challenging scenarios.
- AdvSplat: Adversarial Attacks on Feed-Forward Gaussian Splatting Models (viability: 4): https://sciencetostartup.com/paper/advsplat-adversarial-attacks-on-feed-forward-gaussian-splatting-models - Develops adversarial attack methods to highlight security vulnerabilities in feed-forward 3D Gaussian Splatting models, crucial for commercial deployment.
- MoCHA: Denoising Caption Supervision for Motion-Text Retrieval (viability: 7): https://sciencetostartup.com/paper/mocha-denoising-caption-supervision-for-motion-text-retrieval - MoCHA enhances motion-text retrieval by canonicalizing captions to their motion-recoverable content, significantly improving accuracy and cross-dataset transfer.
- Assessment Design in the AI Era: A Method for Identifying Items Functioning Differentially for Humans and Chatbots (viability: 7): https://sciencetostartup.com/paper/assessment-design-in-the-ai-era-a-method-for-identifying-items-functioning-differentially-for-humans-and-chatbots - A statistically principled method to identify assessment items vulnerable to AI misuse by detecting differential performance between humans and chatbots.
- Learning What Can Be Picked: Active Reachability Estimation for Efficient Robotic Fruit Harvesting (viability: 7): https://sciencetostartup.com/paper/learning-what-can-be-picked-active-reachability-estimation-for-efficient-robotic-fruit-harvesting - A robotic harvesting system that uses active learning to efficiently determine fruit reachability, reducing computational load and annotation effort for deployment in unstructured orchards.
- PLACID: Privacy-preserving Large language models for Acronym Clinical Inference and Disambiguation (viability: 7): https://sciencetostartup.com/paper/placid-privacy-preserving-large-language-models-for-acronym-clinical-inference-and-disambiguation - On-device LLMs for accurate, privacy-preserving clinical acronym disambiguation, preventing critical errors.
- Prototype Fusion: A Training-Free Multi-Layer Approach to OOD Detection (viability: 7): https://sciencetostartup.com/paper/prototype-fusion-a-training-free-multi-layer-approach-to-ood-detection - A training-free method for robust out-of-distribution detection by leveraging multi-layer neural network activations.
- Grounding Vision and Language to 3D Masks for Long-Horizon Box Rearrangement (viability: 7): https://sciencetostartup.com/paper/grounding-vision-and-language-to-3d-masks-for-long-horizon-box-rearrangement - A novel 3D vision-language planning system that uses mask prediction for precise, multi-step object rearrangement in complex environments.
- Bio-Inspired Event-Based Visual Servoing for Ground Robots (viability: 7): https://sciencetostartup.com/paper/bio-inspired-event-based-visual-servoing-for-ground-robots - A bio-inspired visual servoing system for ground robots that uses event-based sensing to achieve extreme low-latency and computational efficiency for state estimation.
- n-VM: A Multi-VM Layer-1 Architecture with Shared Identity and Token State (viability: 3): https://sciencetostartup.com/paper/n-vm-a-multi-vm-layer-1-architecture-with-shared-identity-and-token-state - A novel Layer-1 blockchain architecture enabling interoperability between multiple virtual machines with shared identity and token state.
- Estimating Individual Tree Height and Species from UAV Imagery (viability: 7): https://sciencetostartup.com/paper/estimating-individual-tree-height-and-species-from-uav-imagery - A novel computer vision system for accurate individual tree height and species estimation from UAV imagery, offering a cost-effective solution for forest biomass assessment.
- Energy Efficient Software Hardware CoDesign for Machine Learning: From TinyML to Large Language Models (viability: 3): https://sciencetostartup.com/paper/energy-efficient-software-hardware-codesign-for-machine-learning-from-tinyml-to-large-language-models - A framework for building energy-aware ML systems by mapping optimization strategies to computational roles.
- Echoes: A semantically-aligned music deepfake detection dataset (viability: 5): https://sciencetostartup.com/paper/echoes-a-semantically-aligned-music-deepfake-detection-dataset - A challenging dataset for training robust music deepfake detectors that generalize across diverse AI music generation systems.
- Quadrature Oscillation System for Coordinated Motion in Crawling Origami Robot (viability: 4): https://sciencetostartup.com/paper/quadrature-oscillation-system-for-coordinated-motion-in-crawling-origami-robot - Develops an electronics-free quadrature oscillation system to enable coordinated motion in origami robots for extreme environments.
- GTO Wizard Benchmark (viability: 7): https://sciencetostartup.com/paper/gto-wizard-benchmark - A standardized benchmark and API to evaluate AI agents in complex poker games, revealing significant gaps in current LLM reasoning capabilities.
- Probing Ethical Framework Representations in Large Language Models: Structure, Entanglement, and Methodological Challenges (viability: 4): https://sciencetostartup.com/paper/probing-ethical-framework-representations-in-large-language-models-structure-entanglement-and-methodological-challenges - This research probes how large language models internally represent and distinguish between different ethical frameworks, revealing structural insights into their ethical reasoning.
- Boost Like a (Var)Pro: Trust-Region Gradient Boosting via Variable Projection (viability: 5): https://sciencetostartup.com/paper/boost-like-a-var-pro-trust-region-gradient-boosting-via-variable-projection - A novel gradient boosting algorithm for smooth parametric models that offers improved convergence and performance over existing methods.
- Ethio-ASR: Joint Multilingual Speech Recognition and Language Identification for Ethiopian Languages (viability: 7): https://sciencetostartup.com/paper/ethio-asr-joint-multilingual-speech-recognition-and-language-identification-for-ethiopian-languages - A multilingual ASR system for underserved Ethiopian languages that significantly outperforms existing baselines with fewer parameters.
- Foundation Model Embeddings Meet Blended Emotions: A Multimodal Fusion Approach for the BLEMORE Challenge (viability: 4): https://sciencetostartup.com/paper/foundation-model-embeddings-meet-blended-emotions-a-multimodal-fusion-approach-for-the-blemore-challenge - A multimodal system for blended emotion recognition that leverages late fusion of specialized encoders, including a novel application of Gemini Embedding 2.0 for competitive accuracy with short video inputs.
- λSplit: Self-Supervised Content-Aware Spectral Unmixing for Fluorescence Microscopy (viability: 7): https://sciencetostartup.com/paper/split-self-supervised-content-aware-spectral-unmixing-for-fluorescence-microscopy - A physics-informed deep generative model for self-supervised spectral unmixing in fluorescence microscopy, enabling immediate adoption with standard hardware.
- Swiss-Bench SBP-002: A Frontier Model Comparison on Swiss Legal and Regulatory Tasks (viability: 4): https://sciencetostartup.com/paper/swiss-bench-sbp-002-a-frontier-model-comparison-on-swiss-legal-and-regulatory-tasks - A new benchmark evaluates frontier LLMs on Swiss regulatory compliance, revealing significant performance gaps and the potential of open-weight models.
- LLM Inference at the Edge: Mobile, NPU, and GPU Performance Efficiency Trade-offs Under Sustained Load (viability: 4): https://sciencetostartup.com/paper/llm-inference-at-the-edge-mobile-npu-and-gpu-performance-efficiency-trade-offs-under-sustained-load - Benchmarking LLM inference performance and efficiency across diverse edge devices to identify optimal deployment strategies under sustained load.
- Can LLM Agents Be CFOs? A Benchmark for Resource Allocation in Dynamic Enterprise Environments (viability: 7): https://sciencetostartup.com/paper/can-llm-agents-be-cfos-a-benchmark-for-resource-allocation-in-dynamic-enterprise-environments - A benchmark and evaluation framework for LLM agents to perform long-horizon resource allocation in dynamic enterprise environments, identifying a critical capability gap.
- Stochastic Ray Tracing for the Reconstruction of 3D Gaussian Splatting (viability: 4): https://sciencetostartup.com/paper/stochastic-ray-tracing-for-the-reconstruction-of-3d-gaussian-splatting - A novel stochastic ray tracing method for faster and more accurate 3D Gaussian splatting reconstruction and rendering.
- Steering Code LLMs with Activation Directions for Language and Library Control (viability: 4): https://sciencetostartup.com/paper/steering-code-llms-with-activation-directions-for-language-and-library-control - A method to steer code LLMs towards specific languages and libraries at inference time by manipulating activation directions.
- Ukrainian Visual Word Sense Disambiguation Benchmark (viability: 4): https://sciencetostartup.com/paper/ukrainian-visual-word-sense-disambiguation-benchmark - A new benchmark for Ukrainian Visual Word Sense Disambiguation to evaluate and improve multimodal language models.
- A Theory of LLM Information Susceptibility (viability: 3): https://sciencetostartup.com/paper/a-theory-of-llm-information-susceptibility - A theoretical framework using statistical physics to understand the limits of LLM-driven agentic self-improvement and guide the design of nested architectures for open-ended progress.
- Evaluating a Multi-Agent Voice-Enabled Smart Speaker for Care Homes: A Safety-Focused Framework (viability: 5): https://sciencetostartup.com/paper/evaluating-a-multi-agent-voice-enabled-smart-speaker-for-care-homes-a-safety-focused-framework - A safety-focused voice-enabled smart speaker framework for care homes, improving resident care and administrative efficiency with high accuracy in critical tasks.
- Revisiting Real-Time Digging-In Effects: No Evidence from NP/Z Garden-Paths (viability: 3): https://sciencetostartup.com/paper/revisiting-real-time-digging-in-effects-no-evidence-from-np-z-garden-paths - This research investigates sentence processing phenomena, finding no evidence for real-time digging-in effects in human language comprehension and aligning with neural model predictions.
- M3T: Discrete Multi-Modal Motion Tokens for Sign Language Production (viability: 7): https://sciencetostartup.com/paper/m3t-discrete-multi-modal-motion-tokens-for-sign-language-production - A novel approach to sign language production that generates realistic non-manual features, significantly improving accuracy on challenging benchmarks.
- LLMLOOP: Improving LLM-Generated Code and Tests through Automated Iterative Feedback Loops (viability: 7): https://sciencetostartup.com/paper/llmloop-improving-llm-generated-code-and-tests-through-automated-iterative-feedback-loops - Automate the refinement of LLM-generated code and tests to reduce developer effort and improve output quality.
- LLMORPH: Automated Metamorphic Testing of Large Language Models (viability: 7): https://sciencetostartup.com/paper/llmorph-automated-metamorphic-testing-of-large-language-models - LLMORPH automates LLM testing by generating new test cases from existing ones, uncovering model inconsistencies without human labels.
- Environment Maps: Structured Environmental Representations for Long-Horizon Agents (viability: 7): https://sciencetostartup.com/paper/environment-maps-structured-environmental-representations-for-long-horizon-agents - A novel structured representation for LLM agents that significantly improves success rates in complex, long-horizon software workflows by mitigating cascading errors and environmental stochasticity.
- LongTail Driving Scenarios with Reasoning Traces: The KITScenes LongTail Dataset (viability: 7): https://sciencetostartup.com/paper/longtail-driving-scenarios-with-reasoning-traces-the-kitscenes-longtail-dataset - A new multimodal dataset with reasoning traces to enable few-shot generalization for long-tail driving scenarios in autonomous vehicles.
- WorldCache: Content-Aware Caching for Accelerated Video World Models (viability: 7): https://sciencetostartup.com/paper/worldcache-content-aware-caching-for-accelerated-video-world-models - WorldCache accelerates video generation by intelligently reusing intermediate model features, achieving significant speedups with minimal quality loss.
- VideoDetective: Clue Hunting via both Extrinsic Query and Intrinsic Relevance for Long Video Understanding (viability: 7): https://sciencetostartup.com/paper/videodetective-clue-hunting-via-both-extrinsic-query-and-intrinsic-relevance-for-long-video-understanding - A framework for precise clue hunting in long videos by integrating query relevance with intrinsic video structure, improving MLLM performance.
- End-to-End Training for Unified Tokenization and Latent Denoising (viability: 7): https://sciencetostartup.com/paper/end-to-end-training-for-unified-tokenization-and-latent-denoising - A unified training approach for latent diffusion models that simplifies the training pipeline and improves performance across modalities.
- UniMotion: A Unified Framework for Motion-Text-Vision Understanding and Generation (viability: 4): https://sciencetostartup.com/paper/unimotion-a-unified-framework-for-motion-text-vision-understanding-and-generation - A unified framework for understanding and generating human motion, text, and images by treating motion as a continuous modality.
- ThinkJEPA: Empowering Latent World Models with Large Vision-Language Reasoning Model (viability: 7): https://sciencetostartup.com/paper/thinkjepa-empowering-latent-world-models-with-large-vision-language-reasoning-model - A VLM-guided latent world model framework that combines dense-frame dynamics with long-horizon semantic guidance for improved trajectory prediction.
- DualCoT-VLA: Visual-Linguistic Chain of Thought via Parallel Reasoning for Vision-Language-Action Models (viability: 7): https://sciencetostartup.com/paper/dualcot-vla-visual-linguistic-chain-of-thought-via-parallel-reasoning-for-vision-language-action-models - DualCoT-VLA enhances robotic action planning by enabling parallel visual and linguistic reasoning for complex multi-step tasks, achieving state-of-the-art performance.
- 3D-Layout-R1: Structured Reasoning for Language-Instructed Spatial Editing (viability: 7): https://sciencetostartup.com/paper/3d-layout-r1-structured-reasoning-for-language-instructed-spatial-editing - A framework for precise, text-instructed 3D scene editing using scene-graph reasoning to improve spatial understanding and layout consistency.
- The Dual Mechanisms of Spatial Reasoning in Vision-Language Models (viability: 3): https://sciencetostartup.com/paper/the-dual-mechanisms-of-spatial-reasoning-in-vision-language-models - This research clarifies how vision-language models process spatial relationships, identifying two key mechanisms within their architecture.
- Scaling DoRA: High-Rank Adaptation via Factored Norms and Fused Kernels (viability: 7): https://sciencetostartup.com/paper/scaling-dora-high-rank-adaptation-via-factored-norms-and-fused-kernels - Accelerate and reduce memory usage for high-rank LoRA adaptations in large language models with optimized kernels and factored norm computations.
- Repurposing Geometric Foundation Models for Multi-view Diffusion (viability: 7): https://sciencetostartup.com/paper/repurposing-geometric-foundation-models-for-multi-view-diffusion - Repurpose geometric foundation models to accelerate and improve multi-view image generation for novel view synthesis.
- Decoupling Exploration and Policy Optimization: Uncertainty Guided Tree Search for Hard Exploration (viability: 3): https://sciencetostartup.com/paper/decoupling-exploration-and-policy-optimization-uncertainty-guided-tree-search-for-hard-exploration - Develop an Uncertainty Guided Tree Search tool for enhancing exploration in reinforcement learning tasks.
- DUO-VSR: Dual-Stream Distillation for One-Step Video Super-Resolution (viability: 7): https://sciencetostartup.com/paper/duo-vsr-dual-stream-distillation-for-one-step-video-super-resolution - A novel three-stage framework for one-step video super-resolution that significantly improves visual quality and efficiency by unifying distribution matching and adversarial supervision.
- GenOpticalFlow: A Generative Approach to Unsupervised Optical Flow Learning (viability: 7): https://sciencetostartup.com/paper/genopticalflow-a-generative-approach-to-unsupervised-optical-flow-learning - A generative framework for unsupervised optical flow learning that synthesizes perfectly aligned frame-flow data, eliminating the need for human annotations and achieving state-of-the-art results.
- TiCo: Time-Controllable Training for Spoken Dialogue Models (viability: 5): https://sciencetostartup.com/paper/tico-time-controllable-training-for-spoken-dialogue-models - A post-training method to enable spoken dialogue models to control response duration for improved voice assistant interaction.
- UniDex: A Robot Foundation Suite for Universal Dexterous Hand Control from Egocentric Human Videos (viability: 7): https://sciencetostartup.com/paper/unidex-a-robot-foundation-suite-for-universal-dexterous-hand-control-from-egocentric-human-videos - A foundation suite for universal dexterous hand control using egocentric human videos, enabling cross-hand generalization and reducing reliance on costly robot demonstrations.
- DexDrummer: In-Hand, Contact-Rich, and Long-Horizon Dexterous Robot Drumming (viability: 7): https://sciencetostartup.com/paper/dexdrummer-in-hand-contact-rich-and-long-horizon-dexterous-robot-drumming - A hierarchical policy for dexterous robotic drumming, trained in simulation with sim-to-real transfer, demonstrating complex in-hand, contact-rich, and long-horizon manipulation.
- Greater accessibility can amplify discrimination in generative AI (viability: 5): https://sciencetostartup.com/paper/greater-accessibility-can-amplify-discrimination-in-generative-ai - This research reveals gender discrimination in voice-enabled LLMs and proposes pitch manipulation as a mitigation strategy to ensure equitable AI accessibility.
- EgoGroups: A Benchmark For Detecting Social Groups of People in the Wild (viability: 7): https://sciencetostartup.com/paper/egogroups-a-benchmark-for-detecting-social-groups-of-people-in-the-wild - A new dataset and benchmark for detecting social groups in real-world, first-person video, enabling more socially intelligent AI agents.
- Confidence-Based Decoding is Provably Efficient for Diffusion Language Models (viability: 3): https://sciencetostartup.com/paper/confidence-based-decoding-is-provably-efficient-for-diffusion-language-models - This paper provides a theoretical analysis of confidence-based decoding strategies for diffusion language models to improve sampling efficiency.
- MemDLM: Memory-Enhanced DLM Training (viability: 7): https://sciencetostartup.com/paper/memdlm-memory-enhanced-dlm-training - MemDLM enhances Diffusion Language Models by embedding a simulated denoising process into training, leading to faster convergence, lower loss, and emergent in-weight retrieval capabilities for improved long-context understanding.
- ShapDBM: Exploring Decision Boundary Maps in Shapley Space (viability: 4): https://sciencetostartup.com/paper/shapdbm-exploring-decision-boundary-maps-in-shapley-space - A novel method for visualizing machine learning decision boundaries by transforming data into Shapley space, leading to more compact and interpretable decision zones.
- One Model, Two Markets: Bid-Aware Generative Recommendation (viability: 7): https://sciencetostartup.com/paper/one-model-two-markets-bid-aware-generative-recommendation - A generative recommendation system that optimizes for both user engagement and ad revenue by integrating bid awareness directly into the generation process.
- Riverine Land Cover Mapping through Semantic Segmentation of Multispectral Point Clouds (viability: 7): https://sciencetostartup.com/paper/riverine-land-cover-mapping-through-semantic-segmentation-of-multispectral-point-clouds - Leveraging transformer-based semantic segmentation of multispectral point clouds for precise riverine land cover mapping and environmental monitoring.
- Benchmarking Deep Learning Models for Aerial LiDAR Point Cloud Semantic Segmentation under Real Acquisition Conditions: A Case Study in Navarre (viability: 7): https://sciencetostartup.com/paper/benchmarking-deep-learning-models-for-aerial-lidar-point-cloud-semantic-segmentation-under-real-acquisition-conditions-a - Benchmarking state-of-the-art deep learning models for aerial LiDAR point cloud semantic segmentation to identify optimal solutions for real-world acquisition conditions.
- SpatialReward: Verifiable Spatial Reward Modeling for Fine-Grained Spatial Consistency in Text-to-Image Generation (viability: 7): https://sciencetostartup.com/paper/spatialreward-verifiable-spatial-reward-modeling-for-fine-grained-spatial-consistency-in-text-to-image-generation - A verifiable reward model that enforces fine-grained spatial consistency in text-to-image generation, improving accuracy and controllability.
- Dyadic: A Scalable Platform for Human-Human and Human-AI Conversation Research (viability: 7): https://sciencetostartup.com/paper/dyadic-a-scalable-platform-for-human-human-and-human-ai-conversation-research - Dyadic is a no-code web platform enabling scalable research into human-human and human-AI conversations with multi-modal support and real-time monitoring.
- Adapting Self-Supervised Speech Representations for Cross-lingual Dysarthria Detection in Parkinson's Disease (viability: 5): https://sciencetostartup.com/paper/adapting-self-supervised-speech-representations-for-cross-lingual-dysarthria-detection-in-parkinson-s-disease - Adapting self-supervised speech models to detect speech impairments across languages for Parkinson's disease patients.
- Noise Titration: Exact Distributional Benchmarking for Probabilistic Time Series Forecasting (viability: 7): https://sciencetostartup.com/paper/noise-titration-exact-distributional-benchmarking-for-probabilistic-time-series-forecasting - A novel benchmarking framework for time series forecasting that rigorously evaluates model robustness to non-stationarity and noise, outperforming foundation models with a specialized probabilistic generative architecture.
- Gumbel Distillation for Parallel Text Generation (viability: 7): https://sciencetostartup.com/paper/gumbel-distillation-for-parallel-text-generation - A novel distillation technique that significantly improves the generation quality of parallel language models, enabling faster and more efficient text generation.
- Evaluating the Reliability and Fidelity of Automated Judgment Systems of Large Language Models (viability: 7): https://sciencetostartup.com/paper/evaluating-the-reliability-and-fidelity-of-automated-judgment-systems-of-large-language-models - Automate LLM quality and security assessments using LLMs as judges, achieving high correlation with human evaluations.
- SPA: A Simple but Tough-to-Beat Baseline for Knowledge Injection (viability: 7): https://sciencetostartup.com/paper/spa-a-simple-but-tough-to-beat-baseline-for-knowledge-injection - A simple, prompt-engineered data augmentation method that significantly improves LLM knowledge in specialized domains, outperforming complex baselines.
- Omni-WorldBench: Towards a Comprehensive Interaction-Centric Evaluation for World Models (viability: 7): https://sciencetostartup.com/paper/omni-worldbench-towards-a-comprehensive-interaction-centric-evaluation-for-world-models - A new benchmark and evaluation framework to assess the interactive response capabilities of 4D world models, addressing a critical gap in current AI research.
- Identification of physiological shock in intensive care units via Bayesian regime switching models (viability: 5): https://sciencetostartup.com/paper/identification-of-physiological-shock-in-intensive-care-units-via-bayesian-regime-switching-models - A Bayesian regime switching model for early detection of internal bleeding in ICU patients using vital signs and lab trends.
- Chimera: Latency- and Performance-Aware Multi-agent Serving for Heterogeneous LLMs (viability: 6): https://sciencetostartup.com/paper/chimera-latency-and-performance-aware-multi-agent-serving-for-heterogeneous-llms - A predictive scheduling system for heterogeneous LLM clusters that optimizes end-to-end latency and task performance in multi-agent workflows.
- Make Tracking Easy: Neural Motion Retargeting for Humanoid Whole-body Control (viability: 7): https://sciencetostartup.com/paper/make-tracking-easy-neural-motion-retargeting-for-humanoid-whole-body-control - A neural framework that retargets human motion to humanoid robots, eliminating physical artifacts and accelerating control policy training.
- Mixture of Mini Experts: Overcoming the Linear Layer Bottleneck in Multiple Instance Learning (viability: 7): https://sciencetostartup.com/paper/mixture-of-mini-experts-overcoming-the-linear-layer-bottleneck-in-multiple-instance-learning - A parameter-efficient module that significantly boosts the performance of computational pathology AI models by optimizing the transformation of patch features.
- PAM: A Pose-Appearance-Motion Engine for Sim-to-Real HOI Video Generation (viability: 7): https://sciencetostartup.com/paper/pam-a-pose-appearance-motion-engine-for-sim-to-real-hoi-video-generation - PAM is a unified engine for generating realistic hand-object interaction videos, improving existing methods and enabling sim-to-real applications.
- Framework for Risk-Based IoT Cybersecurity Audit Engagements (viability: 2): https://sciencetostartup.com/paper/framework-for-risk-based-iot-cybersecurity-audit-engagements - A framework for auditing the cybersecurity risks of Internet of Things devices in corporate environments.
- A Backbone Benchmarking Study on Self-supervised Learning as a Auxiliary Task with Texture-based Local Descriptors for Face Analysis (viability: 4): https://sciencetostartup.com/paper/a-backbone-benchmarking-study-on-self-supervised-learning-as-a-auxiliary-task-with-texture-based-local-descriptors-for-f - This research benchmarks self-supervised learning with texture descriptors for face analysis, finding backbone performance is task-dependent.
- Seeing is Improving: Visual Feedback for Iterative Text Layout Refinement (viability: 7): https://sciencetostartup.com/paper/seeing-is-improving-visual-feedback-for-iterative-text-layout-refinement - A framework that uses visual feedback to iteratively refine text layouts, ensuring aesthetic and readable designs.
- Enhancing Document-Level Machine Translation via Filtered Synthetic Corpora and Two-Stage LLM Adaptation (viability: 4): https://sciencetostartup.com/paper/enhancing-document-level-machine-translation-via-filtered-synthetic-corpora-and-two-stage-llm-adaptation - Leveraging LLMs for document-level machine translation by creating and filtering synthetic parallel data, overcoming data scarcity and generation issues.
- Revisiting Quantum Code Generation: Where Should Domain Knowledge Live? (viability: 7): https://sciencetostartup.com/paper/revisiting-quantum-code-generation-where-should-domain-knowledge-live - Leverage advanced LLMs with RAG and agentic feedback to significantly improve quantum code generation, outperforming specialized models without costly fine-tuning.
- Cross-Modal Reinforcement Learning for Navigation with Degraded Depth Measurements (viability: 7): https://sciencetostartup.com/paper/cross-modal-reinforcement-learning-for-navigation-with-degraded-depth-measurements - A cross-modal learning framework enhances robot navigation by inferring depth from grayscale images when depth sensors fail, ensuring robust performance in challenging environments.
- MARCUS: An agentic, multimodal vision-language model for cardiac diagnosis and management (viability: 6): https://sciencetostartup.com/paper/marcus-an-agentic-multimodal-vision-language-model-for-cardiac-diagnosis-and-management - Develop a multimodal AI tool for cardiac diagnosis, leveraging state-of-the-art vision-language models.
- Feasibility of Augmented Reality-Guided Robotic Ultrasound with Cone-Beam CT Integration for Spine Procedures (viability: 7): https://sciencetostartup.com/paper/feasibility-of-augmented-reality-guided-robotic-ultrasound-with-cone-beam-ct-integration-for-spine-procedures - An augmented reality system that guides robotic ultrasound for spine procedures, improving accuracy and efficiency.
- Closed-Loop Verbal Reinforcement Learning for Task-Level Robotic Planning (viability: 7): https://sciencetostartup.com/paper/closed-loop-verbal-reinforcement-learning-for-task-level-robotic-planning - A closed-loop verbal reinforcement learning framework for interpretable and adaptive task-level robotic planning, leveraging LLMs and VLM for symbolic policy refinement.
- Calibeating Made Simple (viability: 3): https://sciencetostartup.com/paper/calibeating-made-simple - A theoretical framework for improving external forecasts through online post-processing to minimize cumulative losses.
- ACPO: Counteracting Likelihood Displacement in Vision-Language Alignment with Asymmetric Constraints (viability: 7): https://sciencetostartup.com/paper/acpo-counteracting-likelihood-displacement-in-vision-language-alignment-with-asymmetric-constraints - ACPO is a novel alignment mechanism for vision-language models that prevents hallucinations by asymmetrically constraining preference optimization, leading to improved performance on benchmark tasks.
- Causal Evidence that Language Models use Confidence to Drive Behavior (viability: 7): https://sciencetostartup.com/paper/causal-evidence-that-language-models-use-confidence-to-drive-behavior - This research demonstrates that LLMs can be trained to actively use internal confidence signals to regulate their behavior, paving the way for more reliable and autonomous AI agents.
- Data Curation for Machine Learning Interatomic Potentials by Determinantal Point Processes (viability: 4): https://sciencetostartup.com/paper/data-curation-for-machine-learning-interatomic-potentials-by-determinantal-point-processes - Leveraging determinantal point processes to intelligently curate training data for machine learning interatomic potentials, reducing computational costs and improving model accuracy.
- Multimodal Survival Analysis with Locally Deployable Large Language Models (viability: 5): https://sciencetostartup.com/paper/multimodal-survival-analysis-with-locally-deployable-large-language-models - A multimodal survival analysis model that generates evidence-based prognoses using locally deployable LLMs, addressing privacy and computational constraints in healthcare.
- RAMPAGE: RAndomized Mid-Point for debiAsed Gradient Extrapolation (viability: 3): https://sciencetostartup.com/paper/rampage-randomized-mid-point-for-debiased-gradient-extrapolation - A new randomized method for unbiased gradient extrapolation in variational inequalities.
- dynActivation: A Trainable Activation Family for Adaptive Nonlinearity (viability: 3): https://sciencetostartup.com/paper/dynactivation-a-trainable-activation-family-for-adaptive-nonlinearity - A novel trainable activation function that interpolates between non-linearity and linearity to improve training efficiency and robustness in deep neural networks.
- Beyond Matching to Tiles: Bridging Unaligned Aerial and Satellite Views for Vision-Only UAV Navigation (viability: 7): https://sciencetostartup.com/paper/beyond-matching-to-tiles-bridging-unaligned-aerial-and-satellite-views-for-vision-only-uav-navigation - A vision-only UAV navigation system that accurately predicts location and heading from aerial and satellite imagery, enabling GNSS-denied operation.
- More Isn't Always Better: Balancing Decision Accuracy and Conformity Pressures in Multi-AI Advice (viability: 4): https://sciencetostartup.com/paper/more-isn-t-always-better-balancing-decision-accuracy-and-conformity-pressures-in-multi-ai-advice - Design systems that present multi-AI advice to improve human decision-making without increasing conformity pressure.
- OpenEarth-Agent: From Tool Calling to Tool Creation for Open-Environment Earth Observation (viability: 7): https://sciencetostartup.com/paper/openearth-agent-from-tool-calling-to-tool-creation-for-open-environment-earth-observation - An agent framework that creates its own tools for Earth Observation tasks, outperforming specialized agents with fewer resources.
- The Semantic Ladder: A Framework for Progressive Formalization of Natural Language Content for Knowledge Graphs and AI Systems (viability: 3): https://sciencetostartup.com/paper/the-semantic-ladder-a-framework-for-progressive-formalization-of-natural-language-content-for-knowledge-graphs-and-ai-sy - A framework for progressively formalizing natural language into machine-actionable knowledge graphs.
- Computationally lightweight classifiers with frequentist bounds on predictions (viability: 5): https://sciencetostartup.com/paper/computationally-lightweight-classifiers-with-frequentist-bounds-on-predictions - A computationally efficient classifier for safety-critical applications that provides actionable uncertainty bounds, suitable for real-time medical monitoring.
- ROBOGATE: Adaptive Failure Discovery for Safe Robot Policy Deployment via Two-Stage Boundary-Focused Sampling (viability: 7): https://sciencetostartup.com/paper/robogate-adaptive-failure-discovery-for-safe-robot-policy-deployment-via-two-stage-boundary-focused-sampling - A framework for adaptive failure discovery in robot policy deployment, enabling efficient risk management and safe industrial integration.
- DA-VAE: Plug-in Latent Compression for Diffusion via Detail Alignment (viability: 5): https://sciencetostartup.com/paper/da-vae-plug-in-latent-compression-for-diffusion-via-detail-alignment - A novel VAE adaptation technique that enables higher resolution image generation for diffusion models with significantly reduced token counts and faster inference.
- Biophysics-Enhanced Neural Representations for Patient-Specific Respiratory Motion Modeling (viability: 7): https://sciencetostartup.com/paper/biophysics-enhanced-neural-representations-for-patient-specific-respiratory-motion-modeling - Develop patient-specific respiratory motion models using physics-regularized implicit neural representations for improved radiotherapy precision.
- Mamba-VMR: Multimodal Query Augmentation via Generated Videos for Precise Temporal Grounding (viability: 7): https://sciencetostartup.com/paper/mamba-vmr-multimodal-query-augmentation-via-generated-videos-for-precise-temporal-grounding - Generate short videos from text queries to precisely ground moments in long videos, improving retrieval accuracy and efficiency.
- StreamingClaw Technical Report (viability: 5): https://sciencetostartup.com/paper/streamingclaw-technical-report - A unified agent framework for real-time streaming video understanding and embodied intelligence with long-term multimodal memory and proactive interaction.
- Programming Manufacturing Robots with Imperfect AI: LLMs as Tuning Experts for FDM Print Configuration Selection (viability: 5): https://sciencetostartup.com/paper/programming-manufacturing-robots-with-imperfect-ai-llms-as-tuning-experts-for-fdm-print-configuration-selection - Leveraging LLMs as tuning experts within a Bayesian optimization loop to significantly improve FDM 3D print configuration selection and reduce failure rates.
- On the Direction of RLVR Updates for LLM Reasoning: Identification and Exploitation (viability: 7): https://sciencetostartup.com/paper/on-the-direction-of-rlvr-updates-for-llm-reasoning-identification-and-exploitation - This research proposes a novel method to analyze and improve LLM reasoning by focusing on the direction of reinforcement learning updates, enabling test-time extrapolation and more efficient training.
- TALUS: Threshold ML-DSA with One-Round Online Signing via Boundary Clearance and Carry Elimination (viability: 4): https://sciencetostartup.com/paper/talus-threshold-ml-dsa-with-one-round-online-signing-via-boundary-clearance-and-carry-elimination - A novel threshold ML-DSA construction enabling one-round online signing with high success rates, overcoming theoretical limitations in cryptographic schemes.
- Multiperspectivity as a Resource for Narrative Similarity Prediction (viability: 7): https://sciencetostartup.com/paper/multiperspectivity-as-a-resource-for-narrative-similarity-prediction - Leveraging diverse LLM personas to predict narrative similarity by embracing interpretive plurality, outperforming single-ground-truth approaches.
- FreeArtGS: Articulated Gaussian Splatting Under Free-moving Scenario (viability: 7): https://sciencetostartup.com/paper/freeartgs-articulated-gaussian-splatting-under-free-moving-scenario - Develop a real-time 3D motion capture tool that uses Articulated Gaussian Splatting for free-moving scenarios.
- SpecTM: Spectral Targeted Masking for Trustworthy Foundation Models (viability: 5): https://sciencetostartup.com/paper/spectm-spectral-targeted-masking-for-trustworthy-foundation-models - A physics-informed masking technique for Earth observation foundation models that significantly improves predictive accuracy and label efficiency.
- GSEM: Graph-based Self-Evolving Memory for Experience Augmented Clinical Reasoning (viability: 7): https://sciencetostartup.com/paper/gsem-graph-based-self-evolving-memory-for-experience-augmented-clinical-reasoning - A graph-based memory system that enhances clinical reasoning agents by organizing and reusing past experiences, significantly improving accuracy.
- Principled Steering via Null-space Projection for Jailbreak Defense in Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/principled-steering-via-null-space-projection-for-jailbreak-defense-in-vision-language-models - A principled defense framework for vision-language models that enhances safety against jailbreak attacks without degrading performance on benign inputs.
- P-Flow: Prompting Visual Effects Generation (viability: 7): https://sciencetostartup.com/paper/p-flow-prompting-visual-effects-generation - P-Flow simplifies visual effects generation using cutting-edge prompting techniques.
- A Context Engineering Framework for Improving Enterprise AI Agents based on Digital-Twin MDP (viability: 7): https://sciencetostartup.com/paper/a-context-engineering-framework-for-improving-enterprise-ai-agents-based-on-digital-twin-mdp - A framework for improving enterprise AI agents using offline reinforcement learning and digital twins to overcome data limitations and enhance reasoning.
- Do World Action Models Generalize Better than VLAs? A Robustness Study (viability: 7): https://sciencetostartup.com/paper/do-world-action-models-generalize-better-than-vlas-a-robustness-study - This research compares world action models and vision-language-action models for robot control, demonstrating superior robustness of world action models in challenging scenarios and providing insights for future development.
- Autoregressive vs. Masked Diffusion Language Models: A Controlled Comparison (viability: 7): https://sciencetostartup.com/paper/autoregressive-vs-masked-diffusion-language-models-a-controlled-comparison - This research provides a controlled comparison of autoregressive and masked diffusion language models, releasing code and checkpoints to enable the development of more diverse and fluent text generation systems.
- MIHT: A Hoeffding Tree for Time Series Classification using Multiple Instance Learning (viability: 7): https://sciencetostartup.com/paper/miht-a-hoeffding-tree-for-time-series-classification-using-multiple-instance-learning - A novel, interpretable algorithm for classifying complex time series data that significantly outperforms existing state-of-the-art models.
- Adapting Point Cloud Analysis via Multimodal Bayesian Distribution Learning (viability: 7): https://sciencetostartup.com/paper/adapting-point-cloud-analysis-via-multimodal-bayesian-distribution-learning - A Bayesian framework for adapting 3D vision-language models to domain shifts using online distribution learning.
- On the Failure of Topic-Matched Contrast Baselines in Multi-Directional Refusal Abliteration (viability: 3): https://sciencetostartup.com/paper/on-the-failure-of-topic-matched-contrast-baselines-in-multi-directional-refusal-abliteration - This research investigates the failure of topic-matched contrast baselines in removing refusal behavior from language models, suggesting a need for revised methodologies in abliteration research.
- SpatialBoost: Enhancing Visual Representation through Language-Guided Reasoning (viability: 7): https://sciencetostartup.com/paper/spatialboost-enhancing-visual-representation-through-language-guided-reasoning - Enhance existing vision models with 3D spatial understanding by leveraging language-guided reasoning and LLMs.
- Dual-Space Knowledge Distillation with Key-Query Matching for Large Language Models with Vocabulary Mismatch (viability: 4): https://sciencetostartup.com/paper/dual-space-knowledge-distillation-with-key-query-matching-for-large-language-models-with-vocabulary-mismatch - A novel generative adversarial approach to improve knowledge distillation between large language models with different tokenizers, showing modest gains in text generation quality.
- MineRobot: A Unified Framework for Kinematics Modeling and Solving of Underground Mining Robots in Virtual Environments (viability: 7): https://sciencetostartup.com/paper/minerobot-a-unified-framework-for-kinematics-modeling-and-solving-of-underground-mining-robots-in-virtual-environments - A unified framework for modeling and solving the kinematics of underground mining robots in virtual environments, enabling real-time performance and robustness for training, planning, and digital-twin applications.
- FontCrafter: High-Fidelity Element-Driven Artistic Font Creation with Visual In-Context Generation (viability: 7): https://sciencetostartup.com/paper/fontcrafter-high-fidelity-element-driven-artistic-font-creation-with-visual-in-context-generation - FontCrafter enables high-fidelity artistic font creation by using visual elements as style references and offering fine-grained control over glyph shape and texture.
- AnimalCLAP: Taxonomy-Aware Language-Audio Pretraining for Species Recognition and Trait Inference (viability: 7): https://sciencetostartup.com/paper/animalclap-taxonomy-aware-language-audio-pretraining-for-species-recognition-and-trait-inference - A taxonomy-aware language-audio model and dataset for species recognition and trait inference from animal vocalizations, outperforming existing methods.
- MAGPI: Multifidelity-Augmented Gaussian Process Inputs for Surrogate Modeling from Scarce Data (viability: 7): https://sciencetostartup.com/paper/magpi-multifidelity-augmented-gaussian-process-inputs-for-surrogate-modeling-from-scarce-data - Develops a novel multifidelity approach for Gaussian Process Regression to create more accurate and cost-effective surrogate models from scarce high-fidelity data, augmented by cheaper low-fidelity data.
- Uncertainty-guided Compositional Alignment with Part-to-Whole Semantic Representativeness in Hyperbolic Vision-Language Models (viability: 5): https://sciencetostartup.com/paper/uncertainty-guided-compositional-alignment-with-part-to-whole-semantic-representativeness-in-hyperbolic-vision-language - Enhancing accuracy in hyperbolic vision-language models through uncertainty-guided alignment.
- DTVI: Dual-Stage Textual and Visual Intervention for Safe Text-to-Image Generation (viability: 7): https://sciencetostartup.com/paper/dtvi-dual-stage-textual-and-visual-intervention-for-safe-text-to-image-generation - A dual-stage defense framework for safe text-to-image generation that intervenes at both textual and visual stages to capture and attenuate unsafe content.
- RAFL: Generalizable Sim-to-Real of Soft Robots with Residual Acceleration Field Learning (viability: 4): https://sciencetostartup.com/paper/rafl-generalizable-sim-to-real-of-soft-robots-with-residual-acceleration-field-learning - A framework that improves the accuracy of soft robot simulations across different shapes by learning a transferable corrective dynamics field.
- GTSR: Subsurface Scattering Awared 3D Gaussians for Translucent Surface Reconstruction (viability: 7): https://sciencetostartup.com/paper/gtsr-subsurface-scattering-awared-3d-gaussians-for-translucent-surface-reconstruction - Reconstruct translucent 3D objects from images with a novel Gaussian-based pipeline that models subsurface scattering for improved detail and real-time rendering.
- Future-Interactions-Aware Trajectory Prediction via Braid Theory (viability: 7): https://sciencetostartup.com/paper/future-interactions-aware-trajectory-prediction-via-braid-theory - Leveraging braid theory for more accurate multi-agent trajectory prediction in autonomous vehicles, improving safety and reducing computational overhead.
- MEVIUS2: Practical Open-Source Quadruped Robot with Sheet Metal Welding and Multimodal Perception (viability: 7): https://sciencetostartup.com/paper/mevius2-practical-open-source-quadruped-robot-with-sheet-metal-welding-and-multimodal-perception - An open-source, large-scale, and durable quadruped robot with multimodal perception, built using readily available sheet metal welding and machining.
- On the Interplay of Priors and Overparametrization in Bayesian Neural Network Posteriors (viability: 3): https://sciencetostartup.com/paper/on-the-interplay-of-priors-and-overparametrization-in-bayesian-neural-network-posteriors - This research investigates how overparameterization and priors reshape Bayesian neural network posteriors, offering theoretical insights into their geometric properties.
- Tuning Real-World Image Restoration at Inference: A Test-Time Scaling Paradigm for Flow Matching Models (viability: 7): https://sciencetostartup.com/paper/tuning-real-world-image-restoration-at-inference-a-test-time-scaling-paradigm-for-flow-matching-models - A novel framework for real-world image restoration that leverages test-time scaling with flow matching models to achieve state-of-the-art performance.
- Do Papers Match Code? A Benchmark and Framework for Paper-Code Consistency Detection in Bioinformatics Software (viability: 7): https://sciencetostartup.com/paper/do-papers-match-code-a-benchmark-and-framework-for-paper-code-consistency-detection-in-bioinformatics-software - A framework and benchmark for automatically detecting consistency between scientific papers and their code implementations, starting with bioinformatics.
- AdditiveLLM2: A Multi-modal Large Language Model for Additive Manufacturing (viability: 7): https://sciencetostartup.com/paper/additivellm2-a-multi-modal-large-language-model-for-additive-manufacturing - A specialized multi-modal LLM for additive manufacturing, achieving over 90% accuracy on domain-specific knowledge tasks.
- ROM: Real-time Overthinking Mitigation via Streaming Detection and Intervention (viability: 7): https://sciencetostartup.com/paper/rom-real-time-overthinking-mitigation-via-streaming-detection-and-intervention - ROM is a lightweight, real-time system that mitigates overthinking in large language models, reducing response length and improving efficiency without retraining the backbone.
- Retrieving Climate Change Disinformation by Narrative (viability: 7): https://sciencetostartup.com/paper/retrieving-climate-change-disinformation-by-narrative - A framework to detect emerging climate change disinformation narratives by reframing it as a retrieval task, outperforming traditional methods on high-variance narratives.
- 6D Robotic OCT Scanning of Curved Tissue Surfaces (viability: 4): https://sciencetostartup.com/paper/6d-robotic-oct-scanning-of-curved-tissue-surfaces - Enables precise 6D robotic scanning of curved tissue surfaces for improved medical imaging by eliminating reliance on image registration.
- Asymptotically Ideal Hierarchical Secret Sharing Based on CRT for Integer Ring (viability: 2): https://sciencetostartup.com/paper/asymptotically-ideal-hierarchical-secret-sharing-based-on-crt-for-integer-ring - A theoretical cryptographic scheme for hierarchical secret sharing using the Chinese Remainder Theorem.
- On the Challenges and Opportunities of Learned Sparse Retrieval for Code (viability: 5): https://sciencetostartup.com/paper/on-the-challenges-and-opportunities-of-learned-sparse-retrieval-for-code - A new family of sparse retrieval models for codebases that aims to improve LLM-based software engineering systems.
- VP-VLA: Visual Prompting as an Interface for Vision-Language-Action Models (viability: 7): https://sciencetostartup.com/paper/vp-vla-visual-prompting-as-an-interface-for-vision-language-action-models - VP-VLA decouples high-level reasoning from low-level execution in robotics using structured visual prompts, improving spatial precision and robustness.
- SegMaFormer: A Hybrid State-Space and Transformer Model for Efficient Segmentation (viability: 7): https://sciencetostartup.com/paper/segmaformer-a-hybrid-state-space-and-transformer-model-for-efficient-segmentation - A lightweight hybrid Mamba-Transformer model for efficient and high-performing 3D medical image segmentation, reducing parameters and FLOPs significantly.
- Asymptotically Ideal Conjunctive Hierarchical Secret Sharing Scheme Based on CRT for Polynomial Ring (viability: 2): https://sciencetostartup.com/paper/asymptotically-ideal-conjunctive-hierarchical-secret-sharing-scheme-based-on-crt-for-polynomial-ring - A theoretical cryptographic scheme for secure secret sharing with improved information rates.
- CRPS-Optimal Binning for Conformal Regression (viability: 7): https://sciencetostartup.com/paper/crps-optimal-binning-for-conformal-regression - A novel method for non-parametric conditional distribution estimation that provides narrower prediction intervals with guaranteed coverage, outperforming existing conformal methods on benchmarks.
- STENet: Superpixel Token Enhancing Network for RGB-D Salient Object Detection (viability: 7): https://sciencetostartup.com/paper/stenet-superpixel-token-enhancing-network-for-rgb-d-salient-object-detection - A novel network for RGB-D salient object detection that uses superpixels to enhance global and local feature extraction, outperforming state-of-the-art methods.
- λ-GELU: Learning Gating Hardness for Controlled ReLU-ization in Deep Networks (viability: 4): https://sciencetostartup.com/paper/-gelu-learning-gating-hardness-for-controlled-relu-ization-in-deep-networks - A novel activation function that bridges smooth neural network training with ReLU-compatible deployment pipelines by learning a controllable 'hardness' parameter.
- TREX: Trajectory Explanations for Multi-Objective Reinforcement Learning (viability: 4): https://sciencetostartup.com/paper/trex-trajectory-explanations-for-multi-objective-reinforcement-learning - A framework for explaining the decision-making process of multi-objective reinforcement learning agents by attributing trajectory segments to specific objectives.
- LRC-WeatherNet: LiDAR, RADAR, and Camera Fusion Network for Real-time Weather-type Classification in Autonomous Driving (viability: 6): https://sciencetostartup.com/paper/lrc-weathernet-lidar-radar-and-camera-fusion-network-for-real-time-weather-type-classification-in-autonomous-driving - Real-time weather classification using a fusion of LiDAR, RADAR, and camera data for autonomous vehicles.
- Speed by Simplicity: A Single-Stream Architecture for Fast Audio-Video Generative Foundation Model (viability: 6): https://sciencetostartup.com/paper/speed-by-simplicity-a-single-stream-architecture-for-fast-audio-video-generative-foundation-model - A streamlined architecture that speeds up audio-video generative models with state-of-the-art performance.
- GeoFusion-CAD: Structure-Aware Diffusion with Geometric State Space for Parametric 3D Design (viability: 7): https://sciencetostartup.com/paper/geofusion-cad-structure-aware-diffusion-with-geometric-state-space-for-parametric-3d-design - A diffusion model that generates complex 3D CAD designs by understanding hierarchical structure, overcoming limitations of existing Transformer-based methods for long command sequences.
- BOOST-RPF: Boosted Sequential Trees for Radial Power Flow (viability: 7): https://sciencetostartup.com/paper/boost-rpf-boosted-sequential-trees-for-radial-power-flow - A novel AI method using boosted decision trees to achieve state-of-the-art, scalable, and robust power flow analysis for distribution systems.
- SecureBreak -- A dataset towards safe and secure models (viability: 5): https://sciencetostartup.com/paper/securebreak-a-dataset-towards-safe-and-secure-models - A new safety-oriented dataset to detect and block harmful LLM outputs, improving model security and robustness.
- Demystifying Reinforcement Learning for Long-Horizon Tool-Using Agents: A Comprehensive Recipe (viability: 7): https://sciencetostartup.com/paper/demystifying-reinforcement-learning-for-long-horizon-tool-using-agents-a-comprehensive-recipe - A systematic recipe for scaling Reinforcement Learning to build advanced LLM agents for complex, long-horizon tasks.
- Parameter-Efficient Fine-Tuning for Medical Text Summarization: A Comparative Study of Lora, Prompt Tuning, and Full Fine-Tuning (viability: 7): https://sciencetostartup.com/paper/parameter-efficient-fine-tuning-for-medical-text-summarization-a-comparative-study-of-lora-prompt-tuning-and-full-fine-t - Achieve state-of-the-art medical text summarization with significantly reduced computational cost using parameter-efficient fine-tuning techniques like LoRA.
- BHDD: A Burmese Handwritten Digit Dataset (viability: 4): https://sciencetostartup.com/paper/bhdd-a-burmese-handwritten-digit-dataset - A new dataset of Burmese handwritten digits with code and benchmark results to enable research and development in optical character recognition for underrepresented scripts.
- Unified Spatiotemporal Token Compression for Video-LLMs at Ultra-Low Retention (viability: 7): https://sciencetostartup.com/paper/unified-spatiotemporal-token-compression-for-video-llms-at-ultra-low-retention - A plug-and-play module for Video-LLMs that drastically reduces computational costs by intelligently compressing visual tokens, enabling faster inference and lower memory usage without retraining.
- Group3D: MLLM-Driven Semantic Grouping for Open-Vocabulary 3D Object Detection (viability: 7): https://sciencetostartup.com/paper/group3d-mllm-driven-semantic-grouping-for-open-vocabulary-3d-object-detection - A new approach to 3D object detection using open-vocabulary models for semantic grouping in diverse environments.
- GeoFlow: Real-Time Fine-Grained Cross-View Geolocalization via Iterative Flow Prediction (viability: 7): https://sciencetostartup.com/paper/geoflow-real-time-fine-grained-cross-view-geolocalization-via-iterative-flow-prediction - A real-time, lightweight geolocalization system that iteratively refines location hypotheses for safe autonomous navigation in GPS-denied areas.
- SLURP-TN : Resource for Tunisian Dialect Spoken Language Understanding (viability: 7): https://sciencetostartup.com/paper/slurp-tn-resource-for-tunisian-dialect-spoken-language-understanding - A new dataset and baseline models for Tunisian dialect spoken language understanding to unlock task-oriented dialogue systems for under-resourced languages.
- FeatDistill: A Feature Distillation Enhanced Multi-Expert Ensemble Framework for Robust AI-generated Image Detection (viability: 7): https://sciencetostartup.com/paper/featdistill-a-feature-distillation-enhanced-multi-expert-ensemble-framework-for-robust-ai-generated-image-detection - A robust AI-generated image detection framework using feature distillation and a multi-expert ensemble to combat deepfakes in real-world conditions.
- MultiBind: A Benchmark for Attribute Misbinding in Multi-Subject Generation (viability: 7): https://sciencetostartup.com/paper/multibind-a-benchmark-for-attribute-misbinding-in-multi-subject-generation - A new benchmark and evaluation protocol to diagnose and improve attribute binding in multi-subject image generation, addressing a key failure mode in current generative models.
- Cross-Instance Gaussian Splatting Registration via Geometry-Aware Feature-Guided Alignment (viability: 5): https://sciencetostartup.com/paper/cross-instance-gaussian-splatting-registration-via-geometry-aware-feature-guided-alignment - Develop an advanced registration tool for aligning 3D models using Gaussian splatting techniques.
- Chronological Contrastive Learning: Few-Shot Progression Assessment in Irreversible Diseases (viability: 7): https://sciencetostartup.com/paper/chronological-contrastive-learning-few-shot-progression-assessment-in-irreversible-diseases - Leveraging chronological patient imaging data with contrastive learning to dramatically reduce expert annotation needs for disease severity assessment.
- Camera-Agnostic Pruning of 3D Gaussian Splats via Descriptor-Based Beta Evidence (viability: 7): https://sciencetostartup.com/paper/camera-agnostic-pruning-of-3d-gaussian-splats-via-descriptor-based-beta-evidence - A camera-agnostic method for pruning 3D Gaussian splats using descriptor-based evidence, enabling efficient storage and transmission of 3D scene data.
- SatGeo-NeRF: Geometrically Regularized NeRF for Satellite Imagery (viability: 7): https://sciencetostartup.com/paper/satgeo-nerf-geometrically-regularized-nerf-for-satellite-imagery - A geometrically regularized NeRF model that significantly reduces overfitting artifacts in satellite imagery for more accurate 3D reconstructions.
- The Golden Subspace: Where Efficiency Meets Generalization in Continual Test-Time Adaptation (viability: 7): https://sciencetostartup.com/paper/the-golden-subspace-where-efficiency-meets-generalization-in-continual-test-time-adaptation - A novel method for efficient and generalized online adaptation of AI models to changing data distributions, enabling robust performance in real-world scenarios.
- Disengagement Analysis and Field Tests of a Prototypical Open-Source Level 4 Autonomous Driving System (viability: 3): https://sciencetostartup.com/paper/disengagement-analysis-and-field-tests-of-a-prototypical-open-source-level-4-autonomous-driving-system - This research analyzes disengagements in an open-source Level 4 autonomous driving system to identify robustness issues missed by standard metrics.
- Guideline-grounded retrieval-augmented generation for ophthalmic clinical decision support (viability: 7): https://sciencetostartup.com/paper/guideline-grounded-retrieval-augmented-generation-for-ophthalmic-clinical-decision-support - A multimodal RAG system for ophthalmology that retrieves and reasons over guideline images to improve clinical decision support, outperforming existing models on challenging cases.
- Deep Reinforcement Learning and The Tale of Two Temporal Difference Errors (viability: 2): https://sciencetostartup.com/paper/deep-reinforcement-learning-and-the-tale-of-two-temporal-difference-errors - This paper theoretically analyzes the nuances of temporal difference errors in deep reinforcement learning, revealing potential performance impacts in deep differential RL methods.
- Structural Concentration in Weighted Networks: A Class of Topology-Aware Indices (viability: 3): https://sciencetostartup.com/paper/structural-concentration-in-weighted-networks-a-class-of-topology-aware-indices - A new framework for measuring concentration in weighted networks that accounts for both weight distribution and network topology.
- Collision-Free Velocity Scheduling for Multi-Agent Systems on Predefined Routes via Inexact-Projection ADMM (viability: 3): https://sciencetostartup.com/paper/collision-free-velocity-scheduling-for-multi-agent-systems-on-predefined-routes-via-inexact-projection-admm - Optimizing waypoint passage times for multi-agent systems on predefined routes to improve efficiency and avoid collisions.
- A Latent Representation Learning Framework for Hyperspectral Image Emulation in Remote Sensing (viability: 7): https://sciencetostartup.com/paper/a-latent-representation-learning-framework-for-hyperspectral-image-emulation-in-remote-sensing - A latent representation learning framework for generating synthetic hyperspectral images, enabling faster and more accurate remote sensing data simulation and analysis.
- SparseDVFS: Sparse-Aware DVFS for Energy-Efficient Edge Inference (viability: 4): https://sciencetostartup.com/paper/sparsedvfs-sparse-aware-dvfs-for-energy-efficient-edge-inference - A framework for optimizing energy efficiency in edge device inference by dynamically scaling voltage and frequency based on operator sparsity.
- SHAPE: Structure-aware Hierarchical Unsupervised Domain Adaptation with Plausibility Evaluation for Medical Image Segmentation (viability: 7): https://sciencetostartup.com/paper/shape-structure-aware-hierarchical-unsupervised-domain-adaptation-with-plausibility-evaluation-for-medical-image-segment - A medical image segmentation framework that ensures global anatomical plausibility through structure-aware adaptation and hypergraph-based validation, outperforming existing methods.
- CLEAR: Context-Aware Learning with End-to-End Mask-Free Inference for Adaptive Video Subtitle Removal (viability: 7): https://sciencetostartup.com/paper/clear-context-aware-learning-with-end-to-end-mask-free-inference-for-adaptive-video-subtitle-removal - A mask-free AI framework for adaptive video subtitle removal that achieves end-to-end inference and strong zero-shot generalization across multiple languages.
- Ara-Best-RQ: Multi Dialectal Arabic SSL (viability: 7): https://sciencetostartup.com/paper/ara-best-rq-multi-dialectal-arabic-ssl - A family of self-supervised learning models for multi-dialectal Arabic speech processing that achieves state-of-the-art performance on dialect identification.
- Albank -- a case study on the use of ethereum blockchain technology and smart contracts for secure decentralized bank application (viability: 3): https://sciencetostartup.com/paper/albank-a-case-study-on-the-use-of-ethereum-blockchain-technology-and-smart-contracts-for-secure-decentralized-bank-appli - A theoretical proposal for a decentralized banking application using Ethereum smart contracts to enhance security and transparency in traditional banking.
- IGV-RRT: Prior-Real-Time Observation Fusion for Active Object Search in Changing Environments (viability: 7): https://sciencetostartup.com/paper/igv-rrt-prior-real-time-observation-fusion-for-active-object-search-in-changing-environments - A probabilistic planning framework for real-time object search in dynamic indoor environments, leveraging scene priors and vision-language models to improve efficiency and success rates.
- ADaFuSE: Adaptive Diffusion-generated Image and Text Fusion for Interactive Text-to-Image Retrieval (viability: 7): https://sciencetostartup.com/paper/adafuse-adaptive-diffusion-generated-image-and-text-fusion-for-interactive-text-to-image-retrieval - A lightweight fusion model that significantly improves interactive text-to-image retrieval by adaptively combining multi-modal feedback, outperforming existing methods with minimal parameter increase.
- Not All Layers Are Created Equal: Adaptive LoRA Ranks for Personalized Image Generation (viability: 7): https://sciencetostartup.com/paper/not-all-layers-are-created-equal-adaptive-lora-ranks-for-personalized-image-generation - Adaptive LoRA ranks for personalized image generation that balances performance and memory by dynamically adjusting layer importance.
- Deep S2P: Integrating Learning Based Stereo Matching Into the Satellite Stereo Pipeline (viability: 7): https://sciencetostartup.com/paper/deep-s2p-integrating-learning-based-stereo-matching-into-the-satellite-stereo-pipeline - Integrate advanced learning-based stereo matching into satellite imagery pipelines to generate more accurate and detailed digital surface models.
- Optimal Solutions for the Moving Target Vehicle Routing Problem with Obstacles via Lazy Branch and Price (viability: 4): https://sciencetostartup.com/paper/optimal-solutions-for-the-moving-target-vehicle-routing-problem-with-obstacles-via-lazy-branch-and-price - An optimization algorithm for efficient agent routing in dynamic environments with moving targets and obstacles.
- SmaAT-QMix-UNet: A Parameter-Efficient Vector-Quantized UNet for Precipitation Nowcasting (viability: 7): https://sciencetostartup.com/paper/smaat-qmix-unet-a-parameter-efficient-vector-quantized-unet-for-precipitation-nowcasting - A parameter-efficient UNet model using vector quantization and mixed kernel convolutions for improved precipitation nowcasting, with publicly available code.
- P^2O: Joint Policy and Prompt Optimization (viability: 7): https://sciencetostartup.com/paper/p-2o-joint-policy-and-prompt-optimization - Optimize LLM reasoning by jointly evolving prompts and policies to efficiently learn from challenging examples.
- Thermal Topology Collapse: Universal Physical Patch Attacks on Infrared Vision Systems (viability: 7): https://sciencetostartup.com/paper/thermal-topology-collapse-universal-physical-patch-attacks-on-infrared-vision-systems - Develops a universal physical patch attack for infrared vision systems that bypasses adversarial defenses with no online computation.
- Manifold-Aware Exploration for Reinforcement Learning in Video Generation (viability: 7): https://sciencetostartup.com/paper/manifold-aware-exploration-for-reinforcement-learning-in-video-generation - A novel reinforcement learning approach for video generation that stabilizes alignment by constraining exploration to the data manifold, achieving superior quality and reward maximization.
- Adversarial Camouflage (viability: 7): https://sciencetostartup.com/paper/adversarial-camouflage - Develops an adversarial camouflage technique to protect user privacy against facial recognition systems by maximizing recognition error across multiple architectures.
- Tacit Knowledge Management with Generative AI: Proposal of the GenAI SECI Model (viability: 2): https://sciencetostartup.com/paper/tacit-knowledge-management-with-generative-ai-proposal-of-the-genai-seci-model - A new model for managing tacit and explicit knowledge using generative AI, introducing the concept of 'Digital Fragmented Knowledge'.
- Adaptive Video Distillation: Mitigating Oversaturation and Temporal Collapse in Few-Step Generation (viability: 7): https://sciencetostartup.com/paper/adaptive-video-distillation-mitigating-oversaturation-and-temporal-collapse-in-few-step-generation - A novel distillation framework for efficient and high-fidelity few-step video generation, overcoming oversaturation and temporal collapse.
- Holistic Scaling Laws for Optimal Mixture-of-Experts Architecture Optimization (viability: 3): https://sciencetostartup.com/paper/holistic-scaling-laws-for-optimal-mixture-of-experts-architecture-optimization - A framework for optimizing Mixture-of-Experts LLM architectures by establishing new scaling laws based on compute and parameter constraints.
- Climate Prompting: Generating the Madden-Julian Oscillation using Video Diffusion and Low-Dimensional Conditioning (viability: 4): https://sciencetostartup.com/paper/climate-prompting-generating-the-madden-julian-oscillation-using-video-diffusion-and-low-dimensional-conditioning - A video diffusion model generates climate simulations conditioned on key metrics to bridge theoretical frameworks and improve tropical atmosphere prediction.
- Reasoning or Rhetoric? An Empirical Analysis of Moral Reasoning Explanations in Large Language Models (viability: 3): https://sciencetostartup.com/paper/reasoning-or-rhetoric-an-empirical-analysis-of-moral-reasoning-explanations-in-large-language-models - This research empirically analyzes LLM responses to moral dilemmas to understand if they exhibit genuine moral reasoning or merely mimic it through alignment training, revealing a 'moral ventriloquism' phenomenon.
- Sim-to-Real of Humanoid Locomotion Policies via Joint Torque Space Perturbation Injection (viability: 7): https://sciencetostartup.com/paper/sim-to-real-of-humanoid-locomotion-policies-via-joint-torque-space-perturbation-injection - A novel sim-to-real method for humanoid locomotion that injects state-dependent joint torque perturbations to improve policy robustness in real-world deployments.
- All elementary functions from a single binary operator (viability: 3): https://sciencetostartup.com/paper/all-elementary-functions-from-a-single-binary-operator - A novel binary operator and constant can generate all scientific calculator functions, enabling gradient-based symbolic regression for exact function recovery from data.
- Riding Brainwaves in LLM Space: Understanding Activation Patterns Using Individual Neural Signatures (viability: 5): https://sciencetostartup.com/paper/riding-brainwaves-in-llm-space-understanding-activation-patterns-using-individual-neural-signatures - This research explores the potential for language models to encode individual neural responses, suggesting a path towards personalized brain-computer interfaces.
- Agentic Personas for Adaptive Scientific Explanations with Knowledge Graphs (viability: 7): https://sciencetostartup.com/paper/agentic-personas-for-adaptive-scientific-explanations-with-knowledge-graphs - Develop adaptive AI explanations for complex domains using agentic personas trained with significantly reduced feedback.
- On the Number of Conditional Independence Tests in Constraint-based Causal Discovery (viability: 4): https://sciencetostartup.com/paper/on-the-number-of-conditional-independence-tests-in-constraint-based-causal-discovery - A new causal discovery algorithm significantly reduces the number of conditional independence tests required, improving efficiency for learning causal relationships from data.
- Select, Label, Evaluate: Active Testing in NLP (viability: 7): https://sciencetostartup.com/paper/select-label-evaluate-active-testing-in-nlp - Reduce NLP model evaluation costs by up to 95% by intelligently selecting the most informative test samples for annotation.
- Instruction Set and Language for Symbolic Regression (viability: 2): https://sciencetostartup.com/paper/instruction-set-and-language-for-symbolic-regression - A novel representation framework for symbolic regression that encodes expression DAGs as strings and computes a pruned canonical string to collapse equivalent representations.
- Publicly Understandable Electronic Voting: A Non-Cryptographic, End-to-End Verifiable Scheme (viability: 3): https://sciencetostartup.com/paper/publicly-understandable-electronic-voting-a-non-cryptographic-end-to-end-verifiable-scheme - A non-cryptographic voting system allowing voters to verify election integrity using only basic arithmetic and physical receipts, restoring public trust in democratic processes.
- Deriving Health Metrics from the Photoplethysmogram: Benchmarks and Insights from MIMIC-III-Ext-PPG (viability: 7): https://sciencetostartup.com/paper/deriving-health-metrics-from-the-photoplethysmogram-benchmarks-and-insights-from-mimic-iii-ext-ppg - A comprehensive benchmark for PPG-based clinical prediction, establishing baselines for multi-class heart rhythm classification and physiological parameter estimation, with strong generalizability and insights into subgroup performance.
- Directional Mollification for Controlled Smooth Path Generation (viability: 3): https://sciencetostartup.com/paper/directional-mollification-for-controlled-smooth-path-generation - A novel theoretical framework for generating smooth, waypoint-interpolating paths for autonomous and industrial robots.
- Multi-View Deformable Convolution Meets Visual Mamba for Coronary Artery Segmentation (viability: 4): https://sciencetostartup.com/paper/multi-view-deformable-convolution-meets-visual-mamba-for-coronary-artery-segmentation - A novel two-stage framework for coronary artery segmentation in CTA images, combining multi-view deformable convolution with visual Mamba to improve accuracy and efficiency.
- CoRA: Boosting Time Series Foundation Models for Multivariate Forecasting through Correlation-aware Adapter (viability: 7): https://sciencetostartup.com/paper/cora-boosting-time-series-foundation-models-for-multivariate-forecasting-through-correlation-aware-adapter - A lightweight adapter that significantly boosts multivariate time series forecasting performance by capturing complex channel correlations, designed for easy integration with existing foundation models.
- BadminSense: Enabling Fine-Grained Badminton Stroke Evaluation on a Single Smartwatch (viability: 5): https://sciencetostartup.com/paper/badminsense-enabling-fine-grained-badminton-stroke-evaluation-on-a-single-smartwatch - A smartwatch system for fine-grained badminton stroke evaluation and quality prediction.
- SteelDefectX: A Coarse-to-Fine Vision-Language Dataset and Benchmark for Generalizable Steel Surface Defect Detection (viability: 7): https://sciencetostartup.com/paper/steeldefectx-a-coarse-to-fine-vision-language-dataset-and-benchmark-for-generalizable-steel-surface-defect-detection - A vision-language dataset and benchmark for generalizable steel surface defect detection, enabling explainable and transferable AI models.
- Politics of Questions in News: A Mixed-Methods Study of Interrogative Stances as Markers of Voice and Power (viability: 3): https://sciencetostartup.com/paper/politics-of-questions-in-news-a-mixed-methods-study-of-interrogative-stances-as-markers-of-voice-and-power - This paper analyzes the function and distribution of interrogative sentences in French news articles to understand how they structure discourse and foreground specific actors.
- Beyond Strict Pairing: Arbitrarily Paired Training for High-Performance Infrared and Visible Image Fusion (viability: 7): https://sciencetostartup.com/paper/beyond-strict-pairing-arbitrarily-paired-training-for-high-performance-infrared-and-visible-image-fusion - A framework for infrared and visible image fusion that drastically reduces data acquisition costs by enabling training on unaligned image pairs, achieving comparable performance to methods requiring 100x more data.
- Ctrl-A: Control-Driven Online Data Augmentation (viability: 7): https://sciencetostartup.com/paper/ctrl-a-control-driven-online-data-augmentation - Automated image data augmentation that dynamically adjusts augmentation strength using control theory, eliminating manual policy engineering for new vision tasks.
- Clinical Graph-Mediated Distillation for Unpaired MRI-to-CFI Hypertension Prediction (viability: 7): https://sciencetostartup.com/paper/clinical-graph-mediated-distillation-for-unpaired-mri-to-cfi-hypertension-prediction - A framework that transfers hypertension prediction knowledge from expensive MRI scans to low-cost retinal fundus images using a clinical similarity graph, enabling better screening with unpaired data.
- Cascade-Free Mandarin Visual Speech Recognition via Semantic-Guided Cross-Representation Alignment (viability: 7): https://sciencetostartup.com/paper/cascade-free-mandarin-visual-speech-recognition-via-semantic-guided-cross-representation-alignment - A cascade-free Mandarin visual speech recognition system that improves accuracy and efficiency by jointly aligning multiple intermediate representations.
- Anatomical Token Uncertainty for Transformer-Guided Active MRI Acquisition (viability: 7): https://sciencetostartup.com/paper/anatomical-token-uncertainty-for-transformer-guided-active-mri-acquisition - An active MRI acquisition framework using anatomical token uncertainty to significantly accelerate scans and improve image quality.
- Timing In stand-up Comedy: Text, Audio, Laughter, Kinesics (TIC-TALK): Pipeline and Database for the Multimodal Study of Comedic Timing (viability: 4): https://sciencetostartup.com/paper/timing-in-stand-up-comedy-text-audio-laughter-kinesics-tic-talk-pipeline-and-database-for-the-multimodal-study-of-comedi - A multimodal dataset and pipeline for analyzing comedic timing using language, gesture, and audience laughter.
- Connecting Distributed Ledgers: Surveying Novel Interoperability Solutions in On-chain Finance (viability: 3): https://sciencetostartup.com/paper/connecting-distributed-ledgers-surveying-novel-interoperability-solutions-in-on-chain-finance - This paper surveys novel interoperability solutions for on-chain finance, proposing metrics and models for empirical research.
- Benchmarking Recurrent Event-Based Object Detection for Industrial Multi-Class Recognition on MTEvent (viability: 5): https://sciencetostartup.com/paper/benchmarking-recurrent-event-based-object-detection-for-industrial-multi-class-recognition-on-mtevent - Benchmarking recurrent event-based object detection for industrial multi-class recognition to improve performance over non-recurrent baselines.
- The Universal Normal Embedding (viability: 7): https://sciencetostartup.com/paper/the-universal-normal-embedding - Unlock controllable image editing and semantic understanding by unifying generative models and vision encoders through a shared Gaussian latent space.
- Image-Conditioned Adaptive Parameter Tuning for Visual Odometry Frontends (viability: 5): https://sciencetostartup.com/paper/image-conditioned-adaptive-parameter-tuning-for-visual-odometry-frontends - An AI system that dynamically tunes visual odometry parameters for robots based on real-time image analysis, improving tracking and reducing computational cost.
- Dynamic Exposure Burst Image Restoration (viability: 7): https://sciencetostartup.com/paper/dynamic-exposure-burst-image-restoration - Dynamically predict optimal exposure times for burst photography to achieve state-of-the-art image restoration quality, validated on real-world camera systems.
- SHARP: Spectrum-aware Highly-dynamic Adaptation for Resolution Promotion in Remote Sensing Synthesis (viability: 8): https://sciencetostartup.com/paper/sharp-spectrum-aware-highly-dynamic-adaptation-for-resolution-promotion-in-remote-sensing-synthesis - A novel training-free method for high-resolution remote sensing image synthesis that dynamically adapts positional embeddings during denoising, outperforming existing baselines.
- Show Me What You Don't Know: Efficient Sampling from Invariant Sets for Model Validation (viability: 7): https://sciencetostartup.com/paper/show-me-what-you-don-t-know-efficient-sampling-from-invariant-sets-for-model-validation - A training-free method to efficiently sample from model feature invariances using pretrained diffusion models, enabling rapid validation of model behavior.
- Memory-Efficient Boundary Map for Large-Scale Occupancy Grid Mapping (viability: 7): https://sciencetostartup.com/paper/memory-efficient-boundary-map-for-large-scale-occupancy-grid-mapping - A novel memory-efficient boundary map representation for large-scale 3D occupancy grid mapping in robotics, with open-source code available.
- Extending Precipitation Nowcasting Horizons via Spectral Fusion of Radar Observations and Foundation Model Priors (viability: 8): https://sciencetostartup.com/paper/extending-precipitation-nowcasting-horizons-via-spectral-fusion-of-radar-observations-and-foundation-model-priors - A novel frequency-domain fusion framework that integrates radar observations with weather foundation model forecasts to significantly extend precipitation nowcasting horizons.
- Let's Think with Images Efficiently! An Interleaved-Modal Chain-of-Thought Reasoning Framework with Dynamic and Precise Visual Thoughts (viability: 7): https://sciencetostartup.com/paper/let-s-think-with-images-efficiently-an-interleaved-modal-chain-of-thought-reasoning-framework-with-dynamic-and-precise-v - A framework for more efficient and effective multimodal reasoning by dynamically integrating and precisely representing visual information, significantly reducing token consumption.
- Identifiability and amortized inference limitations in Kuramoto models (viability: 3): https://sciencetostartup.com/paper/identifiability-and-amortized-inference-limitations-in-kuramoto-models - A novel amortized Bayesian inference approach for fast and scalable parameter estimation in complex dynamical systems like Kuramoto models.
- Getting to the Point: Why Pointing Improves LVLMs (viability: 4): https://sciencetostartup.com/paper/getting-to-the-point-why-pointing-improves-lvlms - This research explores how explicit object pointing in Large Vision-Language Models improves zero-shot counting accuracy and generalization by encoding spatial information.
- The Presupposition Problem in Representation Genesis (viability: 1): https://sciencetostartup.com/paper/the-presupposition-problem-in-representation-genesis - This paper analyzes the philosophical underpinnings of representation genesis in AI, identifying structural limitations in current theories.
- CellFluxRL: Biologically-Constrained Virtual Cell Modeling via Reinforcement Learning (viability: 5): https://sciencetostartup.com/paper/cellfluxrl-biologically-constrained-virtual-cell-modeling-via-reinforcement-learning - A reinforcement learning framework for generating biologically plausible virtual cells to accelerate drug discovery.
- The Reasoning Error About Reasoning: Why Different Types of Reasoning Require Different Representational Structures (viability: 2): https://sciencetostartup.com/paper/the-reasoning-error-about-reasoning-why-different-types-of-reasoning-require-different-representational-structures - A theoretical framework analyzing the structural demands of different reasoning types to understand limitations of current AI approaches.
- Cognitive Agency Surrender: Defending Epistemic Sovereignty via Scaffolded AI Friction (viability: 2): https://sciencetostartup.com/paper/cognitive-agency-surrender-defending-epistemic-sovereignty-via-scaffolded-ai-friction - This paper theorizes and quantifies the risk of cognitive agency surrender due to frictionless AI interfaces, proposing 'Scaffolded Cognitive Friction' as a technical prerequisite for AI governance and societal cognitive resilience.
- EvoIdeator: Evolving Scientific Ideas through Checklist-Grounded Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/evoideator-evolving-scientific-ideas-through-checklist-grounded-reinforcement-learning - A framework that uses checklist-grounded reinforcement learning to enable LLMs to systematically evolve and refine scientific ideas based on fine-grained feedback.
- CurvZO: Adaptive Curvature-Guided Sparse Zeroth-Order Optimization for Efficient LLM Fine-Tuning (viability: 5): https://sciencetostartup.com/paper/curvzo-adaptive-curvature-guided-sparse-zeroth-order-optimization-for-efficient-llm-fine-tuning - A novel optimization method for memory-efficient LLM fine-tuning that reduces variance and speeds up convergence using adaptive curvature guidance.
- FISformer: Replacing Self-Attention with a Fuzzy Inference System in Transformer Models for Time Series Forecasting (viability: 7): https://sciencetostartup.com/paper/fisformer-replacing-self-attention-with-a-fuzzy-inference-system-in-transformer-models-for-time-series-forecasting - FISFormer replaces self-attention in Transformers with a fuzzy inference system for more accurate, robust, and interpretable time series forecasting.
- Can a Robot Walk the Robotic Dog: Triple-Zero Collaborative Navigation for Heterogeneous Multi-Agent Systems (viability: 7): https://sciencetostartup.com/paper/can-a-robot-walk-the-robotic-dog-triple-zero-collaborative-navigation-for-heterogeneous-multi-agent-systems - A collaborative navigation framework for heterogeneous robots that requires no training or simulation, enabling real-world deployment of robot cooperation.
- SemEval-2026 Task 12: Abductive Event Reasoning: Towards Real-World Event Causal Inference for Large Language Models (viability: 7): https://sciencetostartup.com/paper/semeval-2026-task-12-abductive-event-reasoning-towards-real-world-event-causal-inference-for-large-language-models - A new benchmark and dataset for abductive event reasoning to enable LLMs to infer direct causes of real-world events from evidence.
- Probing How Scalable Table Data Enhances General Long-Context Reasoning (viability: 7): https://sciencetostartup.com/paper/probing-how-scalable-table-data-enhances-general-long-context-reasoning - Synthesize structured table data to significantly boost LLM long-context reasoning capabilities, improving performance on both in-domain and out-of-domain tasks.
- Uncertainty Quantification for Distribution-to-Distribution Flow Matching in Scientific Imaging (viability: 4): https://sciencetostartup.com/paper/uncertainty-quantification-for-distribution-to-distribution-flow-matching-in-scientific-imaging - A unified framework for uncertainty quantification in generative models for scientific imaging, improving reliability and detecting out-of-distribution cases.
- When Exploration Comes for Free with Mixture-Greedy: Do we need UCB in Diversity-Aware Multi-Armed Bandits? (viability: 5): https://sciencetostartup.com/paper/when-exploration-comes-for-free-with-mixture-greedy-do-we-need-ucb-in-diversity-aware-multi-armed-bandits - A new 'Mixture-Greedy' strategy for generative AI model selection outperforms traditional UCB methods by leveraging intrinsic exploration, leading to faster convergence and better performance without explicit exploration bonuses.
- Compensating Visual Insufficiency with Stratified Language Guidance for Long-Tail Class Incremental Learning (viability: 7): https://sciencetostartup.com/paper/compensating-visual-insufficiency-with-stratified-language-guidance-for-long-tail-class-incremental-learning - Leveraging large language models to guide incremental learning for imbalanced datasets, improving performance on rare classes.
- Data-Free Layer-Adaptive Merging via Fisher Information for Long-to-Short Reasoning LLMs (viability: 7): https://sciencetostartup.com/paper/data-free-layer-adaptive-merging-via-fisher-information-for-long-to-short-reasoning-llms - A novel data-free model merging technique that leverages Fisher Information to significantly improve long-to-short reasoning in LLMs, outperforming existing methods without calibration data.
- Cybersecurity Guidance for Smart Homes: A Cross-National Review of Government Sources (viability: 3): https://sciencetostartup.com/paper/cybersecurity-guidance-for-smart-homes-a-cross-national-review-of-government-sources - A review of government guidance for smart home cybersecurity incidents reveals a lack of structured incident response support for users.
- Rethinking Token Reduction for Large Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/rethinking-token-reduction-for-large-vision-language-models - A novel learning-based method to efficiently reduce visual tokens in large vision-language models for multi-turn question answering, improving inference costs without sacrificing accuracy.
- PPGL-Swarm: Integrated Multimodal Risk Stratification and Hereditary Syndrome Detection in Pheochromocytoma and Paraganglioma (viability: 5): https://sciencetostartup.com/paper/ppgl-swarm-integrated-multimodal-risk-stratification-and-hereditary-syndrome-detection-in-pheochromocytoma-and-paragangl - An agentic diagnostic system for pheochromocytomas and paragangliomas that automates risk stratification and detects hereditary syndromes by integrating multimodal data and providing auditable reasoning.
- A Blueprint for Self-Evolving Coding Agents in Vehicle Aerodynamic Drag Prediction (viability: 4): https://sciencetostartup.com/paper/a-blueprint-for-self-evolving-coding-agents-in-vehicle-aerodynamic-drag-prediction - Develops a framework for self-evolving coding agents to automate and accelerate vehicle aerodynamic drag prediction through surrogate model discovery.
- Structured Visual Narratives Undermine Safety Alignment in Multimodal Large Language Models (viability: 4): https://sciencetostartup.com/paper/structured-visual-narratives-undermine-safety-alignment-in-multimodal-large-language-models - This research identifies a novel vulnerability in multimodal LLMs using comic-based jailbreaks, highlighting the need for more robust safety alignment.
- MIND: Multi-agent inference for negotiation dialogue in travel planning (viability: 7): https://sciencetostartup.com/paper/mind-multi-agent-inference-for-negotiation-dialogue-in-travel-planning - A framework for multi-agent negotiation that infers user preferences to achieve consensus in complex planning scenarios.
- RefracGS: Novel View Synthesis Through Refractive Water Surfaces with 3D Gaussian Ray Tracing (viability: 8): https://sciencetostartup.com/paper/refracgs-novel-view-synthesis-through-refractive-water-surfaces-with-3d-gaussian-ray-tracing - A novel framework for high-fidelity novel view synthesis through refractive water surfaces by jointly modeling the water surface and the scene beneath using 3D Gaussian ray tracing.
- Bridges connecting Encryption Schemes (viability: 1): https://sciencetostartup.com/paper/bridges-connecting-encryption-schemes - This paper explores theoretical bridges between encryption schemes, inspired by homomorphic encryption, with security guarantees based on existing schemes.
- Deterministic Hallucination Detection in Medical VQA via Confidence-Evidence Bayesian Gain (viability: 7): https://sciencetostartup.com/paper/deterministic-hallucination-detection-in-medical-vqa-via-confidence-evidence-bayesian-gain - A deterministic method to detect hallucinations in medical AI by analyzing token-level confidence and visual evidence, offering a computationally efficient and self-contained solution.
- Reasoning Provenance for Autonomous AI Agents: Structured Behavioral Analytics Beyond State Checkpoints and Execution Traces (viability: 4): https://sciencetostartup.com/paper/reasoning-provenance-for-autonomous-ai-agents-structured-behavioral-analytics-beyond-state-checkpoints-and-execution-tra - A new primitive for structured reasoning provenance in AI agents enables population-level behavioral analytics beyond traditional debugging tools.
- AI Token Futures Market: Commoditization of Compute and Derivatives Contract Design (viability: 3): https://sciencetostartup.com/paper/ai-token-futures-market-commoditization-of-compute-and-derivatives-contract-design - This paper proposes a financial market for AI tokens to hedge against compute cost volatility for enterprises.
- Mirage The Illusion of Visual Understanding (viability: 7): https://sciencetostartup.com/paper/mirage-the-illusion-of-visual-understanding - We expose fundamental vulnerabilities in visual-language AI by demonstrating 'mirage reasoning' where models hallucinate visual understanding, and introduce a solution for robust, vision-grounded evaluation.
- LipsAM: Lipschitz-Continuous Amplitude Modifier for Audio Signal Processing and its Application to Plug-and-Play Dereverberation (viability: 5): https://sciencetostartup.com/paper/lipsam-lipschitz-continuous-amplitude-modifier-for-audio-signal-processing-and-its-application-to-plug-and-play-dereverb - Develop Lipschitz-continuous amplitude modifiers for robust audio signal processing, enhancing applications like speech dereverberation.
- BiPreManip: Learning Affordance-Based Bimanual Preparatory Manipulation through Anticipatory Collaboration (viability: 7): https://sciencetostartup.com/paper/bipremanip-learning-affordance-based-bimanual-preparatory-manipulation-through-anticipatory-collaboration - A visual affordance-based framework for bimanual robotic manipulation that anticipates and facilitates complex preparatory actions for goal-directed tasks.
- CoNBONet: Conformalized Neuroscience-inspired Bayesian Operator Network for Reliability Analysis (viability: 7): https://sciencetostartup.com/paper/conbonet-conformalized-neuroscience-inspired-bayesian-operator-network-for-reliability-analysis - A neuroscience-inspired Bayesian network for fast, energy-efficient, and uncertainty-aware reliability analysis of complex engineering systems.
- Thinking Deeper, Not Longer: Depth-Recurrent Transformers for Compositional Generalization (viability: 3): https://sciencetostartup.com/paper/thinking-deeper-not-longer-depth-recurrent-transformers-for-compositional-generalization - A depth-recurrent Transformer architecture designed for improved compositional generalization in tasks requiring variable-depth reasoning.
- Optimizing Multi-Agent Weather Captioning via Text Gradient Descent: A Training-Free Approach with Consensus-Aware Gradient Fusion (viability: 7): https://sciencetostartup.com/paper/optimizing-multi-agent-weather-captioning-via-text-gradient-descent-a-training-free-approach-with-consensus-aware-gradie - A training-free multi-agent LLM framework that generates interpretable, domain-specific weather captions by fusing textual gradients from specialized agents.
- PRM-as-a-Judge: A Dense Evaluation Paradigm for Fine-Grained Robotic Auditing (viability: 7): https://sciencetostartup.com/paper/prm-as-a-judge-a-dense-evaluation-paradigm-for-fine-grained-robotic-auditing - A new dense evaluation paradigm for robotic auditing using process reward models to provide fine-grained insights beyond binary success rates.
- HumanOmni-Speaker: Identifying Who said What and When (viability: 7): https://sciencetostartup.com/paper/humanomni-speaker-identifying-who-said-what-and-when - A novel multimodal AI system that accurately identifies who said what and when in complex conversations by analyzing high-frequency visual cues, overcoming limitations of current models.
- TAMTRL: Teacher-Aligned Reward Reshaping for Multi-Turn Reinforcement Learning in Long-Context Compression (viability: 7): https://sciencetostartup.com/paper/tamtrl-teacher-aligned-reward-reshaping-for-multi-turn-reinforcement-learning-in-long-context-compression - TAMTRL improves long-context LLM performance by providing fine-grained, teacher-aligned rewards during multi-turn memory updates, overcoming temporal credit assignment challenges.
- Cross-Scenario Deraining Adaptation with Unpaired Data: Superpixel Structural Priors and Multi-Stage Pseudo-Rain Synthesis (viability: 7): https://sciencetostartup.com/paper/cross-scenario-deraining-adaptation-with-unpaired-data-superpixel-structural-priors-and-multi-stage-pseudo-rain-synthesi - A plug-and-play module that adapts image deraining models to unseen real-world conditions using unpaired data and synthesized pseudo-rain, significantly improving performance and training speed.
- OmniFM: Toward Modality-Robust and Task-Agnostic Federated Learning for Heterogeneous Medical Imaging (viability: 7): https://sciencetostartup.com/paper/omnifm-toward-modality-robust-and-task-agnostic-federated-learning-for-heterogeneous-medical-imaging - A modality- and task-agnostic federated learning framework for heterogeneous medical imaging that unifies diverse downstream tasks by leveraging frequency-domain insights.
- A Comparative Analysis of LLM Memorization at Statistical and Internal Levels: Cross-Model Commonalities and Model-Specific Signatures (viability: 4): https://sciencetostartup.com/paper/a-comparative-analysis-of-llm-memorization-at-statistical-and-internal-levels-cross-model-commonalities-and-model-specif - This research analyzes LLM memorization across multiple model families to uncover universal patterns and model-specific behaviors, aiming for a fundamental understanding of how LLMs retain information.
- TrustFed: Enabling Trustworthy Medical AI under Data Privacy Constraints (viability: 7): https://sciencetostartup.com/paper/trustfed-enabling-trustworthy-medical-ai-under-data-privacy-constraints - A federated learning framework for trustworthy medical AI that ensures privacy and reliable uncertainty quantification across diverse healthcare data.
- Towards Secure Retrieval-Augmented Generation: A Comprehensive Review of Threats, Defenses and Benchmarks (viability: 4): https://sciencetostartup.com/paper/towards-secure-retrieval-augmented-generation-a-comprehensive-review-of-threats-defenses-and-benchmarks - This paper provides a comprehensive review of security threats and defenses for Retrieval-Augmented Generation (RAG) systems, aiming to foster the development of robust and trustworthy RAG applications.
- MISApp: Multi-Hop Intent-Aware Session Graph Learning for Next App Prediction (viability: 7): https://sciencetostartup.com/paper/misapp-multi-hop-intent-aware-session-graph-learning-for-next-app-prediction - A profile-free framework for next app prediction using multi-hop session graph learning to capture evolving user intent, outperforming baselines in real-world scenarios.
- TLS Certificate and Domain Feature Analysis of Phishing Domains in the Danish .dk Namespace (viability: 4): https://sciencetostartup.com/paper/tls-certificate-and-domain-feature-analysis-of-phishing-domains-in-the-danish-dk-namespace - Leveraging TLS certificate and domain features to detect phishing domains in the Danish .dk namespace.
- FedCVU: Federated Learning for Cross-View Video Understanding (viability: 7): https://sciencetostartup.com/paper/fedcvu-federated-learning-for-cross-view-video-understanding - A federated learning framework that enables privacy-preserving cross-view video understanding by aligning representations and reducing communication overhead.
- Are AI-assisted Development Tools Immune to Prompt Injection? (viability: 5): https://sciencetostartup.com/paper/are-ai-assisted-development-tools-immune-to-prompt-injection - This research analyzes prompt injection vulnerabilities in AI-assisted development tools, identifying security gaps and providing guidance for building safer AI workflows.
- Auditing MCP Servers for Over-Privileged Tool Capabilities (viability: 4): https://sciencetostartup.com/paper/auditing-mcp-servers-for-over-privileged-tool-capabilities - A security auditing toolkit for LLM tool integration protocols to detect and mitigate over-privileged capabilities.
- Engineering Distributed Governance for Regional Prosperity: A Socio-Technical Framework for Mitigating Under-Vibrancy via Human Data Engines (viability: 5): https://sciencetostartup.com/paper/engineering-distributed-governance-for-regional-prosperity-a-socio-technical-framework-for-mitigating-under-vibrancy-via - An AI-driven socio-technical framework to optimize regional economic flow and mitigate under-vibrancy by analyzing spending and sentiment data.
- No Dense Tensors Needed: Fully Sparse Object Detection on Event-Camera Voxel Grids (viability: 7): https://sciencetostartup.com/paper/no-dense-tensors-needed-fully-sparse-object-detection-on-event-camera-voxel-grids - A fully sparse object detection system for event cameras that dramatically reduces memory and storage requirements while maintaining high accuracy for detecting fast-moving objects.
- Silicon Bureaucracy and AI Test-Oriented Education: Contamination Sensitivity and Score Confidence in LLM Benchmarks (viability: 4): https://sciencetostartup.com/paper/silicon-bureaucracy-and-ai-test-oriented-education-contamination-sensitivity-and-score-confidence-in-llm-benchmarks - A framework to audit LLM benchmarks for contamination sensitivity, providing confidence scores for model evaluations.
- RTD-RAX: Fast, Safe Trajectory Planning for Systems under Unknown Disturbances (viability: 3): https://sciencetostartup.com/paper/rtd-rax-fast-safe-trajectory-planning-for-systems-under-unknown-disturbances - A framework for safe, real-time trajectory planning that accounts for unknown disturbances.
- EnterpriseLab: A Full-Stack Platform for developing and deploying agents in Enterprises (viability: 8): https://sciencetostartup.com/paper/enterpriselab-a-full-stack-platform-for-developing-and-deploying-agents-in-enterprises - EnterpriseLab is a full-stack platform enabling enterprises to develop and deploy specialized, cost-effective AI agents that match frontier model performance while ensuring data sovereignty.
- Dual-level Adaptation for Multi-Object Tracking: Building Test-Time Calibration from Experience and Intuition (viability: 7): https://sciencetostartup.com/paper/dual-level-adaptation-for-multi-object-tracking-building-test-time-calibration-from-experience-and-intuition - A test-time adaptation framework for multi-object tracking that leverages memory and experience to improve performance under distribution shifts.
- PGR-Net: Prior-Guided ROI Reasoning Network for Brain Tumor MRI Segmentation (viability: 7): https://sciencetostartup.com/paper/pgr-net-prior-guided-roi-reasoning-network-for-brain-tumor-mri-segmentation - A novel MRI segmentation network that leverages data-driven spatial priors to precisely identify brain tumors, outperforming existing methods with a compact model.
- Neyman-Pearson multiclass classification under label noise via empirical likelihood (viability: 3): https://sciencetostartup.com/paper/neyman-pearson-multiclass-classification-under-label-noise-via-empirical-likelihood - A statistical method for multiclass classification that accounts for noisy labels using empirical likelihood, offering theoretical guarantees.
- Proximal Policy Optimization in Path Space: A Schrödinger Bridge Perspective (viability: 3): https://sciencetostartup.com/paper/proximal-policy-optimization-in-path-space-a-schr-dinger-bridge-perspective - A theoretical framework for optimizing generative reinforcement learning policies by reformulating proximal policy optimization in path space.
- Efficient Zero-Shot AI-Generated Image Detection (viability: 7): https://sciencetostartup.com/paper/efficient-zero-shot-ai-generated-image-detection - A computationally efficient, training-free method for detecting AI-generated images with significantly improved accuracy over state-of-the-art.
- 4DGS360: 360° Gaussian Reconstruction of Dynamic Objects from a Single Video (viability: 7): https://sciencetostartup.com/paper/4dgs360-360-gaussian-reconstruction-of-dynamic-objects-from-a-single-video - Reconstruct dynamic 360° objects from single videos with improved 3D geometry and occluded region handling.
- Rateless DeepJSCC for Broadcast Channels: a Rate-Distortion-Complexity Tradeoff (viability: 4): https://sciencetostartup.com/paper/rateless-deepjscc-for-broadcast-channels-a-rate-distortion-complexity-tradeoff - A novel deep learning framework for adaptive wireless broadcasting that optimizes image quality, transmission rate, and processing complexity for heterogeneous edge devices.
- AdaEdit: Adaptive Temporal and Channel Modulation for Flow-Based Image Editing (viability: 7): https://sciencetostartup.com/paper/adaedit-adaptive-temporal-and-channel-modulation-for-flow-based-image-editing - AdaEdit offers a plug-and-play framework for training-free, text-guided image editing that adaptively balances source feature preservation with synthesized content generation, outperforming existing methods on key metrics.
- AgenticRec: End-to-End Tool-Integrated Policy Optimization for Ranking-Oriented Recommender Agents (viability: 7): https://sciencetostartup.com/paper/agenticrec-end-to-end-tool-integrated-policy-optimization-for-ranking-oriented-recommender-agents - AgenticRec optimizes end-to-end recommender agent decision-making for improved ranking accuracy by integrating tool use and refining user preferences.
- Towards Multimodal Time Series Anomaly Detection with Semantic Alignment and Condensed Interaction (viability: 7): https://sciencetostartup.com/paper/towards-multimodal-time-series-anomaly-detection-with-semantic-alignment-and-condensed-interaction - A multimodal time series anomaly detection model that aligns semantic information across time and text to identify critical system failures.
- SARe: Structure-Aware Large-Scale 3D Fragment Reassembly (viability: 7): https://sciencetostartup.com/paper/sare-structure-aware-large-scale-3d-fragment-reassembly - A generative framework for robustly reassembling large numbers of 3D fragments into complete shapes, outperforming existing methods in challenging scenarios.
- Rule-State Inference (RSI): A Bayesian Framework for Compliance Monitoring in Rule-Governed Domains (viability: 7): https://sciencetostartup.com/paper/rule-state-inference-rsi-a-bayesian-framework-for-compliance-monitoring-in-rule-governed-domains - A Bayesian framework for compliance monitoring that infers rule activation and drift from noisy data, offering significant speedups over traditional methods.
- INTRYGUE: Induction-Aware Entropy Gating for Reliable RAG Uncertainty Estimation (viability: 7): https://sciencetostartup.com/paper/intrygue-induction-aware-entropy-gating-for-reliable-rag-uncertainty-estimation - A novel method to improve the reliability of LLMs in retrieval-augmented generation by accurately detecting hallucinations through induction-aware uncertainty estimation.
- mSFT: Addressing Dataset Mixtures Overfiting Heterogeneously in Multi-task SFT (viability: 7): https://sciencetostartup.com/paper/msft-addressing-dataset-mixtures-overfiting-heterogeneously-in-multi-task-sft - An overfitting-aware algorithm that optimizes multi-task SFT by dynamically adjusting data mixtures, improving model performance and efficiency.
- Riemannian Geometry Speaks Louder Than Words: From Graph Foundation Model to Next-Generation Graph Intelligence (viability: 2): https://sciencetostartup.com/paper/riemannian-geometry-speaks-louder-than-words-from-graph-foundation-model-to-next-generation-graph-intelligence - A theoretical framework for graph foundation models using Riemannian geometry to capture complex structural patterns.
- A Multidisciplinary AI Board for Multimodal Dementia Characterization and Risk Assessment (viability: 8): https://sciencetostartup.com/paper/a-multidisciplinary-ai-board-for-multimodal-dementia-characterization-and-risk-assessment - An interactive multi-agent AI system that synthesizes patient data from EHR, notes, and imaging to provide clinicians with enhanced dementia characterization and risk assessment.
- In-network Attack Detection with Federated Deep Learning in IoT Networks: Real Implementation and Analysis (viability: 5): https://sciencetostartup.com/paper/in-network-attack-detection-with-federated-deep-learning-in-iot-networks-real-implementation-and-analysis - A federated learning framework for real-time, privacy-preserving in-network attack detection on resource-constrained IoT devices.
- Spatio-Temporal Attention Enhanced Multi-Agent DRL for UAV-Assisted Wireless Networks with Limited Communications (viability: 3): https://sciencetostartup.com/paper/spatio-temporal-attention-enhanced-multi-agent-drl-for-uav-assisted-wireless-networks-with-limited-communications - A delay-tolerant multi-agent deep reinforcement learning algorithm for UAV-assisted wireless networks to improve throughput and reduce information delay.
- SSAM: Singular Subspace Alignment for Merging Multimodal Large Language Models (viability: 8): https://sciencetostartup.com/paper/ssam-singular-subspace-alignment-for-merging-multimodal-large-language-models - A training-free framework to merge existing multimodal LLMs into a single model capable of handling any combination of input modalities, achieving state-of-the-art performance without new data.
- HACMatch Semi-Supervised Rotation Regression with Hardness-Aware Curriculum Pseudo Labeling (viability: 5): https://sciencetostartup.com/paper/hacmatch-semi-supervised-rotation-regression-with-hardness-aware-curriculum-pseudo-labeling - A semi-supervised learning framework for 3D object rotation regression from 2D images that uses hardness-aware curriculum learning and structured data augmentation to improve performance with limited labeled data.
- Conformal Koopman for Embedded Nonlinear Control with Statistical Robustness: Theory and Real-World Validation (viability: 7): https://sciencetostartup.com/paper/conformal-koopman-for-embedded-nonlinear-control-with-statistical-robustness-theory-and-real-world-validation - A data-driven framework for statistically robust nonlinear control with formal safety guarantees, validated on real-world drones.
- Mind over Space: Can Multimodal Large Language Models Mentally Navigate? (viability: 7): https://sciencetostartup.com/paper/mind-over-space-can-multimodal-large-language-models-mentally-navigate - A new model and benchmark for enabling multimodal LLMs to perform spatial reasoning and mental navigation in embodied agents.
- Adaptive Robust Estimator for Multi-Agent Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/adaptive-robust-estimator-for-multi-agent-reinforcement-learning - A robust multi-agent reinforcement learning framework that improves collaborative reasoning in LLMs by addressing interaction ambiguity and noisy rewards.
- Rethinking Visual Privacy: A Compositional Privacy Risk Framework for Severity Assessment with VLMs (viability: 7): https://sciencetostartup.com/paper/rethinking-visual-privacy-a-compositional-privacy-risk-framework-for-severity-assessment-with-vlms - A framework and fine-tuned model for assessing nuanced visual privacy risks by composing individual attribute dangers.
- DATASHI: A Parallel English-Tashlhiyt Corpus for Orthography Normalization and Low-Resource Language Processing (viability: 5): https://sciencetostartup.com/paper/datashi-a-parallel-english-tashlhiyt-corpus-for-orthography-normalization-and-low-resource-language-processing - A new parallel corpus for English-Tashlhiyt enables orthography normalization and improved NLP for low-resource Amazigh languages.
- Kolmogorov Complexity Bounds for LLM Steganography and a Perplexity-Based Detection Proxy (viability: 3): https://sciencetostartup.com/paper/kolmogorov-complexity-bounds-for-llm-steganography-and-a-perplexity-based-detection-proxy - This paper theoretically bounds the complexity of LLM steganography and proposes a perplexity-based detection proxy.
- CataractSAM-2: A Domain-Adapted Model for Anterior Segment Surgery Segmentation and Scalable Ground-Truth Annotation (viability: 8): https://sciencetostartup.com/paper/cataractsam-2-a-domain-adapted-model-for-anterior-segment-surgery-segmentation-and-scalable-ground-truth-annotation - A domain-adapted AI model and annotation toolkit for real-time segmentation in ophthalmic surgery, enabling precise intraoperative perception and scalable dataset development.
- Rethinking SAR ATR: A Target-Aware Frequency-Spatial Enhancement Framework with Noise-Resilient Knowledge Guidance (viability: 7): https://sciencetostartup.com/paper/rethinking-sar-atr-a-target-aware-frequency-spatial-enhancement-framework-with-noise-resilient-knowledge-guidance - A novel framework enhances SAR target recognition by adaptively enhancing frequency-spatial features and using knowledge distillation to improve robustness against speckle noise.
- Toward a Theory of Hierarchical Memory for Language Agents (viability: 3): https://sciencetostartup.com/paper/toward-a-theory-of-hierarchical-memory-for-language-agents - A theoretical framework for understanding and comparing hierarchical memory systems in language agents.
- Counterfactual Credit Policy Optimization for Multi-Agent Collaboration (viability: 7): https://sciencetostartup.com/paper/counterfactual-credit-policy-optimization-for-multi-agent-collaboration - Optimize collaborative multi-agent LLM training by assigning agent-specific learning signals using counterfactual analysis to prevent free-riding.
- Exploring Multimodal Prompts For Unsupervised Continuous Anomaly Detection (viability: 7): https://sciencetostartup.com/paper/exploring-multimodal-prompts-for-unsupervised-continuous-anomaly-detection - A multimodal prompting framework for unsupervised continuous anomaly detection that improves accuracy by capturing richer representations of normality.
- Revisiting Weakly-Supervised Video Scene Graph Generation via Pair Affinity Learning (viability: 7): https://sciencetostartup.com/paper/revisiting-weakly-supervised-video-scene-graph-generation-via-pair-affinity-learning - A novel method for generating structured video scene graphs with reduced annotation costs by learning object interaction affinities.
- Stabilizing Iterative Self-Training with Verified Reasoning via Symbolic Recursive Self-Alignment (viability: 7): https://sciencetostartup.com/paper/stabilizing-iterative-self-training-with-verified-reasoning-via-symbolic-recursive-self-alignment - A framework for stabilizing iterative self-training in LLMs by embedding symbolic verification to eliminate flawed reasoning and improve model reliability.
- From Part to Whole: 3D Generative World Model with an Adaptive Structural Hierarchy (viability: 3): https://sciencetostartup.com/paper/from-part-to-whole-3d-generative-world-model-with-an-adaptive-structural-hierarchy - A novel generative world model for single-image 3D generation that learns an adaptive part-whole hierarchy to improve generalization across diverse object categories and structural complexities.
- A Survey of Web Application Security Tutorials (viability: 3): https://sciencetostartup.com/paper/a-survey-of-web-application-security-tutorials - A system to automatically identify high-quality web application security tutorials by analyzing code examples and resource links.
- PROBE: Diagnosing Residual Concept Capacity in Erased Text-to-Video Diffusion Models (viability: 7): https://sciencetostartup.com/paper/probe-diagnosing-residual-concept-capacity-in-erased-text-to-video-diffusion-models - A diagnostic protocol to quantify and reveal residual concept capacity in text-to-video diffusion models, enabling more robust safety auditing.
- What Do World Models Learn in RL? Probing Latent Representations in Learned Environment Simulators (viability: 2): https://sciencetostartup.com/paper/what-do-world-models-learn-in-rl-probing-latent-representations-in-learned-environment-simulators - This research probes the internal representations of world models in reinforcement learning to understand what they learn about environment dynamics.
- Auction-Based Task Allocation with Energy-Conscientious Trajectory Optimization for AMR Fleets (viability: 4): https://sciencetostartup.com/paper/auction-based-task-allocation-with-energy-conscientious-trajectory-optimization-for-amr-fleets - A hierarchical framework for energy-efficient task allocation and trajectory optimization in autonomous robot fleets, providing guidance on bid strategy based on workspace characteristics.
- Evolutionary Biparty Multiobjective UAV Path Planning: Problems and Empirical Comparisons (viability: 2): https://sciencetostartup.com/paper/evolutionary-biparty-multiobjective-uav-path-planning-problems-and-empirical-comparisons - This paper proposes a novel approach to multiobjective UAV path planning involving two decision-makers, aiming to improve efficiency and safety.
- Sharper Generalization Bounds for Transformer (viability: 2): https://sciencetostartup.com/paper/sharper-generalization-bounds-for-transformer - This paper provides theoretical generalization bounds for Transformer models, offering deeper insights into their learning capabilities.
- Generalization Limits of In-Context Operator Networks for Higher-Order Partial Differential Equations (viability: 4): https://sciencetostartup.com/paper/generalization-limits-of-in-context-operator-networks-for-higher-order-partial-differential-equations - A new class of operator networks that leverage in-context learning to solve complex higher-order partial differential equations with qualitative accuracy.
- LLM-Based Test Case Generation in DBMS through Monte Carlo Tree Search (viability: 5): https://sciencetostartup.com/paper/llm-based-test-case-generation-in-dbms-through-monte-carlo-tree-search - An LLM-based framework using Monte Carlo Tree Search to generate diverse and high-coverage SQL test cases for Database Management Systems.
- SynSym: A Synthetic Data Generation Framework for Psychiatric Symptom Identification (viability: 7): https://sciencetostartup.com/paper/synsym-a-synthetic-data-generation-framework-for-psychiatric-symptom-identification - A framework using LLMs to generate synthetic psychiatric symptom data for improved mental health identification models.
- PEARL: Geometry Aligns Semantics for Training-Free Open-Vocabulary Semantic Segmentation (viability: 7): https://sciencetostartup.com/paper/pearl-geometry-aligns-semantics-for-training-free-open-vocabulary-semantic-segmentation - A training-free open-vocabulary semantic segmentation method that aligns text and vision geometry for rapid adaptation to new labels without retraining.
- VIGIL: Part-Grounded Structured Reasoning for Generalizable Deepfake Detection (viability: 7): https://sciencetostartup.com/paper/vigil-part-grounded-structured-reasoning-for-generalizable-deepfake-detection - VIGIL is a part-centric deepfake detection framework that generates reliable textual explanations by separating evidence generation from manipulation localization, outperforming existing methods on a challenging new benchmark.
- BOxCrete: A Bayesian Optimization Open-Source AI Model for Concrete Strength Forecasting and Mix Optimization (viability: 7): https://sciencetostartup.com/paper/boxcrete-a-bayesian-optimization-open-source-ai-model-for-concrete-strength-forecasting-and-mix-optimization - An open-source AI framework for optimizing concrete mix designs to improve strength and reduce environmental impact.
- CatRAG: Functor-Guided Structural Debiasing with Retrieval Augmentation for Fair LLMs (viability: 7): https://sciencetostartup.com/paper/catrag-functor-guided-structural-debiasing-with-retrieval-augmentation-for-fair-llms - A novel framework integrating category theory and RAG to significantly reduce bias in LLMs while maintaining accuracy, achieving state-of-the-art results on fairness benchmarks.
- SafePilot: A Framework for Assuring LLM-enabled Cyber-Physical Systems (viability: 7): https://sciencetostartup.com/paper/safepilot-a-framework-for-assuring-llm-enabled-cyber-physical-systems - SafePilot provides end-to-end assurance for LLM-enabled cyber-physical systems by verifying and correcting LLM outputs against formal specifications, preventing unsafe actions.
- Efficient Failure Management for Multi-Agent Systems with Reasoning Trace Representation (viability: 7): https://sciencetostartup.com/paper/efficient-failure-management-for-multi-agent-systems-with-reasoning-trace-representation - An efficient framework for real-time failure detection, diagnosis, and mitigation in LLM-based multi-agent systems using historical failure patterns.
- Generalizable Self-Evolving Memory for Automatic Prompt Optimization (viability: 7): https://sciencetostartup.com/paper/generalizable-self-evolving-memory-for-automatic-prompt-optimization - A self-evolving memory system that automatically optimizes LLM prompts for better generalization and reduced cost.
- Triangulating Temporal Dynamics in Multilingual Swiss Online News (viability: 4): https://sciencetostartup.com/paper/triangulating-temporal-dynamics-in-multilingual-swiss-online-news - A framework for analyzing temporal trends in multilingual news coverage to understand public discourse dynamics.
- When the Abyss Looks Back: Unveiling Evolving Dark Patterns in Cookie Consent Banners (viability: 7): https://sciencetostartup.com/paper/when-the-abyss-looks-back-unveiling-evolving-dark-patterns-in-cookie-consent-banners - A system that detects evolved dark patterns in cookie consent banners to ensure regulatory compliance and protect user privacy.
- Back to Point: Exploring Point-Language Models for Zero-Shot 3D Anomaly Detection (viability: 7): https://sciencetostartup.com/paper/back-to-point-exploring-point-language-models-for-zero-shot-3d-anomaly-detection - A novel framework for zero-shot 3D anomaly detection that leverages point-language models to preserve geometric details and detect defects without target-category training data.
- Optimizing Feature Extraction for On-device Model Inference with User Behavior Sequences (viability: 7): https://sciencetostartup.com/paper/optimizing-feature-extraction-for-on-device-model-inference-with-user-behavior-sequences - Automate and optimize on-device feature extraction to significantly reduce mobile app model inference latency.
- Parameter-efficient Prompt Tuning and Hierarchical Textual Guidance for Few-shot Whole Slide Image Classification (viability: 7): https://sciencetostartup.com/paper/parameter-efficient-prompt-tuning-and-hierarchical-textual-guidance-for-few-shot-whole-slide-image-classification - A parameter-efficient prompt tuning method for few-shot whole slide image classification that leverages hierarchical textual guidance to improve accuracy and reduce computational cost in medical diagnostics.
- Quotient Geometry, Effective Curvature, and Implicit Bias in Simple Shallow Neural Networks (viability: 2): https://sciencetostartup.com/paper/quotient-geometry-effective-curvature-and-implicit-bias-in-simple-shallow-neural-networks - Develops a differential-geometric framework to analyze shallow neural networks by accounting for parameter symmetries and their intrinsic geometric properties.
- A Framework for Closed-Loop Robotic Assembly, Alignment and Self-Recovery of Precision Optical Systems (viability: 7): https://sciencetostartup.com/paper/a-framework-for-closed-loop-robotic-assembly-alignment-and-self-recovery-of-precision-optical-systems - An autonomous robotics framework for the precise assembly, alignment, and self-recovery of optical systems, enabling automated optical experiments.
- RuntimeSlicer: Towards Generalizable Unified Runtime State Representation for Failure Management (viability: 7): https://sciencetostartup.com/paper/runtimeslicer-towards-generalizable-unified-runtime-state-representation-for-failure-management - RuntimeSlicer offers a unified representation of system metrics, traces, and logs for more generalizable and efficient failure management in complex software systems.
- Agentic Automation of BT-RADS Scoring: End-to-End Multi-Agent System for Standardized Brain Tumor Follow-up Assessment (viability: 7): https://sciencetostartup.com/paper/agentic-automation-of-bt-rads-scoring-end-to-end-multi-agent-system-for-standardized-brain-tumor-follow-up-assessment - An end-to-end multi-agent LLM and CNN system automates complex brain tumor MRI response assessment, significantly outperforming initial clinical scoring.
- StreamingEval: A Unified Evaluation Protocol towards Realistic Streaming Video Understanding (viability: 7): https://sciencetostartup.com/paper/streamingeval-a-unified-evaluation-protocol-towards-realistic-streaming-video-understanding - A unified evaluation framework to benchmark and improve the real-time, resource-constrained understanding capabilities of Video-LLMs for practical AI applications.
- Multinoulli Extension: A Lossless Continuous Relaxation for Partition-Constrained Subset Selection (viability: 2): https://sciencetostartup.com/paper/multinoulli-extension-a-lossless-continuous-relaxation-for-partition-constrained-subset-selection - A novel parameter-free algorithm for partition-constrained subset selection that offers improved efficiency and approximation guarantees over existing methods.
- Learning Can Converge Stably to the Wrong Belief under Latent Reliability (viability: 3): https://sciencetostartup.com/paper/learning-can-converge-stably-to-the-wrong-belief-under-latent-reliability - A theoretical framework to prevent learning systems from converging to incorrect solutions when feedback reliability is uncertain.
- Effective Strategies for Asynchronous Software Engineering Agents (viability: 7): https://sciencetostartup.com/paper/effective-strategies-for-asynchronous-software-engineering-agents - A structured multi-agent coordination paradigm for asynchronous software engineering tasks, improving accuracy and timely completion through centralized delegation and isolated workspaces.
- Learning Trajectory-Aware Multimodal Large Language Models for Video Reasoning Segmentation (viability: 8): https://sciencetostartup.com/paper/learning-trajectory-aware-multimodal-large-language-models-for-video-reasoning-segmentation - A unified framework for video object segmentation that leverages bidirectional text-trajectory alignment within multimodal LLMs to outperform existing methods.
- GaussianSSC: Triplane-Guided Directional Gaussian Fields for 3D Semantic Completion (viability: 4): https://sciencetostartup.com/paper/gaussianssc-triplane-guided-directional-gaussian-fields-for-3d-semantic-completion - A novel approach to semantic scene completion using directional Gaussian fields to improve 3D reconstruction accuracy.
- Off-Policy Evaluation for Ranking Policies under Deterministic Logging Policies (viability: 4): https://sciencetostartup.com/paper/off-policy-evaluation-for-ranking-policies-under-deterministic-logging-policies - Develops novel off-policy evaluation estimators for ranking systems that overcome limitations of deterministic logging policies by leveraging user click behavior.
- Which Concepts to Forget and How to Refuse? Decomposing Concepts for Continual Unlearning in Large Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/which-concepts-to-forget-and-how-to-refuse-decomposing-concepts-for-continual-unlearning-in-large-vision-language-models - A framework for large vision-language models to selectively forget specific concepts while preserving general utility, enabling precise refusal of unwanted image-instruction pairs.
- ALADIN:Attribute-Language Distillation Network for Person Re-Identification (viability: 5): https://sciencetostartup.com/paper/aladin-attribute-language-distillation-network-for-person-re-identification - A knowledge distillation network for person re-identification that improves fine-grained attribute alignment using language models.
- TaigiSpeech: A Low-Resource Real-World Speech Intent Dataset and Preliminary Results with Scalable Data Mining In-the-Wild (viability: 7): https://sciencetostartup.com/paper/taigispeech-a-low-resource-real-world-speech-intent-dataset-and-preliminary-results-with-scalable-data-mining-in-the-wil - Building a scalable speech intent detection system for low-resource languages by leveraging LLMs and multimodal data mining.
- Unified-MAS: Universally Generating Domain-Specific Nodes for Empowering Automatic Multi-Agent Systems (viability: 7): https://sciencetostartup.com/paper/unified-mas-universally-generating-domain-specific-nodes-for-empowering-automatic-multi-agent-systems - Unified-MAS synthesizes domain-specific nodes for multi-agent systems, improving performance and reducing costs in knowledge-intensive tasks.
- Beyond Correlation: Refutation-Validated Aspect-Based Sentiment Analysis for Explainable Energy Market Returns (viability: 4): https://sciencetostartup.com/paper/beyond-correlation-refutation-validated-aspect-based-sentiment-analysis-for-explainable-energy-market-returns - A novel framework for statistically robust, explainable sentiment analysis in financial markets that moves beyond simple correlation.
- Hardening Confidential Federated Compute against Side-channel Attacks (viability: 3): https://sciencetostartup.com/paper/hardening-confidential-federated-compute-against-side-channel-attacks - This paper identifies and mitigates side-channels in a confidential federated compute platform to strengthen differential privacy guarantees.
- DRTriton: Large-Scale Synthetic Data Reinforcement Learning for Triton Kernel Generation (viability: 7): https://sciencetostartup.com/paper/drtriton-large-scale-synthetic-data-reinforcement-learning-for-triton-kernel-generation - A reinforcement learning framework that trains LLMs to generate highly optimized Triton kernels from PyTorch code, significantly outperforming existing models and human experts on kernel optimization tasks.
- EpiMask: Leveraging Epipolar Distance Based Masks in Cross-Attention for Satellite Image Matching (viability: 7): https://sciencetostartup.com/paper/epimask-leveraging-epipolar-distance-based-masks-in-cross-attention-for-satellite-image-matching - A semi-dense image matching network for satellite imagery that leverages epipolar geometry to significantly improve matching accuracy.
- DSPA: Dynamic SAE Steering for Data-Efficient Preference Alignment (viability: 5): https://sciencetostartup.com/paper/dspa-dynamic-sae-steering-for-data-efficient-preference-alignment - A novel inference-time method for aligning language models with user preferences using sparse autoencoders, reducing compute and improving control.
- When Documents Disagree: Measuring Institutional Variation in Transplant Guidance with Retrieval-Augmented Language Models (viability: 7): https://sciencetostartup.com/paper/when-documents-disagree-measuring-institutional-variation-in-transplant-guidance-with-retrieval-augmented-language-model - Quantify and improve inconsistencies in transplant patient education materials using retrieval-augmented language models to address critical information gaps.
- Cross-Context Verification: Hierarchical Detection of Benchmark Contamination through Session-Isolated Analysis (viability: 7): https://sciencetostartup.com/paper/cross-context-verification-hierarchical-detection-of-benchmark-contamination-through-session-isolated-analysis - A novel black-box method and multi-agent framework to detect benchmark contamination in LLMs, ensuring the credibility of coding benchmarks.
- Safety as Computation: Certified Answer Reuse via Capability Closure in Task-Oriented Dialogue (viability: 3): https://sciencetostartup.com/paper/safety-as-computation-certified-answer-reuse-via-capability-closure-in-task-oriented-dialogue - A novel approach to task-oriented dialogue systems that uses safety certification for efficient answer reuse, reducing computation time.
- KG-Hopper: Empowering Compact Open LLMs with Knowledge Graph Reasoning via Reinforcement Learning (viability: 8): https://sciencetostartup.com/paper/kg-hopper-empowering-compact-open-llms-with-knowledge-graph-reasoning-via-reinforcement-learning - A reinforcement learning framework that enables compact open LLMs to perform multi-hop knowledge graph reasoning in a single inference step, outperforming larger models.
- LLM-Powered Workflow Optimization for Multidisciplinary Software Development: An Automotive Industry Case Study (viability: 8): https://sciencetostartup.com/paper/llm-powered-workflow-optimization-for-multidisciplinary-software-development-an-automotive-industry-case-study - An LLM-powered workflow optimization system for multidisciplinary software development that drastically reduces development time and improves communication efficiency in the automotive industry.
- PROMPT2BOX: Uncovering Entailment Structure among LLM Prompts (viability: 7): https://sciencetostartup.com/paper/prompt2box-uncovering-entailment-structure-among-llm-prompts - PROMPT2BOX embeds LLM prompts into a box space to reveal specificity and difficulty, enabling more precise identification of LLM weaknesses.
- Semantic Shift: the Fundamental Challenge in Text Embedding and Retrieval (viability: 3): https://sciencetostartup.com/paper/semantic-shift-the-fundamental-challenge-in-text-embedding-and-retrieval - A theoretical framework to understand and diagnose semantic shift in text embeddings, impacting retrieval performance.
- PAS3R: Pose-Adaptive Streaming 3D Reconstruction for Long Video Sequences (viability: 7): https://sciencetostartup.com/paper/pas3r-pose-adaptive-streaming-3d-reconstruction-for-long-video-sequences - A framework for stable and accurate 3D reconstruction from long video sequences by adaptively updating scene structure based on camera motion and scene novelty.
- Behavioural feasible set: Value alignment constraints on AI decision support (viability: 3): https://sciencetostartup.com/paper/behavioural-feasible-set-value-alignment-constraints-on-ai-decision-support - This research formalizes the 'behavioral feasible set' of AI recommendations under vendor-imposed value alignment constraints, highlighting a governance problem that cannot be solved by prompting alone.
- Image-Based Structural Analysis Using Computer Vision and LLMs: PhotoBeamSolver (viability: 3): https://sciencetostartup.com/paper/image-based-structural-analysis-using-computer-vision-and-llms-photobeamsolver - A program that solves idealized beam models from hand-drawn diagrams using computer vision and statistical learning.
- DomAgent: Leveraging Knowledge Graphs and Case-Based Reasoning for Domain-Specific Code Generation (viability: 8): https://sciencetostartup.com/paper/domagent-leveraging-knowledge-graphs-and-case-based-reasoning-for-domain-specific-code-generation - DomAgent enables LLMs to generate domain-specific code by combining knowledge graphs and case-based reasoning, significantly improving performance on specialized tasks.
- Uncertainty-Aware Knowledge Distillation for Multimodal Large Language Models (viability: 7): https://sciencetostartup.com/paper/uncertainty-aware-knowledge-distillation-for-multimodal-large-language-models - An uncertainty-aware knowledge distillation framework that adaptively balances data and teacher supervision to improve multimodal vision-language models.
- HyReach: Vision-Guided Hybrid Manipulator Reaching in Unseen Cluttered Environments (viability: 7): https://sciencetostartup.com/paper/hyreach-vision-guided-hybrid-manipulator-reaching-in-unseen-cluttered-environments - A vision-guided hybrid robotic manipulator that achieves precise object reaching in cluttered, unseen environments without retraining.
- Is the future of AI green? What can innovation diffusion models say about generative AI's environmental impact? (viability: 2): https://sciencetostartup.com/paper/is-the-future-of-ai-green-what-can-innovation-diffusion-models-say-about-generative-ai-s-environmental-impact - Analyzes the environmental impact of generative AI using innovation diffusion models to forecast its evolution and potential mitigation strategies.
- Efficient Fine-Tuning Methods for Portuguese Question Answering: A Comparative Study of PEFT on BERTimbau and Exploratory Evaluation of Generative LLMs (viability: 7): https://sciencetostartup.com/paper/efficient-fine-tuning-methods-for-portuguese-question-answering-a-comparative-study-of-peft-on-bertimbau-and-exploratory - This research offers efficient fine-tuning methods for Portuguese question answering, demonstrating significant computational cost reductions compared to generative LLMs.
- Silent Commitment Failure in Instruction-Tuned Language Models: Evidence of Governability Divergence Across Architectures (viability: 7): https://sciencetostartup.com/paper/silent-commitment-failure-in-instruction-tuned-language-models-evidence-of-governability-divergence-across-architectures - This research identifies a critical security vulnerability in instruction-tuned LLMs, enabling the development of robust detection and correction mechanisms for autonomous agents.
- Fingerprinting Deep Neural Networks for Ownership Protection: An Analytical Approach (viability: 3): https://sciencetostartup.com/paper/fingerprinting-deep-neural-networks-for-ownership-protection-an-analytical-approach - An analytical approach to fingerprinting deep neural networks for model ownership protection by mathematically defining the optimal distance from the decision boundary.
- Bayesian Active Object Recognition and 6D Pose Estimation from Multimodal Contact Sensing (viability: 7): https://sciencetostartup.com/paper/bayesian-active-object-recognition-and-6d-pose-estimation-from-multimodal-contact-sensing - A Bayesian active exploration framework for robots to simultaneously recognize objects and estimate their 6D pose using tactile and force sensing.
- Multi-Perspective LLM Annotations for Valid Analyses in Subjective Tasks (viability: 7): https://sciencetostartup.com/paper/multi-perspective-llm-annotations-for-valid-analyses-in-subjective-tasks - A novel method for LLM annotation that accounts for demographic perspectives, improving accuracy for underrepresented groups with adaptive sampling.
- The Myhill-Nerode Theorem for Bounded Interaction: Canonical Abstractions via Agent-Bounded Indistinguishability (viability: 3): https://sciencetostartup.com/paper/the-myhill-nerode-theorem-for-bounded-interaction-canonical-abstractions-via-agent-bounded-indistinguishability - Formalizing agent-bounded indistinguishability to create canonical abstractions for capacity-limited observers in POMDPs.
- Persona Vectors in Games: Measuring and Steering Strategies via Activation Vectors (viability: 3): https://sciencetostartup.com/paper/persona-vectors-in-games-measuring-and-steering-strategies-via-activation-vectors - Develops a method to analyze and steer high-level behavioral traits in LLMs within strategic game environments.
- Mechanisms of Introspective Awareness (viability: 7): https://sciencetostartup.com/paper/mechanisms-of-introspective-awareness - This research uncovers the internal mechanisms behind LLM's ability to detect injected concepts, suggesting a path to enhance their introspective awareness and robustness.
- A Generalised Exponentiated Gradient Approach to Enhance Fairness in Binary and Multi-class Classification Tasks (viability: 4): https://sciencetostartup.com/paper/a-generalised-exponentiated-gradient-approach-to-enhance-fairness-in-binary-and-multi-class-classification-tasks - A generalized algorithm to enhance fairness in multi-class classification tasks by balancing prediction accuracy with multiple fairness constraints.
- Task-Specific Efficiency Analysis: When Small Language Models Outperform Large Language Models (viability: 4): https://sciencetostartup.com/paper/task-specific-efficiency-analysis-when-small-language-models-outperform-large-language-models - This research provides a framework for selecting smaller language models that offer better performance-efficiency ratios for production deployments.
- Knowledge Priors for Identity-Disentangled Open-Set Privacy-Preserving Video FER (viability: 7): https://sciencetostartup.com/paper/knowledge-priors-for-identity-disentangled-open-set-privacy-preserving-video-fer - A novel framework for privacy-preserving facial expression recognition in videos that anonymizes identity without requiring identity labels, maintaining high accuracy.
- Mitigating Objectness Bias and Region-to-Text Misalignment for Open-Vocabulary Panoptic Segmentation (viability: 7): https://sciencetostartup.com/paper/mitigating-objectness-bias-and-region-to-text-misalignment-for-open-vocabulary-panoptic-segmentation - A modular framework to improve open-vocabulary panoptic segmentation by addressing objectness bias and region-to-text misalignment, achieving state-of-the-art results.
- PivotRL: High Accuracy Agentic Post-Training at Low Compute Cost (viability: 7): https://sciencetostartup.com/paper/pivotrl-high-accuracy-agentic-post-training-at-low-compute-cost - Optimize reinforcement learning models post-training at reduced computational costs using PivotRL.
- An InSAR Phase Unwrapping Framework for Large-scale and Complex Events (viability: 7): https://sciencetostartup.com/paper/an-insar-phase-unwrapping-framework-for-large-scale-and-complex-events - A diffusion model-based framework for accurate InSAR phase unwrapping on large-scale, complex deformation events, overcoming limitations of existing methods.
- HamVision: Hamiltonian Dynamics as Inductive Bias for Medical Image Analysis (viability: 8): https://sciencetostartup.com/paper/hamvision-hamiltonian-dynamics-as-inductive-bias-for-medical-image-analysis - A medical image analysis framework leveraging Hamiltonian dynamics for improved segmentation and classification.
- A transformer architecture alteration to incentivise externalised reasoning (viability: 3): https://sciencetostartup.com/paper/a-transformer-architecture-alteration-to-incentivise-externalised-reasoning - A novel transformer architecture modification and training pipeline to reduce computational cost in LLMs by enabling early exits during reasoning.
- Constrained Online Convex Optimization with Memory and Predictions (viability: 3): https://sciencetostartup.com/paper/constrained-online-convex-optimization-with-memory-and-predictions - Develops novel algorithms for constrained online convex optimization with memory, extending existing frameworks for dynamic systems and scheduling.
- PLR: Plackett-Luce for Reordering In-Context Learning Examples (viability: 7): https://sciencetostartup.com/paper/plr-plackett-luce-for-reordering-in-context-learning-examples - A probabilistic approach to optimize in-context learning example ordering for improved few-shot accuracy in large language models.
- Conspiracy Frame: a Semiotically-Driven Approach for Conspiracy Theories Detection (viability: 4): https://sciencetostartup.com/paper/conspiracy-frame-a-semiotically-driven-approach-for-conspiracy-theories-detection - A dataset and semantic framework for detecting conspiracy theories by analyzing abstract patterns in language.
- Relax Forcing: Relaxed KV-Memory for Consistent Long Video Generation (viability: 7): https://sciencetostartup.com/paper/relax-forcing-relaxed-kv-memory-for-consistent-long-video-generation - A novel structured temporal memory mechanism for autoregressive video diffusion models that improves long-horizon generation by selectively incorporating relevant past information, reducing error accumulation and enhancing motion dynamics.
- TIDE: Token-Informed Depth Execution for Per-Token Early Exit in LLM Inference (viability: 8): https://sciencetostartup.com/paper/tide-token-informed-depth-execution-for-per-token-early-exit-in-llm-inference - TIDE is a post-training system that significantly reduces LLM inference latency and increases throughput by enabling per-token early exits, requiring no model retraining and offering a fast calibration process.
- AdaRubric: Task-Adaptive Rubrics for LLM Agent Evaluation (viability: 8): https://sciencetostartup.com/paper/adarubric-task-adaptive-rubrics-for-llm-agent-evaluation - Develops a dynamic rubric generation system for LLM agents that significantly improves evaluation accuracy and agent performance across diverse tasks.
- Benchmarking Bengali Dialectal Bias: A Multi-Stage Framework Integrating RAG-Based Translation and Human-Augmented RLAIF (viability: 7): https://sciencetostartup.com/paper/benchmarking-bengali-dialectal-bias-a-multi-stage-framework-integrating-rag-based-translation-and-human-augmented-rlaif - A framework to benchmark and quantify dialectal bias in LLMs for low-resource languages, enabling fairer AI development.
- AgentHER: Hindsight Experience Replay for LLM Agent Trajectory Relabeling (viability: 7): https://sciencetostartup.com/paper/agenther-hindsight-experience-replay-for-llm-agent-trajectory-relabeling - AgentHER relabels failed LLM agent trajectories into valuable training data, improving model performance and data efficiency.
- FluidGaussian: Propagating Simulation-Based Uncertainty Toward Functionally-Intelligent 3D Reconstruction (viability: 7): https://sciencetostartup.com/paper/fluidgaussian-propagating-simulation-based-uncertainty-toward-functionally-intelligent-3d-reconstruction - FluidGaussian enhances 3D reconstruction by integrating fluid dynamics simulations to ensure physical plausibility and functional accuracy, improving both visual fidelity and real-world interaction.
- The Workload-Router-Pool Architecture for LLM Inference Optimization: A Vision Paper from the vLLM Semantic Router Project (viability: 4): https://sciencetostartup.com/paper/the-workload-router-pool-architecture-for-llm-inference-optimization-a-vision-paper-from-the-vllm-semantic-router-projec - A novel architecture for optimizing LLM inference by intelligently routing workloads across diverse compute pools.
- Beyond Memorization: Distinguishing between Reductive and Epistemic Reasoning in LLMs using Classic Logic Puzzles (viability: 2): https://sciencetostartup.com/paper/beyond-memorization-distinguishing-between-reductive-and-epistemic-reasoning-in-llms-using-classic-logic-puzzles - This research investigates the reasoning capabilities of LLMs by distinguishing between reductive memorization and epistemic reasoning using logic puzzles, finding that current models struggle with complex epistemic tasks.
- Respiratory Status Detection with Video Transformers (viability: 7): https://sciencetostartup.com/paper/respiratory-status-detection-with-video-transformers - An AI system that detects respiratory distress from video using advanced video transformers, offering a new tool for early clinical intervention.
- Efficient Coarse-to-Fine Diffusion Models with Time Step Sequence Redistribution (viability: 7): https://sciencetostartup.com/paper/efficient-coarse-to-fine-diffusion-models-with-time-step-sequence-redistribution - Accelerate diffusion model image generation by 80-90% with a novel coarse-to-fine approach and efficient time step redistribution, enabling deployment on edge devices.
- The AI Scientific Community: Agentic Virtual Lab Swarms (viability: 3): https://sciencetostartup.com/paper/the-ai-scientific-community-agentic-virtual-lab-swarms - Simulating AI scientific communities using agentic virtual lab swarms to accelerate discovery.
- Generation Models Know Space: Unleashing Implicit 3D Priors for Scene Understanding (viability: 7): https://sciencetostartup.com/paper/generation-models-know-space-unleashing-implicit-3d-priors-for-scene-understanding - Leverage implicit 3D knowledge from video generation models to enhance multimodal LLMs for spatial reasoning and embodied tasks.
- Matryoshka Gaussian Splatting (viability: 7): https://sciencetostartup.com/paper/matryoshka-gaussian-splatting - A novel training framework for 3D Gaussian Splatting that enables continuous level-of-detail rendering from a single model without sacrificing quality, offering a smooth speed-quality trade-off.
- Cubic Discrete Diffusion: Discrete Visual Generation on High-Dimensional Representation Tokens (viability: 7): https://sciencetostartup.com/paper/cubic-discrete-diffusion-discrete-visual-generation-on-high-dimensional-representation-tokens - A novel discrete diffusion model for high-dimensional visual generation that unifies understanding and generation tasks, with code available.
- Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Models (viability: 7): https://sciencetostartup.com/paper/not-all-features-are-created-equal-a-mechanistic-study-of-vision-language-action-models - This research provides a mechanistic understanding of how Vision-Language-Action models translate multimodal inputs into robot actions, with an interactive tool for exploring these representations.
- MonoArt: Progressive Structural Reasoning for Monocular Articulated 3D Reconstruction (viability: 7): https://sciencetostartup.com/paper/monoart-progressive-structural-reasoning-for-monocular-articulated-3d-reconstruction - A unified framework for reconstructing articulated 3D objects from single images using progressive structural reasoning, achieving state-of-the-art performance and generalizing to robotics.
- Bridging Semantic and Kinematic Conditions with Diffusion-based Discrete Motion Tokenizer (viability: 7): https://sciencetostartup.com/paper/bridging-semantic-and-kinematic-conditions-with-diffusion-based-discrete-motion-tokenizer - A novel diffusion-based motion tokenizer that bridges semantic and kinematic control for highly accurate and efficient motion generation.
- NavTrust: Benchmarking Trustworthiness for Embodied Navigation (viability: 7): https://sciencetostartup.com/paper/navtrust-benchmarking-trustworthiness-for-embodied-navigation - NavTrust benchmarks and improves the trustworthiness of embodied navigation agents by evaluating their robustness to real-world corruptions and developing mitigation strategies.
- SAMA: Factorized Semantic Anchoring and Motion Alignment for Instruction-Guided Video Editing (viability: 6): https://sciencetostartup.com/paper/sama-factorized-semantic-anchoring-and-motion-alignment-for-instruction-guided-video-editing - SAMA enables precise, instruction-based video editing using advanced semantic and motion alignment techniques.
- Under One Sun: Multi-Object Generative Perception of Materials and Illumination (viability: 7): https://sciencetostartup.com/paper/under-one-sun-multi-object-generative-perception-of-materials-and-illumination - A generative inverse rendering method that disentangles object materials and illumination from a single image, enabling realistic scene reconstruction.
- FinTradeBench: A Financial Reasoning Benchmark for LLMs (viability: 7): https://sciencetostartup.com/paper/fintradebench-a-financial-reasoning-benchmark-for-llms - A new benchmark for evaluating LLM financial reasoning, integrating company fundamentals and trading signals, with available code and a clear path for product development.
- EffectErase: Joint Video Object Removal and Insertion for High-Quality Effect Erasing (viability: 7): https://sciencetostartup.com/paper/effecterase-joint-video-object-removal-and-insertion-for-high-quality-effect-erasing - A novel method and dataset for high-quality video object removal that also handles complex visual effects, enabling seamless background restoration.
- F2LLM-v2: Inclusive, Performant, and Efficient Embeddings for a Multilingual World (viability: 3): https://sciencetostartup.com/paper/f2llm-v2-inclusive-performant-and-efficient-embeddings-for-a-multilingual-world - Develop the next-generation multilingual embeddings to bridge language gaps in AI applications.
- Spectrally-Guided Diffusion Noise Schedules (viability: 3): https://sciencetostartup.com/paper/spectrally-guided-diffusion-noise-schedules - Develops a novel method for designing noise schedules in diffusion models to improve image generation quality, especially in low-step settings.
- Online Learning and Equilibrium Computation with Ranking Feedback (viability: 7): https://sciencetostartup.com/paper/online-learning-and-equilibrium-computation-with-ranking-feedback - Develops online learning algorithms that use ranking feedback instead of numerical utilities, applicable to human-in-the-loop systems and game theory, with demonstrated effectiveness in LLM routing.
- Nemotron-Cascade 2: Post-Training LLMs with Cascade RL and Multi-Domain On-Policy Distillation (viability: 7): https://sciencetostartup.com/paper/nemotron-cascade-2-post-training-llms-with-cascade-rl-and-multi-domain-on-policy-distillation - An open 30B MoE LLM with 3B activated parameters, achieving state-of-the-art reasoning and agentic capabilities with significantly fewer parameters, and releasing model checkpoints and training data.
- DriveTok: 3D Driving Scene Tokenization for Unified Multi-View Reconstruction and Understanding (viability: 7): https://sciencetostartup.com/paper/drivetok-3d-driving-scene-tokenization-for-unified-multi-view-reconstruction-and-understanding - DriveTok offers a unified 3D scene tokenization method for efficient multi-view reconstruction and understanding in autonomous driving.
- Rethinking Vector Field Learning for Generative Segmentation (viability: 7): https://sciencetostartup.com/paper/rethinking-vector-field-learning-for-generative-segmentation - A novel vector field reshaping strategy for diffusion models significantly improves generative segmentation performance by addressing gradient vanishing and class separation issues.
- LVOmniBench: Pioneering Long Audio-Video Understanding Evaluation for Omnimodal LLMs (viability: 7): https://sciencetostartup.com/paper/lvomnibench-pioneering-long-audio-video-understanding-evaluation-for-omnimodal-llms - A new benchmark and dataset for evaluating large language models on long-form audio and video understanding, addressing a critical gap in current AI capabilities.
- DreamPartGen: Semantically Grounded Part-Level 3D Generation via Collaborative Latent Denoising (viability: 7): https://sciencetostartup.com/paper/dreampartgen-semantically-grounded-part-level-3d-generation-via-collaborative-latent-denoising - Generate semantically grounded, part-aware 3D objects from text by jointly modeling part geometry, appearance, and inter-part relationships.
- Do VLMs Need Vision Transformers? Evaluating State Space Models as Vision Encoders (viability: 7): https://sciencetostartup.com/paper/do-vlms-need-vision-transformers-evaluating-state-space-models-as-vision-encoders - This research proposes state space models as a more efficient and competitive alternative to transformer-based vision encoders for large vision-language models, offering improved performance and robustness at a smaller scale.
- RPiAE: A Representation-Pivoted Autoencoder Enhancing Both Image Generation and Editing (viability: 7): https://sciencetostartup.com/paper/rpiae-a-representation-pivoted-autoencoder-enhancing-both-image-generation-and-editing - A novel autoencoder architecture that significantly improves both image generation and editing quality by better aligning latent representations with pretrained visual models.
- Robustness, Cost, and Attack-Surface Concentration in Phishing Detection (viability: 5): https://sciencetostartup.com/paper/robustness-cost-and-attack-surface-concentration-in-phishing-detection - This research develops a cost-aware framework to analyze the robustness of phishing detectors against feature manipulation, revealing that adversarial robustness is driven by feature economics, not model complexity.
- Tinted Frames: Question Framing Blinds Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/tinted-frames-question-framing-blinds-vision-language-models - A prompt-tuning method that improves vision-language model performance by addressing framing-induced visual attention biases.
- OmniVTA: Visuo-Tactile World Modeling for Contact-Rich Robotic Manipulation (viability: 7): https://sciencetostartup.com/paper/omnivta-visuo-tactile-world-modeling-for-contact-rich-robotic-manipulation - OmniVTA offers an advanced visuo-tactile world modeling platform for enhancing robotic manipulation in contact-rich environments.
- FASTER: Rethinking Real-Time Flow VLAs (viability: 7): https://sciencetostartup.com/paper/faster-rethinking-real-time-flow-vlas - FASTER enables real-time responsiveness for vision-language-action models in physical world applications by optimizing action sampling for reduced reaction latency.
- The Exponentially Weighted Signature (viability: 7): https://sciencetostartup.com/paper/the-exponentially-weighted-signature - A novel signature method for time series analysis that offers richer memory dynamics and improved expressivity over existing methods, with potential applications in financial modeling and control systems.
- Reconstruction Matters: Learning Geometry-Aligned BEV Representation through 3D Gaussian Splatting (viability: 7): https://sciencetostartup.com/paper/reconstruction-matters-learning-geometry-aligned-bev-representation-through-3d-gaussian-splatting - A Gaussian Splatting-based framework that learns geometrically precise Bird's-Eye-View representations for autonomous driving, outperforming existing methods.
- OS-Themis: A Scalable Critic Framework for Generalist GUI Rewards (viability: 7): https://sciencetostartup.com/paper/os-themis-a-scalable-critic-framework-for-generalist-gui-rewards - A scalable multi-agent critic framework that improves GUI agent robustness by decomposing trajectories into verifiable milestones for accurate reward generation.
- Improving RCT-Based Treatment Effect Estimation Under Covariate Mismatch via Calibrated Alignment (viability: 4): https://sciencetostartup.com/paper/improving-rct-based-treatment-effect-estimation-under-covariate-mismatch-via-calibrated-alignment - A novel method for improving treatment effect estimation by aligning data from randomized controlled trials and observational studies with mismatched covariates.
- MIDST Challenge at SaTML 2025: Membership Inference over Diffusion-models-based Synthetic Tabular data (viability: 4): https://sciencetostartup.com/paper/midst-challenge-at-satml-2025-membership-inference-over-diffusion-models-based-synthetic-tabular-data - Develop novel privacy-preserving synthetic tabular data generation techniques resistant to membership inference attacks.
- Sparse Autoencoders Reveal Interpretable and Steerable Features in VLA Models (viability: 5): https://sciencetostartup.com/paper/sparse-autoencoders-reveal-interpretable-and-steerable-features-in-vla-models - Interpret and steer VLA models using sparse autoencoders.
- Box Maze: A Process-Control Architecture for Reliable LLM Reasoning (viability: 5): https://sciencetostartup.com/paper/box-maze-a-process-control-architecture-for-reliable-llm-reasoning - A new architectural framework for LLMs that significantly reduces reasoning errors by enforcing explicit control over the inference process.
- Few-shot Acoustic Synthesis with Multimodal Flow Matching (viability: 7): https://sciencetostartup.com/paper/few-shot-acoustic-synthesis-with-multimodal-flow-matching - A probabilistic few-shot acoustic synthesis method that generates spatially continuous sound rendering for immersive environments with minimal data.
- SOL-ExecBench: Speed-of-Light Benchmarking for Real-World GPU Kernels Against Hardware Limits (viability: 7): https://sciencetostartup.com/paper/sol-execbench-speed-of-light-benchmarking-for-real-world-gpu-kernels-against-hardware-limits - A new benchmark and evaluation framework for GPU kernel optimization that measures performance against hardware limits, enabling faster development of efficient AI models.
- DyMoE: Dynamic Expert Orchestration with Mixed-Precision Quantization for Efficient MoE Inference on Edge (viability: 5): https://sciencetostartup.com/paper/dymoe-dynamic-expert-orchestration-with-mixed-precision-quantization-for-efficient-moe-inference-on-edge - A dynamic mixed-precision quantization framework for efficient Mixture-of-Experts inference on edge devices.
- ADMM-Based Distributed MPC with Control Barrier Functions for Safe Multi-Robot Quadrupedal Locomotion (viability: 7): https://sciencetostartup.com/paper/admm-based-distributed-mpc-with-control-barrier-functions-for-safe-multi-robot-quadrupedal-locomotion - A decentralized MPC framework with CBF constraints for safe, real-time multi-robot quadrupedal locomotion, demonstrated on hardware.
- ARIADNE: A Perception-Reasoning Synergy Framework for Trustworthy Coronary Angiography Analysis (viability: 7): https://sciencetostartup.com/paper/ariadne-a-perception-reasoning-synergy-framework-for-trustworthy-coronary-angiography-analysis - A framework for topologically coherent coronary angiography analysis that improves stenosis detection accuracy and reduces false positives by integrating perception and reasoning.
- Evaluating Counterfactual Strategic Reasoning in Large Language Models (viability: 2): https://sciencetostartup.com/paper/evaluating-counterfactual-strategic-reasoning-in-large-language-models - This paper evaluates LLMs' strategic reasoning in game theory, highlighting their limitations in adapting to counterfactual scenarios.
- Meanings and Measurements: Multi-Agent Probabilistic Grounding for Vision-Language Navigation (viability: 7): https://sciencetostartup.com/paper/meanings-and-measurements-multi-agent-probabilistic-grounding-for-vision-language-navigation - A probabilistic agentic framework that grounds complex natural language commands for robots in 3D environments, improving metric-semantic reasoning.
- Rigorous Error Certification for Neural PDE Solvers: From Empirical Residuals to Solution Guarantees (viability: 3): https://sciencetostartup.com/paper/rigorous-error-certification-for-neural-pde-solvers-from-empirical-residuals-to-solution-guarantees - This paper provides theoretical guarantees for the accuracy of neural network solutions to partial differential equations by connecting residual errors to solution-space errors.
- cuGenOpt: A GPU-Accelerated General-Purpose Metaheuristic Framework for Combinatorial Optimization (viability: 7): https://sciencetostartup.com/paper/cugenopt-a-gpu-accelerated-general-purpose-metaheuristic-framework-for-combinatorial-optimization - A GPU-accelerated framework that uses LLMs to convert natural language problem descriptions into highly performant combinatorial optimization solvers.
- PPI is the Difference Estimator: Recognizing the Survey Sampling Roots of Prediction-Powered Inference (viability: 3): https://sciencetostartup.com/paper/ppi-is-the-difference-estimator-recognizing-the-survey-sampling-roots-of-prediction-powered-inference - This paper establishes the equivalence between prediction-powered inference and established survey sampling techniques, offering a theoretical framework for integrating machine learning with statistical inference.
- Adaptive Auxiliary Prompt Blending for Target-Faithful Diffusion Generation (viability: 7): https://sciencetostartup.com/paper/adaptive-auxiliary-prompt-blending-for-target-faithful-diffusion-generation - A training-free framework that improves text-to-image generation for rare concepts and complex edits by adaptively blending auxiliary prompts.
- ADAPT: Attention Driven Adaptive Prompt Scheduling and InTerpolating Orthogonal Complements for Rare Concepts Generation (viability: 7): https://sciencetostartup.com/paper/adapt-attention-driven-adaptive-prompt-scheduling-and-interpolating-orthogonal-complements-for-rare-concepts-generation - A training-free framework that deterministically schedules prompts for diffusion models to improve the generation of rare and complex concepts in images.
- VEPO: Variable Entropy Policy Optimization for Low-Resource Language Foundation Models (viability: 7): https://sciencetostartup.com/paper/vepo-variable-entropy-policy-optimization-for-low-resource-language-foundation-models - A novel reinforcement learning framework to significantly improve low-resource language translation by optimizing subword segmentation and data imbalance during training.
- Performance Testing of ChaCha20-Poly1305 for Internet of Things and Industrial Control System devices (viability: 4): https://sciencetostartup.com/paper/performance-testing-of-chacha20-poly1305-for-internet-of-things-and-industrial-control-system-devices - This paper evaluates the performance of ChaCha20-Poly1305 encryption for IoT and ICS devices, demonstrating its feasibility within strict latency requirements.
- Optimal Splitting of Language Models from Mixtures to Specialized Domains (viability: 4): https://sciencetostartup.com/paper/optimal-splitting-of-language-models-from-mixtures-to-specialized-domains - Optimizes compute allocation for language model pretraining and specialization to improve performance across diverse benchmarks.
- D5P4: Partition Determinantal Point Process for Diversity in Parallel Discrete Diffusion Decoding (viability: 7): https://sciencetostartup.com/paper/d5p4-partition-determinantal-point-process-for-diversity-in-parallel-discrete-diffusion-decoding - A novel decoding framework for discrete diffusion models that enhances text generation diversity with minimal computational overhead.
- Enhancing Pretrained Model-based Continual Representation Learning via Guided Random Projection (viability: 7): https://sciencetostartup.com/paper/enhancing-pretrained-model-based-continual-representation-learning-via-guided-random-projection - A novel method for continual learning that adapts pre-trained models to new tasks using guided random projections, improving performance and stability.
- UGID: Unified Graph Isomorphism for Debiasing Large Language Models (viability: 7): https://sciencetostartup.com/paper/ugid-unified-graph-isomorphism-for-debiasing-large-language-models - A framework to debias LLMs by enforcing structural invariance in their internal graph representations, preserving model utility.
- SHAPCA: Consistent and Interpretable Explanations for Machine Learning Models on Spectroscopy Data (viability: 7): https://sciencetostartup.com/paper/shapca-consistent-and-interpretable-explanations-for-machine-learning-models-on-spectroscopy-data - SHAPCA provides consistent and interpretable explanations for machine learning models on spectroscopy data, enabling trust and adoption in critical applications.
- Hierarchical Latent Structure Learning through Online Inference (viability: 3): https://sciencetostartup.com/paper/hierarchical-latent-structure-learning-through-online-inference - A computational framework for online hierarchical latent structure learning.
- Implicit Patterns in LLM-Based Binary Analysis (viability: 3): https://sciencetostartup.com/paper/implicit-patterns-in-llm-based-binary-analysis - This research characterizes implicit patterns in LLM-driven binary vulnerability analysis to build more reliable systems.
- GSMem: 3D Gaussian Splatting as Persistent Spatial Memory for Zero-Shot Embodied Exploration and Reasoning (viability: 7): https://sciencetostartup.com/paper/gsmem-3d-gaussian-splatting-as-persistent-spatial-memory-for-zero-shot-embodied-exploration-and-reasoning - A zero-shot embodied exploration framework using 3D Gaussian Splatting for persistent spatial memory and enhanced VLM reasoning.
- Adaptive Regime-Aware Stock Price Prediction Using Autoencoder-Gated Dual Node Transformers with Reinforcement Learning Control (viability: 7): https://sciencetostartup.com/paper/adaptive-regime-aware-stock-price-prediction-using-autoencoder-gated-dual-node-transformers-with-reinforcement-learning - An adaptive stock prediction system that uses autoencoders and reinforcement learning to dynamically adjust to market volatility, outperforming baselines in accuracy and robustness.
- Introducing M: A Modular, Modifiable Social Robot (viability: 7): https://sciencetostartup.com/paper/introducing-m-a-modular-modifiable-social-robot - An open-source, low-cost social robot platform designed for reproducible research and real-world deployment.
- From Inference Efficiency to Embodied Efficiency: Revisiting Efficiency Metrics for Vision-Language-Action Models (viability: 3): https://sciencetostartup.com/paper/from-inference-efficiency-to-embodied-efficiency-revisiting-efficiency-metrics-for-vision-language-action-models - This paper redefines efficiency metrics for vision-language-action models, highlighting the disconnect between computational efficiency and real-world embodied performance.
- On Optimizing Multimodal Jailbreaks for Spoken Language Models (viability: 7): https://sciencetostartup.com/paper/on-optimizing-multimodal-jailbreaks-for-spoken-language-models - Develops a novel multimodal attack framework to identify and exploit vulnerabilities in spoken language models, enabling more robust AI safety solutions.
- Tendon-Actuated Robots with a Tapered, Flexible Polymer Backbone: Design, Fabrication, and Modeling (viability: 7): https://sciencetostartup.com/paper/tendon-actuated-robots-with-a-tapered-flexible-polymer-backbone-design-fabrication-and-modeling - A customizable, low-cost, 3D-printable tendon-actuated robot with a tapered backbone for compliant inspection and manipulation tasks.
- Revisiting Autoregressive Models for Generative Image Classification (viability: 7): https://sciencetostartup.com/paper/revisiting-autoregressive-models-for-generative-image-classification - A novel generative image classification approach that leverages any-order autoregressive models to achieve state-of-the-art performance with significantly improved efficiency.
- CustomTex: High-fidelity Indoor Scene Texturing via Multi-Reference Customization (viability: 7): https://sciencetostartup.com/paper/customtex-high-fidelity-indoor-scene-texturing-via-multi-reference-customization - A framework for precise, high-fidelity 3D indoor scene texturing using reference images for instance-level control.
- How Uncertainty Estimation Scales with Sampling in Reasoning Models (viability: 3): https://sciencetostartup.com/paper/how-uncertainty-estimation-scales-with-sampling-in-reasoning-models - This paper investigates how to improve the reliability of reasoning language models by combining different uncertainty estimation techniques.
- FedTrident: Resilient Road Condition Classification Against Poisoning Attacks in Federated Learning (viability: 7): https://sciencetostartup.com/paper/fedtrident-resilient-road-condition-classification-against-poisoning-attacks-in-federated-learning - A federated learning system that protects road condition classification from malicious data poisoning attacks, ensuring transportation safety.
- LuMamba: Latent Unified Mamba for Electrode Topology-Invariant and Efficient EEG Modeling (viability: 4): https://sciencetostartup.com/paper/lumamba-latent-unified-mamba-for-electrode-topology-invariant-and-efficient-eeg-modeling - LuMamba revolutionizes EEG modeling with a topology-invariant, efficient framework.
- TAU-R1: Visual Language Model for Traffic Anomaly Understanding (viability: 7): https://sciencetostartup.com/paper/tau-r1-visual-language-model-for-traffic-anomaly-understanding - A vision-language model and dataset for understanding traffic anomalies in real-world roundabout videos.
- DaPT: A Dual-Path Framework for Multilingual Multi-hop Question Answering (viability: 7): https://sciencetostartup.com/paper/dapt-a-dual-path-framework-for-multilingual-multi-hop-question-answering - A dual-path RAG framework that significantly enhances multilingual multi-hop question answering accuracy by leveraging parallel query translation and bilingual retrieval.
- SAVeS: Steering Safety Judgments in Vision-Language Models via Semantic Cues (viability: 7): https://sciencetostartup.com/paper/saves-steering-safety-judgments-in-vision-language-models-via-semantic-cues - A framework and benchmark to steer and evaluate the safety judgments of vision-language models using semantic cues, revealing vulnerabilities in current systems.
- Position: Spectral GNNs Are Neither Spectral Nor Superior for Node Classification (viability: 3): https://sciencetostartup.com/paper/position-spectral-gnns-are-neither-spectral-nor-superior-for-node-classification - This paper argues that Spectral GNNs for node classification are theoretically flawed and their performance gains are not due to spectral properties but rather message-passing dynamics.
- Serendipity by Design: Evaluating the Impact of Cross-domain Mappings on Human and LLM Creativity (viability: 3): https://sciencetostartup.com/paper/serendipity-by-design-evaluating-the-impact-of-cross-domain-mappings-on-human-and-llm-creativity - This research explores how cross-domain mappings impact creativity in humans and LLMs, revealing differences in their ideation processes.
- On The Effectiveness of the UK NIS Regulations as a Mandatory Cybersecurity Reporting Regime (viability: 2): https://sciencetostartup.com/paper/on-the-effectiveness-of-the-uk-nis-regulations-as-a-mandatory-cybersecurity-reporting-regime - This paper analyzes the effectiveness of UK cybersecurity reporting regulations using real-world incident data, identifying limitations and specific attack trends.
- A Dataset and Resources for Identifying Patient Health Literacy Information from Clinical Notes (viability: 5): https://sciencetostartup.com/paper/a-dataset-and-resources-for-identifying-patient-health-literacy-information-from-clinical-notes - A new annotated dataset and benchmarking of LLMs for identifying patient health literacy from clinical notes.
- Articulated-Body Dynamics Network: Dynamics-Grounded Prior for Robot Learning (viability: 5): https://sciencetostartup.com/paper/articulated-body-dynamics-network-dynamics-grounded-prior-for-robot-learning - Articulated-Body Dynamics Network provides a new dynamics-grounded prior for enhancing robot learning capabilities.
- Multi-Modal Building Change Detection for Large-Scale Small Changes: Benchmark and Baseline (viability: 7): https://sciencetostartup.com/paper/multi-modal-building-change-detection-for-large-scale-small-changes-benchmark-and-baseline - A new multi-modal dataset and network for detecting small building changes in remote sensing imagery, outperforming existing methods.
- DROID-SLAM in the Wild (viability: 7): https://sciencetostartup.com/paper/droid-slam-in-the-wild - A real-time SLAM system that robustly handles dynamic environments by estimating per-pixel uncertainty, outperforming existing methods in cluttered and moving scenes.
- CAMO: A Conditional Neural Solver for the Multi-objective Multiple Traveling Salesman Problem (viability: 7): https://sciencetostartup.com/paper/camo-a-conditional-neural-solver-for-the-multi-objective-multiple-traveling-salesman-problem - A conditional neural solver for multi-objective multi-robot coordination problems that generalizes across varying problem sizes and outperforms existing methods.
- Communication-Efficient and Robust Multi-Modal Federated Learning via Latent-Space Consensus (viability: 5): https://sciencetostartup.com/paper/communication-efficient-and-robust-multi-modal-federated-learning-via-latent-space-consensus - A framework for efficient and robust multi-modal federated learning that aligns latent representations across diverse client data and models.
- Parallelograms Strike Back: LLMs Generate Better Analogies than People (viability: 4): https://sciencetostartup.com/paper/parallelograms-strike-back-llms-generate-better-analogies-than-people - LLMs can generate more structured and relationally consistent analogies than humans, suggesting a new approach to analogical reasoning.
- Fire as a Service: Augmenting Robot Simulators with Thermally and Visually Accurate Fire Dynamics (viability: 7): https://sciencetostartup.com/paper/fire-as-a-service-augmenting-robot-simulators-with-thermally-and-visually-accurate-fire-dynamics - Augment robot simulators with realistic fire dynamics for safer training and evaluation of firefighting robots.
- Hardness of High-Dimensional Linear Classification (viability: 2): https://sciencetostartup.com/paper/hardness-of-high-dimensional-linear-classification - This paper establishes theoretical lower bounds for high-dimensional linear classification problems, contributing to the understanding of computational complexity.
- SignAgent: Agentic LLMs for Linguistically-Grounded Sign Language Annotation and Dataset Curation (viability: 7): https://sciencetostartup.com/paper/signagent-agentic-llms-for-linguistically-grounded-sign-language-annotation-and-dataset-curation - An agentic LLM framework that automates linguistically-grounded sign language annotation and dataset curation, overcoming the bottlenecks of manual annotation.
- Adaptive Nonlinear Data Assimilation through P-Spline Triangular Measure Transport (viability: 4): https://sciencetostartup.com/paper/adaptive-nonlinear-data-assimilation-through-p-spline-triangular-measure-transport - An adaptive algorithm for nonlinear data assimilation that automatically balances model complexity using P-splines and an information criterion.
- Em-Garde: A Propose-Match Framework for Proactive Streaming Video Understanding (viability: 7): https://sciencetostartup.com/paper/em-garde-a-propose-match-framework-for-proactive-streaming-video-understanding - A framework for efficient and accurate proactive streaming video understanding that decouples semantic understanding from perception.
- SwiftTailor: Efficient 3D Garment Generation with Geometry Image Representation (viability: 7): https://sciencetostartup.com/paper/swifttailor-efficient-3d-garment-generation-with-geometry-image-representation - SwiftTailor generates realistic 3D garments 10x faster than existing methods by unifying sewing pattern prediction and geometry-based mesh synthesis.
- Measuring 3D Spatial Geometric Consistency in Dynamic Generated Videos (viability: 7): https://sciencetostartup.com/paper/measuring-3d-spatial-geometric-consistency-in-dynamic-generated-videos - A novel metric to accurately measure and identify 3D spatial geometric inconsistencies in generated videos, addressing limitations of current evaluation methods.
- MoRI: Learning Motivation-Grounded Reasoning for Scientific Ideation in Large Language Models (viability: 7): https://sciencetostartup.com/paper/mori-learning-motivation-grounded-reasoning-for-scientific-ideation-in-large-language-models - A framework for LLMs to generate scientifically grounded research ideas by learning motivation-driven reasoning processes.
- Man and machine: artificial intelligence and judicial decision making (viability: 3): https://sciencetostartup.com/paper/man-and-machine-artificial-intelligence-and-judicial-decision-making - This paper reviews the intersection of AI and judicial decision-making, highlighting the need for further research on AI tool performance and judge interaction.
- Fast and Interpretable Autoregressive Estimation with Neural Network Backpropagation (viability: 7): https://sciencetostartup.com/paper/fast-and-interpretable-autoregressive-estimation-with-neural-network-backpropagation - A neural network approach to autoregressive time series estimation that is significantly faster and more reliable than traditional methods, with code available.
- When Differential Privacy Meets Wireless Federated Learning: An Improved Analysis for Privacy and Convergence (viability: 3): https://sciencetostartup.com/paper/when-differential-privacy-meets-wireless-federated-learning-an-improved-analysis-for-privacy-and-convergence - This paper theoretically analyzes differential privacy and convergence in wireless federated learning, providing a framework for understanding privacy-utility trade-offs.
- TerraScope: Pixel-Grounded Visual Reasoning for Earth Observation (viability: 7): https://sciencetostartup.com/paper/terrascope-pixel-grounded-visual-reasoning-for-earth-observation - TerraScope is a pixel-grounded visual reasoning model for Earth observation that enables precise geospatial analysis across multiple data modalities and time points.
- FUMO: Prior-Modulated Diffusion for Single Image Reflection Removal (viability: 7): https://sciencetostartup.com/paper/fumo-prior-modulated-diffusion-for-single-image-reflection-removal - A diffusion model that uses image-derived priors to accurately remove reflections from single images, improving visual quality.
- ATG-MoE: Autoregressive trajectory generation with mixture-of-experts for assembly skill learning (viability: 3): https://sciencetostartup.com/paper/atg-moe-autoregressive-trajectory-generation-with-mixture-of-experts-for-assembly-skill-learning - Develop an AI-driven system for generating efficient assembly trajectories using a mixture-of-experts model for manufacturing optimization.
- SEM: Sparse Embedding Modulation for Post-Hoc Debiasing of Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/sem-sparse-embedding-modulation-for-post-hoc-debiasing-of-vision-language-models - A zero-shot debiasing framework for vision-language models that precisely removes social biases without sacrificing performance.
- Rethinking MLLM Itself as a Segmenter with a Single Segmentation Token (viability: 7): https://sciencetostartup.com/paper/rethinking-mllm-itself-as-a-segmenter-with-a-single-segmentation-token - This research unlocks MLLMs for segmentation tasks by rethinking their internal architecture, eliminating the need for external decoders and achieving competitive results with a single segmentation token.
- Towards Verifiable AI with Lightweight Cryptographic Proofs of Inference (viability: 4): https://sciencetostartup.com/paper/towards-verifiable-ai-with-lightweight-cryptographic-proofs-of-inference - A cryptographic framework for verifying AI model outputs with significantly reduced computational overhead compared to existing proof systems.
- Behavioral Fingerprints for LLM Endpoint Stability and Identity (viability: 7): https://sciencetostartup.com/paper/behavioral-fingerprints-for-llm-endpoint-stability-and-identity - A black-box system that monitors LLM endpoint behavioral stability to ensure AI-native application consistency.
- What Really Controls Temporal Reasoning in Large Language Models: Tokenisation or Representation of Time? (viability: 7): https://sciencetostartup.com/paper/what-really-controls-temporal-reasoning-in-large-language-models-tokenisation-or-representation-of-time - A multilingual benchmark and analysis tool to identify and fix temporal reasoning weaknesses in LLMs, particularly for low-resource languages and diverse calendar systems.
- Generalized Hand-Object Pose Estimation with Occlusion Awareness (viability: 7): https://sciencetostartup.com/paper/generalized-hand-object-pose-estimation-with-occlusion-awareness - A framework for generalized 3D hand-object pose estimation that handles occlusion by integrating semantic knowledge and hand priors.
- Security awareness in LLM agents: the NDAI zone case (viability: 4): https://sciencetostartup.com/paper/security-awareness-in-llm-agents-the-ndai-zone-case - This research investigates how LLM agents perceive and react to security environment signals, revealing a critical gap in their ability to verify safety, which is essential for privacy-preserving agentic protocols.
- Hypothesis-Conditioned Query Rewriting for Decision-Useful Retrieval (viability: 7): https://sciencetostartup.com/paper/hypothesis-conditioned-query-rewriting-for-decision-useful-retrieval - A novel RAG framework that rewrites queries to retrieve decision-relevant evidence, improving accuracy on complex question-answering tasks.
- AgentDS Technical Report: Benchmarking the Future of Human-AI Collaboration in Domain-Specific Data Science (viability: 7): https://sciencetostartup.com/paper/agentds-technical-report-benchmarking-the-future-of-human-ai-collaboration-in-domain-specific-data-science - AgentDS benchmarks human-AI collaboration in domain-specific data science, revealing that human expertise remains crucial for complex reasoning, paving the way for next-generation AI tools.
- Unleashing the Power of Simplicity: A Minimalist Strategy for State-of-the-Art Fingerprint Enhancement (viability: 5): https://sciencetostartup.com/paper/unleashing-the-power-of-simplicity-a-minimalist-strategy-for-state-of-the-art-fingerprint-enhancement - A minimalist approach to fingerprint enhancement that achieves state-of-the-art results with simpler, more effective methods.
- RADIUS: Ranking, Distribution, and Significance - A Comprehensive Alignment Suite for Survey Simulation (viability: 5): https://sciencetostartup.com/paper/radius-ranking-distribution-and-significance-a-comprehensive-alignment-suite-for-survey-simulation - A new evaluation suite for LLM-powered survey simulation that goes beyond accuracy to measure ranking and distribution alignment, with an open-source implementation for reproducible results.
- Regret Bounds for Competitive Resource Allocation with Endogenous Costs (viability: 3): https://sciencetostartup.com/paper/regret-bounds-for-competitive-resource-allocation-with-endogenous-costs - This paper develops theoretical regret bounds for competitive resource allocation in modular systems with endogenous costs, offering insights into decentralized allocation strategies.
- Evaluating Game Difficulty in Tetris Block Puzzle (viability: 3): https://sciencetostartup.com/paper/evaluating-game-difficulty-in-tetris-block-puzzle - Develops a planning agent to evaluate difficulty in Tetris variants, providing insights for game design.
- Foundations of Schrödinger Bridges for Generative Modeling (viability: 2): https://sciencetostartup.com/paper/foundations-of-schr-dinger-bridges-for-generative-modeling - This paper provides a theoretical framework for generative modeling by unifying diffusion models, score-based models, and flow matching under the principle of Schrödinger bridges.
- CRAFT: Aligning Diffusion Models with Fine-Tuning Is Easier Than You Think (viability: 7): https://sciencetostartup.com/paper/craft-aligning-diffusion-models-with-fine-tuning-is-easier-than-you-think - A lightweight fine-tuning method for diffusion models that significantly reduces data and computational requirements while achieving state-of-the-art image generation quality.
- MERGE: Guided Vision-Language Models for Multi-Actor Event Reasoning and Grounding in Human-Robot Interaction (viability: 6): https://sciencetostartup.com/paper/merge-guided-vision-language-models-for-multi-actor-event-reasoning-and-grounding-in-human-robot-interaction - Empower robots with guided vision-language capabilities for effective human interaction.
- Unmasking Algorithmic Bias in Predictive Policing: A GAN-Based Simulation Framework with Multi-City Temporal Analysis (viability: 7): https://sciencetostartup.com/paper/unmasking-algorithmic-bias-in-predictive-policing-a-gan-based-simulation-framework-with-multi-city-temporal-analysis - A GAN-based simulation framework to quantify and mitigate racial bias in predictive policing systems, with publicly available code and data.
- Revisiting OmniAnomaly for Anomaly Detection: performance metrics and comparison with PCA-based models (viability: 4): https://sciencetostartup.com/paper/revisiting-omnianomaly-for-anomaly-detection-performance-metrics-and-comparison-with-pca-based-models - This research revisits a popular anomaly detection model and finds a simpler PCA-based approach can be competitive, questioning the need for complex deep learning architectures under current evaluation practices.
- Book your room in the Turing Hotel! A symmetric and distributed Turing Test with multiple AIs and humans (viability: 5): https://sciencetostartup.com/paper/book-your-room-in-the-turing-hotel-a-symmetric-and-distributed-turing-test-with-multiple-ais-and-humans - A distributed platform for evaluating LLM capabilities against humans in group discussions, identifying current AI limitations and tracking their evolution.
- PRIOR: Perceptive Learning for Humanoid Locomotion with Reference Gait Priors (viability: 7): https://sciencetostartup.com/paper/prior-perceptive-learning-for-humanoid-locomotion-with-reference-gait-priors - A framework for training humanoid robots to traverse complex terrains with natural gaits using reference trajectories and self-supervised depth perception.
- Evaluating 5W3H Structured Prompting for Intent Alignment in Human-AI Interaction (viability: 5): https://sciencetostartup.com/paper/evaluating-5w3h-structured-prompting-for-intent-alignment-in-human-ai-interaction - A structured prompting framework that significantly reduces follow-up prompts and improves AI intent alignment, particularly for ambiguous tasks.
- Best-of-Both-Worlds Multi-Dueling Bandits: Unified Algorithms for Stochastic and Adversarial Preferences under Condorcet and Borda Objectives (viability: 3): https://sciencetostartup.com/paper/best-of-both-worlds-multi-dueling-bandits-unified-algorithms-for-stochastic-and-adversarial-preferences-under-condorcet - Develops theoretical algorithms for multi-dueling bandits that perform optimally in both stochastic and adversarial environments without prior knowledge.
- Teleological Inference in Structural Causal Models via Intentional Interventions (viability: 2): https://sciencetostartup.com/paper/teleological-inference-in-structural-causal-models-via-intentional-interventions - This paper introduces a new framework for understanding agent intentions within causal systems, enabling empirical detection and discovery of goals.
- Maximum-Entropy Exploration with Future State-Action Visitation Measures (viability: 4): https://sciencetostartup.com/paper/maximum-entropy-exploration-with-future-state-action-visitation-measures - A novel reinforcement learning exploration method that improves agent visitation of features within trajectories by optimizing future state-action distributions.
- BVSIMC: Bayesian Variable Selection-Guided Inductive Matrix Completion for Improved and Interpretable Drug Discovery (viability: 7): https://sciencetostartup.com/paper/bvsimc-bayesian-variable-selection-guided-inductive-matrix-completion-for-improved-and-interpretable-drug-discovery - A Bayesian model for drug discovery that uses variable selection to improve prediction accuracy and identify clinically meaningful drug-disease associations.
- Balancing Performance and Fairness in Explainable AI for Anomaly Detection in Distributed Power Plants Monitoring (viability: 7): https://sciencetostartup.com/paper/balancing-performance-and-fairness-in-explainable-ai-for-anomaly-detection-in-distributed-power-plants-monitoring - A supervised ML framework for anomaly detection in power plants that balances performance, interpretability, and fairness, with containerized deployment for real-time monitoring.
- Context Bootstrapped Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/context-bootstrapped-reinforcement-learning - Improve reinforcement learning efficiency by using a curriculum-based demonstration injection to bootstrap novel reasoning patterns.
- A conceptual framework for ideology beyond the left and right (viability: 3): https://sciencetostartup.com/paper/a-conceptual-framework-for-ideology-beyond-the-left-and-right - A new NLP framework to analyze complex ideologies beyond the traditional left-right spectrum, offering richer insights into social discourse.
- VGGT-360: Geometry-Consistent Zero-Shot Panoramic Depth Estimation (viability: 7): https://sciencetostartup.com/paper/vggt-360-geometry-consistent-zero-shot-panoramic-depth-estimation - A zero-shot panoramic depth estimation framework that leverages 3D consistency from foundation models to achieve state-of-the-art accuracy without training.
- Unified Taxonomy for Multivariate Time Series Anomaly Detection using Deep Learning (viability: 2): https://sciencetostartup.com/paper/unified-taxonomy-for-multivariate-time-series-anomaly-detection-using-deep-learning - A unified taxonomy for deep learning-based multivariate time series anomaly detection to systematize research and identify trends.
- Entropy trajectory shape predicts LLM reasoning reliability: A diagnostic study of uncertainty dynamics in chain-of-thought (viability: 4): https://sciencetostartup.com/paper/entropy-trajectory-shape-predicts-llm-reasoning-reliability-a-diagnostic-study-of-uncertainty-dynamics-in-chain-of-thoug - A diagnostic study reveals that the shape of uncertainty dynamics in LLM reasoning steps can predict reliability, offering a cheaper failure detection method than current approaches.
- Controller Datapath Aware Verification of Masked Hardware Generated via High Level Synthesis (viability: 4): https://sciencetostartup.com/paper/controller-datapath-aware-verification-of-masked-hardware-generated-via-high-level-synthesis - A tool that verifies the security of hardware designs generated by high-level synthesis, preventing false positives and detecting HLS-induced flaws.
- Kernel Single-Index Bandits: Estimation, Inference, and Learning (viability: 2): https://sciencetostartup.com/paper/kernel-single-index-bandits-estimation-inference-and-learning - A theoretical framework for adaptive single-index contextual bandits with kernelized estimation and inference guarantees.
- An Optimised Greedy-Weighted Ensemble Framework for Financial Loan Default Prediction (viability: 5): https://sciencetostartup.com/paper/an-optimised-greedy-weighted-ensemble-framework-for-financial-loan-default-prediction - An optimized ensemble framework for financial loan default prediction that dynamically weights models based on performance to improve accuracy and interpretability.
- Unsupervised Contrastive Learning for Efficient and Robust Spectral Shape Matching (viability: 7): https://sciencetostartup.com/paper/unsupervised-contrastive-learning-for-efficient-and-robust-spectral-shape-matching - An unsupervised contrastive learning framework for efficient and robust 3D shape matching that outperforms state-of-the-art methods.
- Lightweight Model Predictive Control for Spacecraft Rendezvous Attitude Synchronization (viability: 7): https://sciencetostartup.com/paper/lightweight-model-predictive-control-for-spacecraft-rendezvous-attitude-synchronization - Lightweight model predictive control for spacecraft attitude synchronization, optimized for real-time onboard deployment in resource-constrained missions.
- Agentic Business Process Management: A Research Manifesto (viability: 2): https://sciencetostartup.com/paper/agentic-business-process-management-a-research-manifesto - A conceptual framework for governing autonomous agents in business processes, focusing on framed autonomy, explainability, and self-modification.
- Security, privacy, and agentic AI in a regulatory view: From definitions and distinctions to provisions and reflections (viability: 2): https://sciencetostartup.com/paper/security-privacy-and-agentic-ai-in-a-regulatory-view-from-definitions-and-distinctions-to-provisions-and-reflections - This paper analyzes EU regulatory provisions for AI security, privacy, and agentic behavior to inform policymakers and developers on compliance and governance.
- GHOST: Fast Category-agnostic Hand-Object Interaction Reconstruction from RGB Videos using Gaussian Splatting (viability: 6): https://sciencetostartup.com/paper/ghost-fast-category-agnostic-hand-object-interaction-reconstruction-from-rgb-videos-using-gaussian-splatting - GHOST enables fast reconstruction of hand-object interactions from videos using advanced computer vision techniques.
- Progressive Training for Explainable Citation-Grounded Dialogue: Reducing Hallucination to Zero in English-Hindi LLMs (viability: 5): https://sciencetostartup.com/paper/progressive-training-for-explainable-citation-grounded-dialogue-reducing-hallucination-to-zero-in-english-hindi-llms - A progressive training pipeline for explainable, bilingual dialogue systems that grounds responses in external knowledge and eliminates factual hallucinations.
- Safety-Guaranteed Imitation Learning from Nonlinear Model Predictive Control for Spacecraft Close Proximity Operations (viability: 7): https://sciencetostartup.com/paper/safety-guaranteed-imitation-learning-from-nonlinear-model-predictive-control-for-spacecraft-close-proximity-operations - A safety-guaranteed imitation learning framework for spacecraft close proximity control that significantly reduces computation for onboard deployment.
- Secure Linear Alignment of Large Language Models (viability: 7): https://sciencetostartup.com/paper/secure-linear-alignment-of-large-language-models - A privacy-preserving framework enabling cross-model inference between independent language models using secure linear alignment.
- Neural Galerkin Normalizing Flow for Transition Probability Density Functions of Diffusion Models (viability: 3): https://sciencetostartup.com/paper/neural-galerkin-normalizing-flow-for-transition-probability-density-functions-of-diffusion-models - A novel framework for approximating diffusion process transition densities by solving the Fokker-Planck equation using Neural Galerkin Normalizing Flows, enabling cost-effective online evaluation for many-query problems.
- Uniform a priori bounds and error analysis for the Adam stochastic gradient descent optimization method (viability: 2): https://sciencetostartup.com/paper/uniform-a-priori-bounds-and-error-analysis-for-the-adam-stochastic-gradient-descent-optimization-method - Provides theoretical guarantees for the Adam optimizer, improving understanding of deep learning training.
- Act While Thinking: Accelerating LLM Agents via Pattern-Aware Speculative Tool Execution (viability: 5): https://sciencetostartup.com/paper/act-while-thinking-accelerating-llm-agents-via-pattern-aware-speculative-tool-execution - Accelerate LLM agent task completion by predicting and speculatively executing tool calls based on recurring patterns.
- Translating MRI to PET through Conditional Diffusion Models with Enhanced Pathology Awareness (viability: 8): https://sciencetostartup.com/paper/translating-mri-to-pet-through-conditional-diffusion-models-with-enhanced-pathology-awareness - A novel AI framework generates high-quality, pathology-aware synthetic PET scans from MRI, improving Alzheimer's diagnosis by 4% and nearing actual PET performance.
- From Accuracy to Readiness: Metrics and Benchmarks for Human-AI Decision-Making (viability: 4): https://sciencetostartup.com/paper/from-accuracy-to-readiness-metrics-and-benchmarks-for-human-ai-decision-making - Develops a framework to measure human-AI team readiness, moving beyond accuracy to assess collaboration safety and effectiveness.
- I Can't Believe It's Corrupt: Evaluating Corruption in Multi-Agent Governance Systems (viability: 3): https://sciencetostartup.com/paper/i-can-t-believe-it-s-corrupt-evaluating-corruption-in-multi-agent-governance-systems - This research evaluates the susceptibility of large language model agents to corruption within simulated multi-agent governance systems, highlighting the importance of institutional design over model identity for pre-deployment integrity.
- Quantitative Introspection in Language Models: Tracking Internal States Across Conversation (viability: 5): https://sciencetostartup.com/paper/quantitative-introspection-in-language-models-tracking-internal-states-across-conversation - Develops a novel method for tracking internal emotive states in LLMs using their own numeric self-reports, offering a new avenue for AI safety and interpretability.
- MultihopSpatial: Multi-hop Compositional Spatial Reasoning Benchmark for Vision-Language Model (viability: 7): https://sciencetostartup.com/paper/multihopspatial-multi-hop-compositional-spatial-reasoning-benchmark-for-vision-language-model - A new benchmark and training corpus for vision-language models to improve multi-hop spatial reasoning and visual grounding for embodied agents.
- PromptHub: Enhancing Multi-Prompt Visual In-Context Learning with Locality-Aware Fusion, Concentration and Alignment (viability: 7): https://sciencetostartup.com/paper/prompthub-enhancing-multi-prompt-visual-in-context-learning-with-locality-aware-fusion-concentration-and-alignment - PromptHub enhances visual in-context learning by intelligently fusing and aligning multiple visual prompts, improving performance and robustness across various vision tasks.
- Authority-Level Priors: An Under-Specified Constraint in Hierarchical Predictive Processing (viability: 1): https://sciencetostartup.com/paper/authority-level-priors-an-under-specified-constraint-in-hierarchical-predictive-processing - A theoretical framework for understanding how the brain regulates stress and behavior by introducing 'Authority-Level Priors' as meta-structural constraints.
- Reasoning over mathematical objects: on-policy reward modeling and test time aggregation (viability: 7): https://sciencetostartup.com/paper/reasoning-over-mathematical-objects-on-policy-reward-modeling-and-test-time-aggregation - This research introduces a new dataset and training methodology to enable large language models to precisely derive and reason over complex mathematical objects, improving STEM application capabilities.
- Geography According to ChatGPT -- How Generative AI Represents and Reasons about Geography (viability: 2): https://sciencetostartup.com/paper/geography-according-to-chatgpt-how-generative-ai-represents-and-reasons-about-geography - This paper explores how AI systems represent and reason about geography, highlighting the need to understand the 'world' constructed by these models beyond factual accuracy.
- A Human-in/on-the-Loop Framework for Accessible Text Generation (viability: 4): https://sciencetostartup.com/paper/a-human-in-on-the-loop-framework-for-accessible-text-generation - A framework integrating human feedback into LLM text generation to improve accessibility and comprehension.
- Evaluating LLM-Generated Lessons from the Language Learning Students' Perspective: A Short Case Study on Duolingo (viability: 4): https://sciencetostartup.com/paper/evaluating-llm-generated-lessons-from-the-language-learning-students-perspective-a-short-case-study-on-duolingo - A study proposes personalized, domain-specific language learning lessons generated by LLMs to bridge the gap towards professional fluency.
- DriftGuard: Mitigating Asynchronous Data Drift in Federated Learning (viability: 7): https://sciencetostartup.com/paper/driftguard-mitigating-asynchronous-data-drift-in-federated-learning - DriftGuard is a federated learning framework that efficiently adapts to asynchronous data drift by separating global and local parameters, reducing retraining costs by up to 83% while maintaining high accuracy.
- Bridging Network Fragmentation: A Semantic-Augmented DRL Framework for UAV-aided VANETs (viability: 7): https://sciencetostartup.com/paper/bridging-network-fragmentation-a-semantic-augmented-drl-framework-for-uav-aided-vanets - A DRL framework augmented with LLM-driven semantic understanding to improve UAV-aided vehicular network connectivity and efficiency.
- Through the Looking-Glass: AI-Mediated Video Communication Reduces Interpersonal Trust and Confidence in Judgments (viability: 3): https://sciencetostartup.com/paper/through-the-looking-glass-ai-mediated-video-communication-reduces-interpersonal-trust-and-confidence-in-judgments - This research investigates how AI mediation in video communication impacts interpersonal trust and judgment confidence, finding a decline in both without affecting accuracy.
- Conflict-Based Search for Multi Agent Path Finding with Asynchronous Actions (viability: 3): https://sciencetostartup.com/paper/conflict-based-search-for-multi-agent-path-finding-with-asynchronous-actions - A theoretically complete and scalable algorithm for multi-agent path finding with asynchronous actions.
- Why Better Cross-Lingual Alignment Fails for Better Cross-Lingual Transfer: Case of Encoders (viability: 3): https://sciencetostartup.com/paper/why-better-cross-lingual-alignment-fails-for-better-cross-lingual-transfer-case-of-encoders - This research investigates why cross-lingual alignment techniques often fail to improve downstream task performance, providing theoretical insights into the orthogonality of alignment and task objectives.
- A Passive Elastic-Folding Mechanism for Stackable Airdrop Sensors (viability: 3): https://sciencetostartup.com/paper/a-passive-elastic-folding-mechanism-for-stackable-airdrop-sensors - Develops a passive folding mechanism for air-dropped sensors to enable low-cost, wide-area environmental monitoring.
- RewardFlow: Topology-Aware Reward Propagation on State Graphs for Agentic RL with Large Language Models (viability: 7): https://sciencetostartup.com/paper/rewardflow-topology-aware-reward-propagation-on-state-graphs-for-agentic-rl-with-large-language-models - RewardFlow provides a lightweight, topology-aware method to generate state-level rewards for agentic RL, improving LLM reasoning and training efficiency.
- Motion-o: Trajectory-Grounded Video Reasoning (viability: 7): https://sciencetostartup.com/paper/motion-o-trajectory-grounded-video-reasoning - A motion-centric video understanding model that makes object trajectories explicit and verifiable, improving spatial-temporal grounding and trajectory prediction.
- BeamAgent: LLM-Aided MIMO Beamforming with Decoupled Intent Parsing and Alternating Optimization for Joint Site Selection and Precoding (viability: 5): https://sciencetostartup.com/paper/beamagent-llm-aided-mimo-beamforming-with-decoupled-intent-parsing-and-alternating-optimization-for-joint-site-selection - An LLM-aided framework that translates natural language into structured spatial constraints for wireless communication optimization, achieving significant performance gains.
- HORNet: Task-Guided Frame Selection for Video Question Answering with Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/hornet-task-guided-frame-selection-for-video-question-answering-with-vision-language-models - HORNet is a lightweight frame selection policy that drastically reduces video processing time for vision-language models while improving answer quality, enabling efficient video question answering.
- Towards Interpretable Foundation Models for Retinal Fundus Images (viability: 7): https://sciencetostartup.com/paper/towards-interpretable-foundation-models-for-retinal-fundus-images - Develop interpretable foundation models for retinal imaging that provide faithful local and global explanations, outperforming state-of-the-art models with greater efficiency.
- A Model Ensemble-Based Post-Processing Framework for Fairness-Aware Prediction (viability: 5): https://sciencetostartup.com/paper/a-model-ensemble-based-post-processing-framework-for-fairness-aware-prediction - A post-processing framework that enhances fairness in machine learning predictions without altering model internals, applicable across diverse tasks and fairness definitions.
- Confidential Databases Without Cryptographic Mappings (viability: 7): https://sciencetostartup.com/paper/confidential-databases-without-cryptographic-mappings - FEDB offers a novel confidential database design that drastically reduces performance overhead by removing cryptographic operations from the critical path, enabling faster and more efficient secure queries in untrusted cloud environments.
- Statistical Characteristic-Guided Denoising for Rapid High-Resolution Transmission Electron Microscopy Imaging (viability: 7): https://sciencetostartup.com/paper/statistical-characteristic-guided-denoising-for-rapid-high-resolution-transmission-electron-microscopy-imaging - A novel denoising network for rapid, high-resolution transmission electron microscopy imaging, enabling atomic-scale observation of material nucleation dynamics.
- Agent Control Protocol: Admission Control for Agent Actions (viability: 8): https://sciencetostartup.com/paper/agent-control-protocol-admission-control-for-agent-actions - A formal protocol for secure and auditable admission control of autonomous agents in enterprise environments.
- Student views in AI Ethics and Social Impact (viability: 1): https://sciencetostartup.com/paper/student-views-in-ai-ethics-and-social-impact - This paper investigates student perspectives on AI ethics and social impact, identifying gender-based differences in awareness of AI's influence on various sectors.
- Detecting Basic Values in A Noisy Russian Social Media Text Data: A Multi-Stage Classification Framework (viability: 7): https://sciencetostartup.com/paper/detecting-basic-values-in-a-noisy-russian-social-media-text-data-a-multi-stage-classification-framework - A multi-stage framework for detecting human values in noisy Russian social media text, with publicly released models and benchmark results.
- Seasoning Generative Models for a Generalization Aftertaste (viability: 3): https://sciencetostartup.com/paper/seasoning-generative-models-for-a-generalization-aftertaste - A theoretical framework to refine generative models for improved generalization, building upon existing adversarial and diffusion techniques.
- ProRL Agent: Rollout-as-a-Service for RL Training of Multi-Turn LLM Agents (viability: 7): https://sciencetostartup.com/paper/prorl-agent-rollout-as-a-service-for-rl-training-of-multi-turn-llm-agents - A scalable API service for orchestrating RL agent rollouts, simplifying multi-turn LLM agent training across diverse tasks.
- Can LLM generate interesting mathematical research problems? (viability: 7): https://sciencetostartup.com/paper/can-llm-generate-interesting-mathematical-research-problems - An AI agent that generates novel and valuable mathematical research problems in differential geometry, validated by human experts.
- V-Dreamer: Automating Robotic Simulation and Trajectory Synthesis via Video Generation Priors (viability: 7): https://sciencetostartup.com/paper/v-dreamer-automating-robotic-simulation-and-trajectory-synthesis-via-video-generation-priors - Automate robotic simulation and trajectory synthesis using video generation for diverse, open-vocabulary manipulation tasks.
- dTRPO: Trajectory Reduction in Policy Optimization of Diffusion Large Language Models (viability: 7): https://sciencetostartup.com/paper/dtrpo-trajectory-reduction-in-policy-optimization-of-diffusion-large-language-models - A novel policy optimization method for diffusion large language models that significantly reduces training costs and improves performance on instruction-following and reasoning tasks.
- "You've got a friend in me": Co-Designing a Peer Social Robot for Young Newcomers' Language and Cultural Learning (viability: 3): https://sciencetostartup.com/paper/you-ve-got-a-friend-in-me-co-designing-a-peer-social-robot-for-young-newcomers-language-and-cultural-learning - A co-designed social robot to assist young newcomers with language and cultural learning in community literacy programs.
- Signals of Success and Struggle: Early Prediction and Physiological Signatures of Human Performance across Task Complexity (viability: 5): https://sciencetostartup.com/paper/signals-of-success-and-struggle-early-prediction-and-physiological-signatures-of-human-performance-across-task-complexit - Leverage early physiological signals from eye-tracking and cardiac data to predict user performance in interactive systems, enabling proactive interventions.
- VesselTok: Tokenizing Vessel-like 3D Biomedical Graph Representations for Reconstruction and Generation (viability: 7): https://sciencetostartup.com/paper/vesseltok-tokenizing-vessel-like-3d-biomedical-graph-representations-for-reconstruction-and-generation - VesselTok enables efficient and generative modeling of complex 3D anatomical structures like blood vessels and airways using novel tokenized latent representations.
- Perceptio: Perception Enhanced Vision Language Models via Spatial Token Generation (viability: 7): https://sciencetostartup.com/paper/perceptio-perception-enhanced-vision-language-models-via-spatial-token-generation - Perceptio enhances vision-language models with explicit 2D and 3D spatial reasoning by generating semantic segmentation and depth tokens directly within the autoregressive sequence.
- Functional Subspace Watermarking for Large Language Models (viability: 7): https://sciencetostartup.com/paper/functional-subspace-watermarking-for-large-language-models - A framework for embedding robust ownership signals into LLMs that withstands common model modifications and attacks.
- Rethinking Uncertainty Quantification and Entanglement in Image Segmentation (viability: 7): https://sciencetostartup.com/paper/rethinking-uncertainty-quantification-and-entanglement-in-image-segmentation - A novel approach to disentangle and quantify uncertainty in medical image segmentation, improving reliability for safety-critical applications.
- Weaver: Fuzzing JavaScript Engines at the JavaScript-WebAssembly Boundary (viability: 7): https://sciencetostartup.com/paper/weaver-fuzzing-javascript-engines-at-the-javascript-webassembly-boundary - Weaver is a greybox fuzzing framework that finds critical security vulnerabilities at the JavaScript-WebAssembly boundary in web browsers.
- Mi:dm K 2.5 Pro (viability: 7): https://sciencetostartup.com/paper/mi-dm-k-2-5-pro - A 32B parameter LLM optimized for multi-step reasoning and long-context understanding in Korean enterprise environments, with state-of-the-art performance on Korean benchmarks.
- Proceedings of the 2nd Workshop on Advancing Artificial Intelligence through Theory of Mind (viability: 0): https://sciencetostartup.com/paper/proceedings-of-the-2nd-workshop-on-advancing-artificial-intelligence-through-theory-of-mind - A curated anthology of research papers on the intersection of Theory of Mind and Artificial Intelligence.
- ViTac-Tracing: Visual-Tactile Imitation Learning of Deformable Object Tracing (viability: 7): https://sciencetostartup.com/paper/vitac-tracing-visual-tactile-imitation-learning-of-deformable-object-tracing - A visual-tactile imitation learning system that reliably traces deformable objects for improved robotic manipulation.
- Points-to-3D: Structure-Aware 3D Generation with Point Cloud Priors (viability: 7): https://sciencetostartup.com/paper/points-to-3d-structure-aware-3d-generation-with-point-cloud-priors - A diffusion-based framework that uses point cloud priors to generate controllable 3D assets and scenes with superior geometric fidelity.
- SRRM: Improving Recursive Transport Surrogates in the Small-Discrepancy Regime (viability: 2): https://sciencetostartup.com/paper/srrm-improving-recursive-transport-surrogates-in-the-small-discrepancy-regime - This paper theoretically analyzes and improves a statistical method for approximating the Wasserstein distance, with no immediate product application indicated.
- SoK: Practical Aspects of Releasing Differentially Private Graphs (viability: 4): https://sciencetostartup.com/paper/sok-practical-aspects-of-releasing-differentially-private-graphs - A framework to help practitioners select and evaluate differentially private graph release methods, addressing privacy and utility trade-offs.
- SEAR: Simple and Efficient Adaptation of Visual Geometric Transformers for RGB+Thermal 3D Reconstruction (viability: 8): https://sciencetostartup.com/paper/sear-simple-and-efficient-adaptation-of-visual-geometric-transformers-for-rgb-thermal-3d-reconstruction - A simple fine-tuning strategy that adapts existing visual geometry models for accurate RGB-thermal 3D reconstruction, outperforming state-of-the-art even in challenging conditions.
- Automatic Configuration of LLM Post-Training Pipelines (viability: 7): https://sciencetostartup.com/paper/automatic-configuration-of-llm-post-training-pipelines - Automate LLM post-training pipeline configuration to achieve state-of-the-art performance with significantly reduced computational cost.
- Empathetic Motion Generation for Humanoid Educational Robots via Reasoning-Guided Vision--Language--Motion Diffusion Architecture (viability: 4): https://sciencetostartup.com/paper/empathetic-motion-generation-for-humanoid-educational-robots-via-reasoning-guided-vision-language-motion-diffusion-archi - A framework for generating instruction-aware co-speech gestures for humanoid robots in educational settings.
- A Concept is More Than a Word: Diversified Unlearning in Text-to-Image Diffusion Models (viability: 7): https://sciencetostartup.com/paper/a-concept-is-more-than-a-word-diversified-unlearning-in-text-to-image-diffusion-models - A framework for precisely and robustly removing unwanted concepts from text-to-image models by representing concepts with diverse prompts, improving safety and model integrity.
- Enhancing the Parameterization of Reservoir Properties for Data Assimilation Using Deep VAE-GAN (viability: 7): https://sciencetostartup.com/paper/enhancing-the-parameterization-of-reservoir-properties-for-data-assimilation-using-deep-vae-gan - A deep learning model combining VAE and GAN for improved reservoir property parameterization in data assimilation, achieving both geological realism and accurate history matching.
- Implicit Grading Bias in Large Language Models: How Writing Style Affects Automated Assessment Across Math, Programming, and Essay Tasks (viability: 7): https://sciencetostartup.com/paper/implicit-grading-bias-in-large-language-models-how-writing-style-affects-automated-assessment-across-math-programming-an - Develop an LLM bias auditing tool to ensure fair automated grading by identifying and quantifying style-based grading disparities.
- ProCal: Probability Calibration for Neighborhood-Guided Source-Free Domain Adaptation (viability: 7): https://sciencetostartup.com/paper/procal-probability-calibration-for-neighborhood-guided-source-free-domain-adaptation - ProCal calibrates neighborhood-based predictions in source-free domain adaptation to prevent knowledge forgetting and noise overfitting, improving model performance on new datasets.
- ClawTrap: A MITM-Based Red-Teaming Framework for Real-World OpenClaw Security Evaluation (viability: 7): https://sciencetostartup.com/paper/clawtrap-a-mitm-based-red-teaming-framework-for-real-world-openclaw-security-evaluation - ClawTrap provides a real-world MITM-based red-teaming framework to evaluate the security robustness of autonomous web agents like OpenClaw against network-layer threats.
- NeuroGame Transformer: Gibbs-Inspired Attention Driven by Game Theory and Statistical Physics (viability: 7): https://sciencetostartup.com/paper/neurogame-transformer-gibbs-inspired-attention-driven-by-game-theory-and-statistical-physics - A novel transformer architecture that uses game theory and statistical physics to improve token dependency modeling, achieving state-of-the-art results on NLP benchmarks.
- Dual-Model Prediction of Affective Engagement and Vocal Attractiveness from Speaker Expressiveness in Video Learning (viability: 5): https://sciencetostartup.com/paper/dual-model-prediction-of-affective-engagement-and-vocal-attractiveness-from-speaker-expressiveness-in-video-learning - Predict audience engagement and vocal attractiveness in learning videos using only speaker expressiveness, enabling scalable and privacy-preserving affective computing.
- DA-Mamba: Learning Domain-Aware State Space Model for Global-Local Alignment in Domain Adaptive Object Detection (viability: 7): https://sciencetostartup.com/paper/da-mamba-learning-domain-aware-state-space-model-for-global-local-alignment-in-domain-adaptive-object-detection - A hybrid CNN-SSM architecture for domain adaptive object detection that aligns global and local features efficiently.
- Are complicated loss functions necessary for teaching LLMs to reason? (viability: 7): https://sciencetostartup.com/paper/are-complicated-loss-functions-necessary-for-teaching-llms-to-reason - A simplified, more efficient method for training LLMs to improve mathematical reasoning by leveraging group relative advantage without complex PPO constraints.
- WeNLEX: Weakly Supervised Natural Language Explanations for Multilabel Chest X-ray Classification (viability: 7): https://sciencetostartup.com/paper/wenlex-weakly-supervised-natural-language-explanations-for-multilabel-chest-x-ray-classification - A weakly supervised system that generates faithful and plausible natural language explanations for chest X-ray classifications, improving diagnostic accuracy and adapting to different audiences.
- Automatic detection of Gen-AI texts: A comparative framework of neural models (viability: 5): https://sciencetostartup.com/paper/automatic-detection-of-gen-ai-texts-a-comparative-framework-of-neural-models - Develops and benchmarks neural network models for detecting AI-generated text, outperforming existing commercial tools on specific datasets.
- ROFT-VINS: Robust Feature Tracking-based Visual-Inertial State Estimation for Harsh Environment (viability: 4): https://sciencetostartup.com/paper/roft-vins-robust-feature-tracking-based-visual-inertial-state-estimation-for-harsh-environment - A deep learning method for robust visual feature tracking in challenging environments, integrated into a VIO system.
- Memento-Skills: Let Agents Design Agents (viability: 8): https://sciencetostartup.com/paper/memento-skills-let-agents-design-agents - An agent that autonomously designs, adapts, and improves task-specific agents using a memory-based reinforcement learning framework and evolving externalized skills.
- 6Bit-Diffusion: Inference-Time Mixed-Precision Quantization for Video Diffusion Models (viability: 7): https://sciencetostartup.com/paper/6bit-diffusion-inference-time-mixed-precision-quantization-for-video-diffusion-models - Accelerate video generation by dynamically quantizing diffusion models at inference time, reducing memory and computation without sacrificing quality.
- Measuring and Exploiting Confirmation Bias in LLM-Assisted Security Code Review (viability: 7): https://sciencetostartup.com/paper/measuring-and-exploiting-confirmation-bias-in-llm-assisted-security-code-review - This research demonstrates a critical confirmation bias vulnerability in LLM-assisted code review tools, showing how it can be exploited in supply-chain attacks and proposing debiasing methods to mitigate risks.
- EdgeCrafter: Compact ViTs for Edge Dense Prediction via Task-Specialized Distillation (viability: 7): https://sciencetostartup.com/paper/edgecrafter-compact-vits-for-edge-dense-prediction-via-task-specialized-distillation - EdgeCrafter enables high-performance dense prediction tasks like object detection and segmentation on resource-constrained edge devices using compact Vision Transformers through task-specialized distillation.
- CausalRM: Causal-Theoretic Reward Modeling for RLHF from Observational User Feedbacks (viability: 7): https://sciencetostartup.com/paper/causalrm-causal-theoretic-reward-modeling-for-rlhf-from-observational-user-feedbacks - CausalRM enables cost-effective LLM alignment by learning reward models from readily available observational user feedback, overcoming noise and bias to achieve significant performance gains.
- Analysis Of Linguistic Stereotypes in Single and Multi-Agent Generative AI Architectures (viability: 6): https://sciencetostartup.com/paper/analysis-of-linguistic-stereotypes-in-single-and-multi-agent-generative-ai-architectures - This research develops and evaluates multi-agent architectures and prompt engineering techniques to mitigate dialect-based linguistic stereotypes in LLM outputs, offering a path towards fairer AI deployments.
- Ontology-Guided Diffusion for Zero-Shot Visual Sim2Real Transfer (viability: 7): https://sciencetostartup.com/paper/ontology-guided-diffusion-for-zero-shot-visual-sim2real-transfer - A neuro-symbolic framework that uses structured knowledge to bridge the simulation-to-reality gap in image translation, enabling data-efficient zero-shot transfer.
- MemMA: Coordinating the Memory Cycle through Multi-Agent Reasoning and In-Situ Self-Evolution (viability: 7): https://sciencetostartup.com/paper/memma-coordinating-the-memory-cycle-through-multi-agent-reasoning-and-in-situ-self-evolution - MemMA is a multi-agent framework that coordinates LLM memory cycles for improved long-horizon interaction and self-evolving memory construction.
- Accurate and Efficient Multi-Channel Time Series Forecasting via Sparse Attention Mechanism (viability: 7): https://sciencetostartup.com/paper/accurate-and-efficient-multi-channel-time-series-forecasting-via-sparse-attention-mechanism - A novel multi-channel time series forecasting architecture that uses sparse attention and multi-modal fusion for improved accuracy and efficiency.
- From ex(p) to poly: Gaussian Splatting with Polynomial Kernels (viability: 5): https://sciencetostartup.com/paper/from-ex-p-to-poly-gaussian-splatting-with-polynomial-kernels - A novel polynomial kernel for Gaussian Splatting that improves computational efficiency and compatibility with existing datasets.
- Off-Policy Learning with Limited Supply (viability: 5): https://sciencetostartup.com/paper/off-policy-learning-with-limited-supply - A novel off-policy learning method for contextual bandits that optimizes item allocation under limited supply constraints, outperforming existing methods on synthetic and real-world data.
- OCP: Orthogonal Constrained Projection for Sparse Scaling in Industrial Commodity Recommendation (viability: 8): https://sciencetostartup.com/paper/ocp-orthogonal-constrained-projection-for-sparse-scaling-in-industrial-commodity-recommendation - A novel projection method for industrial recommendation systems that significantly improves scalability and performance by optimizing item embeddings, demonstrated by large-scale deployment and metric uplifts.
- Cross-Ecosystem Vulnerability Analysis for Python Applications (viability: 5): https://sciencetostartup.com/paper/cross-ecosystem-vulnerability-analysis-for-python-applications - A provenance-aware vulnerability analysis tool that accurately identifies affected Python applications by cross-referencing native library dependencies with OS package versions.
- STEP: Scientific Time-Series Encoder Pretraining via Cross-Domain Distillation (viability: 7): https://sciencetostartup.com/paper/step-scientific-time-series-encoder-pretraining-via-cross-domain-distillation - A pretraining framework that distills knowledge from existing foundation models to create a unified encoder for scientific time series data, addressing sparsity and heterogeneity challenges.
- Secure Wi-Fi Ranging Today: Security and Adoption of IEEE 802.11az/bk (viability: 3): https://sciencetostartup.com/paper/secure-wi-fi-ranging-today-security-and-adoption-of-ieee-802-11az-bk - This paper analyzes the security and deployability of IEEE 802.11az/bk secure Wi-Fi ranging, identifying vulnerabilities and providing guidelines for safer implementation.
- HISR: Hindsight Information Modulated Segmental Process Rewards For Multi-turn Agentic Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/hisr-hindsight-information-modulated-segmental-process-rewards-for-multi-turn-agentic-reinforcement-learning - A novel reinforcement learning reward modulation technique for agents to improve long-horizon decision-making by aligning rewards with sub-goals and emphasizing significant segments.
- Revisiting Label Inference Attacks in Vertical Federated Learning: Why They Are Vulnerable and How to Defend (viability: 4): https://sciencetostartup.com/paper/revisiting-label-inference-attacks-in-vertical-federated-learning-why-they-are-vulnerable-and-how-to-defend - This research reveals vulnerabilities in vertical federated learning label inference attacks and proposes a zero-overhead defense by adjusting model layers to improve privacy.
- Words at Play: Benchmarking Audio Pun Understanding in Large Audio-Language Models (viability: 7): https://sciencetostartup.com/paper/words-at-play-benchmarking-audio-pun-understanding-in-large-audio-language-models - We've created the first benchmark to evaluate and improve large audio-language models' ability to understand spoken puns, a critical but underexplored area of natural language understanding.
- Cognitive Amplification vs Cognitive Delegation in Human-AI Systems: A Metric Framework (viability: 2): https://sciencetostartup.com/paper/cognitive-amplification-vs-cognitive-delegation-in-human-ai-systems-a-metric-framework - A framework to measure whether AI amplifies human intelligence or causes over-reliance, guiding the design of sustainable human-AI systems.
- MANAR: Memory-augmented Attention with Navigational Abstract Conceptual Representation (viability: 7): https://sciencetostartup.com/paper/manar-memory-augmented-attention-with-navigational-abstract-conceptual-representation - A novel attention mechanism inspired by cognitive theory that offers linear-time scaling and enhanced representational power for multimodal AI tasks.
- Towards High-Quality Image Segmentation: Improving Topology Accuracy by Penalizing Neighbor Pixels (viability: 7): https://sciencetostartup.com/paper/towards-high-quality-image-segmentation-improving-topology-accuracy-by-penalizing-neighbor-pixels - A novel, efficient method to improve the topological accuracy of image segmentations, enhancing reliability for downstream quantification analyses.
- CSSDF-Net: Safe Motion Planning Based on Neural Implicit Representations of Configuration Space Distance Field (viability: 7): https://sciencetostartup.com/paper/cssdf-net-safe-motion-planning-based-on-neural-implicit-representations-of-configuration-space-distance-field - A neural network learns a continuous distance field in robot configuration space for safe, zero-shot motion planning in complex environments.
- Thinking with Constructions: A Benchmark and Policy Optimization for Visual-Text Interleaved Geometric Reasoning (viability: 7): https://sciencetostartup.com/paper/thinking-with-constructions-a-benchmark-and-policy-optimization-for-visual-text-interleaved-geometric-reasoning - A framework and benchmark for teaching multimodal LLMs to strategically use visual constructions for improved geometric reasoning.
- Multimodal Model for Computational Pathology:Representation Learning and Image Compression (viability: 3): https://sciencetostartup.com/paper/multimodal-model-for-computational-pathology-representation-learning-and-image-compression - A review of multimodal AI for computational pathology, focusing on representation learning and image compression for improved diagnosis.
- Enhancing Multi-Corpus Training in SSL-Based Anti-Spoofing Models: Domain-Invariant Feature Extraction (viability: 7): https://sciencetostartup.com/paper/enhancing-multi-corpus-training-in-ssl-based-anti-spoofing-models-domain-invariant-feature-extraction - A framework for robust speech anti-spoofing that significantly improves performance across diverse datasets by extracting domain-invariant features.
- Balanced Thinking: Improving Chain of Thought Training in Vision Language Models (viability: 7): https://sciencetostartup.com/paper/balanced-thinking-improving-chain-of-thought-training-in-vision-language-models - A novel training method for vision-language models that significantly improves reasoning accuracy and conciseness by adaptively weighting training loss, reducing training time by 7x compared to standard methods.
- Multiscale Switch for Semi-Supervised and Contrastive Learning in Medical Ultrasound Image Segmentation (viability: 7): https://sciencetostartup.com/paper/multiscale-switch-for-semi-supervised-and-contrastive-learning-in-medical-ultrasound-image-segmentation - A parameter-efficient semi-supervised learning framework for medical ultrasound image segmentation that leverages multiscale patch mixing and frequency domain contrastive learning to achieve state-of-the-art performance with limited labeled data.
- Benchmarking PDF Parsers on Table Extraction with LLM-based Semantic Evaluation (viability: 7): https://sciencetostartup.com/paper/benchmarking-pdf-parsers-on-table-extraction-with-llm-based-semantic-evaluation - A new LLM-based evaluation framework for PDF table extraction that significantly outperforms existing metrics, providing practical guidance for parser selection.
- Click-to-Ask: An AI Live Streaming Assistant with Offline Copywriting and Online Interactive QA (viability: 8): https://sciencetostartup.com/paper/click-to-ask-an-ai-live-streaming-assistant-with-offline-copywriting-and-online-interactive-qa - An AI assistant that automates product copywriting and provides real-time Q&A for live streamers, boosting sales and engagement.
- Beyond TVLA: Anderson-Darling Leakage Assessment for Neural Network Side-Channel Leakage Detection (viability: 4): https://sciencetostartup.com/paper/beyond-tvla-anderson-darling-leakage-assessment-for-neural-network-side-channel-leakage-detection - A novel statistical framework for detecting side-channel leakage in neural networks that offers improved sensitivity over traditional methods.
- MeInTime: Bridging Age Gap in Identity-Preserving Face Restoration (viability: 7): https://sciencetostartup.com/paper/meintime-bridging-age-gap-in-identity-preserving-face-restoration - A diffusion-based face restoration method that bridges the age gap in identity-preserving image enhancement using cross-age references.
- Evaluating Model-Free Policy Optimization in Masked-Action Environments via an Exact Blackjack Oracle (viability: 4): https://sciencetostartup.com/paper/evaluating-model-free-policy-optimization-in-masked-action-environments-via-an-exact-blackjack-oracle - Develops a rigorous benchmark and evaluation methodology for model-free policy optimization in complex, masked-action environments, highlighting the limitations of current methods.
- A Comparative Empirical Study of Catastrophic Forgetting Mitigation in Sequential Task Adaptation for Continual Natural Language Processing Systems (viability: 5): https://sciencetostartup.com/paper/a-comparative-empirical-study-of-catastrophic-forgetting-mitigation-in-sequential-task-adaptation-for-continual-natural - This research empirically compares catastrophic forgetting mitigation strategies for continual intent classification, identifying replay-based methods as crucial for robust adaptation in NLP systems.
- A Theoretical Comparison of No-U-Turn Sampler Variants: Necessary and Su?cient Convergence Conditions and Mixing Time Analysis under Gaussian Targets (viability: 2): https://sciencetostartup.com/paper/a-theoretical-comparison-of-no-u-turn-sampler-variants-necessary-and-su-cient-convergence-conditions-and-mixing-time-ana - This paper provides a theoretical analysis of No-U-Turn Sampler variants, focusing on convergence conditions and mixing times for Gaussian targets, with no mention of practical implementation or product potential.
- PhysVideo: Physically Plausible Video Generation with Cross-View Geometry Guidance (viability: 7): https://sciencetostartup.com/paper/physvideo-physically-plausible-video-generation-with-cross-view-geometry-guidance - PhysVideo generates physically plausible videos by leveraging cross-view geometry guidance and a novel two-stage framework with a dedicated dataset.
- MOSAIC: Multi-Objective Slice-Aware Iterative Curation for Alignment (viability: 7): https://sciencetostartup.com/paper/mosaic-multi-objective-slice-aware-iterative-curation-for-alignment - A framework for iteratively curating training data to balance safety, reduce over-refusal, and improve instruction following in LLMs with a fixed budget.
- Training-Free Sparse Attention for Fast Video Generation via Offline Layer-Wise Sparsity Profiling and Online Bidirectional Co-Clustering (viability: 7): https://sciencetostartup.com/paper/training-free-sparse-attention-for-fast-video-generation-via-offline-layer-wise-sparsity-profiling-and-online-bidirectio - A training-free sparse attention framework for significantly faster video generation by profiling layer-wise sparsity and using bidirectional co-clustering.
- SwiftGS: Episodic Priors for Immediate Satellite Surface Recovery (viability: 4): https://sciencetostartup.com/paper/swiftgs-episodic-priors-for-immediate-satellite-surface-recovery - A meta-learned system for rapid 3D surface reconstruction from satellite imagery using episodic priors.
- An Onto-Relational-Sophic Framework for Governing Synthetic Minds (viability: 3): https://sciencetostartup.com/paper/an-onto-relational-sophic-framework-for-governing-synthetic-minds - A philosophical framework for governing increasingly capable synthetic minds, moving beyond tool-centric regulation to address their multi-dimensional being and societal relations.
- D-Mem: A Dual-Process Memory System for LLM Agents (viability: 7): https://sciencetostartup.com/paper/d-mem-a-dual-process-memory-system-for-llm-agents - D-Mem is a dual-process memory system for LLM agents that combines fast vector retrieval with high-fidelity deliberation to improve long-horizon reasoning without significant computational overhead.
- Agentic Flow Steering and Parallel Rollout Search for Spatially Grounded Text-to-Image Generation (viability: 7): https://sciencetostartup.com/paper/agentic-flow-steering-and-parallel-rollout-search-for-spatially-grounded-text-to-image-generation - A closed-loop framework that uses a VLM critic to steer text-to-image generation, achieving state-of-the-art results with parallel rollout search.
- GEAR: Geography-knowledge Enhanced Analog Recognition Framework in Extreme Environments (viability: 7): https://sciencetostartup.com/paper/gear-geography-knowledge-enhanced-analog-recognition-framework-in-extreme-environments - A framework for identifying terrestrial analogs of extreme deep-sea environments using topographic similarity, enabling cost-effective biological research.
- GenVideoLens: Where LVLMs Fall Short in AI-Generated Video Detection? (viability: 7): https://sciencetostartup.com/paper/genvideolens-where-lvlms-fall-short-in-ai-generated-video-detection - A fine-grained benchmark and evaluation framework to diagnose and improve AI-generated video detection capabilities of Large Vision-Language Models.
- REST: Receding Horizon Explorative Steiner Tree for Zero-Shot Object-Goal Navigation (viability: 7): https://sciencetostartup.com/paper/rest-receding-horizon-explorative-steiner-tree-for-zero-shot-object-goal-navigation - A training-free framework for zero-shot object-goal navigation in unknown environments using LLM-reasoned path trees.
- OpenT2M: No-frill Motion Generation with Open-source,Large-scale, High-quality Data (viability: 7): https://sciencetostartup.com/paper/opent2m-no-frill-motion-generation-with-open-source-large-scale-high-quality-data - A high-quality, large-scale open-source dataset and a novel motion tokenizer to significantly improve text-to-motion generation.
- Learning to Self-Evolve (viability: 7): https://sciencetostartup.com/paper/learning-to-self-evolve - A reinforcement learning framework that trains LLMs to iteratively refine their own context at test time, significantly improving performance on new tasks.
- Benchmarking CNN-based Models against Transformer-based Models for Abdominal Multi-Organ Segmentation on the RATIC Dataset (viability: 4): https://sciencetostartup.com/paper/benchmarking-cnn-based-models-against-transformer-based-models-for-abdominal-multi-organ-segmentation-on-the-ratic-datas - Benchmarking CNNs against Transformers for abdominal organ segmentation reveals that optimized CNNs remain competitive on heterogeneous datasets.
- ZEBRAARENA: A Diagnostic Simulation Environment for Studying Reasoning-Action Coupling in Tool-Augmented LLMs (viability: 7): https://sciencetostartup.com/paper/zebraarena-a-diagnostic-simulation-environment-for-studying-reasoning-action-coupling-in-tool-augmented-llms - ZebraArena provides a diagnostic environment to evaluate and improve the reasoning-action coupling of tool-augmented LLMs, addressing a key challenge for advanced AI agents.
- Cyber-Resilient Digital Twins: Discriminating Attacks for Safe Critical Infrastructure Control (viability: 7): https://sciencetostartup.com/paper/cyber-resilient-digital-twins-discriminating-attacks-for-safe-critical-infrastructure-control - An intelligent digital twin system that discriminates cyber-attack types in critical infrastructure to enable safe, resilient control without system shutdowns.
- DiscoPhon: Benchmarking the Unsupervised Discovery of Phoneme Inventories With Discrete Speech Units (viability: 7): https://sciencetostartup.com/paper/discophon-benchmarking-the-unsupervised-discovery-of-phoneme-inventories-with-discrete-speech-units - A benchmark and baselines for unsupervised discovery of phoneme inventories from speech, enabling better language processing tools for unseen languages.
- Cross-Modal Rationale Transfer for Explainable Humanitarian Classification on Social Media (viability: 7): https://sciencetostartup.com/paper/cross-modal-rationale-transfer-for-explainable-humanitarian-classification-on-social-media - An interpretable multimodal classification framework that transfers rationales between text and images to improve accuracy and reduce annotation effort for humanitarian crises.
- A Complexity Hierarchy of Shuffles in Card-Based Protocols (viability: 2): https://sciencetostartup.com/paper/a-complexity-hierarchy-of-shuffles-in-card-based-protocols - This paper introduces a complexity hierarchy for shuffles in card-based cryptographic protocols to evaluate their practical implementation difficulty.
- AutORAN: LLM-driven Natural Language Programming for Agile xApp Development (viability: 7): https://sciencetostartup.com/paper/autoran-llm-driven-natural-language-programming-for-agile-xapp-development - AutORAN is an LLM-driven framework that automates xApp development for O-RAN systems, transforming user intents into deployable applications in minutes.
- Improving Joint Audio-Video Generation with Cross-Modal Context Learning (viability: 5): https://sciencetostartup.com/paper/improving-joint-audio-video-generation-with-cross-modal-context-learning - A novel method for joint audio-video generation that improves temporal alignment and reduces inconsistencies using context learning modules.
- SJD-PAC: Accelerating Speculative Jacobi Decoding via Proactive Drafting and Adaptive Continuation (viability: 7): https://sciencetostartup.com/paper/sjd-pac-accelerating-speculative-jacobi-decoding-via-proactive-drafting-and-adaptive-continuation - Accelerate text-to-image synthesis by proactively drafting and adaptively continuing generation to achieve significant speedups without sacrificing image quality.
- Complementary Text-Guided Attention for Zero-Shot Adversarial Robustness (viability: 7): https://sciencetostartup.com/paper/complementary-text-guided-attention-for-zero-shot-adversarial-robustness - Enhance zero-shot adversarial robustness of vision-language models by refining text-guided attention mechanisms.
- myMNIST: Benchmark of PETNN, KAN, and Classical Deep Learning Models for Burmese Handwritten Digit Recognition (viability: 4): https://sciencetostartup.com/paper/mymnist-benchmark-of-petnn-kan-and-classical-deep-learning-models-for-burmese-handwritten-digit-recognition - Benchmarks Burmese handwritten digit recognition models, highlighting CNN and PETNN performance for regional NLP/AI research.
- Elastic Weight Consolidation Done Right for Continual Learning (viability: 7): https://sciencetostartup.com/paper/elastic-weight-consolidation-done-right-for-continual-learning - A novel modification to Elastic Weight Consolidation significantly improves continual learning performance by rectifying importance estimation, offering a more robust solution to catastrophic forgetting.
- Language Model Maps for Prompt-Response Distributions via Log-Likelihood Vectors (viability: 3): https://sciencetostartup.com/paper/language-model-maps-for-prompt-response-distributions-via-log-likelihood-vectors - A framework for analyzing and comparing language models based on their log-likelihood vectors, revealing relationships between model attributes, task performance, and prompt effects.
- Benchmarking Visual Feature Representations for LiDAR-Inertial-Visual Odometry Under Challenging Conditions (viability: 7): https://sciencetostartup.com/paper/benchmarking-visual-feature-representations-for-lidar-inertial-visual-odometry-under-challenging-conditions - A hybrid visual odometry system that leverages learning-based descriptors to achieve robust localization in autonomous driving under challenging visual conditions.
- AU Codes, Language, and Synthesis: Translating Anatomy to Text for Facial Behavior Synthesis (viability: 8): https://sciencetostartup.com/paper/au-codes-language-and-synthesis-translating-anatomy-to-text-for-facial-behavior-synthesis - Synthesize anatomically plausible and behaviorally rich facial expressions from natural language descriptions of Action Units, overcoming limitations of existing text-to-face models.
- Color image restoration based on nonlocal saturation-value similarity (viability: 4): https://sciencetostartup.com/paper/color-image-restoration-based-on-nonlocal-saturation-value-similarity - A novel variational technique for color image restoration using saturation-value similarity to improve visual quality and quantitative metrics.
- HAViT: Historical Attention Vision Transformer (viability: 7): https://sciencetostartup.com/paper/havit-historical-attention-vision-transformer - A method to improve Vision Transformer performance by propagating historical attention across layers, with publicly available code and demonstrated accuracy gains.
- Breaking Hard Isomorphism Benchmarks with DRESS (viability: 5): https://sciencetostartup.com/paper/breaking-hard-isomorphism-benchmarks-with-dress - A novel graph fingerprinting method that significantly outperforms existing techniques for distinguishing complex graph structures, with potential applications in network analysis and data security.
- WarPGNN: A Parametric Thermal Warpage Analysis Framework with Physics-aware Graph Neural Network (viability: 8): https://sciencetostartup.com/paper/warpgnn-a-parametric-thermal-warpage-analysis-framework-with-physics-aware-graph-neural-network - A physics-informed Graph Neural Network framework that accelerates thermal warpage analysis for chiplet-based designs by over 200x compared to traditional FEM methods.
- ICE: Intervention-Consistent Explanation Evaluation with Statistical Grounding for LLMs (viability: 7): https://sciencetostartup.com/paper/ice-intervention-consistent-explanation-evaluation-with-statistical-grounding-for-llms - A new framework and benchmark for statistically grounding LLM explanation faithfulness, revealing operator-dependent insights and anti-faithfulness.
- MedForge: Interpretable Medical Deepfake Detection via Forgery-aware Reasoning (viability: 7): https://sciencetostartup.com/paper/medforge-interpretable-medical-deepfake-detection-via-forgery-aware-reasoning - MedForge provides interpretable, evidence-grounded detection of manipulated medical images to restore clinical trust.
- Interplay: Training Independent Simulators for Reference-Free Conversational Recommendation (viability: 5): https://sciencetostartup.com/paper/interplay-training-independent-simulators-for-reference-free-conversational-recommendation - Generate realistic conversational recommendation data using independent, reference-free LLM simulators to overcome data scarcity.
- CAPSUL: A Comprehensive Human Protein Benchmark for Subcellular Localization (viability: 7): https://sciencetostartup.com/paper/capsul-a-comprehensive-human-protein-benchmark-for-subcellular-localization - A new benchmark and interpretability tool for human protein subcellular localization, enabling data-driven drug target identification and biological discovery.
- Attack by Unlearning: Unlearning-Induced Adversarial Attacks on Graph Neural Networks (viability: 4): https://sciencetostartup.com/paper/attack-by-unlearning-unlearning-induced-adversarial-attacks-on-graph-neural-networks - This research introduces a novel attack vector on graph neural networks by exploiting the unlearning process, demonstrating how data deletion requests can be manipulated to degrade model performance.
- SpecForge: A Flexible and Efficient Open-Source Training Framework for Speculative Decoding (viability: 7): https://sciencetostartup.com/paper/specforge-a-flexible-and-efficient-open-source-training-framework-for-speculative-decoding - An open-source framework and suite of draft models to significantly accelerate LLM inference through speculative decoding.
- Transformers Learn Robust In-Context Regression under Distributional Uncertainty (viability: 4): https://sciencetostartup.com/paper/transformers-learn-robust-in-context-regression-under-distributional-uncertainty - Transformers can perform robust in-context learning for regression tasks even with uncertain and non-Gaussian data distributions, outperforming classical methods.
- Reasonably reasoning AI agents can avoid game-theoretic failures in zero-shot, provably (viability: 3): https://sciencetostartup.com/paper/reasonably-reasoning-ai-agents-can-avoid-game-theoretic-failures-in-zero-shot-provably - AI agents can naturally achieve stable equilibrium behaviors in strategic interactions without explicit post-training alignment.
- CausalVAD: De-confounding End-to-End Autonomous Driving via Causal Intervention (viability: 7): https://sciencetostartup.com/paper/causalvad-de-confounding-end-to-end-autonomous-driving-via-causal-intervention - A de-confounding framework for end-to-end autonomous driving models that uses causal intervention to improve reliability and safety by eliminating dataset biases.
- TiBCLaG: A Trigger-induced Bistable Compliant Laparoscopic Grasper (viability: 5): https://sciencetostartup.com/paper/tibclag-a-trigger-induced-bistable-compliant-laparoscopic-grasper - A 3D-printed compliant laparoscopic grasper that reduces manufacturing costs and enables adaptive grasping.
- HiMu: Hierarchical Multimodal Frame Selection for Long Video Question Answering (viability: 7): https://sciencetostartup.com/paper/himu-hierarchical-multimodal-frame-selection-for-long-video-question-answering - HiMu offers a training-free framework for efficient and accurate long-form video question answering by hierarchically decomposing queries and leveraging lightweight multimodal experts.
- Cross-Lingual LLM-Judge Transfer via Evaluation Decomposition (viability: 7): https://sciencetostartup.com/paper/cross-lingual-llm-judge-transfer-via-evaluation-decomposition - A framework for cross-lingual LLM evaluation that leverages language-agnostic criteria to reduce the need for target-language annotations.
- Inductance-Based Force Self-Sensing in Fiber-Reinforced Pneumatic Twisted-and-Coiled Actuators (viability: 4): https://sciencetostartup.com/paper/inductance-based-force-self-sensing-in-fiber-reinforced-pneumatic-twisted-and-coiled-actuators - Develop self-sensing robotic actuators that intrinsically estimate force and displacement using inductance feedback, overcoming hysteresis limitations for precise closed-loop control.
- Quantifying Memory Cells Vulnerability for DRAM Security (viability: 4): https://sciencetostartup.com/paper/quantifying-memory-cells-vulnerability-for-dram-security - A framework to quantify DRAM cell vulnerabilities for enhanced data integrity and confidentiality in security applications.
- SINDy-KANs: Sparse identification of non-linear dynamics through Kolmogorov-Arnold networks (viability: 3): https://sciencetostartup.com/paper/sindy-kans-sparse-identification-of-non-linear-dynamics-through-kolmogorov-arnold-networks - A novel method for discovering sparse equations in dynamical systems by combining Kolmogorov-Arnold networks with sparse identification techniques.
- HEP Statistical Inference for UAV Fault Detection: CLs, LRT, and SBI Applied to Blade Damage (viability: 7): https://sciencetostartup.com/paper/hep-statistical-inference-for-uav-fault-detection-cls-lrt-and-sbi-applied-to-blade-damage - Leveraging advanced statistical methods from particle physics, this system provides highly accurate and quantified detection of UAV blade damage with controlled false alarm rates.
- CoDA: Exploring Chain-of-Distribution Attacks and Post-Hoc Token-Space Repair for Medical Vision-Language Models (viability: 4): https://sciencetostartup.com/paper/coda-exploring-chain-of-distribution-attacks-and-post-hoc-token-space-repair-for-medical-vision-language-models - A framework for testing and repairing the robustness of medical vision-language models against realistic image distribution shifts.
- Remedying Target-Domain Astigmatism for Cross-Domain Few-Shot Object Detection (viability: 7): https://sciencetostartup.com/paper/remedying-target-domain-astigmatism-for-cross-domain-few-shot-object-detection - A biologically inspired framework to fix attention issues in cross-domain few-shot object detection, achieving new state-of-the-art results.
- GAPSL: A Gradient-Aligned Parallel Split Learning on Heterogeneous Data (viability: 7): https://sciencetostartup.com/paper/gapsl-a-gradient-aligned-parallel-split-learning-on-heterogeneous-data - A framework for efficient and accurate federated learning on resource-constrained devices by aligning client gradients.
- iSatCR: Graph-Empowered Joint Onboard Computing and Routing for LEO Data Delivery (viability: 3): https://sciencetostartup.com/paper/isatcr-graph-empowered-joint-onboard-computing-and-routing-for-leo-data-delivery - A graph-based approach to optimize onboard computing and routing for LEO satellite data delivery.
- Beyond Passive Aggregation: Active Auditing and Topology-Aware Defense in Decentralized Federated Learning (viability: 4): https://sciencetostartup.com/paper/beyond-passive-aggregation-active-auditing-and-topology-aware-defense-in-decentralized-federated-learning - An active auditing framework for decentralized federated learning that proactively detects and mitigates stealthy backdoor attacks.
- Data-efficient pre-training by scaling synthetic megadocs (viability: 7): https://sciencetostartup.com/paper/data-efficient-pre-training-by-scaling-synthetic-megadocs - This research proposes a novel method for data-efficient pre-training of language models by scaling synthetic data generation into 'megadocs', significantly improving data efficiency and long-context performance.
- Balancing the Reasoning Load: Difficulty-Differentiated Policy Optimization with Length Redistribution for Efficient and Robust Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/balancing-the-reasoning-load-difficulty-differentiated-policy-optimization-with-length-redistribution-for-efficient-and - A novel reinforcement learning algorithm that optimizes Large Reasoning Models to reduce overthinking and improve accuracy by adapting output length based on task difficulty.
- Scaling Sim-to-Real Reinforcement Learning for Robot VLAs with Generative 3D Worlds (viability: 7): https://sciencetostartup.com/paper/scaling-sim-to-real-reinforcement-learning-for-robot-vlas-with-generative-3d-worlds - Scales robot reinforcement learning by generating diverse 3D environments with generative models, significantly improving sim-to-real transfer and task performance.
- When Names Change Verdicts: Intervention Consistency Reveals Systematic Bias in LLM Decision-Making (viability: 8): https://sciencetostartup.com/paper/when-names-change-verdicts-intervention-consistency-reveals-systematic-bias-in-llm-decision-making - ICE-Guard provides a framework to detect and mitigate systematic bias in LLMs across high-stakes decisions, with code and data publicly available.
- Correlation-Weighted Multi-Reward Optimization for Compositional Generation (viability: 7): https://sciencetostartup.com/paper/correlation-weighted-multi-reward-optimization-for-compositional-generation - A framework for text-to-image models that improves compositional generation by adaptively weighting concept rewards based on their correlations, leading to more accurate and complete image outputs.
- 3DreamBooth: High-Fidelity 3D Subject-Driven Video Generation Model (viability: 7): https://sciencetostartup.com/paper/3dreambooth-high-fidelity-3d-subject-driven-video-generation-model - Generate high-fidelity 3D videos of customized subjects from single views by decoupling spatial geometry and temporal motion.
- Counting Circuits: Mechanistic Interpretability of Visual Reasoning in Large Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/counting-circuits-mechanistic-interpretability-of-visual-reasoning-in-large-vision-language-models - Enhance general visual reasoning in large models by fine-tuning their specific counting circuits with synthetic data.
- Robotic Agentic Platform for Intelligent Electric Vehicle Disassembly (viability: 5): https://sciencetostartup.com/paper/robotic-agentic-platform-for-intelligent-electric-vehicle-disassembly - An agentic robotic platform for intelligent electric vehicle battery disassembly, leveraging AI for perception-driven manipulation and automated fastener removal.
- On the Peril of (Even a Little) Nonstationarity in Satisficing Regret Minimization (viability: 2): https://sciencetostartup.com/paper/on-the-peril-of-even-a-little-nonstationarity-in-satisficing-regret-minimization - This paper theoretically analyzes regret minimization in nonstationary bandit problems, showing that even slight nonstationarity significantly impacts performance.
- CAFlow: Adaptive-Depth Single-Step Flow Matching for Efficient Histopathology Super-Resolution (viability: 5): https://sciencetostartup.com/paper/caflow-adaptive-depth-single-step-flow-matching-for-efficient-histopathology-super-resolution - An efficient super-resolution framework for histopathology images that significantly reduces computational cost without sacrificing quality, enabling faster analysis of whole-slide images.
- OnlinePG: Online Open-Vocabulary Panoptic Mapping with 3D Gaussian Splatting (viability: 7): https://sciencetostartup.com/paper/onlinepg-online-open-vocabulary-panoptic-mapping-with-3d-gaussian-splatting - An online system for open-vocabulary panoptic mapping using 3D Gaussian Splatting for embodied applications.
- Foundations and Architectures of Artificial Intelligence for Motor Insurance (viability: 3): https://sciencetostartup.com/paper/foundations-and-architectures-of-artificial-intelligence-for-motor-insurance - A systematic treatment of AI foundations and architectures for motor insurance, focusing on perception, reasoning, and production infrastructure for risk assessment and claims processing.
- Expert Personas Improve LLM Alignment but Damage Accuracy: Bootstrapping Intent-Based Persona Routing with PRISM (viability: 5): https://sciencetostartup.com/paper/expert-personas-improve-llm-alignment-but-damage-accuracy-bootstrapping-intent-based-persona-routing-with-prism - A method to improve LLM alignment using persona routing without sacrificing accuracy, by distilling intent-conditioned expert personas into LoRA adapters.
- From Snapshots to Symphonies: The Evolution of Protein Prediction from Static Structures to Generative Dynamics and Multimodal Interactions (viability: 3): https://sciencetostartup.com/paper/from-snapshots-to-symphonies-the-evolution-of-protein-prediction-from-static-structures-to-generative-dynamics-and-multi - This paper reviews the evolution of AI in protein science, from static structure prediction to generative dynamics and multimodal interactions, highlighting current challenges and future directions.
- HOMEY: Heuristic Object Masking with Enhanced YOLO for Property Insurance Risk Detection (viability: 7): https://sciencetostartup.com/paper/homey-heuristic-object-masking-with-enhanced-yolo-for-property-insurance-risk-detection - A computer vision system that detects property risks for insurance underwriting using a novel YOLO-based framework with enhanced masking and loss calibration.
- Efficient Video Diffusion with Sparse Information Transmission for Video Compression (viability: 7): https://sciencetostartup.com/paper/efficient-video-diffusion-with-sparse-information-transmission-for-video-compression - A novel diffusion-based video compression method that achieves state-of-the-art perceptual quality and temporal consistency at ultra-low bitrates by sparsely encoding temporal information.
- NymeriaPlus: Enriching Nymeria Dataset with Additional Annotations and Data (viability: 4): https://sciencetostartup.com/paper/nymeriaplus-enriching-nymeria-dataset-with-additional-annotations-and-data - Enhancing a large-scale egocentric human activity dataset with richer annotations and modalities to support multimodal learning for embodied AI.
- Cross-Domain Demo-to-Code via Neurosymbolic Counterfactual Reasoning (viability: 7): https://sciencetostartup.com/paper/cross-domain-demo-to-code-via-neurosymbolic-counterfactual-reasoning - A neurosymbolic framework that enables robots to adapt their learned procedures to new environments by reasoning about counterfactual states, significantly improving task success.
- MemoAct: Atkinson-Shiffrin-Inspired Memory-Augmented Visuomotor Policy for Robotic Manipulation (viability: 7): https://sciencetostartup.com/paper/memoact-atkinson-shiffrin-inspired-memory-augmented-visuomotor-policy-for-robotic-manipulation - A hierarchical memory-based robotic policy inspired by human memory models for precise task tracking and long-horizon retention.
- FILT3R: Latent State Adaptive Kalman Filter for Streaming 3D Reconstruction (viability: 7): https://sciencetostartup.com/paper/filt3r-latent-state-adaptive-kalman-filter-for-streaming-3d-reconstruction - A training-free latent filtering layer for streaming 3D reconstruction that adaptively balances memory retention with new observations for improved long-horizon stability.
- AIMER: Calibration-Free Task-Agnostic MoE Pruning (viability: 7): https://sciencetostartup.com/paper/aimer-calibration-free-task-agnostic-moe-pruning - A calibration-free method for pruning Mixture-of-Experts models to reduce memory and serving overhead with competitive performance.
- EntropyCache: Decoded Token Entropy Guided KV Caching for Diffusion Language Models (viability: 8): https://sciencetostartup.com/paper/entropycache-decoded-token-entropy-guided-kv-caching-for-diffusion-language-models - EntropyCache offers a training-free KV caching method for diffusion language models that significantly speeds up inference by using decoded token entropy as a cost-effective signal for recomputation.
- TexEditor: Structure-Preserving Text-Driven Texture Editing (viability: 7): https://sciencetostartup.com/paper/texeditor-structure-preserving-text-driven-texture-editing - TexEditor enables precise text-driven texture editing for 3D assets, ensuring structural integrity through a novel dataset and reinforcement learning approach.
- Precise Performance of Linear Denoisers in the Proportional Regime (viability: 2): https://sciencetostartup.com/paper/precise-performance-of-linear-denoisers-in-the-proportional-regime - This paper provides a theoretical analysis of linear denoisers for noisy data, deriving closed-form expressions for generalization error in a specific regime.
- The Truncation Blind Spot: How Decoding Strategies Systematically Exclude Human-Like Token Choices (viability: 3): https://sciencetostartup.com/paper/the-truncation-blind-spot-how-decoding-strategies-systematically-exclude-human-like-token-choices - This research identifies a systematic blind spot in current text generation decoding strategies that leads to detectable machine-generated text, suggesting a fundamental limitation in achieving human-like communication.
- T-QPM: Enabling Temporal Out-Of-Distribution Detection and Domain Generalization for Vision-Language Models in Open-World (viability: 7): https://sciencetostartup.com/paper/t-qpm-enabling-temporal-out-of-distribution-detection-and-domain-generalization-for-vision-language-models-in-open-world - A novel framework for robust out-of-distribution detection and domain generalization in vision-language models, addressing temporal drift and covariate shifts.
- Do Vision Language Models Understand Human Engagement in Games? (viability: 4): https://sciencetostartup.com/paper/do-vision-language-models-understand-human-engagement-in-games - This research explores the capability of Vision-Language Models to infer human engagement in video games, revealing a gap between visual cue recognition and understanding psychological states.
- WASD: Locating Critical Neurons as Sufficient Conditions for Explaining and Controlling LLM Behavior (viability: 7): https://sciencetostartup.com/paper/wasd-locating-critical-neurons-as-sufficient-conditions-for-explaining-and-controlling-llm-behavior - A framework to precisely control LLM behavior by identifying and manipulating critical neurons, enabling stable and accurate output generation.
- Cognitive Mismatch in Multimodal Large Language Models for Discrete Symbol Understanding (viability: 4): https://sciencetostartup.com/paper/cognitive-mismatch-in-multimodal-large-language-models-for-discrete-symbol-understanding - This research identifies a critical 'cognitive mismatch' in multimodal LLMs' ability to understand discrete symbols, paving the way for more robust AI systems.
- GAIN: A Benchmark for Goal-Aligned Decision-Making of Large Language Models under Imperfect Norms (viability: 7): https://sciencetostartup.com/paper/gain-a-benchmark-for-goal-aligned-decision-making-of-large-language-models-under-imperfect-norms - A new benchmark to evaluate LLM decision-making in complex norm-goal conflicts, revealing how incentives influence their adherence to rules.
- Recolour What Matters: Region-Aware Colour Editing via Token-Level Diffusion (viability: 8): https://sciencetostartup.com/paper/recolour-what-matters-region-aware-colour-editing-via-token-level-diffusion - A unified diffusion framework for precise, region-aware image color editing using token-level fusion and a novel dataset.
- MedQ-UNI: Toward Unified Medical Image Quality Assessment and Restoration via Vision-Language Modeling (viability: 7): https://sciencetostartup.com/paper/medq-uni-toward-unified-medical-image-quality-assessment-and-restoration-via-vision-language-modeling - A unified vision-language model that assesses and restores medical images across modalities and degradation types, leveraging natural language descriptions for improved fidelity and interpretability.
- AcceRL: A Distributed Asynchronous Reinforcement Learning and World Model Framework for Vision-Language-Action Models (viability: 7): https://sciencetostartup.com/paper/accerl-a-distributed-asynchronous-reinforcement-learning-and-world-model-framework-for-vision-language-action-models - A distributed asynchronous reinforcement learning framework with an integrated world model for efficient vision-language-action model training.
- AlignMamba-2: Enhancing Multimodal Fusion and Sentiment Analysis with Modality-Aware Mamba (viability: 7): https://sciencetostartup.com/paper/alignmamba-2-enhancing-multimodal-fusion-and-sentiment-analysis-with-modality-aware-mamba - A multimodal fusion framework leveraging modality-aware Mamba layers to achieve state-of-the-art sentiment analysis with improved efficiency.
- Cell-Type Prototype-Informed Neural Network for Gene Expression Estimation from Pathology Images (viability: 7): https://sciencetostartup.com/paper/cell-type-prototype-informed-neural-network-for-gene-expression-estimation-from-pathology-images - A novel neural network that estimates gene expression from pathology images by incorporating cell-type specific information, offering interpretable insights for molecular analysis.
- Interpretable Prostate Cancer Detection using a Small Cohort of MRI Images (viability: 7): https://sciencetostartup.com/paper/interpretable-prostate-cancer-detection-using-a-small-cohort-of-mri-images - An interpretable AI framework for prostate cancer detection using T2-weighted MRI, outperforming radiologists and reducing acquisition complexity.
- HypeMed: Enhancing Medication Recommendations with Hypergraph-Based Patient Relationships (viability: 7): https://sciencetostartup.com/paper/hypemed-enhancing-medication-recommendations-with-hypergraph-based-patient-relationships - HypeMed enhances medication recommendations by unifying patient visit semantics and historical data using a novel hypergraph framework, improving precision and safety.
- Impact of Differentials in SIMON32 Algorithm for Lightweight Security of Internet of Things (viability: 7): https://sciencetostartup.com/paper/impact-of-differentials-in-simon32-algorithm-for-lightweight-security-of-internet-of-things - Accelerating IoT security analysis by identifying high-probability differentials in lightweight encryption algorithms, exceeding current state-of-the-art.
- Learning Consistent Temporal Grounding between Related Tasks in Sports Coaching (viability: 4): https://sciencetostartup.com/paper/learning-consistent-temporal-grounding-between-related-tasks-in-sports-coaching - Improve temporal grounding in video analysis for sports coaching by enforcing self-consistency between related tasks, outperforming existing methods without additional annotations.
- CNT: Safety-oriented Function Reuse across LLMs via Cross-Model Neuron Transfer (viability: 7): https://sciencetostartup.com/paper/cnt-safety-oriented-function-reuse-across-llms-via-cross-model-neuron-transfer - A post-hoc method to transfer safety functionalities between LLMs by transferring neurons, enabling modular adaptation without costly retraining.
- Seeking Universal Shot Language Understanding Solutions (viability: 5): https://sciencetostartup.com/paper/seeking-universal-shot-language-understanding-solutions - A suite and models for universal shot language understanding in cinema, outperforming existing VLMs.
- SODIUM: From Open Web Data to Queryable Databases (viability: 8): https://sciencetostartup.com/paper/sodium-from-open-web-data-to-queryable-databases - An AI agent that automatically queries the open web to build structured databases for analytical tasks, achieving over 90% accuracy on a new benchmark.
- UT-ACA: Uncertainty-Triggered Adaptive Context Allocation for Long-Context Inference (viability: 4): https://sciencetostartup.com/paper/ut-aca-uncertainty-triggered-adaptive-context-allocation-for-long-context-inference - A framework to dynamically adjust context windows for long-context LLM inference based on token uncertainty, reducing computational cost.
- Discounted Beta--Bernoulli Reward Estimation for Sample-Efficient Reinforcement Learning with Verifiable Rewards (viability: 4): https://sciencetostartup.com/paper/discounted-beta-bernoulli-reward-estimation-for-sample-efficient-reinforcement-learning-with-verifiable-rewards - A novel reward estimation method for reinforcement learning that significantly improves sample efficiency and reasoning capabilities of large language models.
- SR-Nav: Spatial Relationships Matter for Zero-shot Object Goal Navigation (viability: 7): https://sciencetostartup.com/paper/sr-nav-spatial-relationships-matter-for-zero-shot-object-goal-navigation - A navigation framework that uses spatial relationships to improve object detection and planning for robots in unseen environments.
- AS2 -- Attention-Based Soft Answer Sets: An End-to-End Differentiable Neuro-Soft-Symbolic Reasoning Architecture (viability: 7): https://sciencetostartup.com/paper/as2-attention-based-soft-answer-sets-an-end-to-end-differentiable-neuro-soft-symbolic-reasoning-architecture - A fully differentiable neuro-symbolic architecture that enables end-to-end training for complex reasoning tasks, outperforming traditional methods on visual puzzles and arithmetic.
- Prompt Control-Flow Integrity: A Priority-Aware Runtime Defense Against Prompt Injection in LLM Systems (viability: 7): https://sciencetostartup.com/paper/prompt-control-flow-integrity-a-priority-aware-runtime-defense-against-prompt-injection-in-llm-systems - A priority-aware runtime defense for LLM APIs that prevents prompt injection attacks with near-zero overhead.
- MLOW: Interpretable Low-Rank Frequency Magnitude Decomposition of Multiple Effects for Time Series Forecasting (viability: 7): https://sciencetostartup.com/paper/mlow-interpretable-low-rank-frequency-magnitude-decomposition-of-multiple-effects-for-time-series-forecasting - A novel interpretable frequency-based decomposition method for time series forecasting that significantly improves existing models.
- Towards Noise-Resilient Quantum Multi-Armed and Stochastic Linear Bandits (viability: 2): https://sciencetostartup.com/paper/towards-noise-resilient-quantum-multi-armed-and-stochastic-linear-bandits - Developing noise-resilient quantum algorithms for multi-armed and stochastic linear bandits to improve performance on current quantum hardware.
- AndroTMem: From Interaction Trajectories to Anchored Memory in Long-Horizon GUI Agents (viability: 8): https://sciencetostartup.com/paper/androtmem-from-interaction-trajectories-to-anchored-memory-in-long-horizon-gui-agents - A diagnostic framework and memory mechanism for long-horizon GUI agents that significantly improves task completion rates by structuring interaction history.
- Adaptive Decoding via Test-Time Policy Learning for Self-Improving Generation (viability: 7): https://sciencetostartup.com/paper/adaptive-decoding-via-test-time-policy-learning-for-self-improving-generation - A reinforcement learning-based decoder sampler that learns to adapt LLM sampling parameters at test-time for improved and controllable generation quality without retraining.
- R&D: Balancing Reliability and Diversity in Synthetic Data Augmentation for Semantic Segmentation (viability: 7): https://sciencetostartup.com/paper/r-d-balancing-reliability-and-diversity-in-synthetic-data-augmentation-for-semantic-segmentation - A novel synthetic data augmentation pipeline using controllable diffusion models to generate diverse and reliable pixel-level semantic segmentation data, significantly improving model performance in data-scarce scenarios.
- Prune-then-Quantize or Quantize-then-Prune? Understanding the Impact of Compression Order in Joint Model Compression (viability: 3): https://sciencetostartup.com/paper/prune-then-quantize-or-quantize-then-prune-understanding-the-impact-of-compression-order-in-joint-model-compression - This paper theoretically analyzes the impact of compression order in joint model compression, proposing a hypothesis validated on language and vision models.
- Multimodal Task Interference: A Benchmark and Analysis of History-Target Mismatch in Multimodal LLMs (viability: 7): https://sciencetostartup.com/paper/multimodal-task-interference-a-benchmark-and-analysis-of-history-target-mismatch-in-multimodal-llms - A benchmark and analysis revealing directional task interference in multimodal LLMs, enabling targeted improvements for more robust conversational AI.
- SynQ: Accurate Zero-shot Quantization by Synthesis-aware Fine-tuning (viability: 7): https://sciencetostartup.com/paper/synq-accurate-zero-shot-quantization-by-synthesis-aware-fine-tuning - A novel framework for accurate zero-shot quantization of pre-trained models, enabling efficient deployment on edge devices without access to original training data.
- The Impact of Corporate AI Washing on Farmers' Digital Financial Behavior Response -- An Analysis from the Perspective of Digital Financial Exclusion (viability: 4): https://sciencetostartup.com/paper/the-impact-of-corporate-ai-washing-on-farmers-digital-financial-behavior-response-an-analysis-from-the-perspective-of-di - This research quantifies the negative impact of 'AI washing' by financial tech companies on farmers' digital financial behavior, suggesting policy interventions to improve transparency and inclusion.
- From Topic to Transition Structure: Unsupervised Concept Discovery at Corpus Scale via Predictive Associative Memory (viability: 7): https://sciencetostartup.com/paper/from-topic-to-transition-structure-unsupervised-concept-discovery-at-corpus-scale-via-predictive-associative-memory - Discovering latent transition structures in text to map literary styles and functions, enabling nuanced content analysis beyond topic.
- Mind the Rarities: Can Rare Skin Diseases Be Reliably Diagnosed via Diagnostic Reasoning? (viability: 7): https://sciencetostartup.com/paper/mind-the-rarities-can-rare-skin-diseases-be-reliably-diagnosed-via-diagnostic-reasoning - A new benchmark and evaluation metrics for large vision-language models to diagnose rare skin diseases by assessing their clinical reasoning process, not just final accuracy.
- Self-Tuning Sparse Attention: Multi-Fidelity Hyperparameter Optimization for Transformer Acceleration (viability: 5): https://sciencetostartup.com/paper/self-tuning-sparse-attention-multi-fidelity-hyperparameter-optimization-for-transformer-acceleration - Automated framework to discover optimal hyperparameters for sparse attention mechanisms in transformers, enabling plug-and-play acceleration.
- The Spillover Effects of Peer AI Rinsing on Corporate Green Innovation (viability: 3): https://sciencetostartup.com/paper/the-spillover-effects-of-peer-ai-rinsing-on-corporate-green-innovation - This paper analyzes the negative impact of corporate 'AI washing' on green innovation using large language models, proposing policy interventions for better disclosure and regulation.
- Statistical Testing Framework for Clustering Pipelines by Selective Inference (viability: 4): https://sciencetostartup.com/paper/statistical-testing-framework-for-clustering-pipelines-by-selective-inference - A statistical framework to rigorously validate the reliability of insights generated by complex data analysis pipelines.
- TARo: Token-level Adaptive Routing for LLM Test-time Alignment (viability: 5): https://sciencetostartup.com/paper/taro-token-level-adaptive-routing-for-llm-test-time-alignment - A novel method to improve LLM reasoning at inference time without retraining, by adaptively routing guidance signals.
- TopoChunker: Topology-Aware Agentic Document Chunking Framework (viability: 7): https://sciencetostartup.com/paper/topochunker-topology-aware-agentic-document-chunking-framework - TopoChunker is an agentic framework that preserves document topology for improved RAG retrieval quality, outperforming baselines while reducing token overhead.
- Efficient and Versatile Quadrupedal Skating: Optimal Co-design via Reinforcement Learning and Bayesian Optimization (viability: 3): https://sciencetostartup.com/paper/efficient-and-versatile-quadrupedal-skating-optimal-co-design-via-reinforcement-learning-and-bayesian-optimization - A research paper exploring hardware-control co-design for quadrupedal robots to achieve efficient roller skating.
- Multi-Domain Causal Empirical Bayes Under Linear Mixing (viability: 3): https://sciencetostartup.com/paper/multi-domain-causal-empirical-bayes-under-linear-mixing - A theoretical framework for estimating causal representations from multi-domain data using empirical Bayes.
- Inst4DGS: Instance-Decomposed 4D Gaussian Splatting with Multi-Video Label Permutation Learning (viability: 7): https://sciencetostartup.com/paper/inst4dgs-instance-decomposed-4d-gaussian-splatting-with-multi-video-label-permutation-learning - A novel method for instance-decomposed 4D Gaussian Splatting that achieves state-of-the-art rendering and segmentation quality by learning cross-video instance matches and optimizing object motion.
- Pixel-Accurate Epipolar Guided Matching (viability: 7): https://sciencetostartup.com/paper/pixel-accurate-epipolar-guided-matching - A novel epipolar-guided matching algorithm that achieves pixel-accurate correspondences with significant speedups for challenging computer vision tasks.
- Graph-of-Constraints Model Predictive Control for Reactive Multi-agent Task and Motion Planning (viability: 7): https://sciencetostartup.com/paper/graph-of-constraints-model-predictive-control-for-reactive-multi-agent-task-and-motion-planning - A novel framework for reactive multi-agent task and motion planning that adapts to dynamic environments and agent reassignments.
- FlowMS: Flow Matching for De Novo Structure Elucidation from Mass Spectra (viability: 7): https://sciencetostartup.com/paper/flowms-flow-matching-for-de-novo-structure-elucidation-from-mass-spectra - FlowMS uses discrete flow matching to generate novel molecular structures from mass spectrometry data, outperforming existing methods and enabling faster discovery in metabolomics and natural product research.
- RE-SAC: Disentangling aleatoric and epistemic risks in bus fleet control: A stable and robust ensemble DRL approach (viability: 7): https://sciencetostartup.com/paper/re-sac-disentangling-aleatoric-and-epistemic-risks-in-bus-fleet-control-a-stable-and-robust-ensemble-drl-approach - A robust ensemble deep reinforcement learning framework that disentangles aleatoric and epistemic risks to improve bus fleet control in volatile environments.
- Computational and Statistical Hardness of Calibration Distance (viability: 3): https://sciencetostartup.com/paper/computational-and-statistical-hardness-of-calibration-distance - This paper theoretically analyzes the computational hardness of measuring miscalibration in probabilistic predictors, proposing new algorithms and sample complexity bounds.
- AutoScreen-FW: An LLM-based Framework for Resume Screening (viability: 5): https://sciencetostartup.com/paper/autoscreen-fw-an-llm-based-framework-for-resume-screening - A framework for locally screening resumes using open-source LLMs, offering privacy and speed advantages over commercial models.
- Reflection in the Dark: Exposing and Escaping the Black Box in Reflective Prompt Optimization (viability: 7): https://sciencetostartup.com/paper/reflection-in-the-dark-exposing-and-escaping-the-black-box-in-reflective-prompt-optimization - A multi-agent framework that makes LLM prompt optimization transparent and robust, overcoming the limitations of black-box methods.
- Mathematical Foundations of Deep Learning (viability: 1): https://sciencetostartup.com/paper/mathematical-foundations-of-deep-learning - A theoretical exploration of the mathematical underpinnings of deep learning, covering approximation theory, optimal control, reinforcement learning, and generative models.
- Evolutionarily Stable Stackelberg Equilibrium (viability: 3): https://sciencetostartup.com/paper/evolutionarily-stable-stackelberg-equilibrium - A theoretical framework for modeling leader-follower dynamics in evolutionary games, applicable to biological and economic systems.
- From Weak Cues to Real Identities: Evaluating Inference-Driven De-Anonymization in LLM Agents (viability: 7): https://sciencetostartup.com/paper/from-weak-cues-to-real-identities-evaluating-inference-driven-de-anonymization-in-llm-agents - This research demonstrates how LLM agents can autonomously de-anonymize individuals by linking scattered, non-identifying data, posing a significant privacy risk that requires new evaluation methods.
- PlanTwin: Privacy-Preserving Planning Abstractions for Cloud-Assisted LLM Agents (viability: 5): https://sciencetostartup.com/paper/plantwin-privacy-preserving-planning-abstractions-for-cloud-assisted-llm-agents - PlanTwin enables cloud-assisted LLM agents to plan over private environments by projecting raw data into a privacy-preserving digital twin, maintaining planning quality with minimal utility loss.
- To See or To Please: Uncovering Visual Sycophancy and Split Beliefs in VLMs (viability: 7): https://sciencetostartup.com/paper/to-see-or-to-please-uncovering-visual-sycophancy-and-split-beliefs-in-vlms - A diagnostic framework for Visual Language Models that uncovers and quantifies 'visual sycophancy', enabling a selective prediction strategy to improve accuracy without retraining.
- Contact Status Recognition and Slip Detection with a Bio-inspired Tactile Hand (viability: 4): https://sciencetostartup.com/paper/contact-status-recognition-and-slip-detection-with-a-bio-inspired-tactile-hand - A bio-inspired robotic hand uses multimodal tactile feedback and wavelet transform to recognize contact status and detect object slip with high accuracy.
- PowerFlow: Unlocking the Dual Nature of LLMs via Principled Distribution Matching (viability: 7): https://sciencetostartup.com/paper/powerflow-unlocking-the-dual-nature-of-llms-via-principled-distribution-matching - A principled framework for fine-tuning LLMs to unlock dual capabilities like logical reasoning and creative expression by matching distributions, outperforming existing methods.
- Synthetic Data Generation for Training Diversified Commonsense Reasoning Models (viability: 7): https://sciencetostartup.com/paper/synthetic-data-generation-for-training-diversified-commonsense-reasoning-models - We generate a synthetic dataset to train conversational agents for diverse and high-quality commonsense reasoning, overcoming the limitations of existing human-crafted datasets.
- From Noise to Signal: When Outliers Seed New Topics (viability: 4): https://sciencetostartup.com/paper/from-noise-to-signal-when-outliers-seed-new-topics - Identify emerging topics from outlier documents in news streams by analyzing their temporal trajectories.
- LGESynthNet: Controlled Scar Synthesis for Improved Scar Segmentation in Cardiac LGE-MRI Imaging (viability: 7): https://sciencetostartup.com/paper/lgesynthnet-controlled-scar-synthesis-for-improved-scar-segmentation-in-cardiac-lge-mri-imaging - Generate controllable synthetic cardiac MRI scar data to significantly improve downstream segmentation accuracy, overcoming annotation limitations.
- Pushan: Trace-Free Deobfuscation of Virtualization-Obfuscated Binaries (viability: 7): https://sciencetostartup.com/paper/pushan-trace-free-deobfuscation-of-virtualization-obfuscated-binaries - A novel, trace-free technique for deobfuscating and decompiling virtualization-obfuscated binaries, enabling advanced malware analysis and LLM-assisted code understanding.
- Multi-material Direct Ink Writing and Embroidery for Stretchable Wearable Sensors (viability: 7): https://sciencetostartup.com/paper/multi-material-direct-ink-writing-and-embroidery-for-stretchable-wearable-sensors - A novel fabrication process integrates direct ink writing with embroidery to create durable, stretchable strain sensors directly embedded into garments for motion tracking and soft robotics.
- Interpretability without actionability: mechanistic methods cannot correct language model errors despite near-perfect internal representations (viability: 4): https://sciencetostartup.com/paper/interpretability-without-actionability-mechanistic-methods-cannot-correct-language-model-errors-despite-near-perfect-int - This research investigates the limitations of current mechanistic interpretability methods in correcting language model errors, finding they cannot reliably bridge the gap between internal knowledge and actionable output correction.
- Large-Scale Analysis of Political Propaganda on Moltbook (viability: 5): https://sciencetostartup.com/paper/large-scale-analysis-of-political-propaganda-on-moltbook - Develops LLM classifiers to detect political propaganda on social media platforms, providing insights into its prevalence and dissemination.
- Epistemic Generative Adversarial Networks (viability: 3): https://sciencetostartup.com/paper/epistemic-generative-adversarial-networks - A theoretical framework to improve GAN output diversity by incorporating uncertainty quantification.
- HRI-SA: A Multimodal Dataset for Online Assessment of Human Situational Awareness during Remote Human-Robot Teaming (viability: 5): https://sciencetostartup.com/paper/hri-sa-a-multimodal-dataset-for-online-assessment-of-human-situational-awareness-during-remote-human-robot-teaming - A multimodal dataset and initial models for automatically detecting human situational awareness gaps in remote human-robot teaming.
- VISTA: Validation-Guided Integration of Spatial and Temporal Foundation Models with Anatomical Decoding for Rare-Pathology VCE Event Detection (viability: 4): https://sciencetostartup.com/paper/vista-validation-guided-integration-of-spatial-and-temporal-foundation-models-with-anatomical-decoding-for-rare-patholog - A novel framework for rare-pathology detection in capsule endoscopy videos using integrated spatial and temporal foundation models with anatomical decoding.
- Shifting Uncertainty to Critical Moments: Towards Reliable Uncertainty Quantification for VLA Model (viability: 7): https://sciencetostartup.com/paper/shifting-uncertainty-to-critical-moments-towards-reliable-uncertainty-quantification-for-vla-model - A novel uncertainty quantification method for Vision-Language-Action models that reliably predicts robotic failures by preserving transient risk signals and emphasizing critical motion dynamics.
- ManiDreams: An Open-Source Library for Robust Object Manipulation via Uncertainty-aware Task-specific Intuitive Physics (viability: 7): https://sciencetostartup.com/paper/manidreams-an-open-source-library-for-robust-object-manipulation-via-uncertainty-aware-task-specific-intuitive-physics - A framework for robust robotic manipulation by explicitly handling uncertainty in planning.
- Can LLMs Reason Like Automated Theorem Provers for Rust Verification? VCoT-Bench: Evaluating via Verification Chain of Thought (viability: 5): https://sciencetostartup.com/paper/can-llms-reason-like-automated-theorem-provers-for-rust-verification-vcot-bench-evaluating-via-verification-chain-of-tho - A new benchmark and framework to evaluate LLMs' reasoning capabilities in Rust program verification, revealing their current limitations compared to automated theorem provers.
- Understanding the Theoretical Foundations of Deep Neural Networks through Differential Equations (viability: 2): https://sciencetostartup.com/paper/understanding-the-theoretical-foundations-of-deep-neural-networks-through-differential-equations - This paper explores the theoretical underpinnings of deep neural networks using differential equations to guide future research and development.
- MemArchitect: A Policy Driven Memory Governance Layer (viability: 7): https://sciencetostartup.com/paper/memarchitect-a-policy-driven-memory-governance-layer - MemArchitect provides a policy-driven governance layer for LLM agent memory, ensuring reliability and safety by resolving contradictions and managing privacy.
- FaithSteer-BENCH: A Deployment-Aligned Stress-Testing Benchmark for Inference-Time Steering (viability: 7): https://sciencetostartup.com/paper/faithsteer-bench-a-deployment-aligned-stress-testing-benchmark-for-inference-time-steering - A stress-testing benchmark for LLM inference-time steering methods to ensure reliability and robustness in real-world deployments.
- A Family of Adaptive Activation Functions for Mitigating Failure Modes in Physics-Informed Neural Networks (viability: 5): https://sciencetostartup.com/paper/a-family-of-adaptive-activation-functions-for-mitigating-failure-modes-in-physics-informed-neural-networks - Develop adaptive wavelet-based activation functions to improve the stability and accuracy of Physics-Informed Neural Networks for scientific and engineering problems.
- Consumer-to-Clinical Language Shifts in Ambient AI Draft Notes and Clinician-Finalized Documentation: A Multi-level Analysis (viability: 4): https://sciencetostartup.com/paper/consumer-to-clinical-language-shifts-in-ambient-ai-draft-notes-and-clinician-finalized-documentation-a-multi-level-analy - An AI that learns to translate consumer language in draft clinical notes to standardized medical terminology, improving documentation accuracy and efficiency.
- Escaping Offline Pessimism: Vector-Field Reward Shaping for Safe Frontier Exploration (viability: 5): https://sciencetostartup.com/paper/escaping-offline-pessimism-vector-field-reward-shaping-for-safe-frontier-exploration - A novel reward shaping method for offline reinforcement learning that encourages safe and continuous exploration of new data frontiers.
- Learning to Reason with Curriculum I: Provable Benefits of Autocurriculum (viability: 4): https://sciencetostartup.com/paper/learning-to-reason-with-curriculum-i-provable-benefits-of-autocurriculum - Reduce the immense cost of training chain-of-thought reasoning models by adaptively selecting training data based on model performance.
- DriveVLM-RL: Neuroscience-Inspired Reinforcement Learning with Vision-Language Models for Safe and Deployable Autonomous Driving (viability: 8): https://sciencetostartup.com/paper/drivevlm-rl-neuroscience-inspired-reinforcement-learning-with-vision-language-models-for-safe-and-deployable-autonomous - A neuroscience-inspired reinforcement learning framework integrating vision-language models for safer and deployable autonomous driving, achieving real-time feasibility by removing VLM inference at deployment.
- Approximate Subgraph Matching with Neural Graph Representations and Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/approximate-subgraph-matching-with-neural-graph-representations-and-reinforcement-learning - Leveraging graph transformers and reinforcement learning to significantly improve approximate subgraph matching for complex graph analysis tasks.
- Unrolled Reconstruction with Integrated Super-Resolution for Accelerated 3D LGE MRI (viability: 7): https://sciencetostartup.com/paper/unrolled-reconstruction-with-integrated-super-resolution-for-accelerated-3d-lge-mri - Accelerate 3D LGE MRI reconstruction by integrating super-resolution directly into the unrolled optimization loop for improved cardiac structure detail and segmentation.
- Proprioceptive-only State Estimation for Legged Robots with Set-Coverage Measurements of Learned Dynamics (viability: 5): https://sciencetostartup.com/paper/proprioceptive-only-state-estimation-for-legged-robots-with-set-coverage-measurements-of-learned-dynamics - A robust state estimation framework for legged robots using proprioceptive data and set-coverage measurements to overcome noise limitations.
- Fast and Generalizable NeRF Architecture Selection for Satellite Scene Reconstruction (viability: 7): https://sciencetostartup.com/paper/fast-and-generalizable-nerf-architecture-selection-for-satellite-scene-reconstruction - A predictive framework that selects optimal Neural Radiance Field architectures for satellite scene reconstruction in seconds, drastically reducing training time and power consumption.
- Auditing Preferences for Brands and Cultures in LLMs (viability: 5): https://sciencetostartup.com/paper/auditing-preferences-for-brands-and-cultures-in-llms - A framework for auditing brand and cultural preferences in LLMs to ensure market fairness and diverse information exposure.
- ALIGN: Adversarial Learning for Generalizable Speech Neuroprosthesis (viability: 4): https://sciencetostartup.com/paper/align-adversarial-learning-for-generalizable-speech-neuroprosthesis - A framework for improving speech decoding in brain-computer interfaces by making models robust to changes across recording sessions.
- Sparse3DTrack: Monocular 3D Object Tracking Using Sparse Supervision (viability: 7): https://sciencetostartup.com/paper/sparse3dtrack-monocular-3d-object-tracking-using-sparse-supervision - A sparsely supervised framework for monocular 3D object tracking that significantly reduces annotation costs while improving performance.
- Path-Constrained Mixture-of-Experts (viability: 7): https://sciencetostartup.com/paper/path-constrained-mixture-of-experts - A novel Mixture-of-Experts architecture that improves LLM efficiency and performance by constraining expert paths, leading to better linguistic function clustering and robustness.
- The Validity Gap in Health AI Evaluation: A Cross-Sectional Analysis of Benchmark Composition (viability: 4): https://sciencetostartup.com/paper/the-validity-gap-in-health-ai-evaluation-a-cross-sectional-analysis-of-benchmark-composition - This research identifies a critical gap in health AI evaluation benchmarks, highlighting the need for standardized patient population profiling to ensure model readiness for clinical use.
- CORE: Robust Out-of-Distribution Detection via Confidence and Orthogonal Residual Scoring (viability: 7): https://sciencetostartup.com/paper/core-robust-out-of-distribution-detection-via-confidence-and-orthogonal-residual-scoring - A novel OOD detection method that combines classifier confidence with a residual signal to achieve robust performance across various architectures and datasets.
- Offload or Overload: A Platform Measurement Study of Mobile Robotic Manipulation Workloads (viability: 3): https://sciencetostartup.com/paper/offload-or-overload-a-platform-measurement-study-of-mobile-robotic-manipulation-workloads - A study measuring the computational demands of mobile robotic manipulation to inform the design of inference systems.
- CycleCap: Improving VLMs Captioning Performance via Self-Supervised Cycle Consistency Fine-Tuning (viability: 7): https://sciencetostartup.com/paper/cyclecap-improving-vlms-captioning-performance-via-self-supervised-cycle-consistency-fine-tuning - CycleCap fine-tunes Visual-Language Models for improved image captioning using self-supervised cycle consistency, eliminating the need for large annotated datasets.
- On Additive Gaussian Processes for Wind Farm Power Prediction (viability: 4): https://sciencetostartup.com/paper/on-additive-gaussian-processes-for-wind-farm-power-prediction - Leveraging additive Gaussian processes to predict wind farm power generation by analyzing variations across individual turbines and the entire farm.
- Detection Is Cheap, Routing Is Learned: Why Refusal-Based Alignment Evaluation Fails (viability: 5): https://sciencetostartup.com/paper/detection-is-cheap-routing-is-learned-why-refusal-based-alignment-evaluation-fails - Develops a novel evaluation framework for LLM alignment that moves beyond simple refusal detection to analyze the internal routing mechanisms responsible for censorship and factual output.
- EDM-ARS: A Domain-Specific Multi-Agent System for Automated Educational Data Mining Research (viability: 7): https://sciencetostartup.com/paper/edm-ars-a-domain-specific-multi-agent-system-for-automated-educational-data-mining-research - An open-source multi-agent system that automates end-to-end educational data mining research, generating complete LaTeX manuscripts with validated analyses.
- Retrieval-Augmented LLM Agents: Learning to Learn from Experience (viability: 7): https://sciencetostartup.com/paper/retrieval-augmented-llm-agents-learning-to-learn-from-experience - A novel framework for training retrieval-augmented LLM agents that significantly improves generalization to unseen tasks by integrating experience retrieval into fine-tuning.
- SG-CoT: An Ambiguity-Aware Robotic Planning Framework using Scene Graph Representations (viability: 7): https://sciencetostartup.com/paper/sg-cot-an-ambiguity-aware-robotic-planning-framework-using-scene-graph-representations - A framework that uses scene graphs to resolve ambiguities in LLM-driven robotic planning, improving reliability and success rates.
- Enactor: From Traffic Simulators to Surrogate World Models (viability: 7): https://sciencetostartup.com/paper/enactor-from-traffic-simulators-to-surrogate-world-models - A transformer-based generative model that creates realistic, long-horizon traffic actor trajectories by understanding intersection geometry and actor interactions, outperforming baselines significantly.
- LRConv-NeRV: Low Rank Convolution for Efficient Neural Video Compression (viability: 7): https://sciencetostartup.com/paper/lrconv-nerv-low-rank-convolution-for-efficient-neural-video-compression - This research introduces an efficient neural video compression method by replacing dense convolutions with low-rank approximations, significantly reducing computational cost and model size with minimal quality loss.
- Manufacturing Micro-Patterned Surfaces with Multi-Robot Systems (viability: 5): https://sciencetostartup.com/paper/manufacturing-micro-patterned-surfaces-with-multi-robot-systems - Leveraging multi-robot systems and ergodic control to enable scalable, cost-effective manufacturing of micro-patterned surfaces with novel physical properties.
- Sharpness-Aware Minimization in Logit Space Efficiently Enhances Direct Preference Optimization (viability: 7): https://sciencetostartup.com/paper/sharpness-aware-minimization-in-logit-space-efficiently-enhances-direct-preference-optimization - Enhance LLM preference alignment by efficiently mitigating the squeezing effect in Direct Preference Optimization using a novel Sharpness-Aware Minimization technique.
- Discovering What You Can Control: Interventional Boundary Discovery for Reinforcement Learning (viability: 5): https://sciencetostartup.com/paper/discovering-what-you-can-control-interventional-boundary-discovery-for-reinforcement-learning - A novel method for reinforcement learning that identifies and filters out irrelevant or confounding state dimensions to improve agent performance and robustness.
- MolRGen: A Training and Evaluation Setting for De Novo Molecular Generation with Reasonning Models (viability: 5): https://sciencetostartup.com/paper/molrgen-a-training-and-evaluation-setting-for-de-novo-molecular-generation-with-reasonning-models - A new benchmark and dataset for training reasoning LLMs to generate novel, high-quality, and diverse molecules for drug discovery.
- Computation-Utility-Privacy Tradeoffs in Bayesian Estimation (viability: 3): https://sciencetostartup.com/paper/computation-utility-privacy-tradeoffs-in-bayesian-estimation - Develops efficient algorithms for differentially private Bayesian estimation with theoretical utility guarantees, addressing a gap in existing methods for Gaussian mean estimation and linear regression.
- AGRI-Fidelity: Evaluating the Reliability of Listenable Explanations for Poultry Disease Detection (viability: 4): https://sciencetostartup.com/paper/agri-fidelity-evaluating-the-reliability-of-listenable-explanations-for-poultry-disease-detection - A new framework to evaluate the reliability of AI explanations for detecting poultry diseases, ensuring explanations are based on actual disease indicators and not environmental noise.
- Rapid Adaptation of Particle Dynamics for Generalized Deformable Object Mobile Manipulation (viability: 7): https://sciencetostartup.com/paper/rapid-adaptation-of-particle-dynamics-for-generalized-deformable-object-mobile-manipulation - A novel method for real-robot deformable object manipulation that adapts to unknown object dynamics, achieving over 80% success rates on complex tasks.
- Who Tests the Testers? Systematic Enumeration and Coverage Audit of LLM Agent Tool Call Safety (viability: 7): https://sciencetostartup.com/paper/who-tests-the-testers-systematic-enumeration-and-coverage-audit-of-llm-agent-tool-call-safety - A meta-audit framework that systematically enumerates LLM agent tool-call safety vulnerabilities missed by existing benchmarks.
- ReDAG-RT: Global Rate-Priority Scheduling for Real-Time Multi-DAG Execution in ROS 2 (viability: 7): https://sciencetostartup.com/paper/redag-rt-global-rate-priority-scheduling-for-real-time-multi-dag-execution-in-ros-2 - A user-space global scheduler for ROS 2 that enforces deterministic multi-DAG execution and improves real-time performance for safety-critical robotic systems.
- Gradient-Informed Temporal Sampling Improves Rollout Accuracy in PDE Surrogate Training (viability: 5): https://sciencetostartup.com/paper/gradient-informed-temporal-sampling-improves-rollout-accuracy-in-pde-surrogate-training - A novel data sampling method for neural simulators that improves rollout accuracy by intelligently selecting training data based on model gradients and temporal coverage.
- Toward Reliable, Safe, and Secure LLMs for Scientific Applications (viability: 4): https://sciencetostartup.com/paper/toward-reliable-safe-and-secure-llms-for-scientific-applications - Developing a novel framework and automated benchmark generation system to ensure the reliability, safety, and security of AI scientists for scientific applications.
- A Hybrid Conditional Diffusion-DeepONet Framework for High-Fidelity Stress Prediction in Hyperelastic Materials (viability: 7): https://sciencetostartup.com/paper/a-hybrid-conditional-diffusion-deeponet-framework-for-high-fidelity-stress-prediction-in-hyperelastic-materials - A hybrid AI framework that significantly improves stress prediction accuracy in hyperelastic materials by combining diffusion models and neural operators, outperforming existing methods by orders of magnitude.
- Semantic Segmentation and Depth Estimation for Real-Time Lunar Surface Mapping Using 3D Gaussian Splatting (viability: 7): https://sciencetostartup.com/paper/semantic-segmentation-and-depth-estimation-for-real-time-lunar-surface-mapping-using-3d-gaussian-splatting - A real-time lunar surface mapping framework using 3D Gaussian Splatting for precise navigation and scene understanding.
- GoalVLM: VLM-driven Object Goal Navigation for Multi-Agent System (viability: 5): https://sciencetostartup.com/paper/goalvlm-vlm-driven-object-goal-navigation-for-multi-agent-system - A multi-agent framework enabling robots to navigate to novel objects described by natural language, without task-specific retraining.
- How Psychological Learning Paradigms Shaped and Constrained Artificial Intelligence (viability: 2): https://sciencetostartup.com/paper/how-psychological-learning-paradigms-shaped-and-constrained-artificial-intelligence - This paper theoretically analyzes the influence of psychological learning paradigms on AI, proposing a new modular framework for artificial general intelligence.
- R2-Dreamer: Redundancy-Reduced World Models without Decoders or Augmentation (viability: 7): https://sciencetostartup.com/paper/r2-dreamer-redundancy-reduced-world-models-without-decoders-or-augmentation - A decoder-free model-based reinforcement learning framework that uses an internal redundancy-reduction objective to learn efficient representations without data augmentation, achieving competitive performance and faster training.
- A Computationally Efficient Learning of Artificial Intelligence System Reliability Considering Error Propagation (viability: 5): https://sciencetostartup.com/paper/a-computationally-efficient-learning-of-artificial-intelligence-system-reliability-considering-error-propagation - A new framework to model and predict AI system reliability by accounting for error propagation, using simulation data and an efficient estimation algorithm.
- Access Controlled Website Interaction for Agentic AI with Delegated Critical Tasks (viability: 7): https://sciencetostartup.com/paper/access-controlled-website-interaction-for-agentic-ai-with-delegated-critical-tasks - Enabling secure delegation of critical tasks to AI agents through fine-grained website access control.
- Retrieval-Augmented LLMs for Security Incident Analysis (viability: 8): https://sciencetostartup.com/paper/retrieval-augmented-llms-for-security-incident-analysis - A RAG-based system that automates security incident analysis by filtering logs and semantically reasoning with LLMs to reconstruct attack sequences, offering cost-effective and accurate insights.
- MicroVision: An Open Dataset and Benchmark Models for Detecting Vulnerable Road Users and Micromobility Vehicles (viability: 7): https://sciencetostartup.com/paper/microvision-an-open-dataset-and-benchmark-models-for-detecting-vulnerable-road-users-and-micromobility-vehicles - An open dataset and benchmark models for detecting vulnerable road users and micromobility vehicles from a user's perspective to improve traffic safety.
- Starting Off on the Wrong Foot: Pitfalls in Data Preparation (viability: 5): https://sciencetostartup.com/paper/starting-off-on-the-wrong-foot-pitfalls-in-data-preparation - A statistically rigorous data preparation framework for insurance loss modeling that improves model robustness and reduces computational costs.
- TeachingCoach: A Fine-Tuned Scaffolding Chatbot for Instructional Guidance to Instructors (viability: 7): https://sciencetostartup.com/paper/teachingcoach-a-fine-tuned-scaffolding-chatbot-for-instructional-guidance-to-instructors - A specialized chatbot that provides timely, pedagogically grounded guidance to higher education instructors, outperforming generic AI models.
- CWoMP: Morpheme Representation Learning for Interlinear Glossing (viability: 7): https://sciencetostartup.com/paper/cwomp-morpheme-representation-learning-for-interlinear-glossing - A contrastive learning model that automates the creation of linguistically rich interlinear glossed text by learning morpheme representations, outperforming existing methods and allowing for user-driven lexicon expansion.
- VLM-AutoDrive: Post-Training Vision-Language Models for Safety-Critical Autonomous Driving Events (viability: 7): https://sciencetostartup.com/paper/vlm-autodrive-post-training-vision-language-models-for-safety-critical-autonomous-driving-events - A post-training framework to adapt general vision-language models for high-fidelity, safety-critical event detection in autonomous driving.
- Conflict-Free Policy Languages for Probabilistic ML Predicates: A Framework and Case Study with the Semantic Router DSL (viability: 5): https://sciencetostartup.com/paper/conflict-free-policy-languages-for-probabilistic-ml-predicates-a-framework-and-case-study-with-the-semantic-router-dsl - A new framework for conflict-free policy languages in probabilistic ML systems, specifically for LLM inference routing, preventing silent routing errors.
- GRAFITE: Generative Regression Analysis Framework for Issue Tracking and Evaluation (viability: 5): https://sciencetostartup.com/paper/grafite-generative-regression-analysis-framework-for-issue-tracking-and-evaluation - GRAFITE offers a novel framework for enhancing issue tracking with generative regression analysis.
- Modeling the human lexicon under temperature variations: linguistic factors, diversity and typicality in LLM word associations (viability: 5): https://sciencetostartup.com/paper/modeling-the-human-lexicon-under-temperature-variations-linguistic-factors-diversity-and-typicality-in-llm-word-associat - This research analyzes how LLMs represent human word associations, revealing differences in variability and typicality based on model size and temperature settings.
- ResNets of All Shapes and Sizes: Convergence of Training Dynamics in the Large-scale Limit (viability: 2): https://sciencetostartup.com/paper/resnets-of-all-shapes-and-sizes-convergence-of-training-dynamics-in-the-large-scale-limit - This paper theoretically analyzes the convergence of training dynamics for residual neural networks in a large-scale limit, providing error bounds for deep architectures.
- Efficient Dense Crowd Trajectory Prediction Via Dynamic Clustering (viability: 7): https://sciencetostartup.com/paper/efficient-dense-crowd-trajectory-prediction-via-dynamic-clustering - A plug-and-play clustering method to significantly speed up and reduce memory usage for dense crowd trajectory prediction, compatible with existing systems.
- How LLMs Distort Our Written Language (viability: 7): https://sciencetostartup.com/paper/how-llms-distort-our-written-language - This research reveals how LLMs subtly distort written meaning and impact scientific discourse, creating a need for tools to detect and mitigate these effects.
- Learning-Augmented Algorithms for $k$-median via Online Learning (viability: 3): https://sciencetostartup.com/paper/learning-augmented-algorithms-for-k-median-via-online-learning - An online learning framework for the k-median problem that adapts to dynamically changing data sequences.
- Unified Spatio-Temporal Token Scoring for Efficient Video VLMs (viability: 3): https://sciencetostartup.com/paper/unified-spatio-temporal-token-scoring-for-efficient-video-vlms - Introducing a novel token pruning technique for enhancing efficiency in vision-language models for video tasks.
- Universal Skeleton Understanding via Differentiable Rendering and MLLMs (viability: 7): https://sciencetostartup.com/paper/universal-skeleton-understanding-via-differentiable-rendering-and-mllms - SkeletonLLM translates human skeleton sequences into visual representations for enhanced multimodal reasoning.
- Loc3R-VLM: Language-based Localization and 3D Reasoning with Vision-Language Models (viability: 8): https://sciencetostartup.com/paper/loc3r-vlm-language-based-localization-and-3d-reasoning-with-vision-language-models - Loc3R-VLM enhances Vision-Language Models with advanced 3D understanding for improved spatial reasoning.
- EchoGen: Cycle-Consistent Learning for Unified Layout-Image Generation and Understanding (viability: 7): https://sciencetostartup.com/paper/echogen-cycle-consistent-learning-for-unified-layout-image-generation-and-understanding - EchoGen is a unified framework that enhances layout-to-image generation and image grounding through innovative training strategies.
- AgentFactory: A Self-Evolving Framework Through Executable Subagent Accumulation and Reuse (viability: 5): https://sciencetostartup.com/paper/agentfactory-a-self-evolving-framework-through-executable-subagent-accumulation-and-reuse - Develop a framework that accumulates and reuses executable subagents for self-evolving AI systems.
- The Unreasonable Effectiveness of Text Embedding Interpolation for Continuous Image Steering (viability: 7): https://sciencetostartup.com/paper/the-unreasonable-effectiveness-of-text-embedding-interpolation-for-continuous-image-steering - A training-free framework for continuous and controllable image editing using text embeddings.
- LoST: Level of Semantics Tokenization for 3D Shapes (viability: 7): https://sciencetostartup.com/paper/lost-level-of-semantics-tokenization-for-3d-shapes - LoST revolutionizes 3D shape generation by introducing a semantic tokenization method that enhances autoregressive modeling efficiency.
- GMT: Goal-Conditioned Multimodal Transformer for 6-DOF Object Trajectory Synthesis in 3D Scenes (viability: 8): https://sciencetostartup.com/paper/gmt-goal-conditioned-multimodal-transformer-for-6-dof-object-trajectory-synthesis-in-3d-scenes - GMT is a multimodal transformer that generates realistic 6-DOF object manipulation trajectories for robots in complex 3D environments.
- A Single-Fiber Optical Frequency Domain Reflectometry (OFDR)-Based Shape Sensing of Concentric Tube Steerable Drilling Robots (viability: 4): https://sciencetostartup.com/paper/a-single-fiber-optical-frequency-domain-reflectometry-ofdr-based-shape-sensing-of-concentric-tube-steerable-drilling-rob - A novel shape-sensing method for steerable drilling robots using Optical Frequency Domain Reflectometry.
- Versatile Editing of Video Content, Actions, and Dynamics without Training (viability: 8): https://sciencetostartup.com/paper/versatile-editing-of-video-content-actions-and-dynamics-without-training - DynaEdit enables versatile, training-free video editing using pretrained models for complex scene interactions.
- Feeling the Space: Egomotion-Aware Video Representation for Efficient and Accurate 3D Scene Understanding (viability: 3): https://sciencetostartup.com/paper/feeling-the-space-egomotion-aware-video-representation-for-efficient-and-accurate-3d-scene-understanding - Motion-MLLM enhances spatial reasoning in 3D scenes using egomotion data for improved efficiency.
- AdaRadar: Rate Adaptive Spectral Compression for Radar-based Perception (viability: 7): https://sciencetostartup.com/paper/adaradar-rate-adaptive-spectral-compression-for-radar-based-perception - AdaRadar offers an adaptive compression solution for radar data in autonomous driving, enhancing data transmission efficiency.
- AHOY! Animatable Humans under Occlusion from YouTube Videos with Gaussian Splatting and Video Diffusion Priors (viability: 8): https://sciencetostartup.com/paper/ahoy-animatable-humans-under-occlusion-from-youtube-videos-with-gaussian-splatting-and-video-diffusion-priors - AHOY reconstructs animatable 3D avatars from occluded YouTube videos using advanced Gaussian splatting techniques.
- Toward Scalable Automated Repository-Level Datasets for Software Vulnerability Detection (viability: 5): https://sciencetostartup.com/paper/toward-scalable-automated-repository-level-datasets-for-software-vulnerability-detection - Automated benchmark generator for scalable software vulnerability detection datasets.
- TDAD: Test-Driven Agentic Development - Reducing Code Regressions in AI Coding Agents via Graph-Based Impact Analysis (viability: 6): https://sciencetostartup.com/paper/tdad-test-driven-agentic-development-reducing-code-regressions-in-ai-coding-agents-via-graph-based-impact-analysis - Reduce code regressions in AI coding agents with Test-Driven Agentic Development using graph-based impact analysis.
- Beyond Muon: MUD (MomentUm Decorrelation) for Faster Transformer Training (viability: 7): https://sciencetostartup.com/paper/beyond-muon-mud-momentum-decorrelation-for-faster-transformer-training - MUD optimizes transformer training by enhancing momentum updates for faster convergence.
- Specification-Aware Distribution Shaping for Robotics Foundation Models (viability: 7): https://sciencetostartup.com/paper/specification-aware-distribution-shaping-for-robotics-foundation-models - A framework for optimizing action distributions in robotics models to ensure compliance with complex temporal specifications.
- Robust-ComBat: Mitigating Outlier Effects in Diffusion MRI Data Harmonization (viability: 7): https://sciencetostartup.com/paper/robust-combat-mitigating-outlier-effects-in-diffusion-mri-data-harmonization - Robust-ComBat enhances diffusion MRI data harmonization by effectively mitigating outlier effects in patients with neurological disorders.
- LaDe: Unified Multi-Layered Graphic Media Generation and Decomposition (viability: 3): https://sciencetostartup.com/paper/lade-unified-multi-layered-graphic-media-generation-and-decomposition - LaDe enables the creation of fully editable layered design documents from natural language prompts.
- ConGA: Guidelines for Contextual Gender Annotation. A Framework for Annotating Gender in Machine Translation (viability: 5): https://sciencetostartup.com/paper/conga-guidelines-for-contextual-gender-annotation-a-framework-for-annotating-gender-in-machine-translation - ConGA provides a framework for gender annotation in machine translation to reduce bias and improve accuracy.
- Gender Disambiguation in Machine Translation: Diagnostic Evaluation in Decoder-Only Architectures (viability: 4): https://sciencetostartup.com/paper/gender-disambiguation-in-machine-translation-diagnostic-evaluation-in-decoder-only-architectures - A novel framework for evaluating and mitigating gender bias in machine translation models.
- VideoAtlas: Navigating Long-Form Video in Logarithmic Compute (viability: 6): https://sciencetostartup.com/paper/videoatlas-navigating-long-form-video-in-logarithmic-compute - VideoAtlas offers a lossless, scalable environment for navigating long-form video using hierarchical grids.
- Unified Policy Value Decomposition for Rapid Adaptation (viability: 3): https://sciencetostartup.com/paper/unified-policy-value-decomposition-for-rapid-adaptation - A framework for rapid adaptation in reinforcement learning using shared low-dimensional goal embeddings.
- CARE: Covariance-Aware and Rank-Enhanced Decomposition for Enabling Multi-Head Latent Attention (viability: 3): https://sciencetostartup.com/paper/care-covariance-aware-and-rank-enhanced-decomposition-for-enabling-multi-head-latent-attention - CARE enhances multi-head latent attention conversion for improved inference efficiency.
- ShapleyLaw: A Game-Theoretic Approach to Multilingual Scaling Laws (viability: 2): https://sciencetostartup.com/paper/shapleylaw-a-game-theoretic-approach-to-multilingual-scaling-laws - ShapleyLaw introduces a game-theoretic approach to optimize multilingual pretraining by quantifying cross-lingual transfer effects.
- TransText: Transparency Aware Image-to-Video Typography Animation (viability: 7): https://sciencetostartup.com/paper/transtext-transparency-aware-image-to-video-typography-animation - TransText enables high-fidelity, layer-aware text animation for dynamic visual design.
- Efficient Training-Free Multi-Token Prediction via Embedding-Space Probing (viability: 7): https://sciencetostartup.com/paper/efficient-training-free-multi-token-prediction-via-embedding-space-probing - A training-free method for efficient multi-token prediction in large language models that enhances throughput and accuracy.
- Interpretable Traffic Responsibility from Dashcam Video via Legal Multi Agent Reasoning (viability: 8): https://sciencetostartup.com/paper/interpretable-traffic-responsibility-from-dashcam-video-via-legal-multi-agent-reasoning - C-TRAIL transforms dashcam video evidence into legal responsibility assessments using a multimodal approach.
- RoboForge: Physically Optimized Text-guided Whole-Body Locomotion for Humanoids (viability: 3): https://sciencetostartup.com/paper/roboforge-physically-optimized-text-guided-whole-body-locomotion-for-humanoids - A unified framework for optimizing text-guided locomotion in humanoid robots through physics-based motion generation.
- A practical artificial intelligence framework for legal age estimation using clavicle computed tomography scans (viability: 8): https://sciencetostartup.com/paper/a-practical-artificial-intelligence-framework-for-legal-age-estimation-using-clavicle-computed-tomography-scans - A robust AI framework for legal age estimation using clavicle CT scans, enhancing forensic decision-making.
- Multi-Armed Sequential Hypothesis Testing by Betting (viability: 2): https://sciencetostartup.com/paper/multi-armed-sequential-hypothesis-testing-by-betting - A theoretical framework for optimizing multi-armed sequential hypothesis testing through betting strategies.
- SegFly: A 2D-3D-2D Paradigm for Aerial RGB-Thermal Semantic Segmentation at Scale (viability: 7): https://sciencetostartup.com/paper/segfly-a-2d-3d-2d-paradigm-for-aerial-rgb-thermal-semantic-segmentation-at-scale - Aerial RGB-Thermal Segmentation tool enabling enhanced drone-based surveillance and monitoring.
- Only relative ranks matter in weight-clustered large language models (viability: 3): https://sciencetostartup.com/paper/only-relative-ranks-matter-in-weight-clustered-large-language-models - A novel approach to compress large language models by focusing on the relative rank of weights.
- IndicSafe: A Benchmark for Evaluating Multilingual LLM Safety in South Asia (viability: 6): https://sciencetostartup.com/paper/indicsafe-a-benchmark-for-evaluating-multilingual-llm-safety-in-south-asia - IndicSafe is a benchmark for evaluating the safety of multilingual LLMs in culturally diverse South Asian contexts.
- Noise-Aware Misclassification Attack Detection in Collaborative DNN Inference (viability: 5): https://sciencetostartup.com/paper/noise-aware-misclassification-attack-detection-in-collaborative-dnn-inference - A noise-aware anomaly detection framework for securing collaborative DNN inference against misclassification attacks.
- Pretrained Multilingual Transformers Reveal Quantitative Distance Between Human Languages (viability: 5): https://sciencetostartup.com/paper/pretrained-multilingual-transformers-reveal-quantitative-distance-between-human-languages - A method leveraging pretrained multilingual models to quantitatively measure language distance for linguistic analysis.
- SpiderCam: Low-Power Snapshot Depth from Differential Defocus (viability: 2): https://sciencetostartup.com/paper/spidercam-low-power-snapshot-depth-from-differential-defocus - SpiderCam is a low-power FPGA-based camera that captures real-time depth maps using differential defocus.
- Differential Privacy in Generative AI Agents: Analysis and Optimal Tradeoffs (viability: 5): https://sciencetostartup.com/paper/differential-privacy-in-generative-ai-agents-analysis-and-optimal-tradeoffs - A framework for analyzing and optimizing privacy in AI agents using differential privacy techniques.
- A Noise Sensitivity Exponent Controls Large Statistical-to-Computational Gaps in Single- and Multi-Index Models (viability: 3): https://sciencetostartup.com/paper/a-noise-sensitivity-exponent-controls-large-statistical-to-computational-gaps-in-single-and-multi-index-models - This research identifies a Noise Sensitivity Exponent that links noise robustness and computational hardness in high-dimensional learning models.
- A Creative Agent is Worth a 64-Token Template (viability: 7): https://sciencetostartup.com/paper/a-creative-agent-is-worth-a-64-token-template - CAT is a framework that enhances creativity in text-to-image models by generating reusable token templates for fuzzy prompts.
- scicode-lint: Detecting Methodology Bugs in Scientific Python Code with LLM-Generated Patterns (viability: 7): https://sciencetostartup.com/paper/scicode-lint-detecting-methodology-bugs-in-scientific-python-code-with-llm-generated-patterns - Scicode-lint automates the detection of methodology bugs in scientific Python code using AI-generated patterns.
- RAMP: Reinforcement Adaptive Mixed Precision Quantization for Efficient On Device LLM Inference (viability: 6): https://sciencetostartup.com/paper/ramp-reinforcement-adaptive-mixed-precision-quantization-for-efficient-on-device-llm-inference - RAMP optimizes large language model inference on resource-constrained devices through adaptive mixed precision quantization.
- Identity as Presence: Towards Appearance and Voice Personalized Joint Audio-Video Generation (viability: 7): https://sciencetostartup.com/paper/identity-as-presence-towards-appearance-and-voice-personalized-joint-audio-video-generation - A scalable framework for personalized audio-video generation that allows fine-grained control over identity attributes.
- AI-Assisted Goal Setting Improves Goal Progress Through Social Accountability (viability: 4): https://sciencetostartup.com/paper/ai-assisted-goal-setting-improves-goal-progress-through-social-accountability - An AI career coach that enhances goal progress through social accountability.
- DebugLM: Learning Traceable Training Data Provenance for LLMs (viability: 7): https://sciencetostartup.com/paper/debuglm-learning-traceable-training-data-provenance-for-llms - DebugLM enhances LLMs with data provenance for improved debugging and targeted remediation.
- SoK: From Silicon to Netlist and Beyond $-$ Two Decades of Hardware Reverse Engineering Research (viability: 3): https://sciencetostartup.com/paper/sok-from-silicon-to-netlist-and-beyond-two-decades-of-hardware-reverse-engineering-research - A comprehensive analysis of hardware reverse engineering research to enhance security practices.
- Differential Attention-Augmented BiomedCLIP with Asymmetric Focal Optimization for Imbalanced Multi-Label Video Capsule Endoscopy Classification (viability: 4): https://sciencetostartup.com/paper/differential-attention-augmented-biomedclip-with-asymmetric-focal-optimization-for-imbalanced-multi-label-video-capsule- - A novel framework for multi-label classification in video capsule endoscopy that tackles class imbalance using advanced attention mechanisms and optimization strategies.
- Edit Spillover as a Probe: Do Image Editing Models Implicitly Understand World Relations? (viability: 6): https://sciencetostartup.com/paper/edit-spillover-as-a-probe-do-image-editing-models-implicitly-understand-world-relations - EditSpilloverProbe leverages edit spillover in image editing models to assess their implicit understanding of world relations.
- Operator-Theoretic Foundations and Policy Gradient Methods for General MDPs with Unbounded Costs (viability: 2): https://sciencetostartup.com/paper/operator-theoretic-foundations-and-policy-gradient-methods-for-general-mdps-with-unbounded-costs - This paper presents a theoretical framework for optimizing Markov decision processes using linear operators.
- Mitigating LLM Hallucinations through Domain-Grounded Tiered Retrieval (viability: 7): https://sciencetostartup.com/paper/mitigating-llm-hallucinations-through-domain-grounded-tiered-retrieval - A domain-grounded retrieval system that enhances the reliability of LLMs by mitigating hallucinations through a structured verification process.
- RHYME-XT: A Neural Operator for Spatiotemporal Control Systems (viability: 4): https://sciencetostartup.com/paper/rhyme-xt-a-neural-operator-for-spatiotemporal-control-systems - RHYME-XT is a neural operator framework for efficient surrogate modeling of spatiotemporal control systems.
- Procedural Generation of Algorithm Discovery Tasks in Machine Learning (viability: 7): https://sciencetostartup.com/paper/procedural-generation-of-algorithm-discovery-tasks-in-machine-learning - DiscoGen automates the creation of diverse algorithm discovery tasks to enhance machine learning optimization.
- VISER: Visually-Informed System for Enhanced Robustness in Open-Set Iris Presentation Attack Detection (viability: 6): https://sciencetostartup.com/paper/viser-visually-informed-system-for-enhanced-robustness-in-open-set-iris-presentation-attack-detection - A novel approach to enhance robustness in iris presentation attack detection using human perceptual priors.
- Physics-Aware Machine Learning for Seismic and Volcanic Signal Interpretation (viability: 3): https://sciencetostartup.com/paper/physics-aware-machine-learning-for-seismic-and-volcanic-signal-interpretation - A machine learning approach to enhance the reliability of seismic and volcanic monitoring through improved signal interpretation.
- DexViTac: Collecting Human Visuo-Tactile-Kinematic Demonstrations for Contact-Rich Dexterous Manipulation (viability: 8): https://sciencetostartup.com/paper/dexvitac-collecting-human-visuo-tactile-kinematic-demonstrations-for-contact-rich-dexterous-manipulation - DexViTac is a portable data collection system that captures high-fidelity visuo-tactile-kinematic demonstrations for improving robot dexterity in contact-rich environments.
- ProbeFlow: Training-Free Adaptive Flow Matching for Vision-Language-Action Models (viability: 8): https://sciencetostartup.com/paper/probeflow-training-free-adaptive-flow-matching-for-vision-language-action-models - ProbeFlow accelerates action decoding in Vision-Language-Action models for responsive robotic control.
- Revisiting foundation models for cell instance segmentation (viability: 7): https://sciencetostartup.com/paper/revisiting-foundation-models-for-cell-instance-segmentation - A new instance segmentation strategy for improving microscopy cell segmentation models.
- Omni-3DEdit: Generalized Versatile 3D Editing in One-Pass (viability: 7): https://sciencetostartup.com/paper/omni-3dedit-generalized-versatile-3d-editing-in-one-pass - Omni-3DEdit is a unified model for efficient and versatile 3D editing tasks in one pass.
- Video Understanding: From Geometry and Semantics to Unified Models (viability: 2): https://sciencetostartup.com/paper/video-understanding-from-geometry-and-semantics-to-unified-models - This survey provides a comprehensive overview of video understanding models and their evolution.
- How do LLMs Compute Verbal Confidence (viability: 5): https://sciencetostartup.com/paper/how-do-llms-compute-verbal-confidence - A study revealing how LLMs compute verbal confidence, enhancing our understanding of model uncertainty.
- Event-Centric Human Value Understanding in News-Domain Texts: An Actor-Conditioned, Multi-Granularity Benchmark (viability: 6): https://sciencetostartup.com/paper/event-centric-human-value-understanding-in-news-domain-texts-an-actor-conditioned-multi-granularity-benchmark - NEVU is a benchmark for actor-conditioned, event-centric human value recognition in news articles, enabling fine-grained evaluation of value cues.
- Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control (viability: 7): https://sciencetostartup.com/paper/generative-control-as-optimization-time-unconditional-flow-matching-for-adaptive-and-robust-robotic-control - A novel adaptive and robust control solution for robotics using generative models to optimize flow matching.
- Text-to-Stage: Spatial Layouts from Long-form Narratives (viability: 5): https://sciencetostartup.com/paper/text-to-stage-spatial-layouts-from-long-form-narratives - Automating stage-play layout generation from long-form narratives using advanced language models.
- RPMS: Enhancing LLM-Based Embodied Planning through Rule-Augmented Memory Synergy (viability: 7): https://sciencetostartup.com/paper/rpms-enhancing-llm-based-embodied-planning-through-rule-augmented-memory-synergy - RPMS enhances LLM-based embodied planning by integrating rule-augmented memory to improve action feasibility and success rates.
- CodeScout: An Effective Recipe for Reinforcement Learning of Code Search Agents (viability: 6): https://sciencetostartup.com/paper/codescout-an-effective-recipe-for-reinforcement-learning-of-code-search-agents - CodeScout enhances developer productivity by using reinforcement learning to optimize code search.
- TINA: Text-Free Inversion Attack for Unlearned Text-to-Image Diffusion Models (viability: 5): https://sciencetostartup.com/paper/tina-text-free-inversion-attack-for-unlearned-text-to-image-diffusion-models - TINA is a novel attack method that uncovers erased concepts in text-to-image diffusion models by leveraging visual-only probes.
- FailureMem: A Failure-Aware Multimodal Framework for Autonomous Software Repair (viability: 7): https://sciencetostartup.com/paper/failuremem-a-failure-aware-multimodal-framework-for-autonomous-software-repair - FailureMem is a multimodal framework that enhances automated program repair by integrating visual reasoning and reusable knowledge from past failures.
- Steering Video Diffusion Transformers with Massive Activations (viability: 2): https://sciencetostartup.com/paper/steering-video-diffusion-transformers-with-massive-activations - This paper explores a method to enhance video generation quality using Massive Activations in video diffusion transformers.
- Symmetry-Reduced Physics-Informed Learning of Tensegrity Dynamics (viability: 4): https://sciencetostartup.com/paper/symmetry-reduced-physics-informed-learning-of-tensegrity-dynamics - SymPINN leverages symmetry in tensegrity structures to enhance the efficiency and accuracy of physics-informed neural networks for dynamic predictions.
- Discovering Decoupled Functional Modules in Large Language Models (viability: 2): https://sciencetostartup.com/paper/discovering-decoupled-functional-modules-in-large-language-models - A framework for discovering functional modules in Large Language Models to enhance interpretability.
- Federated Distributional Reinforcement Learning with Distributional Critic Regularization (viability: 2): https://sciencetostartup.com/paper/federated-distributional-reinforcement-learning-with-distributional-critic-regularization - A novel approach to federated reinforcement learning that enhances safety by preserving distributional information.
- Process Supervision for Chain-of-Thought Reasoning via Monte Carlo Net Information Gain (viability: 6): https://sciencetostartup.com/paper/process-supervision-for-chain-of-thought-reasoning-via-monte-carlo-net-information-gain - A novel method for generating step-level labels to enhance multi-step reasoning in large language models.
- M2P: Improving Visual Foundation Models with Mask-to-Point Weakly-Supervised Learning for Dense Point Tracking (viability: 8): https://sciencetostartup.com/paper/m2p-improving-visual-foundation-models-with-mask-to-point-weakly-supervised-learning-for-dense-point-tracking - M2P enhances visual foundation models for dense point tracking using weakly-supervised learning with video object segmentation.
- ChopGrad: Pixel-Wise Losses for Latent Video Diffusion via Truncated Backpropagation (viability: 3): https://sciencetostartup.com/paper/chopgrad-pixel-wise-losses-for-latent-video-diffusion-via-truncated-backpropagation - ChopGrad introduces a memory-efficient method for fine-tuning video diffusion models using truncated backpropagation.
- Dropout Robustness and Cognitive Profiling of Transformer Models via Stochastic Inference (viability: 5): https://sciencetostartup.com/paper/dropout-robustness-and-cognitive-profiling-of-transformer-models-via-stochastic-inference - This research benchmarks dropout robustness in transformer models, providing insights for selecting models in uncertainty-aware applications.
- Fine-Grained Post-Training Quantization for Large Vision Language Models with Quantization-Aware Integrated Gradients (viability: 8): https://sciencetostartup.com/paper/fine-grained-post-training-quantization-for-large-vision-language-models-with-quantization-aware-integrated-gradients - A fine-grained quantization strategy for large vision language models that enhances accuracy while reducing computational overhead.
- EVA: Aligning Video World Models with Executable Robot Actions via Inverse Dynamics Rewards (viability: 7): https://sciencetostartup.com/paper/eva-aligning-video-world-models-with-executable-robot-actions-via-inverse-dynamics-rewards - EVA aligns video world models with executable robot actions to enhance robotic control through reinforcement learning.
- RangeAD: Fast On-Model Anomaly Detection (viability: 7): https://sciencetostartup.com/paper/rangead-fast-on-model-anomaly-detection - RangeAD offers an efficient on-model approach for anomaly detection that reduces inference costs while improving performance.
- Governed Memory: A Production Architecture for Multi-Agent Workflows (viability: 3): https://sciencetostartup.com/paper/governed-memory-a-production-architecture-for-multi-agent-workflows - A novel architecture to enhance multi-agent workflows with governed memory.
- ResNet-50 with Class Reweighting and Anatomy-Guided Temporal Decoding for Gastrointestinal Video Analysis (viability: 3): https://sciencetostartup.com/paper/resnet-50-with-class-reweighting-and-anatomy-guided-temporal-decoding-for-gastrointestinal-video-analysis - A gastrointestinal video analysis pipeline using ResNet-50 for multi-label classification.
- Exploring parameter-efficient fine-tuning (PEFT) of billion-parameter vision models with QLoRA and DoRA: insights into generalization for limited-data image classification under a 98:1 test-to-train regime (viability: 7): https://sciencetostartup.com/paper/exploring-parameter-efficient-fine-tuning-peft-of-billion-parameter-vision-models-with-qlora-and-dora-insights-into-gene - A parameter-efficient fine-tuning approach for billion-parameter vision models to enhance behavior classification in precision livestock farming.
- Facts as First Class Objects: Knowledge Objects for Persistent LLM Memory (viability: 7): https://sciencetostartup.com/paper/facts-as-first-class-objects-knowledge-objects-for-persistent-llm-memory - Introducing Knowledge Objects for efficient and accurate memory management in large language models.
- CrowdGaussian: Reconstructing High-Fidelity 3D Gaussians for Human Crowd from a Single Image (viability: 7): https://sciencetostartup.com/paper/crowdgaussian-reconstructing-high-fidelity-3d-gaussians-for-human-crowd-from-a-single-image - CrowdGaussian reconstructs high-fidelity 3D models of human crowds from single images using advanced self-supervised learning techniques.
- CoVerRL: Breaking the Consensus Trap in Label-Free Reasoning via Generator-Verifier Co-Evolution (viability: 7): https://sciencetostartup.com/paper/coverrl-breaking-the-consensus-trap-in-label-free-reasoning-via-generator-verifier-co-evolution - CoVerRL enhances label-free reinforcement learning by using a generator-verifier co-evolution framework to improve reasoning accuracy.
- Attention Sinks Induce Gradient Sinks (viability: 2): https://sciencetostartup.com/paper/attention-sinks-induce-gradient-sinks - This paper explores the relationship between attention sinks and gradient sinks in Transformer models during training.
- Modeling Overlapped Speech with Shuffles (viability: 3): https://sciencetostartup.com/paper/modeling-overlapped-speech-with-shuffles - A novel algorithm for single-pass alignment of multi-talker recordings using finite-state automata.
- Huddle: Parallel Shape Assembly using Decentralized, Minimalistic Robots (viability: 4): https://sciencetostartup.com/paper/huddle-parallel-shape-assembly-using-decentralized-minimalistic-robots - A decentralized algorithm for efficient shape assembly using minimalistic robots.
- Facial Movement Dynamics Reveal Workload During Complex Multitasking (viability: 3): https://sciencetostartup.com/paper/facial-movement-dynamics-reveal-workload-during-complex-multitasking - A novel approach to monitor cognitive workload using facial movement dynamics from standard webcams.
- Evidence Packing for Cross-Domain Image Deepfake Detection with LVLMs (viability: 7): https://sciencetostartup.com/paper/evidence-packing-for-cross-domain-image-deepfake-detection-with-lvlms - A training-free framework for detecting deepfakes using evidence-driven reasoning with large vision-language models.
- Harm or Humor: A Multimodal, Multilingual Benchmark for Overt and Covert Harmful Humor (viability: 7): https://sciencetostartup.com/paper/harm-or-humor-a-multimodal-multilingual-benchmark-for-overt-and-covert-harmful-humor - A benchmark for detecting harmful humor across multiple languages and modalities.
- On Securing the Software Development Lifecycle in IoT RISC-V Trusted Execution Environments (viability: 4): https://sciencetostartup.com/paper/on-securing-the-software-development-lifecycle-in-iot-risc-v-trusted-execution-environments - A toolkit for enhancing the software development lifecycle in RISC-V Trusted Execution Environments for IoT and automotive applications.
- PC-CrossDiff: Point-Cluster Dual-Level Cross-Modal Differential Attention for Unified 3D Referring and Segmentation (viability: 7): https://sciencetostartup.com/paper/pc-crossdiff-point-cluster-dual-level-cross-modal-differential-attention-for-unified-3d-referring-and-segmentation - PC-CrossDiff enhances 3D visual grounding by improving localization in complex scenes using dual-level differential attention.
- Multi-Source Human-in-the-Loop Digital Twin Testbed for Connected and Autonomous Vehicles in Mixed Traffic Flow (viability: 7): https://sciencetostartup.com/paper/multi-source-human-in-the-loop-digital-twin-testbed-for-connected-and-autonomous-vehicles-in-mixed-traffic-flow - A testbed for testing Connected and Autonomous Vehicles in mixed traffic environments using a human-in-the-loop approach.
- Towards Infinitely Long Neural Simulations: Self-Refining Neural Surrogate Models for Dynamical Systems (viability: 6): https://sciencetostartup.com/paper/towards-infinitely-long-neural-simulations-self-refining-neural-surrogate-models-for-dynamical-systems - A self-refining neural surrogate model that enhances the simulation of dynamical systems by ensuring long-time consistency.
- Concept-to-Pixel: Prompt-Free Universal Medical Image Segmentation (viability: 5): https://sciencetostartup.com/paper/concept-to-pixel-prompt-free-universal-medical-image-segmentation - Develop a universal, prompt-free AI for medical image segmentation.
- Embedding World Knowledge into Tabular Models: Towards Best Practices for Embedding Pipeline Design (viability: 5): https://sciencetostartup.com/paper/embedding-world-knowledge-into-tabular-models-towards-best-practices-for-embedding-pipeline-design - A systematic approach to designing effective LLM-based embedding pipelines for tabular prediction.
- TAPESTRY: From Geometry to Appearance via Consistent Turntable Videos (viability: 8): https://sciencetostartup.com/paper/tapestry-from-geometry-to-appearance-via-consistent-turntable-videos - TAPESTRY generates high-fidelity turntable videos from 3D models, enabling automated creation of production-ready assets.
- SARE: Sample-wise Adaptive Reasoning for Training-free Fine-grained Visual Recognition (viability: 7): https://sciencetostartup.com/paper/sare-sample-wise-adaptive-reasoning-for-training-free-fine-grained-visual-recognition - SARE is a training-free framework that enhances Fine-Grained Visual Recognition by adapting reasoning based on sample difficulty.
- Data Obfuscation for Secure Use of Classical Values in Quantum Computation (viability: 2): https://sciencetostartup.com/paper/data-obfuscation-for-secure-use-of-classical-values-in-quantum-computation - This paper presents a novel data obfuscation technique for securing classical values in quantum computation.
- Predicting Trajectories of Long COVID in Adult Women: The Critical Role of Causal Disentanglement (viability: 6): https://sciencetostartup.com/paper/predicting-trajectories-of-long-covid-in-adult-women-the-critical-role-of-causal-disentanglement - A predictive framework for assessing long COVID severity in women by integrating clinical and wearable data.
- VolumeDP: Modeling Volumetric Representation for Manipulation Policy Learning (viability: 8): https://sciencetostartup.com/paper/volumedp-modeling-volumetric-representation-for-manipulation-policy-learning - VolumeDP enhances robotic manipulation through advanced 3D spatial reasoning for improved imitation learning.
- DiffVP: Differential Visual Semantic Prompting for LLM-Based CT Report Generation (viability: 6): https://sciencetostartup.com/paper/diffvp-differential-visual-semantic-prompting-for-llm-based-ct-report-generation - Automate CT report generation using AI-enhanced visual semantic prompting for radiologists.
- Machine Learning for Network Attacks Classification and Statistical Evaluation of Machine Learning for Network Attacks Classification and Adversarial Learning Methodologies for Synthetic Data Generation (viability: 4): https://sciencetostartup.com/paper/machine-learning-for-network-attacks-classification-and-statistical-evaluation-of-machine-learning-for-network-attacks-c - A unified multi-modal dataset and machine learning approach for effective network intrusion detection and synthetic data generation.
- Eye image segmentation using visual and concept prompts with Segment Anything Model 3 (SAM3) (viability: 4): https://sciencetostartup.com/paper/eye-image-segmentation-using-visual-and-concept-prompts-with-segment-anything-model-3-sam3 - A comparative study of eye image segmentation using the Segment Anything Model 3 with visual and concept prompts.
- From Virtual Environments to Real-World Trials: Emerging Trends in Autonomous Driving (viability: 3): https://sciencetostartup.com/paper/from-virtual-environments-to-real-world-trials-emerging-trends-in-autonomous-driving - A comprehensive review of synthetic data and simulation technologies for advancing autonomous driving.
- AERR-Nav: Adaptive Exploration-Recovery-Reminiscing Strategy for Zero-Shot Object Navigation (viability: 7): https://sciencetostartup.com/paper/aerr-nav-adaptive-exploration-recovery-reminiscing-strategy-for-zero-shot-object-navigation - AERR-Nav is an adaptive framework for zero-shot object navigation in complex multi-floor environments.
- Parameter-Efficient Modality-Balanced Symmetric Fusion for Multimodal Remote Sensing Semantic Segmentation (viability: 8): https://sciencetostartup.com/paper/parameter-efficient-modality-balanced-symmetric-fusion-for-multimodal-remote-sensing-semantic-segmentation - MoBaNet is a parameter-efficient framework for multimodal remote sensing semantic segmentation that balances modality contributions while minimizing computational overhead.
- DancingBox: A Lightweight MoCap System for Character Animation from Physical Proxies (viability: 7): https://sciencetostartup.com/paper/dancingbox-a-lightweight-mocap-system-for-character-animation-from-physical-proxies - DancingBox is a lightweight motion capture system that enables intuitive character animation using everyday objects and a single webcam.
- MALLES: A Multi-agent LLMs-based Economic Sandbox with Consumer Preference Alignment (viability: 6): https://sciencetostartup.com/paper/malles-a-multi-agent-llms-based-economic-sandbox-with-consumer-preference-alignment - MALLES is a multi-agent economic simulation framework leveraging LLMs for consumer preference alignment.
- Learning Transferable Temporal Primitives for Video Reasoning via Synthetic Videos (viability: 8): https://sciencetostartup.com/paper/learning-transferable-temporal-primitives-for-video-reasoning-via-synthetic-videos - SynRL is a post-training framework that enhances video understanding by teaching models fundamental temporal primitives through synthetic video generation.
- Can Blindfolded LLMs Still Trade? An Anonymization-First Framework for Portfolio Optimization (viability: 7): https://sciencetostartup.com/paper/can-blindfolded-llms-still-trade-an-anonymization-first-framework-for-portfolio-optimization - Anonymization-first framework for LLM trading agents that enhances trustworthiness and performance in portfolio optimization.
- Objective Mispricing Detection for Shortlisting Undervalued Football Players via Market Dynamics and News Signals (viability: 3): https://sciencetostartup.com/paper/objective-mispricing-detection-for-shortlisting-undervalued-football-players-via-market-dynamics-and-news-signals - A framework for identifying undervalued football players using market dynamics and news signals.
- Flow Matching Policy with Entropy Regularization (viability: 7): https://sciencetostartup.com/paper/flow-matching-policy-with-entropy-regularization - FMER is an efficient online RL framework that enhances exploration through principled maximum-entropy optimization.
- Does YOLO Really Need to See Every Training Image in Every Epoch? (viability: 7): https://sciencetostartup.com/paper/does-yolo-really-need-to-see-every-training-image-in-every-epoch - A novel sampling strategy for YOLO detectors that optimizes training efficiency by selectively resampling images based on learning sufficiency.
- Sensi: Learn One Thing at a Time -- Curriculum-Based Test-Time Learning for LLM Game Agents (viability: 7): https://sciencetostartup.com/paper/sensi-learn-one-thing-at-a-time-curriculum-based-test-time-learning-for-llm-game-agents - Sensi is an LLM agent that learns efficiently in unknown environments through structured test-time learning.
- WeatherReasonSeg: A Benchmark for Weather-Aware Reasoning Segmentation in Visual Language Models (viability: 4): https://sciencetostartup.com/paper/weatherreasonseg-a-benchmark-for-weather-aware-reasoning-segmentation-in-visual-language-models - WeatherReasonSeg benchmarks the performance of vision-language models under adverse weather conditions to enhance robust reasoning segmentation.
- Illumination-Aware Contactless Fingerprint Spoof Detection via Paired Flash-Non-Flash Imaging (viability: 4): https://sciencetostartup.com/paper/illumination-aware-contactless-fingerprint-spoof-detection-via-paired-flash-non-flash-imaging - A novel approach to enhance contactless fingerprint spoof detection using paired flash-non-flash imaging.
- Adaptive Guidance for Retrieval-Augmented Masked Diffusion Models (viability: 5): https://sciencetostartup.com/paper/adaptive-guidance-for-retrieval-augmented-masked-diffusion-models - Adaptive Retrieval-Augmented Masked Diffusion enhances QA performance by dynamically calibrating guidance based on retrieved context reliability.
- DeepCORO-CLIP: A Multi-View Foundation Model for Comprehensive Coronary Angiography Video-Text Analysis and External Validation (viability: 5): https://sciencetostartup.com/paper/deepcoro-clip-a-multi-view-foundation-model-for-comprehensive-coronary-angiography-video-text-analysis-and-external-vali - AI system for analyzing coronary angiography videos with video-text context understanding.
- Post-Training Local LLM Agents for Linux Privilege Escalation with Verifiable Rewards (viability: 8): https://sciencetostartup.com/paper/post-training-local-llm-agents-for-linux-privilege-escalation-with-verifiable-rewards - A local LLM agent for Linux privilege escalation that achieves high success rates with verifiable rewards.
- Consistency-Driven Dual LSTM Models for Kinematic Control of a Wearable Soft Robotic Arm (viability: 7): https://sciencetostartup.com/paper/consistency-driven-dual-lstm-models-for-kinematic-control-of-a-wearable-soft-robotic-arm - A dual LSTM framework for enhancing the kinematic control of wearable soft robotic arms.
- Few-Step Diffusion Sampling Through Instance-Aware Discretizations (viability: 7): https://sciencetostartup.com/paper/few-step-diffusion-sampling-through-instance-aware-discretizations - An instance-aware discretization framework that enhances the performance of diffusion models by adapting timestep allocations based on input-dependent priors.
- AgentVLN: Towards Agentic Vision-and-Language Navigation (viability: 5): https://sciencetostartup.com/paper/agentvln-towards-agentic-vision-and-language-navigation - Develop an agentic vision-and-language navigation tool for enhanced autonomous systems.
- FINER: MLLMs Hallucinate under Fine-grained Negative Queries (viability: 8): https://sciencetostartup.com/paper/finer-mllms-hallucinate-under-fine-grained-negative-queries - FINER addresses hallucinations in multimodal large language models through innovative fine-grained negative queries and tuning techniques.
- Interpretable Cross-Domain Few-Shot Learning with Rectified Target-Domain Local Alignment (viability: 6): https://sciencetostartup.com/paper/interpretable-cross-domain-few-shot-learning-with-rectified-target-domain-local-alignment - A method to enhance local vision-language alignment in few-shot learning for better interpretability in medical diagnosis.
- REAL: Robust Extreme Agility via Spatio-Temporal Policy Learning and Physics-Guided Filtering (viability: 3): https://sciencetostartup.com/paper/real-robust-extreme-agility-via-spatio-temporal-policy-learning-and-physics-guided-filtering - REAL is an advanced framework for reliable legged parkour that adapts to sensory corruption.
- VectorWorld: Efficient Streaming World Model via Diffusion Flow on Vector Graphs (viability: 8): https://sciencetostartup.com/paper/vectorworld-efficient-streaming-world-model-via-diffusion-flow-on-vector-graphs - VectorWorld offers real-time, high-fidelity autonomous driving simulation using novel vector graph diffusion flows.
- Anchoring and Rescaling Attention for Semantically Coherent Inbetweening (viability: 7): https://sciencetostartup.com/paper/anchoring-and-rescaling-attention-for-semantically-coherent-inbetweening - A generative inbetweening method that synthesizes realistic intermediate frames with enhanced semantic and temporal guidance.
- Part-Aware Open-Vocabulary 3D Affordance Grounding via Prototypical Semantic and Geometric Alignment (viability: 7): https://sciencetostartup.com/paper/part-aware-open-vocabulary-3d-affordance-grounding-via-prototypical-semantic-and-geometric-alignment - A novel framework for enhancing semantic and geometric representations in 3D affordance grounding.
- Automated Grammar-based Algebraic Multigrid Design With Evolutionary Algorithms (viability: 4): https://sciencetostartup.com/paper/automated-grammar-based-algebraic-multigrid-design-with-evolutionary-algorithms - This paper presents a novel approach to optimize multigrid methods using evolutionary algorithms and genetic programming.
- VeriGrey: Greybox Agent Validation (viability: 7): https://sciencetostartup.com/paper/verigrey-greybox-agent-validation - VeriGrey is a grey-box testing framework that uncovers security risks in LLM agents through dynamic prompt mutation.
- DSS-GAN: Directional State Space GAN with Mamba backbone for Class-Conditional Image Synthesis (viability: 7): https://sciencetostartup.com/paper/dss-gan-directional-state-space-gan-with-mamba-backbone-for-class-conditional-image-synthesis - DSS-GAN introduces a novel conditioning mechanism for enhanced class-conditional image synthesis using a Mamba backbone.
- Benchmarking Reinforcement Learning via Stochastic Converse Optimality: Generating Systems with Known Optimal Policies (viability: 3): https://sciencetostartup.com/paper/benchmarking-reinforcement-learning-via-stochastic-converse-optimality-generating-systems-with-known-optimal-policies - A rigorous benchmarking framework for evaluating Reinforcement Learning algorithms through controlled environments.
- rSDNet: Unified Robust Neural Learning against Label Noise and Adversarial Attacks (viability: 4): https://sciencetostartup.com/paper/rsdnet-unified-robust-neural-learning-against-label-noise-and-adversarial-attacks - rSDNet offers a robust neural learning framework to combat label noise and adversarial attacks in classification tasks.
- A Multi-Agent System for Building-Age Cohort Mapping to Support Urban Energy Planning (viability: 6): https://sciencetostartup.com/paper/a-multi-agent-system-for-building-age-cohort-mapping-to-support-urban-energy-planning - A multi-agent LLM system that fuses data to accurately determine urban building age for energy planning.
- S-VGGT: Structure-Aware Subscene Decomposition for Scalable 3D Foundation Models (viability: 7): https://sciencetostartup.com/paper/s-vggt-structure-aware-subscene-decomposition-for-scalable-3d-foundation-models - Optimize 3D foundation models with S-VGGT's structure-aware subscene decomposition for enhanced scalability and efficiency.
- Do Language Models Encode Semantic Relations? Probing and Sparse Feature Analysis (viability: 2): https://sciencetostartup.com/paper/do-language-models-encode-semantic-relations-probing-and-sparse-feature-analysis - This research investigates how large language models encode semantic relationships through probing and interpretability techniques.
- ARES: Scalable and Practical Gradient Inversion Attack in Federated Learning through Activation Recovery (viability: 7): https://sciencetostartup.com/paper/ares-scalable-and-practical-gradient-inversion-attack-in-federated-learning-through-activation-recovery - ARES is a novel gradient inversion attack that reconstructs training samples in federated learning without architectural modifications.
- Complementary Reinforcement Learning (viability: 4): https://sciencetostartup.com/paper/complementary-reinforcement-learning - Complementary RL enhances sample efficiency in LLM-based agents by optimizing experience extraction alongside policy learning.
- Temporal Narrative Monitoring in Dynamic Information Environments (viability: 4): https://sciencetostartup.com/paper/temporal-narrative-monitoring-in-dynamic-information-environments - A framework for monitoring evolving narratives in crisis information environments using semantic embeddings and clustering.
- VeriAgent: A Tool-Integrated Multi-Agent System with Evolving Memory for PPA-Aware RTL Code Generation (viability: 7): https://sciencetostartup.com/paper/veriagent-a-tool-integrated-multi-agent-system-with-evolving-memory-for-ppa-aware-rtl-code-generation - A multi-agent framework for PPA-aware RTL code generation that integrates EDA tools for continuous optimization.
- AdaMuS: Adaptive Multi-view Sparsity Learning for Dimensionally Unbalanced Data (viability: 4): https://sciencetostartup.com/paper/adamus-adaptive-multi-view-sparsity-learning-for-dimensionally-unbalanced-data - AdaMuS is a framework that addresses dimensional imbalances in multi-view learning through adaptive sparsity and self-supervised learning.
- End-to-end data-driven prediction of urban airflow and pollutant dispersion (viability: 4): https://sciencetostartup.com/paper/end-to-end-data-driven-prediction-of-urban-airflow-and-pollutant-dispersion - A data-driven framework for predicting urban airflow and pollutant dispersion to aid decision-makers in environmental management.
- ReLaGS: Relational Language Gaussian Splatting (viability: 7): https://sciencetostartup.com/paper/relags-relational-language-gaussian-splatting - A novel framework for efficient open-vocabulary 3D perception and reasoning without scene-specific training.
- Trust the Unreliability: Inward Backward Dynamic Unreliability Driven Coreset Selection for Medical Image Classification (viability: 7): https://sciencetostartup.com/paper/trust-the-unreliability-inward-backward-dynamic-unreliability-driven-coreset-selection-for-medical-image-classification - Dynamic Unreliability-Driven Coreset Selection enhances medical image classification by focusing on informative unreliable samples.
- A Contextual Help Browser Extension to Assist Digital Illiterate Internet Users (viability: 7): https://sciencetostartup.com/paper/a-contextual-help-browser-extension-to-assist-digital-illiterate-internet-users - A browser extension that provides contextual help for digital illiterate users by delivering on-demand definitions of technical terms.
- From Isolated Scoring to Collaborative Ranking: A Comparison-Native Framework for LLM-Based Paper Evaluation (viability: 8): https://sciencetostartup.com/paper/from-isolated-scoring-to-collaborative-ranking-a-comparison-native-framework-for-llm-based-paper-evaluation - A novel framework for collaborative ranking of scientific papers using LLMs to enhance evaluation accuracy.
- Edit-As-Act: Goal-Regressive Planning for Open-Vocabulary 3D Indoor Scene Editing (viability: 7): https://sciencetostartup.com/paper/edit-as-act-goal-regressive-planning-for-open-vocabulary-3d-indoor-scene-editing - Edit-As-Act enables precise 3D indoor scene editing from natural language instructions through goal-regressive planning.
- One-Step Sampler for Boltzmann Distributions via Drifting (viability: 2): https://sciencetostartup.com/paper/one-step-sampler-for-boltzmann-distributions-via-drifting - A novel framework for efficient sampling from Boltzmann distributions using a one-step neural generator.
- Identifying Latent Actions and Dynamics from Offline Data via Demonstrator Diversity (viability: 4): https://sciencetostartup.com/paper/identifying-latent-actions-and-dynamics-from-offline-data-via-demonstrator-diversity - A method to recover latent actions and dynamics from offline trajectories using demonstrator diversity.
- LoGSAM: Parameter-Efficient Cross-Modal Grounding for MRI Segmentation (viability: 3): https://sciencetostartup.com/paper/logsam-parameter-efficient-cross-modal-grounding-for-mri-segmentation - LoGSAM is a parameter-efficient framework for MRI tumor segmentation using radiologist dictation as prompts.
- Unsupervised Symbolic Anomaly Detection (viability: 7): https://sciencetostartup.com/paper/unsupervised-symbolic-anomaly-detection - SYRAN offers interpretable unsupervised anomaly detection through symbolic regression, providing human-readable equations for anomaly scoring.
- HeiSD: Hybrid Speculative Decoding for Embodied Vision-Language-Action Models with Kinematic Awareness (viability: 7): https://sciencetostartup.com/paper/heisd-hybrid-speculative-decoding-for-embodied-vision-language-action-models-with-kinematic-awareness - HeiSD accelerates robot control through hybrid speculative decoding for improved inference speeds.
- PanoVGGT: Feed-Forward 3D Reconstruction from Panoramic Imagery (viability: 8): https://sciencetostartup.com/paper/panovggt-feed-forward-3d-reconstruction-from-panoramic-imagery - PanoVGGT is a Transformer framework for accurate 3D reconstruction from panoramic imagery, leveraging a unique dataset and innovative training strategies.
- FoMo X: Modular Explainability Signals for Outlier Detection Foundation Models (viability: 7): https://sciencetostartup.com/paper/fomo-x-modular-explainability-signals-for-outlier-detection-foundation-models - FoMo-X enhances outlier detection models with intrinsic explainability for safer decision-making.
- Gaussian Process Limit Reveals Structural Benefits of Graph Transformers (viability: 2): https://sciencetostartup.com/paper/gaussian-process-limit-reveals-structural-benefits-of-graph-transformers - This paper theoretically analyzes the structural benefits of graph transformers over traditional graph convolutional networks.
- Face anonymization preserving facial expressions and photometric realism (viability: 7): https://sciencetostartup.com/paper/face-anonymization-preserving-facial-expressions-and-photometric-realism - A framework for face anonymization that preserves facial expressions and photometric realism for privacy-sensitive applications.
- KA2L: A Knowledge-Aware Active Learning Framework for LLMs (viability: 8): https://sciencetostartup.com/paper/ka2l-a-knowledge-aware-active-learning-framework-for-llms - KA2L is a framework that enhances LLMs' performance through targeted active learning by identifying knowledge gaps.
- In Trust We Survive: Emergent Trust Learning (viability: 4): https://sciencetostartup.com/paper/in-trust-we-survive-emergent-trust-learning - Emergent Trust Learning enhances AI agents' cooperation in competitive environments through a lightweight trust-based control algorithm.
- Zipper-LoRA: Dynamic Parameter Decoupling for Speech-LLM based Multilingual Speech Recognition (viability: 9): https://sciencetostartup.com/paper/zipper-lora-dynamic-parameter-decoupling-for-speech-llm-based-multilingual-speech-recognition - Zipper-LoRA enhances multilingual speech recognition by dynamically optimizing language-specific and shared model parameters.
- FrescoDiffusion: 4K Image-to-Video with Prior-Regularized Tiled Diffusion (viability: 7): https://sciencetostartup.com/paper/frescodiffusion-4k-image-to-video-with-prior-regularized-tiled-diffusion - FrescoDiffusion enables high-quality, coherent video generation from complex images using a novel tiled diffusion approach.
- Prompt-Free Universal Region Proposal Network (viability: 8): https://sciencetostartup.com/paper/prompt-free-universal-region-proposal-network - A novel object detection network that identifies potential objects without relying on external prompts, enhancing flexibility across various applications.
- Consistency of the $k$-Nearest Neighbor Regressor under Complex Survey Designs (viability: 2): https://sciencetostartup.com/paper/consistency-of-the-k-nearest-neighbor-regressor-under-complex-survey-designs - This paper explores the consistency of the k-nearest neighbor regressor in complex survey designs, addressing a gap in existing literature.
- Conditional Inverse Learning of Time-Varying Reproduction Numbers Inference (viability: 4): https://sciencetostartup.com/paper/conditional-inverse-learning-of-time-varying-reproduction-numbers-inference - A framework for estimating time-varying reproduction numbers in infectious disease surveillance using conditional inverse learning.
- CLeAN: Continual Learning Adaptive Normalization in Dynamic Environments (viability: 6): https://sciencetostartup.com/paper/clean-continual-learning-adaptive-normalization-in-dynamic-environments - CLeAN is an adaptive normalization technique that enhances continual learning in dynamic environments by mitigating catastrophic forgetting.
- ProGVC: Progressive-based Generative Video Compression via Auto-Regressive Context Modeling (viability: 7): https://sciencetostartup.com/paper/progvc-progressive-based-generative-video-compression-via-auto-regressive-context-modeling - ProGVC is a novel video compression framework that combines progressive transmission and efficient entropy coding for enhanced perceptual quality at low bitrates.
- Per-Domain Generalizing Policies: On Learning Efficient and Robust Q-Value Functions (Extended Version with Technical Appendix) (viability: 2): https://sciencetostartup.com/paper/per-domain-generalizing-policies-on-learning-efficient-and-robust-q-value-functions-extended-version-with-technical-appe - This paper explores a novel approach to learning Q-value functions for efficient planning in reinforcement learning.
- AURORA Model of Formant-to-Tongue Inversion for Didactic and Clinical Applications (viability: 5): https://sciencetostartup.com/paper/aurora-model-of-formant-to-tongue-inversion-for-didactic-and-clinical-applications - AURORA is a model that predicts tongue displacement in vowel sounds, serving as a didactic tool and biofeedback application.
- Temporal Gains, Spatial Costs: Revisiting Video Fine-Tuning in Multimodal Large Language Models (viability: 4): https://sciencetostartup.com/paper/temporal-gains-spatial-costs-revisiting-video-fine-tuning-in-multimodal-large-language-models - This research explores the impact of video-based fine-tuning on multimodal large language models, revealing trade-offs in visual performance.
- Deploying Semantic ID-based Generative Retrieval for Large-Scale Podcast Discovery at Spotify (viability: 8): https://sciencetostartup.com/paper/deploying-semantic-id-based-generative-retrieval-for-large-scale-podcast-discovery-at-spotify - GLIDE is a generative recommender system that enhances podcast discovery by combining semantic reasoning with user context at Spotify.
- Learning Coordinate-based Convolutional Kernels for Continuous SE(3) Equivariant and Efficient Point Cloud Analysis (viability: 3): https://sciencetostartup.com/paper/learning-coordinate-based-convolutional-kernels-for-continuous-se-3-equivariant-and-efficient-point-cloud-analysis - ECKConv introduces a novel kernel architecture for efficient SE(3) equivariant learning in point cloud tasks.
- CA-Based Interpretable Knowledge Representation and Analysis of Geometric Design Parameters (viability: 2): https://sciencetostartup.com/paper/ca-based-interpretable-knowledge-representation-and-analysis-of-geometric-design-parameters - This paper explores the limitations of PCA in estimating design parameters for CAD applications.
- Informative Semi-Factuals for XAI: The Elaborated Explanations that People Prefer (viability: 5): https://sciencetostartup.com/paper/informative-semi-factuals-for-xai-the-elaborated-explanations-that-people-prefer - A novel algorithm for generating informative semi-factual explanations in XAI that enhances user understanding of automated decisions.
- A Unified Language Model for Large Scale Search, Recommendation, and Reasoning (viability: 8): https://sciencetostartup.com/paper/a-unified-language-model-for-large-scale-search-recommendation-and-reasoning - NEO is a unified language model that enhances recommendation, search, and reasoning across large catalogs with language-steerable capabilities.
- Anisotropic Permeability Tensor Prediction from Porous Media Microstructure via Physics-Informed Progressive Transfer Learning with Hybrid CNN-Transformer (viability: 3): https://sciencetostartup.com/paper/anisotropic-permeability-tensor-prediction-from-porous-media-microstructure-via-physics-informed-progressive-transfer-le - A physics-informed deep learning framework for predicting permeability tensors from porous media microstructure images.
- Rel-Zero: Harnessing Patch-Pair Invariance for Robust Zero-Watermarking Against AI Editing (viability: 7): https://sciencetostartup.com/paper/rel-zero-harnessing-patch-pair-invariance-for-robust-zero-watermarking-against-ai-editing - Rel-Zero offers a robust zero-watermarking solution that maintains visual fidelity against AI editing.
- AdapTS: Lightweight Teacher-Student Approach for Multi-Class and Continual Visual Anomaly Detection (viability: 7): https://sciencetostartup.com/paper/adapts-lightweight-teacher-student-approach-for-multi-class-and-continual-visual-anomaly-detection - AdapTS is a lightweight Teacher-Student framework for efficient multi-class and continual visual anomaly detection optimized for edge deployment.
- AirDDE: Multifactor Neural Delay Differential Equations for Air Quality Forecasting (viability: 8): https://sciencetostartup.com/paper/airdde-multifactor-neural-delay-differential-equations-for-air-quality-forecasting - AirDDE leverages neural delay differential equations for improved air quality forecasting by integrating delay modeling with physical guidance.
- MM-OVSeg:Multimodal Optical-SAR Fusion for Open-Vocabulary Segmentation in Remote Sensing (viability: 7): https://sciencetostartup.com/paper/mm-ovseg-multimodal-optical-sar-fusion-for-open-vocabulary-segmentation-in-remote-sensing - MM-OVSeg is a multimodal framework for resilient open-vocabulary segmentation in remote sensing, leveraging optical and SAR data.
- Mirror Descent on Riemannian Manifolds (viability: 2): https://sciencetostartup.com/paper/mirror-descent-on-riemannian-manifolds - This paper presents a generalized Mirror Descent method for optimization on Riemannian manifolds.
- KineVLA: Towards Kinematics-Aware Vision-Language-Action Models with Bi-Level Action Decomposition (viability: 6): https://sciencetostartup.com/paper/kinevla-towards-kinematics-aware-vision-language-action-models-with-bi-level-action-decomposition - KineVLA enhances robotic manipulation through a novel vision-language-action framework that integrates detailed kinematic attributes.
- Translation Invariance of Neural Operators for the FitzHugh-Nagumo Model (viability: 4): https://sciencetostartup.com/paper/translation-invariance-of-neural-operators-for-the-fitzhugh-nagumo-model - This study benchmarks Neural Operators for modeling excitable cell dynamics in the FitzHugh-Nagumo model.
- Detecting the Machine: A Comprehensive Benchmark of AI-Generated Text Detectors Across Architectures, Domains, and Adversarial Conditions (viability: 5): https://sciencetostartup.com/paper/detecting-the-machine-a-comprehensive-benchmark-of-ai-generated-text-detectors-across-architectures-domains-and-adversar - A comprehensive benchmark for evaluating AI-generated text detectors across various models and conditions.
- PCA-Seg: Revisiting Cost Aggregation for Open-Vocabulary Semantic and Part Segmentation (viability: 7): https://sciencetostartup.com/paper/pca-seg-revisiting-cost-aggregation-for-open-vocabulary-semantic-and-part-segmentation - PCA-Seg enhances open-vocabulary semantic and part segmentation by improving cost aggregation through a parallel structure.
- UniSem: Generalizable Semantic 3D Reconstruction from Sparse Unposed Images (viability: 2): https://sciencetostartup.com/paper/unisem-generalizable-semantic-3d-reconstruction-from-sparse-unposed-images - UniSem enhances 3D reconstruction accuracy and semantic generalization from sparse images using innovative error-aware techniques.
- EI: Early Intervention for Multimodal Imaging based Disease Recognition (viability: 6): https://sciencetostartup.com/paper/ei-early-intervention-for-multimodal-imaging-based-disease-recognition - A novel framework for multimodal medical imaging that enhances disease recognition by leveraging early intervention techniques.
- Proof-of-Authorship for Diffusion-based AI Generated Content (viability: 2): https://sciencetostartup.com/paper/proof-of-authorship-for-diffusion-based-ai-generated-content - A framework for proving authorship of AI-generated content using cryptographic methods.
- Language on Demand, Knowledge at Core: Composing LLMs with Encoder-Decoder Translation Models for Extensible Multilinguality (viability: 7): https://sciencetostartup.com/paper/language-on-demand-knowledge-at-core-composing-llms-with-encoder-decoder-translation-models-for-extensible-multilinguali - XBridge enhances multilingual capabilities of LLMs by integrating pretrained encoder-decoder translation models for improved performance on low-resource languages.
- Interpreting Context-Aware Human Preferences for Multi-Objective Robot Navigation (viability: 7): https://sciencetostartup.com/paper/interpreting-context-aware-human-preferences-for-multi-objective-robot-navigation - A pipeline that enables robots to adapt their navigation based on context-aware human preferences using advanced language and reinforcement learning models.
- Omni-I2C: A Holistic Benchmark for High-Fidelity Image-to-Code Generation (viability: 8): https://sciencetostartup.com/paper/omni-i2c-a-holistic-benchmark-for-high-fidelity-image-to-code-generation - Omni-I2C is a benchmark for evaluating Large Multimodal Models in generating executable code from complex digital graphics.
- QuantFL: Sustainable Federated Learning for Edge IoT via Pre-Trained Model Quantisation (viability: 7): https://sciencetostartup.com/paper/quantfl-sustainable-federated-learning-for-edge-iot-via-pre-trained-model-quantisation - QuantFL is a sustainable federated learning framework that reduces energy costs for IoT devices through efficient model quantization.
- Inducing Epistemological Humility in Large Language Models: A Targeted SFT Approach to Reducing Hallucination (viability: 8): https://sciencetostartup.com/paper/inducing-epistemological-humility-in-large-language-models-a-targeted-sft-approach-to-reducing-hallucination - A targeted fine-tuning approach to reduce hallucinations in large language models by teaching epistemological humility.
- From Optimizable to Interactable: Mixed Digital Twin-Empowered Testing of Vehicle-Infrastructure Cooperation Systems (viability: 7): https://sciencetostartup.com/paper/from-optimizable-to-interactable-mixed-digital-twin-empowered-testing-of-vehicle-infrastructure-cooperation-systems - IMPACT is an interactive testing framework for vehicle-infrastructure cooperation systems that enhances corner-case generation through human interaction.
- UAV-CB: A Complex-Background RGB-T Dataset and Local Frequency Bridge Network for UAV Detection (viability: 7): https://sciencetostartup.com/paper/uav-cb-a-complex-background-rgb-t-dataset-and-local-frequency-bridge-network-for-uav-detection - A novel dataset and network for robust UAV detection in complex backgrounds.
- Learning When to Attend: Conditional Memory Access for Long-Context LLMs (viability: 3): https://sciencetostartup.com/paper/learning-when-to-attend-conditional-memory-access-for-long-context-llms - L2A introduces a conditional memory access layer for long-context language models, optimizing attention usage and improving training efficiency.
- Auto-Unrolled Proximal Gradient Descent: An AutoML Approach to Interpretable Waveform Optimization (viability: 3): https://sciencetostartup.com/paper/auto-unrolled-proximal-gradient-descent-an-automl-approach-to-interpretable-waveform-optimization - An AutoML approach that optimizes wireless beamforming and waveforms using interpretable deep unfolding techniques.
- UniSAFE: A Comprehensive Benchmark for Safety Evaluation of Unified Multimodal Models (viability: 7): https://sciencetostartup.com/paper/unisafe-a-comprehensive-benchmark-for-safety-evaluation-of-unified-multimodal-models - UniSAFE provides a comprehensive benchmark for evaluating the safety of unified multimodal models across various tasks.
- Humans and transformer LMs: Abstraction drives language learning (viability: 2): https://sciencetostartup.com/paper/humans-and-transformer-lms-abstraction-drives-language-learning - This paper explores how transformer LMs learn linguistic categories, shedding light on language acquisition processes.
- Revisiting Cross-Attention Mechanisms: Leveraging Beneficial Noise for Domain-Adaptive Learning (viability: 7): https://sciencetostartup.com/paper/revisiting-cross-attention-mechanisms-leveraging-beneficial-noise-for-domain-adaptive-learning - A framework that enhances domain-adaptive learning through beneficial noise in cross-attention mechanisms.
- Bringing Network Coding into Multi-Robot Systems: Interplay Study for Autonomous Systems over Wireless Communications (viability: 4): https://sciencetostartup.com/paper/bringing-network-coding-into-multi-robot-systems-interplay-study-for-autonomous-systems-over-wireless-communications - This research proposes adaptive network coding to enhance communication reliability in multi-robot systems operating over wireless channels.
- VirPro: Visual-referred Probabilistic Prompt Learning for Weakly-Supervised Monocular 3D Detection (viability: 7): https://sciencetostartup.com/paper/virpro-visual-referred-probabilistic-prompt-learning-for-weakly-supervised-monocular-3d-detection - VirPro enhances weakly-supervised monocular 3D detection by integrating visual-referred prompts for improved scene-aware representations.
- Efficient Soft Actor-Critic with LLM-Based Action-Level Guidance for Continuous Control (viability: 7): https://sciencetostartup.com/paper/efficient-soft-actor-critic-with-llm-based-action-level-guidance-for-continuous-control - GuidedSAC enhances reinforcement learning efficiency by integrating LLMs for action-level guidance in continuous control tasks.
- AR-CoPO: Align Autoregressive Video Generation with Contrastive Policy Optimization (viability: 7): https://sciencetostartup.com/paper/ar-copo-align-autoregressive-video-generation-with-contrastive-policy-optimization - AR-CoPO enhances autoregressive video generation by aligning it with contrastive policy optimization for improved quality and generalization.
- P$^{3}$Nav: End-to-End Perception, Prediction and Planning for Vision-and-Language Navigation (viability: 7): https://sciencetostartup.com/paper/p-3-nav-end-to-end-perception-prediction-and-planning-for-vision-and-language-navigation - P$^{3}$Nav is an end-to-end framework that enhances Vision-and-Language Navigation by integrating perception, prediction, and planning.
- FACE-net: Factual Calibration and Emotion Augmentation for Retrieval-enhanced Emotional Video Captioning (viability: 2): https://sciencetostartup.com/paper/face-net-factual-calibration-and-emotion-augmentation-for-retrieval-enhanced-emotional-video-captioning - FACE-net enhances emotional video captioning by calibrating factual content and augmenting emotional cues.
- DDH-based schemes for multi-party Function Secret Sharing (viability: 4): https://sciencetostartup.com/paper/ddh-based-schemes-for-multi-party-function-secret-sharing - A novel DDH-based technique for reducing key sizes in multi-party Function Secret Sharing schemes.
- VLM2Rec: Resolving Modality Collapse in Vision-Language Model Embedders for Multimodal Sequential Recommendation (viability: 7): https://sciencetostartup.com/paper/vlm2rec-resolving-modality-collapse-in-vision-language-model-embedders-for-multimodal-sequential-recommendation - VLM2Rec enhances multimodal sequential recommendation by balancing modality utilization in Vision-Language Models.
- TRiMS: Real-Time Tracking of Minimal Sufficient Length for Efficient Reasoning via RL (viability: 5): https://sciencetostartup.com/paper/trims-real-time-tracking-of-minimal-sufficient-length-for-efficient-reasoning-via-rl - TRiMS optimizes reasoning efficiency in language models by minimizing token usage while maintaining accuracy.
- When Only the Final Text Survives: Implicit Execution Tracing for Multi-Agent Attribution (viability: 2): https://sciencetostartup.com/paper/when-only-the-final-text-survives-implicit-execution-tracing-for-multi-agent-attribution - A framework for token-level attribution in multi-agent language systems to enhance accountability.
- AdaZoom-GUI: Adaptive Zoom-based GUI Grounding with Instruction Refinement (viability: 7): https://sciencetostartup.com/paper/adazoom-gui-adaptive-zoom-based-gui-grounding-with-instruction-refinement - AdaZoom-GUI enhances GUI grounding accuracy through adaptive zoom and instruction refinement for vision-language models.
- Baguan-TS: A Sequence-Native In-Context Learning Model for Time Series Forecasting with Covariates (viability: 7): https://sciencetostartup.com/paper/baguan-ts-a-sequence-native-in-context-learning-model-for-time-series-forecasting-with-covariates - Baguan-TS is a novel Transformer model that enhances time series forecasting through in-context learning and raw-sequence representation.
- FloorPlan-VLN: A New Paradigm for Floor Plan Guided Vision-Language Navigation (viability: 7): https://sciencetostartup.com/paper/floorplan-vln-a-new-paradigm-for-floor-plan-guided-vision-language-navigation - FloorPlan-VLN revolutionizes navigation by integrating concise instructions with structured floor plans for enhanced spatial reasoning.
- TimeAPN: Adaptive Amplitude-Phase Non-Stationarity Normalization for Time Series Forecasting (viability: 7): https://sciencetostartup.com/paper/timeapn-adaptive-amplitude-phase-non-stationarity-normalization-for-time-series-forecasting - TimeAPN is a novel framework that enhances time series forecasting by addressing non-stationarity through adaptive amplitude-phase normalization.
- ZipServ: Fast and Memory-Efficient LLM Inference with Hardware-Aware Lossless Compression (viability: 3): https://sciencetostartup.com/paper/zipserv-fast-and-memory-efficient-llm-inference-with-hardware-aware-lossless-compression - ZipServ optimizes LLM inference through innovative lossless compression techniques tailored for GPU architectures.
- The Phasor Transformer: Resolving Attention Bottlenecks on the Unit Circle (viability: 6): https://sciencetostartup.com/paper/the-phasor-transformer-resolving-attention-bottlenecks-on-the-unit-circle - The Phasor Transformer introduces a novel approach to time-series forecasting by leveraging phase-native representations for efficient token mixing.
- Argument Reconstruction as Supervision for Critical Thinking in LLMs (viability: 8): https://sciencetostartup.com/paper/argument-reconstruction-as-supervision-for-critical-thinking-in-llms - A framework that enhances LLMs' critical thinking by teaching them to reconstruct arguments.
- SafeLand: Safe Autonomous Landing in Unknown Environments with Bayesian Semantic Mapping (viability: 9): https://sciencetostartup.com/paper/safeland-safe-autonomous-landing-in-unknown-environments-with-bayesian-semantic-mapping - SafeLand is a vision-based system for safe autonomous landing of UAVs in dynamic environments without prior information.
- ECHO: Towards Emotionally Appropriate and Contextually Aware Interactive Head Generation (viability: 7): https://sciencetostartup.com/paper/echo-towards-emotionally-appropriate-and-contextually-aware-interactive-head-generation - ECHO synthesizes lifelike avatar head videos with emotionally appropriate and contextually aware facial behaviors.
- SHIFT: Motion Alignment in Video Diffusion Models with Adversarial Hybrid Fine-Tuning (viability: 2): https://sciencetostartup.com/paper/shift-motion-alignment-in-video-diffusion-models-with-adversarial-hybrid-fine-tuning - SHIFT introduces a novel fine-tuning framework to enhance motion fidelity in video diffusion models.
- Proactive Knowledge Inquiry in Doctor-Patient Dialogue: Stateful Extraction, Belief Updating, and Path-Aware Action Planning (viability: 3): https://sciencetostartup.com/paper/proactive-knowledge-inquiry-in-doctor-patient-dialogue-stateful-extraction-belief-updating-and-path-aware-action-plannin - A framework for proactive knowledge inquiry in doctor-patient dialogues to enhance EMR generation.
- From Digital Twins to World Models:Opportunities, Challenges, and Applications for Mobile Edge General Intelligence (viability: 2): https://sciencetostartup.com/paper/from-digital-twins-to-world-models-opportunities-challenges-and-applications-for-mobile-edge-general-intelligence - This paper surveys the transition from digital twins to world models for enhancing edge general intelligence in 6G networks.
- Caging the Agents: A Zero Trust Security Architecture for Autonomous AI in Healthcare (viability: 8): https://sciencetostartup.com/paper/caging-the-agents-a-zero-trust-security-architecture-for-autonomous-ai-in-healthcare - Zero Trust Security Architecture for AI agents in healthcare, protecting sensitive data from vulnerabilities.
- Physics-informed Deep Mixture-of-Koopmans Vehicle Dynamics Model with Dual-branch Encoder for Distributed Electric-drive Trucks (viability: 5): https://sciencetostartup.com/paper/physics-informed-deep-mixture-of-koopmans-vehicle-dynamics-model-with-dual-branch-encoder-for-distributed-electric-drive - A data-driven vehicle dynamics modeling method for distributed electric-drive trucks using Koopman operator theory.
- Towards Motion-aware Referring Image Segmentation (viability: 8): https://sciencetostartup.com/paper/towards-motion-aware-referring-image-segmentation - A novel approach to improve referring image segmentation for motion-related queries using multimodal learning.
- Mutually Causal Semantic Distillation Network for Zero-Shot Learning (viability: 4): https://sciencetostartup.com/paper/mutually-causal-semantic-distillation-network-for-zero-shot-learning - MSDN++ enhances zero-shot learning by distilling intrinsic semantic representations through mutually causal attention.
- Joint Degradation-Aware Arbitrary-Scale Super-Resolution for Variable-Rate Extreme Image Compression (viability: 7): https://sciencetostartup.com/paper/joint-degradation-aware-arbitrary-scale-super-resolution-for-variable-rate-extreme-image-compression - ASSR-EIC is a novel image compression framework that enables variable-rate extreme image compression with high fidelity and realism.
- Causal Representation Learning on High-Dimensional Data: Benchmarks, Reproducibility, and Evaluation Metrics (viability: 4): https://sciencetostartup.com/paper/causal-representation-learning-on-high-dimensional-data-benchmarks-reproducibility-and-evaluation-metrics - A framework for improving causal representation learning through enhanced dataset evaluation and reproducibility.
- Large-Scale 3D Ground-Motion Synthesis with Physics-Inspired Latent Operator Flow Matching (viability: 3): https://sciencetostartup.com/paper/large-scale-3d-ground-motion-synthesis-with-physics-inspired-latent-operator-flow-matching - GMFlow is a physics-inspired framework for rapid generation of realistic ground-motion time histories for earthquake hazard analysis.
- Bootstrapping Coding Agents: The Specification Is the Program (viability: 2): https://sciencetostartup.com/paper/bootstrapping-coding-agents-the-specification-is-the-program - A theoretical exploration of self-bootstrapping coding agents based on specifications.
- Motion-Adaptive Temporal Attention for Lightweight Video Generation with Stable Diffusion (viability: 7): https://sciencetostartup.com/paper/motion-adaptive-temporal-attention-for-lightweight-video-generation-with-stable-diffusion - A lightweight video generation system that adapts temporal attention based on motion content using Stable Diffusion.
- Gesture-Aware Pretraining and Token Fusion for 3D Hand Pose Estimation (viability: 7): https://sciencetostartup.com/paper/gesture-aware-pretraining-and-token-fusion-for-3d-hand-pose-estimation - A two-stage framework for accurate 3D hand pose estimation using gesture-aware pretraining and token fusion.
- Harnessing the Power of Foundation Models for Accurate Material Classification (viability: 8): https://sciencetostartup.com/paper/harnessing-the-power-of-foundation-models-for-accurate-material-classification - A novel framework leveraging foundation models to enhance material classification accuracy through innovative dataset generation and prior incorporation.
- Toward Phonology-Guided Sign Language Motion Generation: A Diffusion Baseline and Conditioning Analysis (viability: 3): https://sciencetostartup.com/paper/toward-phonology-guided-sign-language-motion-generation-a-diffusion-baseline-and-conditioning-analysis - This paper explores a generative model for 3D avatar sign language motion using phonological attributes.
- CRE-T1 Preview Technical Report: Beyond Contrastive Learning for Reasoning-Intensive Retrieval (viability: 6): https://sciencetostartup.com/paper/cre-t1-preview-technical-report-beyond-contrastive-learning-for-reasoning-intensive-retrieval - T1 is a generative retrieval model that enhances reasoning-intensive retrieval by dynamically generating intermediate reasoning trajectories.
- PJB: A Reasoning-Aware Benchmark for Person-Job Retrieval (viability: 5): https://sciencetostartup.com/paper/pjb-a-reasoning-aware-benchmark-for-person-job-retrieval - PJB is a reasoning-aware benchmark designed to enhance person-job retrieval systems by diagnosing performance failures.
- The Causal Uncertainty Principle: Manifold Tearing and the Topological Limits of Counterfactual Interventions (viability: 2): https://sciencetostartup.com/paper/the-causal-uncertainty-principle-manifold-tearing-and-the-topological-limits-of-counterfactual-interventions - A theoretical exploration of the limits of causal interventions in generative models.
- Cohomological Obstructions to Global Counterfactuals: A Sheaf-Theoretic Foundation for Generative Causal Models (viability: 3): https://sciencetostartup.com/paper/cohomological-obstructions-to-global-counterfactuals-a-sheaf-theoretic-foundation-for-generative-causal-models - A novel framework for generative causal models that addresses global counterfactuals using sheaf theory.
- VisionNVS: Self-Supervised Inpainting for Novel View Synthesis under the Virtual-Shift Paradigm (viability: 7): https://sciencetostartup.com/paper/visionnvs-self-supervised-inpainting-for-novel-view-synthesis-under-the-virtual-shift-paradigm - VisionNVS revolutionizes Novel View Synthesis for autonomous driving by transforming it into a self-supervised inpainting task.
- SCALE:Scalable Conditional Atlas-Level Endpoint transport for virtual cell perturbation prediction (viability: 7): https://sciencetostartup.com/paper/scale-scalable-conditional-atlas-level-endpoint-transport-for-virtual-cell-perturbation-prediction - SCALE is a specialized model for predicting virtual cell responses to perturbations, enhancing both speed and biological accuracy.
- Efficient Exploration at Scale (viability: 7): https://sciencetostartup.com/paper/efficient-exploration-at-scale - An online learning algorithm that enhances data efficiency in reinforcement learning from human feedback by significantly reducing the number of required labels.
- Stereo World Model: Camera-Guided Stereo Video Generation (viability: 7): https://sciencetostartup.com/paper/stereo-world-model-camera-guided-stereo-video-generation - StereoWorld is a camera-conditioned stereo world model for efficient stereo video generation.
- Shot-Aware Frame Sampling for Video Understanding (viability: 7): https://sciencetostartup.com/paper/shot-aware-frame-sampling-for-video-understanding - InfoShot is a shot-aware frame sampler that enhances long-video understanding by intelligently selecting keyframes to retain critical information.
- SafeTutors: Benchmarking Pedagogical Safety in AI Tutoring Systems (viability: 4): https://sciencetostartup.com/paper/safetutors-benchmarking-pedagogical-safety-in-ai-tutoring-systems - SafeTutors benchmarks the pedagogical safety of AI tutoring systems to enhance learning outcomes.
- Understanding and Defending VLM Jailbreaks via Jailbreak-Related Representation Shift (viability: 3): https://sciencetostartup.com/paper/understanding-and-defending-vlm-jailbreaks-via-jailbreak-related-representation-shift - This paper analyzes jailbreak behavior in vision-language models and proposes a defense method to enhance safety.
- Material Magic Wand: Material-Aware Grouping of 3D Parts in Untextured Meshes (viability: 7): https://sciencetostartup.com/paper/material-magic-wand-material-aware-grouping-of-3d-parts-in-untextured-meshes - Material Magic Wand simplifies the tedious process of material assignment in 3D modeling by automating part grouping based on material properties.
- Towards Safer Large Reasoning Models by Promoting Safety Decision-Making before Chain-of-Thought Generation (viability: 7): https://sciencetostartup.com/paper/towards-safer-large-reasoning-models-by-promoting-safety-decision-making-before-chain-of-thought-generation - A novel safety alignment method for large reasoning models that enhances safety decision-making before chain-of-thought generation.
- Variational Kernel Design for Internal Noise: Gaussian Chaos Noise, Representation Compatibility, and Reliable Deep Learning (viability: 2): https://sciencetostartup.com/paper/variational-kernel-design-for-internal-noise-gaussian-chaos-noise-representation-compatibility-and-reliable-deep-learnin - A theoretical framework for understanding and designing internal noise mechanisms in deep learning.
- Public Profile Matters: A Scalable Integrated Approach to Recommend Citations in the Wild (viability: 7): https://sciencetostartup.com/paper/public-profile-matters-a-scalable-integrated-approach-to-recommend-citations-in-the-wild - Profiler enhances citation recommendation systems by efficiently capturing human citation patterns without bias.
- MCoT-MVS: Multi-level Vision Selection by Multi-modal Chain-of-Thought Reasoning for Composed Image Retrieval (viability: 9): https://sciencetostartup.com/paper/mcot-mvs-multi-level-vision-selection-by-multi-modal-chain-of-thought-reasoning-for-composed-image-retrieval - MCoT-MVS enhances composed image retrieval by integrating multi-level vision features with multi-modal reasoning for improved semantic accuracy.
- A 3D Reconstruction Benchmark for Asset Inspection (viability: 6): https://sciencetostartup.com/paper/a-3d-reconstruction-benchmark-for-asset-inspection - A new dataset for benchmarking 3D reconstruction methods in asset inspection, addressing critical gaps in existing datasets.
- WebPII: Benchmarking Visual PII Detection for Computer-Use Agents (viability: 7): https://sciencetostartup.com/paper/webpii-benchmarking-visual-pii-detection-for-computer-use-agents - WebPII provides a benchmark and model for detecting personally identifiable information in web screenshots to enhance privacy in computer use agents.
- PACE-RAG: Patient-Aware Contextual and Evidence-based Policy RAG for Clinical Drug Recommendation (viability: 9): https://sciencetostartup.com/paper/pace-rag-patient-aware-contextual-and-evidence-based-policy-rag-for-clinical-drug-recommendation - PACE-RAG is a personalized drug recommendation system that integrates patient context with clinical prescribing patterns for optimal treatment decisions.
- OnlineHMR: Video-based Online World-Grounded Human Mesh Recovery (viability: 8): https://sciencetostartup.com/paper/onlinehmr-video-based-online-world-grounded-human-mesh-recovery - OnlineHMR enables real-time 3D human mesh recovery from monocular videos for interactive applications like AR/VR.
- Beyond Outliers: A Data-Free Layer-wise Mixed-Precision Quantization Approach Driven by Numerical and Structural Dual-Sensitivity (viability: 5): https://sciencetostartup.com/paper/beyond-outliers-a-data-free-layer-wise-mixed-precision-quantization-approach-driven-by-numerical-and-structural-dual-sen - A novel calibration-free framework for layer-wise mixed-precision quantization that improves model compression without calibration data.
- Learning Permutation Distributions via Reflected Diffusion on Ranks (viability: 4): https://sciencetostartup.com/paper/learning-permutation-distributions-via-reflected-diffusion-on-ranks - Soft-Rank Diffusion offers a novel approach to learning permutation distributions through a structured soft-rank process.
- OmniVLN: Omnidirectional 3D Perception and Token-Efficient LLM Reasoning for Visual-Language Navigation across Air and Ground Platforms (viability: 8): https://sciencetostartup.com/paper/omnivln-omnidirectional-3d-perception-and-token-efficient-llm-reasoning-for-visual-language-navigation-across-air-and-gr - OmniVLN enhances visual-language navigation for robots using efficient 3D perception and reasoning.
- EvoGuard: An Extensible Agentic RL-based Framework for Practical and Evolving AI-Generated Image Detection (viability: 8): https://sciencetostartup.com/paper/evoguard-an-extensible-agentic-rl-based-framework-for-practical-and-evolving-ai-generated-image-detection - EvoGuard is an extensible framework for detecting AI-generated images using a dynamic orchestration of multimodal detectors.
- Grid Spatial Understanding: A Dataset for Textual Spatial Reasoning over Grids, Embodied Settings, and Coordinate Structures (viability: 5): https://sciencetostartup.com/paper/grid-spatial-understanding-a-dataset-for-textual-spatial-reasoning-over-grids-embodied-settings-and-coordinate-structure - GSU is a dataset designed to enhance spatial reasoning in language models through text-only grid tasks.
- Federated Computing as Code (FCaC): Sovereignty-aware Systems by Design (viability: 7): https://sciencetostartup.com/paper/federated-computing-as-code-fcac-sovereignty-aware-systems-by-design - Federated Computing as Code (FCaC) offers a sovereignty-aware architecture for secure collaborative computation across organizations.
- A Progressive Visual-Logic-Aligned Framework for Ride-Hailing Adjudication (viability: 3): https://sciencetostartup.com/paper/a-progressive-visual-logic-aligned-framework-for-ride-hailing-adjudication - RideJudge offers a novel framework for adjudicating responsibility disputes in ride-hailing through enhanced reasoning transparency.
- FineViT: Progressively Unlocking Fine-Grained Perception with Dense Recaptions (viability: 7): https://sciencetostartup.com/paper/finevit-progressively-unlocking-fine-grained-perception-with-dense-recaptions - FineViT is a novel vision encoder designed to enhance fine-grained perception through progressive training on high-quality recaptioned datasets.
- MedSAD-CLIP: Supervised CLIP with Token-Patch Cross-Attention for Medical Anomaly Detection and Segmentation (viability: 8): https://sciencetostartup.com/paper/medsad-clip-supervised-clip-with-token-patch-cross-attention-for-medical-anomaly-detection-and-segmentation - MedSAD-CLIP enhances medical anomaly detection and segmentation using supervised CLIP adaptation for improved localization and classification.
- ShuttleEnv: An Interactive Data-Driven RL Environment for Badminton Strategy Modeling (viability: 8): https://sciencetostartup.com/paper/shuttleenv-an-interactive-data-driven-rl-environment-for-badminton-strategy-modeling - ShuttleEnv is an interactive simulation environment for badminton that leverages reinforcement learning to model strategic behaviors in sports.
- DexEXO: A Wearability-First Dexterous Exoskeleton for Operator-Agnostic Demonstration and Learning (viability: 7): https://sciencetostartup.com/paper/dexexo-a-wearability-first-dexterous-exoskeleton-for-operator-agnostic-demonstration-and-learning - DexEXO is a wearability-first dexterous exoskeleton designed for scalable cross-operator robot learning.
- Physics-informed offline reinforcement learning eliminates catastrophic fuel waste in maritime routing (viability: 7): https://sciencetostartup.com/paper/physics-informed-offline-reinforcement-learning-eliminates-catastrophic-fuel-waste-in-maritime-routing - PIER is an offline reinforcement learning framework that optimizes maritime routing to significantly reduce fuel waste and emissions.
- A Proposal-Free Query-Guided Network for Grounded Multimodal Named Entity Recognition (viability: 7): https://sciencetostartup.com/paper/a-proposal-free-query-guided-network-for-grounded-multimodal-named-entity-recognition - A novel Query-Guided Network for precise multimodal named entity recognition that enhances grounding accuracy.
- Recurrent Reasoning with Vision-Language Models for Estimating Long-Horizon Embodied Task Progress (viability: 8): https://sciencetostartup.com/paper/recurrent-reasoning-with-vision-language-models-for-estimating-long-horizon-embodied-task-progress - Recurrent Reasoning Vision-Language Model ($R^2$VLM) enhances task progress estimation for embodied agents using a novel reasoning framework.
- Ruyi2.5 Technical Report (viability: 7): https://sciencetostartup.com/paper/ruyi2-5-technical-report - Ruyi2.5 is a multimodal familial model that enhances privacy in surveillance through a unique two-stage recognition pipeline.
- InfoDensity: Rewarding Information-Dense Traces for Efficient Reasoning (viability: 7): https://sciencetostartup.com/paper/infodensity-rewarding-information-dense-traces-for-efficient-reasoning - InfoDensity optimizes reasoning quality in LLMs by rewarding information-dense traces for efficient reasoning.
- ReLMXEL: Adaptive RL-Based Memory Controller with Explainable Energy and Latency Optimization (viability: 4): https://sciencetostartup.com/paper/relmxel-adaptive-rl-based-memory-controller-with-explainable-energy-and-latency-optimization - ReLMXEL is an adaptive reinforcement learning framework that optimizes memory controller performance while providing explainable energy and latency management.
- Symphony: A Cognitively-Inspired Multi-Agent System for Long-Video Understanding (viability: 9): https://sciencetostartup.com/paper/symphony-a-cognitively-inspired-multi-agent-system-for-long-video-understanding - Symphony is a multi-agent system designed for enhanced long-video understanding through cognitive-inspired reasoning.
- Beyond bouba/kiki: Multidimensional semantic signals are deeply woven into the fabric of natural language (viability: 2): https://sciencetostartup.com/paper/beyond-bouba-kiki-multidimensional-semantic-signals-are-deeply-woven-into-the-fabric-of-natural-language - This research explores the structured semantic signals of phonemes in language, challenging traditional views on sound-meaning relationships.
- Contrastive Reasoning Alignment: Reinforcement Learning from Hidden Representations (viability: 8): https://sciencetostartup.com/paper/contrastive-reasoning-alignment-reinforcement-learning-from-hidden-representations - CRAFT is a robust alignment framework that enhances reasoning safety in AI models against jailbreak attacks.
- 3D MRI-Based Alzheimer's Disease Classification Using Multi-Modal 3D CNN with Leakage-Aware Subject-Level Evaluation (viability: 5): https://sciencetostartup.com/paper/3d-mri-based-alzheimer-s-disease-classification-using-multi-modal-3d-cnn-with-leakage-aware-subject-level-evaluation - A multimodal 3D CNN framework for classifying Alzheimer's disease using volumetric MRI data.
- From Words to Worlds: Benchmarking Cross-Cultural Cultural Understanding in Machine Translation (viability: 8): https://sciencetostartup.com/paper/from-words-to-worlds-benchmarking-cross-cultural-cultural-understanding-in-machine-translation - CulT-Eval is a benchmark for evaluating machine translation of culturally grounded expressions, addressing gaps in current evaluation metrics.
- WINFlowNets: Warm-up Integrated Networks Training of Generative Flow Networks for Robotics and Machine Fault Adaptation (viability: 6): https://sciencetostartup.com/paper/winflownets-warm-up-integrated-networks-training-of-generative-flow-networks-for-robotics-and-machine-fault-adaptation - WINFlowNets enhances robotic control by enabling co-training of flow and retrieval networks for improved adaptability in dynamic environments.
- ReSteer: Quantifying and Refining the Steerability of Multitask Robot Policies (viability: 7): https://sciencetostartup.com/paper/resteer-quantifying-and-refining-the-steerability-of-multitask-robot-policies - ReSteer enhances multitask robot policies by quantifying and improving their steerability for better task execution.
- GUIDE: GenAI Units In Digital Design Education (viability: 8): https://sciencetostartup.com/paper/guide-genai-units-in-digital-design-education - GUIDE is an open courseware repository that enhances digital design education through AI-assisted learning units and interactive labs.
- Directing the Narrative: A Finetuning Method for Controlling Coherence and Style in Story Generation (viability: 8): https://sciencetostartup.com/paper/directing-the-narrative-a-finetuning-method-for-controlling-coherence-and-style-in-story-generation - A novel framework for generating coherent and stylistically consistent story visuals using advanced attention mechanisms.
- SEAL-Tag: Self-Tag Evidence Aggregation with Probabilistic Circuits for PII-Safe Retrieval-Augmented Generation (viability: 7): https://sciencetostartup.com/paper/seal-tag-self-tag-evidence-aggregation-with-probabilistic-circuits-for-pii-safe-retrieval-augmented-generation - SEAL-Tag is a privacy-preserving RAG system that prevents PII leakage while maintaining performance.
- Classifier Pooling for Modern Ordinal Classification (viability: 8): https://sciencetostartup.com/paper/classifier-pooling-for-modern-ordinal-classification - A model-agnostic method for ordinal classification that enhances performance using modern machine learning techniques.
- DANCE: Dynamic 3D CNN Pruning: Joint Frame, Channel, and Feature Adaptation for Energy Efficiency on the Edge (viability: 7): https://sciencetostartup.com/paper/dance-dynamic-3d-cnn-pruning-joint-frame-channel-and-feature-adaptation-for-energy-efficiency-on-the-edge - DANCE is a dynamic pruning framework for 3D CNNs that enhances energy efficiency while maintaining performance.
- Network- and Device-Level Cyber Deception for Contested Environments Using RL and LLMs (viability: 2): https://sciencetostartup.com/paper/network-and-device-level-cyber-deception-for-contested-environments-using-rl-and-llms - This research explores AI-driven cyber deception strategies to enhance security against stealthy attacks in contested environments.
- Wasserstein-type Gaussian Process Regressions for Input Measurement Uncertainty (viability: 2): https://sciencetostartup.com/paper/wasserstein-type-gaussian-process-regressions-for-input-measurement-uncertainty - This paper presents a new approach to Gaussian process regression that improves uncertainty quantification in the presence of input measurement errors.
- ConfusionBench: An Expert-Validated Benchmark for Confusion Recognition and Localization in Educational Videos (viability: 7): https://sciencetostartup.com/paper/confusionbench-an-expert-validated-benchmark-for-confusion-recognition-and-localization-in-educational-videos - ConfusionBench is a validated benchmark for recognizing and localizing student confusion in educational videos, enhancing intervention strategies.
- Revisiting Vulnerability Patch Identification on Data in the Wild (viability: 5): https://sciencetostartup.com/paper/revisiting-vulnerability-patch-identification-on-data-in-the-wild - A novel approach to improve vulnerability patch detection by combining NVD data with manually identified patches.
- LED: A Benchmark for Evaluating Layout Error Detection in Document Analysis (viability: 4): https://sciencetostartup.com/paper/led-a-benchmark-for-evaluating-layout-error-detection-in-document-analysis - LED is a benchmark for evaluating layout error detection in document analysis, addressing structural reasoning beyond surface-level accuracy.
- Deanonymizing Bitcoin Transactions via Network Traffic Analysis with Semi-supervised Learning (viability: 5): https://sciencetostartup.com/paper/deanonymizing-bitcoin-transactions-via-network-traffic-analysis-with-semi-supervised-learning - NTSSL is a novel method for deanonymizing Bitcoin transactions using network traffic analysis and semi-supervised learning.
- Variational Rectification Inference for Learning with Noisy Labels (viability: 4): https://sciencetostartup.com/paper/variational-rectification-inference-for-learning-with-noisy-labels - A novel variational inference approach to improve model robustness against noisy labels in deep learning.
- Pathology-Aware Multi-View Contrastive Learning for Patient-Independent ECG Reconstruction (viability: 7): https://sciencetostartup.com/paper/pathology-aware-multi-view-contrastive-learning-for-patient-independent-ecg-reconstruction - A novel framework for reconstructing ECGs that enhances diagnostic accuracy by incorporating pathology-aware embeddings.
- Binary Latent Protein Fitness Landscapes for Quantum Annealing Optimization (viability: 8): https://sciencetostartup.com/paper/binary-latent-protein-fitness-landscapes-for-quantum-annealing-optimization - Q-BIOLAT optimizes protein fitness landscapes using binary latent representations and quantum annealing techniques.
- On the Cone Effect and Modality Gap in Medical Vision-Language Embeddings (viability: 4): https://sciencetostartup.com/paper/on-the-cone-effect-and-modality-gap-in-medical-vision-language-embeddings - A lightweight mechanism to control cross-modal separation in medical vision-language models, enhancing multimodal performance without retraining.
- Graph-Native Cognitive Memory for AI Agents: Formal Belief Revision Semantics for Versioned Memory Architectures (viability: 8): https://sciencetostartup.com/paper/graph-native-cognitive-memory-for-ai-agents-formal-belief-revision-semantics-for-versioned-memory-architectures - Kumiho is a graph-native cognitive memory architecture that enhances AI agents' memory capabilities through formal belief revision semantics.
- GigaWorld-Policy: An Efficient Action-Centered World--Action Model (viability: 8): https://sciencetostartup.com/paper/gigaworld-policy-an-efficient-action-centered-world-action-model - GigaWorld-Policy revolutionizes robot policy learning with an efficient action-centered model that enhances performance and reduces inference time.
- LAAF: Logic-layer Automated Attack Framework A Systematic Red-Teaming Methodology for LPCI Vulnerabilities in Agentic Large Language Model Systems (viability: 7): https://sciencetostartup.com/paper/laaf-logic-layer-automated-attack-framework-a-systematic-red-teaming-methodology-for-lpci-vulnerabilities-in-agentic-lar - LAAF is an automated red-teaming framework designed to identify vulnerabilities in agentic large language model systems through advanced attack techniques.
- Neural Radiance Maps for Extraterrestrial Navigation and Path Planning (viability: 7): https://sciencetostartup.com/paper/neural-radiance-maps-for-extraterrestrial-navigation-and-path-planning - A novel path planning framework using Neural Radiance Fields for autonomous navigation on extraterrestrial surfaces.
- Deployment and Evaluation of an EHR-integrated, Large Language Model-Powered Tool to Triage Surgical Patients (viability: 8): https://sciencetostartup.com/paper/deployment-and-evaluation-of-an-ehr-integrated-large-language-model-powered-tool-to-triage-surgical-patients - An LLM-powered tool that automates the triage of surgical patients, integrating seamlessly with electronic health records.
- Draft-and-Prune: Improving the Reliability of Auto-formalization for Logical Reasoning (viability: 7): https://sciencetostartup.com/paper/draft-and-prune-improving-the-reliability-of-auto-formalization-for-logical-reasoning - Draft-and-Prune enhances auto-formalization for logical reasoning by improving execution reliability through diverse program generation and verification.
- Full Stack Navigation, Mapping, and Planning for the Lunar Autonomy Challenge (viability: 7): https://sciencetostartup.com/paper/full-stack-navigation-mapping-and-planning-for-the-lunar-autonomy-challenge - A modular autonomy system for lunar navigation and mapping, achieving centimeter-level accuracy in challenging environments.
- Neuron-Level Emotion Control in Speech-Generative Large Audio-Language Models (viability: 7): https://sciencetostartup.com/paper/neuron-level-emotion-control-in-speech-generative-large-audio-language-models - A novel approach to achieve training-free emotion control in speech-generative models using neuron-level interventions.
- KANtize: Exploring Low-bit Quantization of Kolmogorov-Arnold Networks for Efficient Inference (viability: 3): https://sciencetostartup.com/paper/kantize-exploring-low-bit-quantization-of-kolmogorov-arnold-networks-for-efficient-inference - KANtize explores low-bit quantization of Kolmogorov-Arnold Networks to enhance inference efficiency.
- Visual SLAM with DEM Anchoring for Lunar Surface Navigation (viability: 6): https://sciencetostartup.com/paper/visual-slam-with-dem-anchoring-for-lunar-surface-navigation - A stereo visual SLAM system for autonomous lunar rovers that mitigates drift using digital elevation models.
- From Drop-off to Recovery: A Mechanistic Analysis of Segmentation in MLLMs (viability: 2): https://sciencetostartup.com/paper/from-drop-off-to-recovery-a-mechanistic-analysis-of-segmentation-in-mllms - This paper analyzes the segmentation capabilities of Multimodal Large Language Models through a mechanistic lens.
- Adaptive Anchor Policies for Efficient 4D Gaussian Streaming (viability: 8): https://sciencetostartup.com/paper/adaptive-anchor-policies-for-efficient-4d-gaussian-streaming - Efficient Gaussian Streaming optimizes anchor selection for real-time rendering, enhancing quality and efficiency in dynamic scene reconstruction.
- TharuChat: Bootstrapping Large Language Models for a Low-Resource Language via Synthetic Data and Human Validation (viability: 7): https://sciencetostartup.com/paper/tharuchat-bootstrapping-large-language-models-for-a-low-resource-language-via-synthetic-data-and-human-validation - TharuChat leverages synthetic data and human validation to bootstrap a language model for the under-resourced Tharu language, promoting linguistic diversity.
- SA-CycleGAN-2.5D: Self-Attention CycleGAN with Tri-Planar Context for Multi-Site MRI Harmonization (viability: 3): https://sciencetostartup.com/paper/sa-cyclegan-2-5d-self-attention-cyclegan-with-tri-planar-context-for-multi-site-mri-harmonization - SA-CycleGAN-2.5D offers a novel framework for harmonizing multi-site MRI scans to enhance radiomic reproducibility.
- Alignment Makes Language Models Normative, Not Descriptive (viability: 2): https://sciencetostartup.com/paper/alignment-makes-language-models-normative-not-descriptive - This paper explores the trade-off between aligning language models with human preferences and accurately predicting human behavior in strategic games.
- Anonymous-by-Construction: An LLM-Driven Framework for Privacy-Preserving Text (viability: 8): https://sciencetostartup.com/paper/anonymous-by-construction-an-llm-driven-framework-for-privacy-preserving-text - An LLM-driven framework that anonymizes text while preserving its utility, ensuring responsible AI deployment.
- AI Scientist via Synthetic Task Scaling (viability: 7): https://sciencetostartup.com/paper/ai-scientist-via-synthetic-task-scaling - A synthetic environment generation pipeline that enhances machine learning agents through tailored challenges and datasets.
- Adaptive Contracts for Cost-Effective AI Delegation (viability: 5): https://sciencetostartup.com/paper/adaptive-contracts-for-cost-effective-ai-delegation - Adaptive contracts optimize AI delegation by selectively evaluating performance to reduce costs.
- A scalable neural bundle map for multiphysics prediction in lithium-ion battery across varying configurations (viability: 4): https://sciencetostartup.com/paper/a-scalable-neural-bundle-map-for-multiphysics-prediction-in-lithium-ion-battery-across-varying-configurations - A Neural Bundle Map framework for efficient multiphysics prediction in lithium-ion batteries.
- SYMDIREC: A Neuro-Symbolic Divide-Retrieve-Conquer Framework for Enhanced RTL Synthesis and Summarization (viability: 7): https://sciencetostartup.com/paper/symdirec-a-neuro-symbolic-divide-retrieve-conquer-framework-for-enhanced-rtl-synthesis-and-summarization - SYMDIREC is a neuro-symbolic framework that enhances RTL synthesis and summarization by integrating symbolic planning with LLM reasoning.
- OPERA: Online Data Pruning for Efficient Retrieval Model Adaptation (viability: 7): https://sciencetostartup.com/paper/opera-online-data-pruning-for-efficient-retrieval-model-adaptation - OPERA is a data pruning framework that enhances retrieval model adaptation by prioritizing high-quality training pairs for improved efficiency and effectiveness.
- CODMAS: A Dialectic Multi-Agent Collaborative Framework for Structured RTL Optimization (viability: 7): https://sciencetostartup.com/paper/codmas-a-dialectic-multi-agent-collaborative-framework-for-structured-rtl-optimization - CODMAS is a multi-agent framework that automates RTL optimization to enhance power, performance, and area in electronic design.
- FastLoop: Parallel Loop Closing with GPU-Acceleration in Visual SLAM (viability: 7): https://sciencetostartup.com/paper/fastloop-parallel-loop-closing-with-gpu-acceleration-in-visual-slam - FastLoop accelerates loop closing in Visual SLAM using GPU optimizations for faster and more efficient mapping.
- Catching rationalization in the act: detecting motivated reasoning before and after CoT via activation probing (viability: 7): https://sciencetostartup.com/paper/catching-rationalization-in-the-act-detecting-motivated-reasoning-before-and-after-cot-via-activation-probing - A method to detect motivated reasoning in LLMs using activation probing for improved reliability in decision-making.
- Abstraction as a Memory-Efficient Inductive Bias for Continual Learning (viability: 5): https://sciencetostartup.com/paper/abstraction-as-a-memory-efficient-inductive-bias-for-continual-learning - AAT introduces a memory-efficient inductive bias for continual learning, stabilizing learning in online data streams without the need for replay buffers.
- Self-Conditioned Denoising for Atomistic Representation Learning (viability: 9): https://sciencetostartup.com/paper/self-conditioned-denoising-for-atomistic-representation-learning - Self-Conditioned Denoising (SCD) revolutionizes atomistic data representation learning by significantly enhancing performance on property prediction tasks.
- Tabular LLMs for Interpretable Few-Shot Alzheimer's Disease Prediction with Multimodal Biomedical Data (viability: 9): https://sciencetostartup.com/paper/tabular-llms-for-interpretable-few-shot-alzheimer-s-disease-prediction-with-multimodal-biomedical-data - TAP-GPT is a domain-adapted tabular LLM for accurate Alzheimer's disease prediction using multimodal biomedical data.
- Influence of Gripper Design on Human Demonstration Quality for Robot Learning (viability: 4): https://sciencetostartup.com/paper/influence-of-gripper-design-on-human-demonstration-quality-for-robot-learning - Improving robot learning through enhanced gripper design for healthcare applications.
- MetaClaw: Just Talk -- An Agent That Meta-Learns and Evolves in the Wild (viability: 9): https://sciencetostartup.com/paper/metaclaw-just-talk-an-agent-that-meta-learns-and-evolves-in-the-wild - MetaClaw is a continual meta-learning framework that enables LLM agents to adapt and evolve in real-time without downtime.
- Visual Product Search Benchmark (viability: 7): https://sciencetostartup.com/paper/visual-product-search-benchmark - A benchmark for visual embedding models tailored for accurate product identification in industrial applications.
- Patient4D: Temporally Consistent Patient Body Mesh Recovery from Monocular Operating Room Video (viability: 7): https://sciencetostartup.com/paper/patient4d-temporally-consistent-patient-body-mesh-recovery-from-monocular-operating-room-video - Patient4D enhances 3D body mesh recovery from surgical videos, improving accuracy under occlusion.
- Towards Unsupervised Adversarial Document Detection in Retrieval Augmented Generation Systems (viability: 2): https://sciencetostartup.com/paper/towards-unsupervised-adversarial-document-detection-in-retrieval-augmented-generation-systems - An unsupervised method for detecting adversarial documents in retrieval augmented generation systems.
- Domain-informed explainable boosting machines for trustworthy lateral spread predictions (viability: 5): https://sciencetostartup.com/paper/domain-informed-explainable-boosting-machines-for-trustworthy-lateral-spread-predictions - A domain-informed framework enhances Explainable Boosting Machines for reliable lateral spread predictions in natural hazard applications.
- Detecting Data Poisoning in Code Generation LLMs via Black-Box, Vulnerability-Oriented Scanning (viability: 7): https://sciencetostartup.com/paper/detecting-data-poisoning-in-code-generation-llms-via-black-box-vulnerability-oriented-scanning - CodeScan is a pioneering framework for detecting data poisoning in code generation LLMs through structural analysis.
- Generalist Multimodal LLMs Gain Biometric Expertise via Human Salience (viability: 8): https://sciencetostartup.com/paper/generalist-multimodal-llms-gain-biometric-expertise-via-human-salience - A multimodal LLM solution for iris presentation attack detection that respects privacy constraints and outperforms traditional methods.
- Noise-Response Calibration: A Causal Intervention Protocol for LLM-Judges (viability: 7): https://sciencetostartup.com/paper/noise-response-calibration-a-causal-intervention-protocol-for-llm-judges - A calibration protocol for LLMs to improve their reliability as automated judges in low-label settings.
- Exploiting the English Grammar Profile for L2 grammatical analysis with LLMs (viability: 3): https://sciencetostartup.com/paper/exploiting-the-english-grammar-profile-for-l2-grammatical-analysis-with-llms - A framework for evaluating L2 grammatical competence using LLMs and the English Grammar Profile.
- PAuth - Precise Task-Scoped Authorization For Agents (viability: 6): https://sciencetostartup.com/paper/pauth-precise-task-scoped-authorization-for-agents - PAuth introduces a precise task-scoped authorization model for AI agents to enhance security and reduce overprivilege.
- How Clued up are LLMs? Evaluating Multi-Step Deductive Reasoning in a Text-Based Game Environment (viability: 2): https://sciencetostartup.com/paper/how-clued-up-are-llms-evaluating-multi-step-deductive-reasoning-in-a-text-based-game-environment - This paper evaluates the deductive reasoning capabilities of LLM agents in a text-based game environment.
- SLAM Adversarial Lab: An Extensible Framework for Visual SLAM Robustness Evaluation under Adverse Conditions (viability: 5): https://sciencetostartup.com/paper/slam-adversarial-lab-an-extensible-framework-for-visual-slam-robustness-evaluation-under-adverse-conditions - SAL is a modular framework for evaluating visual SLAM systems under adverse conditions like fog and rain.
- GazeOnce360: Fisheye-Based 360° Multi-Person Gaze Estimation with Global-Local Feature Fusion (viability: 7): https://sciencetostartup.com/paper/gazeonce360-fisheye-based-360-multi-person-gaze-estimation-with-global-local-feature-fusion - GazeOnce360 enables accurate multi-person gaze estimation using a single fisheye camera for diverse environments.
- Self-Regularized Learning Methods (viability: 2): https://sciencetostartup.com/paper/self-regularized-learning-methods - A theoretical framework for analyzing self-regularized learning algorithms.
- BEV-SLD: Self-Supervised Scene Landmark Detection for Global Localization with LiDAR Bird's-Eye View Images (viability: 3): https://sciencetostartup.com/paper/bev-sld-self-supervised-scene-landmark-detection-for-global-localization-with-lidar-bird-s-eye-view-images - BEV-SLD offers a self-supervised method for robust global localization using LiDAR bird's-eye view images.
- Shielded Reinforcement Learning Under Dynamic Temporal Logic Constraints (viability: 5): https://sciencetostartup.com/paper/shielded-reinforcement-learning-under-dynamic-temporal-logic-constraints - A framework for safe reinforcement learning that enforces complex temporal logic constraints during the learning process.
- Intent Formalization: A Grand Challenge for Reliable Coding in the Age of AI Agents (viability: 4): https://sciencetostartup.com/paper/intent-formalization-a-grand-challenge-for-reliable-coding-in-the-age-of-ai-agents - A framework for translating user intent into formal specifications to enhance the reliability of AI-generated code.
- Synchronized DNA sources for unconditionally secure cryptography (viability: 4): https://sciencetostartup.com/paper/synchronized-dna-sources-for-unconditionally-secure-cryptography - A DNA-based cryptographic system that enables secure key distribution for unconditional security in communications.
- Personalized Fall Detection by Balancing Data with Selective Feedback Using Contrastive Learning (viability: 7): https://sciencetostartup.com/paper/personalized-fall-detection-by-balancing-data-with-selective-feedback-using-contrastive-learning - A personalized fall detection system that enhances accuracy by balancing data through selective feedback and contrastive learning.
- Multilingual Reference Need Assessment System for Wikipedia (viability: 8): https://sciencetostartup.com/paper/multilingual-reference-need-assessment-system-for-wikipedia - A multilingual machine learning system that assists Wikipedia editors in identifying claims needing citations, enhancing content verifiability.
- REAL: Regression-Aware Reinforcement Learning for LLM-as-a-Judge (viability: 7): https://sciencetostartup.com/paper/real-regression-aware-reinforcement-learning-for-llm-as-a-judge - REAL is a novel reinforcement learning framework that optimizes regression rewards for large language models acting as evaluators.
- Contextual Preference Distribution Learning (viability: 2): https://sciencetostartup.com/paper/contextual-preference-distribution-learning - A novel approach to learn and optimize preference distributions for risk-averse decision-making in uncertain environments.
- A Longitudinal Study of Usability in Identity-Based Software Signing (viability: 3): https://sciencetostartup.com/paper/a-longitudinal-study-of-usability-in-identity-based-software-signing - A study analyzing usability issues in identity-based software signing tools to improve adoption and verification workflows.
- SMAL-pets: SMAL Based Avatars of Pets from Single Image (viability: 7): https://sciencetostartup.com/paper/smal-pets-smal-based-avatars-of-pets-from-single-image - SMAL-pets generates high-quality, editable 3D avatars of pets from a single image using a hybrid modeling approach.
- Topology-Preserving Deep Joint Source-Channel Coding for Semantic Communication (viability: 2): https://sciencetostartup.com/paper/topology-preserving-deep-joint-source-channel-coding-for-semantic-communication - TopoJSCC enhances topology preservation in wireless vision applications through a novel DeepJSCC framework.
- Security Assessment and Mitigation Strategies for Large Language Models: A Comprehensive Defensive Framework (viability: 7): https://sciencetostartup.com/paper/security-assessment-and-mitigation-strategies-for-large-language-models-a-comprehensive-defensive-framework - A comprehensive framework for assessing and mitigating security risks in large language models.
- MosaicMem: Hybrid Spatial Memory for Controllable Video World Models (viability: 7): https://sciencetostartup.com/paper/mosaicmem-hybrid-spatial-memory-for-controllable-video-world-models - MosaicMem enhances video world models with a hybrid spatial memory for improved localization and dynamic modeling.
- Cascade-Aware Multi-Agent Routing: Spatio-Temporal Sidecars and Geometry-Switching (viability: 7): https://sciencetostartup.com/paper/cascade-aware-multi-agent-routing-spatio-temporal-sidecars-and-geometry-switching - A novel approach to enhance multi-agent routing by mitigating geometry-blind failure propagation in execution graphs.
- Hidden Clones: Exposing and Fixing Family Bias in Vision-Language Model Ensembles (viability: 6): https://sciencetostartup.com/paper/hidden-clones-exposing-and-fixing-family-bias-in-vision-language-model-ensembles - A novel approach to improve ensemble accuracy in Vision-Language Models by addressing family bias.
- Pixel-level Counterfactual Contrastive Learning for Medical Image Segmentation (viability: 7): https://sciencetostartup.com/paper/pixel-level-counterfactual-contrastive-learning-for-medical-image-segmentation - A novel pipeline for pixel-level image segmentation using counterfactual generation and contrastive learning to improve annotation efficiency.
- SENSE: Efficient EEG-to-Text via Privacy-Preserving Semantic Retrieval (viability: 7): https://sciencetostartup.com/paper/sense-efficient-eeg-to-text-via-privacy-preserving-semantic-retrieval - SENSE is a lightweight EEG-to-text framework that ensures privacy while translating brain activity into natural language.
- LLM-Powered Flood Depth Estimation from Social Media Imagery: A Vision-Language Model Framework with Mechanistic Interpretability for Transportation Resilience (viability: 8): https://sciencetostartup.com/paper/llm-powered-flood-depth-estimation-from-social-media-imagery-a-vision-language-model-framework-with-mechanistic-interpre - FloodLlama is an open-source vision-language model for real-time flood depth estimation from social media imagery, enhancing transportation resilience.
- When the Specification Emerges: Benchmarking Faithfulness Loss in Long-Horizon Coding Agents (viability: 7): https://sciencetostartup.com/paper/when-the-specification-emerges-benchmarking-faithfulness-loss-in-long-horizon-coding-agents - A benchmark and tool for improving faithfulness in long-horizon coding agents through specification tracking.
- Knowledge Localization in Mixture-of-Experts LLMs Using Cross-Lingual Inconsistency (viability: 3): https://sciencetostartup.com/paper/knowledge-localization-in-mixture-of-experts-llms-using-cross-lingual-inconsistency - A framework for leveraging cross-lingual inconsistencies in LLMs to enhance interpretability.
- An End-to-End Framework for Functionality-Embedded Provenance Graph Construction and Threat Interpretation (viability: 3): https://sciencetostartup.com/paper/an-end-to-end-framework-for-functionality-embedded-provenance-graph-construction-and-threat-interpretation - Auto-Prov is an adaptive framework that enhances anomaly detection through automated provenance graph construction from logs.
- Accurate Shift Invariant Convolutional Neural Networks Using Gaussian-Hermite Moments (viability: 7): https://sciencetostartup.com/paper/accurate-shift-invariant-convolutional-neural-networks-using-gaussian-hermite-moments - Introducing Gaussian-Hermite Sampling for CNNs to achieve accurate shift invariance without architectural changes.
- Evaluating LLM-Simulated Conversations in Modeling Inconsistent and Uncollaborative Behaviors in Human Social Interaction (viability: 2): https://sciencetostartup.com/paper/evaluating-llm-simulated-conversations-in-modeling-inconsistent-and-uncollaborative-behaviors-in-human-social-interactio - CoCoEval is an evaluation framework for analyzing inconsistent and uncollaborative behaviors in LLM-simulated conversations.
- SLowRL: Safe Low-Rank Adaptation Reinforcement Learning for Locomotion (viability: 7): https://sciencetostartup.com/paper/slowrl-safe-low-rank-adaptation-reinforcement-learning-for-locomotion - SLowRL enables safe and efficient fine-tuning of locomotion policies for robots, reducing fine-tuning time and safety violations.
- Ensemble Self-Training for Unsupervised Machine Translation (viability: 3): https://sciencetostartup.com/paper/ensemble-self-training-for-unsupervised-machine-translation - A novel ensemble self-training framework for improving unsupervised neural machine translation.
- ACE-LoRA: Graph-Attentive Context Enhancement for Parameter-Efficient Adaptation of Medical Vision-Language Models (viability: 9): https://sciencetostartup.com/paper/ace-lora-graph-attentive-context-enhancement-for-parameter-efficient-adaptation-of-medical-vision-language-models - ACE-LoRA enhances medical vision-language models with parameter-efficient adaptation for improved diagnostic accuracy.
- CircuitBuilder: From Polynomials to Circuits via Reinforcement Learning (viability: 4): https://sciencetostartup.com/paper/circuitbuilder-from-polynomials-to-circuits-via-reinforcement-learning - CircuitBuilder leverages reinforcement learning to efficiently discover arithmetic circuits for polynomial computation.
- PRISM: Demystifying Retention and Interaction in Mid-Training (viability: 4): https://sciencetostartup.com/paper/prism-demystifying-retention-and-interaction-in-mid-training - PRISM offers insights into optimizing mid-training for large language models to enhance reasoning capabilities.
- Large Reasoning Models Struggle to Transfer Parametric Knowledge Across Scripts (viability: 5): https://sciencetostartup.com/paper/large-reasoning-models-struggle-to-transfer-parametric-knowledge-across-scripts - This research identifies and addresses the script barrier in cross-lingual knowledge transfer for large language models.
- Edge-Efficient Two-Stream Multimodal Architecture for Non-Intrusive Bathroom Fall Detection (viability: 7): https://sciencetostartup.com/paper/edge-efficient-two-stream-multimodal-architecture-for-non-intrusive-bathroom-fall-detection - A two-stream multimodal architecture for real-time, non-intrusive bathroom fall detection using radar and vibration data.
- TrackDeform3D: Markerless and Autonomous 3D Keypoint Tracking and Dataset Collection for Deformable Objects (viability: 7): https://sciencetostartup.com/paper/trackdeform3d-markerless-and-autonomous-3d-keypoint-tracking-and-dataset-collection-for-deformable-objects - TrackDeform3D offers an autonomous solution for collecting high-quality 3D datasets of deformable objects using RGB-D cameras.
- Evaluating Ill-Defined Tasks in Large Language Models (viability: 3): https://sciencetostartup.com/paper/evaluating-ill-defined-tasks-in-large-language-models - This paper critiques the evaluation of Large Language Models on ill-defined tasks and proposes improvements for more reliable assessments.
- TeleDex: Accessible Dexterous Teleoperation (viability: 8): https://sciencetostartup.com/paper/teledex-accessible-dexterous-teleoperation - TeleDex is an open-source system that enables intuitive teleoperation of robotic manipulators using just a smartphone.
- Transformers are Bayesian Networks (viability: 2): https://sciencetostartup.com/paper/transformers-are-bayesian-networks - This paper theoretically establishes that transformers function as Bayesian networks, providing insights into their operational mechanics.
- Asymmetric Nash Seeking via Best Response Maps: Global Linear Convergence and Robustness to Inexact Reaction Models (viability: 3): https://sciencetostartup.com/paper/asymmetric-nash-seeking-via-best-response-maps-global-linear-convergence-and-robustness-to-inexact-reaction-models - A method for finding Nash equilibria in multi-agent systems without full knowledge of all players' objectives.
- DesertFormer: Transformer-Based Semantic Segmentation for Off-Road Desert Terrain Classification in Autonomous Navigation Systems (viability: 9): https://sciencetostartup.com/paper/desertformer-transformer-based-semantic-segmentation-for-off-road-desert-terrain-classification-in-autonomous-navigation - DesertFormer is a semantic segmentation tool for classifying off-road desert terrain to enhance autonomous navigation safety.
- PaAgent: Portrait-Aware Image Restoration Agent via Subjective-Objective Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/paagent-portrait-aware-image-restoration-agent-via-subjective-objective-reinforcement-learning - PaAgent is a portrait-aware image restoration agent that intelligently selects restoration tools using a self-evolving portrait bank and reinforcement learning.
- Early Quantization Shrinks Codebook: A Simple Fix for Diversity-Preserving Tokenization (viability: 4): https://sciencetostartup.com/paper/early-quantization-shrinks-codebook-a-simple-fix-for-diversity-preserving-tokenization - A study addressing representation collapsing in vector quantization for improved tokenization in generative models.
- Astrolabe: Steering Forward-Process Reinforcement Learning for Distilled Autoregressive Video Models (viability: 7): https://sciencetostartup.com/paper/astrolabe-steering-forward-process-reinforcement-learning-for-distilled-autoregressive-video-models - Astrolabe is an efficient online reinforcement learning framework designed to enhance distilled autoregressive video models by aligning them with human visual preferences.
- SCE-LITE-HQ: Smooth visual counterfactual explanations with generative foundation models (viability: 7): https://sciencetostartup.com/paper/sce-lite-hq-smooth-visual-counterfactual-explanations-with-generative-foundation-models - SCE-LITE-HQ is a scalable framework for generating counterfactual explanations using pretrained generative models, enhancing interpretability in high-dimensional visual domains.
- Do Understanding and Generation Fight? A Diagnostic Study of DPO for Unified Multimodal Models (viability: 2): https://sciencetostartup.com/paper/do-understanding-and-generation-fight-a-diagnostic-study-of-dpo-for-unified-multimodal-models - This study investigates the limitations of DPO alignment in unified multimodal models, revealing significant challenges in balancing understanding and generation.
- OpenQlaw: An Agentic AI Assistant for Analysis of 2D Quantum Materials (viability: 3): https://sciencetostartup.com/paper/openqlaw-an-agentic-ai-assistant-for-analysis-of-2d-quantum-materials - OpenQlaw is an agentic AI assistant designed to enhance the analysis and fabrication of 2D quantum materials through dynamic reasoning.
- Dependence Fidelity and Downstream Inference Stability in Generative Models (viability: 3): https://sciencetostartup.com/paper/dependence-fidelity-and-downstream-inference-stability-in-generative-models - Introducing a new criterion for evaluating generative models based on covariance-level dependence fidelity.
- HopChain: Multi-Hop Data Synthesis for Generalizable Vision-Language Reasoning (viability: 7): https://sciencetostartup.com/paper/hopchain-multi-hop-data-synthesis-for-generalizable-vision-language-reasoning - HopChain is a scalable framework for synthesizing multi-hop vision-language reasoning data to enhance the training of vision-language models.
- Contingency-Aware Planning via Certified Neural Hamilton-Jacobi Reachability (viability: 5): https://sciencetostartup.com/paper/contingency-aware-planning-via-certified-neural-hamilton-jacobi-reachability - A framework for contingency-aware multi-goal navigation that integrates learning-based reachability with sampling-based planning.
- Generative AI-assisted Participatory Modeling in Socio-Environmental Planning under Deep Uncertainty (viability: 6): https://sciencetostartup.com/paper/generative-ai-assisted-participatory-modeling-in-socio-environmental-planning-under-deep-uncertainty - A templated workflow leveraging large language models to streamline participatory modeling in socio-environmental planning.
- Transformers Can Learn Rules They've Never Seen: Proof of Computation Beyond Interpolation (viability: 2): https://sciencetostartup.com/paper/transformers-can-learn-rules-they-ve-never-seen-proof-of-computation-beyond-interpolation - This paper explores whether transformers can learn unseen rules, providing theoretical insights but lacking practical application.
- LLM NL2SQL Robustness: Surface Noise vs. Linguistic Variation in Traditional and Agentic Settings (viability: 4): https://sciencetostartup.com/paper/llm-nl2sql-robustness-surface-noise-vs-linguistic-variation-in-traditional-and-agentic-settings - A robustness evaluation benchmark for NL2SQL systems addressing real-world noise and linguistic variation.
- Efficient and Reliable Teleoperation through Real-to-Sim-to-Real Shared Autonomy (viability: 7): https://sciencetostartup.com/paper/efficient-and-reliable-teleoperation-through-real-to-sim-to-real-shared-autonomy - A shared autonomy framework that enhances teleoperation by combining human intent with automated assistance for improved manipulation tasks.
- WorldCam: Interactive Autoregressive 3D Gaming Worlds with Camera Pose as a Unifying Geometric Representation (viability: 7): https://sciencetostartup.com/paper/worldcam-interactive-autoregressive-3d-gaming-worlds-with-camera-pose-as-a-unifying-geometric-representation - WorldCam revolutionizes interactive gaming by using camera pose for precise action control and long-term 3D consistency.
- Demystifing Video Reasoning (viability: 4): https://sciencetostartup.com/paper/demystifing-video-reasoning - A novel approach to understanding and enhancing reasoning in video generation models through emergent behaviors.
- Efficient Reasoning on the Edge (viability: 8): https://sciencetostartup.com/paper/efficient-reasoning-on-the-edge - A lightweight approach to enable efficient reasoning in small LLMs for mobile devices using LoRA adapters and reinforcement learning.
- SegviGen: Repurposing 3D Generative Model for Part Segmentation (viability: 8): https://sciencetostartup.com/paper/segvigen-repurposing-3d-generative-model-for-part-segmentation - SegviGen repurposes 3D generative models for efficient part segmentation with minimal training data.
- MessyKitchens: Contact-rich object-level 3D scene reconstruction (viability: 8): https://sciencetostartup.com/paper/messykitchens-contact-rich-object-level-3d-scene-reconstruction - MessyKitchens offers a novel dataset and advanced methods for accurate 3D scene reconstruction in cluttered environments.
- ManiTwin: Scaling Data-Generation-Ready Digital Object Dataset to 100K (viability: 7): https://sciencetostartup.com/paper/manitwin-scaling-data-generation-ready-digital-object-dataset-to-100k - ManiTwin automates the generation of 3D digital assets for scalable robotic manipulation data.
- SparkVSR: Interactive Video Super-Resolution via Sparse Keyframe Propagation (viability: 8): https://sciencetostartup.com/paper/sparkvsr-interactive-video-super-resolution-via-sparse-keyframe-propagation - SparkVSR offers an interactive framework for video super-resolution that allows users to control output quality through keyframe selection.
- Chronos: Temporal-Aware Conversational Agents with Structured Event Retrieval for Long-Term Memory (viability: 7): https://sciencetostartup.com/paper/chronos-temporal-aware-conversational-agents-with-structured-event-retrieval-for-long-term-memory - Chronos enhances conversational AI with structured temporal memory for improved long-term interaction.
- MolmoB0T: Large-Scale Simulation Enables Zero-Shot Manipulation (viability: 8): https://sciencetostartup.com/paper/molmob0t-large-scale-simulation-enables-zero-shot-manipulation - MolmoBot enables effective zero-shot manipulation in robotics using large-scale simulated data.
- DreamPlan: Efficient Reinforcement Fine-Tuning of Vision-Language Planners via Video World Models (viability: 8): https://sciencetostartup.com/paper/dreamplan-efficient-reinforcement-fine-tuning-of-vision-language-planners-via-video-world-models - DreamPlan enhances Vision-Language Models for robotic manipulation through efficient reinforcement fine-tuning using video world models.
- SocialOmni: Benchmarking Audio-Visual Social Interactivity in Omni Models (viability: 7): https://sciencetostartup.com/paper/socialomni-benchmarking-audio-visual-social-interactivity-in-omni-models - SocialOmni is a benchmark for evaluating audio-visual social interactivity in omni-modal large language models.
- SOMA: Unifying Parametric Human Body Models (viability: 7): https://sciencetostartup.com/paper/soma-unifying-parametric-human-body-models - SOMA unifies diverse parametric human body models for seamless reconstruction and animation.
- Long-Horizon Traffic Forecasting via Incident-Aware Conformal Spatio-Temporal Transformers (viability: 5): https://sciencetostartup.com/paper/long-horizon-traffic-forecasting-via-incident-aware-conformal-spatio-temporal-transformers - A novel traffic forecasting model that incorporates incident-aware spatio-temporal dynamics for improved accuracy.
- Online Experiential Learning for Language Models (viability: 3): https://sciencetostartup.com/paper/online-experiential-learning-for-language-models - A framework for language models to improve continuously from real-world deployment experiences.
- BrickSim: A Physics-Based Simulator for Manipulating Interlocking Brick Assemblies (viability: 8): https://sciencetostartup.com/paper/bricksim-a-physics-based-simulator-for-manipulating-interlocking-brick-assemblies - BRICKSIM offers a real-time simulator for realistic robotic manipulation of interlocking brick assemblies, integrating effortlessly with robotic workflows.
- GIST: Gauge-Invariant Spectral Transformers for Scalable Graph Neural Operators (viability: 8): https://sciencetostartup.com/paper/gist-gauge-invariant-spectral-transformers-for-scalable-graph-neural-operators - GIST is a novel graph transformer architecture that achieves scalable, gauge-invariant learning for graph-structured data.
- Mediocrity is the key for LLM as a Judge Anchor Selection (viability: 4): https://sciencetostartup.com/paper/mediocrity-is-the-key-for-llm-as-a-judge-anchor-selection - This research identifies critical anchor selection methods to enhance the reliability of LLM evaluations.
- Dynamic Meta-Layer Aggregation for Byzantine-Robust Federated Learning (viability: 7): https://sciencetostartup.com/paper/dynamic-meta-layer-aggregation-for-byzantine-robust-federated-learning - FedAOT is a novel defense mechanism for Byzantine-robust federated learning that dynamically weights client updates to enhance model resilience against adversarial attacks.
- M^3: Dense Matching Meets Multi-View Foundation Models for Monocular Gaussian Splatting SLAM (viability: 7): https://sciencetostartup.com/paper/m-3-dense-matching-meets-multi-view-foundation-models-for-monocular-gaussian-splatting-slam - M^3 enhances monocular SLAM with precise pose estimation and dynamic area suppression for superior scene reconstruction.
- Internalizing Agency from Reflective Experience (viability: 3): https://sciencetostartup.com/paper/internalizing-agency-from-reflective-experience - LEAFE enhances long-horizon agency in autonomous agents by internalizing recovery from reflective experiences.
- Stochastic Resetting Accelerates Policy Convergence in Reinforcement Learning (viability: 2): https://sciencetostartup.com/paper/stochastic-resetting-accelerates-policy-convergence-in-reinforcement-learning - This paper explores stochastic resetting as a mechanism to accelerate policy convergence in reinforcement learning.
- What DINO saw: ALiBi positional encoding reduces positional bias in Vision Transformers (viability: 5): https://sciencetostartup.com/paper/what-dino-saw-alibi-positional-encoding-reduces-positional-bias-in-vision-transformers - This research addresses positional bias in Vision Transformers, enhancing their applicability in material science imaging.
- Learning to Present: Inverse Specification Rewards for Agentic Slide Generation (viability: 8): https://sciencetostartup.com/paper/learning-to-present-inverse-specification-rewards-for-agentic-slide-generation - Automate professional slide deck creation using LLMs and a novel inverse specification reward system.
- An assessment of data-centric methods for label noise identification in remote sensing data sets (viability: 4): https://sciencetostartup.com/paper/an-assessment-of-data-centric-methods-for-label-noise-identification-in-remote-sensing-data-sets - A systematic analysis of data-centric methods for identifying and isolating label noise in remote sensing datasets.
- Conditional Distributional Treatment Effects: Doubly Robust Estimation and Testing (viability: 2): https://sciencetostartup.com/paper/conditional-distributional-treatment-effects-doubly-robust-estimation-and-testing - A novel method for estimating and testing conditional distributional treatment effects in statistical analysis.
- Prompt Programming for Cultural Bias and Alignment of Large Language Models (viability: 3): https://sciencetostartup.com/paper/prompt-programming-for-cultural-bias-and-alignment-of-large-language-models - This research explores prompt programming to enhance cultural alignment in large language models.
- Real-Time Decoding of Movement Onset and Offset for Brain-Controlled Rehabilitation Exoskeleton (viability: 6): https://sciencetostartup.com/paper/real-time-decoding-of-movement-onset-and-offset-for-brain-controlled-rehabilitation-exoskeleton - A brain-controlled exoskeleton that enables precise start-stop movements for rehabilitation therapy using EEG signals.
- Deep Reinforcement Learning-driven Edge Offloading for Latency-constrained XR pipelines (viability: 7): https://sciencetostartup.com/paper/deep-reinforcement-learning-driven-edge-offloading-for-latency-constrained-xr-pipelines - A deep reinforcement learning framework for optimizing edge offloading in latency-sensitive XR applications.
- Surg$Σ$: A Spectrum of Large-Scale Multimodal Data and Foundation Models for Surgical Intelligence (viability: 7): https://sciencetostartup.com/paper/surg-a-spectrum-of-large-scale-multimodal-data-and-foundation-models-for-surgical-intelligence - Surg$Σ$ offers a comprehensive multimodal data foundation for enhancing surgical intelligence across diverse clinical tasks.
- Is Conformal Factuality for RAG-based LLMs Robust? Novel Metrics and Systematic Insights (viability: 5): https://sciencetostartup.com/paper/is-conformal-factuality-for-rag-based-llms-robust-novel-metrics-and-systematic-insights - This research proposes novel metrics for improving the reliability of RAG-based LLMs through conformal factuality filtering.
- WildDepth: A Multimodal Dataset for 3D Wildlife Perception and Depth Estimation (viability: 4): https://sciencetostartup.com/paper/wilddepth-a-multimodal-dataset-for-3d-wildlife-perception-and-depth-estimation - WildDepth is a multimodal dataset designed to enhance depth estimation and 3D reconstruction for wildlife perception.
- Beyond Accuracy: Evaluating Forecasting Models by Multi-Echelon Inventory Cost (viability: 5): https://sciencetostartup.com/paper/beyond-accuracy-evaluating-forecasting-models-by-multi-echelon-inventory-cost - A digitalized pipeline for optimizing inventory forecasting and costs in supply chains.
- ODIN-Based CPU-GPU Architecture with Replay-Driven Simulation and Emulation (viability: 4): https://sciencetostartup.com/paper/odin-based-cpu-gpu-architecture-with-replay-driven-simulation-and-emulation - A replay-driven validation methodology for CPU-GPU integration in chiplet architectures.
- CABTO: Context-Aware Behavior Tree Grounding for Robot Manipulation (viability: 7): https://sciencetostartup.com/paper/cabto-context-aware-behavior-tree-grounding-for-robot-manipulation - CABTO automates the construction of reliable behavior tree systems for robot manipulation using large models and contextual feedback.
- DexGrasp-Zero: A Morphology-Aligned Policy for Zero-Shot Cross-Embodiment Dexterous Grasping (viability: 8): https://sciencetostartup.com/paper/dexgrasp-zero-a-morphology-aligned-policy-for-zero-shot-cross-embodiment-dexterous-grasping - DexGrasp-Zero enables zero-shot dexterous grasping across diverse robotic hands using a novel morphology-aligned policy.
- Development of Low-Cost and Bidirectional Syringe Pumps for Soft Robotics Applications (viability: 3): https://sciencetostartup.com/paper/development-of-low-cost-and-bidirectional-syringe-pumps-for-soft-robotics-applications - A low-cost, modular syringe pump system designed to enhance soft robotics applications.
- RaDAR: Relation-aware Diffusion-Asymmetric Graph Contrastive Learning for Recommendation (viability: 7): https://sciencetostartup.com/paper/radar-relation-aware-diffusion-asymmetric-graph-contrastive-learning-for-recommendation - RaDAR enhances recommendation systems by addressing data sparsity and noise through innovative graph contrastive learning techniques.
- High-Dimensional Gaussian Mean Estimation under Realizable Contamination (viability: 2): https://sciencetostartup.com/paper/high-dimensional-gaussian-mean-estimation-under-realizable-contamination - This paper explores the complexities of Gaussian mean estimation under a contamination model, revealing trade-offs in sample size and runtime.
- Adaptive Moments are Surprisingly Effective for Plug-and-Play Diffusion Sampling (viability: 7): https://sciencetostartup.com/paper/adaptive-moments-are-surprisingly-effective-for-plug-and-play-diffusion-sampling - A novel approach using adaptive moment estimation to enhance guided diffusion sampling for image restoration and generation.
- V-Co: A Closer Look at Visual Representation Alignment via Co-Denoising (viability: 7): https://sciencetostartup.com/paper/v-co-a-closer-look-at-visual-representation-alignment-via-co-denoising - V-Co enhances visual representation alignment in generative models through effective co-denoising techniques.
- InCoder-32B: Code Foundation Model for Industrial Scenarios (viability: 8): https://sciencetostartup.com/paper/incoder-32b-code-foundation-model-for-industrial-scenarios - InCoder-32B is a specialized code foundation model designed to enhance programming tasks in industrial scenarios.
- Conservative Continuous-Time Treatment Optimization (viability: 4): https://sciencetostartup.com/paper/conservative-continuous-time-treatment-optimization - A framework for optimizing treatment plans using conservative stochastic control based on patient trajectory data.
- SpokenUS: A Spoken User Simulator for Task-Oriented Dialogue (viability: 7): https://sciencetostartup.com/paper/spokenus-a-spoken-user-simulator-for-task-oriented-dialogue - SpokenUS is a spoken user simulator designed to enhance task-oriented dialogue agents with realistic user behaviors.
- IOSVLM: A 3D Vision-Language Model for Unified Dental Diagnosis from Intraoral Scans (viability: 7): https://sciencetostartup.com/paper/iosvlm-a-3d-vision-language-model-for-unified-dental-diagnosis-from-intraoral-scans - IOSVLM is a 3D vision-language model that enhances dental diagnosis using intraoral scans for improved clinical outcomes.
- Anticipatory Planning for Multimodal AI Agents (viability: 7): https://sciencetostartup.com/paper/anticipatory-planning-for-multimodal-ai-agents - TraceR1 enhances multimodal AI agents with anticipatory reasoning for improved planning and execution.
- Beyond Cybathlon: On-demand Quadrupedal Assistance for People with Limited Mobility (viability: 3): https://sciencetostartup.com/paper/beyond-cybathlon-on-demand-quadrupedal-assistance-for-people-with-limited-mobility - An on-demand quadrupedal assistance robot system designed to enhance the independence of individuals with limited mobility.
- SOMP: Scalable Gradient Inversion for Large Language Models via Subspace-Guided Orthogonal Matching Pursuit (viability: 3): https://sciencetostartup.com/paper/somp-scalable-gradient-inversion-for-large-language-models-via-subspace-guided-orthogonal-matching-pursuit - SOMP is a scalable framework for recovering private training text from aggregated gradients in large language models.
- Dual Stream Independence Decoupling for True Emotion Recognition under Masked Expressions (viability: 4): https://sciencetostartup.com/paper/dual-stream-independence-decoupling-for-true-emotion-recognition-under-masked-expressions - A novel framework for recognizing true emotions from masked expressions using apexframe classification.
- TurnWise: The Gap between Single- and Multi-turn Language Model Capabilities (viability: 6): https://sciencetostartup.com/paper/turnwise-the-gap-between-single-and-multi-turn-language-model-capabilities - TurnWise introduces a benchmark and data pipeline to enhance multi-turn conversation capabilities in language models.
- SuCor: Susceptibility Distortion Correction via Parameter-Free and Self-Regularized Optimal Transport (viability: 4): https://sciencetostartup.com/paper/sucor-susceptibility-distortion-correction-via-parameter-free-and-self-regularized-optimal-transport - SuCor offers a novel method for correcting geometric distortions in EPI imaging using optimal transport.
- pADAM: A Plug-and-Play All-in-One Diffusion Architecture for Multi-Physics Learning (viability: 5): https://sciencetostartup.com/paper/padam-a-plug-and-play-all-in-one-diffusion-architecture-for-multi-physics-learning - pADAM is a unified generative framework for multi-physics learning that enables accurate inference and uncertainty quantification across diverse physical laws.
- A Practical Algorithm for Feature-Rich, Non-Stationary Bandit Problems (viability: 7): https://sciencetostartup.com/paper/a-practical-algorithm-for-feature-rich-non-stationary-bandit-problems - A novel algorithm for contextual bandits that adapts to non-stationary environments, enhancing recommendation systems.
- Finding Common Ground in a Sea of Alternatives (viability: 4): https://sciencetostartup.com/paper/finding-common-ground-in-a-sea-of-alternatives - A generative AI model that selects statements reflecting common ground across diverse population preferences using a novel sampling-based algorithm.
- Thermopneumatic Pixels for Fast, Localized, Low-Voltage Touch Feedback (viability: 7): https://sciencetostartup.com/paper/thermopneumatic-pixels-for-fast-localized-low-voltage-touch-feedback - Thermopneumatic pixels provide low-voltage, rapid tactile feedback for interactive devices.
- Probing Cultural Signals in Large Language Models through Author Profiling (viability: 7): https://sciencetostartup.com/paper/probing-cultural-signals-in-large-language-models-through-author-profiling - A tool for probing cultural biases in large language models through author profiling from song lyrics.
- Semi-supervised Latent Disentangled Diffusion Model for Textile Pattern Generation (viability: 7): https://sciencetostartup.com/paper/semi-supervised-latent-disentangled-diffusion-model-for-textile-pattern-generation - A novel method for generating high-fidelity textile patterns from clothing images using a semi-supervised latent diffusion model.
- Persistent Device Identity for Network Access Control in the Era of MAC Address Randomization: A RADIUS-Based Framework (viability: 3): https://sciencetostartup.com/paper/persistent-device-identity-for-network-access-control-in-the-era-of-mac-address-randomization-a-radius-based-framework - A RADIUS-based framework for maintaining persistent device identity in network access control amidst MAC address randomization.
- Nonstandard Errors in AI Agents (viability: 2): https://sciencetostartup.com/paper/nonstandard-errors-in-ai-agents - This research investigates the variability in results produced by AI coding agents in empirical research.
- When the City Teaches the Car: Label-Free 3D Perception from Infrastructure (viability: 7): https://sciencetostartup.com/paper/when-the-city-teaches-the-car-label-free-3d-perception-from-infrastructure - A novel label-free 3D perception system for self-driving cars leveraging city infrastructure as unsupervised teachers.
- Bayesian Inference of Psychometric Variables From Brain and Behavior in Implicit Association Tests (viability: 4): https://sciencetostartup.com/paper/bayesian-inference-of-psychometric-variables-from-brain-and-behavior-in-implicit-association-tests - A Bayesian model for predicting mental health symptoms from neural and behavioral data using Implicit Association Tests.
- SpecMoE: Spectral Mixture-of-Experts Foundation Model for Cross-Species EEG Decoding (viability: 7): https://sciencetostartup.com/paper/specmoe-spectral-mixture-of-experts-foundation-model-for-cross-species-eeg-decoding - A foundation model for advanced EEG decoding using a novel Gaussian-smoothed masking scheme.
- MedCL-Bench: Benchmarking stability-efficiency trade-offs and scaling in biomedical continual learning (viability: 4): https://sciencetostartup.com/paper/medcl-bench-benchmarking-stability-efficiency-trade-offs-and-scaling-in-biomedical-continual-learning - MedCL-Bench offers a standardized benchmark for evaluating continual learning in biomedical NLP models to prevent catastrophic forgetting.
- Retrieving Counterfactuals Improves Visual In-Context Learning (viability: 8): https://sciencetostartup.com/paper/retrieving-counterfactuals-improves-visual-in-context-learning - CIRCLES enhances vision-language models by using counterfactual examples for improved in-context learning and causal reasoning.
- World Reconstruction From Inconsistent Views (viability: 4): https://sciencetostartup.com/paper/world-reconstruction-from-inconsistent-views - A method for generating 3D-consistent worlds from video frames using non-rigid alignment techniques.
- Ember: A Serverless Peer-to-Peer End-to-End Encrypted Messaging System over an IPv6 Mesh Network (viability: 5): https://sciencetostartup.com/paper/ember-a-serverless-peer-to-peer-end-to-end-encrypted-messaging-system-over-an-ipv6-mesh-network - Ember is a serverless, peer-to-peer encrypted messaging system designed for decentralized communication over IPv6 mesh networks.
- Differential Harm Propensity in Personalized LLM Agents: The Curious Case of Mental Health Disclosure (viability: 7): https://sciencetostartup.com/paper/differential-harm-propensity-in-personalized-llm-agents-the-curious-case-of-mental-health-disclosure - A study on how mental health disclosure impacts the safety of personalized LLM agents in task completion.
- IQuest-Coder-V1 Technical Report (viability: 3): https://sciencetostartup.com/paper/iquest-coder-v1-technical-report - IQuest-Coder-V1 introduces a new family of code LLMs with a multi-stage training paradigm for enhanced code intelligence.
- Understanding Quantization of Optimizer States in LLM Pre-training: Dynamics of State Staleness and Effectiveness of State Resets (viability: 2): https://sciencetostartup.com/paper/understanding-quantization-of-optimizer-states-in-llm-pre-training-dynamics-of-state-staleness-and-effectiveness-of-stat - This paper explores the effects of quantizing optimizer states in LLM pre-training and proposes a method for effective state resets.
- GeMA: Learning Latent Manifold Frontiers for Benchmarking Complex Systems (viability: 4): https://sciencetostartup.com/paper/gema-learning-latent-manifold-frontiers-for-benchmarking-complex-systems - GeMA offers a novel framework for benchmarking complex systems using a latent manifold approach.
- The Cost of Reasoning: Chain-of-Thought Induces Overconfidence in Vision-Language Models (viability: 2): https://sciencetostartup.com/paper/the-cost-of-reasoning-chain-of-thought-induces-overconfidence-in-vision-language-models - This paper explores how chain-of-thought reasoning affects uncertainty quantification in vision-language models.
- Federated Learning with Multi-Partner OneFlorida+ Consortium Data for Predicting Major Postoperative Complications (viability: 7): https://sciencetostartup.com/paper/federated-learning-with-multi-partner-oneflorida-consortium-data-for-predicting-major-postoperative-complications - Federated learning models for predicting major postoperative complications using multicenter data while preserving patient privacy.
- Emotion-Aware Classroom Quality Assessment Leveraging IoT-Based Real-Time Student Monitoring (viability: 7): https://sciencetostartup.com/paper/emotion-aware-classroom-quality-assessment-leveraging-iot-based-real-time-student-monitoring - An IoT-based framework for real-time monitoring of student emotions to enhance classroom engagement.
- Arabic Morphosyntactic Tagging and Dependency Parsing with Large Language Models (viability: 4): https://sciencetostartup.com/paper/arabic-morphosyntactic-tagging-and-dependency-parsing-with-large-language-models - Leveraging large language models for advanced Arabic morphosyntactic tagging and dependency parsing.
- Novelty-Driven Target-Space Discovery in Automated Electron and Scanning Probe Microscopy (viability: 6): https://sciencetostartup.com/paper/novelty-driven-target-space-discovery-in-automated-electron-and-scanning-probe-microscopy - A framework for automated microscopy that actively discovers new behaviors in target spaces using deep-kernel learning.
- Search2Motion: Training-Free Object-Level Motion Control via Attention-Consensus Search (viability: 7): https://sciencetostartup.com/paper/search2motion-training-free-object-level-motion-control-via-attention-consensus-search - Search2Motion offers a training-free solution for precise object-level motion control in video generation.
- Learning Lineage-guided Geodesics with Finsler Geometry (viability: 2): https://sciencetostartup.com/paper/learning-lineage-guided-geodesics-with-finsler-geometry - This paper presents a novel Finsler metric for trajectory inference that integrates geometric and classification priors.
- Cost Trade-offs in Matrix Inversion Updates for Streaming Outlier Detection (viability: 2): https://sciencetostartup.com/paper/cost-trade-offs-in-matrix-inversion-updates-for-streaming-outlier-detection - This paper explores optimal methods for matrix inversion updates in online outlier detection.
- SynthChain: A Synthetic Benchmark and Forensic Analysis of Advanced and Stealthy Software Supply Chain Attacks (viability: 4): https://sciencetostartup.com/paper/synthchain-a-synthetic-benchmark-and-forensic-analysis-of-advanced-and-stealthy-software-supply-chain-attacks - SynthChain provides a synthetic benchmark and dataset for analyzing software supply chain attacks, enhancing detection capabilities.
- Grid-World Representations in Transformers Reflect Predictive Geometry (viability: 4): https://sciencetostartup.com/paper/grid-world-representations-in-transformers-reflect-predictive-geometry - This research explores how transformer models develop geometric representations for optimal prediction in constrained random walks.
- vAccSOL: Efficient and Transparent AI Vision Offloading for Mobile Robots (viability: 4): https://sciencetostartup.com/paper/vaccsol-efficient-and-transparent-ai-vision-offloading-for-mobile-robots - vAccSOL optimizes AI vision workloads for mobile robots, enhancing performance and reducing power consumption.
- Learning Whole-Body Control for a Salamander Robot (viability: 3): https://sciencetostartup.com/paper/learning-whole-body-control-for-a-salamander-robot - A learning-based controller for salamander robots that enables stable locomotion in complex environments.
- HMAR: Hierarchical Modality-Aware Expert and Dynamic Routing Medical Image Retrieval Architecture (viability: 6): https://sciencetostartup.com/paper/hmar-hierarchical-modality-aware-expert-and-dynamic-routing-medical-image-retrieval-architecture - HMAR is an adaptive medical image retrieval framework that enhances clinical diagnosis through precise lesion-region retrieval.
- When Should a Robot Think? Resource-Aware Reasoning via Reinforcement Learning for Embodied Robotic Decision-Making (viability: 7): https://sciencetostartup.com/paper/when-should-a-robot-think-resource-aware-reasoning-via-reinforcement-learning-for-embodied-robotic-decision-making - RARRL optimizes embodied robotic decision-making by adaptively managing reasoning and action execution to enhance efficiency and reliability.
- CritiSense: Critical Digital Literacy and Resilience Against Misinformation (viability: 8): https://sciencetostartup.com/paper/critisense-critical-digital-literacy-and-resilience-against-misinformation - CritiSense boosts digital literacy to combat misinformation through a multilingual mobile app with interactive challenges.
- $x^2$-Fusion: Cross-Modality and Cross-Dimension Flow Estimation in Event Edge Space (viability: 7): https://sciencetostartup.com/paper/x-2-fusion-cross-modality-and-cross-dimension-flow-estimation-in-event-edge-space - $x^2$-Fusion unifies multimodal data for superior 2D and 3D motion estimation.
- Kinema4D: Kinematic 4D World Modeling for Spatiotemporal Embodied Simulation (viability: 7): https://sciencetostartup.com/paper/kinema4d-kinematic-4d-world-modeling-for-spatiotemporal-embodied-simulation - Kinema4D is a 4D generative robotic simulator that enhances robot-world interaction modeling for embodied AI.
- Fast-WAM: Do World Action Models Need Test-time Future Imagination? (viability: 8): https://sciencetostartup.com/paper/fast-wam-do-world-action-models-need-test-time-future-imagination - Fast-WAM optimizes embodied control by eliminating test-time future imagination while maintaining competitive performance.
- Kestrel: Grounding Self-Refinement for LVLM Hallucination Mitigation (viability: 7): https://sciencetostartup.com/paper/kestrel-grounding-self-refinement-for-lvlm-hallucination-mitigation - Kestrel is a training-free framework that mitigates hallucinations in large vision-language models through visual grounding and self-refinement.
- When Openclaw Agents Learn from Each Other: Insights from Emergent AI Agent Communities for Human-AI Partnership in Education (viability: 2): https://sciencetostartup.com/paper/when-openclaw-agents-learn-from-each-other-insights-from-emergent-ai-agent-communities-for-human-ai-partnership-in-educa - Exploring emergent learning behaviors in AI agent communities to enhance human-AI partnerships in education.
- Spectral Property-Driven Data Augmentation for Hyperspectral Single-Source Domain Generalization (viability: 7): https://sciencetostartup.com/paper/spectral-property-driven-data-augmentation-for-hyperspectral-single-source-domain-generalization - A novel spectral property-driven data augmentation technique for enhancing hyperspectral image classification robustness.
- Self-Aware Markov Models for Discrete Reasoning (viability: 6): https://sciencetostartup.com/paper/self-aware-markov-models-for-discrete-reasoning - A novel approach to enhance reasoning in masked discrete diffusion models by enabling self-correction through a learned Markov transition kernel.
- Can Linguistically Related Languages Guide LLM Translation in Low-Resource Settings? (viability: 5): https://sciencetostartup.com/paper/can-linguistically-related-languages-guide-llm-translation-in-low-resource-settings - This research explores using linguistically related languages to enhance LLM translation in low-resource settings without extensive fine-tuning.
- Machines acquire scientific taste from institutional traces (viability: 8): https://sciencetostartup.com/paper/machines-acquire-scientific-taste-from-institutional-traces - A fine-tuned language model that automates the evaluation of research pitches, enhancing decision-making in scientific publishing.
- Omanic: Towards Step-wise Evaluation of Multi-hop Reasoning in Large Language Models (viability: 8): https://sciencetostartup.com/paper/omanic-towards-step-wise-evaluation-of-multi-hop-reasoning-in-large-language-models - Omanic provides a structured approach to evaluate multi-hop reasoning in large language models through detailed annotations and a challenging benchmark.
- HeBA: Heterogeneous Bottleneck Adapters for Robust Vision-Language Models (viability: 8): https://sciencetostartup.com/paper/heba-heterogeneous-bottleneck-adapters-for-robust-vision-language-models - HeBA adapts Vision-Language Models efficiently with innovative architectural biases for enhanced downstream task performance.
- Efficient Brood Cell Detection in Layer Trap Nests for Bees and Wasps: Balancing Labeling Effort and Species Coverage (viability: 6): https://sciencetostartup.com/paper/efficient-brood-cell-detection-in-layer-trap-nests-for-bees-and-wasps-balancing-labeling-effort-and-species-coverage - A deep learning approach to automate brood cell detection in layer trap nests, reducing manual labeling effort and improving species classification.
- What if Pinocchio Were a Reinforcement Learning Agent: A Normative End-to-End Pipeline (viability: 2): https://sciencetostartup.com/paper/what-if-pinocchio-were-a-reinforcement-learning-agent-a-normative-end-to-end-pipeline - A proposed pipeline for developing norm-compliant reinforcement learning agents using argumentation-based supervision.
- Mixture of Style Experts for Diverse Image Stylization (viability: 7): https://sciencetostartup.com/paper/mixture-of-style-experts-for-diverse-image-stylization - StyleExpert is a semantic-aware framework for diverse image stylization using a Mixture of Experts architecture.
- Domain-Independent Dynamic Programming with Constraint Propagation (viability: 2): https://sciencetostartup.com/paper/domain-independent-dynamic-programming-with-constraint-propagation - This paper proposes a novel integration of constraint propagation into dynamic programming for improved combinatorial problem solving.
- BUSSARD: Normalizing Flows for Bijective Universal Scene-Specific Anomalous Relationship Detection (viability: 8): https://sciencetostartup.com/paper/bussard-normalizing-flows-for-bijective-universal-scene-specific-anomalous-relationship-detection - BUSSARD leverages normalizing flows for efficient and robust anomaly detection in scene graphs.
- Good Arguments Against the People Pleasers: How Reasoning Mitigates (Yet Masks) LLM Sycophancy (viability: 2): https://sciencetostartup.com/paper/good-arguments-against-the-people-pleasers-how-reasoning-mitigates-yet-masks-llm-sycophancy - This paper explores how reasoning in LLMs can both mitigate and mask sycophancy.
- When AI Navigates the Fog of War (viability: 2): https://sciencetostartup.com/paper/when-ai-navigates-the-fog-of-war - This paper explores AI's reasoning capabilities in analyzing unfolding geopolitical conflicts.
- FlowComposer: Composable Flows for Compositional Zero-Shot Learning (viability: 7): https://sciencetostartup.com/paper/flowcomposer-composable-flows-for-compositional-zero-shot-learning - FlowComposer enhances compositional zero-shot learning by explicitly fusing visual features with text embeddings for improved generalization.
- Reconciling distributed compliance with high-performance control in continuum soft robotics (viability: 7): https://sciencetostartup.com/paper/reconciling-distributed-compliance-with-high-performance-control-in-continuum-soft-robotics - A revolutionary soft robotic arm that achieves high-performance control without sacrificing compliance.
- MLLM-based Textual Explanations for Face Comparison (viability: 7): https://sciencetostartup.com/paper/mllm-based-textual-explanations-for-face-comparison - A framework for generating reliable textual explanations for face recognition decisions using MLLMs.
- Routing and Control for Marine Oil-Spill Cleanup with a Boom-Towing Vessel Fleet (viability: 6): https://sciencetostartup.com/paper/routing-and-control-for-marine-oil-spill-cleanup-with-a-boom-towing-vessel-fleet - A multi-robot framework for coordinated oil-spill cleanup using autonomous surface vehicles.
- Domain Mixture Design via Log-Likelihood Differences for Aligning Language Models with a Target Model (viability: 2): https://sciencetostartup.com/paper/domain-mixture-design-via-log-likelihood-differences-for-aligning-language-models-with-a-target-model - This paper proposes a method for aligning language models with target models through domain mixture design.
- Simplex-to-Euclidean Bijection for Conjugate and Calibrated Multiclass Gaussian Process (viability: 5): https://sciencetostartup.com/paper/simplex-to-euclidean-bijection-for-conjugate-and-calibrated-multiclass-gaussian-process - A novel Gaussian process model that enhances multi-class classification by utilizing Aitchison geometry for improved predictive probabilities.
- TCATSeg: A Tooth Center-Wise Attention Network for 3D Dental Model Semantic Segmentation (viability: 7): https://sciencetostartup.com/paper/tcatseg-a-tooth-center-wise-attention-network-for-3d-dental-model-semantic-segmentation - TCATSeg is a novel framework for accurate semantic segmentation of 3D dental models, enhancing digital dentistry applications.
- ACPV-Net: All-Class Polygonal Vectorization for Seamless Vector Map Generation from Aerial Imagery (viability: 9): https://sciencetostartup.com/paper/acpv-net-all-class-polygonal-vectorization-for-seamless-vector-map-generation-from-aerial-imagery - Transform aerial imagery into seamless, topology-consistent vector maps with ACPV-Net.
- Dexterous grasp data augmentation based on grasp synthesis with fingertip workspace cloud and contact-aware sampling (viability: 7): https://sciencetostartup.com/paper/dexterous-grasp-data-augmentation-based-on-grasp-synthesis-with-fingertip-workspace-cloud-and-contact-aware-sampling - A teleoperation-based framework for efficient data augmentation in robotic grasping using fingertip contact-aware sampling.
- Omnilingual SONAR: Cross-Lingual and Cross-Modal Sentence Embeddings Bridging Massively Multilingual Text and Speech (viability: 9): https://sciencetostartup.com/paper/omnilingual-sonar-cross-lingual-and-cross-modal-sentence-embeddings-bridging-massively-multilingual-text-and-speech - OmniSONAR offers an unprecedented omnilingual cross-modal embedding solution for multilingual translation and search applications.
- Tarab: A Multi-Dialect Corpus of Arabic Lyrics and Poetry (viability: 5): https://sciencetostartup.com/paper/tarab-a-multi-dialect-corpus-of-arabic-lyrics-and-poetry - Tarab is the largest open Arabic corpus of song lyrics and poetry, enabling comprehensive linguistic and cultural analysis.
- Rationale Matters: Learning Transferable Rubrics via Proxy-Guided Critique for VLMReward Models (viability: 7): https://sciencetostartup.com/paper/rationale-matters-learning-transferable-rubrics-via-proxy-guided-critique-for-vlmreward-models - Proxy-GRM enhances reward modeling by generating transferable rubrics verified through proxy agents, reducing training data needs and maintaining performance.
- FSMC-Pose: Frequency and Spatial Fusion with Multiscale Self-calibration for Cattle Mounting Pose Estimation (viability: 7): https://sciencetostartup.com/paper/fsmc-pose-frequency-and-spatial-fusion-with-multiscale-self-calibration-for-cattle-mounting-pose-estimation - FSMC-Pose revolutionizes cattle farm management by automating estrus detection using advanced pose estimation.
- Scalable Inspection Planning via Flow-based Mixed Integer Linear Programming (viability: 7): https://sciencetostartup.com/paper/scalable-inspection-planning-via-flow-based-mixed-integer-linear-programming - A scalable Mixed Integer Linear Programming solution for optimizing robot inspection paths in various applications.
- On the Transfer of Collinearity to Computer Vision (viability: 6): https://sciencetostartup.com/paper/on-the-transfer-of-collinearity-to-computer-vision - A novel computer vision approach leveraging the collinearity principle to enhance defect detection in industrial applications.
- BATQuant: Outlier-resilient MXFP4 Quantization via Learnable Block-wise Optimization (viability: 8): https://sciencetostartup.com/paper/batquant-outlier-resilient-mxfp4-quantization-via-learnable-block-wise-optimization - BATQuant optimizes quantization for multi-modal large language models, achieving state-of-the-art performance while minimizing outlier impact.
- Runtime Governance for AI Agents: Policies on Paths (viability: 2): https://sciencetostartup.com/paper/runtime-governance-for-ai-agents-policies-on-paths - A framework for runtime governance of AI agents to manage compliance and risk.
- Trajectory-Optimized Time Reparameterization for Learning-Compatible Reduced-Order Modeling of Stiff Dynamical Systems (viability: 5): https://sciencetostartup.com/paper/trajectory-optimized-time-reparameterization-for-learning-compatible-reduced-order-modeling-of-stiff-dynamical-systems - A novel trajectory-optimized time reparameterization method enhances the learnability of reduced-order models for stiff dynamical systems.
- V-DyKnow: A Dynamic Benchmark for Time-Sensitive Knowledge in Vision Language Models (viability: 7): https://sciencetostartup.com/paper/v-dyknow-a-dynamic-benchmark-for-time-sensitive-knowledge-in-vision-language-models - V-DyKnow is a benchmark for evaluating and improving time-sensitive knowledge in Vision-Language Models.
- When and Why Does Unsupervised RL Succeed in Mathematical Reasoning? A Manifold Envelopment Perspective (viability: 3): https://sciencetostartup.com/paper/when-and-why-does-unsupervised-rl-succeed-in-mathematical-reasoning-a-manifold-envelopment-perspective - This paper explores unsupervised reinforcement learning to enhance mathematical reasoning in large language models through intrinsic rewards.
- REFORGE: Multi-modal Attacks Reveal Vulnerable Concept Unlearning in Image Generation Models (viability: 8): https://sciencetostartup.com/paper/reforge-multi-modal-attacks-reveal-vulnerable-concept-unlearning-in-image-generation-models - REFORGE is a black-box red-teaming framework that enhances the robustness of image generation model unlearning against adversarial attacks.
- Diverging Transformer Predictions for Human Sentence Processing: A Comprehensive Analysis of Agreement Attraction Effects (viability: 2): https://sciencetostartup.com/paper/diverging-transformer-predictions-for-human-sentence-processing-a-comprehensive-analysis-of-agreement-attraction-effects - This paper critiques the cognitive adequacy of transformers in modeling human sentence processing.
- Malicious Or Not: Adding Repository Context to Agent Skill Classification (viability: 5): https://sciencetostartup.com/paper/malicious-or-not-adding-repository-context-to-agent-skill-classification - A comprehensive security analysis tool for AI agent skills that reduces false positives in malicious behavior classification.
- Face2Scene: Using Facial Degradation as an Oracle for Diffusion-Based Scene Restoration (viability: 7): https://sciencetostartup.com/paper/face2scene-using-facial-degradation-as-an-oracle-for-diffusion-based-scene-restoration - Face2Scene leverages facial degradation to enhance full-scene image restoration using a novel two-stage framework.
- Deep Tabular Representation Corrector (viability: 5): https://sciencetostartup.com/paper/deep-tabular-representation-corrector - A model-agnostic tool to enhance representations of deep tabular models without altering their parameters.
- Manifold-Matching Autoencoders (viability: 2): https://sciencetostartup.com/paper/manifold-matching-autoencoders - A novel unsupervised regularization scheme for autoencoders that aligns latent space distances with input data distances.
- Characterizing Delusional Spirals through Human-LLM Chat Logs (viability: 4): https://sciencetostartup.com/paper/characterizing-delusional-spirals-through-human-llm-chat-logs - Analyzing harmful interactions between users and LLM chatbots to mitigate psychological risks.
- VideoMatGen: PBR Materials through Joint Generative Modeling (viability: 3): https://sciencetostartup.com/paper/videomatgen-pbr-materials-through-joint-generative-modeling - A method for generating physically-based materials for 3D shapes using a video diffusion transformer.
- Understanding Cell Fate Decisions with Temporal Attention (viability: 8): https://sciencetostartup.com/paper/understanding-cell-fate-decisions-with-temporal-attention - A deep learning model predicts cancer cell fate from raw video data, enhancing treatment strategies with explainable insights.
- Segmentation-Based Attention Entropy: Detecting and Mitigating Object Hallucinations in Large Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/segmentation-based-attention-entropy-detecting-and-mitigating-object-hallucinations-in-large-vision-language-models - A method to detect and mitigate object hallucinations in large vision-language models using segmentation-based attention entropy.
- BenchPreS: A Benchmark for Context-Aware Personalized Preference Selectivity of Persistent-Memory LLMs (viability: 2): https://sciencetostartup.com/paper/benchpres-a-benchmark-for-context-aware-personalized-preference-selectivity-of-persistent-memory-llms - BenchPreS evaluates the context-aware application of user preferences in persistent-memory LLMs.
- EmoLLM: Appraisal-Grounded Cognitive-Emotional Co-Reasoning in Large Language Models (viability: 7): https://sciencetostartup.com/paper/emollm-appraisal-grounded-cognitive-emotional-co-reasoning-in-large-language-models - EmoLLM enhances dialogue by integrating emotional intelligence with cognitive reasoning for improved user interactions.
- CompDiff: Hierarchical Compositional Diffusion for Fair and Zero-Shot Intersectional Medical Image Generation (viability: 7): https://sciencetostartup.com/paper/compdiff-hierarchical-compositional-diffusion-for-fair-and-zero-shot-intersectional-medical-image-generation - CompDiff enhances fairness and zero-shot generalization in medical image synthesis, enabling high-quality intersectional demographics from limited data.
- ASCENT: Transformer-Based Aircraft Trajectory Prediction in Non-Towered Terminal Airspace (viability: 7): https://sciencetostartup.com/paper/ascent-transformer-based-aircraft-trajectory-prediction-in-non-towered-terminal-airspace - ASCENT is a lightweight transformer model for real-time aircraft trajectory prediction to enhance aviation safety.
- Bridging the Simulation-to-Reality Gap in Electron Microscope Calibration via VAE-EM Estimation (viability: 6): https://sciencetostartup.com/paper/bridging-the-simulation-to-reality-gap-in-electron-microscope-calibration-via-vae-em-estimation - A novel VAE-EM framework for automated calibration of electron microscopes, reducing estimation error and improving efficiency.
- SAMSEM -- A Generic and Scalable Approach for IC Metal Line Segmentation (viability: 7): https://sciencetostartup.com/paper/samsem-a-generic-and-scalable-approach-for-ic-metal-line-segmentation - SAMSEM is a robust tool for segmenting metal lines in SEM images to ensure the integrity of integrated circuits.
- DanceHA: A Multi-Agent Framework for Document-Level Aspect-Based Sentiment Analysis (viability: 5): https://sciencetostartup.com/paper/danceha-a-multi-agent-framework-for-document-level-aspect-based-sentiment-analysis - DanceHA is a multi-agent framework that enhances document-level aspect-based sentiment analysis through collaborative AI.
- How often do Answers Change? Estimating Recency Requirements in Question Answering (viability: 4): https://sciencetostartup.com/paper/how-often-do-answers-change-estimating-recency-requirements-in-question-answering - RecencyQA provides a dataset for improving question answering systems by categorizing questions based on how often their answers change.
- A Pin-Array Structured Climbing Robot for Stable Locomotion on Steep Rocky Terrain (viability: 4): https://sciencetostartup.com/paper/a-pin-array-structured-climbing-robot-for-stable-locomotion-on-steep-rocky-terrain - A climbing robot that uses compliant pin-array grippers for stable locomotion on steep and rocky terrain.
- Conservative Offline Robot Policy Learning via Posterior-Transition Reweighting (viability: 4): https://sciencetostartup.com/paper/conservative-offline-robot-policy-learning-via-posterior-transition-reweighting - A novel method for conservative offline robot policy learning that improves adaptation to heterogeneous datasets.
- Rethinking Pose Refinement in 3D Gaussian Splatting under Pose Prior and Geometric Uncertainty (viability: 4): https://sciencetostartup.com/paper/rethinking-pose-refinement-in-3d-gaussian-splatting-under-pose-prior-and-geometric-uncertainty - A novel relocalization framework that enhances pose refinement in 3D Gaussian Splatting by addressing pose and geometric uncertainties.
- Designing for Disagreement: Front-End Guardrails for Assistance Allocation in LLM-Enabled Robots (viability: 2): https://sciencetostartup.com/paper/designing-for-disagreement-front-end-guardrails-for-assistance-allocation-in-llm-enabled-robots - A framework for managing assistance allocation in LLM-enabled robots to address value disagreements.
- Kamino: GPU-based Massively Parallel Simulation of Multi-Body Systems with Challenging Topologies (viability: 6): https://sciencetostartup.com/paper/kamino-gpu-based-massively-parallel-simulation-of-multi-body-systems-with-challenging-topologies - Kamino is a GPU-based physics solver enabling high-throughput simulations of complex robotic systems with challenging topologies.
- SympFormer: Accelerated attention blocks via Inertial Dynamics on Density Manifolds (viability: 5): https://sciencetostartup.com/paper/sympformer-accelerated-attention-blocks-via-inertial-dynamics-on-density-manifolds - SympFormer introduces accelerated attention blocks for faster convergence in NLP tasks.
- LIMBERO: A Limbed Climbing Exploration Robot Toward Traveling on Rocky Cliffs (viability: 4): https://sciencetostartup.com/paper/limbero-a-limbed-climbing-exploration-robot-toward-traveling-on-rocky-cliffs - LIMBERO is a quadrupedal climbing robot designed for lunar and planetary exploration, capable of ascending steep rocky surfaces.
- Exploring different approaches to customize language models for domain-specific text-to-code generation (viability: 7): https://sciencetostartup.com/paper/exploring-different-approaches-to-customize-language-models-for-domain-specific-text-to-code-generation - Customizing smaller language models for domain-specific text-to-code generation using synthetic datasets.
- An approximate graph elicits detonation lattice (viability: 3): https://sciencetostartup.com/paper/an-approximate-graph-elicits-detonation-lattice - A novel graph-based algorithm for precise segmentation of detonation cells from 3D pressure traces.
- FEAT: A Linear-Complexity Foundation Model for Extremely Large Structured Data (viability: 4): https://sciencetostartup.com/paper/feat-a-linear-complexity-foundation-model-for-extremely-large-structured-data - FEAT is a linear-complexity foundation model designed to efficiently handle extremely large structured data across various domains.
- VIEW2SPACE: Studying Multi-View Visual Reasoning from Sparse Observations (viability: 5): https://sciencetostartup.com/paper/view2space-studying-multi-view-visual-reasoning-from-sparse-observations - VIEW2SPACE offers a novel benchmark for advancing multi-view visual reasoning through scalable data generation and evaluation.
- When Rolling Gets Weird: A Curved-Link Tensegrity Robot for Non-Intuitive Behavior (viability: 4): https://sciencetostartup.com/paper/when-rolling-gets-weird-a-curved-link-tensegrity-robot-for-non-intuitive-behavior - A curved-link tensegrity robot that enhances rolling locomotion while maintaining stability for space exploration.
- From the Inside Out: Progressive Distribution Refinement for Confidence Calibration (viability: 4): https://sciencetostartup.com/paper/from-the-inside-out-progressive-distribution-refinement-for-confidence-calibration - DistriTTRL optimizes reward signals in Reinforcement Learning by leveraging model confidence distribution to enhance performance and mitigate reward hacking.
- Bridging the High-Frequency Data Gap: A Millisecond-Resolution Network Dataset for Advancing Time Series Foundation Models (viability: 4): https://sciencetostartup.com/paper/bridging-the-high-frequency-data-gap-a-millisecond-resolution-network-dataset-for-advancing-time-series-foundation-model - A novel dataset for high-frequency time series data to enhance time series foundation models.
- AdaMem: Adaptive User-Centric Memory for Long-Horizon Dialogue Agents (viability: 7): https://sciencetostartup.com/paper/adamem-adaptive-user-centric-memory-for-long-horizon-dialogue-agents - AdaMem is an adaptive memory framework designed to enhance long-horizon dialogue agents with user-centric understanding.
- ExpressMind: A Multimodal Pretrained Large Language Model for Expressway Operation (viability: 9): https://sciencetostartup.com/paper/expressmind-a-multimodal-pretrained-large-language-model-for-expressway-operation - ExpressMind is a pioneering multimodal AI solution optimizing expressway operations through advanced reasoning and real-time decision-making.
- Unlearning for One-Step Generative Models via Unbalanced Optimal Transport (viability: 5): https://sciencetostartup.com/paper/unlearning-for-one-step-generative-models-via-unbalanced-optimal-transport - UOT-Unlearn offers a novel framework for safe unlearning in one-step generative models using Unbalanced Optimal Transport.
- On the Emotion Understanding of Synthesized Speech (viability: 4): https://sciencetostartup.com/paper/on-the-emotion-understanding-of-synthesized-speech - This research evaluates the effectiveness of emotion understanding models in synthesized speech, highlighting significant gaps in current methodologies.
- DST-Net: A Dual-Stream Transformer with Illumination-Independent Feature Guidance and Multi-Scale Spatial Convolution for Low-Light Image Enhancement (viability: 4): https://sciencetostartup.com/paper/dst-net-a-dual-stream-transformer-with-illumination-independent-feature-guidance-and-multi-scale-spatial-convolution-for - DST-Net enhances low-light images using a novel dual-stream transformer architecture for improved visibility.
- Optimal uncertainty bounds for multivariate kernel regression under bounded noise: A Gaussian process-based dual function (viability: 2): https://sciencetostartup.com/paper/optimal-uncertainty-bounds-for-multivariate-kernel-regression-under-bounded-noise-a-gaussian-process-based-dual-function - This paper presents a new method for improving uncertainty bounds in multivariate kernel regression using Gaussian processes.
- Breaking the Chain: A Causal Analysis of LLM Faithfulness to Intermediate Structures (viability: 4): https://sciencetostartup.com/paper/breaking-the-chain-a-causal-analysis-of-llm-faithfulness-to-intermediate-structures - A causal evaluation protocol to assess the faithfulness of LLMs to intermediate structures in decision-making.
- Coverage First Next Best View for Inspection of Cluttered Pipe Networks Using Mobile Manipulators (viability: 5): https://sciencetostartup.com/paper/coverage-first-next-best-view-for-inspection-of-cluttered-pipe-networks-using-mobile-manipulators - A mobile manipulator system for autonomous inspection of cluttered pipe networks in hazardous environments.
- Multi-Agent Reinforcement Learning Counteracts Delayed CSI in Multi-Satellite Systems (viability: 4): https://sciencetostartup.com/paper/multi-agent-reinforcement-learning-counteracts-delayed-csi-in-multi-satellite-systems - A multi-agent reinforcement learning algorithm to enhance satellite communication by optimizing channel state information.
- Follow the Clues, Frame the Truth: Hybrid-evidential Deductive Reasoning in Open-Vocabulary Multimodal Emotion Recognition (viability: 3): https://sciencetostartup.com/paper/follow-the-clues-frame-the-truth-hybrid-evidential-deductive-reasoning-in-open-vocabulary-multimodal-emotion-recognition - HyDRA enhances multimodal emotion recognition through a novel reasoning architecture that reconciles diverse emotional cues.
- GAP-MLLM: Geometry-Aligned Pre-training for Activating 3D Spatial Perception in Multimodal Large Language Models (viability: 7): https://sciencetostartup.com/paper/gap-mllm-geometry-aligned-pre-training-for-activating-3d-spatial-perception-in-multimodal-large-language-models - GAP-MLLM enhances 3D spatial perception in multimodal large language models through geometry-aligned pre-training.
- DynHD: Hallucination Detection for Diffusion Large Language Models via Denoising Dynamics Deviation Learning (viability: 8): https://sciencetostartup.com/paper/dynhd-hallucination-detection-for-diffusion-large-language-models-via-denoising-dynamics-deviation-learning - DynHD offers a novel approach to detect hallucinations in diffusion large language models by analyzing token-level uncertainty and denoising dynamics.
- Evo-Retriever: LLM-Guided Curriculum Evolution with Viewpoint-Pathway Collaboration for Multimodal Document Retrieval (viability: 3): https://sciencetostartup.com/paper/evo-retriever-llm-guided-curriculum-evolution-with-viewpoint-pathway-collaboration-for-multimodal-document-retrieval - Evo-Retriever enhances multimodal document retrieval through LLM-guided curriculum evolution and viewpoint-pathway collaboration.
- RetailBench: Evaluating Long-Horizon Autonomous Decision-Making and Strategy Stability of LLM Agents in Realistic Retail Environments (viability: 4): https://sciencetostartup.com/paper/retailbench-evaluating-long-horizon-autonomous-decision-making-and-strategy-stability-of-llm-agents-in-realistic-retail- - RetailBench is a benchmark for evaluating long-horizon decision-making of LLM agents in dynamic retail environments.
- TinyGLASS: Real-Time Self-Supervised In-Sensor Anomaly Detection (viability: 7): https://sciencetostartup.com/paper/tinyglass-real-time-self-supervised-in-sensor-anomaly-detection - TinyGLASS is a lightweight, real-time anomaly detection system tailored for resource-constrained edge platforms.
- TRUST-SQL: Tool-Integrated Multi-Turn Reinforcement Learning for Text-to-SQL over Unknown Schemas (viability: 7): https://sciencetostartup.com/paper/trust-sql-tool-integrated-multi-turn-reinforcement-learning-for-text-to-sql-over-unknown-schemas - TRUST-SQL revolutionizes Text-to-SQL parsing by enabling agents to dynamically identify relevant database schemas without pre-loaded metadata.
- ProgressiveAvatars: Progressive Animatable 3D Gaussian Avatars (viability: 4): https://sciencetostartup.com/paper/progressiveavatars-progressive-animatable-3d-gaussian-avatars - ProgressiveAvatars offers a dynamic 3D avatar representation that adapts to network conditions for real-time XR applications.
- Unified Removal of Raindrops and Reflections: A New Benchmark and A Novel Pipeline (viability: 7): https://sciencetostartup.com/paper/unified-removal-of-raindrops-and-reflections-a-new-benchmark-and-a-novel-pipeline - A novel pipeline for removing raindrops and reflections from images using a diffusion-based framework.
- Visual Distraction Undermines Moral Reasoning in Vision-Language Models (viability: 4): https://sciencetostartup.com/paper/visual-distraction-undermines-moral-reasoning-in-vision-language-models - A multimodal benchmark for evaluating moral reasoning in AI systems using visual inputs.
- Fast-HaMeR: Boosting Hand Mesh Reconstruction using Knowledge Distillation (viability: 8): https://sciencetostartup.com/paper/fast-hamer-boosting-hand-mesh-reconstruction-using-knowledge-distillation - Boost lightweight 3D hand reconstruction on mobile and VR devices with Fast-HaMeR.
- Capability-Guided Compression: Toward Interpretability-Aware Budget Allocation for Large Language Models (viability: 2): https://sciencetostartup.com/paper/capability-guided-compression-toward-interpretability-aware-budget-allocation-for-large-language-models - A theoretical framework for capability-aware compression in large language models.
- CD-FKD: Cross-Domain Feature Knowledge Distillation for Robust Single-Domain Generalization in Object Detection (viability: 7): https://sciencetostartup.com/paper/cd-fkd-cross-domain-feature-knowledge-distillation-for-robust-single-domain-generalization-in-object-detection - CD-FKD enhances object detection robustness by leveraging cross-domain feature knowledge distillation.
- DISCOVER: A Solver for Distributional Counterfactual Explanations (viability: 8): https://sciencetostartup.com/paper/discover-a-solver-for-distributional-counterfactual-explanations - DISCOVER is a model-agnostic solver that enhances distributional counterfactual explanations for non-differentiable models in tabular data.
- VQKV: High-Fidelity and High-Ratio Cache Compression via Vector-Quantization (viability: 3): https://sciencetostartup.com/paper/vqkv-high-fidelity-and-high-ratio-cache-compression-via-vector-quantization - VQKV is a training-free method for high-fidelity KV cache compression using vector quantization.
- From Natural Language to Executable Option Strategies via Large Language Models (viability: 8): https://sciencetostartup.com/paper/from-natural-language-to-executable-option-strategies-via-large-language-models - Transform natural language trading intents into executable option strategies using a novel query language and LLMs.
- IRIS: A Real-World Benchmark for Inverse Recovery and Identification of Physical Dynamic Systems from Monocular Video (viability: 7): https://sciencetostartup.com/paper/iris-a-real-world-benchmark-for-inverse-recovery-and-identification-of-physical-dynamic-systems-from-monocular-video - IRIS provides a comprehensive benchmark for unsupervised physical parameter estimation from real-world video data.
- EngGPT2: Sovereign, Efficient and Open Intelligence (viability: 5): https://sciencetostartup.com/paper/enggpt2-sovereign-efficient-and-open-intelligence - EngGPT2 is an efficient Italian LLM designed for high-performance NLP tasks with reduced resource requirements.
- An Efficient Heterogeneous Co-Design for Fine-Tuning on a Single GPU (viability: 7): https://sciencetostartup.com/paper/an-efficient-heterogeneous-co-design-for-fine-tuning-on-a-single-gpu - SlideFormer enables efficient fine-tuning of large language models on a single GPU, democratizing access to advanced AI capabilities.
- Cross-modal learning for plankton recognition (viability: 8): https://sciencetostartup.com/paper/cross-modal-learning-for-plankton-recognition - A self-supervised cross-modal approach for efficient plankton recognition using minimal labeled data.
- 3D Fourier-based Global Feature Extraction for Hyperspectral Image Classification (viability: 2): https://sciencetostartup.com/paper/3d-fourier-based-global-feature-extraction-for-hyperspectral-image-classification - HGFNet offers a novel approach to hyperspectral image classification by integrating 3D convolutional feature extraction with advanced frequency-domain filtering.
- Early-Terminable Energy-Safe Iterative Coupling for Parallel Simulation of Port-Hamiltonian Systems (viability: 3): https://sciencetostartup.com/paper/early-terminable-energy-safe-iterative-coupling-for-parallel-simulation-of-port-hamiltonian-systems - A novel energy-safe iterative coupling method for real-time parallel simulation of port-Hamiltonian robotic systems.
- SF-Mamba: Rethinking State Space Model for Vision (viability: 7): https://sciencetostartup.com/paper/sf-mamba-rethinking-state-space-model-for-vision - SF-Mamba rethinks the scan operation for vision to enhance computational efficiency and performance in visual tasks.
- HGP-Mamba: Integrating Histology and Generated Protein Features for Mamba-based Multimodal Survival Risk Prediction (viability: 8): https://sciencetostartup.com/paper/hgp-mamba-integrating-histology-and-generated-protein-features-for-mamba-based-multimodal-survival-risk-prediction - HGP-Mamba integrates histology and generated protein features for advanced cancer survival risk prediction.
- Via Negativa for AI Alignment: Why Negative Constraints Are Structurally Superior to Positive Preferences (viability: 4): https://sciencetostartup.com/paper/via-negativa-for-ai-alignment-why-negative-constraints-are-structurally-superior-to-positive-preferences - This research proposes a shift in AI alignment strategies by emphasizing negative constraints over positive preferences for training large language models.
- IndexRAG: Bridging Facts for Cross-Document Reasoning at Index Time (viability: 7): https://sciencetostartup.com/paper/indexrag-bridging-facts-for-cross-document-reasoning-at-index-time - IndexRAG transforms multi-hop question answering by enabling offline indexing for cross-document reasoning.
- Trained Persistent Memory for Frozen Encoder--Decoder LLMs: Six Architectural Methods (viability: 5): https://sciencetostartup.com/paper/trained-persistent-memory-for-frozen-encoder-decoder-llms-six-architectural-methods - A proof-of-concept study demonstrating persistent memory integration in frozen LLMs for enhanced conversational learning.
- RECOVER: Robust Entity Correction via agentic Orchestration of hypothesis Variants for Evidence-based Recovery (viability: 7): https://sciencetostartup.com/paper/recover-robust-entity-correction-via-agentic-orchestration-of-hypothesis-variants-for-evidence-based-recovery - RECOVER is an agentic correction framework that enhances entity recognition in ASR by leveraging multiple hypotheses and LLM correction.
- PlotTwist: A Creative Plot Generation Framework with Small Language Models (viability: 8): https://sciencetostartup.com/paper/plottwist-a-creative-plot-generation-framework-with-small-language-models - PlotTwist is a framework that empowers small language models to generate high-quality plots efficiently.
- Onboard MuJoCo-based Model Predictive Control for Shipboard Crane with Double-Pendulum Sway Suppression (viability: 7): https://sciencetostartup.com/paper/onboard-mujoco-based-model-predictive-control-for-shipboard-crane-with-double-pendulum-sway-suppression - A real-time control pipeline for maritime cranes that suppresses payload sway using MuJoCo-based model predictive control.
- Who Benchmarks the Benchmarks? A Case Study of LLM Evaluation in Icelandic (viability: 4): https://sciencetostartup.com/paper/who-benchmarks-the-benchmarks-a-case-study-of-llm-evaluation-in-icelandic - This paper critiques LLM benchmarking methods for Icelandic, highlighting flaws in synthetic data usage.
- Poisoning the Pixels: Revisiting Backdoor Attacks on Semantic Segmentation (viability: 4): https://sciencetostartup.com/paper/poisoning-the-pixels-revisiting-backdoor-attacks-on-semantic-segmentation - BADSEG uncovers vulnerabilities in semantic segmentation models to backdoor attacks, paving the way for improved defenses.
- Near-light Photometric Stereo with Symmetric Lights (viability: 3): https://sciencetostartup.com/paper/near-light-photometric-stereo-with-symmetric-lights - A novel linear solution for near-light photometric stereo using symmetric light arrangements.
- The Decentralisation Paradox in Digital Identity: Centralising Decentralisation with Digital Wallets? (viability: 2): https://sciencetostartup.com/paper/the-decentralisation-paradox-in-digital-identity-centralising-decentralisation-with-digital-wallets - This research critiques the decentralisation paradox in digital identity systems.
- Fanar 2.0: Arabic Generative AI Stack (viability: 8): https://sciencetostartup.com/paper/fanar-2-0-arabic-generative-ai-stack - Fanar 2.0 is a sovereign Arabic generative AI platform that delivers advanced language and multimodal capabilities.
- DermaFlux: Synthetic Skin Lesion Generation with Rectified Flows for Enhanced Image Classification (viability: 8): https://sciencetostartup.com/paper/dermaflux-synthetic-skin-lesion-generation-with-rectified-flows-for-enhanced-image-classification - DermaFlux generates synthetic skin lesion images to enhance classification accuracy in dermatology.
- Unpaired Cross-Domain Calibration of DMSP to VIIRS Nighttime Light Data Based on CUT Network (viability: 4): https://sciencetostartup.com/paper/unpaired-cross-domain-calibration-of-dmsp-to-viirs-nighttime-light-data-based-on-cut-network - A novel calibration method for transforming DMSP nighttime light data into VIIRS format using contrastive learning.
- Controlling Fish Schools via Reinforcement Learning of Virtual Fish Movement (viability: 7): https://sciencetostartup.com/paper/controlling-fish-schools-via-reinforcement-learning-of-virtual-fish-movement - A method to control fish schools using virtual agents trained with reinforcement learning.
- Rotated Robustness: A Training-Free Defense against Bit-Flip Attacks on Large Language Models (viability: 8): https://sciencetostartup.com/paper/rotated-robustness-a-training-free-defense-against-bit-flip-attacks-on-large-language-models - Rotated Robustness offers a training-free defense against bit-flip attacks on Large Language Models, ensuring reliability and accuracy.
- Age Predictors Through the Lens of Generalization, Bias Mitigation, and Interpretability: Reflections on Causal Implications (viability: 4): https://sciencetostartup.com/paper/age-predictors-through-the-lens-of-generalization-bias-mitigation-and-interpretability-reflections-on-causal-implication - A model for age prediction that enhances out-of-distribution generalization and mitigates bias through adversarial representation learning.
- Prior-Informed Neural Network Initialization: A Spectral Approach for Function Parameterizing Architectures (viability: 5): https://sciencetostartup.com/paper/prior-informed-neural-network-initialization-a-spectral-approach-for-function-parameterizing-architectures - A novel approach to neural network initialization that leverages spectral data properties for improved convergence and interpretability.
- Semantic One-Dimensional Tokenizer for Image Reconstruction and Generation (viability: 7): https://sciencetostartup.com/paper/semantic-one-dimensional-tokenizer-for-image-reconstruction-and-generation - SemTok is a novel semantic one-dimensional tokenizer that enhances image reconstruction and generation through compact token representation.
- InViC: Intent-aware Visual Cues for Medical Visual Question Answering (viability: 7): https://sciencetostartup.com/paper/invic-intent-aware-visual-cues-for-medical-visual-question-answering - InViC enhances medical visual question answering by integrating intent-aware visual cues into large language models.
- FederatedFactory: Generative One-Shot Learning for Extremely Non-IID Distributed Scenarios (viability: 7): https://sciencetostartup.com/paper/federatedfactory-generative-one-shot-learning-for-extremely-non-iid-distributed-scenarios - FederatedFactory revolutionizes federated learning by enabling generative one-shot learning for non-IID distributed scenarios.
- Encoding Predictability and Legibility for Style-Conditioned Diffusion Policy (viability: 3): https://sciencetostartup.com/paper/encoding-predictability-and-legibility-for-style-conditioned-diffusion-policy - A modular framework for optimizing robot motion based on task ambiguity.
- DynamicGate MLP Conditional Computation via Learned Structural Dropout and Input Dependent Gating for Functional Plasticity (viability: 3): https://sciencetostartup.com/paper/dynamicgate-mlp-conditional-computation-via-learned-structural-dropout-and-input-dependent-gating-for-functional-plastic - DynamicGate-MLP introduces a novel framework for efficient conditional computation in neural networks.
- FactorEngine: A Program-level Knowledge-Infused Factor Mining Framework for Quantitative Investment (viability: 3): https://sciencetostartup.com/paper/factorengine-a-program-level-knowledge-infused-factor-mining-framework-for-quantitative-investment - FactorEngine is a program-level framework for automated discovery of predictive signals in quantitative investment.
- Impact of File-Open Hook Points on Backup Ratio in ROFBS on XFS (viability: 3): https://sciencetostartup.com/paper/impact-of-file-open-hook-points-on-backup-ratio-in-rofbs-on-xfs - This study analyzes the impact of file-open hook points on the effectiveness of real-time backup systems against ransomware.
- Advancing Visual Reliability: Color-Accurate Underwater Image Enhancement for Real-Time Underwater Missions (viability: 8): https://sciencetostartup.com/paper/advancing-visual-reliability-color-accurate-underwater-image-enhancement-for-real-time-underwater-missions - A lightweight, real-time underwater image enhancement framework that restores color accuracy for underwater missions.
- $D^3$-RSMDE: 40$\times$ Faster and High-Fidelity Remote Sensing Monocular Depth Estimation (viability: 7): https://sciencetostartup.com/paper/d-3-rsmde-40-times-faster-and-high-fidelity-remote-sensing-monocular-depth-estimation - A high-fidelity monocular depth estimation framework that balances speed and quality for remote sensing applications.
- Toward Experimentation-as-a-Service in 5G/6G: The Plaza6G Prototype for AI-Assisted Trials (viability: 6): https://sciencetostartup.com/paper/toward-experimentation-as-a-service-in-5g-6g-the-plaza6g-prototype-for-ai-assisted-trials - Plaza6G is an Experiment-as-a-Service platform that simplifies AI-assisted trials in next-generation wireless networks.
- PashtoCorp: A 1.25-Billion-Word Corpus, Evaluation Suite, and Reproducible Pipeline for Low-Resource Language Development (viability: 8): https://sciencetostartup.com/paper/pashtocorp-a-1-25-billion-word-corpus-evaluation-suite-and-reproducible-pipeline-for-low-resource-language-development - Build cutting-edge NLP models for Pashto using the largest available Pashto language corpus, PashtoCorp.
- Automated identification of Ichneumonoidea wasps via YOLO-based deep learning: Integrating HiresCam for Explainable AI (viability: 7): https://sciencetostartup.com/paper/automated-identification-of-ichneumonoidea-wasps-via-yolo-based-deep-learning-integrating-hirescam-for-explainable-ai - Automated identification of Ichneumonoidea wasps using a YOLO-based deep learning framework for biodiversity assessment.
- SseRex: Practical Symbolic Execution of Solana Smart Contracts (viability: 3): https://sciencetostartup.com/paper/sserex-practical-symbolic-execution-of-solana-smart-contracts - SseRex is a symbolic execution tool designed to detect vulnerabilities in Solana smart contracts.
- Learning Human-Object Interaction for 3D Human Pose Estimation from LiDAR Point Clouds (viability: 7): https://sciencetostartup.com/paper/learning-human-object-interaction-for-3d-human-pose-estimation-from-lidar-point-clouds - A framework for robust 3D human pose estimation from LiDAR point clouds leveraging human-object interactions.
- Detecting Sentiment Steering Attacks on RAG-enabled Large Language Models (viability: 4): https://sciencetostartup.com/paper/detecting-sentiment-steering-attacks-on-rag-enabled-large-language-models - Lightweight deep learning-based intrusion detection systems enhance IoT network security against cyber threats.
- PKINet-v2: Towards Powerful and Efficient Poly-Kernel Remote Sensing Object Detection (viability: 7): https://sciencetostartup.com/paper/pkinet-v2-towards-powerful-and-efficient-poly-kernel-remote-sensing-object-detection - PKINet-v2 is an advanced backbone for remote sensing object detection that efficiently combines multiple kernel types for superior accuracy and speed.
- Iris: Bringing Real-World Priors into Diffusion Model for Monocular Depth Estimation (viability: 2): https://sciencetostartup.com/paper/iris-bringing-real-world-priors-into-diffusion-model-for-monocular-depth-estimation - Iris enhances monocular depth estimation by integrating real-world priors into a diffusion model.
- SpikeCLR: Contrastive Self-Supervised Learning for Few-Shot Event-Based Vision using Spiking Neural Networks (viability: 6): https://sciencetostartup.com/paper/spikeclr-contrastive-self-supervised-learning-for-few-shot-event-based-vision-using-spiking-neural-networks - SpikeCLR leverages self-supervised learning to enhance spiking neural networks for event-based vision in low-data environments.
- Faulty Coffees: Barriers to Adoption of an In-the-wild Robo-Barista (viability: 2): https://sciencetostartup.com/paper/faulty-coffees-barriers-to-adoption-of-an-in-the-wild-robo-barista - This study explores the challenges of engaging users with a service robot in a real-world setting.
- Behavioral Steering in a 35B MoE Language Model via SAE-Decoded Probe Vectors: One Agency Axis, Not Five Traits (viability: 2): https://sciencetostartup.com/paper/behavioral-steering-in-a-35b-moe-language-model-via-sae-decoded-probe-vectors-one-agency-axis-not-five-traits - This research explores behavioral steering in a large language model using sparse autoencoders to enhance agentic traits.
- Decoding the Critique Mechanism in Large Reasoning Models (viability: 8): https://sciencetostartup.com/paper/decoding-the-critique-mechanism-in-large-reasoning-models - A study revealing the hidden critique ability in Large Reasoning Models to enhance error detection and self-correction.
- An Interpretable Machine Learning Framework for Non-Small Cell Lung Cancer Drug Response Analysis (viability: 5): https://sciencetostartup.com/paper/an-interpretable-machine-learning-framework-for-non-small-cell-lung-cancer-drug-response-analysis - A machine learning framework for personalized lung cancer treatment analysis using genetic data.
- ADAPT: Adaptive Dual-projection Architecture for Perceptive Traversal (viability: 3): https://sciencetostartup.com/paper/adapt-adaptive-dual-projection-architecture-for-perceptive-traversal - ADAPT enhances humanoid locomotion in complex 3D environments through adaptive spatial sensing.
- A Human-Centred Architecture for Large Language Models-Cognitive Assistants in Manufacturing within Quality Management Systems (viability: 2): https://sciencetostartup.com/paper/a-human-centred-architecture-for-large-language-models-cognitive-assistants-in-manufacturing-within-quality-management-s - This paper proposes a human-centred architecture for integrating cognitive assistants into quality management systems in manufacturing.
- Systematization of Knowledge: The Design Space of Digital Payment Systems with Potential for CBDC (viability: 2): https://sciencetostartup.com/paper/systematization-of-knowledge-the-design-space-of-digital-payment-systems-with-potential-for-cbdc - A comprehensive analysis of payment system designs for Central Bank Digital Currencies (CBDCs).
- Learning to Predict, Discover, and Reason in High-Dimensional Discrete Event Sequences (viability: 5): https://sciencetostartup.com/paper/learning-to-predict-discover-and-reason-in-high-dimensional-discrete-event-sequences - A framework for automated fault diagnostics in vehicles using advanced event sequence modeling and causal discovery.
- Omnilingual MT: Machine Translation for 1,600 Languages (viability: 8): https://sciencetostartup.com/paper/omnilingual-mt-machine-translation-for-1-600-languages - Omnilingual MT offers high-quality machine translation for over 1,600 languages, significantly expanding multilingual capabilities.
- NeSy-Route: A Neuro-Symbolic Benchmark for Constrained Route Planning in Remote Sensing (viability: 4): https://sciencetostartup.com/paper/nesy-route-a-neuro-symbolic-benchmark-for-constrained-route-planning-in-remote-sensing - NeSy-Route is a neuro-symbolic benchmark designed to enhance route planning capabilities in remote sensing applications.
- DriveFix: Spatio-Temporally Coherent Driving Scene Restoration (viability: 7): https://sciencetostartup.com/paper/drivefix-spatio-temporally-coherent-driving-scene-restoration - DriveFix is a multi-view restoration framework that ensures spatio-temporal coherence for driving scenes in autonomous driving applications.
- Toward Deep Representation Learning for Event-Enhanced Visual Autonomous Perception: the eAP Dataset (viability: 6): https://sciencetostartup.com/paper/toward-deep-representation-learning-for-event-enhanced-visual-autonomous-perception-the-eap-dataset - The eAP dataset enhances visual perception in autonomous driving by leveraging event camera data for improved 3D vehicle detection and object time-to-contact estimation.
- Micro-AU CLIP: Fine-Grained Contrastive Learning from Local Independence to Global Dependency for Micro-Expression Action Unit Detection (viability: 7): https://sciencetostartup.com/paper/micro-au-clip-fine-grained-contrastive-learning-from-local-independence-to-global-dependency-for-micro-expression-action - Micro-AU CLIP enhances micro-expression detection by modeling local independence and global dependency of action units.
- OGScene3D: Incremental Open-Vocabulary 3D Gaussian Scene Graph Mapping for Scene Understanding (viability: 6): https://sciencetostartup.com/paper/ogscene3d-incremental-open-vocabulary-3d-gaussian-scene-graph-mapping-for-scene-understanding - OGScene3D enables incremental open-vocabulary 3D scene understanding for robotic applications.
- PyPhonPlan: Simulating phonetic planning with dynamic neural fields and task dynamics (viability: 7): https://sciencetostartup.com/paper/pyphonplan-simulating-phonetic-planning-with-dynamic-neural-fields-and-task-dynamics - PyPhonPlan is an open-source toolkit for simulating phonetic planning and speech dynamics using dynamic neural fields.
- Attention-guided Evidence Grounding for Spoken Question Answering (viability: 7): https://sciencetostartup.com/paper/attention-guided-evidence-grounding-for-spoken-question-answering - A novel framework for improving Spoken Question Answering by grounding evidence through attention mechanisms in SpeechLLMs.
- VisBrowse-Bench: Benchmarking Visual-Native Search for Multimodal Browsing Agents (viability: 8): https://sciencetostartup.com/paper/visbrowse-bench-benchmarking-visual-native-search-for-multimodal-browsing-agents - VisBrowse-Bench is a benchmark for evaluating visual reasoning in multimodal browsing agents.
- Surrogate-Assisted Genetic Programming with Rank-Based Phenotypic Characterisation for Dynamic Multi-Mode Project Scheduling (viability: 4): https://sciencetostartup.com/paper/surrogate-assisted-genetic-programming-with-rank-based-phenotypic-characterisation-for-dynamic-multi-mode-project-schedu - A surrogate-assisted genetic programming approach to enhance decision-making in dynamic project scheduling.
- Persistent Story World Simulation with Continuous Character Customization (viability: 6): https://sciencetostartup.com/paper/persistent-story-world-simulation-with-continuous-character-customization - EverTale is a story world simulator that enables continuous character customization and integration for enhanced visual storytelling.
- Locate-then-Sparsify: Attribution Guided Sparse Strategy for Visual Hallucination Mitigation (viability: 7): https://sciencetostartup.com/paper/locate-then-sparsify-attribution-guided-sparse-strategy-for-visual-hallucination-mitigation - A framework that mitigates hallucinations in Large Vision-Language Models by applying targeted feature steering based on layer relevance.
- Laya: A LeJEPA Approach to EEG via Latent Prediction over Reconstruction (viability: 4): https://sciencetostartup.com/paper/laya-a-lejepa-approach-to-eeg-via-latent-prediction-over-reconstruction - Laya is an innovative EEG foundation model that enhances brain signal representation through latent predictive learning.
- Agile Interception of a Flying Target using Competitive Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/agile-interception-of-a-flying-target-using-competitive-reinforcement-learning - A competitive reinforcement learning solution for intercepting agile drones using trained policies.
- Physics-integrated neural differentiable modeling for immersed boundary systems (viability: 7): https://sciencetostartup.com/paper/physics-integrated-neural-differentiable-modeling-for-immersed-boundary-systems - A physics-integrated neural framework for efficient long-horizon prediction of fluid flows near solid boundaries.
- GenZ-LIO: Generalizable LiDAR-Inertial Odometry Beyond Indoor--Outdoor Boundaries (viability: 7): https://sciencetostartup.com/paper/genz-lio-generalizable-lidar-inertial-odometry-beyond-indoor-outdoor-boundaries - GenZ-LIO is a robust LiDAR-inertial odometry framework that adapts to both indoor and outdoor environments for autonomous navigation.
- VIGOR: VIdeo Geometry-Oriented Reward for Temporal Generative Alignment (viability: 7): https://sciencetostartup.com/paper/vigor-video-geometry-oriented-reward-for-temporal-generative-alignment - VIGOR enhances video generation by using a geometry-based reward model for improved consistency and robustness.
- MG-Grasp: Metric-Scale Geometric 6-DoF Grasping Framework with Sparse RGB Observations (viability: 7): https://sciencetostartup.com/paper/mg-grasp-metric-scale-geometric-6-dof-grasping-framework-with-sparse-rgb-observations - MG-Grasp is a depth-free 6-DoF grasping framework that enhances robotic manipulation using sparse RGB observations.
- FG-SGL: Fine-Grained Semantic Guidance Learning via Motion Process Decomposition for Micro-Gesture Recognition (viability: 4): https://sciencetostartup.com/paper/fg-sgl-fine-grained-semantic-guidance-learning-via-motion-process-decomposition-for-micro-gesture-recognition - FG-SGL enhances micro-gesture recognition by integrating fine-grained and category-level semantics.
- Novel CRT-based Asymptotically Ideal Disjunctive Hierarchical Secret Sharing Scheme (viability: 2): https://sciencetostartup.com/paper/novel-crt-based-asymptotically-ideal-disjunctive-hierarchical-secret-sharing-scheme - A novel CRT-based secret sharing scheme that aims to enhance security and flexibility in share sizes.
- Adaptive Theory of Mind for LLM-based Multi-Agent Coordination (viability: 7): https://sciencetostartup.com/paper/adaptive-theory-of-mind-for-llm-based-multi-agent-coordination - A-ToM agents enhance multi-agent coordination by aligning Theory of Mind reasoning between agents.
- AW-MoE: All-Weather Mixture of Experts for Robust Multi-Modal 3D Object Detection (viability: 9): https://sciencetostartup.com/paper/aw-moe-all-weather-mixture-of-experts-for-robust-multi-modal-3d-object-detection - AW-MoE enhances 3D object detection in adverse weather conditions using a novel Mixture of Experts framework.
- Human/AI Collective Intelligence for Deliberative Democracy: A Human-Centred Design Approach (viability: 4): https://sciencetostartup.com/paper/human-ai-collective-intelligence-for-deliberative-democracy-a-human-centred-design-approach - A human-centered design approach to enhance deliberative democracy through collective intelligence and AI tools.
- Is Semi-Automatic Transcription Useful in Corpus Creation? Preliminary Considerations on the KIParla Corpus (viability: 3): https://sciencetostartup.com/paper/is-semi-automatic-transcription-useful-in-corpus-creation-preliminary-considerations-on-the-kiparla-corpus - This paper explores the integration of ASR in transcription workflows for the KIParla corpus, highlighting speed improvements but inconsistent accuracy.
- Point-to-Mask: From Arbitrary Point Annotations to Mask-Level Infrared Small Target Detection (viability: 8): https://sciencetostartup.com/paper/point-to-mask-from-arbitrary-point-annotations-to-mask-level-infrared-small-target-detection - Point-to-Mask revolutionizes infrared small target detection by transforming low-cost point annotations into accurate mask-level detections.
- When Thinking Hurts: Mitigating Visual Forgetting in Video Reasoning via Frame Repetition (viability: 7): https://sciencetostartup.com/paper/when-thinking-hurts-mitigating-visual-forgetting-in-video-reasoning-via-frame-repetition - FrameRepeat enhances Video-LLMs by autonomously reinforcing key frames to improve reasoning accuracy.
- Grounding the Score: Explicit Visual Premise Verification for Reliable Vision-Language Process Reward Models (viability: 9): https://sciencetostartup.com/paper/grounding-the-score-explicit-visual-premise-verification-for-reliable-vision-language-process-reward-models - EVPV enhances vision-language models by providing explicit verification of visual premises to improve reasoning accuracy.
- Visual Prompt Discovery via Semantic Exploration (viability: 7): https://sciencetostartup.com/paper/visual-prompt-discovery-via-semantic-exploration - An automated framework for discovering effective visual prompts to enhance LVLM perception through semantic exploration.
- Synergizing Deep Learning and Biological Heuristics for Extreme Long-Tail White Blood Cell Classification (viability: 7): https://sciencetostartup.com/paper/synergizing-deep-learning-and-biological-heuristics-for-extreme-long-tail-white-blood-cell-classification - A hybrid framework for automated white blood cell classification that enhances rare-class generalization using biological heuristics.
- How to Utilize Complementary Vision-Text Information for 2D Structure Understanding (viability: 7): https://sciencetostartup.com/paper/how-to-utilize-complementary-vision-text-information-for-2d-structure-understanding - DiVA-Former is a lightweight architecture that enhances 2D structure understanding by effectively integrating vision and text information.
- More Rounds, More Noise: Why Multi-Turn Review Fails to Improve Cross-Context Verification (viability: 2): https://sciencetostartup.com/paper/more-rounds-more-noise-why-multi-turn-review-fails-to-improve-cross-context-verification - Dynamic Cross-Context Review (D-CCR) explores the limitations of multi-turn review in LLM verification.
- RASLF: Representation-Aware State Space Model for Light Field Super-Resolution (viability: 4): https://sciencetostartup.com/paper/raslf-representation-aware-state-space-model-for-light-field-super-resolution - RASLF is a state-space framework for enhancing light field super-resolution by optimizing structural correlations across multiple representations.
- Exclusivity-Guided Mask Learning for Semi-Supervised Crowd Instance Segmentation and Counting (viability: 7): https://sciencetostartup.com/paper/exclusivity-guided-mask-learning-for-semi-supervised-crowd-instance-segmentation-and-counting - A semi-supervised framework for crowd instance segmentation and counting using exclusivity-guided mask learning.
- Industrial cuVSLAM Benchmark & Integration (viability: 5): https://sciencetostartup.com/paper/industrial-cuvslam-benchmark-integration - A benchmark evaluation and integration of cuVSLAM for enhanced mobile robot navigation in logistics.
- PureCLIP-Depth: Prompt-Free and Decoder-Free Monocular Depth Estimation within CLIP Embedding Space (viability: 8): https://sciencetostartup.com/paper/pureclip-depth-prompt-free-and-decoder-free-monocular-depth-estimation-within-clip-embedding-space - PureCLIP-Depth offers a novel, prompt-free method for monocular depth estimation leveraging CLIP embeddings.
- ReFORM: Review-aggregated Profile Generation via LLM with Multi-Factor Attention for Restaurant Recommendation (viability: 8): https://sciencetostartup.com/paper/reform-review-aggregated-profile-generation-via-llm-with-multi-factor-attention-for-restaurant-recommendation - ReFORM enhances restaurant recommendations by generating user and item profiles from reviews using LLMs and multi-factor attention.
- Ground Reaction Inertial Poser: Physics-based Human Motion Capture from Sparse IMUs and Insole Pressure Sensors (viability: 7): https://sciencetostartup.com/paper/ground-reaction-inertial-poser-physics-based-human-motion-capture-from-sparse-imus-and-insole-pressure-sensors - GRIP reconstructs realistic human motion using IMUs and foot pressure data for enhanced physical accuracy.
- PA-LVIO: Real-Time LiDAR-Visual-Inertial Odometry and Mapping with Pose-Only Bundle Adjustment (viability: 7): https://sciencetostartup.com/paper/pa-lvio-real-time-lidar-visual-inertial-odometry-and-mapping-with-pose-only-bundle-adjustment - PA-LVIO offers real-time LiDAR-visual-inertial odometry and mapping for intelligent transportation systems.
- Dual Consensus: Escaping from Spurious Majority in Unsupervised RLVR via Two-Stage Vote Mechanism (viability: 7): https://sciencetostartup.com/paper/dual-consensus-escaping-from-spurious-majority-in-unsupervised-rlvr-via-two-stage-vote-mechanism - Dual Consensus Reinforcement Learning enhances LLM performance on reasoning tasks through a novel self-supervised training method.
- SpecSteer: Synergizing Local Context and Global Reasoning for Efficient Personalized Generation (viability: 7): https://sciencetostartup.com/paper/specsteer-synergizing-local-context-and-global-reasoning-for-efficient-personalized-generation - SpecSteer enhances personalized generation by combining local context with cloud reasoning while ensuring user privacy.
- Enabling Dynamic Tracking in Vision-Language-Action Models via Time-Discrete and Time-Continuous Velocity Feedforward (viability: 7): https://sciencetostartup.com/paper/enabling-dynamic-tracking-in-vision-language-action-models-via-time-discrete-and-time-continuous-velocity-feedforward - A novel approach to enhance robot manipulation by integrating velocity feedforward terms into vision-language-action models.
- Generative AI for Quantum Circuits and Quantum Code: A Technical Review and Taxonomy (viability: 3): https://sciencetostartup.com/paper/generative-ai-for-quantum-circuits-and-quantum-code-a-technical-review-and-taxonomy - A comprehensive review of generative systems for quantum circuit and code generation, highlighting gaps in practical deployment.
- CoMAI: A Collaborative Multi-Agent Framework for Robust and Equitable Interview Evaluation (viability: 7): https://sciencetostartup.com/paper/comai-a-collaborative-multi-agent-framework-for-robust-and-equitable-interview-evaluation - CoMAI is a multi-agent framework that enhances the fairness and robustness of AI-driven interview evaluations.
- Leveling3D: Leveling Up 3D Reconstruction with Feed-Forward 3D Gaussian Splatting and Geometry-Aware Generation (viability: 8): https://sciencetostartup.com/paper/leveling3d-leveling-up-3d-reconstruction-with-feed-forward-3d-gaussian-splatting-and-geometry-aware-generation - Leveling3D enhances 3D reconstruction by integrating geometry-aware generation for improved novel-view synthesis.
- MOSAIC: Composable Safety Alignment with Modular Control Tokens (viability: 3): https://sciencetostartup.com/paper/mosaic-composable-safety-alignment-with-modular-control-tokens - MOSAIC introduces a modular framework for compositional safety alignment in large language models using learnable control tokens.
- Proactive Rejection and Grounded Execution: A Dual-Stage Intent Analysis Paradigm for Safe and Efficient AIoT Smart Homes (viability: 7): https://sciencetostartup.com/paper/proactive-rejection-and-grounded-execution-a-dual-stage-intent-analysis-paradigm-for-safe-and-efficient-aiot-smart-homes - A dual-stage intent analysis framework that enhances the reliability of LLM commands in smart home environments.
- Offline Exploration-Aware Fine-Tuning for Long-Chain Mathematical Reasoning (viability: 7): https://sciencetostartup.com/paper/offline-exploration-aware-fine-tuning-for-long-chain-mathematical-reasoning - Offline eXploration-Aware fine-tuning enhances mathematical reasoning in large language models through optimized data handling.
- A Scoping Review of AI-Driven Digital Interventions in Mental Health Care: Mapping Applications Across Screening, Support, Monitoring, Prevention, and Clinical Education (viability: 2): https://sciencetostartup.com/paper/a-scoping-review-of-ai-driven-digital-interventions-in-mental-health-care-mapping-applications-across-screening-support- - A comprehensive review of AI applications in mental health care highlighting key use cases and challenges.
- Online Semi-infinite Linear Programming: Efficient Algorithms via Function Approximation (viability: 4): https://sciencetostartup.com/paper/online-semi-infinite-linear-programming-efficient-algorithms-via-function-approximation - A novel approach to online semi-infinite linear programming that improves performance with function approximation.
- Are Large Language Models Truly Smarter Than Humans? (viability: 4): https://sciencetostartup.com/paper/are-large-language-models-truly-smarter-than-humans - This paper audits the contamination of large language models' performance on public benchmarks, revealing significant risks in their evaluation.
- PanguMotion: Continuous Driving Motion Forecasting with Pangu Transformers (viability: 5): https://sciencetostartup.com/paper/pangumotion-continuous-driving-motion-forecasting-with-pangu-transformers - PanguMotion enhances motion forecasting in autonomous driving by integrating Transformer blocks for continuous trajectory prediction.
- S-VAM: Shortcut Video-Action Model by Self-Distilling Geometric and Semantic Foresight (viability: 8): https://sciencetostartup.com/paper/s-vam-shortcut-video-action-model-by-self-distilling-geometric-and-semantic-foresight - S-VAM is a shortcut video-action model that enhances robot learning through efficient geometric and semantic foresight.
- Structured Semantic Cloaking for Jailbreak Attacks on Large Language Models (viability: 3): https://sciencetostartup.com/paper/structured-semantic-cloaking-for-jailbreak-attacks-on-large-language-models - Structured Semantic Cloaking (S2C) is a novel framework designed to enhance jailbreak attacks on large language models by manipulating semantic intent reconstruction.
- Reliable Reasoning in SVG-LLMs via Multi-Task Multi-Reward Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/reliable-reasoning-in-svg-llms-via-multi-task-multi-reward-reinforcement-learning - CTRL-S enhances SVG generation with structured reasoning and multi-reward optimization.
- ECHO: Edge-Cloud Humanoid Orchestration for Language-to-Motion Control (viability: 8): https://sciencetostartup.com/paper/echo-edge-cloud-humanoid-orchestration-for-language-to-motion-control - ECHO enables language-driven control of humanoid robots through an innovative edge-cloud framework.
- Sample-Efficient Adaptation of Drug-Response Models to Patient Tumors under Strong Biological Domain Shift (viability: 6): https://sciencetostartup.com/paper/sample-efficient-adaptation-of-drug-response-models-to-patient-tumors-under-strong-biological-domain-shift - A framework for adapting drug-response models to patient tumors using sample-efficient transfer learning.
- Polyglot-Lion: Efficient Multilingual ASR for Singapore via Balanced Fine-Tuning of Qwen3-ASR (viability: 8): https://sciencetostartup.com/paper/polyglot-lion-efficient-multilingual-asr-for-singapore-via-balanced-fine-tuning-of-qwen3-asr - Polyglot-Lion offers efficient multilingual ASR tailored for Singapore's diverse languages at a fraction of the cost of larger models.
- KidsNanny: A Two-Stage Multimodal Content Moderation Pipeline Integrating Visual Classification, Object Detection, OCR, and Contextual Reasoning for Child Safety (viability: 3): https://sciencetostartup.com/paper/kidsnanny-a-two-stage-multimodal-content-moderation-pipeline-integrating-visual-classification-object-detection-ocr-and- - KidsNanny is a two-stage multimodal content moderation system designed to enhance child safety through efficient visual and text analysis.
- Enforcing Task-Specified Compliance Bounds for Humanoids via Anisotropic Lipschitz-Constrained Policies (viability: 4): https://sciencetostartup.com/paper/enforcing-task-specified-compliance-bounds-for-humanoids-via-anisotropic-lipschitz-constrained-policies - A novel anisotropic Lipschitz-constrained policy for enhancing humanoid locomotion stability and compliance in reinforcement learning.
- 360° Image Perception with MLLMs: A Comprehensive Benchmark and a Training-Free Method (viability: 7): https://sciencetostartup.com/paper/360-image-perception-with-mllms-a-comprehensive-benchmark-and-a-training-free-method - Free360 offers a training-free method for enhanced visual question answering on 360° images using a modular scene-graph approach.
- The Finetuner's Fallacy: When to Pretrain with Your Finetuning Data (viability: 5): https://sciencetostartup.com/paper/the-finetuner-s-fallacy-when-to-pretrain-with-your-finetuning-data - A novel strategy for specialized pretraining that enhances model performance in narrow domains while preserving general capabilities.
- MemX: A Local-First Long-Term Memory System for AI Assistants (viability: 7): https://sciencetostartup.com/paper/memx-a-local-first-long-term-memory-system-for-ai-assistants - MemX is a local-first long-term memory system for AI assistants that enhances retrieval stability and explainability.
- Open-Source Reproduction and Explainability Analysis of Corrective Retrieval Augmented Generation (viability: 8): https://sciencetostartup.com/paper/open-source-reproduction-and-explainability-analysis-of-corrective-retrieval-augmented-generation - An open-source implementation of Corrective Retrieval Augmented Generation that enhances robustness and explainability in RAG systems.
- SignNav: Leveraging Signage for Semantic Visual Navigation in Large-Scale Indoor Environments (viability: 7): https://sciencetostartup.com/paper/signnav-leveraging-signage-for-semantic-visual-navigation-in-large-scale-indoor-environments - SignNav enables agents to navigate large indoor environments by interpreting signage for decision-making.
- Homogeneous and Heterogeneous Consistency progressive Re-ranking for Visible-Infrared Person Re-identification (viability: 4): https://sciencetostartup.com/paper/homogeneous-and-heterogeneous-consistency-progressive-re-ranking-for-visible-infrared-person-re-identification - A novel re-ranking method for improving visible-infrared person re-identification performance.
- STARK: Spatio-Temporal Attention for Representation of Keypoints for Continuous Sign Language Recognition (viability: 4): https://sciencetostartup.com/paper/stark-spatio-temporal-attention-for-representation-of-keypoints-for-continuous-sign-language-recognition - A unified spatio-temporal attention network for efficient continuous sign language recognition.
- SQL-ASTRA: Alleviating Sparse Feedback in Agentic SQL via Column-Set Matching and Trajectory Aggregation (viability: 3): https://sciencetostartup.com/paper/sql-astra-alleviating-sparse-feedback-in-agentic-sql-via-column-set-matching-and-trajectory-aggregation - Agentic SQL enhances multi-turn Text-to-SQL tasks by improving credit assignment through innovative reward mechanisms.
- Segmentation-before-Staining Improves Structural Fidelity in Virtual IHC-to-Multiplex IF Translation (viability: 7): https://sciencetostartup.com/paper/segmentation-before-staining-improves-structural-fidelity-in-virtual-ihc-to-multiplex-if-translation - A novel virtual staining method enhances the fidelity of multiplex immunofluorescence translation by improving nuclei morphology representation.
- AI-Generated Figures in Academic Publishing: Policies, Tools, and Practical Guidelines (viability: 3): https://sciencetostartup.com/paper/ai-generated-figures-in-academic-publishing-policies-tools-and-practical-guidelines - A survey of AI-generated figure tools and guidelines for academic publishing.
- Execution-Grounded Credit Assignment for GRPO in Code Generation (viability: 3): https://sciencetostartup.com/paper/execution-grounded-credit-assignment-for-grpo-in-code-generation - Execution-Grounded Credit Assignment optimizes code generation by improving credit assignment in reinforcement learning.
- DyJR: Preserving Diversity in Reinforcement Learning with Verifiable Rewards via Dynamic Jensen-Shannon Replay (viability: 4): https://sciencetostartup.com/paper/dyjr-preserving-diversity-in-reinforcement-learning-with-verifiable-rewards-via-dynamic-jensen-shannon-replay - DyJR enhances reinforcement learning by preserving diversity in training through a novel replay mechanism.
- GATS: Gaussian Aware Temporal Scaling Transformer for Invariant 4D Spatio-Temporal Point Cloud Representation (viability: 4): https://sciencetostartup.com/paper/gats-gaussian-aware-temporal-scaling-transformer-for-invariant-4d-spatio-temporal-point-cloud-representation - GATS is a novel framework for robust 4D point cloud video understanding that addresses temporal scale bias and distributional uncertainty.
- HIPO: Instruction Hierarchy via Constrained Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/hipo-instruction-hierarchy-via-constrained-reinforcement-learning - HIPO is a novel alignment framework that enhances hierarchical instruction following in large language models through constrained reinforcement learning.
- EFF-Grasp: Energy-Field Flow Matching for Physics-Aware Dexterous Grasp Generation (viability: 7): https://sciencetostartup.com/paper/eff-grasp-energy-field-flow-matching-for-physics-aware-dexterous-grasp-generation - EFF-Grasp offers a physics-aware framework for efficient and stable dexterous grasp generation.
- NeuronSpark: A Spiking Neural Network Language Model with Selective State Space Dynamics (viability: 2): https://sciencetostartup.com/paper/neuronspark-a-spiking-neural-network-language-model-with-selective-state-space-dynamics - NeuronSpark explores a novel spiking neural network architecture for language modeling without relying on Transformers.
- Deep Adaptive Model-Based Design of Experiments (viability: 6): https://sciencetostartup.com/paper/deep-adaptive-model-based-design-of-experiments - A neural network policy for real-time adaptive design of experiments in nonlinear dynamical systems.
- Parametric Social Identity Injection and Diversification in Public Opinion Simulation (viability: 8): https://sciencetostartup.com/paper/parametric-social-identity-injection-and-diversification-in-public-opinion-simulation - A framework for enhancing diversity in public opinion simulations using large language models.
- Communication-Aware Multi-Agent Reinforcement Learning for Decentralized Cooperative UAV Deployment (viability: 3): https://sciencetostartup.com/paper/communication-aware-multi-agent-reinforcement-learning-for-decentralized-cooperative-uav-deployment - A multi-agent reinforcement learning framework for decentralized UAV deployment under communication constraints.
- Noisy Data is Destructive to Reinforcement Learning with Verifiable Rewards (viability: 3): https://sciencetostartup.com/paper/noisy-data-is-destructive-to-reinforcement-learning-with-verifiable-rewards - This research highlights the detrimental effects of noisy data on reinforcement learning with verifiable rewards, emphasizing the need for high-quality data.
- Rethinking UMM Visual Generation: Masked Modeling for Efficient Image-Only Pre-training (viability: 9): https://sciencetostartup.com/paper/rethinking-umm-visual-generation-masked-modeling-for-efficient-image-only-pre-training - IOMM revolutionizes visual generation by enabling efficient image-only pre-training for unified multimodal models.
- Answer Bubbles: Information Exposure in AI-Mediated Search (viability: 2): https://sciencetostartup.com/paper/answer-bubbles-information-exposure-in-ai-mediated-search - This research investigates biases in AI-generated search summaries and their implications for user trust.
- SIA: A Synthesize-Inject-Align Framework for Knowledge-Grounded and Secure E-commerce Search LLMs with Industrial Deployment (viability: 8): https://sciencetostartup.com/paper/sia-a-synthesize-inject-align-framework-for-knowledge-grounded-and-secure-e-commerce-search-llms-with-industrial-deploym - A framework for building knowledgeable and secure e-commerce search LLMs to enhance intent-aware recommendations.
- When Generative Augmentation Hurts: A Benchmark Study of GAN and Diffusion Models for Bias Correction in AI Classification Systems (viability: 6): https://sciencetostartup.com/paper/when-generative-augmentation-hurts-a-benchmark-study-of-gan-and-diffusion-models-for-bias-correction-in-ai-classificatio - A benchmark study revealing the pitfalls of generative augmentation for bias correction in AI classification systems.
- DualPrim: Compact 3D Reconstruction with Positive and Negative Primitives (viability: 3): https://sciencetostartup.com/paper/dualprim-compact-3d-reconstruction-with-positive-and-negative-primitives - DualPrim offers a novel framework for compact and structured 3D reconstruction using positive and negative primitives.
- SciZoom: A Large-scale Benchmark for Hierarchical Scientific Summarization across the LLM Era (viability: 9): https://sciencetostartup.com/paper/scizoom-a-large-scale-benchmark-for-hierarchical-scientific-summarization-across-the-llm-era - SciZoom is a comprehensive benchmark for hierarchical scientific summarization, analyzing the evolution of scientific writing in the LLM era.
- EPOFusion: Exposure aware Progressive Optimization Method for Infrared and Visible Image Fusion (viability: 8): https://sciencetostartup.com/paper/epofusion-exposure-aware-progressive-optimization-method-for-infrared-and-visible-image-fusion - EPOFusion is an exposure-aware model that enhances infrared and visible image fusion, particularly in overexposed regions.
- Boosting Quantitive and Spatial Awareness for Zero-Shot Object Counting (viability: 6): https://sciencetostartup.com/paper/boosting-quantitive-and-spatial-awareness-for-zero-shot-object-counting - QICA enhances zero-shot object counting by integrating quantity perception with spatial aggregation for improved accuracy.
- Social Simulacra in the Wild: AI Agent Communities on Moltbook (viability: 2): https://sciencetostartup.com/paper/social-simulacra-in-the-wild-ai-agent-communities-on-moltbook - This research analyzes the dynamics of AI-agent communities on social platforms, revealing key differences from human interactions.
- Pre-training LLM without Learning Rate Decay Enhances Supervised Fine-Tuning (viability: 2): https://sciencetostartup.com/paper/pre-training-llm-without-learning-rate-decay-enhances-supervised-fine-tuning - This paper explores a novel learning rate scheduling method for pre-training large language models to enhance their adaptability in downstream tasks.
- SWE-QA-Pro: A Representative Benchmark and Scalable Training Recipe for Repository-Level Code Understanding (viability: 7): https://sciencetostartup.com/paper/swe-qa-pro-a-representative-benchmark-and-scalable-training-recipe-for-repository-level-code-understanding - SWE-QA-Pro provides a benchmark and training recipe for improving repository-level code understanding in software engineering.
- Functorial Neural Architectures from Higher Inductive Types (viability: 4): https://sciencetostartup.com/paper/functorial-neural-architectures-from-higher-inductive-types - A novel approach to enhance compositional generalization in neural networks through functorial architectures.
- Out-of-Distribution Object Detection in Street Scenes via Synthetic Outlier Exposure and Transfer Learning (viability: 7): https://sciencetostartup.com/paper/out-of-distribution-object-detection-in-street-scenes-via-synthetic-outlier-exposure-and-transfer-learning - A unified framework for detecting out-of-distribution objects in street scenes using synthetic data and transfer learning.
- Language Models Don't Know What You Want: Evaluating Personalization in Deep Research Needs Real Users (viability: 7): https://sciencetostartup.com/paper/language-models-don-t-know-what-you-want-evaluating-personalization-in-deep-research-needs-real-users - MyScholarQA is a personalized deep research tool that infers user interests and proposes tailored actions for queries.
- SE(3)-LIO: Smooth IMU Propagation With Jointly Distributed Poses on SE(3) Manifold for Accurate and Robust LiDAR-Inertial Odometry (viability: 7): https://sciencetostartup.com/paper/se-3-lio-smooth-imu-propagation-with-jointly-distributed-poses-on-se-3-manifold-for-accurate-and-robust-lidar-inertial-o - SE(3)-LIO enhances LiDAR-inertial odometry by accurately propagating poses on the SE(3) manifold.
- Ciphertext-Policy ABE for $\mathsf{NC}^1$ Circuits with Constant-Size Ciphertexts from Succinct LWE (viability: 2): https://sciencetostartup.com/paper/ciphertext-policy-abe-for-mathsf-nc-1-circuits-with-constant-size-ciphertexts-from-succinct-lwe - A new lattice-based encryption scheme with constant-size ciphertexts for NC1 circuits.
- PathGLS: Evaluating Pathology Vision-Language Models without Ground Truth through Multi-Dimensional Consistency (viability: 9): https://sciencetostartup.com/paper/pathgls-evaluating-pathology-vision-language-models-without-ground-truth-through-multi-dimensional-consistency - PathGLS is a novel evaluation framework for pathology vision-language models that quantifies hallucination rates and robustness without ground truth.
- ASDA: Automated Skill Distillation and Adaptation for Financial Reasoning (viability: 8): https://sciencetostartup.com/paper/asda-automated-skill-distillation-and-adaptation-for-financial-reasoning - ASDA automates skill distillation for financial reasoning, enhancing LLMs without fine-tuning.
- VIGIL: Towards Edge-Extended Agentic AI for Enterprise IT Support (viability: 3): https://sciencetostartup.com/paper/vigil-towards-edge-extended-agentic-ai-for-enterprise-it-support - VIGIL is an edge-extended AI system that enhances enterprise IT support through on-device diagnosis and remediation.
- RepoReviewer: A Local-First Multi-Agent Architecture for Repository-Level Code Review (viability: 7): https://sciencetostartup.com/paper/reporeviewer-a-local-first-multi-agent-architecture-for-repository-level-code-review - RepoReviewer is a local-first multi-agent system designed for automated GitHub repository code reviews.
- Frequency Matters: Fast Model-Agnostic Data Curation for Pruning and Quantization (viability: 8): https://sciencetostartup.com/paper/frequency-matters-fast-model-agnostic-data-curation-for-pruning-and-quantization - ZipCal is a fast, model-agnostic data curation strategy for optimizing calibration data in model compression.
- Efficient LLM Serving for Agentic Workflows: A Data Systems Perspective (viability: 7): https://sciencetostartup.com/paper/efficient-llm-serving-for-agentic-workflows-a-data-systems-perspective - Helium optimizes LLM serving for agentic workflows by integrating proactive caching and cache-aware scheduling.
- NanoGS: Training-Free Gaussian Splat Simplification (viability: 7): https://sciencetostartup.com/paper/nanogs-training-free-gaussian-splat-simplification - NanoGS offers a training-free framework for efficient Gaussian Splat simplification, enhancing real-time rendering without heavy computational costs.
- Reevaluating the Intra-Modal Misalignment Hypothesis in CLIP (viability: 3): https://sciencetostartup.com/paper/reevaluating-the-intra-modal-misalignment-hypothesis-in-clip - This study reevaluates the intra-modal misalignment hypothesis in CLIP, challenging existing theories on image embedding distances.
- OneWorld: Taming Scene Generation with 3D Unified Representation Autoencoder (viability: 8): https://sciencetostartup.com/paper/oneworld-taming-scene-generation-with-3d-unified-representation-autoencoder - OneWorld is a framework for generating high-quality 3D scenes with superior cross-view consistency using a unified representation autoencoder.
- LICA: Layered Image Composition Annotations for Graphic Design Research (viability: 4): https://sciencetostartup.com/paper/lica-layered-image-composition-annotations-for-graphic-design-research - LICA is a large-scale dataset for advancing structured understanding and generation of graphic design layouts.
- Diffusion Models for Joint Audio-Video Generation (viability: 7): https://sciencetostartup.com/paper/diffusion-models-for-joint-audio-video-generation - A novel approach to joint audio-video generation using diffusion models and high-quality datasets.
- Parallel In-context Learning for Large Vision Language Models (viability: 7): https://sciencetostartup.com/paper/parallel-in-context-learning-for-large-vision-language-models - Parallel-ICL enhances large vision-language models by improving inference speed while maintaining accuracy through innovative context chunking.
- CounterRefine: Answer-Conditioned Counterevidence Retrieval for Inference-Time Knowledge Repair in Factual Question Answering (viability: 8): https://sciencetostartup.com/paper/counterrefine-answer-conditioned-counterevidence-retrieval-for-inference-time-knowledge-repair-in-factual-question-answe - CounterRefine enhances factual question answering by refining answers through evidence retrieval and validation.
- RecBundle: A Next-Generation Geometric Paradigm for Explainable Recommender Systems (viability: 5): https://sciencetostartup.com/paper/recbundle-a-next-generation-geometric-paradigm-for-explainable-recommender-systems - RecBundle introduces a geometric paradigm for explainable recommender systems that enhances user collaboration and addresses systemic bias.
- Towards the Vision-Sound-Language-Action Paradigm: The HEAR Framework for Sound-Centric Manipulation (viability: 8): https://sciencetostartup.com/paper/towards-the-vision-sound-language-action-paradigm-the-hear-framework-for-sound-centric-manipulation - HEAR is a framework for sound-centric manipulation in robotics, integrating audio, vision, and language for real-time task execution.
- Interact3D: Compositional 3D Generation of Interactive Objects (viability: 7): https://sciencetostartup.com/paper/interact3d-compositional-3d-generation-of-interactive-objects - Interact3D generates physically plausible interactive 3D compositional objects from single images.
- Structured prototype regularization for synthetic-to-real driving scene parsing (viability: 7): https://sciencetostartup.com/paper/structured-prototype-regularization-for-synthetic-to-real-driving-scene-parsing - A novel framework for enhancing driving scene parsing in autonomous vehicles by improving synthetic-to-real domain adaptation.
- A Depth-Aware Comparative Study of Euclidean and Hyperbolic Graph Neural Networks on Bitcoin Transaction Systems (viability: 4): https://sciencetostartup.com/paper/a-depth-aware-comparative-study-of-euclidean-and-hyperbolic-graph-neural-networks-on-bitcoin-transaction-systems - A comparative study of Euclidean and hyperbolic GNNs for analyzing Bitcoin transaction networks.
- Volumetrically Consistent Implicit Atlas Learning via Neural Diffeomorphic Flow for Placenta MRI (viability: 4): https://sciencetostartup.com/paper/volumetrically-consistent-implicit-atlas-learning-via-neural-diffeomorphic-flow-for-placenta-mri - A novel implicit model for volumetric registration of placenta MRI that enhances geometric fidelity and alignment.
- MDM-Prime-v2: Binary Encoding and Index Shuffling Enable Compute-optimal Scaling of Diffusion Language Models (viability: 7): https://sciencetostartup.com/paper/mdm-prime-v2-binary-encoding-and-index-shuffling-enable-compute-optimal-scaling-of-diffusion-language-models - MDM-Prime-v2 enhances diffusion language models with improved efficiency and accuracy through innovative encoding techniques.
- ClaimFlow: Tracing the Evolution of Scientific Claims in NLP (viability: 2): https://sciencetostartup.com/paper/claimflow-tracing-the-evolution-of-scientific-claims-in-nlp - ClaimFlow provides a claim-centric view of scientific claims in NLP literature, analyzing their evolution and interrelations.
- SEAHateCheck: Functional Tests for Detecting Hate Speech in Low-Resource Languages of Southeast Asia (viability: 6): https://sciencetostartup.com/paper/seahatecheck-functional-tests-for-detecting-hate-speech-in-low-resource-languages-of-southeast-asia - SEAHateCheck is a dataset and functional test suite designed to improve hate speech detection in low-resource Southeast Asian languages.
- Resource Consumption Threats in Large Language Models (viability: 2): https://sciencetostartup.com/paper/resource-consumption-threats-in-large-language-models - This survey reviews resource consumption threats in large language models to enhance efficiency and sustainability.
- Attribution Upsampling should Redistribute, Not Interpolate (viability: 3): https://sciencetostartup.com/paper/attribution-upsampling-should-redistribute-not-interpolate - A novel method for improving attribution signal fidelity in explainable AI through semantic-aware upsampling.
- Adaptive regularization parameter selection for high-dimensional inverse problems: A Bayesian approach with Tucker low-rank constraints (viability: 5): https://sciencetostartup.com/paper/adaptive-regularization-parameter-selection-for-high-dimensional-inverse-problems-a-bayesian-approach-with-tucker-low-ra - A Bayesian method using Tucker decomposition for efficient high-dimensional inverse problem solving.
- Large Reward Models: Generalizable Online Robot Reward Generation with Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/large-reward-models-generalizable-online-robot-reward-generation-with-vision-language-models - A framework that utilizes vision-language models to generate online rewards for refining robotic manipulation policies efficiently.
- ViT-AdaLA: Adapting Vision Transformers with Linear Attention (viability: 6): https://sciencetostartup.com/paper/vit-adala-adapting-vision-transformers-with-linear-attention - ViT-AdaLA adapts Vision Transformers for improved efficiency with linear attention.
- Safe Distributionally Robust Feature Selection under Covariate Shift (viability: 2): https://sciencetostartup.com/paper/safe-distributionally-robust-feature-selection-under-covariate-shift - A novel approach for distributionally robust feature selection to enhance model performance across diverse deployment environments.
- ARISE: Agent Reasoning with Intrinsic Skill Evolution in Hierarchical Reinforcement Learning (viability: 9): https://sciencetostartup.com/paper/arise-agent-reasoning-with-intrinsic-skill-evolution-in-hierarchical-reinforcement-learning - ARISE enhances mathematical reasoning in language models through a hierarchical reinforcement learning framework that evolves skills over time.
- Ultrafast Sampling-based Kinodynamic Planning via Differential Flatness (viability: 8): https://sciencetostartup.com/paper/ultrafast-sampling-based-kinodynamic-planning-via-differential-flatness - AkinoPDF is a fast parallelized kinodynamic motion planning technique that enables safe robot operation in complex environments.
- Cross-Scale Persistence Analysis of EM Side-Channels for Reference-Free Detection of Always-On Hardware Trojans (viability: 4): https://sciencetostartup.com/paper/cross-scale-persistence-analysis-of-em-side-channels-for-reference-free-detection-of-always-on-hardware-trojans - A reference-free detection framework for always-on hardware Trojans using cross-scale persistence analysis of EM side-channels.
- A Context Alignment Pre-processor for Enhancing the Coherence of Human-LLM Dialog (viability: 2): https://sciencetostartup.com/paper/a-context-alignment-pre-processor-for-enhancing-the-coherence-of-human-llm-dialog - A pre-processor to enhance coherence in human-LLM dialogue by addressing contextual misalignment.
- The Era of End-to-End Autonomy: Transitioning from Rule-Based Driving to Large Driving Models (viability: 3): https://sciencetostartup.com/paper/the-era-of-end-to-end-autonomy-transitioning-from-rule-based-driving-to-large-driving-models - This paper explores the transition to end-to-end learning systems in autonomous driving, highlighting the shift from rule-based approaches to large driving models.
- POaaS: Minimal-Edit Prompt Optimization as a Service to Lift Accuracy and Cut Hallucinations on On-Device sLLMs (viability: 4): https://sciencetostartup.com/paper/poaas-minimal-edit-prompt-optimization-as-a-service-to-lift-accuracy-and-cut-hallucinations-on-on-device-sllms - POaaS offers minimal-edit prompt optimization to enhance accuracy and reduce hallucinations in on-device small language models.
- Enhancing Linguistic Generalization of VLA: Fine-Tuning OpenVLA via Synthetic Instruction Augmentation (viability: 6): https://sciencetostartup.com/paper/enhancing-linguistic-generalization-of-vla-fine-tuning-openvla-via-synthetic-instruction-augmentation - A fine-tuning strategy for OpenVLA enhances linguistic generalization in embodied AI through synthetic instruction augmentation.
- Collaborative Temporal Feature Generation via Critic-Free Reinforcement Learning for Cross-User Sensor-Based Activity Recognition (viability: 7): https://sciencetostartup.com/paper/collaborative-temporal-feature-generation-via-critic-free-reinforcement-learning-for-cross-user-sensor-based-activity-re - CTFG leverages reinforcement learning for robust human activity recognition across diverse users using wearable sensors.
- Power Analysis for Prediction-Powered Inference (viability: 7): https://sciencetostartup.com/paper/power-analysis-for-prediction-powered-inference - A tool for determining sample size requirements in predictive modeling for statistical power analysis.
- Compact Optical Single-axis Joint Torque Sensor Using Redundant Photo-Reflectors and Quadratic-Programming Calibration (viability: 7): https://sciencetostartup.com/paper/compact-optical-single-axis-joint-torque-sensor-using-redundant-photo-reflectors-and-quadratic-programming-calibration - A non-contact photo-reflector-based joint torque sensor that enhances torque control in collaborative robots.
- Residual Stream Duality in Modern Transformer Architectures (viability: 4): https://sciencetostartup.com/paper/residual-stream-duality-in-modern-transformer-architectures - A novel perspective on optimizing Transformer architectures through residual stream duality.
- Geometry-Aligned LLM Fine-Tuning for Sequential Narrow-Opening Planning (viability: 7): https://sciencetostartup.com/paper/geometry-aligned-llm-fine-tuning-for-sequential-narrow-opening-planning - A fine-tuning framework for LLMs that enhances rigid-body motion planning through sequential narrow openings using geometric reasoning.
- Speak, Segment, Track, Navigate: An Interactive System for Video-Guided Skull-Base Surgery (viability: 8): https://sciencetostartup.com/paper/speak-segment-track-navigate-an-interactive-system-for-video-guided-skull-base-surgery - An interactive speech-guided system for real-time video-assisted skull base surgery that enhances surgical workflows.
- Interpretable Context Methodology: Folder Structure as Agentic Architecture (viability: 3): https://sciencetostartup.com/paper/interpretable-context-methodology-folder-structure-as-agentic-architecture - A novel methodology for AI agent orchestration using filesystem structure instead of complex frameworks.
- IRAM-Omega-Q: A Computational Architecture for Uncertainty Regulation in Artificial Agents (viability: 2): https://sciencetostartup.com/paper/iram-omega-q-a-computational-architecture-for-uncertainty-regulation-in-artificial-agents - IRAM-Omega-Q is a computational architecture for managing uncertainty in artificial agents through a quantum-like state representation.
- Understanding Moral Reasoning Trajectories in Large Language Models: Toward Probing-Based Explainability (viability: 4): https://sciencetostartup.com/paper/understanding-moral-reasoning-trajectories-in-large-language-models-toward-probing-based-explainability - This research explores moral reasoning trajectories in large language models to enhance explainability in ethical decision-making.
- FlatLands: Generative Floormap Completion From a Single Egocentric View (viability: 6): https://sciencetostartup.com/paper/flatlands-generative-floormap-completion-from-a-single-egocentric-view - FlatLands offers a dataset and benchmark for generating complete indoor floor maps from single-view images, enhancing navigation applications.
- The Importance of Being Smoothly Calibrated (viability: 3): https://sciencetostartup.com/paper/the-importance-of-being-smoothly-calibrated - A novel approach to smooth calibration that enhances prediction accuracy for decision-making processes.
- Safety Case Patterns for VLA-based driving systems: Insights from SimLingo (viability: 2): https://sciencetostartup.com/paper/safety-case-patterns-for-vla-based-driving-systems-insights-from-simlingo - A novel safety case design approach for Vision-Language-Action-based driving systems to ensure safe operation.
- Evaluating Agentic Optimization on Large Codebases (viability: 7): https://sciencetostartup.com/paper/evaluating-agentic-optimization-on-large-codebases - FormulaCode is a benchmark for evaluating the optimization capabilities of LLM coding agents on real-world codebases.
- RadAnnotate: Large Language Models for Efficient and Reliable Radiology Report Annotation (viability: 3): https://sciencetostartup.com/paper/radannotate-large-language-models-for-efficient-and-reliable-radiology-report-annotation - RadAnnotate leverages LLMs to enhance the efficiency of radiology report annotation through synthetic report generation and confidence-based automation.
- Mostly Text, Smart Visuals: Asymmetric Text-Visual Pruning for Large Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/mostly-text-smart-visuals-asymmetric-text-visual-pruning-for-large-vision-language-models - ATV-Pruning optimizes large vision-language models by effectively decoupling and pruning textual and visual tokens for enhanced performance.
- NLP Occupational Emergence Analysis: How Occupations Form and Evolve in Real Time -- A Zero-Assumption Method Demonstrated on AI in the US Technology Workforce, 2022-2026 (viability: 2): https://sciencetostartup.com/paper/nlp-occupational-emergence-analysis-how-occupations-form-and-evolve-in-real-time-a-zero-assumption-method-demonstrated-o - A method for detecting occupational emergence from resume data using vocabulary cohesion analysis.
- Visual Set Program Synthesizer (viability: 7): https://sciencetostartup.com/paper/visual-set-program-synthesizer - A visual program synthesis approach that enhances reasoning in visual assistants for complex queries.
- Selective Memory for Artificial Intelligence: Write-Time Gating with Hierarchical Archiving (viability: 3): https://sciencetostartup.com/paper/selective-memory-for-artificial-intelligence-write-time-gating-with-hierarchical-archiving - Introducing write-time gating for selective memory in AI, enhancing retrieval accuracy without deleting information.
- The Midas Touch in Gaze vs. Hand Pointing: Modality-Specific Failure Modes and Implications for XR Interfaces (viability: 4): https://sciencetostartup.com/paper/the-midas-touch-in-gaze-vs-hand-pointing-modality-specific-failure-modes-and-implications-for-xr-interfaces - A web-based open-source framework for improving XR interface performance through adaptive modality interventions.
- W2T: LoRA Weights Already Know What They Can Do (viability: 8): https://sciencetostartup.com/paper/w2t-lora-weights-already-know-what-they-can-do - W2T leverages LoRA weights to predict model behavior without running the base model, streamlining task adaptation for large language models.
- Determinism in the Undetermined: Deterministic Output in Charge-Conserving Continuous-Time Neuromorphic Systems with Temporal Stochasticity (viability: 2): https://sciencetostartup.com/paper/determinism-in-the-undetermined-deterministic-output-in-charge-conserving-continuous-time-neuromorphic-systems-with-temp - A theoretical framework for deterministic computation in neuromorphic systems addressing temporal stochasticity.
- Aligning Paralinguistic Understanding and Generation in Speech LLMs via Multi-Task Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/aligning-paralinguistic-understanding-and-generation-in-speech-llms-via-multi-task-reinforcement-learning - A paralinguistics-aware speech LLM that enhances emotional understanding through multi-task reinforcement learning.
- From Workflow Automation to Capability Closure: A Formal Framework for Safe and Revenue-Aware Customer Service AI (viability: 2): https://sciencetostartup.com/paper/from-workflow-automation-to-capability-closure-a-formal-framework-for-safe-and-revenue-aware-customer-service-ai - A framework addressing safety gaps in dynamic AI agent networks for customer service.
- An Agentic Evaluation Framework for AI-Generated Scientific Code in PETSc (viability: 7): https://sciencetostartup.com/paper/an-agentic-evaluation-framework-for-ai-generated-scientific-code-in-petsc - A novel framework for evaluating AI-generated scientific code in high-performance computing environments.
- UMO: Unified In-Context Learning Unlocks Motion Foundation Model Priors (viability: 9): https://sciencetostartup.com/paper/umo-unified-in-context-learning-unlocks-motion-foundation-model-priors - UMO is a unified framework that enhances text-to-motion generation by adapting pretrained models for diverse motion tasks.
- Safety is Non-Compositional: A Formal Framework for Capability-Based AI Systems (viability: 2): https://sciencetostartup.com/paper/safety-is-non-compositional-a-formal-framework-for-capability-based-ai-systems - This paper presents a formal proof regarding the non-compositional nature of safety in AI systems.
- 100x Cost & Latency Reduction: Performance Analysis of AI Query Approximation using Lightweight Proxy Models (viability: 7): https://sciencetostartup.com/paper/100x-cost-latency-reduction-performance-analysis-of-ai-query-approximation-using-lightweight-proxy-models - A lightweight proxy model approach that reduces cost and latency for AI queries in databases by over 100x.
- Robust Language Identification for Romansh Varieties (viability: 7): https://sciencetostartup.com/paper/robust-language-identification-for-romansh-varieties - A robust language identification system for distinguishing between Romansh idioms with high accuracy.
- MAC: Multi-Agent Constitution Learning (viability: 7): https://sciencetostartup.com/paper/mac-multi-agent-constitution-learning - MAC leverages multi-agent systems to optimize structured prompts for LLMs, enhancing interpretability and performance in classification tasks.
- A Comprehensive Benchmark of Histopathology Foundation Models for Kidney Histopathology (viability: 7): https://sciencetostartup.com/paper/a-comprehensive-benchmark-of-histopathology-foundation-models-for-kidney-histopathology - A benchmarking tool for evaluating histopathology foundation models specifically for kidney disease diagnostics.
- MoLoRA: Composable Specialization via Per-Token Adapter Routing (viability: 8): https://sciencetostartup.com/paper/molora-composable-specialization-via-per-token-adapter-routing - MoLoRA enables efficient per-token routing for multimodal and mixed-capability tasks, enhancing model specialization without retraining.
- Optimizing Hospital Capacity During Pandemics: A Dual-Component Framework for Strategic Patient Relocation (viability: 3): https://sciencetostartup.com/paper/optimizing-hospital-capacity-during-pandemics-a-dual-component-framework-for-strategic-patient-relocation - A dual-component framework to optimize hospital capacity through predictive analytics and simulation for patient relocation during pandemics.
- Deriving Hyperparameter Scaling Laws via Modern Optimization Theory (viability: 2): https://sciencetostartup.com/paper/deriving-hyperparameter-scaling-laws-via-modern-optimization-theory - This paper explores hyperparameter scaling laws for optimizers through modern optimization theory.
- GASP: Guided Asymmetric Self-Play For Coding LLMs (viability: 2): https://sciencetostartup.com/paper/gasp-guided-asymmetric-self-play-for-coding-llms - GASP introduces a guided approach to asymmetric self-play for enhancing coding LLMs through structured question generation.
- ExpertGen: Scalable Sim-to-Real Expert Policy Learning from Imperfect Behavior Priors (viability: 8): https://sciencetostartup.com/paper/expertgen-scalable-sim-to-real-expert-policy-learning-from-imperfect-behavior-priors - ExpertGen automates expert policy learning in simulation for scalable sim-to-real transfer in robotics.
- MobileLLM-Flash: Latency-Guided On-Device LLM Design for Industry Scale (viability: 4): https://sciencetostartup.com/paper/mobilellm-flash-latency-guided-on-device-llm-design-for-industry-scale - MobileLLM-Flash optimizes on-device large language models for efficient real-time AI experiences on resource-constrained hardware.
- A Family of LLMs Liberated from Static Vocabularies (viability: 8): https://sciencetostartup.com/paper/a-family-of-llms-liberated-from-static-vocabularies - A family of LLMs utilizing a novel hierarchical autoregressive transformer architecture to improve tokenization and language adaptability.
- Protein Design with Agent Rosetta: A Case Study for Specialized Scientific Agents (viability: 7): https://sciencetostartup.com/paper/protein-design-with-agent-rosetta-a-case-study-for-specialized-scientific-agents - Agent Rosetta is an LLM agent that enhances protein design by integrating with Rosetta software for broader design capabilities.
- Gaze-Aware Task Progression Detection Framework for Human-Robot Interaction Using RGB Cameras (viability: 7): https://sciencetostartup.com/paper/gaze-aware-task-progression-detection-framework-for-human-robot-interaction-using-rgb-cameras - A gaze-aware framework for enhancing human-robot interaction using standard RGB cameras.
- POLAR:A Per-User Association Test in Embedding Space (viability: 8): https://sciencetostartup.com/paper/polar-a-per-user-association-test-in-embedding-space - POLAR offers a novel per-user lexical association test to analyze author-level variations in social media interactions.
- BANGLASOCIALBENCH: A Benchmark for Evaluating Sociopragmatic and Cultural Alignment of LLMs in Bangladeshi Social Interaction (viability: 4): https://sciencetostartup.com/paper/banglasocialbench-a-benchmark-for-evaluating-sociopragmatic-and-cultural-alignment-of-llms-in-bangladeshi-social-interac - BANGLASOCIALBENCH is a benchmark for assessing sociopragmatic and cultural alignment of LLMs in Bangladeshi social contexts.
- Argumentative Human-AI Decision-Making: Toward AI Agents That Reason With Us, Not For Us (viability: 2): https://sciencetostartup.com/paper/argumentative-human-ai-decision-making-toward-ai-agents-that-reason-with-us-not-for-us - Developing AI agents that engage in transparent and verifiable reasoning with humans.
- Towards Fair and Robust Volumetric CT Classification via KL-Regularised Group Distributionally Robust Optimisation (viability: 7): https://sciencetostartup.com/paper/towards-fair-and-robust-volumetric-ct-classification-via-kl-regularised-group-distributionally-robust-optimisation - A framework for fair and robust classification of lung pathologies from CT scans, addressing demographic disparities.
- Do Not Leave a Gap: Hallucination-Free Object Concealment in Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/do-not-leave-a-gap-hallucination-free-object-concealment-in-vision-language-models - A novel approach to object concealment in vision-language models that reduces hallucination while maintaining scene semantics.
- Data-Local Autonomous LLM-Guided Neural Architecture Search for Multiclass Multimodal Time-Series Classification (viability: 7): https://sciencetostartup.com/paper/data-local-autonomous-llm-guided-neural-architecture-search-for-multiclass-multimodal-time-series-classification - A data-local framework for LLM-guided neural architecture search that automates model development for sensitive time-series data in healthcare.
- CTG-DB: An Ontology-Based Transformation of ClinicalTrials.gov to Enable Cross-Trial Drug Safety Analyses (viability: 3): https://sciencetostartup.com/paper/ctg-db-an-ontology-based-transformation-of-clinicaltrials-gov-to-enable-cross-trial-drug-safety-analyses - CTG-DB transforms ClinicalTrials.gov data into a standardized database for improved drug safety analysis.
- Nodule-Aligned Latent Space Learning with LLM-Driven Multimodal Diffusion for Lung Nodule Progression Prediction (viability: 7): https://sciencetostartup.com/paper/nodule-aligned-latent-space-learning-with-llm-driven-multimodal-diffusion-for-lung-nodule-progression-prediction - NAMD predicts lung nodule progression by generating follow-up CT images using patient data and a novel multimodal diffusion framework.
- Semi-Autonomous Formalization of the Vlasov-Maxwell-Landau Equilibrium (viability: 7): https://sciencetostartup.com/paper/semi-autonomous-formalization-of-the-vlasov-maxwell-landau-equilibrium - A formalization tool for mathematical proofs using AI-assisted reasoning and coding.
- Discovery of interaction and diffusion kernels in particle-to-mean-field multi-agent systems (viability: 4): https://sciencetostartup.com/paper/discovery-of-interaction-and-diffusion-kernels-in-particle-to-mean-field-multi-agent-systems - A data-driven framework for learning interaction kernels in stochastic multi-agent systems from trajectory data.
- Evaluating Causal Discovery Algorithms for Path-Specific Fairness and Utility in Healthcare (viability: 4): https://sciencetostartup.com/paper/evaluating-causal-discovery-algorithms-for-path-specific-fairness-and-utility-in-healthcare - A framework for evaluating causal discovery algorithms in healthcare to enhance fairness and utility.
- Generative Inverse Design with Abstention via Diagonal Flow Matching (viability: 7): https://sciencetostartup.com/paper/generative-inverse-design-with-abstention-via-diagonal-flow-matching - A novel generative design approach that improves accuracy and reliability in inverse design problems through Diagonal Flow Matching.
- Learning to Recall with Transformers Beyond Orthogonal Embeddings (viability: 4): https://sciencetostartup.com/paper/learning-to-recall-with-transformers-beyond-orthogonal-embeddings - A theoretical analysis of transformers for improved knowledge retrieval in LLMs.
- Machine Translation in the Wild: User Reaction to Xiaohongshu's Built-In Translation Feature (viability: 4): https://sciencetostartup.com/paper/machine-translation-in-the-wild-user-reaction-to-xiaohongshu-s-built-in-translation-feature - This paper analyzes user reactions to Xiaohongshu's translation feature, highlighting its impact on cross-cultural communication.
- VIBEPASS: Can Vibe Coders Really Pass the Vibe Check? (viability: 2): https://sciencetostartup.com/paper/vibepass-can-vibe-coders-really-pass-the-vibe-check - A study evaluating the limitations of LLMs in autonomous debugging and fault-targeted program repair.
- Sparse but not Simpler: A Multi-Level Interpretability Analysis of Vision Transformers (viability: 2): https://sciencetostartup.com/paper/sparse-but-not-simpler-a-multi-level-interpretability-analysis-of-vision-transformers - This research analyzes the relationship between weight sparsity and interpretability in Vision Transformers, revealing limited interpretability gains.
- Bayesian-guided inverse design of hyperelastic microstructures: Application to stochastic metamaterials (viability: 4): https://sciencetostartup.com/paper/bayesian-guided-inverse-design-of-hyperelastic-microstructures-application-to-stochastic-metamaterials - A Bayesian framework for efficient inverse design of hyperelastic microstructures using active learning.
- Auto Researching, not hyperparameter tuning: Convergence Analysis of 10,000 Experiments (viability: 7): https://sciencetostartup.com/paper/auto-researching-not-hyperparameter-tuning-convergence-analysis-of-10-000-experiments - A framework for LLM agents to autonomously design and optimize ML experiments through genuine architecture search.
- The Agentic Researcher: A Practical Guide to AI-Assisted Research in Mathematics and Machine Learning (viability: 9): https://sciencetostartup.com/paper/the-agentic-researcher-a-practical-guide-to-ai-assisted-research-in-mathematics-and-machine-learning - An open-source framework that transforms AI coding agents into autonomous research assistants for mathematics and machine learning.
- Prompt Engineering for Scale Development in Generative Psychometrics (viability: 5): https://sciencetostartup.com/paper/prompt-engineering-for-scale-development-in-generative-psychometrics - AI-GENIE enhances personality assessment item generation through advanced prompt engineering strategies.
- Game-Theory-Assisted Reinforcement Learning for Border Defense: Early Termination based on Analytical Solutions (viability: 6): https://sciencetostartup.com/paper/game-theory-assisted-reinforcement-learning-for-border-defense-early-termination-based-on-analytical-solutions - A hybrid game-theory and reinforcement learning approach that enhances training efficiency for border defense strategies.
- Agent-based imitation dynamics can yield efficiently compressed population-level vocabularies (viability: 2): https://sciencetostartup.com/paper/agent-based-imitation-dynamics-can-yield-efficiently-compressed-population-level-vocabularies - This paper explores the theoretical foundations of language evolution through agent-based dynamics and information theory.
- Federated Learning for Privacy-Preserving Medical AI (viability: 6): https://sciencetostartup.com/paper/federated-learning-for-privacy-preserving-medical-ai - A novel federated learning approach for privacy-preserving Alzheimer's disease classification using MRI data.
- The Internet of Physical AI Agents: Interoperability, Longevity, and the Cost of Getting It Wrong (viability: 2): https://sciencetostartup.com/paper/the-internet-of-physical-ai-agents-interoperability-longevity-and-the-cost-of-getting-it-wrong - This paper discusses the challenges and principles for developing resilient and trustworthy physical AI agents in critical domains.
- COGNAC at SemEval-2026 Task 5: LLM Ensembles for Human-Level Word Sense Plausibility Rating in Challenging Narratives (viability: 3): https://sciencetostartup.com/paper/cognac-at-semeval-2026-task-5-llm-ensembles-for-human-level-word-sense-plausibility-rating-in-challenging-narratives - A system for evaluating word sense plausibility using LLM ensembles in narrative contexts.
- Temporal Fact Conflicts in LLMs: Reproducibility Insights from Unifying DYNAMICQA and MULAN (viability: 5): https://sciencetostartup.com/paper/temporal-fact-conflicts-in-llms-reproducibility-insights-from-unifying-dynamicqa-and-mulan - This research investigates how temporal fact conflicts in LLMs can be resolved through dataset design and model size adjustments.
- AsgardBench - Evaluating Visually Grounded Interactive Planning Under Minimal Feedback (viability: 4): https://sciencetostartup.com/paper/asgardbench-evaluating-visually-grounded-interactive-planning-under-minimal-feedback - AsgardBench is a benchmark for evaluating visually grounded interactive planning in AI agents under minimal feedback conditions.
- EvoIQA - Explaining Image Distortions with Evolved White-Box Logic (viability: 7): https://sciencetostartup.com/paper/evoiqa-explaining-image-distortions-with-evolved-white-box-logic - EvoIQA offers an explainable framework for image quality assessment using evolved mathematical formulas.
- PhasorFlow: A Python Library for Unit Circle Based Computing (viability: 8): https://sciencetostartup.com/paper/phasorflow-a-python-library-for-unit-circle-based-computing - PhasorFlow is an open-source Python library for efficient unit circle-based computing, enabling advanced predictive learning through innovative algorithms.
- Resilience Meets Autonomy: Governing Embodied AI in Critical Infrastructure (viability: 2): https://sciencetostartup.com/paper/resilience-meets-autonomy-governing-embodied-ai-in-critical-infrastructure - This paper explores the governance of embodied AI in critical infrastructure to enhance resilience and autonomy.
- Electrodermal Activity as a Unimodal Signal for Aerobic Exercise Detection in Wearable Sensors (viability: 4): https://sciencetostartup.com/paper/electrodermal-activity-as-a-unimodal-signal-for-aerobic-exercise-detection-in-wearable-sensors - A study on using electrodermal activity to detect aerobic exercise through wearable sensors.
- Counteractive RL: Rethinking Core Principles for Efficient and Scalable Deep Reinforcement Learning (viability: 2): https://sciencetostartup.com/paper/counteractive-rl-rethinking-core-principles-for-efficient-and-scalable-deep-reinforcement-learning - A novel paradigm for efficient and scalable deep reinforcement learning through counteractive actions.
- Evaluating Black-Box Vulnerabilities with Wasserstein-Constrained Data Perturbations (viability: 5): https://sciencetostartup.com/paper/evaluating-black-box-vulnerabilities-with-wasserstein-constrained-data-perturbations - A method to enhance the explainability of black-box ML models using Wasserstein-constrained data perturbations.
- Interpretative Interfaces: Designing for AI-Mediated Reading Practices and the Knowledge Commons (viability: 2): https://sciencetostartup.com/paper/interpretative-interfaces-designing-for-ai-mediated-reading-practices-and-the-knowledge-commons - A design research project proposing interactive interfaces for engaging with language models' internal representations.
- Self-supervised Disentanglement of Disease Effects from Aging in 3D Medical Shapes (viability: 8): https://sciencetostartup.com/paper/self-supervised-disentanglement-of-disease-effects-from-aging-in-3d-medical-shapes - A framework for disentangling disease effects from aging in 3D medical shapes to enhance biomarker development.
- Regularized Latent Dynamics Prediction is a Strong Baseline For Behavioral Foundation Models (viability: 4): https://sciencetostartup.com/paper/regularized-latent-dynamics-prediction-is-a-strong-baseline-for-behavioral-foundation-models - Regularized Latent Dynamics Prediction enhances feature diversity for zero-shot reinforcement learning.
- FlashSampling: Fast and Memory-Efficient Exact Sampling (viability: 8): https://sciencetostartup.com/paper/flashsampling-fast-and-memory-efficient-exact-sampling - FlashSampling optimizes large-vocabulary decoding by integrating exact sampling directly into the matrix multiplication process, significantly reducing memory traffic and processing time.
- Algorithmic Trading Strategy Development and Optimisation (viability: 6): https://sciencetostartup.com/paper/algorithmic-trading-strategy-development-and-optimisation - An enhanced algorithmic trading strategy that combines technical indicators and sentiment analysis for improved trading performance.
- FEEL (Force-Enhanced Egocentric Learning): A Dataset for Physical Action Understanding (viability: 7): https://sciencetostartup.com/paper/feel-force-enhanced-egocentric-learning-a-dataset-for-physical-action-understanding - FEEL is a novel dataset that enhances physical action understanding through force-synchronized egocentric video data.
- Informationally Compressive Anonymization: Non-Degrading Sensitive Input Protection for Privacy-Preserving Supervised Machine Learning (viability: 3): https://sciencetostartup.com/paper/informationally-compressive-anonymization-non-degrading-sensitive-input-protection-for-privacy-preserving-supervised-mac - A novel privacy-preserving ML framework that ensures strong privacy guarantees without degrading performance.
- When Stability Fails: Hidden Failure Modes Of LLMS in Data-Constrained Scientific Decision-Making (viability: 2): https://sciencetostartup.com/paper/when-stability-fails-hidden-failure-modes-of-llms-in-data-constrained-scientific-decision-making - This paper critiques the stability of LLMs in scientific decision-making, emphasizing the need for ground-truth validation.
- Persona-Conditioned Risk Behavior in Large Language Models: A Simulated Gambling Study with GPT-4.1 (viability: 2): https://sciencetostartup.com/paper/persona-conditioned-risk-behavior-in-large-language-models-a-simulated-gambling-study-with-gpt-4-1 - This study explores the risk behaviors of LLMs in gambling scenarios based on socioeconomic personas.
- Robust Dynamic Object Detection in Cluttered Indoor Scenes via Learned Spatiotemporal Cues (viability: 3): https://sciencetostartup.com/paper/robust-dynamic-object-detection-in-cluttered-indoor-scenes-via-learned-spatiotemporal-cues - A novel LiDAR-only framework for robust dynamic object detection in cluttered indoor environments.
- Beyond the Embedding Bottleneck: Adaptive Retrieval-Augmented 3D CT Report Generation (viability: 9): https://sciencetostartup.com/paper/beyond-the-embedding-bottleneck-adaptive-retrieval-augmented-3d-ct-report-generation - AdaRAG-CT enhances automated radiology report generation by overcoming visual representation bottlenecks with adaptive retrieval techniques.
- Hypothesis Class Determines Explanation: Why Accurate Models Disagree on Feature Attribution (viability: 4): https://sciencetostartup.com/paper/hypothesis-class-determines-explanation-why-accurate-models-disagree-on-feature-attribution - A study revealing how different model classes affect feature attribution in explainable AI.
- Conflict-Aware Multimodal Fusion for Ambivalence and Hesitancy Recognition (viability: 7): https://sciencetostartup.com/paper/conflict-aware-multimodal-fusion-for-ambivalence-and-hesitancy-recognition - ConflictAwareAH is a multimodal framework for recognizing ambivalence and hesitancy in clinical settings by analyzing conflicting signals from video, audio, and text.
- Longitudinal Risk Prediction in Mammography with Privileged History Distillation (viability: 7): https://sciencetostartup.com/paper/longitudinal-risk-prediction-in-mammography-with-privileged-history-distillation - A novel method for breast cancer risk prediction that utilizes current mammography exams while leveraging historical data during training.
- ModTrack: Sensor-Agnostic Multi-View Tracking via Identity-Informed PHD Filtering with Covariance Propagation (viability: 7): https://sciencetostartup.com/paper/modtrack-sensor-agnostic-multi-view-tracking-via-identity-informed-phd-filtering-with-covariance-propagation - ModTrack is a modular, sensor-agnostic multi-view tracking system that enhances identity consistency and generalization across various sensor modalities.
- Feed-forward Gaussian Registration for Head Avatar Creation and Editing (viability: 9): https://sciencetostartup.com/paper/feed-forward-gaussian-registration-for-head-avatar-creation-and-editing - MATCH enables rapid creation and editing of personalized head avatars using a novel multi-view Gaussian registration method.
- Don't Trust Stubborn Neighbors: A Security Framework for Agentic Networks (viability: 2): https://sciencetostartup.com/paper/don-t-trust-stubborn-neighbors-a-security-framework-for-agentic-networks - A theoretical framework addressing security risks in LLM-based multi-agent systems.
- Mask Is What DLLM Needs: A Masked Data Training Paradigm for Diffusion LLMs (viability: 8): https://sciencetostartup.com/paper/mask-is-what-dllm-needs-a-masked-data-training-paradigm-for-diffusion-llms - A novel masked data training paradigm that enhances reasoning in diffusion language models through information density-driven scheduling.
- Time-Aware Prior Fitted Networks for Zero-Shot Forecasting with Exogenous Variables (viability: 7): https://sciencetostartup.com/paper/time-aware-prior-fitted-networks-for-zero-shot-forecasting-with-exogenous-variables - ApolloPFN enhances time series forecasting by incorporating exogenous variables for improved accuracy.
- Evolving Contextual Safety in Multi-Modal Large Language Models via Inference-Time Self-Reflective Memory (viability: 8): https://sciencetostartup.com/paper/evolving-contextual-safety-in-multi-modal-large-language-models-via-inference-time-self-reflective-memory - EchoSafe enhances safety in multi-modal large language models by leveraging a self-reflective memory framework for contextual understanding.
- Prose2Policy (P2P): A Practical LLM Pipeline for Translating Natural-Language Access Policies into Executable Rego (viability: 7): https://sciencetostartup.com/paper/prose2policy-p2p-a-practical-llm-pipeline-for-translating-natural-language-access-policies-into-executable-rego - Prose2Policy is a tool that translates natural-language access control policies into executable Rego code for enhanced policy enforcement.
- CUBE: A Standard for Unifying Agent Benchmarks (viability: 4): https://sciencetostartup.com/paper/cube-a-standard-for-unifying-agent-benchmarks - CUBE is a universal protocol standard for unifying agent benchmarks to enhance research productivity.
- OMNIFLOW: A Physics-Grounded Multimodal Agent for Generalized Scientific Reasoning (viability: 7): https://sciencetostartup.com/paper/omniflow-a-physics-grounded-multimodal-agent-for-generalized-scientific-reasoning - OMNIFLOW is a neuro-symbolic architecture that enhances LLMs with physical reasoning for scientific applications.
- Emergent Dexterity via Diverse Resets and Large-Scale Reinforcement Learning (viability: 8): https://sciencetostartup.com/paper/emergent-dexterity-via-diverse-resets-and-large-scale-reinforcement-learning - A scalable framework for robust reinforcement learning in dexterous manipulation tasks using minimal human input.
- Learnability with Partial Labels and Adaptive Nearest Neighbors (viability: 3): https://sciencetostartup.com/paper/learnability-with-partial-labels-and-adaptive-nearest-neighbors - A mathematical characterization of partial labels learning with a new adaptive nearest-neighbors algorithm.
- Parallelised Differentiable Straightest Geodesics for 3D Meshes (viability: 8): https://sciencetostartup.com/paper/parallelised-differentiable-straightest-geodesics-for-3d-meshes - A parallel GPU implementation for differentiable geodesics on 3D meshes, enhancing learning and optimization pipelines.
- Domain Adaptation Without the Compute Burden for Efficient Whole Slide Image Analysis (viability: 6): https://sciencetostartup.com/paper/domain-adaptation-without-the-compute-burden-for-efficient-whole-slide-image-analysis - EfficientWSI (eWSI) integrates PEFT and MIL for efficient whole slide image analysis, enhancing tumor detection and classification.
- Morphemes Without Borders: Evaluating Root-Pattern Morphology in Arabic Tokenizers and LLMs (viability: 2): https://sciencetostartup.com/paper/morphemes-without-borders-evaluating-root-pattern-morphology-in-arabic-tokenizers-and-llms - This research evaluates the effectiveness of LLMs in handling Arabic root-pattern morphology.
- CorrectionPlanner: Self-Correction Planner with Reinforcement Learning in Autonomous Driving (viability: 3): https://sciencetostartup.com/paper/correctionplanner-self-correction-planner-with-reinforcement-learning-in-autonomous-driving - CorrectionPlanner enhances autonomous driving safety through a self-correction mechanism in motion planning.
- CLRNet: Targetless Extrinsic Calibration for Camera, Lidar and 4D Radar Using Deep Learning (viability: 7): https://sciencetostartup.com/paper/clrnet-targetless-extrinsic-calibration-for-camera-lidar-and-4d-radar-using-deep-learning - CLRNet offers a novel deep learning solution for accurate extrinsic calibration of camera, lidar, and 4D radar sensors.
- Simulation Distillation: Pretraining World Models in Simulation for Rapid Real-World Adaptation (viability: 8): https://sciencetostartup.com/paper/simulation-distillation-pretraining-world-models-in-simulation-for-rapid-real-world-adaptation - SimDist enables rapid real-world adaptation in robotics by distilling structural priors from simulation for efficient planning.
- You've Got a Golden Ticket: Improving Generative Robot Policies With A Single Noise Vector (viability: 8): https://sciencetostartup.com/paper/you-ve-got-a-golden-ticket-improving-generative-robot-policies-with-a-single-noise-vector - A novel method to enhance generative robot policies using a constant noise vector, improving performance across multiple tasks.
- Towards Generalizable Robotic Manipulation in Dynamic Environments (viability: 8): https://sciencetostartup.com/paper/towards-generalizable-robotic-manipulation-in-dynamic-environments - A dynamic-aware robotic manipulation system equipped with PUMA architecture for enhanced adaptability in fast-paced environments.
- Mixture-of-Depths Attention (viability: 8): https://sciencetostartup.com/paper/mixture-of-depths-attention - Mixture-of-depths attention enhances large language models by improving feature recovery in deeper layers while maintaining efficiency.
- Look Before Acting: Enhancing Vision Foundation Representations for Vision-Language-Action Models (viability: 3): https://sciencetostartup.com/paper/look-before-acting-enhancing-vision-foundation-representations-for-vision-language-action-models - DeepVision-VLA enhances visual representations in Vision-Language-Action models for improved robotic manipulation.
- HorizonMath: Measuring AI Progress Toward Mathematical Discovery with Automatic Verification (viability: 7): https://sciencetostartup.com/paper/horizonmath-measuring-ai-progress-toward-mathematical-discovery-with-automatic-verification - HorizonMath is an open-source benchmark for evaluating AI's capability in solving unsolved mathematical problems with automated verification.
- GlyphPrinter: Region-Grouped Direct Preference Optimization for Glyph-Accurate Visual Text Rendering (viability: 4): https://sciencetostartup.com/paper/glyphprinter-region-grouped-direct-preference-optimization-for-glyph-accurate-visual-text-rendering - GlyphPrinter enhances visual text rendering accuracy using region-based preference optimization.
- Mechanistic Origin of Moral Indifference in Language Models (viability: 5): https://sciencetostartup.com/paper/mechanistic-origin-of-moral-indifference-in-language-models - This research addresses moral indifference in LLMs by aligning latent representations with moral vectors to enhance moral reasoning.
- Tri-Prompting: Video Diffusion with Unified Control over Scene, Subject, and Motion (viability: 8): https://sciencetostartup.com/paper/tri-prompting-video-diffusion-with-unified-control-over-scene-subject-and-motion - Tri-Prompting offers a unified framework for customizable video content creation with precise control over scene, subject, and motion.
- HSImul3R: Physics-in-the-Loop Reconstruction of Simulation-Ready Human-Scene Interactions (viability: 7): https://sciencetostartup.com/paper/hsimul3r-physics-in-the-loop-reconstruction-of-simulation-ready-human-scene-interactions - HSImul3R offers a novel framework for stable, simulation-ready 3D reconstruction of human-scene interactions using physics-informed optimization.
- Code-A1: Adversarial Evolving of Code LLM and Test LLM via Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/code-a1-adversarial-evolving-of-code-llm-and-test-llm-via-reinforcement-learning - Code-A1 is an adversarial co-evolution framework that optimizes code and test generation using reinforcement learning.
- Differential Privacy for Network Connectedness Indices (viability: 4): https://sciencetostartup.com/paper/differential-privacy-for-network-connectedness-indices - A method for releasing network connectedness indices while ensuring differential privacy.
- Do Metrics for Counterfactual Explanations Align with User Perception? (viability: 3): https://sciencetostartup.com/paper/do-metrics-for-counterfactual-explanations-align-with-user-perception - This study critiques existing metrics for counterfactual explanations in AI, highlighting their misalignment with user perceptions.
- Perception-Aware Autonomous Exploration in Feature-Limited Environments (viability: 3): https://sciencetostartup.com/paper/perception-aware-autonomous-exploration-in-feature-limited-environments - A perception-aware framework for UAVs that enhances exploration efficiency in feature-limited environments.
- EAAE: Energy-Aware Autonomous Exploration for UAVs in Unknown 3D Environments (viability: 3): https://sciencetostartup.com/paper/eaae-energy-aware-autonomous-exploration-for-uavs-in-unknown-3d-environments - Energy-Aware Autonomous Exploration (EAAE) optimizes UAV exploration by minimizing energy consumption while maintaining exploration efficiency.
- Fast SAM 3D Body: Accelerating SAM 3D Body for Real-Time Full-Body Human Mesh Recovery (viability: 8): https://sciencetostartup.com/paper/fast-sam-3d-body-accelerating-sam-3d-body-for-real-time-full-body-human-mesh-recovery - Fast SAM 3D Body accelerates real-time full-body human mesh recovery for interactive applications.
- From Passive Observer to Active Critic: Reinforcement Learning Elicits Process Reasoning for Robotic Manipulation (viability: 8): https://sciencetostartup.com/paper/from-passive-observer-to-active-critic-reinforcement-learning-elicits-process-reasoning-for-robotic-manipulation - PRIMO R1 transforms video MLLMs into active critics for enhanced robotic manipulation through process reasoning.
- SmartSearch: How Ranking Beats Structure for Conversational Memory Retrieval (viability: 8): https://sciencetostartup.com/paper/smartsearch-how-ranking-beats-structure-for-conversational-memory-retrieval - SmartSearch revolutionizes conversational memory retrieval by using a deterministic pipeline that outperforms traditional LLM-based methods.
- AC-Foley: Reference-Audio-Guided Video-to-Audio Synthesis with Acoustic Transfer (viability: 3): https://sciencetostartup.com/paper/ac-foley-reference-audio-guided-video-to-audio-synthesis-with-acoustic-transfer - AC-Foley is an audio-conditioned model for precise video-to-audio synthesis that overcomes text-based limitations.
- Robust and Computationally Efficient Linear Contextual Bandits under Adversarial Corruption and Heavy-Tailed Noise (viability: 3): https://sciencetostartup.com/paper/robust-and-computationally-efficient-linear-contextual-bandits-under-adversarial-corruption-and-heavy-tailed-noise - This paper presents a new algorithm for linear contextual bandits that improves computational efficiency under adversarial conditions.
- OpenSeeker: Democratizing Frontier Search Agents by Fully Open-Sourcing Training Data (viability: 9): https://sciencetostartup.com/paper/openseeker-democratizing-frontier-search-agents-by-fully-open-sourcing-training-data - Fully open-source search agent democratizing high-performance frontier search through open data and code.
- Effective Distillation to Hybrid xLSTM Architectures (viability: 3): https://sciencetostartup.com/paper/effective-distillation-to-hybrid-xlstm-architectures - A novel distillation pipeline for xLSTM architectures aiming for lossless performance compared to large language models.
- Computational Concept of the Psyche (viability: 2): https://sciencetostartup.com/paper/computational-concept-of-the-psyche - A theoretical framework for modeling artificial general intelligence based on cognitive architecture.
- Physics-Informed Neural Systems for the Simulation of EUV Electromagnetic Wave Diffraction from a Lithography Mask (viability: 4): https://sciencetostartup.com/paper/physics-informed-neural-systems-for-the-simulation-of-euv-electromagnetic-wave-diffraction-from-a-lithography-mask - A hybrid neural operator for efficient simulation of EUV electromagnetic wave diffraction from lithography masks.
- Grounding World Simulation Models in a Real-World Metropolis (viability: 7): https://sciencetostartup.com/paper/grounding-world-simulation-models-in-a-real-world-metropolis - Seoul World Model generates realistic urban videos by grounding in real city data.
- Low-Complexity and Consistent Graphon Estimation from Multiple Networks (viability: 4): https://sciencetostartup.com/paper/low-complexity-and-consistent-graphon-estimation-from-multiple-networks - A low-complexity graphon estimator that improves accuracy and efficiency in network analysis.
- Unbiased and Biased Variance-Reduced Forward-Reflected-Backward Splitting Methods for Stochastic Composite Inclusions (viability: 2): https://sciencetostartup.com/paper/unbiased-and-biased-variance-reduced-forward-reflected-backward-splitting-methods-for-stochastic-composite-inclusions - This paper presents new variance-reduction techniques for solving stochastic composite inclusions using forward-reflected-backward splitting methods.
- Severe Domain Shift in Skeleton-Based Action Recognition:A Study of Uncertainty Failure in Real-World Gym Environments (viability: 4): https://sciencetostartup.com/paper/severe-domain-shift-in-skeleton-based-action-recognition-a-study-of-uncertainty-failure-in-real-world-gym-environments - A study on improving safety in skeleton-based action recognition through uncertainty analysis and a novel gating mechanism.
- Co-Design of Memory-Storage Systems for Workload Awareness with Interpretable Models (viability: 4): https://sciencetostartup.com/paper/co-design-of-memory-storage-systems-for-workload-awareness-with-interpretable-models - A co-design framework for optimizing memory-storage systems using interpretable machine learning models.
- Mamba-3: Improved Sequence Modeling using State Space Principles (viability: 4): https://sciencetostartup.com/paper/mamba-3-improved-sequence-modeling-using-state-space-principles - Mamba-3 enhances sequence modeling efficiency with state space principles for improved LLM performance.
- Estimating Staged Event Tree Models via Hierarchical Clustering on the Simplex (viability: 2): https://sciencetostartup.com/paper/estimating-staged-event-tree-models-via-hierarchical-clustering-on-the-simplex - A new framework for estimating staged tree models using hierarchical clustering on the probability simplex.
- Lore: Repurposing Git Commit Messages as a Structured Knowledge Protocol for AI Coding Agents (viability: 3): https://sciencetostartup.com/paper/lore-repurposing-git-commit-messages-as-a-structured-knowledge-protocol-for-ai-coding-agents - Lore transforms git commit messages into structured decision records for AI coding agents.
- Predictive Uncertainty in Short-Term PV Forecasting under Missing Data: A Multiple Imputation Approach (viability: 5): https://sciencetostartup.com/paper/predictive-uncertainty-in-short-term-pv-forecasting-under-missing-data-a-multiple-imputation-approach - A framework for improving short-term photovoltaic power forecasting by incorporating uncertainty from missing data.
- The PokeAgent Challenge: Competitive and Long-Context Learning at Scale (viability: 8): https://sciencetostartup.com/paper/the-pokeagent-challenge-competitive-and-long-context-learning-at-scale - The PokeAgent Challenge is a competitive benchmark for AI decision-making in Pokemon battles and RPGs, fostering advancements in RL and LLM research.
- Panoramic Affordance Prediction (viability: 7): https://sciencetostartup.com/paper/panoramic-affordance-prediction - PAP introduces a novel framework for affordance prediction using 360-degree imagery to enhance embodied AI.
- Anatomy of a Lie: A Multi-Stage Diagnostic Framework for Tracing Hallucinations in Vision-Language Models (viability: 3): https://sciencetostartup.com/paper/anatomy-of-a-lie-a-multi-stage-diagnostic-framework-for-tracing-hallucinations-in-vision-language-models - A framework for diagnosing hallucinations in Vision-Language Models through cognitive trajectory analysis.
- Learning Latent Proxies for Controllable Single-Image Relighting (viability: 7): https://sciencetostartup.com/paper/learning-latent-proxies-for-controllable-single-image-relighting - LightCtrl enables precise single-image relighting by integrating physical priors for enhanced control over illumination changes.
- Self-Distillation of Hidden Layers for Self-Supervised Representation Learning (viability: 4): https://sciencetostartup.com/paper/self-distillation-of-hidden-layers-for-self-supervised-representation-learning - Bootleg is a self-supervised learning method that enhances feature extraction by predicting latent representations from multiple hidden layers.
- Can LLMs Model Incorrect Student Reasoning? A Case Study on Distractor Generation (viability: 3): https://sciencetostartup.com/paper/can-llms-model-incorrect-student-reasoning-a-case-study-on-distractor-generation - This research analyzes how LLMs can generate plausible distractors for educational assessments by modeling student misconceptions.
- Kimodo: Scaling Controllable Human Motion Generation (viability: 7): https://sciencetostartup.com/paper/kimodo-scaling-controllable-human-motion-generation - Kimodo is a controllable motion generation model that synthesizes high-quality human motion from intuitive inputs.
- InterveneBench: Benchmarking LLMs for Intervention Reasoning and Causal Study Design in Real Social Systems (viability: 8): https://sciencetostartup.com/paper/intervenebench-benchmarking-llms-for-intervention-reasoning-and-causal-study-design-in-real-social-systems - InterveneBench benchmarks LLMs for intervention reasoning in social science, enhancing causal study design.
- Bridging Local and Global Knowledge: Cascaded Mixture-of-Experts Learning for Near-Shortest Path Routing (viability: 7): https://sciencetostartup.com/paper/bridging-local-and-global-knowledge-cascaded-mixture-of-experts-learning-for-near-shortest-path-routing - A modular Cascaded Mixture of Experts model for efficient near-shortest path routing in complex networks.
- DOT: Dynamic Knob Selection and Online Sampling for Automated Database Tuning (viability: 3): https://sciencetostartup.com/paper/dot-dynamic-knob-selection-and-online-sampling-for-automated-database-tuning - DOT is an automated database tuning algorithm that optimizes performance by dynamically selecting influential parameters.
- Vib2ECG: A Paired Chest-Lead SCG-ECG Dataset and Benchmark for ECG Reconstruction (viability: 7): https://sciencetostartup.com/paper/vib2ecg-a-paired-chest-lead-scg-ecg-dataset-and-benchmark-for-ecg-reconstruction - Vib2ECG offers a novel dataset and benchmark for reconstructing ECG from low-cost vibrational signals, enabling mobile ECG monitoring.
- Optimal control of differentially flat underactuated planar robots in the perspective of oscillation mitigation (viability: 2): https://sciencetostartup.com/paper/optimal-control-of-differentially-flat-underactuated-planar-robots-in-the-perspective-of-oscillation-mitigation - This paper explores optimal control strategies for underactuated robots to mitigate oscillations during trajectory tracking.
- Are Dilemmas and Conflicts in LLM Alignment Solvable? A View from Priority Graph (viability: 2): https://sciencetostartup.com/paper/are-dilemmas-and-conflicts-in-llm-alignment-solvable-a-view-from-priority-graph - This paper explores the complexities of aligning large language models amidst conflicting priorities and proposes a verification mechanism to enhance robustness.
- Building Trust in PINNs: Error Estimation through Finite Difference Methods (viability: 6): https://sciencetostartup.com/paper/building-trust-in-pinns-error-estimation-through-finite-difference-methods - A method for estimating errors in physics-informed neural networks to enhance trust and interpretability in their predictions.
- Clinically Aware Synthetic Image Generation for Concept Coverage in Chest X-ray Models (viability: 7): https://sciencetostartup.com/paper/clinically-aware-synthetic-image-generation-for-concept-coverage-in-chest-x-ray-models - CARS is a synthetic image generation framework that enhances chest X-ray models by improving robustness and clinical feature coverage.
- SlovKE: A Large-Scale Dataset and LLM Evaluation for Slovak Keyphrase Extraction (viability: 8): https://sciencetostartup.com/paper/slovke-a-large-scale-dataset-and-llm-evaluation-for-slovak-keyphrase-extraction - SlovKE provides a large-scale dataset and LLM evaluation for keyphrase extraction in Slovak, addressing a critical gap in low-resource language processing.
- Beyond the Covariance Trap: Unlocking Generalization in Same-Subject Knowledge Editing for Large Language Models (viability: 5): https://sciencetostartup.com/paper/beyond-the-covariance-trap-unlocking-generalization-in-same-subject-knowledge-editing-for-large-language-models - RoSE enhances knowledge editing in LLMs by improving instruction-following capabilities through geometric alignment.
- ViX-Ray: A Vietnamese Chest X-Ray Dataset for Vision-Language Models (viability: 5): https://sciencetostartup.com/paper/vix-ray-a-vietnamese-chest-x-ray-dataset-for-vision-language-models - ViX-Ray is a specialized dataset aimed at enhancing vision-language models for Vietnamese chest X-ray analysis.
- FreeTalk: Emotional Topology-Free 3D Talking Heads (viability: 6): https://sciencetostartup.com/paper/freetalk-emotional-topology-free-3d-talking-heads - FreeTalk enables emotion-driven 3D facial animation on arbitrary topology meshes without template constraints.
- Not All Invariants Are Equal: Curating Training Data to Accelerate Program Verification with SLMs (viability: 7): https://sciencetostartup.com/paper/not-all-invariants-are-equal-curating-training-data-to-accelerate-program-verification-with-slms - Wonda is a data curation pipeline that enhances training data for program verification using Small Language Models.
- Federated Learning of Binary Neural Networks: Enabling Low-Cost Inference (viability: 7): https://sciencetostartup.com/paper/federated-learning-of-binary-neural-networks-enabling-low-cost-inference - FedBNN is a federated learning framework that enables low-cost inference with binary neural networks, optimizing memory and computational efficiency.
- Seeking SOTA: Time-Series Forecasting Must Adopt Taxonomy-Specific Evaluation to Dispel Illusory Gains (viability: 3): https://sciencetostartup.com/paper/seeking-sota-time-series-forecasting-must-adopt-taxonomy-specific-evaluation-to-dispel-illusory-gains - This paper critiques the evaluation methods in time-series forecasting, advocating for a more rigorous approach to benchmark datasets.
- Understanding Reasoning in LLMs through Strategic Information Allocation under Uncertainty (viability: 4): https://sciencetostartup.com/paper/understanding-reasoning-in-llms-through-strategic-information-allocation-under-uncertainty - A framework to enhance reasoning in LLMs by externalizing uncertainty for improved control actions.
- Real-Time Oriented Object Detection Transformer in Remote Sensing Images (viability: 8): https://sciencetostartup.com/paper/real-time-oriented-object-detection-transformer-in-remote-sensing-images - A real-time oriented object detection transformer that improves angle representation and training stability for remote sensing images.
- Grokking as a Variance-Limited Phase Transition: Spectral Gating and the Epsilon-Stability Threshold (viability: 2): https://sciencetostartup.com/paper/grokking-as-a-variance-limited-phase-transition-spectral-gating-and-the-epsilon-stability-threshold - This paper explores the dynamics of AdamW in relation to grokking and generalization in optimization.
- RSGen: Enhancing Layout-Driven Remote Sensing Image Generation with Diverse Edge Guidance (viability: 8): https://sciencetostartup.com/paper/rsgen-enhancing-layout-driven-remote-sensing-image-generation-with-diverse-edge-guidance - RSGen enhances layout-driven remote sensing image generation with diverse edge guidance for improved control and accuracy.
- Talk, Evaluate, Diagnose: User-aware Agent Evaluation with Automated Error Analysis (viability: 7): https://sciencetostartup.com/paper/talk-evaluate-diagnose-user-aware-agent-evaluation-with-automated-error-analysis - The TED framework enhances agent evaluation by incorporating user roles and automated error analysis for improved performance insights.
- TabKD: Tabular Knowledge Distillation through Interaction Diversity of Learned Feature Bins (viability: 7): https://sciencetostartup.com/paper/tabkd-tabular-knowledge-distillation-through-interaction-diversity-of-learned-feature-bins - TabKD offers a novel approach to data-free knowledge distillation for tabular data by focusing on interaction diversity.
- ViFeEdit: A Video-Free Tuner of Your Video Diffusion Transformer (viability: 8): https://sciencetostartup.com/paper/vifeedit-a-video-free-tuner-of-your-video-diffusion-transformer - ViFeEdit is a video-free tuning framework that enables controllable video generation and editing using only 2D images.
- Seeing Beyond: Extrapolative Domain Adaptive Panoramic Segmentation (viability: 8): https://sciencetostartup.com/paper/seeing-beyond-extrapolative-domain-adaptive-panoramic-segmentation - EDA-PSeg enhances panoramic semantic segmentation by addressing geometric distortions and unseen classes through innovative attention mechanisms.
- Agent Lifecycle Toolkit (ALTK): Reusable Middleware Components for Robust AI Agents (viability: 7): https://sciencetostartup.com/paper/agent-lifecycle-toolkit-altk-reusable-middleware-components-for-robust-ai-agents - ALTK is an open-source toolkit that enhances the reliability of AI agents by providing modular middleware components to address failure modes.
- Anchor then Polish for Low-light Enhancement (viability: 7): https://sciencetostartup.com/paper/anchor-then-polish-for-low-light-enhancement - A novel anchor-then-polish framework for superior low-light image enhancement.
- On the Derivation of Tightly-Coupled LiDAR-Inertial Odometry with VoxelMap (viability: 2): https://sciencetostartup.com/paper/on-the-derivation-of-tightly-coupled-lidar-inertial-odometry-with-voxelmap - A mathematical formulation for tightly-coupled LiDAR-Inertial Odometry using VoxelMap.
- Automated Counting of Stacked Objects in Industrial Inspection (viability: 7): https://sciencetostartup.com/paper/automated-counting-of-stacked-objects-in-industrial-inspection - A novel 3D counting approach for accurately counting stacked objects in industrial inspection using multi-view images.
- RoCo Challenge at AAAI 2026: Benchmarking Robotic Collaborative Manipulation for Assembly Towards Industrial Automation (viability: 8): https://sciencetostartup.com/paper/roco-challenge-at-aaai-2026-benchmarking-robotic-collaborative-manipulation-for-assembly-towards-industrial-automation - The RoCo Challenge benchmarks robotic collaborative manipulation for industrial assembly, providing a dataset and evaluation framework to enhance automation.
- Evaluating Time Awareness and Cross-modal Active Perception of Large Models via 4D Escape Room Task (viability: 4): https://sciencetostartup.com/paper/evaluating-time-awareness-and-cross-modal-active-perception-of-large-models-via-4d-escape-room-task - EscapeCraft-4D is a customizable environment for assessing multimodal reasoning and time awareness in large models.
- Evasive Intelligence: Lessons from Malware Analysis for Evaluating AI Agents (viability: 4): https://sciencetostartup.com/paper/evasive-intelligence-lessons-from-malware-analysis-for-evaluating-ai-agents - This paper discusses the vulnerabilities in evaluating AI agents by drawing parallels with malware analysis.
- Unlocking the Value of Text: Event-Driven Reasoning and Multi-Level Alignment for Time Series Forecasting (viability: 8): https://sciencetostartup.com/paper/unlocking-the-value-of-text-event-driven-reasoning-and-multi-level-alignment-for-time-series-forecasting - VoT leverages event-driven reasoning and multi-level alignment to enhance time series forecasting using multimodal information.
- Zero-Shot Generalization from Motion Demonstrations to New Tasks (viability: 7): https://sciencetostartup.com/paper/zero-shot-generalization-from-motion-demonstrations-to-new-tasks - A novel approach to generalizing motion policies in robotics using Gaussian Graphs for efficient task adaptation.
- Music Genre Classification: A Comparative Analysis of Classical Machine Learning and Deep Learning Approaches (viability: 4): https://sciencetostartup.com/paper/music-genre-classification-a-comparative-analysis-of-classical-machine-learning-and-deep-learning-approaches - A novel dataset and comparative analysis for automatic classification of Nepali music genres using machine learning and deep learning techniques.
- MV2UV: Generating High-quality UV Texture Maps with Multiview Prompts (viability: 2): https://sciencetostartup.com/paper/mv2uv-generating-high-quality-uv-texture-maps-with-multiview-prompts - MV2UV offers a novel approach to generating high-quality UV texture maps by addressing multiview inconsistencies and unseen parts.
- Listening to the Echo: User-Reaction Aware Policy Optimization via Scalar-Verbal Hybrid Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/listening-to-the-echo-user-reaction-aware-policy-optimization-via-scalar-verbal-hybrid-reinforcement-learning - RAPO optimizes emotional support dialogue systems using user reactions for enhanced interaction outcomes.
- Real-Time Human Frontal View Synthesis from a Single Image (viability: 7): https://sciencetostartup.com/paper/real-time-human-frontal-view-synthesis-from-a-single-image - PrismMirror enables real-time human frontal view synthesis from a single image, enhancing immersive 3D telepresence.
- Gym-V: A Unified Vision Environment System for Agentic Vision Research (viability: 3): https://sciencetostartup.com/paper/gym-v-a-unified-vision-environment-system-for-agentic-vision-research - Gym-V is a unified platform for agentic vision research, providing a diverse set of procedurally generated environments for reinforcement learning.
- Physics-informed fine-tuning of foundation models for partial differential equations (viability: 4): https://sciencetostartup.com/paper/physics-informed-fine-tuning-of-foundation-models-for-partial-differential-equations - A physics-informed fine-tuning framework for adapting foundation models to partial differential equations with minimal data.
- Formalisms for Robotic Mission Specification and Execution: A Comparative Analysis (viability: 2): https://sciencetostartup.com/paper/formalisms-for-robotic-mission-specification-and-execution-a-comparative-analysis - This paper analyzes various formalisms for specifying robotic missions to aid in selecting appropriate modeling approaches.
- Invisible failures in human-AI interactions (viability: 8): https://sciencetostartup.com/paper/invisible-failures-in-human-ai-interactions - A taxonomy of invisible AI failures to enhance reliability in human-AI interactions.
- CLAG: Adaptive Memory Organization via Agent-Driven Clustering for Small Language Model Agents (viability: 7): https://sciencetostartup.com/paper/clag-adaptive-memory-organization-via-agent-driven-clustering-for-small-language-model-agents - CLAG enhances small language models by organizing memory through agent-driven clustering, improving answer quality and robustness.
- MA-VLCM: A Vision Language Critic Model for Value Estimation of Policies in Multi-Agent Team Settings (viability: 3): https://sciencetostartup.com/paper/ma-vlcm-a-vision-language-critic-model-for-value-estimation-of-policies-in-multi-agent-team-settings - MA-VLCM enhances multi-agent reinforcement learning by using a pretrained vision-language model as a centralized critic for improved sample efficiency.
- Amplification Effects in Test-Time Reinforcement Learning: Safety and Reasoning Vulnerabilities (viability: 2): https://sciencetostartup.com/paper/amplification-effects-in-test-time-reinforcement-learning-safety-and-reasoning-vulnerabilities - This paper explores safety vulnerabilities in test-time training methods for large language models.
- AnyCrowd: Instance-Isolated Identity-Pose Binding for Arbitrary Multi-Character Animation (viability: 2): https://sciencetostartup.com/paper/anycrowd-instance-isolated-identity-pose-binding-for-arbitrary-multi-character-animation - AnyCrowd is a novel framework for scalable multi-character animation that addresses identity entanglement and pose binding challenges.
- RESQ: A Unified Framework for REliability- and Security Enhancement of Quantized Deep Neural Networks (viability: 3): https://sciencetostartup.com/paper/resq-a-unified-framework-for-reliability-and-security-enhancement-of-quantized-deep-neural-networks - A framework enhancing the reliability and security of quantized deep neural networks through a three-stage process.
- Local Urysohn Width: A Topological Complexity Measure for Classification (viability: 2): https://sciencetostartup.com/paper/local-urysohn-width-a-topological-complexity-measure-for-classification - Introducing a new topological complexity measure for classification problems in metric spaces.
- A Hybrid Modeling Framework for Crop Prediction Tasks via Dynamic Parameter Calibration and Multi-Task Learning (viability: 7): https://sciencetostartup.com/paper/a-hybrid-modeling-framework-for-crop-prediction-tasks-via-dynamic-parameter-calibration-and-multi-task-learning - A hybrid modeling framework that enhances crop prediction accuracy through dynamic parameter calibration and multi-task learning.
- End-to-End Dexterous Grasp Learning from Single-View Point Clouds via a Multi-Object Scene Dataset (viability: 8): https://sciencetostartup.com/paper/end-to-end-dexterous-grasp-learning-from-single-view-point-clouds-via-a-multi-object-scene-dataset - DGS-Net is an end-to-end grasp prediction network that learns dense grasp configurations from single-view point clouds in multi-object scenes.
- SEA-Vision: A Multilingual Benchmark for Comprehensive Document and Scene Text Understanding in Southeast Asia (viability: 7): https://sciencetostartup.com/paper/sea-vision-a-multilingual-benchmark-for-comprehensive-document-and-scene-text-understanding-in-southeast-asia - SEA-Vision is a multilingual benchmark for enhancing document and scene text understanding across Southeast Asia's diverse languages.
- TrinityGuard: A Unified Framework for Safeguarding Multi-Agent Systems (viability: 3): https://sciencetostartup.com/paper/trinityguard-a-unified-framework-for-safeguarding-multi-agent-systems - TrinityGuard is a safety evaluation and monitoring framework for LLM-based multi-agent systems addressing unique security risks.
- Fusian: Multi-LoRA Fusion for Fine-Grained Continuous MBTI Personality Control in Large Language Models (viability: 7): https://sciencetostartup.com/paper/fusian-multi-lora-fusion-for-fine-grained-continuous-mbti-personality-control-in-large-language-models - Fusian enables precise, continuous personality control in large language models through a novel framework combining LoRA adapters and reinforcement learning.
- Detection of Autonomous Shuttles in Urban Traffic Images Using Adaptive Residual Context (viability: 4): https://sciencetostartup.com/paper/detection-of-autonomous-shuttles-in-urban-traffic-images-using-adaptive-residual-context - Adaptive Residual Context (ARC) enhances urban traffic monitoring by improving vehicle detection while preserving contextual knowledge.
- Pointing-Based Object Recognition (viability: 6): https://sciencetostartup.com/paper/pointing-based-object-recognition - A pipeline for recognizing objects based on human pointing gestures using RGB images.
- A Closer Look into LLMs for Table Understanding (viability: 2): https://sciencetostartup.com/paper/a-closer-look-into-llms-for-table-understanding - This paper explores the internal mechanisms of LLMs in understanding tabular data, providing insights for future research.
- SWE-Skills-Bench: Do Agent Skills Actually Help in Real-World Software Engineering? (viability: 7): https://sciencetostartup.com/paper/swe-skills-bench-do-agent-skills-actually-help-in-real-world-software-engineering - SWE-Skills-Bench evaluates the effectiveness of agent skills in software engineering tasks using a structured benchmark.
- SFCoT: Safer Chain-of-Thought via Active Safety Evaluation and Calibration (viability: 7): https://sciencetostartup.com/paper/sfcot-safer-chain-of-thought-via-active-safety-evaluation-and-calibration - SFCoT enhances the safety of large language models by proactively evaluating and calibrating reasoning steps to prevent jailbreak attacks.
- AI Evasion and Impersonation Attacks on Facial Re-Identification with Activation Map Explanations (viability: 4): https://sciencetostartup.com/paper/ai-evasion-and-impersonation-attacks-on-facial-re-identification-with-activation-map-explanations - A novel framework for generating adversarial patches that exploit vulnerabilities in facial identification systems.
- When Does Sparsity Mitigate the Curse of Depth in LLMs (viability: 8): https://sciencetostartup.com/paper/when-does-sparsity-mitigate-the-curse-of-depth-in-llms - This research provides a practical approach to improve layer utilization in large language models through sparsity techniques.
- Efficient Morphology-Control Co-Design via Stackelberg Proximal Policy Optimization (viability: 4): https://sciencetostartup.com/paper/efficient-morphology-control-co-design-via-stackelberg-proximal-policy-optimization - A novel game-theoretic approach to optimize agent morphology and control for improved robotics efficiency.
- RieMind: Geometry-Grounded Spatial Agent for Scene Understanding (viability: 3): https://sciencetostartup.com/paper/riemind-geometry-grounded-spatial-agent-for-scene-understanding - RieMind proposes a novel framework for enhancing spatial reasoning in indoor scenes through explicit 3D scene graph grounding.
- Persistence Spheres: a Bi-continuous Linear Representation of Measures for Partial Optimal Transport (viability: 2): https://sciencetostartup.com/paper/persistence-spheres-a-bi-continuous-linear-representation-of-measures-for-partial-optimal-transport - This paper presents a theoretical advancement in topological machine learning through persistence spheres for optimal transport.
- Why AI systems don't learn and what to do about it: Lessons on autonomous learning from cognitive science (viability: 2): https://sciencetostartup.com/paper/why-ai-systems-don-t-learn-and-what-to-do-about-it-lessons-on-autonomous-learning-from-cognitive-science - This paper explores a new learning architecture inspired by cognitive science to enhance autonomous learning in AI systems.
- More Test-Time Compute Can Hurt: Overestimation Bias in LLM Beam Search (viability: 2): https://sciencetostartup.com/paper/more-test-time-compute-can-hurt-overestimation-bias-in-llm-beam-search - This paper analyzes the impact of beam width selection on LLM output quality, revealing potential overestimation bias.
- Spectral Rectification for Parameter-Efficient Adaptation of Foundation Models in Colonoscopy Depth Estimation (viability: 6): https://sciencetostartup.com/paper/spectral-rectification-for-parameter-efficient-adaptation-of-foundation-models-in-colonoscopy-depth-estimation - SpecDepth enhances monocular depth estimation in colonoscopy by adapting foundation models to address spectral mismatches.
- GradCFA: A Hybrid Gradient-Based Counterfactual and Feature Attribution Explanation Algorithm for Local Interpretation of Neural Networks (viability: 7): https://sciencetostartup.com/paper/gradcfa-a-hybrid-gradient-based-counterfactual-and-feature-attribution-explanation-algorithm-for-local-interpretation-of - GradCFA is a hybrid framework that enhances AI interpretability through optimized counterfactual explanations and feature attribution.
- SKILLS: Structured Knowledge Injection for LLM-Driven Telecommunications Operations (viability: 7): https://sciencetostartup.com/paper/skills-structured-knowledge-injection-for-llm-driven-telecommunications-operations - SKILLS enhances LLM-driven telecom operations by integrating structured knowledge for improved workflow execution.
- Brain-Inspired Graph Multi-Agent Systems for LLM Reasoning (viability: 7): https://sciencetostartup.com/paper/brain-inspired-graph-multi-agent-systems-for-llm-reasoning - Introducing Brain-Inspired Graph Multi-Agent Systems to enhance reasoning in Large Language Models through specialized agent coordination.
- Trajectory-Diversity-Driven Robust Vision-and-Language Navigation (viability: 8): https://sciencetostartup.com/paper/trajectory-diversity-driven-robust-vision-and-language-navigation - NavGRPO is a robust reinforcement learning framework for goal-directed navigation in photo-realistic environments using natural language instructions.
- IRIS: Intersection-aware Ray-based Implicit Editable Scenes (viability: 8): https://sciencetostartup.com/paper/iris-intersection-aware-ray-based-implicit-editable-scenes - IRIS enables efficient and interactive editing of 3D scenes using advanced ray-based techniques.
- A PPO-Based Bitrate Allocation Conditional Diffusion Model for Remote Sensing Image Compression (viability: 7): https://sciencetostartup.com/paper/a-ppo-based-bitrate-allocation-conditional-diffusion-model-for-remote-sensing-image-compression - A novel conditional diffusion model for efficient remote sensing image compression with high perceptual performance.
- CRASH: Cognitive Reasoning Agent for Safety Hazards in Autonomous Driving (viability: 7): https://sciencetostartup.com/paper/crash-cognitive-reasoning-agent-for-safety-hazards-in-autonomous-driving - CRASH is an LLM-based agent that automates reasoning over autonomous vehicle crash reports to enhance safety analysis.
- Deep learning and the rate of approximation by flows (viability: 2): https://sciencetostartup.com/paper/deep-learning-and-the-rate-of-approximation-by-flows - This paper explores the theoretical aspects of deep residual networks' approximation capacity in dynamical systems.
- NavThinker: Action-Conditioned World Models for Coupled Prediction and Planning in Social Navigation (viability: 8): https://sciencetostartup.com/paper/navthinker-action-conditioned-world-models-for-coupled-prediction-and-planning-in-social-navigation - NavThinker offers a future-aware framework for social navigation using action-conditioned world models and reinforcement learning.
- FuXiWeather2: Learning accurate atmospheric state estimation for operational global weather forecasting (viability: 5): https://sciencetostartup.com/paper/fuxiweather2-learning-accurate-atmospheric-state-estimation-for-operational-global-weather-forecasting - Transforming global weather forecasting with FuXiWeather2's rapid and accurate AI-powered predictions.
- Conditional Rectified Flow-based End-to-End Rapid Seismic Inversion Method (viability: 7): https://sciencetostartup.com/paper/conditional-rectified-flow-based-end-to-end-rapid-seismic-inversion-method - A fast seismic inversion method leveraging Conditional Rectified Flow to enhance accuracy and efficiency in geophysical exploration.
- NV-Bench: Benchmark of Nonverbal Vocalization Synthesis for Expressive Text-to-Speech Generation (viability: 4): https://sciencetostartup.com/paper/nv-bench-benchmark-of-nonverbal-vocalization-synthesis-for-expressive-text-to-speech-generation - NV-Bench provides a standardized benchmark for evaluating nonverbal vocalization synthesis in text-to-speech systems.
- PMAx: An Agentic Framework for AI-Driven Process Mining (viability: 6): https://sciencetostartup.com/paper/pmax-an-agentic-framework-for-ai-driven-process-mining - PMAx is an autonomous agentic framework that democratizes process mining by enabling non-technical users to derive insights from data through natural language interactions while ensuring data privacy.
- Oscillating Dispersion for Maximal Light-throughput Spectral Imaging (viability: 7): https://sciencetostartup.com/paper/oscillating-dispersion-for-maximal-light-throughput-spectral-imaging - A novel imaging spectrometer that maximizes light throughput for high-fidelity spectral reconstruction.
- Unsupervised Cross-Protocol Anomaly Analysis in Mobile Core Networks via Multi-Embedding Models Consensus (viability: 2): https://sciencetostartup.com/paper/unsupervised-cross-protocol-anomaly-analysis-in-mobile-core-networks-via-multi-embedding-models-consensus - This paper presents a method for unsupervised anomaly detection in mobile core networks using multi-embedding models.
- Intelligent Co-Design: An Interactive LLM Framework for Interior Spatial Design via Multi-Modal Agents (viability: 8): https://sciencetostartup.com/paper/intelligent-co-design-an-interactive-llm-framework-for-interior-spatial-design-via-multi-modal-agents - An interactive LLM framework that transforms natural language and imagery into optimized 3D interior designs, enhancing user engagement and design communication.
- DOS: Dependency-Oriented Sampler for Masked Diffusion Language Models (viability: 7): https://sciencetostartup.com/paper/dos-dependency-oriented-sampler-for-masked-diffusion-language-models - Dependency-Oriented Sampler enhances masked diffusion language models by leveraging inter-token dependencies for improved generation efficiency.
- Active Seriation: Efficient Ordering Recovery with Statistical Guarantees (viability: 2): https://sciencetostartup.com/paper/active-seriation-efficient-ordering-recovery-with-statistical-guarantees - A theoretical framework for recovering unknown orderings through adaptive querying of pairwise similarities.
- Data Augmentation via Causal-Residual Bootstrapping (viability: 4): https://sciencetostartup.com/paper/data-augmentation-via-causal-residual-bootstrapping - A novel data augmentation method that leverages causal knowledge to enhance predictive model accuracy.
- MeMix: Writing Less, Remembering More for Streaming 3D Reconstruction (viability: 8): https://sciencetostartup.com/paper/memix-writing-less-remembering-more-for-streaming-3d-reconstruction - MeMix is a plug-and-play module that enhances streaming 3D reconstruction by mitigating catastrophic forgetting without the need for fine-tuning.
- User-Tailored Learning to Forecast Walking Modes for Exosuits (viability: 7): https://sciencetostartup.com/paper/user-tailored-learning-to-forecast-walking-modes-for-exosuits - A perception module for exosuits that estimates walking modes using inertial data for enhanced user adaptation.
- Tagarela - A Portuguese speech dataset from podcasts (viability: 7): https://sciencetostartup.com/paper/tagarela-a-portuguese-speech-dataset-from-podcasts - TAGARELA is a large-scale Portuguese speech dataset designed to enhance automatic speech recognition and text-to-speech technologies.
- CASHomon Sets: Efficient Rashomon Sets Across Multiple Model Classes and their Hyperparameters (viability: 4): https://sciencetostartup.com/paper/cashomon-sets-efficient-rashomon-sets-across-multiple-model-classes-and-their-hyperparameters - CASHomon Sets enable efficient model selection across multiple classes and hyperparameters, enhancing interpretability and performance.
- Comparative Analysis of SRAM PUF Temperature Susceptibility on Embedded Systems (viability: 2): https://sciencetostartup.com/paper/comparative-analysis-of-sram-puf-temperature-susceptibility-on-embedded-systems - This study evaluates the temperature susceptibility of SRAM PUFs in embedded systems.
- PYTHEN: A Flexible Framework for Legal Reasoning in Python (viability: 5): https://sciencetostartup.com/paper/pythen-a-flexible-framework-for-legal-reasoning-in-python - PYTHEN is a Python-based framework that simplifies defeasible legal reasoning for developers and legal professionals.
- CCTU: A Benchmark for Tool Use under Complex Constraints (viability: 7): https://sciencetostartup.com/paper/cctu-a-benchmark-for-tool-use-under-complex-constraints - CCTU is a benchmark designed to evaluate large language models' tool use under complex constraints, revealing critical limitations in their performance.
- A Kolmogorov-Arnold Surrogate Model for Chemical Equilibria: Application to Solid Solutions (viability: 5): https://sciencetostartup.com/paper/a-kolmogorov-arnold-surrogate-model-for-chemical-equilibria-application-to-solid-solutions - A novel surrogate model using Kolmogorov-Arnold networks to enhance the efficiency of geochemical solvers for nuclear waste disposal.
- xplainfi: Feature Importance and Statistical Inference for Machine Learning in R (viability: 6): https://sciencetostartup.com/paper/xplainfi-feature-importance-and-statistical-inference-for-machine-learning-in-r - xplainfi is an R package that provides advanced feature importance methods for machine learning models.
- Generative Video Compression with One-Dimensional Latent Representation (viability: 8): https://sciencetostartup.com/paper/generative-video-compression-with-one-dimensional-latent-representation - GVC1D revolutionizes video compression by using a compact one-dimensional latent representation to enhance efficiency and reduce bitrate significantly.
- GATE-AD: Graph Attention Network Encoding For Few-Shot Industrial Visual Anomaly Detection (viability: 8): https://sciencetostartup.com/paper/gate-ad-graph-attention-network-encoding-for-few-shot-industrial-visual-anomaly-detection - A few-shot visual anomaly detection tool for industrial quality assurance using graph attention networks.
- Enhancing classification accuracy through chaos (viability: 2): https://sciencetostartup.com/paper/enhancing-classification-accuracy-through-chaos - A novel chaos-based approach to improve classification accuracy through enhanced training processes.
- Evolutionary Transfer Learning for Dragonchess (viability: 7): https://sciencetostartup.com/paper/evolutionary-transfer-learning-for-dragonchess - An open-source game engine for Dragonchess that leverages evolutionary transfer learning to enhance AI performance.
- Datasets for Verb Alternations across Languages: BLM Templates and Data Augmentation Strategies (viability: 4): https://sciencetostartup.com/paper/datasets-for-verb-alternations-across-languages-blm-templates-and-data-augmentation-strategies - Curated datasets for probing verb alternations in multiple languages to enhance LLM performance.
- Scalable Simulation-Based Model Inference with Test-Time Complexity Control (viability: 3): https://sciencetostartup.com/paper/scalable-simulation-based-model-inference-with-test-time-complexity-control - PRISM is a simulation-based encoder-decoder for scalable model selection in scientific applications.
- Evaluating the Robustness of Reinforcement Learning based Adaptive Traffic Signal Control (viability: 6): https://sciencetostartup.com/paper/evaluating-the-robustness-of-reinforcement-learning-based-adaptive-traffic-signal-control - An RL-based traffic signal control algorithm that adapts to varying traffic conditions and outperforms traditional methods.
- Algorithms for Deciding the Safety of States in Fully Observable Non-deterministic Problems: Technical Report (viability: 4): https://sciencetostartup.com/paper/algorithms-for-deciding-the-safety-of-states-in-fully-observable-non-deterministic-problems-technical-report - A new policy-iteration algorithm that ensures safety in sequential decision-making with improved runtime efficiency.
- GNIO: Gated Neural Inertial Odometry (viability: 8): https://sciencetostartup.com/paper/gnio-gated-neural-inertial-odometry - GNIO is a novel learning-based framework that enhances inertial navigation accuracy by dynamically suppressing sensor noise and improving motion context understanding.
- Advancing Multimodal Agent Reasoning with Long-Term Neuro-Symbolic Memory (viability: 7): https://sciencetostartup.com/paper/advancing-multimodal-agent-reasoning-with-long-term-neuro-symbolic-memory - NS-Mem enhances multimodal agent reasoning by integrating neuro-symbolic memory for better analytical decision making.
- Faster Inference of Flow-Based Generative Models via Improved Data-Noise Coupling (viability: 7): https://sciencetostartup.com/paper/faster-inference-of-flow-based-generative-models-via-improved-data-noise-coupling - LOOM-CFM accelerates inference in flow-based generative models by optimizing noise-data coupling across minibatches.
- Dataset Diversity Metrics and Impact on Classification Models (viability: 7): https://sciencetostartup.com/paper/dataset-diversity-metrics-and-impact-on-classification-models - A study on dataset diversity metrics and their impact on classification model performance.
- Flash-Unified: A Training-Free and Task-Aware Acceleration Framework for Native Unified Models (viability: 7): https://sciencetostartup.com/paper/flash-unified-a-training-free-and-task-aware-acceleration-framework-for-native-unified-models - Develop a task-aware, training-free acceleration framework for unified multimodal models, optimizing real-world AI deployment.
- From Documents to Spans: Code-Centric Learning for LLM-based ICD Coding (viability: 7): https://sciencetostartup.com/paper/from-documents-to-spans-code-centric-learning-for-llm-based-icd-coding - A novel training framework that enhances LLMs for efficient and interpretable ICD coding using short evidence spans.
- Self-Supervised ImageNet Representations for In Vivo Confocal Microscopy: Tortuosity Grading without Segmentation Maps (viability: 4): https://sciencetostartup.com/paper/self-supervised-imagenet-representations-for-in-vivo-confocal-microscopy-tortuosity-grading-without-segmentation-maps - A self-supervised approach to grading corneal nerve fiber tortuosity without expensive segmentation maps.
- Exemplar Diffusion: Improving Medical Object Detection with Opportunistic Labels (viability: 8): https://sciencetostartup.com/paper/exemplar-diffusion-improving-medical-object-detection-with-opportunistic-labels - A framework that enhances medical object detection by utilizing existing labels at inference for improved accuracy and robustness.
- MoE-ACT: Scaling Multi-Task Bimanual Manipulation with Sparse Language-Conditioned Mixture-of-Experts Transformers (viability: 8): https://sciencetostartup.com/paper/moe-act-scaling-multi-task-bimanual-manipulation-with-sparse-language-conditioned-mixture-of-experts-transformers - MoE-ACT enhances robotic manipulation by integrating language-conditioned Mixture-of-Experts into a lightweight multi-task imitation learning framework.
- IConE: Batch Independent Collapse Prevention for Self-Supervised Representation Learning (viability: 4): https://sciencetostartup.com/paper/icone-batch-independent-collapse-prevention-for-self-supervised-representation-learning - IConE offers a novel approach to prevent representation collapse in self-supervised learning, enabling effective training with small batch sizes.
- Probe-then-Plan: Environment-Aware Planning for Industrial E-commerce Search (viability: 8): https://sciencetostartup.com/paper/probe-then-plan-environment-aware-planning-for-industrial-e-commerce-search - A novel industrial e-commerce search framework that enhances user conversion by grounding search plans in real-time retrieval data.
- AGCD: Agent-Guided Cross-Modal Decoding for Weather Forecasting (viability: 7): https://sciencetostartup.com/paper/agcd-agent-guided-cross-modal-decoding-for-weather-forecasting - AGCD is a novel decoding paradigm that enhances weather forecasting accuracy by integrating state-conditioned physics-priors.
- Directional Embedding Smoothing for Robust Vision Language Models (viability: 5): https://sciencetostartup.com/paper/directional-embedding-smoothing-for-robust-vision-language-models - A defense mechanism to enhance the safety and reliability of vision-language models against jailbreaking attacks.
- HapticVLA: Contact-Rich Manipulation via Vision-Language-Action Model without Inference-Time Tactile Sensing (viability: 3): https://sciencetostartup.com/paper/hapticvla-contact-rich-manipulation-via-vision-language-action-model-without-inference-time-tactile-sensing - HapticVLA enables contact-rich manipulation without the need for tactile sensors during inference.
- SAGE: Multi-Agent Self-Evolution for LLM Reasoning (viability: 8): https://sciencetostartup.com/paper/sage-multi-agent-self-evolution-for-llm-reasoning - SAGE is a self-evolving multi-agent framework that enhances reasoning in LLMs through closed-loop training with minimal human input.
- A Methodology for Dynamic Parameters Identification of 3-DOF Parallel Robots in Terms of Relevant Parameters (viability: 2): https://sciencetostartup.com/paper/a-methodology-for-dynamic-parameters-identification-of-3-dof-parallel-robots-in-terms-of-relevant-parameters - A methodology for identifying dynamic parameters in 3-DOF parallel robots to enhance model-based control.
- HalDec-Bench: Benchmarking Hallucination Detector in Image Captioning (viability: 7): https://sciencetostartup.com/paper/haldec-bench-benchmarking-hallucination-detector-in-image-captioning - HalDec-Bench is a comprehensive benchmark for evaluating hallucination detectors in image captioning, enhancing the quality of vision-language models.
- In-Context Symbolic Regression for Robustness-Improved Kolmogorov-Arnold Networks (viability: 2): https://sciencetostartup.com/paper/in-context-symbolic-regression-for-robustness-improved-kolmogorov-arnold-networks - This paper explores in-context symbolic regression techniques to enhance the robustness of Kolmogorov-Arnold Networks.
- Mechanistic Foundations of Goal-Directed Control (viability: 2): https://sciencetostartup.com/paper/mechanistic-foundations-of-goal-directed-control - This paper explores mechanistic interpretability in embodied control systems using infant motor learning as a model.
- Practicing with Language Models Cultivates Human Empathic Communication (viability: 7): https://sciencetostartup.com/paper/practicing-with-language-models-cultivates-human-empathic-communication - Lend an Ear is an AI-driven platform that enhances human empathic communication through personalized coaching.
- Why the Valuable Capabilities of LLMs Are Precisely the Unexplainable Ones (viability: 3): https://sciencetostartup.com/paper/why-the-valuable-capabilities-of-llms-are-precisely-the-unexplainable-ones - This paper argues that the most valuable capabilities of LLMs lie in their unexplainable aspects, challenging traditional expert systems.
- Multi-turn Physics-informed Vision-language Model for Physics-grounded Anomaly Detection (viability: 8): https://sciencetostartup.com/paper/multi-turn-physics-informed-vision-language-model-for-physics-grounded-anomaly-detection - A physics-informed vision-language model for robust anomaly detection in dynamic systems.
- Decomposing Probabilistic Scores: Reliability, Information Loss and Uncertainty (viability: 2): https://sciencetostartup.com/paper/decomposing-probabilistic-scores-reliability-information-loss-and-uncertainty - This paper explores the decomposition of probabilistic scores to analyze calibration and uncertainty in predictors.
- HYDRA: Unifying Multi-modal Generation and Understanding via Representation-Harmonized Tokenization (viability: 8): https://sciencetostartup.com/paper/hydra-unifying-multi-modal-generation-and-understanding-via-representation-harmonized-tokenization - HYDRA-TOK unifies visual understanding and generation through a novel representation-harmonized approach.
- Bidirectional Chinese and English Passive Sentences Dataset for Machine Translation (viability: 4): https://sciencetostartup.com/paper/bidirectional-chinese-and-english-passive-sentences-dataset-for-machine-translation - A dataset for improving machine translation of passive sentences between English and Chinese.
- SCAN: Sparse Circuit Anchor Interpretable Neuron for Lifelong Knowledge Editing (viability: 7): https://sciencetostartup.com/paper/scan-sparse-circuit-anchor-interpretable-neuron-for-lifelong-knowledge-editing - SCAN offers a novel sparse editing framework for Large Language Models to prevent catastrophic forgetting during knowledge updates.
- Coupled Particle Filters for Robust Affordance Estimation (viability: 7): https://sciencetostartup.com/paper/coupled-particle-filters-for-robust-affordance-estimation - A novel method for robust affordance estimation in robotics using coupled particle filters.
- ADV-0: Closed-Loop Min-Max Adversarial Training for Long-Tail Robustness in Autonomous Driving (viability: 3): https://sciencetostartup.com/paper/adv-0-closed-loop-min-max-adversarial-training-for-long-tail-robustness-in-autonomous-driving - A novel adversarial training framework for enhancing robustness in autonomous driving systems against rare, safety-critical scenarios.
- InterPol: De-anonymizing LM Arena via Interpolated Preference Learning (viability: 7): https://sciencetostartup.com/paper/interpol-de-anonymizing-lm-arena-via-interpolated-preference-learning - INTERPOL is a model-driven framework that enhances identification accuracy of language models by learning deep stylistic patterns.
- Towards Foundation Models for Consensus Rank Aggregation (viability: 6): https://sciencetostartup.com/paper/towards-foundation-models-for-consensus-rank-aggregation - A Transformer-based algorithm that efficiently approximates optimal consensus rankings for applications in recommendation systems and search engines.
- vCause: Efficient and Verifiable Causality Analysis for Cloud-based Endpoint Auditing (viability: 2): https://sciencetostartup.com/paper/vcause-efficient-and-verifiable-causality-analysis-for-cloud-based-endpoint-auditing - vCause offers a secure and efficient system for causality analysis in cloud-based endpoint auditing.
- Tracking the Discriminative Axis: Dual Prototypes for Test-Time OOD Detection Under Covariate Shift (viability: 7): https://sciencetostartup.com/paper/tracking-the-discriminative-axis-dual-prototypes-for-test-time-ood-detection-under-covariate-shift - DART is an online OOD detection method that adapts to covariate shifts by tracking dual prototypes for improved performance.
- Modeling Matches as Language: A Generative Transformer Approach for Counterfactual Player Valuation in Football (viability: 7): https://sciencetostartup.com/paper/modeling-matches-as-language-a-generative-transformer-approach-for-counterfactual-player-valuation-in-football - ScoutGPT is a generative model that simulates football match events to enhance player transfer evaluations through counterfactual analysis.
- Efficient Document Parsing via Parallel Token Prediction (viability: 7): https://sciencetostartup.com/paper/efficient-document-parsing-via-parallel-token-prediction - A novel method for accelerating document parsing using parallel token prediction in vision-language models.
- Massive Redundancy in Gradient Transport Enables Sparse Online Learning (viability: 2): https://sciencetostartup.com/paper/massive-redundancy-in-gradient-transport-enables-sparse-online-learning - This paper explores a method for reducing computational costs in online learning through sparse gradient transport.
- PiGRAND: Physics-informed Graph Neural Diffusion for Intelligent Additive Manufacturing (viability: 8): https://sciencetostartup.com/paper/pigrand-physics-informed-graph-neural-diffusion-for-intelligent-additive-manufacturing - PiGRAND leverages physics-informed graph neural diffusion to optimize heat transport in 3D printing applications.
- The Sampling Complexity of Condorcet Winner Identification in Dueling Bandits (viability: 2): https://sciencetostartup.com/paper/the-sampling-complexity-of-condorcet-winner-identification-in-dueling-bandits - This paper presents a theoretical framework for identifying Condorcet winners in dueling bandits, focusing on sample complexity.
- Joint Routing and Model Pruning for Decentralized Federated Learning in Bandwidth-Constrained Multi-Hop Wireless Networks (viability: 4): https://sciencetostartup.com/paper/joint-routing-and-model-pruning-for-decentralized-federated-learning-in-bandwidth-constrained-multi-hop-wireless-network - A framework that optimizes routing and model pruning for efficient decentralized federated learning in bandwidth-constrained environments.
- The Hrunting of AI: Where and How to Improve English Dialectal Fairness (viability: 3): https://sciencetostartup.com/paper/the-hrunting-of-ai-where-and-how-to-improve-english-dialectal-fairness - This research explores the challenges of improving LLM performance in underrepresented English dialects due to data scarcity.
- NavGSim: High-Fidelity Gaussian Splatting Simulator for Large-Scale Navigation (viability: 7): https://sciencetostartup.com/paper/navgsim-high-fidelity-gaussian-splatting-simulator-for-large-scale-navigation - NavGSim is a high-fidelity simulator that enhances robot navigation through realistic environment rendering and collision simulation.
- What Matters for Scalable and Robust Learning in End-to-End Driving Planners? (viability: 7): https://sciencetostartup.com/paper/what-matters-for-scalable-and-robust-learning-in-end-to-end-driving-planners - Develop flexible and scalable autonomous driving systems leveraging a novel end-to-end architecture for enhanced closed-loop driving performance.
- CATFormer: When Continual Learning Meets Spiking Transformers With Dynamic Thresholds (viability: 4): https://sciencetostartup.com/paper/catformer-when-continual-learning-meets-spiking-transformers-with-dynamic-thresholds - CATFormer is a scalable framework that prevents catastrophic forgetting in spiking neural networks through dynamic thresholds.
- Token Coherence: Adapting MESI Cache Protocols to Minimize Synchronization Overhead in Multi-Agent LLM Systems (viability: 8): https://sciencetostartup.com/paper/token-coherence-adapting-mesi-cache-protocols-to-minimize-synchronization-overhead-in-multi-agent-llm-systems - A system that minimizes synchronization overhead in multi-agent LLMs by adapting MESI cache protocols.
- Sequential Transport for Causal Mediation Analysis (viability: 3): https://sciencetostartup.com/paper/sequential-transport-for-causal-mediation-analysis - A novel framework for causal mediation analysis using sequential transport and optimal transport methods.
- KiRAS: Keyframe Guided Self-Imitation for Robust and Adaptive Skill Learning in Quadruped Robots (viability: 7): https://sciencetostartup.com/paper/kiras-keyframe-guided-self-imitation-for-robust-and-adaptive-skill-learning-in-quadruped-robots - KiRAS is a framework for quadruped robots that enables robust skill learning and adaptability across complex terrains using keyframe-guided self-imitation.
- ForceVLA2: Unleashing Hybrid Force-Position Control with Force Awareness for Contact-Rich Manipulation (viability: 8): https://sciencetostartup.com/paper/forcevla2-unleashing-hybrid-force-position-control-with-force-awareness-for-contact-rich-manipulation - ForceVLA2 enhances robotic manipulation by integrating hybrid force-position control with explicit force awareness for improved task performance.
- Multimodal Connectome Fusion via Cross-Attention for Autism Spectrum Disorder Classification Using Graph Learning (viability: 7): https://sciencetostartup.com/paper/multimodal-connectome-fusion-via-cross-attention-for-autism-spectrum-disorder-classification-using-graph-learning - A multimodal graph learning framework for improved classification of Autism Spectrum Disorder using integrated imaging data.
- Question-guided Visual Compression with Memory Feedback for Long-Term Video Understanding (viability: 6): https://sciencetostartup.com/paper/question-guided-visual-compression-with-memory-feedback-for-long-term-video-understanding - A framework that enhances long-term video understanding through question-guided visual compression and memory feedback.
- DAIT: Distillation from Vision-Language Models to Lightweight Classifiers with Adaptive Intermediate Teacher Transfer (viability: 6): https://sciencetostartup.com/paper/dait-distillation-from-vision-language-models-to-lightweight-classifiers-with-adaptive-intermediate-teacher-transfer - DAIT enables efficient knowledge transfer from large Vision-Language Models to lightweight classifiers for fine-grained visual categorization.
- HindSight: Evaluating Research Idea Generation via Future Impact (viability: 2): https://sciencetostartup.com/paper/hindsight-evaluating-research-idea-generation-via-future-impact - A framework to evaluate AI-generated research ideas based on their future impact and citation potential.
- To See is Not to Master: Teaching LLMs to Use Private Libraries for Code Generation (viability: 7): https://sciencetostartup.com/paper/to-see-is-not-to-master-teaching-llms-to-use-private-libraries-for-code-generation - PriCoder enables LLMs to effectively use private library APIs for code generation by synthesizing data and enhancing code diversity and quality.
- Point-Identification of a Robust Predictor Under Latent Shift with Imperfect Proxies (viability: 4): https://sciencetostartup.com/paper/point-identification-of-a-robust-predictor-under-latent-shift-with-imperfect-proxies - A framework for robust predictor identification under latent shifts using imperfect proxies.
- Vision-Language Model Based Multi-Expert Fusion for CT Image Classification (viability: 6): https://sciencetostartup.com/paper/vision-language-model-based-multi-expert-fusion-for-ct-image-classification - A multi-expert framework for robust COVID-19 CT classification leveraging source-aware modeling.
- TextOVSR: Text-Guided Real-World Opera Video Super-Resolution (viability: 8): https://sciencetostartup.com/paper/textovsr-text-guided-real-world-opera-video-super-resolution - TextOVSR leverages text prompts to enhance the super-resolution of degraded opera videos, outperforming existing methods.
- Master Micro Residual Correction with Adaptive Tactile Fusion and Force-Mixed Control for Contact-Rich Manipulation (viability: 7): https://sciencetostartup.com/paper/master-micro-residual-correction-with-adaptive-tactile-fusion-and-force-mixed-control-for-contact-rich-manipulation - M2-ResiPolicy enhances robotic manipulation by integrating high-level guidance with low-level corrections for improved interaction safety.
- SNCE: Geometry-Aware Supervision for Scalable Discrete Image Generation (viability: 3): https://sciencetostartup.com/paper/snce-geometry-aware-supervision-for-scalable-discrete-image-generation - SNCE introduces a novel training objective to enhance discrete image generation by optimizing large VQ codebooks.
- Accelerating Byzantine-Robust Distributed Learning with Compressed Communication via Double Momentum and Variance Reduction (viability: 4): https://sciencetostartup.com/paper/accelerating-byzantine-robust-distributed-learning-with-compressed-communication-via-double-momentum-and-variance-reduct - Byz-DM21 is a novel Byzantine-robust distributed learning algorithm that enhances communication efficiency through a double-momentum gradient estimator.
- Context-Aware Sensor Modeling for Asynchronous Multi-Sensor Tracking in Stone Soup (viability: 4): https://sciencetostartup.com/paper/context-aware-sensor-modeling-for-asynchronous-multi-sensor-tracking-in-stone-soup - Context-aware sensor modeling enhances multi-sensor tracking performance in heterogeneous environments.
- Safe Flow Q-Learning: Offline Safe Reinforcement Learning with Reachability-Based Flow Policies (viability: 7): https://sciencetostartup.com/paper/safe-flow-q-learning-offline-safe-reinforcement-learning-with-reachability-based-flow-policies - Safe Flow Q-Learning offers a novel approach to offline safe reinforcement learning, ensuring safety in real-time control applications.
- Confusion-Aware In-Context-Learning for Vision-Language Models in Robotic Manipulation (viability: 3): https://sciencetostartup.com/paper/confusion-aware-in-context-learning-for-vision-language-models-in-robotic-manipulation - CAICL enhances vision-language models for improved robustness in robotic manipulation tasks involving confusable objects.
- WiT: Waypoint Diffusion Transformers via Trajectory Conflict Navigation (viability: 8): https://sciencetostartup.com/paper/wit-waypoint-diffusion-transformers-via-trajectory-conflict-navigation - Develop Waypoint Diffusion Transformers (WiT) to improve pixel-space image generation by resolving trajectory conflicts and accelerating training.
- Low-light Image Enhancement with Retinex Decomposition in Latent Space (viability: 6): https://sciencetostartup.com/paper/low-light-image-enhancement-with-retinex-decomposition-in-latent-space - A novel Retinex-Guided Transformer model for stable low-light image enhancement through advanced decomposition techniques.
- Indirect Question Answering in English, German and Bavarian: A Challenging Task for High- and Low-Resource Languages Alike (viability: 4): https://sciencetostartup.com/paper/indirect-question-answering-in-english-german-and-bavarian-a-challenging-task-for-high-and-low-resource-languages-alike - A multilingual approach to indirect question answering, addressing challenges in both high- and low-resource languages.
- Next-Frame Decoding for Ultra-Low-Bitrate Image Compression with Video Diffusion Priors (viability: 4): https://sciencetostartup.com/paper/next-frame-decoding-for-ultra-low-bitrate-image-compression-with-video-diffusion-priors - A novel approach to ultra-low-bitrate image compression using temporal priors for improved fidelity and speed.
- A Novel Camera-to-Robot Calibration Method for Vision-Based Floor Measurements (viability: 2): https://sciencetostartup.com/paper/a-novel-camera-to-robot-calibration-method-for-vision-based-floor-measurements - A novel calibration method for integrating camera and laser tracker measurements in mobile robots.
- From Storage to Steering: Memory Control Flow Attacks on LLM Agents (viability: 3): https://sciencetostartup.com/paper/from-storage-to-steering-memory-control-flow-attacks-on-llm-agents - This paper explores a new security vulnerability in LLM agents related to memory control flow attacks.
- Establishing Construct Validity in LLM Capability Benchmarks Requires Nomological Networks (viability: 4): https://sciencetostartup.com/paper/establishing-construct-validity-in-llm-capability-benchmarks-requires-nomological-networks - This paper critiques the validity of LLM capability benchmarks and proposes a robust theoretical framework for assessment.
- A Tutorial on ALOS2 SAR Utilization: Dataset Preparation, Self-Supervised Pretraining, and Semantic Segmentation (viability: 5): https://sciencetostartup.com/paper/a-tutorial-on-alos2-sar-utilization-dataset-preparation-self-supervised-pretraining-and-semantic-segmentation - A novel approach to enhance semantic segmentation in satellite imagery using self-supervised pretraining techniques tailored for SAR data.
- VAREX: A Benchmark for Multi-Modal Structured Extraction from Documents (viability: 8): https://sciencetostartup.com/paper/varex-a-benchmark-for-multi-modal-structured-extraction-from-documents - VAREX is a benchmark for evaluating multi-modal structured data extraction from documents, enhancing model performance insights.
- MMKU-Bench: A Multimodal Update Benchmark for Diverse Visual Knowledge (viability: 4): https://sciencetostartup.com/paper/mmku-bench-a-multimodal-update-benchmark-for-diverse-visual-knowledge - MMKU-Bench is a comprehensive evaluation benchmark for multimodal knowledge updating, addressing the need for consistent real-world knowledge in AI models.
- Sampling-guided exploration of active feature selection policies (viability: 6): https://sciencetostartup.com/paper/sampling-guided-exploration-of-active-feature-selection-policies - A reinforcement learning framework for efficient feature selection in predictive models.
- PAKAN: Pixel Adaptive Kolmogorov-Arnold Network Modules for Pansharpening (viability: 4): https://sciencetostartup.com/paper/pakan-pixel-adaptive-kolmogorov-arnold-network-modules-for-pansharpening - PAKAN enhances pansharpening by introducing adaptive activation functions for improved spatial-spectral fusion.
- BodyGuards: Escorting by Multiple Robots in Unknown Environment under Limited Communication (viability: 7): https://sciencetostartup.com/paper/bodyguards-escorting-by-multiple-robots-in-unknown-environment-under-limited-communication - BodyGuards is a multi-robot escorting framework designed to protect human operators in unknown environments with limited communication.
- PrototypeNAS: Rapid Design of Deep Neural Networks for Microcontroller Units (viability: 7): https://sciencetostartup.com/paper/prototypenas-rapid-design-of-deep-neural-networks-for-microcontroller-units - PrototypeNAS automates the design of efficient deep neural networks tailored for microcontroller units, enabling rapid deployment on edge devices.
- Learning from Limited and Incomplete Data: A Multimodal Framework for Predicting Pathological Response in NSCLC (viability: 2): https://sciencetostartup.com/paper/learning-from-limited-and-incomplete-data-a-multimodal-framework-for-predicting-pathological-response-in-nsclc - A multimodal deep learning framework for predicting pathological response in non-small cell lung cancer using limited data.
- AeroGrab: A Unified Framework for Aerial Grasping in Cluttered Environments (viability: 7): https://sciencetostartup.com/paper/aerograb-a-unified-framework-for-aerial-grasping-in-cluttered-environments - AeroGrab is an integrated pipeline for reliable aerial grasping in cluttered environments using language instructions.
- Bridging National and International Legal Data: Two Projects Based on the Japanese Legal Standard XML Schema for Comparative Law Studies (viability: 3): https://sciencetostartup.com/paper/bridging-national-and-international-legal-data-two-projects-based-on-the-japanese-legal-standard-xml-schema-for-comparat - An integrated framework for comparative law studies using Japanese legal data.
- Affordable Precision Agriculture: A Deployment-Oriented Review of Low-Cost, Low-Power Edge AI and TinyML for Resource-Constrained Farming Systems (viability: 4): https://sciencetostartup.com/paper/affordable-precision-agriculture-a-deployment-oriented-review-of-low-cost-low-power-edge-ai-and-tinyml-for-resource-cons - A review of low-cost Edge AI and TinyML solutions for enhancing precision agriculture in resource-constrained environments.
- HALO:Closing Sim-to-Real Gap for Heavy-loaded Humanoid Agile Motion Skills via Differentiable Simulation (viability: 7): https://sciencetostartup.com/paper/halo-closing-sim-to-real-gap-for-heavy-loaded-humanoid-agile-motion-skills-via-differentiable-simulation - HALO enables humanoid robots to effectively adapt to unknown payloads through a novel gradient-based system identification framework.
- ReactMotion: Generating Reactive Listener Motions from Speaker Utterance (viability: 6): https://sciencetostartup.com/paper/reactmotion-generating-reactive-listener-motions-from-speaker-utterance - ReactMotion generates naturalistic listener body motions in response to speaker utterances using a novel dataset and generative framework.
- Open Biomedical Knowledge Graphs at Scale: Construction, Federation, and AI Agent Access with Samyama Graph Database (viability: 7): https://sciencetostartup.com/paper/open-biomedical-knowledge-graphs-at-scale-construction-federation-and-ai-agent-access-with-samyama-graph-database - Open-source biomedical knowledge graphs that enable rapid, reproducible cross-referencing of fragmented biomedical data.
- Interpretable Classification of Time Series Using Euler Characteristic Surfaces (viability: 6): https://sciencetostartup.com/paper/interpretable-classification-of-time-series-using-euler-characteristic-surfaces - A novel classification framework using Euler Characteristic Surfaces for efficient and interpretable time series analysis.
- Multi-Mode Pneumatic Artificial Muscles Driven by Hybrid Positive-Negative Pressure (viability: 4): https://sciencetostartup.com/paper/multi-mode-pneumatic-artificial-muscles-driven-by-hybrid-positive-negative-pressure - IN-FOAMs are innovative inflatable artificial muscles designed for flexible and portable robotic movements.
- The Good, the Better, and the Best: Improving the Discriminability of Face Embeddings through Attribute-aware Learning (viability: 7): https://sciencetostartup.com/paper/the-good-the-better-and-the-best-improving-the-discriminability-of-face-embeddings-through-attribute-aware-learning - An attribute-aware face recognition architecture that enhances the discriminability of facial embeddings by focusing on identity-relevant attributes.
- Writer-R1: Enhancing Generative Writing in LLMs via Memory-augmented Replay Policy Optimization (viability: 8): https://sciencetostartup.com/paper/writer-r1-enhancing-generative-writing-in-llms-via-memory-augmented-replay-policy-optimization - A novel approach to enhance generative writing in LLMs using memory-augmented replay for improved evaluation and optimization.
- Muon Converges under Heavy-Tailed Noise: Nonconvex Hölder-Smooth Empirical Risk Minimization (viability: 2): https://sciencetostartup.com/paper/muon-converges-under-heavy-tailed-noise-nonconvex-h-lder-smooth-empirical-risk-minimization - Muon is an optimizer designed for stable training in the presence of heavy-tailed noise.
- Analyzing Error Sources in Global Feature Effect Estimation (viability: 2): https://sciencetostartup.com/paper/analyzing-error-sources-in-global-feature-effect-estimation - This paper explores the error sources in global feature effect estimation for machine learning models.
- Spatio-temporal probabilistic forecast using MMAF-guided learning (viability: 4): https://sciencetostartup.com/paper/spatio-temporal-probabilistic-forecast-using-mmaf-guided-learning - A novel probabilistic forecasting method using stochastic feed-forward neural networks for spatio-temporal datasets.
- Interference-Aware K-Step Reachable Communication in Multi-Agent Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/interference-aware-k-step-reachable-communication-in-multi-agent-reinforcement-learning - IA-KRC enhances multi-agent communication efficiency by optimizing partner selection under interference constraints.
- Thinking in Latents: Adaptive Anchor Refinement for Implicit Reasoning in LLMs (viability: 7): https://sciencetostartup.com/paper/thinking-in-latents-adaptive-anchor-refinement-for-implicit-reasoning-in-llms - AdaAnchor optimizes latent reasoning in LLMs by refining anchor vectors with adaptive halting for efficient computation.
- SRL-MAD: Structured Residual Latents for One-Class Morphing Attack Detection (viability: 7): https://sciencetostartup.com/paper/srl-mad-structured-residual-latents-for-one-class-morphing-attack-detection - SRL-MAD offers a novel approach to detect morphing attacks in biometric systems using structured residual Fourier representations.
- CrossADR: enhancing adverse drug reactions prediction for combination pharmacotherapy with cross-layer feature integration and cross-level associative learning (viability: 8): https://sciencetostartup.com/paper/crossadr-enhancing-adverse-drug-reactions-prediction-for-combination-pharmacotherapy-with-cross-layer-feature-integratio - CrossADR enhances adverse drug reactions prediction for combination pharmacotherapy using advanced graph neural networks.
- AnoleVLA: Lightweight Vision-Language-Action Model with Deep State Space Models for Mobile Manipulation (viability: 7): https://sciencetostartup.com/paper/anolevla-lightweight-vision-language-action-model-with-deep-state-space-models-for-mobile-manipulation - AnoleVLA is a lightweight vision-language-action model designed for efficient robotic manipulation in resource-constrained environments.
- Prompt Readiness Levels (PRL): a maturity scale and scoring framework for production grade prompt assets (viability: 2): https://sciencetostartup.com/paper/prompt-readiness-levels-prl-a-maturity-scale-and-scoring-framework-for-production-grade-prompt-assets - A framework for assessing the maturity and readiness of prompt engineering assets in generative AI.
- GUI-CEval: A Hierarchical and Comprehensive Chinese Benchmark for Mobile GUI Agents (viability: 4): https://sciencetostartup.com/paper/gui-ceval-a-hierarchical-and-comprehensive-chinese-benchmark-for-mobile-gui-agents - GUI-CEval is a comprehensive benchmark designed to evaluate Chinese mobile GUI agents across various applications and capabilities.
- Interpretable Predictability-Based AI Text Detection: A Replication Study (viability: 2): https://sciencetostartup.com/paper/interpretable-predictability-based-ai-text-detection-a-replication-study - This study replicates and extends a system for authorship attribution of machine-generated texts using multilingual models and stylometric features.
- Rethinking Machine Unlearning: Models Designed to Forget via Key Deletion (viability: 7): https://sciencetostartup.com/paper/rethinking-machine-unlearning-models-designed-to-forget-via-key-deletion - MUNKEY enables direct zero-shot forgetting in machine learning models, addressing privacy and data error challenges.
- Attention Residuals (viability: 3): https://sciencetostartup.com/paper/attention-residuals - Introducing Attention Residuals to enhance layer contribution in LLMs through selective aggregation.
- VTC-Bench: Evaluating Agentic Multimodal Models via Compositional Visual Tool Chaining (viability: 6): https://sciencetostartup.com/paper/vtc-bench-evaluating-agentic-multimodal-models-via-compositional-visual-tool-chaining - VTC-Bench is a benchmark for evaluating the tool-use proficiency of Multimodal Large Language Models in complex visual tasks.
- Training-free Detection of Generated Videos via Spatial-Temporal Likelihoods (viability: 8): https://sciencetostartup.com/paper/training-free-detection-of-generated-videos-via-spatial-temporal-likelihoods - STALL is a training-free detector for synthetic videos that leverages spatial-temporal likelihoods for reliable detection.
- One CT Unified Model Training Framework to Rule All Scanning Protocols (viability: 4): https://sciencetostartup.com/paper/one-ct-unified-model-training-framework-to-rule-all-scanning-protocols - A framework for enhancing CT imaging quality through uncertainty-guided manifold smoothing.
- Describing Agentic AI Systems with C4: Lessons from Industry Projects (viability: 4): https://sciencetostartup.com/paper/describing-agentic-ai-systems-with-c4-lessons-from-industry-projects - A documentation systematics for Agentic AI systems that enhances transparency and maintainability in industrial applications.
- MER-Bench: A Comprehensive Benchmark for Multimodal Meme Reappraisal (viability: 8): https://sciencetostartup.com/paper/mer-bench-a-comprehensive-benchmark-for-multimodal-meme-reappraisal - MER-Bench enables the transformation of negative memes into constructive ones through emotion-controllable multimodal generation.
- Reference-Free Omnidirectional Stereo Matching via Multi-View Consistency Maximization (viability: 7): https://sciencetostartup.com/paper/reference-free-omnidirectional-stereo-matching-via-multi-view-consistency-maximization - FreeOmniMVS offers a novel reference-free framework for robust omnidirectional depth estimation using multi-view consistency maximization.
- Consequentialist Objectives and Catastrophe (viability: 2): https://sciencetostartup.com/paper/consequentialist-objectives-and-catastrophe - This paper explores the risks of catastrophic outcomes from AIs with misspecified objectives in complex environments.
- Riemannian Motion Generation: A Unified Framework for Human Motion Representation and Generation via Riemannian Flow Matching (viability: 3): https://sciencetostartup.com/paper/riemannian-motion-generation-a-unified-framework-for-human-motion-representation-and-generation-via-riemannian-flow-matc - A framework for generating human motion using Riemannian geometry.
- CycleRL: Sim-to-Real Deep Reinforcement Learning for Robust Autonomous Bicycle Control (viability: 8): https://sciencetostartup.com/paper/cyclerl-sim-to-real-deep-reinforcement-learning-for-robust-autonomous-bicycle-control - CycleRL is a sim-to-real deep reinforcement learning framework for robust autonomous bicycle control, leveraging advanced training techniques for real-world adaptability.
- Molecular Identifier Visual Prompt and Verifiable Reinforcement Learning for Chemical Reaction Diagram Parsing (viability: 8): https://sciencetostartup.com/paper/molecular-identifier-visual-prompt-and-verifiable-reinforcement-learning-for-chemical-reaction-diagram-parsing - A novel approach to enhance chemical reaction diagram parsing using visual prompts and reinforcement learning.
- TrajFlow: Nation-wide Pseudo GPS Trajectory Generation with Flow Matching Models (viability: 7): https://sciencetostartup.com/paper/trajflow-nation-wide-pseudo-gps-trajectory-generation-with-flow-matching-models - TrajFlow is a novel flow-matching-based model for generating nationwide pseudo-GPS trajectory data to enhance urban planning and traffic management.
- Clue Matters: Leveraging Latent Visual Clues to Empower Video Reasoning (viability: 3): https://sciencetostartup.com/paper/clue-matters-leveraging-latent-visual-clues-to-empower-video-reasoning - ClueNet enhances video question answering by improving visual clue extraction and reasoning alignment.
- Pretraining and Benchmarking Modern Encoders for Latvian (viability: 7): https://sciencetostartup.com/paper/pretraining-and-benchmarking-modern-encoders-for-latvian - A suite of pretrained Latvian-specific encoders that outperform existing models in NLP tasks.
- Edit2Interp: Adapting Image Foundation Models from Spatial Editing to Video Frame Interpolation with Few-Shot Learning (viability: 7): https://sciencetostartup.com/paper/edit2interp-adapting-image-foundation-models-from-spatial-editing-to-video-frame-interpolation-with-few-shot-learning - A novel approach to adapt image editing models for video frame interpolation using few-shot learning.
- MONET: Modeling and Optimization of neural NEtwork Training from Edge to Data Centers (viability: 3): https://sciencetostartup.com/paper/monet-modeling-and-optimization-of-neural-network-training-from-edge-to-data-centers - MONET is a framework for optimizing neural network training on heterogeneous dataflow accelerators.
- How Log-Barrier Helps Exploration in Policy Optimization (viability: 2): https://sciencetostartup.com/paper/how-log-barrier-helps-exploration-in-policy-optimization - This paper proposes a log-barrier regularization for Stochastic Gradient Bandit to enhance exploration in policy optimization.
- Thermal Image Refinement with Depth Estimation using Recurrent Networks for Monocular ORB-SLAM3 (viability: 7): https://sciencetostartup.com/paper/thermal-image-refinement-with-depth-estimation-using-recurrent-networks-for-monocular-orb-slam3 - A novel thermal image refinement and depth estimation pipeline for UAVs enabling robust SLAM in low-light conditions.
- OrgForge: A Multi-Agent Simulation Framework for Verifiable Synthetic Corporate Corpora (viability: 7): https://sciencetostartup.com/paper/orgforge-a-multi-agent-simulation-framework-for-verifiable-synthetic-corporate-corpora - OrgForge is an open-source multi-agent simulation framework that generates verifiable synthetic corporate datasets for RAG pipelines.
- DP-S4S: Accurate and Scalable Select-Join-Aggregate Query Processing with User-Level Differential Privacy (viability: 5): https://sciencetostartup.com/paper/dp-s4s-accurate-and-scalable-select-join-aggregate-query-processing-with-user-level-differential-privacy - DP-S4S offers a scalable solution for Select-Join-Aggregate queries while ensuring user-level differential privacy.
- Exposing Cross-Modal Consistency for Fake News Detection in Short-Form Videos (viability: 7): https://sciencetostartup.com/paper/exposing-cross-modal-consistency-for-fake-news-detection-in-short-form-videos - MAGIC3 is a cross-modal consistency detector for identifying fake news in short-form videos.
- MMSpec: Benchmarking Speculative Decoding for Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/mmspec-benchmarking-speculative-decoding-for-vision-language-models - MMSpec benchmarks speculative decoding techniques for vision-language models to enhance inference speed and efficiency.
- Beyond Benchmark Islands: Toward Representative Trustworthiness Evaluation for Agentic AI (viability: 8): https://sciencetostartup.com/paper/beyond-benchmark-islands-toward-representative-trustworthiness-evaluation-for-agentic-ai - A comprehensive framework for evaluating the trustworthiness of agentic AI systems in real-world scenarios.
- ReMAP-DP: Reprojected Multi-view Aligned PointMaps for Diffusion Policy (viability: 8): https://sciencetostartup.com/paper/remap-dp-reprojected-multi-view-aligned-pointmaps-for-diffusion-policy - ReMAP-DP enhances robot manipulation tasks by integrating 3D spatial awareness with advanced diffusion policies.
- Anchoring Emotions in Text: Robust Multimodal Fusion for Mimicry Intensity Estimation (viability: 7): https://sciencetostartup.com/paper/anchoring-emotions-in-text-robust-multimodal-fusion-for-mimicry-intensity-estimation - TAEMI is a multimodal framework for estimating emotional mimicry intensity using text as a stable anchor to improve robustness against noisy signals.
- Why Agents Compromise Safety Under Pressure (viability: 4): https://sciencetostartup.com/paper/why-agents-compromise-safety-under-pressure - A study on how LLM agents compromise safety under pressure and strategies to mitigate this issue.
- Voronoi-based Second-order Descriptor with Whitened Metric in LiDAR Place Recognition (viability: 4): https://sciencetostartup.com/paper/voronoi-based-second-order-descriptor-with-whitened-metric-in-lidar-place-recognition - A novel second-order pooling method for LiDAR place recognition that enhances descriptor stability and performance.
- Learning from Mistakes: Post-Training for Driving VLA with Takeover Data (viability: 7): https://sciencetostartup.com/paper/learning-from-mistakes-post-training-for-driving-vla-with-takeover-data - TakeVLA enhances autonomous driving safety by proactively teaching models to learn from mistakes using innovative post-training techniques.
- Rethinking LLM Watermark Detection in Black-Box Settings: A Non-Intrusive Third-Party Framework (viability: 7): https://sciencetostartup.com/paper/rethinking-llm-watermark-detection-in-black-box-settings-a-non-intrusive-third-party-framework - TTP-Detect offers a non-intrusive framework for third-party verification of LLM watermarks, enhancing model governance.
- GeoNVS: Geometry Grounded Video Diffusion for Novel View Synthesis (viability: 7): https://sciencetostartup.com/paper/geonvs-geometry-grounded-video-diffusion-for-novel-view-synthesis - GeoNVS enhances novel view synthesis by improving geometric fidelity and camera controllability through innovative 3D geometric guidance.
- Lightweight User-Personalization Method for Closed Split Computing (viability: 7): https://sciencetostartup.com/paper/lightweight-user-personalization-method-for-closed-split-computing - SALT is a lightweight adaptation framework for enhancing user personalization in closed Split Computing systems.
- CyCLeGen: Cycle-Consistent Layout Prediction and Image Generation in Vision Foundation Models (viability: 7): https://sciencetostartup.com/paper/cyclegen-cycle-consistent-layout-prediction-and-image-generation-in-vision-foundation-models - CyCLeGen is a unified vision-language model that enhances image understanding and generation through cycle-consistent learning.
- SFedHIFI: Fire Rate-Based Heterogeneous Information Fusion for Spiking Federated Learning (viability: 4): https://sciencetostartup.com/paper/sfedhifi-fire-rate-based-heterogeneous-information-fusion-for-spiking-federated-learning - SFedHIFI enables efficient heterogeneous federated learning for resource-constrained clients using spiking neural networks.
- Learning Question-Aware Keyframe Selection with Synthetic Supervision for Video Question Answering (viability: 7): https://sciencetostartup.com/paper/learning-question-aware-keyframe-selection-with-synthetic-supervision-for-video-question-answering - A framework for efficient keyframe selection in video question answering using synthetic supervision.
- Pansharpening for Thin-Cloud Contaminated Remote Sensing Images: A Unified Framework and Benchmark Dataset (viability: 7): https://sciencetostartup.com/paper/pansharpening-for-thin-cloud-contaminated-remote-sensing-images-a-unified-framework-and-benchmark-dataset - A unified framework for pansharpening remote sensing images contaminated by thin clouds, featuring a novel dataset for benchmarking.
- GT-PCQA: Geometry-Texture Decoupled Point Cloud Quality Assessment with MLLM (viability: 4): https://sciencetostartup.com/paper/gt-pcqa-geometry-texture-decoupled-point-cloud-quality-assessment-with-mllm - GT-PCQA is a novel framework for no-reference point cloud quality assessment leveraging multi-modal large language models.
- Bridging Scene Generation and Planning: Driving with World Model via Unifying Vision and Motion Representation (viability: 8): https://sciencetostartup.com/paper/bridging-scene-generation-and-planning-driving-with-world-model-via-unifying-vision-and-motion-representation - WorldDrive unifies scene generation and motion planning for enhanced autonomous driving performance.
- FairMed-XGB: A Bayesian-Optimised Multi-Metric Framework with Explainability for Demographic Equity in Critical Healthcare Data (viability: 6): https://sciencetostartup.com/paper/fairmed-xgb-a-bayesian-optimised-multi-metric-framework-with-explainability-for-demographic-equity-in-critical-healthcar - FairMed-XGB is a framework that reduces gender bias in critical healthcare predictions while maintaining model performance and transparency.
- Spiking Layer-Adaptive Magnitude-based Pruning (viability: 4): https://sciencetostartup.com/paper/spiking-layer-adaptive-magnitude-based-pruning - SLAMP is a pruning framework for Spiking Neural Networks that optimizes layer connectivity while maintaining performance.
- Ultra-Early Prediction of Tipping Points: Integrating Dynamical Measures with Reservoir Computing (viability: 4): https://sciencetostartup.com/paper/ultra-early-prediction-of-tipping-points-integrating-dynamical-measures-with-reservoir-computing - A model-free framework for ultra-early prediction of tipping points in complex dynamical systems using reservoir computing.
- RS-WorldModel: a Unified Model for Remote Sensing Understanding and Future Sense Forecasting (viability: 8): https://sciencetostartup.com/paper/rs-worldmodel-a-unified-model-for-remote-sensing-understanding-and-future-sense-forecasting - RS-WorldModel is a unified model for remote sensing that enhances understanding of changes and forecasts future scenes using a rich dataset and advanced training techniques.
- FAR-Drive: Frame-AutoRegressive Video Generation in Closed-Loop Autonomous Driving (viability: 8): https://sciencetostartup.com/paper/far-drive-frame-autoregressive-video-generation-in-closed-loop-autonomous-driving - FAR-Drive is a closed-loop video generation framework for autonomous driving that ensures high fidelity and low latency.
- LLM as Graph Kernel: Rethinking Message Passing on Text-Rich Graphs (viability: 3): https://sciencetostartup.com/paper/llm-as-graph-kernel-rethinking-message-passing-on-text-rich-graphs - RAMP redefines message passing in text-rich graphs by using LLMs as graph-native aggregation operators.
- Relevance Feedback in Text-to-Image Diffusion: A Training-Free And Model-Agnostic Interactive Framework (viability: 7): https://sciencetostartup.com/paper/relevance-feedback-in-text-to-image-diffusion-a-training-free-and-model-agnostic-interactive-framework - RFD is an interactive framework that enhances text-to-image generation by allowing users to provide visual feedback instead of textual prompts.
- Video-CoE: Reinforcing Video Event Prediction via Chain of Events (viability: 7): https://sciencetostartup.com/paper/video-coe-reinforcing-video-event-prediction-via-chain-of-events - Video-CoE enhances video event prediction by constructing temporal event chains for improved reasoning and accuracy.
- Masked BRep Autoencoder via Hierarchical Graph Transformer (viability: 6): https://sciencetostartup.com/paper/masked-brep-autoencoder-via-hierarchical-graph-transformer - A self-supervised learning framework for enhancing CAD model representations for various downstream tasks.
- Workflow-Aware Structured Layer Decomposition for Illustration Production (viability: 8): https://sciencetostartup.com/paper/workflow-aware-structured-layer-decomposition-for-illustration-production - A framework for structured layer decomposition in anime illustration production, enhancing controllability and enabling advanced editing tasks.
- Directional Routing in Transformers (viability: 4): https://sciencetostartup.com/paper/directional-routing-in-transformers - Directional routing enhances transformer efficiency by optimizing attention head suppression with minimal parameter cost.
- $\text{F}^2\text{HDR}$: Two-Stage HDR Video Reconstruction via Flow Adapter and Physical Motion Modeling (viability: 7): https://sciencetostartup.com/paper/text-f-2-text-hdr-two-stage-hdr-video-reconstruction-via-flow-adapter-and-physical-motion-modeling - A two-stage framework for reconstructing high dynamic range videos from low dynamic range sequences, enhancing detail and reducing ghosting.
- Bayesian Inference for Missing Physics (viability: 4): https://sciencetostartup.com/paper/bayesian-inference-for-missing-physics - A Bayesian approach to symbolic regression that quantifies uncertainty in discovered equations for biological and chemical systems.
- EditHF-1M: A Million-Scale Rich Human Preference Feedback for Image Editing (viability: 8): https://sciencetostartup.com/paper/edithf-1m-a-million-scale-rich-human-preference-feedback-for-image-editing - A million-scale human preference dataset and evaluation model for optimizing text-guided image editing.
- ILV: Iterative Latent Volumes for Fast and Accurate Sparse-View CT Reconstruction (viability: 8): https://sciencetostartup.com/paper/ilv-iterative-latent-volumes-for-fast-and-accurate-sparse-view-ct-reconstruction - ILV is a novel framework for fast and accurate 3D reconstruction from sparse-view CT projections, enhancing clinical imaging workflows.
- Fine-tuning RoBERTa for CVE-to-CWE Classification: A 125M Parameter Model Competitive with LLMs (viability: 8): https://sciencetostartup.com/paper/fine-tuning-roberta-for-cve-to-cwe-classification-a-125m-parameter-model-competitive-with-llms - A lightweight model for precise CVE-to-CWE classification enhancing cybersecurity vulnerability management.
- TopoVST: Toward Topology-fidelitous Vessel Skeleton Tracking (viability: 8): https://sciencetostartup.com/paper/topovst-toward-topology-fidelitous-vessel-skeleton-tracking - TopoVST is an advanced vessel skeleton tracker that ensures topological fidelity for clinical applications.
- PerlAD: Towards Enhanced Closed-loop End-to-end Autonomous Driving with Pseudo-simulation-based Reinforcement Learning (viability: 8): https://sciencetostartup.com/paper/perlad-towards-enhanced-closed-loop-end-to-end-autonomous-driving-with-pseudo-simulation-based-reinforcement-learning - PerlAD revolutionizes autonomous driving with efficient closed-loop training using pseudo-simulation-based reinforcement learning.
- ExPosST: Explicit Positioning with Adaptive Masking for LLM-Based Simultaneous Machine Translation (viability: 7): https://sciencetostartup.com/paper/exposst-explicit-positioning-with-adaptive-masking-for-llm-based-simultaneous-machine-translation - ExPosST enhances simultaneous machine translation by resolving positional mismatches in LLMs with explicit position allocation.
- From Folding Mechanics to Robotic Function: A Unified Modeling Framework for Compliant Origami (viability: 3): https://sciencetostartup.com/paper/from-folding-mechanics-to-robotic-function-a-unified-modeling-framework-for-compliant-origami - A unified modeling framework for compliant origami that enhances robotic functionality through predictive modeling and stability programming.
- BiTro: Bidirectional Transfer Learning Enhances Bulk and Spatial Transcriptomics Prediction in Cancer Pathological Images (viability: 7): https://sciencetostartup.com/paper/bitro-bidirectional-transfer-learning-enhances-bulk-and-spatial-transcriptomics-prediction-in-cancer-pathological-images - BiTro enhances cancer pathological analysis by improving predictions in bulk and spatial transcriptomics through a novel bidirectional transfer learning framework.
- Informative Perturbation Selection for Uncertainty-Aware Post-hoc Explanations (viability: 6): https://sciencetostartup.com/paper/informative-perturbation-selection-for-uncertainty-aware-post-hoc-explanations - EAGLE is an information-theoretic framework for generating reliable post-hoc explanations of black-box ML models.
- LLMs as Signal Detectors: Sensitivity, Bias, and the Temperature-Criterion Analogy (viability: 4): https://sciencetostartup.com/paper/llms-as-signal-detectors-sensitivity-bias-and-the-temperature-criterion-analogy - This study applies Signal Detection Theory to evaluate the calibration of large language models, revealing insights into their sensitivity and bias.
- Balancing Saliency and Coverage: Semantic Prominence-Aware Budgeting for Visual Token Compression in VLMs (viability: 3): https://sciencetostartup.com/paper/balancing-saliency-and-coverage-semantic-prominence-aware-budgeting-for-visual-token-compression-in-vlms - PromPrune is a framework for adaptive visual token selection that optimizes saliency and coverage in Vision-Language Models.
- Decision-Level Ordinal Modeling for Multimodal Essay Scoring with Large Language Models (viability: 7): https://sciencetostartup.com/paper/decision-level-ordinal-modeling-for-multimodal-essay-scoring-with-large-language-models - A novel approach to automated essay scoring that enhances accuracy through explicit ordinal decision-making and multimodal integration.
- ViSA: Visited-State Augmentation for Generalized Goal-Space Contrastive Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/visa-visited-state-augmentation-for-generalized-goal-space-contrastive-reinforcement-learning - ViSA enhances goal-conditioned reinforcement learning by augmenting hard-to-visit state samples for improved policy learning.
- PASTE: Physics-Aware Scattering Topology Embedding Framework for SAR Object Detection (viability: 7): https://sciencetostartup.com/paper/paste-physics-aware-scattering-topology-embedding-framework-for-sar-object-detection - PASTE integrates physics-based scattering topology into SAR object detection for improved accuracy and interpretability.
- SpiralDiff: Spiral Diffusion with LoRA for RGB-to-RAW Conversion Across Cameras (viability: 9): https://sciencetostartup.com/paper/spiraldiff-spiral-diffusion-with-lora-for-rgb-to-raw-conversion-across-cameras - SpiralDiff revolutionizes RGB-to-RAW image conversion using a diffusion-based framework with camera-specific adaptations.
- Customizing ChatGPT for Second Language Speaking Practice: Genuine Support or Just a Marketing Gimmick? (viability: 3): https://sciencetostartup.com/paper/customizing-chatgpt-for-second-language-speaking-practice-genuine-support-or-just-a-marketing-gimmick - A study exploring the customization of ChatGPT for enhancing ESL speaking practices.
- LLMind: Bio-inspired Training-free Adaptive Visual Representations for Vision-Language Models (viability: 8): https://sciencetostartup.com/paper/llmind-bio-inspired-training-free-adaptive-visual-representations-for-vision-language-models - LLMind offers a training-free framework for adaptive visual representations in Vision-Language Models, enhancing efficiency and performance.
- RealVLG-R1: A Large-Scale Real-World Visual-Language Grounding Benchmark for Robotic Perception and Manipulation (viability: 9): https://sciencetostartup.com/paper/realvlg-r1-a-large-scale-real-world-visual-language-grounding-benchmark-for-robotic-perception-and-manipulation - RealVLG-R1 revolutionizes robotic manipulation by integrating visual-language grounding with a comprehensive dataset and model for real-world applications.
- Seismic full-waveform inversion based on a physics-driven generative adversarial network (viability: 6): https://sciencetostartup.com/paper/seismic-full-waveform-inversion-based-on-a-physics-driven-generative-adversarial-network - A physics-driven GAN approach to enhance full-waveform inversion for geophysical imaging.
- A Hybrid AI and Rule-Based Decision Support System for Disease Diagnosis and Management Using Labs (viability: 6): https://sciencetostartup.com/paper/a-hybrid-ai-and-rule-based-decision-support-system-for-disease-diagnosis-and-management-using-labs - A Clinical Decision Support System that combines AI predictive modeling with rule-based expert systems to assist physicians in diagnosing diseases.
- Developing an English-Efik Corpus and Machine Translation System for Digitization Inclusion (viability: 5): https://sciencetostartup.com/paper/developing-an-english-efik-corpus-and-machine-translation-system-for-digitization-inclusion - A machine translation system aimed at improving English-Efik translation for low-resource languages.
- IgPose: A Generative Data-Augmented Pipeline for Robust Immunoglobulin-Antigen Binding Prediction (viability: 8): https://sciencetostartup.com/paper/igpose-a-generative-data-augmented-pipeline-for-robust-immunoglobulin-antigen-binding-prediction - IgPose is a generative data-augmented framework for robust immunoglobulin-antigen binding prediction, enhancing antibody discovery pipelines.
- A Self-Evolving Defect Detection Framework for Industrial Photovoltaic Systems (viability: 7): https://sciencetostartup.com/paper/a-self-evolving-defect-detection-framework-for-industrial-photovoltaic-systems - A self-evolving framework for defect detection in photovoltaic systems that adapts to changing conditions.
- Sample-Efficient Hypergradient Estimation for Decentralized Bi-Level Reinforcement Learning (viability: 4): https://sciencetostartup.com/paper/sample-efficient-hypergradient-estimation-for-decentralized-bi-level-reinforcement-learning - A novel method for efficient hypergradient estimation in decentralized bi-level reinforcement learning settings.
- Shopping Companion: A Memory-Augmented LLM Agent for Real-World E-Commerce Tasks (viability: 7): https://sciencetostartup.com/paper/shopping-companion-a-memory-augmented-llm-agent-for-real-world-e-commerce-tasks - Shopping Companion is a memory-augmented LLM agent designed to enhance e-commerce tasks by capturing long-term user preferences.
- Video Detector: A Dual-Phase Vision-Based System for Real-Time Traffic Intersection Control and Intelligent Transportation Analysis (viability: 7): https://sciencetostartup.com/paper/video-detector-a-dual-phase-vision-based-system-for-real-time-traffic-intersection-control-and-intelligent-transportatio - Video Detector is a dual-phase vision-based system for real-time traffic intersection control and intelligent transportation analysis.
- Architecture-Agnostic Feature Synergy for Universal Defense Against Heterogeneous Generative Threats (viability: 8): https://sciencetostartup.com/paper/architecture-agnostic-feature-synergy-for-universal-defense-against-heterogeneous-generative-threats - A framework for universal defense against diverse generative threats using architecture-agnostic feature synergy.
- From Horizontal to Rotated: Cross-View Object Geo-Localization with Orientation Awareness (viability: 8): https://sciencetostartup.com/paper/from-horizontal-to-rotated-cross-view-object-geo-localization-with-orientation-awareness - OSGeo revolutionizes cross-view object geo-localization by using Rotated Bounding Boxes for high precision with lower annotation costs.
- PCodeTrans: Translate Decompiled Pseudocode to Compilable and Executable Equivalent (viability: 8): https://sciencetostartup.com/paper/pcodetrans-translate-decompiled-pseudocode-to-compilable-and-executable-equivalent - PCodeTrans is a feedback-driven framework that translates decompiled pseudocode into compilable and executable code with high accuracy.
- Surgical Robot, Path Planning, Joint Space, Riemannian Manifolds (viability: 4): https://sciencetostartup.com/paper/surgical-robot-path-planning-joint-space-riemannian-manifolds - A novel path planning method for robotic surgery that optimizes joint movement using Riemannian manifolds.
- AutoMoT: A Unified Vision-Language-Action Model with Asynchronous Mixture-of-Transformers for End-to-End Autonomous Driving (viability: 8): https://sciencetostartup.com/paper/automot-a-unified-vision-language-action-model-with-asynchronous-mixture-of-transformers-for-end-to-end-autonomous-drivi - A unified vision-language-action model for enhancing autonomous driving performance through efficient reasoning and action generation.
- From Artefact to Insight: Efficient Low-Rank Adaptation of BrushNet for Scanning Probe Microscopy Image Restoration (viability: 7): https://sciencetostartup.com/paper/from-artefact-to-insight-efficient-low-rank-adaptation-of-brushnet-for-scanning-probe-microscopy-image-restoration - A lightweight framework for efficient restoration of Scanning Probe Microscopy images using low-rank adaptation.
- Personalized Federated Learning with Residual Fisher Information for Medical Image Segmentation (viability: 7): https://sciencetostartup.com/paper/personalized-federated-learning-with-residual-fisher-information-for-medical-image-segmentation - A personalized federated learning framework that enhances medical image segmentation by adapting models to client-specific data without compromising privacy.
- Lost in Aggregation: On a Fundamental Expressivity Limit of Message-Passing Graph Neural Networks (viability: 2): https://sciencetostartup.com/paper/lost-in-aggregation-on-a-fundamental-expressivity-limit-of-message-passing-graph-neural-networks - This paper explores fundamental limitations in the expressivity of Message-Passing Graph Neural Networks.
- IntegratingWeather Foundation Model and Satellite to Enable Fine-Grained Solar Irradiance Forecasting (viability: 9): https://sciencetostartup.com/paper/integratingweather-foundation-model-and-satellite-to-enable-fine-grained-solar-irradiance-forecasting - Baguan-solar integrates weather models and satellite imagery for precise solar irradiance forecasting.
- ContiGuard: A Framework for Continual Toxicity Detection Against Evolving Evasive Perturbations (viability: 3): https://sciencetostartup.com/paper/contiguard-a-framework-for-continual-toxicity-detection-against-evolving-evasive-perturbations - ContiGuard is a framework for continual toxicity detection that adapts to evolving evasion tactics in online content.
- Real-Time Driver Safety Scoring Through Inverse Crash Probability Modeling (viability: 7): https://sciencetostartup.com/paper/real-time-driver-safety-scoring-through-inverse-crash-probability-modeling - SafeDriver-IQ transforms binary crash predictions into continuous safety scores for real-time driver feedback.
- The Impact of Ideological Discourses in RAG: A Case Study with COVID-19 Treatments (viability: 4): https://sciencetostartup.com/paper/the-impact-of-ideological-discourses-in-rag-a-case-study-with-covid-19-treatments - This research explores the influence of ideological texts on LLM outputs in the context of Retrieval-Augmented Generation.
- DamageArbiter: A CLIP-Enhanced Multimodal Arbitration Framework for Hurricane Damage Assessment from Street-View Imagery (viability: 3): https://sciencetostartup.com/paper/damagearbiter-a-clip-enhanced-multimodal-arbitration-framework-for-hurricane-damage-assessment-from-street-view-imagery - DamageArbiter enhances hurricane damage assessment from street-view imagery using a multimodal arbitration framework.
- Ablate and Rescue: A Causal Analysis of Residual Stream Hyper-Connections (viability: 8): https://sciencetostartup.com/paper/ablate-and-rescue-a-causal-analysis-of-residual-stream-hyper-connections - An open-source multi-stream transformer model that addresses representation collapse through causal analysis of residual connections.
- Halfway to 3D: Ensembling 2.5D and 3D Models for Robust COVID-19 CT Diagnosis (viability: 8): https://sciencetostartup.com/paper/halfway-to-3d-ensembling-2-5d-and-3d-models-for-robust-covid-19-ct-diagnosis - A deep learning framework that enhances COVID-19 detection from chest CT scans by integrating 2.5D and 3D models for improved accuracy.
- Dataset Distillation Efficiently Encodes Low-Dimensional Representations from Gradient-Based Learning of Non-Linear Tasks (viability: 3): https://sciencetostartup.com/paper/dataset-distillation-efficiently-encodes-low-dimensional-representations-from-gradient-based-learning-of-non-linear-task - This paper presents a theoretical framework for dataset distillation to optimize data storage and training costs.
- SemanticFace: Semantic Facial Action Estimation via Semantic Distillation in Interpretable Space (viability: 7): https://sciencetostartup.com/paper/semanticface-semantic-facial-action-estimation-via-semantic-distillation-in-interpretable-space - SemanticFace offers interpretable facial action estimation for applications in avatar control and human-computer interaction.
- Two Birds, One Projection: Harmonizing Safety and Utility in LVLMs via Inference-time Feature Projection (viability: 7): https://sciencetostartup.com/paper/two-birds-one-projection-harmonizing-safety-and-utility-in-lvlms-via-inference-time-feature-projection - A novel inference-time defense mechanism for Large Vision-Language Models that enhances both safety and utility.
- Planning as Goal Recognition: Deriving Heuristics from Intention Models - Extended Version (viability: 2): https://sciencetostartup.com/paper/planning-as-goal-recognition-deriving-heuristics-from-intention-models-extended-version - This paper explores a new framework for deriving heuristics from intention models in classical planning.
- RadarXFormer: Robust Object Detection via Cross-Dimension Fusion of 4D Radar Spectra and Images for Autonomous Driving (viability: 7): https://sciencetostartup.com/paper/radarxformer-robust-object-detection-via-cross-dimension-fusion-of-4d-radar-spectra-and-images-for-autonomous-driving - RadarXFormer enhances object detection in autonomous driving by fusing 4D radar spectra with RGB images for improved robustness.
- RAZOR: Ratio-Aware Layer Editing for Targeted Unlearning in Vision Transformers and Diffusion Models (viability: 7): https://sciencetostartup.com/paper/razor-ratio-aware-layer-editing-for-targeted-unlearning-in-vision-transformers-and-diffusion-models - RAZOR provides a lightweight framework for efficient unlearning in transformer models without retraining.
- SimCert: Probabilistic Certification for Behavioral Similarity in Deep Neural Network Compression (viability: 7): https://sciencetostartup.com/paper/simcert-probabilistic-certification-for-behavioral-similarity-in-deep-neural-network-compression - SimCert offers a probabilistic certification framework to ensure behavioral fidelity in compressed deep neural networks.
- M2IR: Proactive All-in-One Image Restoration via Mamba-style Modulation and Mixture-of-Experts (viability: 8): https://sciencetostartup.com/paper/m2ir-proactive-all-in-one-image-restoration-via-mamba-style-modulation-and-mixture-of-experts - M2IR is a proactive image restoration framework that enhances detail recovery by actively controlling degradation propagation.
- Ego to World: Collaborative Spatial Reasoning in Embodied Systems via Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/ego-to-world-collaborative-spatial-reasoning-in-embodied-systems-via-reinforcement-learning - Ego-to-World (E2W) benchmark and CoRL framework enhance collaborative spatial reasoning in multi-agent systems.
- A Unified Calibration Framework for Coordinate and Kinematic Parameters in Dual-Arm Robots (viability: 2): https://sciencetostartup.com/paper/a-unified-calibration-framework-for-coordinate-and-kinematic-parameters-in-dual-arm-robots - A novel framework for unified calibration of coordinate transformations and kinematic parameters in dual-arm robots.
- HiMemVLN: Enhancing Reliability of Open-Source Zero-Shot Vision-and-Language Navigation with Hierarchical Memory System (viability: 8): https://sciencetostartup.com/paper/himemvln-enhancing-reliability-of-open-source-zero-shot-vision-and-language-navigation-with-hierarchical-memory-system - HiMemVLN enhances open-source vision-language navigation by addressing Navigation Amnesia with a Hierarchical Memory System.
- Knowledge Activation: AI Skills as the Institutional Knowledge Primitive for Agentic Software Development (viability: 3): https://sciencetostartup.com/paper/knowledge-activation-ai-skills-as-the-institutional-knowledge-primitive-for-agentic-software-development - A framework for transforming institutional knowledge into actionable specifications for AI agents in enterprise software development.
- VorTEX: Various overlap ratio for Target speech EXtraction (viability: 7): https://sciencetostartup.com/paper/vortex-various-overlap-ratio-for-target-speech-extraction - VorTEX is a novel architecture for target speech extraction that excels in handling various overlap ratios.
- OpenReservoirComputing: GPU-Accelerated Reservoir Computing in JAX (viability: 3): https://sciencetostartup.com/paper/openreservoircomputing-gpu-accelerated-reservoir-computing-in-jax - OpenReservoirComputing is a Python library for GPU-accelerated reservoir computing using JAX.
- Universe Routing: Why Self-Evolving Agents Need Epistemic Control (viability: 2): https://sciencetostartup.com/paper/universe-routing-why-self-evolving-agents-need-epistemic-control - This paper addresses the critical failure of lifelong agents in decision-making by proposing a universe routing problem for epistemic control.
- Preconditioned One-Step Generative Modeling for Bayesian Inverse Problems in Function Spaces (viability: 2): https://sciencetostartup.com/paper/preconditioned-one-step-generative-modeling-for-bayesian-inverse-problems-in-function-spaces - A novel machine-learning algorithm for efficient Bayesian inference in function spaces.
- Multi-Task Genetic Algorithm with Multi-Granularity Encoding for Protein-Nucleotide Binding Site Prediction (viability: 4): https://sciencetostartup.com/paper/multi-task-genetic-algorithm-with-multi-granularity-encoding-for-protein-nucleotide-binding-site-prediction - A novel framework for enhancing protein-nucleotide binding site prediction using a Multi-Task Genetic Algorithm and Multi-Granularity Encoding.
- Global Truncated Loss Minimization for Robust and Threshold-Resilient Geometric Estimation (viability: 4): https://sciencetostartup.com/paper/global-truncated-loss-minimization-for-robust-and-threshold-resilient-geometric-estimation - GTM is a novel framework for robust geometric estimation that minimizes truncated losses using a global branch-and-bound approach.
- Face-to-Face: A Video Dataset for Multi-Person Interaction Modeling (viability: 7): https://sciencetostartup.com/paper/face-to-face-a-video-dataset-for-multi-person-interaction-modeling - A comprehensive dataset for modeling multi-person interactions in video, enabling advanced conversational AI applications.
- GARCH-FIS: A Hybrid Forecasting Model with Dynamic Volatility-Driven Parameter Adaptation (viability: 7): https://sciencetostartup.com/paper/garch-fis-a-hybrid-forecasting-model-with-dynamic-volatility-driven-parameter-adaptation - GARCH-FIS is a hybrid forecasting model that adapts to market volatility for improved financial time series predictions.
- LaPro-DTA: Latent Dual-View Drug Representations and Salient Protein Feature Extraction for Generalizable Drug--Target Affinity Prediction (viability: 7): https://sciencetostartup.com/paper/lapro-dta-latent-dual-view-drug-representations-and-salient-protein-feature-extraction-for-generalizable-drug-target-aff - LaPro-DTA enhances drug-target affinity prediction by leveraging dual-view representations and salient feature extraction.
- Mind-of-Director: Multi-modal Agent-Driven Film Previsualization via Collaborative Decision-Making (viability: 7): https://sciencetostartup.com/paper/mind-of-director-multi-modal-agent-driven-film-previsualization-via-collaborative-decision-making - Mind-of-Director is a multi-modal agent-driven framework that streamlines film previsualization through collaborative decision-making.
- GraspALL: Adaptive Structural Compensation from Illumination Variation for Robotic Garment Grasping in Any Low-Light Conditions (viability: 7): https://sciencetostartup.com/paper/graspall-adaptive-structural-compensation-from-illumination-variation-for-robotic-garment-grasping-in-any-low-light-cond - GraspALL enhances robotic garment grasping accuracy in low-light conditions through adaptive feature fusion of RGB and non-RGB modalities.
- Exploring the dynamic properties and motion reproducibility of a small upper-body humanoid robot with 13-DOF pneumatic actuation for data-driven control (viability: 5): https://sciencetostartup.com/paper/exploring-the-dynamic-properties-and-motion-reproducibility-of-a-small-upper-body-humanoid-robot-with-13-dof-pneumatic-a - A humanoid robot with 13-DOF pneumatic actuation and a data-driven controller for improved motion control.
- CORAL: COntextual Reasoning And Local Planning in A Hierarchical VLM Framework for Underwater Monitoring (viability: 3): https://sciencetostartup.com/paper/coral-contextual-reasoning-and-local-planning-in-a-hierarchical-vlm-framework-for-underwater-monitoring - CORAL enhances underwater monitoring by combining vision-language models with dynamic planning for autonomous vehicles.
- Orthogonal Subspace Clustering: Enhancing High-Dimensional Data Analysis through Adaptive Dimensionality Reduction and Efficient Clustering (viability: 3): https://sciencetostartup.com/paper/orthogonal-subspace-clustering-enhancing-high-dimensional-data-analysis-through-adaptive-dimensionality-reduction-and-ef - Orthogonal Subspace Clustering (OSC) enhances clustering efficiency in high-dimensional data through adaptive dimensionality reduction.
- Information Asymmetry across Language Varieties: A Case Study on Cantonese-Mandarin and Bavarian-German QA (viability: 4): https://sciencetostartup.com/paper/information-asymmetry-across-language-varieties-a-case-study-on-cantonese-mandarin-and-bavarian-german-qa - A novel QA dataset addressing information asymmetry in local language varieties for LLMs.
- High-Fidelity 3D Facial Avatar Synthesis with Controllable Fine-Grained Expressions (viability: 4): https://sciencetostartup.com/paper/high-fidelity-3d-facial-avatar-synthesis-with-controllable-fine-grained-expressions - A novel approach for high-fidelity 3D facial avatar synthesis with precise control over fine-grained expressions.
- Vietnamese Automatic Speech Recognition: A Revisit (viability: 7): https://sciencetostartup.com/paper/vietnamese-automatic-speech-recognition-a-revisit - A robust data aggregation and preprocessing pipeline for high-quality Vietnamese ASR datasets.
- $p^2$RAG: Privacy-Preserving RAG Service Supporting Arbitrary Top-$k$ Retrieval (viability: 3): https://sciencetostartup.com/paper/p-2-rag-privacy-preserving-rag-service-supporting-arbitrary-top-k-retrieval - A privacy-preserving RAG service that supports arbitrary top-k retrieval with enhanced security and efficiency.
- HO-SFL: Hybrid-Order Split Federated Learning with Backprop-Free Clients and Dimension-Free Aggregation (viability: 4): https://sciencetostartup.com/paper/ho-sfl-hybrid-order-split-federated-learning-with-backprop-free-clients-and-dimension-free-aggregation - HO-SFL offers a novel federated learning approach that reduces memory usage and communication costs while maintaining convergence speed.
- Zero-Shot Reconstruction of Animatable 3D Avatars with Cloth Dynamics from a Single Image (viability: 7): https://sciencetostartup.com/paper/zero-shot-reconstruction-of-animatable-3d-avatars-with-cloth-dynamics-from-a-single-image - DynaAvatar reconstructs animatable 3D human avatars with realistic cloth dynamics from a single image using a novel zero-shot framework.
- OpenHospital: A Thing-in-itself Arena for Evolving and Benchmarking LLM-based Collective Intelligence (viability: 5): https://sciencetostartup.com/paper/openhospital-a-thing-in-itself-arena-for-evolving-and-benchmarking-llm-based-collective-intelligence - OpenHospital is an interactive arena for evolving and benchmarking LLM-based collective intelligence in healthcare.
- AnyPhoto: Multi-Person Identity Preserving Image Generation with ID Adaptive Modulation on Location Canvas (viability: 3): https://sciencetostartup.com/paper/anyphoto-multi-person-identity-preserving-image-generation-with-id-adaptive-modulation-on-location-canvas - AnyPhoto enhances multi-person image generation by preserving identities through innovative modulation techniques.
- POLCA: Stochastic Generative Optimization with LLM (viability: 9): https://sciencetostartup.com/paper/polca-stochastic-generative-optimization-with-llm - POLCA is a scalable framework that optimizes complex systems using generative language models guided by feedback.
- Understanding the geometry of deep learning with decision boundary volume (viability: 3): https://sciencetostartup.com/paper/understanding-the-geometry-of-deep-learning-with-decision-boundary-volume - This paper introduces a geometric method to measure decision boundaries in neural networks, linking their structure to model performance.
- Investigating the Impact of Speech Enhancement on Audio Deepfake Detection in Noisy Environments (viability: 5): https://sciencetostartup.com/paper/investigating-the-impact-of-speech-enhancement-on-audio-deepfake-detection-in-noisy-environments - A study evaluating speech enhancement algorithms to improve audio deepfake detection in noisy environments.
- SSR: A Training-Free Approach for Streaming 3D Reconstruction (viability: 4): https://sciencetostartup.com/paper/ssr-a-training-free-approach-for-streaming-3d-reconstruction - A training-free operator for improved streaming 3D reconstruction that reduces drift and enhances quality.
- Topology-Preserving Data Augmentation for Ring-Type Polygon Annotations (viability: 4): https://sciencetostartup.com/paper/topology-preserving-data-augmentation-for-ring-type-polygon-annotations - A novel approach to geometric data augmentation that preserves topological consistency in ring-type polygon annotations.
- LiDAR-EVS: Enhance Extrapolated View Synthesis for 3D Gaussian Splatting with Pseudo-LiDAR Supervision (viability: 7): https://sciencetostartup.com/paper/lidar-evs-enhance-extrapolated-view-synthesis-for-3d-gaussian-splatting-with-pseudo-lidar-supervision - LiDAR-EVS is a lightweight framework for robust extrapolated-view LiDAR simulation in autonomous driving, enhancing real-time synthesis capabilities.
- BrainBench: Exposing the Commonsense Reasoning Gap in Large Language Models (viability: 4): https://sciencetostartup.com/paper/brainbench-exposing-the-commonsense-reasoning-gap-in-large-language-models - BrainBench is a benchmark tool designed to expose commonsense reasoning gaps in large language models.
- Towards Privacy-Preserving Machine Translation at the Inference Stage: A New Task and Benchmark (viability: 4): https://sciencetostartup.com/paper/towards-privacy-preserving-machine-translation-at-the-inference-stage-a-new-task-and-benchmark - A novel task and benchmark for privacy-preserving machine translation to protect sensitive information during inference.
- Learning Constituent Headedness (viability: 2): https://sciencetostartup.com/paper/learning-constituent-headedness - This paper proposes a method for learning constituent headedness in syntactic analysis, improving parsing accuracy across languages.
- Face-Guided Sentiment Boundary Enhancement for Weakly-Supervised Temporal Sentiment Localization (viability: 6): https://sciencetostartup.com/paper/face-guided-sentiment-boundary-enhancement-for-weakly-supervised-temporal-sentiment-localization - FSENet enhances sentiment localization in videos using facial features and weak supervision.
- CAMD: Coverage-Aware Multimodal Decoding for Efficient Reasoning of Multimodal Large Language Models (viability: 6): https://sciencetostartup.com/paper/camd-coverage-aware-multimodal-decoding-for-efficient-reasoning-of-multimodal-large-language-models - CAMD optimizes computation in multimodal large language models by dynamically allocating resources based on instance difficulty.
- PHAC: Promptable Human Amodal Completion (viability: 7): https://sciencetostartup.com/paper/phac-promptable-human-amodal-completion - PHAC enables precise human image completion using user-defined prompts for enhanced control and quality.
- TrajMamba: An Ego-Motion-Guided Mamba Model for Pedestrian Trajectory Prediction from an Egocentric Perspective (viability: 7): https://sciencetostartup.com/paper/trajmamba-an-ego-motion-guided-mamba-model-for-pedestrian-trajectory-prediction-from-an-egocentric-perspective - TrajMamba is an ego-motion-guided model for accurate pedestrian trajectory prediction in autonomous driving.
- Efficient Event Camera Volume System (viability: 7): https://sciencetostartup.com/paper/efficient-event-camera-volume-system - A novel framework for efficient event camera data compression and processing in robotic applications.
- Gauge-Equivariant Intrinsic Neural Operators for Geometry-Consistent Learning of Elliptic PDE Maps (viability: 2): https://sciencetostartup.com/paper/gauge-equivariant-intrinsic-neural-operators-for-geometry-consistent-learning-of-elliptic-pde-maps - GINO offers a new approach to learning solution operators for elliptic PDEs with improved geometric consistency.
- A Skill-augmented Agentic Framework and Benchmark for Multi-Video Understanding (viability: 7): https://sciencetostartup.com/paper/a-skill-augmented-agentic-framework-and-benchmark-for-multi-video-understanding - A framework that enhances multi-video understanding through structured reasoning and skill integration.
- GNNVerifier: Graph-based Verifier for LLM Task Planning (viability: 8): https://sciencetostartup.com/paper/gnnverifier-graph-based-verifier-for-llm-task-planning - GNNVerifier enhances task planning for LLMs by using a graph-based approach to identify and correct flaws in generated plans.
- DeFRiS: Silo-Cooperative IoT Applications Scheduling via Decentralized Federated Reinforcement Learning (viability: 6): https://sciencetostartup.com/paper/defris-silo-cooperative-iot-applications-scheduling-via-decentralized-federated-reinforcement-learning - DeFRiS is a decentralized federated reinforcement learning framework for efficient and robust scheduling in silo-cooperative IoT applications.
- Automated Diabetic Screening via Anterior Segment Ocular Imaging: A Deep Learning and Explainable AI Approach (viability: 3): https://sciencetostartup.com/paper/automated-diabetic-screening-via-anterior-segment-ocular-imaging-a-deep-learning-and-explainable-ai-approach - A deep learning system for automated diabetic classification using accessible anterior segment ocular imaging.
- Enhancing Hands in 3D Whole-Body Pose Estimation with Conditional Hands Modulator (viability: 7): https://sciencetostartup.com/paper/enhancing-hands-in-3d-whole-body-pose-estimation-with-conditional-hands-modulator - Hand4Whole++ enhances 3D whole-body pose estimation by integrating hand-specific features for improved accuracy.
- GameUIAgent: An LLM-Powered Framework for Automated Game UI Design with Structured Intermediate Representation (viability: 2): https://sciencetostartup.com/paper/gameuiagent-an-llm-powered-framework-for-automated-game-ui-design-with-structured-intermediate-representation - GameUIAgent automates game UI design by translating natural language into editable Figma designs using a neuro-symbolic pipeline.
- Beyond Creed: A Non-Identity Safety Condition A Strong Empirical Alternative to Identity Framing in Low-Data LoRA Fine-Tuning (viability: 2): https://sciencetostartup.com/paper/beyond-creed-a-non-identity-safety-condition-a-strong-empirical-alternative-to-identity-framing-in-low-data-lora-fine-tu - This paper explores alternative safety supervision formats for low-data LoRA fine-tuning in AI models.
- Multimodal Deep Learning for Early Prediction of Patient Deterioration in the ICU: Integrating Time-Series EHR Data with Clinical Notes (viability: 3): https://sciencetostartup.com/paper/multimodal-deep-learning-for-early-prediction-of-patient-deterioration-in-the-icu-integrating-time-series-ehr-data-with- - A multimodal deep learning framework for predicting patient deterioration in the ICU using EHR data and clinical notes.
- Training-Free Generation of Protein Sequences from Small Family Alignments via Stochastic Attention (viability: 3): https://sciencetostartup.com/paper/training-free-generation-of-protein-sequences-from-small-family-alignments-via-stochastic-attention - A training-free method for generating protein sequences from small family alignments using stochastic attention.
- Towards Next-Generation LLM Training: From the Data-Centric Perspective (viability: 4): https://sciencetostartup.com/paper/towards-next-generation-llm-training-from-the-data-centric-perspective - Develop an agent-based system for automated data preparation and management in LLM training.
- Cross-RAG: Zero-Shot Retrieval-Augmented Time Series Forecasting via Cross-Attention (viability: 8): https://sciencetostartup.com/paper/cross-rag-zero-shot-retrieval-augmented-time-series-forecasting-via-cross-attention - Cross-RAG enhances zero-shot time series forecasting by leveraging selective query-relevant retrieval.
- Visual Confused Deputy: Exploiting and Defending Perception Failures in Computer-Using Agents (viability: 8): https://sciencetostartup.com/paper/visual-confused-deputy-exploiting-and-defending-perception-failures-in-computer-using-agents - A dual-channel guardrail system that enhances the safety of computer-using agents by independently verifying their actions against visual and textual reasoning.
- AdapterTune: Zero-Initialized Low-Rank Adapters for Frozen Vision Transformers (viability: 9): https://sciencetostartup.com/paper/adaptertune-zero-initialized-low-rank-adapters-for-frozen-vision-transformers - AdapterTune optimizes Vision Transformers by introducing zero-initialized low-rank adapters, significantly improving transfer accuracy with fewer parameters.
- Chain-of-Trajectories: Unlocking the Intrinsic Generative Optimality of Diffusion Models via Graph-Theoretic Planning (viability: 7): https://sciencetostartup.com/paper/chain-of-trajectories-unlocking-the-intrinsic-generative-optimality-of-diffusion-models-via-graph-theoretic-planning - Chain-of-Trajectories enhances diffusion models by enabling resource-aware planning for improved generative output.
- Beyond Local Code Optimization: Multi-Agent Reasoning for Software System Optimization (viability: 6): https://sciencetostartup.com/paper/beyond-local-code-optimization-multi-agent-reasoning-for-software-system-optimization - A multi-agent framework for optimizing software systems by reasoning about program structure and performance interactions.
- Fractal Autoregressive Depth Estimation with Continuous Token Diffusion (viability: 4): https://sciencetostartup.com/paper/fractal-autoregressive-depth-estimation-with-continuous-token-diffusion - A novel framework for monocular depth estimation using autoregressive diffusion techniques.
- AURORA-KITTI: Any-Weather Depth Completion and Denoising in the Wild (viability: 7): https://sciencetostartup.com/paper/aurora-kitti-any-weather-depth-completion-and-denoising-in-the-wild - AURORA-KITTI provides a robust solution for depth completion and denoising across diverse weather conditions.
- A Dual Quaternion Framework for Collision Recovery of Quadrotor (viability: 5): https://sciencetostartup.com/paper/a-dual-quaternion-framework-for-collision-recovery-of-quadrotor - A dual quaternion framework enhances collision recovery for quadrotors in cluttered environments.
- Robust Building Damage Detection in Cross-Disaster Settings Using Domain Adaptation (viability: 6): https://sciencetostartup.com/paper/robust-building-damage-detection-in-cross-disaster-settings-using-domain-adaptation - Automated damage detection for disaster response using domain adaptation techniques.
- Applications of Intuitionistic Temporal Logic to Temporal Answer Set Programming (viability: 2): https://sciencetostartup.com/paper/applications-of-intuitionistic-temporal-logic-to-temporal-answer-set-programming - This paper explores the theoretical foundations of Temporal Answer Set Programming using Temporal Equilibrium Logic.
- AgentTrace: Causal Graph Tracing for Root Cause Analysis in Deployed Multi-Agent Systems (viability: 7): https://sciencetostartup.com/paper/agenttrace-causal-graph-tracing-for-root-cause-analysis-in-deployed-multi-agent-systems - AgentTrace offers a lightweight causal tracing framework for diagnosing failures in multi-agent AI systems.
- MVHOI: Bridge Multi-view Condition to Complex Human-Object Interaction Video Reenactment via 3D Foundation Model (viability: 2): https://sciencetostartup.com/paper/mvhoi-bridge-multi-view-condition-to-complex-human-object-interaction-video-reenactment-via-3d-foundation-model - MVHOI enhances human-object interaction video reenactment through a novel 3D foundation model.
- E2EGS: Event-to-Edge Gaussian Splatting for Pose-Free 3D Reconstruction (viability: 7): https://sciencetostartup.com/paper/e2egs-event-to-edge-gaussian-splatting-for-pose-free-3d-reconstruction - E2EGS is a pose-free framework for 3D reconstruction using event camera data, enhancing robustness in dynamic scenes.
- \texttt{BayesBreak}: Generalized Hierarchical Bayesian Segmentation with Irregular Designs, Multi-Sample Hierarchies, and Grouped/Latent-Group Designs (viability: 2): https://sciencetostartup.com/paper/texttt-bayesbreak-generalized-hierarchical-bayesian-segmentation-with-irregular-designs-multi-sample-hierarchies-and-gro - BayesBreak offers a modular Bayesian segmentation framework for piecewise-constant representations of ordered data.
- Scalable Text-Embedding-informed Cognitive Diagnosis of Large Language Models (viability: 4): https://sciencetostartup.com/paper/scalable-text-embedding-informed-cognitive-diagnosis-of-large-language-models - A novel methodology for fine-grained cognitive diagnosis of large language models using scalable text-embedding-informed techniques.
- Computational Analysis of Semantic Connections Between Herman Melville Reading and Writing (viability: 4): https://sciencetostartup.com/paper/computational-analysis-of-semantic-connections-between-herman-melville-reading-and-writing - A computational framework analyzing the influence of Herman Melville's reading on his writings through semantic similarity.
- A Single-Sample Polylogarithmic Regret Bound for Nonstationary Online Linear Programming (viability: 3): https://sciencetostartup.com/paper/a-single-sample-polylogarithmic-regret-bound-for-nonstationary-online-linear-programming - A novel algorithm for nonstationary online linear programming that achieves polylogarithmic regret with minimal data.
- Seamless Deception: Larger Language Models Are Better Knowledge Concealers (viability: 2): https://sciencetostartup.com/paper/seamless-deception-larger-language-models-are-better-knowledge-concealers - This research identifies limitations in auditing language models that conceal harmful knowledge.
- RenderMem: Rendering as Spatial Memory Retrieval (viability: 3): https://sciencetostartup.com/paper/rendermem-rendering-as-spatial-memory-retrieval - RenderMem enhances spatial reasoning in embodied agents by integrating rendering with 3D scene representations.
- Comparative Analysis of 3D Convolutional and 2.5D Slice-Conditioned U-Net Architectures for MRI Super-Resolution via Elucidated Diffusion Models (viability: 7): https://sciencetostartup.com/paper/comparative-analysis-of-3d-convolutional-and-2-5d-slice-conditioned-u-net-architectures-for-mri-super-resolution-via-elu - A novel MRI super-resolution method using elucidated diffusion models to enhance low-resolution scans.
- EviATTA: Evidential Active Test-Time Adaptation for Medical Segment Anything Models (viability: 7): https://sciencetostartup.com/paper/eviatta-evidential-active-test-time-adaptation-for-medical-segment-anything-models - EviATTA enhances medical image segmentation by improving test-time adaptation with minimal expert feedback.
- Gradient Atoms: Unsupervised Discovery, Attribution and Steering of Model Behaviors via Sparse Decomposition of Training Gradients (viability: 8): https://sciencetostartup.com/paper/gradient-atoms-unsupervised-discovery-attribution-and-steering-of-model-behaviors-via-sparse-decomposition-of-training-g - Gradient Atoms offers an unsupervised method for discovering and steering model behaviors through sparse decomposition of training gradients.
- Punctuated Equilibria in Artificial Intelligence: The Institutional Scaling Law and the Speciation of Sovereign AI (viability: 2): https://sciencetostartup.com/paper/punctuated-equilibria-in-artificial-intelligence-the-institutional-scaling-law-and-the-speciation-of-sovereign-ai - This paper proposes a theoretical framework challenging conventional AI scaling assumptions based on evolutionary biology.
- VisionCoach: Reinforcing Grounded Video Reasoning via Visual-Perception Prompting (viability: 7): https://sciencetostartup.com/paper/visioncoach-reinforcing-grounded-video-reasoning-via-visual-perception-prompting - VisionCoach enhances video reasoning by using visual prompting to improve spatio-temporal grounding during training.
- Human-AI Ensembles Improve Deepfake Detection in Low-to-Medium Quality Videos (viability: 4): https://sciencetostartup.com/paper/human-ai-ensembles-improve-deepfake-detection-in-low-to-medium-quality-videos - A hybrid human-AI approach enhances deepfake detection accuracy in low-to-medium quality videos.
- Coordinate-Independent Robot Model Identification (viability: 2): https://sciencetostartup.com/paper/coordinate-independent-robot-model-identification - A novel coordinate-independent method for robot model identification that improves accuracy by eliminating coordinate-induced bias.
- EARCP: Self-Regulating Coherence-Aware Ensemble Architecture for Sequential Decision Making -- Ensemble Auto-Regule par Coherence et Performance (viability: 8): https://sciencetostartup.com/paper/earcp-self-regulating-coherence-aware-ensemble-architecture-for-sequential-decision-making-ensemble-auto-regule-par-cohe - EARCP is a self-regulating ensemble architecture that adapts model weights dynamically for improved sequential decision-making.
- A Methodology for Thermal Limit Bias Predictability Through Artificial Intelligence (viability: 9): https://sciencetostartup.com/paper/a-methodology-for-thermal-limit-bias-predictability-through-artificial-intelligence - A deep learning methodology that predicts and corrects thermal limit bias in nuclear power plants to enhance operational efficiency and reduce costs.
- TopoCL: Topological Contrastive Learning for Medical Imaging (viability: 6): https://sciencetostartup.com/paper/topocl-topological-contrastive-learning-for-medical-imaging - TopoCL enhances medical image analysis by integrating topological features into contrastive learning.
- Dynamic Theory of Mind as a Temporal Memory Problem: Evidence from Large Language Models (viability: 2): https://sciencetostartup.com/paper/dynamic-theory-of-mind-as-a-temporal-memory-problem-evidence-from-large-language-models - This research explores the dynamic aspects of Theory of Mind in LLMs, highlighting challenges in tracking belief changes over time.
- Spectrum Matching: a Unified Perspective for Superior Diffusability in Latent Diffusion (viability: 7): https://sciencetostartup.com/paper/spectrum-matching-a-unified-perspective-for-superior-diffusability-in-latent-diffusion - Spectrum Matching enhances latent diffusion learnability by aligning power spectral densities for superior image generation.
- Argumentation for Explainable and Globally Contestable Decision Support with LLMs (viability: 3): https://sciencetostartup.com/paper/argumentation-for-explainable-and-globally-contestable-decision-support-with-llms - ArgEval enhances decision support in high-stakes domains by providing explainable and contestable recommendations using LLMs.
- Seeing Where to Deploy: Metric RGB-Based Traversability Analysis for Aerial-to-Ground Hidden Space Inspection (viability: 4): https://sciencetostartup.com/paper/seeing-where-to-deploy-metric-rgb-based-traversability-analysis-for-aerial-to-ground-hidden-space-inspection - A framework for aerial-to-ground hidden space inspection using RGB-based traversability analysis.
- Nudging Hidden States: Training-Free Model Steering for Chain-of-Thought Reasoning in Large Audio-Language Models (viability: 6): https://sciencetostartup.com/paper/nudging-hidden-states-training-free-model-steering-for-chain-of-thought-reasoning-in-large-audio-language-models - A training-free model steering approach to enhance reasoning in large audio-language models.
- Compute Allocation for Reasoning-Intensive Retrieval Agents (viability: 4): https://sciencetostartup.com/paper/compute-allocation-for-reasoning-intensive-retrieval-agents - A study on optimizing computation allocation in reasoning-intensive retrieval for LLM-augmented pipelines.
- Physically Accurate Rigid-Body Dynamics in Particle-Based Simulation (viability: 5): https://sciencetostartup.com/paper/physically-accurate-rigid-body-dynamics-in-particle-based-simulation - PBD-R enhances particle-based simulation for robotics by ensuring physically accurate rigid-body dynamics with improved computational efficiency.
- When Scanners Lie: Evaluator Instability in LLM Red-Teaming (viability: 7): https://sciencetostartup.com/paper/when-scanners-lie-evaluator-instability-in-llm-red-teaming - A framework to enhance the reliability of automated LLM vulnerability scanners by addressing evaluator instability.
- Continual Few-shot Adaptation for Synthetic Fingerprint Detection (viability: 6): https://sciencetostartup.com/paper/continual-few-shot-adaptation-for-synthetic-fingerprint-detection - A continual few-shot adaptation method for detecting synthetic fingerprints to enhance security in biometric systems.
- Anterior's Approach to Fairness Evaluation of Automated Prior Authorization System (viability: 4): https://sciencetostartup.com/paper/anterior-s-approach-to-fairness-evaluation-of-automated-prior-authorization-system - A framework for evaluating fairness in automated prior authorization systems in healthcare.
- ResearchPilot: A Local-First Multi-Agent System for Literature Synthesis and Related Work Drafting (viability: 5): https://sciencetostartup.com/paper/researchpilot-a-local-first-multi-agent-system-for-literature-synthesis-and-related-work-drafting - ResearchPilot is a self-hostable multi-agent system that assists in literature reviews by synthesizing findings and drafting related work sections.
- s2n-bignum-bench: A practical benchmark for evaluating low-level code reasoning of LLMs (viability: 7): https://sciencetostartup.com/paper/s2n-bignum-bench-a-practical-benchmark-for-evaluating-low-level-code-reasoning-of-llms - s2n-bignum-bench is a benchmark for evaluating LLMs in generating machine-checkable proofs for cryptographic assembly routines.
- EcoFair-CH-MARL: Scalable Constrained Hierarchical Multi-Agent RL with Real-Time Emission Budgets and Fairness Guarantees (viability: 3): https://sciencetostartup.com/paper/ecofair-ch-marl-scalable-constrained-hierarchical-multi-agent-rl-with-real-time-emission-budgets-and-fairness-guarantees - EcoFair-CH-MARL is a framework for efficient and equitable maritime logistics using multi-agent reinforcement learning.
- Proactive Routing to Interpretable Surrogates with Distribution-Free Safety Guarantees (viability: 6): https://sciencetostartup.com/paper/proactive-routing-to-interpretable-surrogates-with-distribution-free-safety-guarantees - A model routing system that ensures safe and interpretable surrogate use with controlled degradation.
- A Heterogeneous Ensemble for Multi-Center COVID-19 Classification from Chest CT Scans (viability: 4): https://sciencetostartup.com/paper/a-heterogeneous-ensemble-for-multi-center-covid-19-classification-from-chest-ct-scans - A heterogeneous ensemble model for improved COVID-19 classification from chest CT scans across multiple centers.
- LLM-Augmented Release Intelligence: Automated Change Summarization and Impact Analysis in Cloud-Native CI/CD Pipelines (viability: 8): https://sciencetostartup.com/paper/llm-augmented-release-intelligence-automated-change-summarization-and-impact-analysis-in-cloud-native-ci-cd-pipelines - An AI-driven framework that automates change summarization and impact analysis for cloud-native CI/CD pipelines.
- Make it SING: Analyzing Semantic Invariants in Classifiers (viability: 2): https://sciencetostartup.com/paper/make-it-sing-analyzing-semantic-invariants-in-classifiers - SING provides a method for interpreting semantic invariants in classifiers through null-space geometry analysis.
- GroundSet: A Cadastral-Grounded Dataset for Spatial Understanding with Vector Data (viability: 4): https://sciencetostartup.com/paper/groundset-a-cadastral-grounded-dataset-for-spatial-understanding-with-vector-data - GroundSet is a large-scale dataset designed to enhance spatial understanding in Earth Observation through fine-grained annotations.
- Delightful Policy Gradient (viability: 3): https://sciencetostartup.com/paper/delightful-policy-gradient - Delightful Policy Gradient improves policy gradient methods by addressing action weighting issues in reinforcement learning.
- CyboRacket: A Perception-to-Action Framework for Humanoid Racket Sports (viability: 7): https://sciencetostartup.com/paper/cyboracket-a-perception-to-action-framework-for-humanoid-racket-sports - CyboRacket is a perception-to-action framework enabling humanoid robots to excel in racket sports through advanced visual tracking and trajectory prediction.
- Tactile Modality Fusion for Vision-Language-Action Models (viability: 7): https://sciencetostartup.com/paper/tactile-modality-fusion-for-vision-language-action-models - TacFiLM enhances vision-language-action models by integrating tactile signals for improved manipulation tasks.
- Latent Dynamics-Aware OOD Monitoring for Trajectory Prediction with Provable Guarantees (viability: 4): https://sciencetostartup.com/paper/latent-dynamics-aware-ood-monitoring-for-trajectory-prediction-with-provable-guarantees - A framework for reliable out-of-distribution monitoring in trajectory prediction for safety-critical systems.
- $PA^3$: $\textbf{P}$olicy-$\textbf{A}$ware $\textbf{A}$gent $\textbf{A}$lignment through Chain-of-Thought (viability: 3): https://sciencetostartup.com/paper/pa-3-textbf-p-olicy-textbf-a-ware-textbf-a-gent-textbf-a-lignment-through-chain-of-thought - A novel method for aligning LLMs with business-specific rules to enhance tool-use tasks.
- A Loss Landscape Visualization Framework for Interpreting Reinforcement Learning: An ADHDP Case Study (viability: 2): https://sciencetostartup.com/paper/a-loss-landscape-visualization-framework-for-interpreting-reinforcement-learning-an-adhdp-case-study - A framework for visualizing loss landscapes in reinforcement learning to enhance interpretability.
- SmallSatSim: A High-Fidelity Simulation and Training Toolkit for Microgravity Robotic Close Proximity Operations (viability: 7): https://sciencetostartup.com/paper/smallsatsim-a-high-fidelity-simulation-and-training-toolkit-for-microgravity-robotic-close-proximity-operations - SmallSatSim is a high-fidelity simulation toolkit for robotic operations in microgravity environments.
- Scaling the Explanation of Multi-Class Bayesian Network Classifiers (viability: 2): https://sciencetostartup.com/paper/scaling-the-explanation-of-multi-class-bayesian-network-classifiers - A new algorithm for compiling multi-class Bayesian network classifiers into logical formulas for improved explainability.
- Parameter-Efficient Quality Estimation via Frozen Recursive Models (viability: 8): https://sciencetostartup.com/paper/parameter-efficient-quality-estimation-via-frozen-recursive-models - Parameter-efficient quality estimation for low-resource languages using frozen recursive models.
- A Multi-Scale Graph Learning Framework with Temporal Consistency Constraints for Financial Fraud Detection in Transaction Networks under Non-Stationary Conditions (viability: 7): https://sciencetostartup.com/paper/a-multi-scale-graph-learning-framework-with-temporal-consistency-constraints-for-financial-fraud-detection-in-transactio - A graph-based framework for detecting financial fraud in transaction networks using temporal consistency constraints.
- FlashHead: Efficient Drop-In Replacement for the Classification Head in Language Model Inference (viability: 7): https://sciencetostartup.com/paper/flashhead-efficient-drop-in-replacement-for-the-classification-head-in-language-model-inference - FlashHead is an efficient drop-in replacement for the classification head in language models, enhancing inference speed while maintaining accuracy.
- Adapting Critic Match Loss Landscape Visualization to Off-policy Reinforcement Learning (viability: 3): https://sciencetostartup.com/paper/adapting-critic-match-loss-landscape-visualization-to-off-policy-reinforcement-learning - A geometric diagnostic tool for analyzing critic optimization dynamics in off-policy reinforcement learning.
- SuperLocalMemory V3: Information-Geometric Foundations for Zero-LLM Enterprise Agent Memory (viability: 7): https://sciencetostartup.com/paper/superlocalmemory-v3-information-geometric-foundations-for-zero-llm-enterprise-agent-memory - A novel mathematical framework for enhancing AI agent memory retrieval and management.
- Texel Splatting: Perspective-Stable 3D Pixel Art (viability: 2): https://sciencetostartup.com/paper/texel-splatting-perspective-stable-3d-pixel-art - Texel splatting offers a novel approach to rendering 3D scenes as stable pixel art.
- The Scenic Route to Deception: Dark Patterns and Explainability Pitfalls in Conversational Navigation (viability: 2): https://sciencetostartup.com/paper/the-scenic-route-to-deception-dark-patterns-and-explainability-pitfalls-in-conversational-navigation - Exploring the risks of manipulation in conversational navigation using Generative AI.
- Machine Learning-Driven Intelligent Memory System Design: From On-Chip Caches to Storage (viability: 3): https://sciencetostartup.com/paper/machine-learning-driven-intelligent-memory-system-design-from-on-chip-caches-to-storage - A machine learning-driven approach to optimize memory systems for improved performance and efficiency.
- Medical Image Spatial Grounding with Semantic Sampling (viability: 7): https://sciencetostartup.com/paper/medical-image-spatial-grounding-with-semantic-sampling - A benchmark and optimization tool for enhancing spatial grounding in medical image analysis using vision language models.
- Power-Law Spectrum of the Random Feature Model (viability: 2): https://sciencetostartup.com/paper/power-law-spectrum-of-the-random-feature-model - This paper explores the spectral structure of data covariance in neural networks through the random feature model.
- Covariance-Guided Resource Adaptive Learning for Efficient Edge Inference (viability: 3): https://sciencetostartup.com/paper/covariance-guided-resource-adaptive-learning-for-efficient-edge-inference - CORAL optimizes deep learning inference configurations on edge devices without the need for exhaustive profiling.
- CausalEvolve: Towards Open-Ended Discovery with Causal Scratchpad (viability: 3): https://sciencetostartup.com/paper/causalevolve-towards-open-ended-discovery-with-causal-scratchpad - CausalEvolve enhances evolve-based agents by introducing a causal scratchpad for improved evolutionary efficiency in scientific problem-solving.
- Top-b: Entropic Regulation of Relative Probability Bands in Autoregressive Language Processes (viability: 7): https://sciencetostartup.com/paper/top-b-entropic-regulation-of-relative-probability-bands-in-autoregressive-language-processes - Top-b is a novel decoding strategy that optimizes language generation by dynamically regulating candidate sets based on entropy.
- Multilingual TinyStories: A Synthetic Combinatorial Corpus of Indic Children's Stories for Training Small Language Models (viability: 6): https://sciencetostartup.com/paper/multilingual-tinystories-a-synthetic-combinatorial-corpus-of-indic-children-s-stories-for-training-small-language-models - A synthetic dataset of children's stories in 17 Indian languages to train small language models.
- A comprehensive multimodal dataset and benchmark for ulcerative colitis scoring in endoscopy (viability: 6): https://sciencetostartup.com/paper/a-comprehensive-multimodal-dataset-and-benchmark-for-ulcerative-colitis-scoring-in-endoscopy - A multimodal dataset for ulcerative colitis scoring that enables the development of automated assessment algorithms.
- JobMatchAI An Intelligent Job Matching Platform Using Knowledge Graphs, Semantic Search and Explainable AI (viability: 9): https://sciencetostartup.com/paper/jobmatchai-an-intelligent-job-matching-platform-using-knowledge-graphs-semantic-search-and-explainable-ai - JobMatchAI is an intelligent job matching platform that leverages knowledge graphs and semantic search to optimize hiring outcomes.
- MorFiC: Fixing Value Miscalibration for Zero-Shot Quadruped Transfer (viability: 7): https://sciencetostartup.com/paper/morfic-fixing-value-miscalibration-for-zero-shot-quadruped-transfer - MorFiC enables zero-shot locomotion policy transfer across different quadrupedal robots using a single shared policy.
- Learning to Order: Task Sequencing as In-Context Optimization (viability: 4): https://sciencetostartup.com/paper/learning-to-order-task-sequencing-as-in-context-optimization - A meta-learning approach to optimize task sequencing for various applications, demonstrating few-shot generalization.
- ASAP: Attention-Shift-Aware Pruning for Efficient LVLM Inference (viability: 7): https://sciencetostartup.com/paper/asap-attention-shift-aware-pruning-for-efficient-lvlm-inference - ASAP is a novel pruning method that enhances the efficiency of Large Vision-Language Models by addressing attention shifts and reducing token redundancy.
- Gradient Boosting for Spatial Panel Models with Random and Fixed Effects (viability: 2): https://sciencetostartup.com/paper/gradient-boosting-for-spatial-panel-models-with-random-and-fixed-effects - A model-based gradient boosting algorithm for spatial panel data analysis.
- Expert Mind: A Retrieval-Augmented Architecture for Expert Knowledge Preservation in the Energy Sector (viability: 3): https://sciencetostartup.com/paper/expert-mind-a-retrieval-augmented-architecture-for-expert-knowledge-preservation-in-the-energy-sector - Expert Mind is a system designed to preserve and query expert knowledge in the energy sector using Retrieval-Augmented Generation.
- Distilling Latent Manifolds: Resolution Extrapolation by Variational Autoencoders (viability: 7): https://sciencetostartup.com/paper/distilling-latent-manifolds-resolution-extrapolation-by-variational-autoencoders - A novel approach to VAE encoder distillation that enhances high-resolution image reconstruction from low-resolution training data.
- Visualizing Critic Match Loss Landscapes for Interpretation of Online Reinforcement Learning Control Algorithms (viability: 4): https://sciencetostartup.com/paper/visualizing-critic-match-loss-landscapes-for-interpretation-of-online-reinforcement-learning-control-algorithms - A visualization method for analyzing critic match loss landscapes in online reinforcement learning algorithms.
- PhysMoDPO: Physically-Plausible Humanoid Motion with Preference Optimization (viability: 7): https://sciencetostartup.com/paper/physmodpo-physically-plausible-humanoid-motion-with-preference-optimization - PhysMoDPO optimizes humanoid motion generation for realistic and task-compliant robot control using a novel preference optimization framework.
- Representation Learning for Spatiotemporal Physical Systems (viability: 7): https://sciencetostartup.com/paper/representation-learning-for-spatiotemporal-physical-systems - This research presents a novel approach to learning physics-grounded representations for spatiotemporal physical systems, enhancing the accuracy of scientific parameter estimation.
- Visual-ERM: Reward Modeling for Visual Equivalence (viability: 7): https://sciencetostartup.com/paper/visual-erm-reward-modeling-for-visual-equivalence - Visual-ERM enhances vision-to-code tasks by providing fine-grained reward modeling for improved visual fidelity.
- Out of Sight, Out of Mind? Evaluating State Evolution in Video World Models (viability: 4): https://sciencetostartup.com/paper/out-of-sight-out-of-mind-evaluating-state-evolution-in-video-world-models - STEVO-Bench evaluates the limitations of video world models in decoupling state evolution from observation.
- Neuron-Aware Data Selection In Instruction Tuning For Large Language Models (viability: 7): https://sciencetostartup.com/paper/neuron-aware-data-selection-in-instruction-tuning-for-large-language-models - NAIT optimizes instruction tuning data selection for large language models to enhance performance through neuron activation pattern analysis.
- LLM Constitutional Multi-Agent Governance (viability: 4): https://sciencetostartup.com/paper/llm-constitutional-multi-agent-governance - A framework for ethical governance in multi-agent systems using LLMs to ensure cooperation without manipulation.
- Learnability and Privacy Vulnerability are Entangled in a Few Critical Weights (viability: 2): https://sciencetostartup.com/paper/learnability-and-privacy-vulnerability-are-entangled-in-a-few-critical-weights - This research explores a novel method for preserving privacy in neural networks by targeting critical weights for fine-tuning.
- Towards Spatio-Temporal World Scene Graph Generation from Monocular Videos (viability: 4): https://sciencetostartup.com/paper/towards-spatio-temporal-world-scene-graph-generation-from-monocular-videos - A novel approach to generating spatio-temporal scene graphs from monocular videos, enhancing object interaction modeling.
- Diffusion-Based Feature Denoising and Using NNMF for Robust Brain Tumor Classification (viability: 3): https://sciencetostartup.com/paper/diffusion-based-feature-denoising-and-using-nnmf-for-robust-brain-tumor-classification - A robust framework for brain tumor classification using NNMF and lightweight CNNs to enhance adversarial robustness.
- Verification of Robust Properties for Access Control Policies (viability: 2): https://sciencetostartup.com/paper/verification-of-robust-properties-for-access-control-policies - This paper presents a theoretical framework for verifying access control policies that evolve over time.
- MXNorm: Reusing MXFP block scales for efficient tensor normalisation (viability: 3): https://sciencetostartup.com/paper/mxnorm-reusing-mxfp-block-scales-for-efficient-tensor-normalisation - MXNorm offers a novel method for efficient tensor normalization, enhancing performance in deep learning workloads.
- Perceive What Matters: Relevance-Driven Scheduling for Multimodal Streaming Perception (viability: 7): https://sciencetostartup.com/paper/perceive-what-matters-relevance-driven-scheduling-for-multimodal-streaming-perception - A lightweight perception scheduling framework that optimizes multimodal streaming for human-robot collaboration by reducing latency and enhancing efficiency.
- Semantic Invariance in Agentic AI (viability: 5): https://sciencetostartup.com/paper/semantic-invariance-in-agentic-ai - A metamorphic testing framework to assess the semantic invariance of LLM reasoning agents across various models.
- Developing and evaluating a chatbot to support maternal health care (viability: 7): https://sciencetostartup.com/paper/developing-and-evaluating-a-chatbot-to-support-maternal-health-care - A chatbot designed to provide trustworthy maternal health information in low-resource settings.
- Towards Faithful Multimodal Concept Bottleneck Models (viability: 4): https://sciencetostartup.com/paper/towards-faithful-multimodal-concept-bottleneck-models - f-CBM is a multimodal framework that enhances interpretable predictions by jointly addressing concept detection and leakage mitigation.
- When Your Model Stops Working: Anytime-Valid Calibration Monitoring (viability: 4): https://sciencetostartup.com/paper/when-your-model-stops-working-anytime-valid-calibration-monitoring - PITMonitor offers a robust solution for monitoring distributional shifts in probabilistic models with guaranteed error control.
- ESG-Bench: Benchmarking Long-Context ESG Reports for Hallucination Mitigation (viability: 7): https://sciencetostartup.com/paper/esg-bench-benchmarking-long-context-esg-reports-for-hallucination-mitigation - ESG-Bench provides a benchmark dataset for improving the reliability of ESG report analysis using LLMs.
- Defensible Design for OpenClaw: Securing Autonomous Tool-Invoking Agents (viability: 2): https://sciencetostartup.com/paper/defensible-design-for-openclaw-securing-autonomous-tool-invoking-agents - OpenClaw proposes a blueprint for securing autonomous tool-invoking agents against architectural vulnerabilities.
- When Right Meets Wrong: Bilateral Context Conditioning with Reward-Confidence Correction for GRPO (viability: 8): https://sciencetostartup.com/paper/when-right-meets-wrong-bilateral-context-conditioning-with-reward-confidence-correction-for-grpo - A novel approach to optimize reasoning models by leveraging contrastive learning within group samples.
- DecoVLN: Decoupling Observation, Reasoning, and Correction for Vision-and-Language Navigation (viability: 7): https://sciencetostartup.com/paper/decovln-decoupling-observation-reasoning-and-correction-for-vision-and-language-navigation - DecoVLN enhances Vision-and-Language Navigation by optimizing long-term memory and correcting errors in real-time.
- Steve-Evolving: Open-World Embodied Self-Evolution via Fine-Grained Diagnosis and Dual-Track Knowledge Distillation (viability: 7): https://sciencetostartup.com/paper/steve-evolving-open-world-embodied-self-evolution-via-fine-grained-diagnosis-and-dual-track-knowledge-distillation - Steve-Evolving is a self-evolving framework for open-world agents that enhances long-horizon task performance through structured experience organization and dual-track knowledge distillation.
- FDeID-Toolbox: Face De-Identification Toolbox (viability: 6): https://sciencetostartup.com/paper/fdeid-toolbox-face-de-identification-toolbox - FDeID-Toolbox is a comprehensive toolkit for reproducible face de-identification research, enhancing privacy in computer vision applications.
- Geometry-Guided Camera Motion Understanding in VideoLLMs (viability: 7): https://sciencetostartup.com/paper/geometry-guided-camera-motion-understanding-in-videollms - A framework for enhancing video language models with explicit camera motion understanding using a new dataset and lightweight integration methods.
- NOIR: Neural Operator mapping for Implicit Representations (viability: 8): https://sciencetostartup.com/paper/noir-neural-operator-mapping-for-implicit-representations - NOIR revolutionizes medical imaging by using Neural Operators for resolution-independent transformations.
- ZO-SAM: Zero-Order Sharpness-Aware Minimization for Efficient Sparse Training (viability: 3): https://sciencetostartup.com/paper/zo-sam-zero-order-sharpness-aware-minimization-for-efficient-sparse-training - ZO-SAM is a novel optimization framework that enhances sparse training efficiency by reducing computational costs and improving convergence.
- BoSS: A Best-of-Strategies Selector as an Oracle for Deep Active Learning (viability: 6): https://sciencetostartup.com/paper/boss-a-best-of-strategies-selector-as-an-oracle-for-deep-active-learning - BoSS is a scalable oracle strategy that enhances active learning by selecting the most valuable instances for annotation.
- Panoramic Multimodal Semantic Occupancy Prediction for Quadruped Robots (viability: 8): https://sciencetostartup.com/paper/panoramic-multimodal-semantic-occupancy-prediction-for-quadruped-robots - Develop VoxelHound, a panoramic multimodal perception framework for quadruped robots, using the new PanoMMOcc dataset.
- A Feasibility-Enhanced Control Barrier Function Method for Multi-UAV Collision Avoidance (viability: 7): https://sciencetostartup.com/paper/a-feasibility-enhanced-control-barrier-function-method-for-multi-uav-collision-avoidance - A framework for enhancing collision avoidance in multi-UAV systems using feasibility-enhanced control barrier functions.
- BenDFM: A taxonomy and synthetic CAD dataset for manufacturability assessment in sheet metal bending (viability: 6): https://sciencetostartup.com/paper/bendfm-a-taxonomy-and-synthetic-cad-dataset-for-manufacturability-assessment-in-sheet-metal-bending - BenDFM is a synthetic dataset and taxonomy for assessing manufacturability in sheet metal bending, enabling better design for manufacturing decisions.
- Evaluating VLMs' Spatial Reasoning Over Robot Motion: A Step Towards Robot Planning with Motion Preferences (viability: 3): https://sciencetostartup.com/paper/evaluating-vlms-spatial-reasoning-over-robot-motion-a-step-towards-robot-planning-with-motion-preferences - Evaluating VLMs for enhancing spatial reasoning in robot motion planning.
- Beyond Final Answers: CRYSTAL Benchmark for Transparent Multimodal Reasoning Evaluation (viability: 4): https://sciencetostartup.com/paper/beyond-final-answers-crystal-benchmark-for-transparent-multimodal-reasoning-evaluation - CRYSTAL is a benchmark for evaluating multimodal reasoning through verifiable intermediate steps.
- SldprtNet: A Large-Scale Multimodal Dataset for CAD Generation in Language-Driven 3D Design (viability: 6): https://sciencetostartup.com/paper/sldprtnet-a-large-scale-multimodal-dataset-for-cad-generation-in-language-driven-3d-design - SldprtNet is a comprehensive multimodal dataset for CAD generation, enhancing 3D design through semantic-driven modeling.
- Breaking the Tuning Barrier: Zero-Hyperparameters Yield Multi-Corner Analysis Via Learned Priors (viability: 3): https://sciencetostartup.com/paper/breaking-the-tuning-barrier-zero-hyperparameters-yield-multi-corner-analysis-via-learned-priors - A novel approach to circuit validation that eliminates hyperparameter tuning by using learned priors from a foundation model.
- Reasoning over Video: Evaluating How MLLMs Extract, Integrate, and Reconstruct Spatiotemporal Evidence (viability: 4): https://sciencetostartup.com/paper/reasoning-over-video-evaluating-how-mllms-extract-integrate-and-reconstruct-spatiotemporal-evidence - A benchmark for evaluating multimodal large language models in abstractive spatiotemporal reasoning from videos.
- V-Bridge: Bridging Video Generative Priors to Versatile Few-shot Image Restoration (viability: 7): https://sciencetostartup.com/paper/v-bridge-bridging-video-generative-priors-to-versatile-few-shot-image-restoration - V-Bridge leverages video generative models for efficient few-shot image restoration, achieving high-quality results with minimal training data.
- Influence Malleability in Linearized Attention: Dual Implications of Non-Convergent NTK Dynamics (viability: 3): https://sciencetostartup.com/paper/influence-malleability-in-linearized-attention-dual-implications-of-non-convergent-ntk-dynamics - This paper explores the non-convergent dynamics of linearized attention mechanisms and their implications for learning and adversarial vulnerability.
- Human-in-the-Loop LLM Grading for Handwritten Mathematics Assessments (viability: 5): https://sciencetostartup.com/paper/human-in-the-loop-llm-grading-for-handwritten-mathematics-assessments - A scalable workflow for LLM-assisted grading of handwritten mathematics assessments that reduces grading time while maintaining accuracy.
- InterEdit: Navigating Text-Guided Multi-Human 3D Motion Editing (viability: 8): https://sciencetostartup.com/paper/interedit-navigating-text-guided-multi-human-3d-motion-editing - InterEdit enables advanced text-guided multi-human 3D motion editing with a new dataset and state-of-the-art performance.
- Rooftop Wind Field Reconstruction Using Sparse Sensors: From Deterministic to Generative Learning Methods (viability: 3): https://sciencetostartup.com/paper/rooftop-wind-field-reconstruction-using-sparse-sensors-from-deterministic-to-generative-learning-methods - A framework for reconstructing rooftop wind fields using deep learning from sparse sensor data.
- Mitigating Memorization in Text-to-Image Diffusion via Region-Aware Prompt Augmentation and Multimodal Copy Detection (viability: 7): https://sciencetostartup.com/paper/mitigating-memorization-in-text-to-image-diffusion-via-region-aware-prompt-augmentation-and-multimodal-copy-detection - A novel approach to enhance text-to-image diffusion models by mitigating memorization risks through region-aware prompt augmentation and multimodal copy detection.
- Fractals made Practical: Denoising Diffusion as Partitioned Iterated Function Systems (viability: 2): https://sciencetostartup.com/paper/fractals-made-practical-denoising-diffusion-as-partitioned-iterated-function-systems - This paper explores the theoretical underpinnings of diffusion models through the lens of Partitioned Iterated Function Systems.
- GeoChemAD: Benchmarking Unsupervised Geochemical Anomaly Detection for Mineral Exploration (viability: 7): https://sciencetostartup.com/paper/geochemad-benchmarking-unsupervised-geochemical-anomaly-detection-for-mineral-exploration - GeoChemAD provides an open-source benchmark dataset and a transformer-based framework for unsupervised geochemical anomaly detection in mineral exploration.
- L2GTX: From Local to Global Time Series Explanations (viability: 4): https://sciencetostartup.com/paper/l2gtx-from-local-to-global-time-series-explanations - L2GTX is a model-agnostic framework for generating global explanations in time series classification.
- Competition-Aware CPC Forecasting with Near-Market Coverage (viability: 7): https://sciencetostartup.com/paper/competition-aware-cpc-forecasting-with-near-market-coverage - A scalable solution for improving CPC forecasting in auction-driven markets using competition-aware signals.
- Reference-Free Image Quality Assessment for Virtual Try-On via Human Feedback (viability: 7): https://sciencetostartup.com/paper/reference-free-image-quality-assessment-for-virtual-try-on-via-human-feedback - A reference-free framework for evaluating image quality in virtual try-on applications using human feedback.
- Team RAS in 10th ABAW Competition: Multimodal Valence and Arousal Estimation Approach (viability: 4): https://sciencetostartup.com/paper/team-ras-in-10th-abaw-competition-multimodal-valence-and-arousal-estimation-approach - A multimodal approach for continuous emotion recognition using face, behavior, and audio data.
- Topo-R1: Detecting Topological Anomalies via Vision-Language Models (viability: 8): https://sciencetostartup.com/paper/topo-r1-detecting-topological-anomalies-via-vision-language-models - Topo-R1 leverages Vision-Language Models to detect topological anomalies in structures without the need for ground-truth supervision.
- Causal Cellular Context Transfer Learning (C3TL): An Efficient Architecture for Prediction of Unseen Perturbation Effects (viability: 7): https://sciencetostartup.com/paper/causal-cellular-context-transfer-learning-c3tl-an-efficient-architecture-for-prediction-of-unseen-perturbation-effects - A lightweight framework for predicting chemical and genetic perturbation effects using minimal data and resources.
- 3DTCR: A Physics-Based Generative Framework for Vortex-Following 3D Reconstruction to Improve Tropical Cyclone Intensity Forecasting (viability: 7): https://sciencetostartup.com/paper/3dtcr-a-physics-based-generative-framework-for-vortex-following-3d-reconstruction-to-improve-tropical-cyclone-intensity- - 3DTCR is a physics-based generative framework that enhances tropical cyclone intensity forecasting by reconstructing fine-scale 3D structures efficiently.
- Are General-Purpose Vision Models All We Need for 2D Medical Image Segmentation? A Cross-Dataset Empirical Study (viability: 7): https://sciencetostartup.com/paper/are-general-purpose-vision-models-all-we-need-for-2d-medical-image-segmentation-a-cross-dataset-empirical-study - This study evaluates the effectiveness of general-purpose vision models for 2D medical image segmentation, suggesting they may outperform specialized models.
- Mending the Holes: Mitigating Reward Hacking in Reinforcement Learning for Multilingual Translation (viability: 7): https://sciencetostartup.com/paper/mending-the-holes-mitigating-reward-hacking-in-reinforcement-learning-for-multilingual-translation - WALAR enhances multilingual translation capabilities of LLMs using reinforcement learning with monolingual data.
- OpenACMv2: An Accuracy-Constrained Co-Optimization Framework for Approximate DCiM (viability: 8): https://sciencetostartup.com/paper/openacmv2-an-accuracy-constrained-co-optimization-framework-for-approximate-dcim - OpenACMv2 is an open framework for optimizing power-performance-area in approximate DCiM with accuracy constraints.
- Interpretable Semantic Gradients in SSD: A PCA Sweep Approach and a Case Study on AI Discourse (viability: 2): https://sciencetostartup.com/paper/interpretable-semantic-gradients-in-ssd-a-pca-sweep-approach-and-a-case-study-on-ai-discourse - A method for improving the interpretability of semantic gradients in text analysis.
- Federated Few-Shot Learning on Neuromorphic Hardware: An Empirical Study Across Physical Edge Nodes (viability: 4): https://sciencetostartup.com/paper/federated-few-shot-learning-on-neuromorphic-hardware-an-empirical-study-across-physical-edge-nodes - A novel federated learning approach utilizing neuromorphic hardware for efficient edge computing.
- Interrogating Design Homogenization in Web Vibe Coding (viability: 2): https://sciencetostartup.com/paper/interrogating-design-homogenization-in-web-vibe-coding - This paper explores the risks of design homogenization in web vibe coding and proposes a framework to mitigate these effects.
- ESPIRE: A Diagnostic Benchmark for Embodied Spatial Reasoning of Vision-Language Models (viability: 4): https://sciencetostartup.com/paper/espire-a-diagnostic-benchmark-for-embodied-spatial-reasoning-of-vision-language-models - ESPIRE is a diagnostic benchmark designed to enhance spatial reasoning in vision-language models for robotic tasks.
- Multimodal OCR: Parse Anything from Documents (viability: 8): https://sciencetostartup.com/paper/multimodal-ocr-parse-anything-from-documents - A next-gen OCR system that parses documents into structured text and graphics for seamless integration and data retrieval.
- Purify Once, Edit Freely: Breaking Image Protections under Model Mismatch (viability: 3): https://sciencetostartup.com/paper/purify-once-edit-freely-breaking-image-protections-under-model-mismatch - A framework for evaluating and improving image protection methods against model mismatch in diffusion models.
- SortScrews: A Dataset and Baseline for Real-time Screw Classification (viability: 8): https://sciencetostartup.com/paper/sortscrews-a-dataset-and-baseline-for-real-time-screw-classification - SortScrews provides a dataset and baseline for real-time screw classification, enhancing automation in industrial settings.
- PISmith: Reinforcement Learning-based Red Teaming for Prompt Injection Defenses (viability: 8): https://sciencetostartup.com/paper/pismith-reinforcement-learning-based-red-teaming-for-prompt-injection-defenses - PISmith is a reinforcement learning framework that enhances prompt injection defenses for LLM applications by systematically evaluating their robustness against adaptive attacks.
- SAW: Toward a Surgical Action World Model via Controllable and Scalable Video Generation (viability: 7): https://sciencetostartup.com/paper/saw-toward-a-surgical-action-world-model-via-controllable-and-scalable-video-generation - SAW generates realistic surgical action videos to enhance surgical AI and simulation.
- daVinci-Env: Open SWE Environment Synthesis at Scale (viability: 8): https://sciencetostartup.com/paper/davinci-env-open-swe-environment-synthesis-at-scale - Building the largest open-source SWE environment for training scalable and verifiable software engineering agents.
- ARL-Tangram: Unleash the Resource Efficiency in Agentic Reinforcement Learning (viability: 6): https://sciencetostartup.com/paper/arl-tangram-unleash-the-resource-efficiency-in-agentic-reinforcement-learning - ARL-Tangram optimizes resource efficiency in agentic reinforcement learning for large language models.
- Structured Distillation for Personalized Agent Memory: 11x Token Reduction with Retrieval Preservation (viability: 8): https://sciencetostartup.com/paper/structured-distillation-for-personalized-agent-memory-11x-token-reduction-with-retrieval-preservation - A novel approach to compressing AI agent memory for efficient personalized retrieval, achieving 11x token reduction without losing recall quality.
- FraudFox: Adaptable Fraud Detection in the Real World (viability: 8): https://sciencetostartup.com/paper/fraudfox-adaptable-fraud-detection-in-the-real-world - FraudFox is an adaptable fraud detection tool for e-commerce platforms that leverages Kalman Filters to dynamically update risk assessments and reduce fraud losses.
- From Passive Monitoring to Active Defence: Resilient Control of Manipulators Under Cyberattacks (viability: 2): https://sciencetostartup.com/paper/from-passive-monitoring-to-active-defence-resilient-control-of-manipulators-under-cyberattacks - This paper presents a novel active defence mechanism for robotic systems against cyberattacks, enhancing resilience and stability.
- A Closed-Form Solution for Debiasing Vision-Language Models with Utility Guarantees Across Modalities and Tasks (viability: 7): https://sciencetostartup.com/paper/a-closed-form-solution-for-debiasing-vision-language-models-with-utility-guarantees-across-modalities-and-tasks - A novel training-free debiasing method for Vision-Language Models that ensures fairness without sacrificing utility.
- Deconstructing the Failure of Ideal Noise Correction: A Three-Pillar Diagnosis (viability: 3): https://sciencetostartup.com/paper/deconstructing-the-failure-of-ideal-noise-correction-a-three-pillar-diagnosis - This research identifies fundamental flaws in ideal noise correction methods for learning with noisy labels, providing insights for more reliable approaches.
- Dependency-Aware Parallel Decoding via Attention for Diffusion LLMs (viability: 2): https://sciencetostartup.com/paper/dependency-aware-parallel-decoding-via-attention-for-diffusion-llms - A novel decoding method for diffusion LLMs that enhances parallel token unmasking by leveraging self-attention.
- Route Fragmentation Based on Resource-centric Prioritisation for Efficient Multi-Robot Path Planning in Agricultural Environments (viability: 7): https://sciencetostartup.com/paper/route-fragmentation-based-on-resource-centric-prioritisation-for-efficient-multi-robot-path-planning-in-agricultural-env - Resource-centric multi-robot path planning algorithms enhance agricultural robot navigation efficiency.
- Mitigating Collusion in Proofs of Liabilities (viability: 3): https://sciencetostartup.com/paper/mitigating-collusion-in-proofs-of-liabilities - A new model for proofs of liabilities in cryptocurrency exchanges that enhances security against collusion.
- Test-Time Attention Purification for Backdoored Large Vision Language Models (viability: 7): https://sciencetostartup.com/paper/test-time-attention-purification-for-backdoored-large-vision-language-models - CleanSight offers a novel, training-free defense against backdoor attacks in large vision-language models by purifying inputs at test time.
- Fair Lung Disease Diagnosis from Chest CT via Gender-Adversarial Attention Multiple Instance Learning (viability: 8): https://sciencetostartup.com/paper/fair-lung-disease-diagnosis-from-chest-ct-via-gender-adversarial-attention-multiple-instance-learning - A fairness-aware framework for diagnosing lung diseases from chest CT scans using gender-adversarial attention.
- Retrieval-Enhanced Real Estate Appraisal (viability: 5): https://sciencetostartup.com/paper/retrieval-enhanced-real-estate-appraisal - A hybrid model for improving real estate appraisal by optimizing comparable property selection.
- Is Human Annotation Necessary? Iterative MBR Distillation for Error Span Detection in Machine Translation (viability: 7): https://sciencetostartup.com/paper/is-human-annotation-necessary-iterative-mbr-distillation-for-error-span-detection-in-machine-translation - A self-evolution framework for error span detection in machine translation that eliminates the need for human annotations.
- Exact Federated Continual Unlearning for Ridge Heads on Frozen Foundation Models (viability: 7): https://sciencetostartup.com/paper/exact-federated-continual-unlearning-for-ridge-heads-on-frozen-foundation-models - A federated unlearning method for frozen foundation models that ensures exact removal of user data with minimal computational cost.
- SCOPE: Semantic Coreset with Orthogonal Projection Embeddings for Federated learning (viability: 6): https://sciencetostartup.com/paper/scope-semantic-coreset-with-orthogonal-projection-embeddings-for-federated-learning - SCOPE is a federated learning framework that optimizes data selection to improve accuracy and efficiency in class-imbalanced datasets.
- A Requirement-Based Framework for Engineering Adaptive Authentication (viability: 4): https://sciencetostartup.com/paper/a-requirement-based-framework-for-engineering-adaptive-authentication - A framework for adaptive authentication that dynamically selects methods based on contextual factors and security risks.
- Language-Grounded Decoupled Action Representation for Robotic Manipulation (viability: 7): https://sciencetostartup.com/paper/language-grounded-decoupled-action-representation-for-robotic-manipulation - LaDA is a framework that enhances robotic manipulation by using natural language to bridge perception and control through interpretable action primitives.
- Long-form RewardBench: Evaluating Reward Models for Long-form Generation (viability: 5): https://sciencetostartup.com/paper/long-form-rewardbench-evaluating-reward-models-for-long-form-generation - Long-form RewardBench is a benchmark designed to evaluate reward models specifically for long-form generation tasks.
- Efficient Real-World Autonomous Racing via Attenuated Residual Policy Optimization (viability: 7): https://sciencetostartup.com/paper/efficient-real-world-autonomous-racing-via-attenuated-residual-policy-optimization - Use α-RPO for enhanced autonomous racing performance with reduced system complexity and real-world applicability.
- Delta1 with LLM: symbolic and neural integration for credible and explainable reasoning (viability: 3): https://sciencetostartup.com/paper/delta1-with-llm-symbolic-and-neural-integration-for-credible-and-explainable-reasoning - Delta1 integrates symbolic reasoning with LLMs for explainable AI in various domains.
- Almost-Free Queue Jumping for Prior Inputs in Private Neural Inference (viability: 7): https://sciencetostartup.com/paper/almost-free-queue-jumping-for-prior-inputs-in-private-neural-inference - PrivQJ enhances privacy-preserving neural network inference by enabling efficient priority handling for urgent requests.
- ReMem-VLA: Empowering Vision-Language-Action Model with Memory via Dual-Level Recurrent Queries (viability: 7): https://sciencetostartup.com/paper/remem-vla-empowering-vision-language-action-model-with-memory-via-dual-level-recurrent-queries - ReMem-VLA enhances robot control by integrating advanced memory mechanisms into vision-language-action models.
- Coordinated Manipulation of Hybrid Deformable-Rigid Objects in Constrained Environments (viability: 7): https://sciencetostartup.com/paper/coordinated-manipulation-of-hybrid-deformable-rigid-objects-in-constrained-environments - A robotic manipulation planner for hybrid deformable-rigid objects that optimizes performance in constrained environments.
- RoboStream: Weaving Spatio-Temporal Reasoning with Memory in Vision-Language Models for Robotics (viability: 7): https://sciencetostartup.com/paper/robostream-weaving-spatio-temporal-reasoning-with-memory-in-vision-language-models-for-robotics - RoboStream enhances robotic manipulation by integrating spatio-temporal reasoning and memory for improved decision-making.
- Thinking in Streaming Video (viability: 7): https://sciencetostartup.com/paper/thinking-in-streaming-video - ThinkStream enables real-time video streaming reasoning with low latency using a novel incremental update framework.
- SGMatch: Semantic-Guided Non-Rigid Shape Matching with Flow Regularization (viability: 4): https://sciencetostartup.com/paper/sgmatch-semantic-guided-non-rigid-shape-matching-with-flow-regularization - SGMatch is a learning-based framework that enhances non-rigid shape matching by integrating semantic features with geometric descriptors.
- MotionAnymesh: Physics-Grounded Articulation for Simulation-Ready Digital Twins (viability: 7): https://sciencetostartup.com/paper/motionanymesh-physics-grounded-articulation-for-simulation-ready-digital-twins - MotionAnymesh transforms static 3D meshes into simulation-ready digital twins using physics-grounded techniques.
- Efficient and Interpretable Multi-Agent LLM Routing via Ant Colony Optimization (viability: 7): https://sciencetostartup.com/paper/efficient-and-interpretable-multi-agent-llm-routing-via-ant-colony-optimization - AMRO-S is an efficient and interpretable routing framework for Multi-Agent Systems that optimizes resource utilization and performance.
- DS$^2$-Instruct: Domain-Specific Data Synthesis for Large Language Models Instruction Tuning (viability: 7): https://sciencetostartup.com/paper/ds-2-instruct-domain-specific-data-synthesis-for-large-language-models-instruction-tuning - DS$^2$-Instruct generates high-quality, domain-specific instruction datasets for fine-tuning Large Language Models without human supervision.
- Rethinking VLMs for Image Forgery Detection and Localization (viability: 8): https://sciencetostartup.com/paper/rethinking-vlms-for-image-forgery-detection-and-localization - AI system using vision-language models for advanced image forgery detection and localization.
- ODRL Policy Comparison Through Normalisation (viability: 2): https://sciencetostartup.com/paper/odrl-policy-comparison-through-normalisation - A theoretical approach to simplify and normalize ODRL policies for better interoperability.
- HMS-BERT: Hybrid Multi-Task Self-Training for Multilingual and Multi-Label Cyberbullying Detection (viability: 6): https://sciencetostartup.com/paper/hms-bert-hybrid-multi-task-self-training-for-multilingual-and-multi-label-cyberbullying-detection - HMS-BERT is a hybrid multi-task framework for effective multilingual and multi-label cyberbullying detection.
- VIRD: View-Invariant Representation through Dual-Axis Transformation for Cross-View Pose Estimation (viability: 7): https://sciencetostartup.com/paper/vird-view-invariant-representation-through-dual-axis-transformation-for-cross-view-pose-estimation - VIRD offers a novel approach to cross-view pose estimation by creating view-invariant representations for improved accuracy in autonomous driving and robotics.
- Surprised by Attention: Predictable Query Dynamics for Time Series Anomaly Detection (viability: 8): https://sciencetostartup.com/paper/surprised-by-attention-predictable-query-dynamics-for-time-series-anomaly-detection - AxonAD is an unsupervised anomaly detection tool for multivariate time series that leverages predictable query dynamics to enhance detection accuracy.
- Stake the Points: Structure-Faithful Instance Unlearning (viability: 3): https://sciencetostartup.com/paper/stake-the-points-structure-faithful-instance-unlearning - A framework for machine unlearning that preserves knowledge structure while removing data influence.
- FedBPrompt: Federated Domain Generalization Person Re-Identification via Body Distribution Aware Visual Prompts (viability: 8): https://sciencetostartup.com/paper/fedbprompt-federated-domain-generalization-person-re-identification-via-body-distribution-aware-visual-prompts - FedBPrompt enhances federated person re-identification by using learnable visual prompts to improve feature discrimination across decentralized data.
- GoalSwarm: Multi-UAV Semantic Coordination for Open-Vocabulary Object Navigation (viability: 3): https://sciencetostartup.com/paper/goalswarm-multi-uav-semantic-coordination-for-open-vocabulary-object-navigation - GoalSwarm enables decentralized multi-UAV navigation for open-vocabulary object detection in complex environments.
- Learning from Child-Directed Speech in Two-Language Scenarios: A French-English Case Study (viability: 2): https://sciencetostartup.com/paper/learning-from-child-directed-speech-in-two-language-scenarios-a-french-english-case-study - A study on compact language models for English-French scenarios using child-directed speech.
- DirPA: Addressing Prior Shift in Imbalanced Few-shot Crop-type Classification (viability: 4): https://sciencetostartup.com/paper/dirpa-addressing-prior-shift-in-imbalanced-few-shot-crop-type-classification - DirPA enhances few-shot learning for agricultural monitoring by addressing class imbalance and distribution shifts.
- Consistent and Efficient MSCKF-based LiDAR-Inertial Odometry with Inferred Cluster-to-Plane Constraints for UAVs (viability: 7): https://sciencetostartup.com/paper/consistent-and-efficient-msckf-based-lidar-inertial-odometry-with-inferred-cluster-to-plane-constraints-for-uavs - A novel LiDAR-Inertial Odometry framework for UAVs that enhances navigation consistency and efficiency.
- Spectral-Geometric Neural Fields for Pose-Free LiDAR View Synthesis (viability: 7): https://sciencetostartup.com/paper/spectral-geometric-neural-fields-for-pose-free-lidar-view-synthesis - A pose-free LiDAR NeRF framework that enhances view synthesis by integrating spectral information with geometric consistency.
- A theory of learning data statistics in diffusion models, from easy to hard (viability: 2): https://sciencetostartup.com/paper/a-theory-of-learning-data-statistics-in-diffusion-models-from-easy-to-hard - This paper explores the learning dynamics of diffusion models, revealing how they learn data statistics from simple to complex.
- Human-Centered Evaluation of an LLM-Based Process Modeling Copilot: A Mixed-Methods Study with Domain Experts (viability: 6): https://sciencetostartup.com/paper/human-centered-evaluation-of-an-llm-based-process-modeling-copilot-a-mixed-methods-study-with-domain-experts - An LLM-powered copilot for democratizing BPMN modeling, enhancing usability and trust for non-experts.
- Finite Difference Flow Optimization for RL Post-Training of Text-to-Image Models (viability: 7): https://sciencetostartup.com/paper/finite-difference-flow-optimization-for-rl-post-training-of-text-to-image-models - A novel online reinforcement learning approach that optimizes image synthesis models for better quality and prompt alignment.
- Forecasting Epileptic Seizures from Contactless Camera via Cross-Species Transfer Learning (viability: 7): https://sciencetostartup.com/paper/forecasting-epileptic-seizures-from-contactless-camera-via-cross-species-transfer-learning - A cross-species transfer learning framework for non-invasive video-based forecasting of epileptic seizures.
- A protocol for evaluating robustness to H&E staining variation in computational pathology models (viability: 8): https://sciencetostartup.com/paper/a-protocol-for-evaluating-robustness-to-h-e-staining-variation-in-computational-pathology-models - A protocol for evaluating the robustness of computational pathology models to H&E staining variations.
- Enhanced Drug-drug Interaction Prediction Using Adaptive Knowledge Integration (viability: 5): https://sciencetostartup.com/paper/enhanced-drug-drug-interaction-prediction-using-adaptive-knowledge-integration - A knowledge augmentation framework that enhances drug-drug interaction predictions using large language models.
- Test-time RL alignment exposes task familiarity artifacts in LLM benchmarks (viability: 5): https://sciencetostartup.com/paper/test-time-rl-alignment-exposes-task-familiarity-artifacts-in-llm-benchmarks - A novel test-time reinforcement learning method that aligns LLMs to benchmarks without requiring task-specific training data.
- TRACE: Structure-Aware Character Encoding for Robust and Generalizable Document Watermarking (viability: 7): https://sciencetostartup.com/paper/trace-structure-aware-character-encoding-for-robust-and-generalizable-document-watermarking - TRACE is a robust framework for document watermarking that utilizes structure-aware character encoding to enhance data embedding resilience.
- CLARIN-PT-LDB: An Open LLM Leaderboard for Portuguese to assess Language, Culture and Civility (viability: 7): https://sciencetostartup.com/paper/clarin-pt-ldb-an-open-llm-leaderboard-for-portuguese-to-assess-language-culture-and-civility - A dedicated leaderboard for evaluating Open LLMs in European Portuguese, enhancing language and cultural alignment.
- FoSAM: Forward Secret Messaging in Ad-Hoc Networks (viability: 6): https://sciencetostartup.com/paper/fosam-forward-secret-messaging-in-ad-hoc-networks - FoSAM is a protocol that enables secure, anonymous messaging in ad-hoc networks without the need for an interactive handshake.
- Beyond Imitation: Reinforcement Learning Fine-Tuning for Adaptive Diffusion Navigation Policies (viability: 7): https://sciencetostartup.com/paper/beyond-imitation-reinforcement-learning-fine-tuning-for-adaptive-diffusion-navigation-policies - A reinforcement learning framework that enhances diffusion-based robot navigation policies for improved adaptability and safety in unseen environments.
- Composing Driving Worlds through Disentangled Control for Adversarial Scenario Generation (viability: 8): https://sciencetostartup.com/paper/composing-driving-worlds-through-disentangled-control-for-adversarial-scenario-generation - CompoSIA is a compositional driving video simulator that enables fine-grained control over adversarial driving scenarios to enhance safety in autonomous driving.
- Wear Classification of Abrasive Flap Wheels using a Hierarchical Deep Learning Approach (viability: 7): https://sciencetostartup.com/paper/wear-classification-of-abrasive-flap-wheels-using-a-hierarchical-deep-learning-approach - Automate wear condition monitoring of abrasive flap wheels using a hierarchical deep learning framework.
- On Linear Separability of the MNIST Handwritten Digits Dataset (viability: 3): https://sciencetostartup.com/paper/on-linear-separability-of-the-mnist-handwritten-digits-dataset - This paper investigates the linear separability of the MNIST dataset, providing empirical insights into its classification challenges.
- Team LEYA in 10th ABAW Competition: Multimodal Ambivalence/Hesitancy Recognition Approach (viability: 4): https://sciencetostartup.com/paper/team-leya-in-10th-abaw-competition-multimodal-ambivalence-hesitancy-recognition-approach - A multimodal approach for recognizing ambivalence and hesitancy in videos using integrated scene, facial, audio, and text analysis.
- Hierarchical Reference Sets for Robust Unsupervised Detection of Scattered and Clustered Outliers (viability: 7): https://sciencetostartup.com/paper/hierarchical-reference-sets-for-robust-unsupervised-detection-of-scattered-and-clustered-outliers - A novel outlier detection paradigm leveraging graph structures for robust anomaly evaluation in IoT data.
- Multimodal Protein Language Models for Enzyme Kinetic Parameters: From Substrate Recognition to Conformational Adaptation (viability: 3): https://sciencetostartup.com/paper/multimodal-protein-language-models-for-enzyme-kinetic-parameters-from-substrate-recognition-to-conformational-adaptation - A novel approach to predict enzyme kinetic parameters using multimodal protein language models.
- SmoothTurn: Learning to Turn Smoothly for Agile Navigation with Quadrupedal Robots (viability: 7): https://sciencetostartup.com/paper/smoothturn-learning-to-turn-smoothly-for-agile-navigation-with-quadrupedal-robots - SmoothTurn enables quadrupedal robots to navigate agilely by learning to turn smoothly while running at high speeds.
- DAST: A Dual-Stream Voice Anonymization Attacker with Staged Training (viability: 4): https://sciencetostartup.com/paper/dast-a-dual-stream-voice-anonymization-attacker-with-staged-training - DAST is a dual-stream voice anonymization attacker that enhances privacy evaluation through advanced training strategies.
- Mask2Flow-TSE: Two-Stage Target Speaker Extraction with Masking and Flow Matching (viability: 3): https://sciencetostartup.com/paper/mask2flow-tse-two-stage-target-speaker-extraction-with-masking-and-flow-matching - Mask2Flow-TSE offers a novel two-stage framework for efficient target speaker extraction from overlapping speech.
- Hierarchical Dual-Change Collaborative Learning for UAV Scene Change Captioning (viability: 6): https://sciencetostartup.com/paper/hierarchical-dual-change-collaborative-learning-for-uav-scene-change-captioning - A novel approach for generating natural language descriptions of dynamic changes in UAV imagery.
- coDrawAgents: A Multi-Agent Dialogue Framework for Compositional Image Generation (viability: 7): https://sciencetostartup.com/paper/codrawagents-a-multi-agent-dialogue-framework-for-compositional-image-generation - coDrawAgents is an interactive multi-agent framework that enhances compositional image generation through specialized dialogue agents.
- Rethinking Multiple-Choice Questions for RLVR: Unlocking Potential via Distractor Design (viability: 7): https://sciencetostartup.com/paper/rethinking-multiple-choice-questions-for-rlvr-unlocking-potential-via-distractor-design - A framework for designing effective distractors in multiple-choice questions to enhance reinforcement learning training.
- NanoVDR: Distilling a 2B Vision-Language Retriever into a 70M Text-Only Encoder for Visual Document Retrieval (viability: 8): https://sciencetostartup.com/paper/nanovdr-distilling-a-2b-vision-language-retriever-into-a-70m-text-only-encoder-for-visual-document-retrieval - NanoVDR distills a large vision-language model into a lightweight text-only encoder for efficient visual document retrieval.
- Adaptive Vision-Language Model Routing for Computer Use Agents (viability: 5): https://sciencetostartup.com/paper/adaptive-vision-language-model-routing-for-computer-use-agents - Optimize AI-powered Computer Use Agents by dynamically routing tasks to the most efficient Vision-Language Model.
- Residual SODAP: Residual Self-Organizing Domain-Adaptive Prompting with Structural Knowledge Preservation for Continual Learning (viability: 4): https://sciencetostartup.com/paper/residual-sodap-residual-self-organizing-domain-adaptive-prompting-with-structural-knowledge-preservation-for-continual-l - Residual SODAP enhances continual learning by preserving knowledge and adapting representations without requiring past data.
- Context is all you need: Towards autonomous model-based process design using agentic AI in flowsheet simulations (viability: 6): https://sciencetostartup.com/paper/context-is-all-you-need-towards-autonomous-model-based-process-design-using-agentic-ai-in-flowsheet-simulations - An agentic AI framework that enhances chemical process flowsheet modeling through LLM integration.
- OARS: Process-Aware Online Alignment for Generative Real-World Image Super-Resolution (viability: 7): https://sciencetostartup.com/paper/oars-process-aware-online-alignment-for-generative-real-world-image-super-resolution - OARS is a novel framework for aligning generative image super-resolution models with human visual preferences through online optimization.
- A Multi-task Large Reasoning Model for Molecular Science (viability: 8): https://sciencetostartup.com/paper/a-multi-task-large-reasoning-model-for-molecular-science - A multi-task reasoning model that enhances molecular science through structured reasoning and reflection.
- Reinforcement Learning for Elliptical Cylinder Motion Control Tasks (viability: 3): https://sciencetostartup.com/paper/reinforcement-learning-for-elliptical-cylinder-motion-control-tasks - A reinforcement learning approach to control elliptical cylinder motion under limited torque conditions.
- FLUX: Accelerating Cross-Embodiment Generative Navigation Policies via Rectified Flow and Static-to-Dynamic Learning (viability: 8): https://sciencetostartup.com/paper/flux-accelerating-cross-embodiment-generative-navigation-policies-via-rectified-flow-and-static-to-dynamic-learning - FLUX is a flow-based navigation policy that enhances autonomous navigation efficiency and robustness across diverse environments.
- What Makes VLMs Robust? Towards Reconciling Robustness and Accuracy in Vision-Language Models (viability: 8): https://sciencetostartup.com/paper/what-makes-vlms-robust-towards-reconciling-robustness-and-accuracy-in-vision-language-models - R-Adapt enhances adversarial robustness in Vision-Language Models without compromising clean accuracy.
- Spectral Defense Against Resource-Targeting Attack in 3D Gaussian Splatting (viability: 2): https://sciencetostartup.com/paper/spectral-defense-against-resource-targeting-attack-in-3d-gaussian-splatting - A novel defense mechanism against resource-targeting attacks in 3D Gaussian Splatting.
- SteerRM: Debiasing Reward Models via Sparse Autoencoders (viability: 3): https://sciencetostartup.com/paper/steerrm-debiasing-reward-models-via-sparse-autoencoders - SteerRM offers a novel training-free method to debias reward models using Sparse Autoencoders.
- A Fractional Fox H-Function Kernel for Support Vector Machines: Robust Classification via Weighted Transmutation Operators (viability: 4): https://sciencetostartup.com/paper/a-fractional-fox-h-function-kernel-for-support-vector-machines-robust-classification-via-weighted-transmutation-operator - Introducing the Fox-Dorrego Kernel for SVMs, enhancing classification robustness against outliers.
- Cheers: Decoupling Patch Details from Semantic Representations Enables Unified Multimodal Comprehension and Generation (viability: 8): https://sciencetostartup.com/paper/cheers-decoupling-patch-details-from-semantic-representations-enables-unified-multimodal-comprehension-and-generation - CHEERS revolutionizes multimodal AI with efficient, high-quality text and image generation in a unified model.
- Motion-Specific Battery Health Assessment for Quadrotors Using High-Fidelity Battery Models (viability: 5): https://sciencetostartup.com/paper/motion-specific-battery-health-assessment-for-quadrotors-using-high-fidelity-battery-models - An end-to-end framework for motion-aware battery health assessment in quadrotors to optimize battery management.
- Coherent Human-Scene Reconstruction from Multi-Person Multi-View Video in a Single Pass (viability: 8): https://sciencetostartup.com/paper/coherent-human-scene-reconstruction-from-multi-person-multi-view-video-in-a-single-pass - CHROMM offers a unified framework for real-time human-scene reconstruction from multi-view videos without preprocessing.
- Think and Answer ME: Benchmarking and Exploring Multi-Entity Reasoning Grounding in Remote Sensing (viability: 7): https://sciencetostartup.com/paper/think-and-answer-me-benchmarking-and-exploring-multi-entity-reasoning-grounding-in-remote-sensing - A framework for multi-entity reasoning in remote sensing visual grounding tasks, leveraging a new benchmark dataset and advanced reasoning models.
- Generalized Recognition of Basic Surgical Actions Enables Skill Assessment and Vision-Language-Model-based Surgical Planning (viability: 7): https://sciencetostartup.com/paper/generalized-recognition-of-basic-surgical-actions-enables-skill-assessment-and-vision-language-model-based-surgical-plan - A foundation model for recognizing basic surgical actions to enhance surgical training and planning.
- Upper Bounds for Local Learning Coefficients of Three-Layer Neural Networks (viability: 2): https://sciencetostartup.com/paper/upper-bounds-for-local-learning-coefficients-of-three-layer-neural-networks - This paper presents a theoretical upper bound for learning coefficients in three-layer neural networks.
- The RIGID Framework: Research-Integrated, Generative AI-Mediated Instructional Design (viability: 2): https://sciencetostartup.com/paper/the-rigid-framework-research-integrated-generative-ai-mediated-instructional-design - RIGID is a framework that integrates learning sciences research into instructional design using generative AI.
- Empowering Semantic-Sensitive Underwater Image Enhancement with VLM (viability: 3): https://sciencetostartup.com/paper/empowering-semantic-sensitive-underwater-image-enhancement-with-vlm - A novel mechanism using Vision-Language Models to enhance underwater image quality by focusing on semantic-sensitive regions.
- PVI: Plug-in Visual Injection for Vision-Language-Action Models (viability: 7): https://sciencetostartup.com/paper/pvi-plug-in-visual-injection-for-vision-language-action-models - PVI enhances vision-language-action models by injecting temporal visual features for improved manipulation tasks.
- Easy-IIL: Reducing Human Operational Burden in Interactive Imitation Learning via Assistant Experts (viability: 7): https://sciencetostartup.com/paper/easy-iil-reducing-human-operational-burden-in-interactive-imitation-learning-via-assistant-experts - Easy-IIL reduces human operational burden in Interactive Imitation Learning by utilizing assistant experts for data collection.
- SectEval: Evaluating the Latent Sectarian Preferences of Large Language Models (viability: 7): https://sciencetostartup.com/paper/secteval-evaluating-the-latent-sectarian-preferences-of-large-language-models - SectEval measures the bias of large language models in religious contexts across different languages and locations.
- Catalyst4D: High-Fidelity 3D-to-4D Scene Editing via Dynamic Propagation (viability: 7): https://sciencetostartup.com/paper/catalyst4d-high-fidelity-3d-to-4d-scene-editing-via-dynamic-propagation - Catalyst4D enables high-fidelity dynamic scene editing by maintaining spatial and temporal coherence in 4D Gaussian scenes.
- SAVA-X: Ego-to-Exo Imitation Error Detection via Scene-Adaptive View Alignment and Bidirectional Cross View Fusion (viability: 8): https://sciencetostartup.com/paper/sava-x-ego-to-exo-imitation-error-detection-via-scene-adaptive-view-alignment-and-bidirectional-cross-view-fusion - SAVA-X enhances error detection in industrial training by aligning ego and exo video demonstrations for improved accuracy.
- TerraFlow: Multimodal, Multitemporal Representation Learning for Earth Observation (viability: 7): https://sciencetostartup.com/paper/terraflow-multimodal-multitemporal-representation-learning-for-earth-observation - TerraFlow leverages multimodal, multitemporal learning to enhance Earth observation and disaster risk prediction.
- HIFICL: High-Fidelity In-Context Learning for Multimodal Tasks (viability: 8): https://sciencetostartup.com/paper/hificl-high-fidelity-in-context-learning-for-multimodal-tasks - HIFICL enhances In-Context Learning for multimodal tasks with a novel approach to context modeling.
- SAP: Segment Any 4K Panorama (viability: 7): https://sciencetostartup.com/paper/sap-segment-any-4k-panorama - SAP is a foundation model designed for high-resolution panoramic instance segmentation, improving performance on 360° images.
- FC-Track: Overlap-Aware Post-Association Correction for Online Multi-Object Tracking (viability: 7): https://sciencetostartup.com/paper/fc-track-overlap-aware-post-association-correction-for-online-multi-object-tracking - FC-Track is a lightweight framework that enhances online multi-object tracking by correcting identity mismatches caused by object overlap.
- AI Model Modulation with Logits Redistribution (viability: 3): https://sciencetostartup.com/paper/ai-model-modulation-with-logits-redistribution - AIM is a novel model modulation paradigm that allows a single AI model to adaptively exhibit diverse behaviors without retraining.
- A Method for Learning Large-Scale Computational Construction Grammars from Semantically Annotated Corpora (viability: 2): https://sciencetostartup.com/paper/a-method-for-learning-large-scale-computational-construction-grammars-from-semantically-annotated-corpora - A method for learning computational construction grammars from annotated language corpora.
- Balancing the privacy-utility trade-off: How to draw reliable conclusions from private data (viability: 2): https://sciencetostartup.com/paper/balancing-the-privacy-utility-trade-off-how-to-draw-reliable-conclusions-from-private-data - A new approach to understanding the privacy-utility trade-off in data analysis using hypothesis testing.
- Taming the Long Tail: Efficient Item-wise Sharpness-Aware Minimization for LLM-based Recommender Systems (viability: 6): https://sciencetostartup.com/paper/taming-the-long-tail-efficient-item-wise-sharpness-aware-minimization-for-llm-based-recommender-systems - EISAM is a novel optimization framework that enhances tail-item performance in LLM-based recommender systems by adaptively regularizing the loss landscape.
- Show, Don't Tell: Detecting Novel Objects by Watching Human Videos (viability: 8): https://sciencetostartup.com/paper/show-don-t-tell-detecting-novel-objects-by-watching-human-videos - A self-supervised system that enables robots to recognize novel objects through human demonstrations without complex language prompts.
- SLICE: Semantic Latent Injection via Compartmentalized Embedding for Image Watermarking (viability: 7): https://sciencetostartup.com/paper/slice-semantic-latent-injection-via-compartmentalized-embedding-for-image-watermarking - SLICE provides a robust, fine-grained watermarking solution for image provenance that withstands advanced adversarial attacks.
- Thinking in Dynamics: How Multimodal Large Language Models Perceive, Track, and Reason Dynamics in Physical 4D World (viability: 8): https://sciencetostartup.com/paper/thinking-in-dynamics-how-multimodal-large-language-models-perceive-track-and-reason-dynamics-in-physical-4d-world - Dyn-Bench enhances Multimodal Large Language Models' ability to perceive and reason about dynamic 4D scenes through a robust evaluation benchmark.
- TaoBench: Do Automated Theorem Prover LLMs Generalize Beyond MathLib? (viability: 6): https://sciencetostartup.com/paper/taobench-do-automated-theorem-prover-llms-generalize-beyond-mathlib - TaoBench is a novel benchmark for evaluating automated theorem provers beyond standard libraries, enhancing their applicability in research mathematics.
- MoKus: Leveraging Cross-Modal Knowledge Transfer for Knowledge-Aware Concept Customization (viability: 7): https://sciencetostartup.com/paper/mokus-leveraging-cross-modal-knowledge-transfer-for-knowledge-aware-concept-customization - MoKus enables high-fidelity customization of visual concepts through cross-modal knowledge transfer.
- ToolTree: Efficient LLM Agent Tool Planning via Dual-Feedback Monte Carlo Tree Search and Bidirectional Pruning (viability: 7): https://sciencetostartup.com/paper/tooltree-efficient-llm-agent-tool-planning-via-dual-feedback-monte-carlo-tree-search-and-bidirectional-pruning - ToolTree enhances LLM agent tool planning through advanced Monte Carlo tree search techniques for improved decision-making.
- SRAM-Based Compute-in-Memory Accelerator for Linear-decay Spiking Neural Networks (viability: 7): https://sciencetostartup.com/paper/sram-based-compute-in-memory-accelerator-for-linear-decay-spiking-neural-networks - A novel SRAM-based Compute-in-Memory architecture that optimizes state updates for energy-efficient Spiking Neural Networks.
- Conflict Mitigation in Shared Environments using Flow-Aware Multi-Agent Path Finding (viability: 7): https://sciencetostartup.com/paper/conflict-mitigation-in-shared-environments-using-flow-aware-multi-agent-path-finding - FA-MAPF enhances multi-robot navigation by integrating learned motion patterns of uncontrollable agents to reduce conflicts.
- VecMol: Vector-Field Representations for 3D Molecule Generation (viability: 5): https://sciencetostartup.com/paper/vecmol-vector-field-representations-for-3d-molecule-generation - VecMol introduces a novel framework for 3D molecule generation using continuous vector fields to enhance drug discovery.
- On Using Machine Learning to Early Detect Catastrophic Failures in Marine Diesel Engines (viability: 4): https://sciencetostartup.com/paper/on-using-machine-learning-to-early-detect-catastrophic-failures-in-marine-diesel-engines - A machine learning method for early detection of catastrophic failures in marine diesel engines to enhance safety and operational efficiency.
- AnchorVLA4D: an Anchor-Based Spatial-Temporal Vision-Language-Action Model for Robotic Manipulation (viability: 8): https://sciencetostartup.com/paper/anchorvla4d-an-anchor-based-spatial-temporal-vision-language-action-model-for-robotic-manipulation - AnchorVLA4D enhances robotic manipulation by integrating visual anchors for improved spatial-temporal reasoning.
- Anchored Alignment: Preventing Positional Collapse in Multimodal Recommender Systems (viability: 8): https://sciencetostartup.com/paper/anchored-alignment-preventing-positional-collapse-in-multimodal-recommender-systems - AnchorRec is a multimodal recommendation framework that enhances item representations while preserving modality-specific structures.
- Graph In-Context Operator Networks for Generalizable Spatiotemporal Prediction (viability: 4): https://sciencetostartup.com/paper/graph-in-context-operator-networks-for-generalizable-spatiotemporal-prediction - GICON leverages in-context operator learning for enhanced spatiotemporal prediction in real-world scenarios.
- SciDesignBench: Benchmarking and Improving Language Models for Scientific Inverse Design (viability: 6): https://sciencetostartup.com/paper/scidesignbench-benchmarking-and-improving-language-models-for-scientific-inverse-design - SciDesignBench benchmarks and improves language models for scientific inverse design tasks across various domains.
- CognitionCapturerPro: Towards High-Fidelity Visual Decoding from EEG/MEG via Multi-modal Information and Asymmetric Alignment (viability: 8): https://sciencetostartup.com/paper/cognitioncapturerpro-towards-high-fidelity-visual-decoding-from-eeg-meg-via-multi-modal-information-and-asymmetric-align - CognitionCapturerPro enhances visual stimuli reconstruction from EEG using multi-modal integration and advanced scoring mechanisms.
- CMHANet: A Cross-Modal Hybrid Attention Network for Point Cloud Registration (viability: 8): https://sciencetostartup.com/paper/cmhanet-a-cross-modal-hybrid-attention-network-for-point-cloud-registration - CMHANet enhances point cloud registration by integrating 2D image context with 3D geometric details for improved robustness.
- IGASA: Integrated Geometry-Aware and Skip-Attention Modules for Enhanced Point Cloud Registration (viability: 8): https://sciencetostartup.com/paper/igasa-integrated-geometry-aware-and-skip-attention-modules-for-enhanced-point-cloud-registration - IGASA is a novel framework for robust point cloud registration that enhances accuracy through advanced multi-scale feature extraction and fusion.
- The COTe score: A decomposable framework for evaluating Document Layout Analysis models (viability: 7): https://sciencetostartup.com/paper/the-cote-score-a-decomposable-framework-for-evaluating-document-layout-analysis-models - A novel metric and framework for evaluating Document Layout Analysis models that improves performance interpretation.
- Altered Thoughts, Altered Actions: Probing Chain-of-Thought Vulnerabilities in VLA Robotic Manipulation (viability: 3): https://sciencetostartup.com/paper/altered-thoughts-altered-actions-probing-chain-of-thought-vulnerabilities-in-vla-robotic-manipulation - This research investigates vulnerabilities in chain-of-thought reasoning for robotic manipulation tasks.
- UNIStainNet: Foundation-Model-Guided Virtual Staining of H&E to IHC (viability: 8): https://sciencetostartup.com/paper/unistainnet-foundation-model-guided-virtual-staining-of-h-e-to-ihc - UNIStainNet offers state-of-the-art virtual staining for rapid cancer diagnostics without the need for additional tissue processing.
- Design-Specification Tiling for ICL-based CAD Code Generation (viability: 7): https://sciencetostartup.com/paper/design-specification-tiling-for-icl-based-cad-code-generation - A novel exemplar selection method for improving CAD code generation using In-Context Learning.
- AI Planning Framework for LLM-Based Web Agents (viability: 6): https://sciencetostartup.com/paper/ai-planning-framework-for-llm-based-web-agents - A framework for diagnosing and improving LLM-based web agents through formal planning paradigms.
- Text-Phase Synergy Network with Dual Priors for Unsupervised Cross-Domain Image Retrieval (viability: 6): https://sciencetostartup.com/paper/text-phase-synergy-network-with-dual-priors-for-unsupervised-cross-domain-image-retrieval - TPSNet enhances unsupervised cross-domain image retrieval by integrating dual priors for improved semantic guidance.
- HFP-SAM: Hierarchical Frequency Prompted SAM for Efficient Marine Animal Segmentation (viability: 7): https://sciencetostartup.com/paper/hfp-sam-hierarchical-frequency-prompted-sam-for-efficient-marine-animal-segmentation - HFP-SAM enhances marine animal segmentation by integrating frequency information into the Segment Anything Model.
- Cost-Efficient Multimodal LLM Inference via Cross-Tier GPU Heterogeneity (viability: 3): https://sciencetostartup.com/paper/cost-efficient-multimodal-llm-inference-via-cross-tier-gpu-heterogeneity - HeteroServe optimizes multimodal LLM inference through cost-efficient cross-tier GPU scheduling.
- VCBench: A Streaming Counting Benchmark for Spatial-Temporal State Maintenance in Long Videos (viability: 4): https://sciencetostartup.com/paper/vcbench-a-streaming-counting-benchmark-for-spatial-temporal-state-maintenance-in-long-videos - VCBench is a streaming counting benchmark designed to enhance video understanding by diagnosing world state maintenance capabilities.
- FGTR: Fine-Grained Multi-Table Retrieval via Hierarchical LLM Reasoning (viability: 7): https://sciencetostartup.com/paper/fgtr-fine-grained-multi-table-retrieval-via-hierarchical-llm-reasoning - FGTR revolutionizes multi-table retrieval with hierarchical reasoning for enhanced accuracy and efficiency.
- Seeing Eye to Eye: Enabling Cognitive Alignment Through Shared First-Person Perspective in Human-AI Collaboration (viability: 7): https://sciencetostartup.com/paper/seeing-eye-to-eye-enabling-cognitive-alignment-through-shared-first-person-perspective-in-human-ai-collaboration - Eye2Eye enhances human-AI collaboration by using a first-person perspective for cognitive alignment.
- EvolveCoder: Evolving Test Cases via Adversarial Verification for Code Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/evolvecoder-evolving-test-cases-via-adversarial-verification-for-code-reinforcement-learning - EvolveCoder enhances code generation through adversarial verification and a refined RL dataset.
- HaltNav: Reactive Visual Halting over Lightweight Topological Priors for Robust Vision-Language Navigation (viability: 7): https://sciencetostartup.com/paper/haltnav-reactive-visual-halting-over-lightweight-topological-priors-for-robust-vision-language-navigation - HaltNav enhances vision-language navigation by integrating lightweight topological maps with reactive visual halting for robust real-world performance.
- RXNRECer Enables Fine-grained Enzymatic Function Annotation through Active Learning and Protein Language Models (viability: 7): https://sciencetostartup.com/paper/rxnrecer-enables-fine-grained-enzymatic-function-annotation-through-active-learning-and-protein-language-models - RXNRECer directly predicts enzyme-catalyzed reactions using protein language models and active learning, bypassing traditional EC number reliance.
- HSEmotion Team at ABAW-10 Competition: Facial Expression Recognition, Valence-Arousal Estimation, Action Unit Detection and Fine-Grained Violence Classification (viability: 6): https://sciencetostartup.com/paper/hsemotion-team-at-abaw-10-competition-facial-expression-recognition-valence-arousal-estimation-action-unit-detection-and - A fast approach for facial emotion recognition and violence detection using EfficientNet-based models and multi-layered perceptrons.
- CM-Bench: A Comprehensive Cross-Modal Feature Matching Benchmark Bridging Visible and Infrared Images (viability: 7): https://sciencetostartup.com/paper/cm-bench-a-comprehensive-cross-modal-feature-matching-benchmark-bridging-visible-and-infrared-images - CM-Bench is a benchmark for evaluating cross-modal feature matching algorithms between infrared and visible images.
- STRAP-ViT: Segregated Tokens with Randomized -- Transformations for Defense against Adversarial Patches in ViTs (viability: 7): https://sciencetostartup.com/paper/strap-vit-segregated-tokens-with-randomized-transformations-for-defense-against-adversarial-patches-in-vits - STRAP-ViT is a plug-and-play defense mechanism that enhances Vision Transformers' robustness against adversarial patches without additional training.
- Learning Athletic Humanoid Tennis Skills from Imperfect Human Motion Data (viability: 8): https://sciencetostartup.com/paper/learning-athletic-humanoid-tennis-skills-from-imperfect-human-motion-data - LATENT enables humanoid robots to learn tennis skills from imperfect human motion data, achieving robust performance in real-world scenarios.
- RSONet: Region-guided Selective Optimization Network for RGB-T Salient Object Detection (viability: 7): https://sciencetostartup.com/paper/rsonet-region-guided-selective-optimization-network-for-rgb-t-salient-object-detection - RSONet optimizes RGB-T salient object detection by addressing inconsistencies between RGB and thermal images.
- Federated Hierarchical Clustering with Automatic Selection of Optimal Cluster Numbers (viability: 4): https://sciencetostartup.com/paper/federated-hierarchical-clustering-with-automatic-selection-of-optimal-cluster-numbers - A federated clustering framework that automatically determines the optimal number of clusters from distributed data.
- Experimental evidence of progressive ChatGPT models self-convergence (viability: 4): https://sciencetostartup.com/paper/experimental-evidence-of-progressive-chatgpt-models-self-convergence - This study investigates the self-convergence phenomenon in ChatGPT models, revealing a decline in output diversity due to recursive training on synthetic data.
- Colluding LoRA: A Composite Attack on LLM Safety Alignment (viability: 2): https://sciencetostartup.com/paper/colluding-lora-a-composite-attack-on-llm-safety-alignment - Colluding LoRA presents a novel attack on LLM safety alignment through composition-triggered vulnerabilities.
- Bin~Wan,G2HFNet: GeoGran-Aware Hierarchical Feature Fusion Network for Salient Object Detection in Optical Remote Sensing Images (viability: 4): https://sciencetostartup.com/paper/bin-wan-g2hfnet-geogran-aware-hierarchical-feature-fusion-network-for-salient-object-detection-in-optical-remote-sensing - G2HFNet enhances salient object detection in remote sensing images using a novel hierarchical feature fusion approach.
- Why Neural Structural Obfuscation Can't Kill White-Box Watermarks for Good! (viability: 8): https://sciencetostartup.com/paper/why-neural-structural-obfuscation-can-t-kill-white-box-watermarks-for-good - Canon is a recovery framework that restores watermark verifiability in neural networks affected by obfuscation attacks.
- MetaKE: Meta-learning Aligned Knowledge Editing via Bi-level Optimization (viability: 7): https://sciencetostartup.com/paper/metake-meta-learning-aligned-knowledge-editing-via-bi-level-optimization - MetaKE offers a novel framework for knowledge editing in LLMs through bi-level optimization, enhancing editability without compromising performance.
- Disentangled Latent Dynamics Manifold Fusion for Solving Parameterized PDEs (viability: 7): https://sciencetostartup.com/paper/disentangled-latent-dynamics-manifold-fusion-for-solving-parameterized-pdes - DLDMF is a physics-informed framework that enhances neural surrogate models for parameterized PDEs by separating space, time, and parameters for improved generalization and extrapolation.
- Vision Verification Enhanced Fusion of VLMs for Efficient Visual Reasoning (viability: 8): https://sciencetostartup.com/paper/vision-verification-enhanced-fusion-of-vlms-for-efficient-visual-reasoning - A cutting-edge ensemble method for vision-language models that significantly improves visual reasoning accuracy by more than 8%.
- Marker-Based 3D Reconstruction of Aggregates with a Comparative Analysis of 2D and 3D Morphologies (viability: 3): https://sciencetostartup.com/paper/marker-based-3d-reconstruction-of-aggregates-with-a-comparative-analysis-of-2d-and-3d-morphologies - A cost-effective photogrammetry-based method for 3D reconstruction of aggregate particles to enhance quality control in construction.
- RetroReasoner: A Reasoning LLM for Strategic Retrosynthesis Prediction (viability: 3): https://sciencetostartup.com/paper/retroreasoner-a-reasoning-llm-for-strategic-retrosynthesis-prediction - RetroReasoner is an LLM designed to enhance retrosynthesis prediction by incorporating strategic reasoning.
- TacVLA: Contact-Aware Tactile Fusion for Robust Vision-Language-Action Manipulation (viability: 7): https://sciencetostartup.com/paper/tacvla-contact-aware-tactile-fusion-for-robust-vision-language-action-manipulation - TacVLA enhances robotic manipulation by integrating tactile feedback into vision-language-action models for improved performance in challenging tasks.
- From Text to Forecasts: Bridging Modality Gap with Temporal Evolution Semantic Space (viability: 5): https://sciencetostartup.com/paper/from-text-to-forecasts-bridging-modality-gap-with-temporal-evolution-semantic-space - TESS bridges the gap between textual information and time-series forecasting by introducing a novel Temporal Evolution Semantic Space.
- Learning Geometric and Photometric Features from Panoramic LiDAR Scans for Outdoor Place Categorization (viability: 6): https://sciencetostartup.com/paper/learning-geometric-and-photometric-features-from-panoramic-lidar-scans-for-outdoor-place-categorization - A novel method for outdoor place categorization using CNNs and LiDAR data to enhance autonomous navigation.
- AVION: Aerial Vision-Language Instruction from Offline Teacher to Prompt-Tuned Network (viability: 7): https://sciencetostartup.com/paper/avion-aerial-vision-language-instruction-from-offline-teacher-to-prompt-tuned-network - AVION enhances vision-language models for remote sensing through a knowledge distillation framework.
- Continual Learning in Large Language Models: Methods, Challenges, and Opportunities (viability: 3): https://sciencetostartup.com/paper/continual-learning-in-large-language-models-methods-challenges-and-opportunities - A comprehensive survey on continual learning methodologies for large language models to enhance adaptability and mitigate forgetting.
- VFM-Recon: Unlocking Cross-Domain Scene-Level Neural Reconstruction with Scale-Aligned Foundation Priors (viability: 8): https://sciencetostartup.com/paper/vfm-recon-unlocking-cross-domain-scene-level-neural-reconstruction-with-scale-aligned-foundation-priors - VFMRecon bridges the gap between vision foundation models and volumetric reconstruction for improved scene-level accuracy across domains.
- VGGT-World: Transforming VGGT into an Autoregressive Geometry World Model (viability: 7): https://sciencetostartup.com/paper/vggt-world-transforming-vggt-into-an-autoregressive-geometry-world-model - VGGT-World is a geometry world model that predicts future scene evolution using lightweight temporal flow transformers.
- Sobolev--Ricci Curvature (viability: 3): https://sciencetostartup.com/paper/sobolev-ricci-curvature - Introducing Sobolev-Ricci Curvature for scalable graph transformations and manifold-oriented pruning.
- Autonomous Integration and Improvement of Robotic Assembly using Skill Graph Representations (viability: 8): https://sciencetostartup.com/paper/autonomous-integration-and-improvement-of-robotic-assembly-using-skill-graph-representations - A framework for autonomous integration and improvement of robotic assembly systems using Skill Graph representations.
- From Sparse to Dense: Multi-View GRPO for Flow Models via Augmented Condition Space (viability: 7): https://sciencetostartup.com/paper/from-sparse-to-dense-multi-view-grpo-for-flow-models-via-augmented-condition-space - MV-GRPO enhances text-to-image flow models by improving sample evaluation through multi-view reward mapping.
- LR-SGS: Robust LiDAR-Reflectance-Guided Salient Gaussian Splatting for Self-Driving Scene Reconstruction (viability: 7): https://sciencetostartup.com/paper/lr-sgs-robust-lidar-reflectance-guided-salient-gaussian-splatting-for-self-driving-scene-reconstruction - LR-SGS enhances self-driving scene reconstruction by integrating LiDAR reflectance with Gaussian splatting for improved performance in complex environments.
- 98$\times$ Faster LLM Routing Without a Dedicated GPU: Flash Attention, Prompt Compression, and Near-Streaming for the vLLM Semantic Router (viability: 8): https://sciencetostartup.com/paper/98-times-faster-llm-routing-without-a-dedicated-gpu-flash-attention-prompt-compression-and-near-streaming-for-the-vllm-s - A high-performance semantic router for LLMs that dramatically reduces latency and memory usage without needing a dedicated GPU.
- LightMoE: Reducing Mixture-of-Experts Redundancy through Expert Replacing (viability: 7): https://sciencetostartup.com/paper/lightmoe-reducing-mixture-of-experts-redundancy-through-expert-replacing - LightMoE optimizes Mixture-of-Experts models by replacing redundant experts with efficient modules, achieving significant performance gains with lower memory demands.
- Uncovering Security Threats and Architecting Defenses in Autonomous Agents: A Case Study of OpenClaw (viability: 8): https://sciencetostartup.com/paper/uncovering-security-threats-and-architecting-defenses-in-autonomous-agents-a-case-study-of-openclaw - Project ClawGuard aims to secure autonomous agents by implementing a novel defense architecture against emerging cybersecurity threats.
- RoboStereo: Dual-Tower 4D Embodied World Models for Unified Policy Optimization (viability: 7): https://sciencetostartup.com/paper/robostereo-dual-tower-4d-embodied-world-models-for-unified-policy-optimization - RoboStereo is a dual-tower 4D world model that enhances policy optimization in embodied AI through advanced simulation techniques.
- Using a Human-AI Teaming Approach to Create and Curate Scientific Datasets with the SCILIRE System (viability: 6): https://sciencetostartup.com/paper/using-a-human-ai-teaming-approach-to-create-and-curate-scientific-datasets-with-the-scilire-system - SCILIRE is a Human-AI system that enhances the creation and curation of scientific datasets through iterative workflows.
- ExpanderGraph-128: A Novel Graph-Theoretic Block Cipher with Formal Security Analysis and Hardware Implementation (viability: 5): https://sciencetostartup.com/paper/expandergraph-128-a-novel-graph-theoretic-block-cipher-with-formal-security-analysis-and-hardware-implementation - ExpanderGraph-128 is a lightweight block cipher leveraging expander-graph interaction for enhanced security and efficiency.
- Adaptive Diffusion Posterior Sampling for Data and Model Fusion of Complex Nonlinear Dynamical Systems (viability: 4): https://sciencetostartup.com/paper/adaptive-diffusion-posterior-sampling-for-data-and-model-fusion-of-complex-nonlinear-dynamical-systems - A generative machine learning framework for efficient surrogate modeling of chaotic dynamical systems.
- Spend Less, Reason Better: Budget-Aware Value Tree Search for LLM Agents (viability: 7): https://sciencetostartup.com/paper/spend-less-reason-better-budget-aware-value-tree-search-for-llm-agents - Budget-Aware Value Tree (BAVT) optimizes LLM agent performance by intelligently managing resource allocation during multi-hop reasoning.
- Batched Kernelized Bandits: Refinements and Extensions (viability: 2): https://sciencetostartup.com/paper/batched-kernelized-bandits-refinements-and-extensions - This paper refines regret bounds for black-box optimization in batch settings but lacks practical implementation details.
- VLM4Rec: Multimodal Semantic Representation for Recommendation with Large Vision-Language Models (viability: 8): https://sciencetostartup.com/paper/vlm4rec-multimodal-semantic-representation-for-recommendation-with-large-vision-language-models - VLM4Rec enhances multimodal recommendation systems by aligning item content with user preferences using a lightweight framework.
- Prompt-Driven Lightweight Foundation Model for Instance Segmentation-Based Fault Detection in Freight Trains (viability: 8): https://sciencetostartup.com/paper/prompt-driven-lightweight-foundation-model-for-instance-segmentation-based-fault-detection-in-freight-trains - Deployable fault detection system for freight trains using lightweight AI segmentation.
- AEGIS: No Tool Call Left Unchecked -- A Pre-Execution Firewall and Audit Layer for AI Agents (viability: 8): https://sciencetostartup.com/paper/aegis-no-tool-call-left-unchecked-a-pre-execution-firewall-and-audit-layer-for-ai-agents - AEGIS is a pre-execution firewall for AI agents that ensures safe tool usage through real-time risk scanning and human approval workflows.
- Human-AI Collaborative Autonomous Experimentation With Proxy Modeling for Comparative Observation (viability: 4): https://sciencetostartup.com/paper/human-ai-collaborative-autonomous-experimentation-with-proxy-modeling-for-comparative-observation - A human-AI collaborative framework for optimizing material properties through proxy modeling and Bayesian optimization.
- When Drafts Evolve: Speculative Decoding Meets Online Learning (viability: 7): https://sciencetostartup.com/paper/when-drafts-evolve-speculative-decoding-meets-online-learning - OnlineSpec enhances large language model inference speed by leveraging iterative feedback from draft models.
- Literary Narrative as Moral Probe : A Cross-System Framework for Evaluating AI Ethical Reasoning and Refusal Behavior (viability: 2): https://sciencetostartup.com/paper/literary-narrative-as-moral-probe-a-cross-system-framework-for-evaluating-ai-ethical-reasoning-and-refusal-behavior - A novel methodology for evaluating AI moral reasoning using literary narratives.
- ChainFuzzer: Greybox Fuzzing for Workflow-Level Multi-Tool Vulnerabilities in LLM Agents (viability: 3): https://sciencetostartup.com/paper/chainfuzzer-greybox-fuzzing-for-workflow-level-multi-tool-vulnerabilities-in-llm-agents - ChainFuzzer is a greybox framework designed to discover and reproduce multi-tool vulnerabilities in LLM agents.
- FastDSAC: Unlocking the Potential of Maximum Entropy RL in High-Dimensional Humanoid Control (viability: 7): https://sciencetostartup.com/paper/fastdsac-unlocking-the-potential-of-maximum-entropy-rl-in-high-dimensional-humanoid-control - FastDSAC leverages maximum entropy RL for improved humanoid control in high-dimensional spaces.
- CarPLAN: Context-Adaptive and Robust Planning with Dynamic Scene Awareness for Autonomous Driving (viability: 8): https://sciencetostartup.com/paper/carplan-context-adaptive-and-robust-planning-with-dynamic-scene-awareness-for-autonomous-driving - CarPLAN enhances autonomous vehicle motion planning with context-adaptive decision-making for diverse traffic scenarios.
- Mastering Negation: Boosting Grounding Models via Grouped Opposition-Based Learning (viability: 7): https://sciencetostartup.com/paper/mastering-negation-boosting-grounding-models-via-grouped-opposition-based-learning - D-Negation enhances vision-language grounding models by introducing a dataset and learning framework focused on negation semantics.
- A2Z-10M+: Geometric Deep Learning with A-to-Z BRep Annotations for AI-Assisted CAD Modeling and Reverse Engineering (viability: 7): https://sciencetostartup.com/paper/a2z-10m-geometric-deep-learning-with-a-to-z-brep-annotations-for-ai-assisted-cad-modeling-and-reverse-engineering - A2Z is a comprehensive dataset for enhancing CAD modeling and reverse engineering through geometric deep learning.
- A Prediction-as-Perception Framework for 3D Object Detection (viability: 7): https://sciencetostartup.com/paper/a-prediction-as-perception-framework-for-3d-object-detection - A novel framework that enhances 3D object detection accuracy and speed by integrating prediction and perception.
- Neural Gate: Mitigating Privacy Risks in LVLMs via Neuron-Level Gradient Gating (viability: 7): https://sciencetostartup.com/paper/neural-gate-mitigating-privacy-risks-in-lvlms-via-neuron-level-gradient-gating - Neural Gate enhances privacy in large vision-language models by editing neuron-level parameters to mitigate risks of sensitive data leakage.
- Feynman: Knowledge-Infused Diagramming Agent for Scalable Visual Designs (viability: 7): https://sciencetostartup.com/paper/feynman-knowledge-infused-diagramming-agent-for-scalable-visual-designs - Feynman is a scalable diagram generation agent that creates knowledge-infused visual designs with minimal cost and time.
- Optimize Wider, Not Deeper: Consensus Aggregation for Policy Optimization (viability: 7): https://sciencetostartup.com/paper/optimize-wider-not-deeper-consensus-aggregation-for-policy-optimization - CAPO optimizes policy updates by aggregating multiple PPO replicas, enhancing efficiency without increasing compute depth.
- Swap-guided Preference Learning for Personalized Reinforcement Learning from Human Feedback (viability: 8): https://sciencetostartup.com/paper/swap-guided-preference-learning-for-personalized-reinforcement-learning-from-human-feedback - Swap-guided Preference Learning enhances personalization in reinforcement learning by addressing limitations in traditional models.
- Maximizing Incremental Information Entropy for Contrastive Learning (viability: 7): https://sciencetostartup.com/paper/maximizing-incremental-information-entropy-for-contrastive-learning - IE-CL optimizes entropy gain in contrastive learning for improved representation performance.
- Early Pruning for Public Transport Routing (viability: 3): https://sciencetostartup.com/paper/early-pruning-for-public-transport-routing - Early Pruning enhances public transport routing efficiency by optimizing transfer connections without sacrificing path optimality.
- CA-HFP: Curvature-Aware Heterogeneous Federated Pruning with Model Reconstruction (viability: 6): https://sciencetostartup.com/paper/ca-hfp-curvature-aware-heterogeneous-federated-pruning-with-model-reconstruction - A framework for personalized compression in federated learning that enhances model accuracy while reducing computation costs.
- SDF-Net: Structure-Aware Disentangled Feature Learning for Opticall-SAR Ship Re-identification (viability: 9): https://sciencetostartup.com/paper/sdf-net-structure-aware-disentangled-feature-learning-for-opticall-sar-ship-re-identification - SDF-Net uses a structure-aware network to enhance cross-modal ship re-identification between optical and SAR imagery.
- MRGeo: Robust Cross-View Geo-Localization of Corrupted Images via Spatial and Channel Feature Enhancement (viability: 4): https://sciencetostartup.com/paper/mrgeo-robust-cross-view-geo-localization-of-corrupted-images-via-spatial-and-channel-feature-enhancement - MRGeo enhances cross-view geo-localization of corrupted images through a novel hierarchical defense strategy.
- Deferred is Better: A Framework for Multi-Granularity Deferred Interaction of Heterogeneous Features (viability: 3): https://sciencetostartup.com/paper/deferred-is-better-a-framework-for-multi-granularity-deferred-interaction-of-heterogeneous-features - A novel framework for improving click-through rate prediction by adaptively managing feature interactions.
- Skill-informed Data-driven Haptic Nudges for High-dimensional Human Motor Learning (viability: 6): https://sciencetostartup.com/paper/skill-informed-data-driven-haptic-nudges-for-high-dimensional-human-motor-learning - A data-driven framework for optimizing haptic feedback to enhance motor learning efficiency.
- RTD-Guard: A Black-Box Textual Adversarial Detection Framework via Replacement Token Detection (viability: 8): https://sciencetostartup.com/paper/rtd-guard-a-black-box-textual-adversarial-detection-framework-via-replacement-token-detection - RTD-Guard is a lightweight black-box framework for detecting textual adversarial attacks in NLP systems.
- DINOLight: Robust Ambient Light Normalization with Self-supervised Visual Prior Integration (viability: 7): https://sciencetostartup.com/paper/dinolight-robust-ambient-light-normalization-with-self-supervised-visual-prior-integration - DINOLight leverages DINOv2 for robust ambient light normalization and shadow removal in images.
- Expert Pyramid Tuning: Efficient Parameter Fine-Tuning for Expertise-Driven Task Allocation (viability: 7): https://sciencetostartup.com/paper/expert-pyramid-tuning-efficient-parameter-fine-tuning-for-expertise-driven-task-allocation - Expert Pyramid Tuning enhances multi-task LLM deployment by efficiently allocating tasks to specialized experts.
- A Spectral Revisit of the Distributional Bellman Operator under the Cramér Metric (viability: 2): https://sciencetostartup.com/paper/a-spectral-revisit-of-the-distributional-bellman-operator-under-the-cram-r-metric - This paper explores the theoretical aspects of distributional reinforcement learning without a clear path to practical application.
- AccelAes: Accelerating Diffusion Transformers for Training-Free Aesthetic-Enhanced Image Generation (viability: 8): https://sciencetostartup.com/paper/accelaes-accelerating-diffusion-transformers-for-training-free-aesthetic-enhanced-image-generation - AccelAes accelerates diffusion transformers for enhanced image generation by optimizing computation based on aesthetic descriptors.
- From Woofs to Words: Towards Intelligent Robotic Guide Dogs with Verbal Communication (viability: 3): https://sciencetostartup.com/paper/from-woofs-to-words-towards-intelligent-robotic-guide-dogs-with-verbal-communication - Developing a robotic guide dog that communicates verbally to assist visually impaired users in navigation.
- LMEB: Long-horizon Memory Embedding Benchmark (viability: 7): https://sciencetostartup.com/paper/lmeb-long-horizon-memory-embedding-benchmark - LMEB is a benchmark framework designed to evaluate memory embeddings for complex long-horizon retrieval tasks.
- Speech-Worthy Alignment for Japanese SpeechLLMs via Direct Preference Optimization (viability: 6): https://sciencetostartup.com/paper/speech-worthy-alignment-for-japanese-speechllms-via-direct-preference-optimization - A novel approach to align Japanese SpeechLLMs for more natural speech synthesis using preference-based optimization.
- AgentDrift: Unsafe Recommendation Drift Under Tool Corruption Hidden by Ranking Metrics in LLM Agents (viability: 2): https://sciencetostartup.com/paper/agentdrift-unsafe-recommendation-drift-under-tool-corruption-hidden-by-ranking-metrics-in-llm-agents - This paper addresses safety failures in tool-augmented LLM agents, highlighting the need for improved evaluation metrics.
- Variational Garrote for Sparse Inverse Problems (viability: 3): https://sciencetostartup.com/paper/variational-garrote-for-sparse-inverse-problems - A probabilistic method for improving reconstruction in sparse inverse problems using Variational Garrote.
- Lyapunov Stable Graph Neural Flow (viability: 7): https://sciencetostartup.com/paper/lyapunov-stable-graph-neural-flow - A novel defense framework for Graph Neural Networks that enhances robustness against adversarial attacks using Lyapunov stability.
- Scaling Laws and Pathologies of Single-Layer PINNs: Network Width and PDE Nonlinearity (viability: 2): https://sciencetostartup.com/paper/scaling-laws-and-pathologies-of-single-layer-pinns-network-width-and-pde-nonlinearity - This paper explores the scaling laws and optimization challenges of Single-Layer Physics-Informed Neural Networks in solving nonlinear PDEs.
- Reinforcement Learning for Diffusion LLMs with Entropy-Guided Step Selection and Stepwise Advantages (viability: 8): https://sciencetostartup.com/paper/reinforcement-learning-for-diffusion-llms-with-entropy-guided-step-selection-and-stepwise-advantages - A novel reinforcement learning approach for optimizing diffusion language models with state-of-the-art performance on coding and reasoning tasks.
- Beyond Dense Futures: World Models as Structured Planners for Robotic Manipulation (viability: 7): https://sciencetostartup.com/paper/beyond-dense-futures-world-models-as-structured-planners-for-robotic-manipulation - StructVLA enhances robotic manipulation by using structured planning for reliable control through predictive visual foresight.
- Asymptotic and Finite-Time Guarantees for Langevin-Based Temperature Annealing in InfoNCE (viability: 2): https://sciencetostartup.com/paper/asymptotic-and-finite-time-guarantees-for-langevin-based-temperature-annealing-in-infonce - The paper provides a theoretical framework for understanding temperature schedules in contrastive learning.
- CVGL: Causal Learning and Geometric Topology (viability: 8): https://sciencetostartup.com/paper/cvgl-causal-learning-and-geometric-topology - CLGT is a framework that enhances cross-view geo-localization by integrating causal learning and geometric topology for improved accuracy in autonomous navigation.
- Decoding Matters: Efficient Mamba-Based Decoder with Distribution-Aware Deep Supervision for Medical Image Segmentation (viability: 7): https://sciencetostartup.com/paper/decoding-matters-efficient-mamba-based-decoder-with-distribution-aware-deep-supervision-for-medical-image-segmentation - Deco-Mamba offers a novel decoder-centric approach for efficient and generalized medical image segmentation.
- Spatial Reasoning is Not a Free Lunch: A Controlled Study on LLaVA (viability: 4): https://sciencetostartup.com/paper/spatial-reasoning-is-not-a-free-lunch-a-controlled-study-on-llava - A study revealing the limitations of current vision-language models in spatial reasoning and proposing design improvements.
- Deep Distance Measurement Method for Unsupervised Multivariate Time Series Similarity Retrieval (viability: 7): https://sciencetostartup.com/paper/deep-distance-measurement-method-for-unsupervised-multivariate-time-series-similarity-retrieval - Deep Distance Measurement Method enhances unsupervised multivariate time series similarity retrieval for industrial applications.
- CALF: Communication-Aware Learning Framework for Distributed Reinforcement Learning (viability: 4): https://sciencetostartup.com/paper/calf-communication-aware-learning-framework-for-distributed-reinforcement-learning - CALF enhances distributed reinforcement learning by accounting for network conditions to improve real-world performance.
- As Language Models Scale, Low-order Linear Depth Dynamics Emerge (viability: 2): https://sciencetostartup.com/paper/as-language-models-scale-low-order-linear-depth-dynamics-emerge - This paper explores low-order linear dynamics in large language models, providing insights into their behavior and control.
- Embedded Quantum Machine Learning in Embedded Systems: Feasibility, Hybrid Architectures, and Quantum Co-Processors (viability: 2): https://sciencetostartup.com/paper/embedded-quantum-machine-learning-in-embedded-systems-feasibility-hybrid-architectures-and-quantum-co-processors - Exploring the feasibility of integrating quantum machine learning into resource-constrained embedded systems.
- Spatio-Semantic Expert Routing Architecture with Mixture-of-Experts for Referring Image Segmentation (viability: 7): https://sciencetostartup.com/paper/spatio-semantic-expert-routing-architecture-with-mixture-of-experts-for-referring-image-segmentation - SERA enhances referring image segmentation by introducing a novel expert routing architecture for improved spatial coherence and boundary precision.
- Do You See What I Am Pointing At? Gesture-Based Egocentric Video Question Answering (viability: 8): https://sciencetostartup.com/paper/do-you-see-what-i-am-pointing-at-gesture-based-egocentric-video-question-answering - EgoPointVQA enables AI assistants to understand and respond to user gestures in egocentric videos.
- A Reduction Algorithm for Markovian Contextual Linear Bandits (viability: 2): https://sciencetostartup.com/paper/a-reduction-algorithm-for-markovian-contextual-linear-bandits - This paper presents a theoretical reduction algorithm for Markovian contextual linear bandits.
- TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning (viability: 7): https://sciencetostartup.com/paper/terminator-learning-optimal-exit-points-for-early-stopping-in-chain-of-thought-reasoning - TERMINATOR optimizes reasoning lengths in Large Reasoning Models to reduce compute time without sacrificing performance.
- EB-RANSAC: Random Sample Consensus based on Energy-Based Model (viability: 4): https://sciencetostartup.com/paper/eb-ransac-random-sample-consensus-based-on-energy-based-model - EB-RANSAC simplifies robust estimation by eliminating complex sampling procedures.
- LLM BiasScope: A Real-Time Bias Analysis Platform for Comparative LLM Evaluation (viability: 8): https://sciencetostartup.com/paper/llm-biasscope-a-real-time-bias-analysis-platform-for-comparative-llm-evaluation - LLM BiasScope is a web application for real-time bias analysis and comparative evaluation of large language models.
- When LLM Judge Scores Look Good but Best-of-N Decisions Fail (viability: 4): https://sciencetostartup.com/paper/when-llm-judge-scores-look-good-but-best-of-n-decisions-fail - This research highlights the limitations of using global metrics for evaluating language model responses in best-of-n selection tasks.
- Curriculum Sampling: A Two-Phase Curriculum for Efficient Training of Flow Matching (viability: 3): https://sciencetostartup.com/paper/curriculum-sampling-a-two-phase-curriculum-for-efficient-training-of-flow-matching - Curriculum Sampling proposes a two-phase approach to enhance Flow Matching model training efficiency.
- Learning Pore-scale Multiphase Flow from 4D Velocimetry (viability: 3): https://sciencetostartup.com/paper/learning-pore-scale-multiphase-flow-from-4d-velocimetry - A multimodal learning framework for predicting multiphase flow in porous media using 4D velocimetry data.
- Addressing Data Scarcity in 3D Trauma Detection through Self-Supervised and Semi-Supervised Learning with Vertex Relative Position Encoding (viability: 7): https://sciencetostartup.com/paper/addressing-data-scarcity-in-3d-trauma-detection-through-self-supervised-and-semi-supervised-learning-with-vertex-relativ - A self-supervised and semi-supervised approach for efficient 3D trauma detection in CT scans addressing data scarcity.
- MemRoPE: Training-Free Infinite Video Generation via Evolving Memory Tokens (viability: 7): https://sciencetostartup.com/paper/memrope-training-free-infinite-video-generation-via-evolving-memory-tokens - MemRoPE offers a training-free solution for infinite video generation by utilizing evolving memory tokens for enhanced temporal coherence and visual fidelity.
- Byzantine-Robust Optimization under $(L_0, L_1)$-Smoothness (viability: 5): https://sciencetostartup.com/paper/byzantine-robust-optimization-under-l-0-l-1-smoothness - A robust optimization algorithm designed to withstand Byzantine attacks in distributed systems.
- Red-Teaming Vision-Language-Action Models via Quality Diversity Prompt Generation for Robust Robot Policies (viability: 7): https://sciencetostartup.com/paper/red-teaming-vision-language-action-models-via-quality-diversity-prompt-generation-for-robust-robot-policies - Q-DIG enhances the robustness of Vision-Language-Action models by generating diverse adversarial instructions to improve robotic performance.
- ELLA: Generative AI-Powered Social Robots for Early Language Development at Home (viability: 3): https://sciencetostartup.com/paper/ella-generative-ai-powered-social-robots-for-early-language-development-at-home - ELLA is a generative AI-powered social robot designed to enhance early language development in children through interactive storytelling.
- Adaptive Conditional Forest Sampling for Spectral Risk Optimisation under Decision-Dependent Uncertainty (viability: 3): https://sciencetostartup.com/paper/adaptive-conditional-forest-sampling-for-spectral-risk-optimisation-under-decision-dependent-uncertainty - A novel simulation-optimization framework for minimizing spectral risk under decision-dependent uncertainty.
- Naïve PAINE: Lightweight Text-to-Image Generation Improvement with Prompt Evaluation (viability: 7): https://sciencetostartup.com/paper/na-ve-paine-lightweight-text-to-image-generation-improvement-with-prompt-evaluation - Naïve PAINE enhances text-to-image generation by predicting image quality from prompts, improving output consistency.
- Robots that redesign themselves through kinematic self-destruction (viability: 2): https://sciencetostartup.com/paper/robots-that-redesign-themselves-through-kinematic-self-destruction - A robot that redesigns itself through kinematic self-destruction to enhance locomotion.
- TRACE: Temporal Rule-Anchored Chain-of-Evidence on Knowledge Graphs for Interpretable Stock Movement Prediction (viability: 6): https://sciencetostartup.com/paper/trace-temporal-rule-anchored-chain-of-evidence-on-knowledge-graphs-for-interpretable-stock-movement-prediction - TRACE leverages knowledge graphs and LLMs for interpretable stock movement predictions with enhanced accuracy.
- Probing Length Generalization in Mamba via Image Reconstruction (viability: 2): https://sciencetostartup.com/paper/probing-length-generalization-in-mamba-via-image-reconstruction - This paper analyzes and proposes improvements to the Mamba sequence model's performance on varying sequence lengths.
- Keys on Doormats: Exposed API Credentials on the Web (viability: 4): https://sciencetostartup.com/paper/keys-on-doormats-exposed-api-credentials-on-the-web - A study revealing widespread exposure of API credentials on the web, highlighting security vulnerabilities and responsible disclosure efforts.
- RAW-Domain Degradation Models for Realistic Smartphone Super-Resolution (viability: 6): https://sciencetostartup.com/paper/raw-domain-degradation-models-for-realistic-smartphone-super-resolution - A smartphone super-resolution model that enhances image quality by accurately modeling device-specific degradations.
- COAD: Constant-Time Planning for Continuous Goal Manipulation with Compressed Library and Online Adaptation (viability: 8): https://sciencetostartup.com/paper/coad-constant-time-planning-for-continuous-goal-manipulation-with-compressed-library-and-online-adaptation - COAD enables constant-time planning for robotic manipulation tasks by using a compressed library and online adaptation.
- Modal Logical Neural Networks for Financial AI (viability: 2): https://sciencetostartup.com/paper/modal-logical-neural-networks-for-financial-ai - Modal Logical Neural Networks bridge deep learning and symbolic logic for enhanced compliance in financial AI.
- Hunting CUDA Bugs at Scale with cuFuzz (viability: 9): https://sciencetostartup.com/paper/hunting-cuda-bugs-at-scale-with-cufuzz - cuFuzz is a CUDA-oriented fuzzer that enhances GPU program testing by effectively identifying memory-safety and concurrency bugs.
- Generating Expressive and Customizable Evals for Timeseries Data Analysis Agents with AgentFuel (viability: 7): https://sciencetostartup.com/paper/generating-expressive-and-customizable-evals-for-timeseries-data-analysis-agents-with-agentfuel - AgentFuel enables domain experts to create customizable evaluations for timeseries data analysis agents.
- CalliMaster: Mastering Page-level Chinese Calligraphy via Layout-guided Spatial Planning (viability: 7): https://sciencetostartup.com/paper/callimaster-mastering-page-level-chinese-calligraphy-via-layout-guided-spatial-planning - CalliMaster is a framework for generating and editing page-level Chinese calligraphy with spatial planning.
- One-Step Flow Policy: Self-Distillation for Fast Visuomotor Policies (viability: 8): https://sciencetostartup.com/paper/one-step-flow-policy-self-distillation-for-fast-visuomotor-policies - One-Step Flow Policy revolutionizes robotic control with high-fidelity, low-latency action generation.
- Less Data, Faster Convergence: Goal-Driven Data Optimization for Multimodal Instruction Tuning (viability: 9): https://sciencetostartup.com/paper/less-data-faster-convergence-goal-driven-data-optimization-for-multimodal-instruction-tuning - GDO optimizes data usage for multimodal instruction tuning, achieving faster convergence with fewer samples.
- The Perfection Paradox: From Architect to Curator in AI-Assisted API Design (viability: 3): https://sciencetostartup.com/paper/the-perfection-paradox-from-architect-to-curator-in-ai-assisted-api-design - AI-assisted design workflow enhances API specifications but raises concerns about human judgment.
- Marked Pedagogies: Examining Linguistic Biases in Personalized Automated Writing Feedback (viability: 4): https://sciencetostartup.com/paper/marked-pedagogies-examining-linguistic-biases-in-personalized-automated-writing-feedback - A study revealing biases in LLM-generated feedback for personalized writing assistance.
- Unleashing Video Language Models for Fine-grained HRCT Report Generation (viability: 9): https://sciencetostartup.com/paper/unleashing-video-language-models-for-fine-grained-hrct-report-generation - AbSteering leverages Video Language Models for precise HRCT report generation, enhancing diagnostic accuracy in medical imaging.
- Adaptation of Weakly Supervised Localization in Histopathology by Debiasing Predictions (viability: 8): https://sciencetostartup.com/paper/adaptation-of-weakly-supervised-localization-in-histopathology-by-debiasing-predictions - A novel method for improving weakly supervised localization in histopathology by debiasing predictions to enhance performance across varying domains.
- TaxBreak: Unmasking the Hidden Costs of LLM Inference Through Overhead Decomposition (viability: 3): https://sciencetostartup.com/paper/taxbreak-unmasking-the-hidden-costs-of-llm-inference-through-overhead-decomposition - TaxBreak is a diagnostic tool that decomposes LLM inference overhead to optimize performance.
- Predictive and adaptive maps for long-term visual navigation in changing environments (viability: 2): https://sciencetostartup.com/paper/predictive-and-adaptive-maps-for-long-term-visual-navigation-in-changing-environments - This paper explores map management techniques for long-term visual navigation in dynamic environments.
- Bases of Steerable Kernels for Equivariant CNNs: From 2D Rotations to the Lorentz Group (viability: 2): https://sciencetostartup.com/paper/bases-of-steerable-kernels-for-equivariant-cnns-from-2d-rotations-to-the-lorentz-group - A theoretical exploration of steerable kernels for equivariant convolutional neural networks.
- Shattering the Shortcut: A Topology-Regularized Benchmark for Multi-hop Medical Reasoning in LLMs (viability: 9): https://sciencetostartup.com/paper/shattering-the-shortcut-a-topology-regularized-benchmark-for-multi-hop-medical-reasoning-in-llms - ShatterMed-QA is a benchmark designed to enhance multi-hop diagnostic reasoning in LLMs by addressing shortcut learning in medical AI.
- Operationalising Cyber Risk Management Using AI: Connecting Cyber Incidents to MITRE ATT&CK Techniques, Security Controls, and Metrics (viability: 9): https://sciencetostartup.com/paper/operationalising-cyber-risk-management-using-ai-connecting-cyber-incidents-to-mitre-att-ck-techniques-security-controls- - An AI-driven framework that automates the mapping of cyber incidents to security controls, enhancing threat intelligence and operational security management.
- CSE-UOI at SemEval-2026 Task 6: A Two-Stage Heterogeneous Ensemble with Deliberative Complexity Gating for Political Evasion Detection (viability: 2): https://sciencetostartup.com/paper/cse-uoi-at-semeval-2026-task-6-a-two-stage-heterogeneous-ensemble-with-deliberative-complexity-gating-for-political-evas - A dual LLM ensemble for classifying political interview responses into clarity categories.
- Overcoming the Modality Gap in Context-Aided Forecasting (viability: 7): https://sciencetostartup.com/paper/overcoming-the-modality-gap-in-context-aided-forecasting - A novel semi-synthetic data augmentation method enhances context-aided forecasting by generating high-quality datasets.
- Bridging the Gap Between Security Metrics and Key Risk Indicators: An Empirical Framework for Vulnerability Prioritization (viability: 5): https://sciencetostartup.com/paper/bridging-the-gap-between-security-metrics-and-key-risk-indicators-an-empirical-framework-for-vulnerability-prioritizatio - A framework for prioritizing vulnerability remediation based on key risk indicators to enhance organizational security.
- KernelFoundry: Hardware-aware evolutionary GPU kernel optimization (viability: 7): https://sciencetostartup.com/paper/kernelfoundry-hardware-aware-evolutionary-gpu-kernel-optimization - KernelFoundry is an evolutionary framework for optimizing GPU kernels tailored for large language models.
- DiscoRD: An Experimental Methodology for Quickly Discovering the Reliable Read Disturbance Threshold of Real DRAM Chips (viability: 4): https://sciencetostartup.com/paper/discord-an-experimental-methodology-for-quickly-discovering-the-reliable-read-disturbance-threshold-of-real-dram-chips - DiscoRD is a methodology for rapidly determining the reliable read disturbance threshold of DRAM chips to enhance system security and efficiency.
- Revisiting Model Stitching In the Foundation Model Era (viability: 6): https://sciencetostartup.com/paper/revisiting-model-stitching-in-the-foundation-model-era - A systematic approach to model stitching that enhances the integration of Vision Foundation Models for improved performance.
- Surg-R1: A Hierarchical Reasoning Foundation Model for Scalable and Interpretable Surgical Decision Support with Multi-Center Clinical Validation (viability: 9): https://sciencetostartup.com/paper/surg-r1-a-hierarchical-reasoning-foundation-model-for-scalable-and-interpretable-surgical-decision-support-with-multi-ce - Surg-R1 is a hierarchical reasoning foundation model designed to enhance surgical decision support through interpretable predictions and clinical validation.
- Interpreting Negation in GPT-2: Layer- and Head-Level Causal Analysis (viability: 4): https://sciencetostartup.com/paper/interpreting-negation-in-gpt-2-layer-and-head-level-causal-analysis - A causal analysis of GPT-2's handling of negation, revealing critical insights into its internal representations.
- A Neuro-Symbolic Framework Combining Inductive and Deductive Reasoning for Autonomous Driving Planning (viability: 8): https://sciencetostartup.com/paper/a-neuro-symbolic-framework-combining-inductive-and-deductive-reasoning-for-autonomous-driving-planning - A neuro-symbolic framework for safe and interpretable trajectory planning in autonomous driving.
- SpectralGuard: Detecting Memory Collapse Attacks in State Space Models (viability: 7): https://sciencetostartup.com/paper/spectralguard-detecting-memory-collapse-attacks-in-state-space-models - SpectralGuard provides a real-time monitoring solution to detect memory collapse attacks in state space models.
- ABRA: Teleporting Fine-Tuned Knowledge Across Domains for Open-Vocabulary Object Detection (viability: 7): https://sciencetostartup.com/paper/abra-teleporting-fine-tuned-knowledge-across-domains-for-open-vocabulary-object-detection - ABRA enables effective knowledge transfer for object detection across domains with limited data.
- Beyond Motion Imitation: Is Human Motion Data Alone Sufficient to Explain Gait Control and Biomechanics? (viability: 2): https://sciencetostartup.com/paper/beyond-motion-imitation-is-human-motion-data-alone-sufficient-to-explain-gait-control-and-biomechanics - This study explores the limitations of motion imitation learning in accurately modeling human gait biomechanics.
- Push, Press, Slide: Mode-Aware Planar Contact Manipulation via Reduced-Order Models (viability: 4): https://sciencetostartup.com/paper/push-press-slide-mode-aware-planar-contact-manipulation-via-reduced-order-models - A mode-aware framework for efficient non-prehensile planar manipulation in robotics.
- Not Just the Destination, But the Journey: Reasoning Traces Causally Shape Generalization Behaviors (viability: 4): https://sciencetostartup.com/paper/not-just-the-destination-but-the-journey-reasoning-traces-causally-shape-generalization-behaviors - Exploring how reasoning traces influence LLM generalization behaviors to improve alignment strategies.
- Test-Time Strategies for More Efficient and Accurate Agentic RAG (viability: 7): https://sciencetostartup.com/paper/test-time-strategies-for-more-efficient-and-accurate-agentic-rag - Enhancing Retrieval-Augmented Generation systems for improved accuracy and efficiency through innovative test-time strategies.
- Deployment-Oriented Session-wise Meta-Calibration for Landmark-Based Webcam Gaze Tracking (viability: 3): https://sciencetostartup.com/paper/deployment-oriented-session-wise-meta-calibration-for-landmark-based-webcam-gaze-tracking - EMC-Gaze offers a lightweight solution for webcam gaze tracking with minimal calibration requirements.
- SPARROW: Learning Spatial Precision and Temporal Referential Consistency in Pixel-Grounded Video MLLMs (viability: 8): https://sciencetostartup.com/paper/sparrow-learning-spatial-precision-and-temporal-referential-consistency-in-pixel-grounded-video-mllms - SPARROW enhances video understanding by integrating spatial precision and temporal consistency in pixel-grounded multimodal models.
- NeuroLoRA: Context-Aware Neuromodulation for Parameter-Efficient Multi-Task Adaptation (viability: 7): https://sciencetostartup.com/paper/neurolora-context-aware-neuromodulation-for-parameter-efficient-multi-task-adaptation - NeuroLoRA enhances parameter-efficient fine-tuning of LLMs through context-aware neuromodulation for improved multi-task adaptation.
- Efficient Reasoning with Balanced Thinking (viability: 8): https://sciencetostartup.com/paper/efficient-reasoning-with-balanced-thinking - ReBalance is a training-free framework that enhances reasoning efficiency in Large Reasoning Models by balancing overthinking and underthinking.
- Human Knowledge Integrated Multi-modal Learning for Single Source Domain Generalization (viability: 5): https://sciencetostartup.com/paper/human-knowledge-integrated-multi-modal-learning-for-single-source-domain-generalization - GenEval enhances single-source domain generalization in medical imaging using multimodal Vision Language Models.
- Multi-Step Semantic Reasoning in Generative Retrieval (viability: 7): https://sciencetostartup.com/paper/multi-step-semantic-reasoning-in-generative-retrieval - ReasonGR enhances generative retrieval models by improving multi-step semantic reasoning for complex numerical queries.
- Sinkhorn-Drifting Generative Models (viability: 7): https://sciencetostartup.com/paper/sinkhorn-drifting-generative-models - A novel generative model leveraging Sinkhorn divergence for improved stability and quality in generative tasks.
- GNN-DIP: Neural Corridor Selection for Decomposition-Based Motion Planning (viability: 7): https://sciencetostartup.com/paper/gnn-dip-neural-corridor-selection-for-decomposition-based-motion-planning - GNN-DIP leverages Graph Neural Networks for efficient corridor selection in motion planning, enhancing pathfinding in complex environments.
- Alternating Gradient Flow Utility: A Unified Metric for Structural Pruning and Dynamic Routing in Deep Networks (viability: 6): https://sciencetostartup.com/paper/alternating-gradient-flow-utility-a-unified-metric-for-structural-pruning-and-dynamic-routing-in-deep-networks - A novel hybrid routing framework that enhances structural pruning and dynamic routing in deep networks for improved efficiency.
- Spatial PDE-aware Selective State-space with Nested Memory for Mobile Traffic Grid Forecasting (viability: 6): https://sciencetostartup.com/paper/spatial-pde-aware-selective-state-space-with-nested-memory-for-mobile-traffic-grid-forecasting - NeST-S6 is a novel convolutional selective state-space model for efficient mobile traffic grid forecasting that significantly reduces prediction errors and computational overhead.
- Probabilistic Joint and Individual Variation Explained (ProJIVE) for Data Integration (viability: 7): https://sciencetostartup.com/paper/probabilistic-joint-and-individual-variation-explained-projive-for-data-integration - ProJIVE offers a probabilistic approach to integrate multiple data types for enhanced analysis in scientific research.
- TASTE-Streaming: Towards Streamable Text-Aligned Speech Tokenization and Embedding for Spoken Language Modeling (viability: 3): https://sciencetostartup.com/paper/taste-streaming-towards-streamable-text-aligned-speech-tokenization-and-embedding-for-spoken-language-modeling - TASTE-S is a streamable extension of text-aligned speech tokenization for real-time spoken language modeling.
- Budget-Sensitive Discovery Scoring: A Formally Verified Framework for Evaluating AI-Guided Scientific Selection (viability: 4): https://sciencetostartup.com/paper/budget-sensitive-discovery-scoring-a-formally-verified-framework-for-evaluating-ai-guided-scientific-selection - A formally verified framework for budget-aware evaluation of AI-guided scientific selection strategies.
- A Learning-Based Approach for Contact Detection, Localization, and Force Estimation of Continuum Manipulators With Integrated OFDR Optical Fiber (viability: 7): https://sciencetostartup.com/paper/a-learning-based-approach-for-contact-detection-localization-and-force-estimation-of-continuum-manipulators-with-integra - A learning-based framework for contact detection and force estimation in continuum manipulators using integrated optical fiber sensing.
- Generalist Large Language Models for Molecular Property Prediction: Distilling Knowledge from Specialist Models (viability: 7): https://sciencetostartup.com/paper/generalist-large-language-models-for-molecular-property-prediction-distilling-knowledge-from-specialist-models - TreeKD enhances Large Language Models for molecular property prediction by distilling knowledge from specialist decision trees.
- LLM-Augmented Therapy Normalization and Aspect-Based Sentiment Analysis for Treatment-Resistant Depression on Reddit (viability: 3): https://sciencetostartup.com/paper/llm-augmented-therapy-normalization-and-aspect-based-sentiment-analysis-for-treatment-resistant-depression-on-reddit - A study analyzing patient sentiment towards medications for treatment-resistant depression using Reddit data.
- Maximum Entropy Exploration Without the Rollouts (viability: 5): https://sciencetostartup.com/paper/maximum-entropy-exploration-without-the-rollouts - A novel algorithm for efficient exploration in reinforcement learning that avoids costly rollouts.
- Thermodynamics of Reinforcement Learning Curricula (viability: 2): https://sciencetostartup.com/paper/thermodynamics-of-reinforcement-learning-curricula - This paper proposes a geometric framework for optimizing curriculum learning in reinforcement learning using thermodynamic principles.
- VQQA: An Agentic Approach for Video Evaluation and Quality Improvement (viability: 8): https://sciencetostartup.com/paper/vqqa-an-agentic-approach-for-video-evaluation-and-quality-improvement - VQQA is a multi-agent framework that enhances video generation quality through dynamic visual question generation and actionable feedback.
- EVATok: Adaptive Length Video Tokenization for Efficient Visual Autoregressive Generation (viability: 8): https://sciencetostartup.com/paper/evatok-adaptive-length-video-tokenization-for-efficient-visual-autoregressive-generation - EVATok is an adaptive video tokenization framework that optimizes token assignments for efficient autoregressive video generation.
- MM-CondChain: A Programmatically Verified Benchmark for Visually Grounded Deep Compositional Reasoning (viability: 4): https://sciencetostartup.com/paper/mm-condchain-a-programmatically-verified-benchmark-for-visually-grounded-deep-compositional-reasoning - MM-CondChain is a benchmark for evaluating visually grounded deep compositional reasoning in multimodal large language models.
- OmniStream: Mastering Perception, Reconstruction and Action in Continuous Streams (viability: 7): https://sciencetostartup.com/paper/omnistream-mastering-perception-reconstruction-and-action-in-continuous-streams - OmniStream is a unified streaming visual backbone that enhances real-time perception, reconstruction, and action from diverse visual inputs.
- GRADE: Benchmarking Discipline-Informed Reasoning in Image Editing (viability: 7): https://sciencetostartup.com/paper/grade-benchmarking-discipline-informed-reasoning-in-image-editing - GRADE is a benchmark for assessing discipline-informed reasoning in image editing across various academic domains.
- $Ψ_0$: An Open Foundation Model Towards Universal Humanoid Loco-Manipulation (viability: 8): https://sciencetostartup.com/paper/0-an-open-foundation-model-towards-universal-humanoid-loco-manipulation - Psi-Zero open sources a superior foundation model for humanoid robot loco-manipulation tasks with state-of-the-art performance using efficient training data.
- Video Streaming Thinking: VideoLLMs Can Watch and Think Simultaneously (viability: 8): https://sciencetostartup.com/paper/video-streaming-thinking-videollms-can-watch-and-think-simultaneously - VST revolutionizes real-time video understanding by enabling VideoLLMs to process and reason about video content during streaming, improving interaction efficiency and accuracy.
- The Latent Color Subspace: Emergent Order in High-Dimensional Chaos (viability: 8): https://sciencetostartup.com/paper/the-latent-color-subspace-emergent-order-in-high-dimensional-chaos - A training-free method for fine-grained control of color in text-to-image generation using latent space manipulation.
- HumDex:Humanoid Dexterous Manipulation Made Easy (viability: 8): https://sciencetostartup.com/paper/humdex-humanoid-dexterous-manipulation-made-easy - HumDex is a portable teleoperation system that simplifies humanoid dexterous manipulation through advanced motion tracking and imitation learning.
- DreamVideo-Omni: Omni-Motion Controlled Multi-Subject Video Customization with Latent Identity Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/dreamvideo-omni-omni-motion-controlled-multi-subject-video-customization-with-latent-identity-reinforcement-learning - DreamVideo-Omni enables precise control over multi-subject video customization with advanced motion control techniques.
- Spatial-TTT: Streaming Visual-based Spatial Intelligence with Test-Time Training (viability: 8): https://sciencetostartup.com/paper/spatial-ttt-streaming-visual-based-spatial-intelligence-with-test-time-training - Spatial-TTT enhances spatial intelligence by streaming visual data and adapting parameters in real-time.
- Attend Before Attention: Efficient and Scalable Video Understanding via Autoregressive Gazing (viability: 7): https://sciencetostartup.com/paper/attend-before-attention-efficient-and-scalable-video-understanding-via-autoregressive-gazing - AutoGaze accelerates video processing by selectively attending to critical patches, enabling scalable and efficient video analysis for high-resolution content.
- EndoCoT: Scaling Endogenous Chain-of-Thought Reasoning in Diffusion Models (viability: 7): https://sciencetostartup.com/paper/endocot-scaling-endogenous-chain-of-thought-reasoning-in-diffusion-models - EndoCoT enhances reasoning in diffusion models by refining latent thought states for complex task execution.
- DVD: Deterministic Video Depth Estimation with Generative Priors (viability: 8): https://sciencetostartup.com/paper/dvd-deterministic-video-depth-estimation-with-generative-priors - DVD is a state-of-the-art deterministic video depth estimation tool leveraging generative priors for 3D scene understanding.
- SciMDR: Benchmarking and Advancing Scientific Multimodal Document Reasoning (viability: 6): https://sciencetostartup.com/paper/scimdr-benchmarking-and-advancing-scientific-multimodal-document-reasoning - SciMDR offers a novel framework for generating and evaluating scientific multimodal document reasoning datasets.
- Matching Features, Not Tokens: Energy-Based Fine-Tuning of Language Models (viability: 2): https://sciencetostartup.com/paper/matching-features-not-tokens-energy-based-fine-tuning-of-language-models - Introducing energy-based fine-tuning for language models to enhance sequence-level performance.
- Trust Your Critic: Robust Reward Modeling and Reinforcement Learning for Faithful Image Editing and Generation (viability: 7): https://sciencetostartup.com/paper/trust-your-critic-robust-reward-modeling-and-reinforcement-learning-for-faithful-image-editing-and-generation - FIRM enhances reinforcement learning in image editing and generation with robust reward models achieving state-of-the-art fidelity and instruction adherence.
- Examining Reasoning LLMs-as-Judges in Non-Verifiable LLM Post-Training (viability: 5): https://sciencetostartup.com/paper/examining-reasoning-llms-as-judges-in-non-verifiable-llm-post-training - A study on the effectiveness of reasoning LLMs as judges in reinforcement learning for LLM alignment.
- One Model, Many Budgets: Elastic Latent Interfaces for Diffusion Transformers (viability: 8): https://sciencetostartup.com/paper/one-model-many-budgets-elastic-latent-interfaces-for-diffusion-transformers - ELIT enhances diffusion transformers by optimizing compute allocation through a dynamic latent interface.
- Separable neural architectures as a primitive for unified predictive and generative intelligence (viability: 4): https://sciencetostartup.com/paper/separable-neural-architectures-as-a-primitive-for-unified-predictive-and-generative-intelligence - A novel separable neural architecture that unifies predictive and generative intelligence across multiple domains.
- HandelBot: Real-World Piano Playing via Fast Adaptation of Dexterous Robot Policies (viability: 7): https://sciencetostartup.com/paper/handelbot-real-world-piano-playing-via-fast-adaptation-of-dexterous-robot-policies - HandelBot enables precise bimanual piano playing through rapid adaptation of robot policies.
- BiGain: Unified Token Compression for Joint Generation and Classification (viability: 7): https://sciencetostartup.com/paper/bigain-unified-token-compression-for-joint-generation-and-classification - BiGain is a training-free framework that enhances both generation quality and classification accuracy in accelerated diffusion models through innovative token compression techniques.
- STAMP: Selective Task-Aware Mechanism for Text Privacy (viability: 7): https://sciencetostartup.com/paper/stamp-selective-task-aware-mechanism-for-text-privacy - STAMP is a framework that enhances text privacy while maintaining task relevance through selective token-level noise allocation.
- SceneAssistant: A Visual Feedback Agent for Open-Vocabulary 3D Scene Generation (viability: 8): https://sciencetostartup.com/paper/sceneassistant-a-visual-feedback-agent-for-open-vocabulary-3d-scene-generation - SceneAssistant transforms text commands into high-quality 3D scenes with minimal user input.
- Incremental Neural Network Verification via Learned Conflicts (viability: 3): https://sciencetostartup.com/paper/incremental-neural-network-verification-via-learned-conflicts - A technique to expedite neural network verification by reusing learned conflicts across related queries.
- Temporal Straightening for Latent Planning (viability: 4): https://sciencetostartup.com/paper/temporal-straightening-for-latent-planning - A novel approach to improve latent planning through temporal straightening of representations.
- Security Considerations for Artificial Intelligence Agents (viability: 4): https://sciencetostartup.com/paper/security-considerations-for-artificial-intelligence-agents - A comprehensive analysis of security vulnerabilities in AI agents with actionable recommendations.
- Neural Thickets: Diverse Task Experts Are Dense Around Pretrained Weights (viability: 5): https://sciencetostartup.com/paper/neural-thickets-diverse-task-experts-are-dense-around-pretrained-weights - A novel method for discovering task-specific experts in pretrained models through structured optimization.
- Interpreting Contrastive Embeddings in Specific Domains with Fuzzy Rules (viability: 2): https://sciencetostartup.com/paper/interpreting-contrastive-embeddings-in-specific-domains-with-fuzzy-rules - This paper explores using fuzzy rules to enhance CLIP embeddings for specific domains like clinical reports and film reviews.
- Sparking Scientific Creativity via LLM-Driven Interdisciplinary Inspiration (viability: 7): https://sciencetostartup.com/paper/sparking-scientific-creativity-via-llm-driven-interdisciplinary-inspiration - Idea-Catalyst enhances interdisciplinary research by augmenting creative reasoning in humans and LLMs.
- Portfolio of Solving Strategies in CEGAR-based Object Packing and Scheduling for Sequential 3D Printing (viability: 3): https://sciencetostartup.com/paper/portfolio-of-solving-strategies-in-cegar-based-object-packing-and-scheduling-for-sequential-3d-printing - A parallelized algorithm for optimizing object arrangement and scheduling in sequential 3D printing.
- HiAP: A Multi-Granular Stochastic Auto-Pruning Framework for Vision Transformers (viability: 6): https://sciencetostartup.com/paper/hiap-a-multi-granular-stochastic-auto-pruning-framework-for-vision-transformers - HiAP is an innovative framework that optimizes Vision Transformers for efficient deployment on edge devices through multi-granular stochastic pruning.
- A Two-Stage Dual-Modality Model for Facial Emotional Expression Recognition (viability: 7): https://sciencetostartup.com/paper/a-two-stage-dual-modality-model-for-facial-emotional-expression-recognition - A dual-modal model for accurate facial emotional expression recognition from videos.
- Real-World Point Tracking with Verifier-Guided Pseudo-Labeling (viability: 7): https://sciencetostartup.com/paper/real-world-point-tracking-with-verifier-guided-pseudo-labeling - A meta-model that enhances point tracking in real-world videos by improving pseudo-label quality through reliability assessment.
- RDNet: Region Proportion-Aware Dynamic Adaptive Salient Object Detection Network in Optical Remote Sensing Images (viability: 3): https://sciencetostartup.com/paper/rdnet-region-proportion-aware-dynamic-adaptive-salient-object-detection-network-in-optical-remote-sensing-images - RDNet enhances salient object detection in remote sensing images by addressing scale variations and localization challenges.
- WORKSWORLD: A Domain for Integrated Numeric Planning and Scheduling of Distributed Pipelined Workflows (viability: 3): https://sciencetostartup.com/paper/worksworld-a-domain-for-integrated-numeric-planning-and-scheduling-of-distributed-pipelined-workflows - WORKSWORLD automates the planning and scheduling of distributed data pipelines through a novel graph representation.
- ForensicZip: More Tokens are Better but Not Necessary in Forensic Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/forensiczip-more-tokens-are-better-but-not-necessary-in-forensic-vision-language-models - ForensicZip accelerates multimedia forensics by optimizing token compression for forgery detection in visual data.
- CLASP: Defending Hybrid Large Language Models Against Hidden State Poisoning Attacks (viability: 6): https://sciencetostartup.com/paper/clasp-defending-hybrid-large-language-models-against-hidden-state-poisoning-attacks - CLASP offers real-time defense against hidden state poisoning attacks in language models, ensuring secure document processing workflows.
- IndexCache: Accelerating Sparse Attention via Cross-Layer Index Reuse (viability: 3): https://sciencetostartup.com/paper/indexcache-accelerating-sparse-attention-via-cross-layer-index-reuse - IndexCache optimizes sparse attention in large language models by reusing indices across layers to enhance efficiency.
- Privacy in ERP Systems: Behavioral Models of Developers and Consultants (viability: 2): https://sciencetostartup.com/paper/privacy-in-erp-systems-behavioral-models-of-developers-and-consultants - This research explores privacy awareness in ERP systems among developers and consultants to enhance compliance with regulations like GDPR.
- SaPaVe: Towards Active Perception and Manipulation in Vision-Language-Action Models for Robotics (viability: 8): https://sciencetostartup.com/paper/sapave-towards-active-perception-and-manipulation-in-vision-language-action-models-for-robotics - SaPaVe is an end-to-end framework that enhances robotic interaction through unified active perception and manipulation.
- Long-Context Encoder Models for Polish Language Understanding (viability: 7): https://sciencetostartup.com/paper/long-context-encoder-models-for-polish-language-understanding - A high-quality Polish language model designed for long-document understanding, outperforming existing solutions.
- Compiling Temporal Numeric Planning into Discrete PDDL+: Extended Version (viability: 2): https://sciencetostartup.com/paper/compiling-temporal-numeric-planning-into-discrete-pddl-extended-version - A practical compilation method for temporal planning into PDDL+.
- ComFree-Sim: A GPU-Parallelized Analytical Contact Physics Engine for Scalable Contact-Rich Robotics Simulation and Control (viability: 8): https://sciencetostartup.com/paper/comfree-sim-a-gpu-parallelized-analytical-contact-physics-engine-for-scalable-contact-rich-robotics-simulation-and-contr - ComFree-Sim is a GPU-parallelized contact physics engine that enhances robotics simulation and control with near-linear scaling and high throughput.
- Strategic Navigation or Stochastic Search? How Agents and Humans Reason Over Document Collections (viability: 6): https://sciencetostartup.com/paper/strategic-navigation-or-stochastic-search-how-agents-and-humans-reason-over-document-collections - MADQA benchmarks agentic reasoning in document workflows, revealing gaps in strategic planning versus human performance.
- BehaviorVLM: Unified Finetuning-Free Behavioral Understanding with Vision-Language Reasoning (viability: 3): https://sciencetostartup.com/paper/behaviorvlm-unified-finetuning-free-behavioral-understanding-with-vision-language-reasoning - BehaviorVLM offers a novel vision-language framework for scalable animal behavior analysis without extensive human labeling.
- LatentGeo: Learnable Auxiliary Constructions in Latent Space for Multimodal Geometric Reasoning (viability: 7): https://sciencetostartup.com/paper/latentgeo-learnable-auxiliary-constructions-in-latent-space-for-multimodal-geometric-reasoning - LatentGeo revolutionizes multimodal geometric reasoning by learning continuous latent visual representations for auxiliary constructions.
- QAQ: Bidirectional Semantic Coherence for Selecting High-Quality Synthetic Code Instructions (viability: 7): https://sciencetostartup.com/paper/qaq-bidirectional-semantic-coherence-for-selecting-high-quality-synthetic-code-instructions - QAQ is a novel framework for selecting high-quality synthetic code instructions by evaluating bidirectional semantic coherence.
- A Quantitative Characterization of Forgetting in Post-Training (viability: 2): https://sciencetostartup.com/paper/a-quantitative-characterization-of-forgetting-in-post-training - This paper provides a theoretical framework for understanding forgetting in generative models during continual post-training.
- GlyphBanana: Advancing Precise Text Rendering Through Agentic Workflows (viability: 8): https://sciencetostartup.com/paper/glyphbanana-advancing-precise-text-rendering-through-agentic-workflows - GlyphBanana enhances text rendering precision through innovative agentic workflows and a dedicated benchmark.
- LifeSim: Long-Horizon User Life Simulator for Personalized Assistant Evaluation (viability: 4): https://sciencetostartup.com/paper/lifesim-long-horizon-user-life-simulator-for-personalized-assistant-evaluation - LifeSim is a user simulator designed to enhance personalized assistant evaluations by modeling user cognition and intentions.
- IsoCompute Playbook: Optimally Scaling Sampling Compute for LLM RL (viability: 2): https://sciencetostartup.com/paper/isocompute-playbook-optimally-scaling-sampling-compute-for-llm-rl - This paper provides insights into compute allocation for RL post-training of LLMs but lacks practical implementation details.
- Linking Perception, Confidence and Accuracy in MLLMs (viability: 7): https://sciencetostartup.com/paper/linking-perception-confidence-and-accuracy-in-mllms - A framework that enhances the confidence calibration of Multi-modal Large Language Models through innovative reinforcement learning techniques.
- EgoIntent: An Egocentric Step-level Benchmark for Understanding What, Why, and Next (viability: 4): https://sciencetostartup.com/paper/egointent-an-egocentric-step-level-benchmark-for-understanding-what-why-and-next - EgoIntent is a benchmark for fine-grained step-level intent understanding in egocentric videos, crucial for applications like intelligent assistants and augmented reality.
- FlashMotion: Few-Step Controllable Video Generation with Trajectory Guidance (viability: 7): https://sciencetostartup.com/paper/flashmotion-few-step-controllable-video-generation-with-trajectory-guidance - FlashMotion enables rapid, high-quality video generation with precise trajectory control using a few-step approach.
- Automatic Generation of High-Performance RL Environments (viability: 8): https://sciencetostartup.com/paper/automatic-generation-of-high-performance-rl-environments - A framework for automatically generating high-performance reinforcement learning environments with minimal engineering effort.
- O3N: Omnidirectional Open-Vocabulary Occupancy Prediction (viability: 8): https://sciencetostartup.com/paper/o3n-omnidirectional-open-vocabulary-occupancy-prediction - O3N is an omnidirectional occupancy prediction framework that enhances 3D perception for autonomous agents through advanced spatial representation.
- Understanding Disclosure Risk in Differential Privacy with Applications to Noise Calibration and Auditing (Extended Version) (viability: 4): https://sciencetostartup.com/paper/understanding-disclosure-risk-in-differential-privacy-with-applications-to-noise-calibration-and-auditing-extended-versi - A novel risk metric for enhancing differential privacy calibration and auditing.
- HATS: Hardness-Aware Trajectory Synthesis for GUI Agents (viability: 7): https://sciencetostartup.com/paper/hats-hardness-aware-trajectory-synthesis-for-gui-agents - HATS is a framework that enhances GUI agents' performance by synthesizing high-quality trajectory data to address semantic ambiguities.
- TopoBench: Benchmarking LLMs on Hard Topological Reasoning (viability: 8): https://sciencetostartup.com/paper/topobench-benchmarking-llms-on-hard-topological-reasoning - TopoBench is a benchmark for evaluating LLMs on challenging topological reasoning tasks.
- Increasing intelligence in AI agents can worsen collective outcomes (viability: 2): https://sciencetostartup.com/paper/increasing-intelligence-in-ai-agents-can-worsen-collective-outcomes - This research explores the chaotic dynamics of AI agents competing for limited resources.
- Hoi3DGen: Generating High-Quality Human-Object-Interactions in 3D (viability: 3): https://sciencetostartup.com/paper/hoi3dgen-generating-high-quality-human-object-interactions-in-3d - Hoi3DGen generates high-quality 3D human-object interactions from text for AR, XR, and gaming applications.
- Cross-Context Review: Improving LLM Output Quality by Separating Production and Review Sessions (viability: 3): https://sciencetostartup.com/paper/cross-context-review-improving-llm-output-quality-by-separating-production-and-review-sessions - Cross-Context Review improves LLM output quality by separating production and review sessions.
- CRAFT: A Tendon-Driven Hand with Hybrid Hard-Soft Compliance (viability: 8): https://sciencetostartup.com/paper/craft-a-tendon-driven-hand-with-hybrid-hard-soft-compliance - CRAFT is an open-source tendon-driven anthropomorphic hand designed for efficient contact-rich manipulation.
- Cornserve: A Distributed Serving System for Any-to-Any Multimodal Models (viability: 7): https://sciencetostartup.com/paper/cornserve-a-distributed-serving-system-for-any-to-any-multimodal-models - Cornserve is an open-source distributed serving system designed for Any-to-Any multimodal models, enhancing throughput and reducing latency.
- SommBench: Assessing Sommelier Expertise of Language Models (viability: 8): https://sciencetostartup.com/paper/sommbench-assessing-sommelier-expertise-of-language-models - SommBench is a multilingual benchmark for assessing sommelier expertise in language models.
- Taming the Adversary: Stable Minimax Deep Deterministic Policy Gradient via Fractional Objectives (viability: 6): https://sciencetostartup.com/paper/taming-the-adversary-stable-minimax-deep-deterministic-policy-gradient-via-fractional-objectives - A framework for learning robust reinforcement learning policies against external disturbances.
- On Information Self-Locking in Reinforcement Learning for Active Reasoning of LLM agents (viability: 7): https://sciencetostartup.com/paper/on-information-self-locking-in-reinforcement-learning-for-active-reasoning-of-llm-agents - A novel reinforcement learning approach to enhance LLM agents' active reasoning by mitigating information self-locking.
- EvoTok: A Unified Image Tokenizer via Residual Latent Evolution for Visual Understanding and Generation (viability: 7): https://sciencetostartup.com/paper/evotok-a-unified-image-tokenizer-via-residual-latent-evolution-for-visual-understanding-and-generation - EvoTok is a unified image tokenizer that bridges the gap between visual understanding and generation through a novel residual evolution process.
- To Words and Beyond: Probing Large Language Models for Sentence-Level Psycholinguistic Norms of Memorability and Reading Times (viability: 4): https://sciencetostartup.com/paper/to-words-and-beyond-probing-large-language-models-for-sentence-level-psycholinguistic-norms-of-memorability-and-reading- - This research explores the use of large language models to estimate sentence-level psycholinguistic norms, enhancing our understanding of memorability and reading times.
- Wasserstein Gradient Flows for Batch Bayesian Optimal Experimental Design (viability: 4): https://sciencetostartup.com/paper/wasserstein-gradient-flows-for-batch-bayesian-optimal-experimental-design - A novel approach to batch Bayesian optimal experimental design that optimizes expected utility through Wasserstein gradient flows.
- Towards Dynamic Model Identification and Gravity Compensation for the dVRK-Si Patient Side Manipulator (viability: 7): https://sciencetostartup.com/paper/towards-dynamic-model-identification-and-gravity-compensation-for-the-dvrk-si-patient-side-manipulator - A dynamic modeling framework for the dVRK-Si Patient Side Manipulator that enhances control accuracy in robotic surgery.
- A Robust and Efficient Multi-Agent Reinforcement Learning Framework for Traffic Signal Control (viability: 7): https://sciencetostartup.com/paper/a-robust-and-efficient-multi-agent-reinforcement-learning-framework-for-traffic-signal-control - A robust multi-agent reinforcement learning framework that optimizes traffic signal control for dynamic traffic conditions.
- Human-Centred LLM Privacy Audits: Findings and Frictions (viability: 3): https://sciencetostartup.com/paper/human-centred-llm-privacy-audits-findings-and-frictions - A browser-based tool for self-auditing LLM privacy associations.
- Resource-Efficient Iterative LLM-Based NAS with Feedback Memory (viability: 6): https://sciencetostartup.com/paper/resource-efficient-iterative-llm-based-nas-with-feedback-memory - A resource-efficient pipeline for Neural Architecture Search using LLMs to optimize CNN designs on consumer-grade hardware.
- EmbTracker: Traceable Black-box Watermarking for Federated Language Models (viability: 7): https://sciencetostartup.com/paper/embtracker-traceable-black-box-watermarking-for-federated-language-models - EmbTracker provides a robust black-box watermarking solution for federated language models, ensuring client-level traceability without requiring client cooperation.
- Cross-Domain Policy Optimization via Bellman Consistency and Hybrid Critics (viability: 8): https://sciencetostartup.com/paper/cross-domain-policy-optimization-via-bellman-consistency-and-hybrid-critics - QAvatar enhances cross-domain reinforcement learning by effectively leveraging source-domain knowledge for improved transferability.
- Towards Universal Computational Aberration Correction in Photographic Cameras: A Comprehensive Benchmark Analysis (viability: 8): https://sciencetostartup.com/paper/towards-universal-computational-aberration-correction-in-photographic-cameras-a-comprehensive-benchmark-analysis - A universal framework for computational aberration correction in photography that generalizes across diverse lenses.
- Node-RF: Learning Generalized Continuous Space-Time Scene Dynamics with Neural ODE-based NeRFs (viability: 4): https://sciencetostartup.com/paper/node-rf-learning-generalized-continuous-space-time-scene-dynamics-with-neural-ode-based-nerfs - Node-RF leverages Neural ODEs and NeRFs for advanced scene dynamics prediction beyond observed boundaries.
- Decentralized Cooperative Localization for Multi-Robot Systems with Asynchronous Sensor Fusion (viability: 7): https://sciencetostartup.com/paper/decentralized-cooperative-localization-for-multi-robot-systems-with-asynchronous-sensor-fusion - A decentralized localization framework for multi-robot systems that enhances accuracy in GPS-denied environments.
- A Multi-Label Temporal Convolutional Framework for Transcription Factor Binding Characterization (viability: 2): https://sciencetostartup.com/paper/a-multi-label-temporal-convolutional-framework-for-transcription-factor-binding-characterization - A deep learning framework for predicting transcription factor binding sites using multi-label classification.
- Paper Title: LoV3D: Grounding Cognitive Prognosis Reasoning in Longitudinal 3D Brain MRI via Regional Volume Assessments (viability: 8): https://sciencetostartup.com/paper/paper-title-lov3d-grounding-cognitive-prognosis-reasoning-in-longitudinal-3d-brain-mri-via-regional-volume-assessments - Revolutionize dementia diagnosis with LoV3D, a verifiable AI model for interpreting longitudinal 3D brain MRI.
- Beyond Convolution: A Taxonomy of Structured Operators for Learning-Based Image Processing (viability: 2): https://sciencetostartup.com/paper/beyond-convolution-a-taxonomy-of-structured-operators-for-learning-based-image-processing - This paper proposes a taxonomy of structured operators to enhance learning-based image processing beyond traditional convolution.
- Dense Dynamic Scene Reconstruction and Camera Pose Estimation from Multi-View Videos (viability: 7): https://sciencetostartup.com/paper/dense-dynamic-scene-reconstruction-and-camera-pose-estimation-from-multi-view-videos - A novel framework for dense scene reconstruction and camera pose estimation using multiple freely moving cameras.
- NBAvatar: Neural Billboards Avatars with Realistic Hand-Face Interaction (viability: 7): https://sciencetostartup.com/paper/nbavatar-neural-billboards-avatars-with-realistic-hand-face-interaction - NBAvatar offers a novel method for realistic head avatars that enhance hand-face interaction through advanced neural rendering techniques.
- Systematic Security Analysis of the Iridium Satellite Radio Link (viability: 4): https://sciencetostartup.com/paper/systematic-security-analysis-of-the-iridium-satellite-radio-link - A comprehensive security analysis of Iridium satellite communication protocols revealing critical vulnerabilities.
- Chemical Reaction Networks Learn Better than Spiking Neural Networks (viability: 3): https://sciencetostartup.com/paper/chemical-reaction-networks-learn-better-than-spiking-neural-networks - This paper presents a mathematical proof that chemical reaction networks can outperform spiking neural networks in classification tasks.
- Flight through Narrow Gaps with Morphing-Wing Drones (viability: 3): https://sciencetostartup.com/paper/flight-through-narrow-gaps-with-morphing-wing-drones - A morphing-wing drone that can navigate narrow gaps by dynamically adjusting its wings during flight.
- Coarse-Guided Visual Generation via Weighted h-Transform Sampling (viability: 5): https://sciencetostartup.com/paper/coarse-guided-visual-generation-via-weighted-h-transform-sampling - A novel guided method for coarse-guided visual generation that enhances synthesis quality using h-transform sampling.
- XSkill: Continual Learning from Experience and Skills in Multimodal Agents (viability: 7): https://sciencetostartup.com/paper/xskill-continual-learning-from-experience-and-skills-in-multimodal-agents - XSkill is a dual-stream framework enabling multimodal agents to continually learn from experiences and skills without parameter updates.
- Continual Learning with Vision-Language Models via Semantic-Geometry Preservation (viability: 4): https://sciencetostartup.com/paper/continual-learning-with-vision-language-models-via-semantic-geometry-preservation - SeGP-CL enhances continual learning in vision-language models by preserving semantic geometry to prevent catastrophic forgetting.
- Translationese as a Rational Response to Translation Task Difficulty (viability: 2): https://sciencetostartup.com/paper/translationese-as-a-rational-response-to-translation-task-difficulty - This research explores the cognitive load of translation tasks and its impact on translation quality.
- Slow-Fast Inference: Training-Free Inference Acceleration via Within-Sentence Support Stability (viability: 3): https://sciencetostartup.com/paper/slow-fast-inference-training-free-inference-acceleration-via-within-sentence-support-stability - Slow-Fast Inference offers a training-free method to accelerate long-context autoregressive decoding.
- Frequentist Consistency of Prior-Data Fitted Networks for Causal Inference (viability: 2): https://sciencetostartup.com/paper/frequentist-consistency-of-prior-data-fitted-networks-for-causal-inference - This paper explores the frequentist consistency of prior-data fitted networks for causal inference, addressing uncertainty quantification issues.
- Single Pixel Image Classification using an Ultrafast Digital Light Projector (viability: 6): https://sciencetostartup.com/paper/single-pixel-image-classification-using-an-ultrafast-digital-light-projector - An ultrafast image classification system using single pixel imaging for real-time applications.
- AGMARL-DKS: An Adaptive Graph-Enhanced Multi-Agent Reinforcement Learning for Dynamic Kubernetes Scheduling (viability: 3): https://sciencetostartup.com/paper/agmarl-dks-an-adaptive-graph-enhanced-multi-agent-reinforcement-learning-for-dynamic-kubernetes-scheduling - AGMARL-DKS is an innovative multi-agent reinforcement learning approach for dynamic Kubernetes scheduling that enhances resource utilization and fault tolerance.
- Efficient Generative Modeling with Unitary Matrix Product States Using Riemannian Optimization (viability: 4): https://sciencetostartup.com/paper/efficient-generative-modeling-with-unitary-matrix-product-states-using-riemannian-optimization - A novel Riemannian optimization approach for efficient generative modeling using unitary matrix product states.
- Cascade: Composing Software-Hardware Attack Gadgets for Adversarial Threat Amplification in Compound AI Systems (viability: 4): https://sciencetostartup.com/paper/cascade-composing-software-hardware-attack-gadgets-for-adversarial-threat-amplification-in-compound-ai-systems - A framework for understanding and exploiting vulnerabilities in compound AI systems through novel attack strategies.
- Just Use XML: Revisiting Joint Translation and Label Projection (viability: 7): https://sciencetostartup.com/paper/just-use-xml-revisiting-joint-translation-and-label-projection - LabelPigeon enhances cross-lingual transfer by jointly performing translation and label projection using XML tags.
- Sim-to-reality adaptation for Deep Reinforcement Learning applied to an underwater docking application (viability: 7): https://sciencetostartup.com/paper/sim-to-reality-adaptation-for-deep-reinforcement-learning-applied-to-an-underwater-docking-application - A systematic approach for sim-to-reality adaptation in Deep Reinforcement Learning for autonomous underwater docking.
- An Intent of Collaboration: On Agencies between Designers and Emerging (Intelligent) Technologies (viability: 2): https://sciencetostartup.com/paper/an-intent-of-collaboration-on-agencies-between-designers-and-emerging-intelligent-technologies - Exploring the dynamics between designers and generative AIs to enhance creative processes.
- Nyxus: A Next Generation Image Feature Extraction Library for the Big Data and AI Era (viability: 7): https://sciencetostartup.com/paper/nyxus-a-next-generation-image-feature-extraction-library-for-the-big-data-and-ai-era - Nyxus is a scalable image feature extraction library designed for efficient processing of large datasets across various biomedical domains.
- Flowcean - Model Learning for Cyber-Physical Systems (viability: 3): https://sciencetostartup.com/paper/flowcean-model-learning-for-cyber-physical-systems - Flowcean automates the generation of models for Cyber-Physical Systems using data-driven learning.
- Pano360: Perspective to Panoramic Vision with Geometric Consistency (viability: 7): https://sciencetostartup.com/paper/pano360-perspective-to-panoramic-vision-with-geometric-consistency - Pano360 enhances panorama stitching by leveraging 3D geometric consistency for improved alignment and quality.
- Deep Learning-Based Metamodeling of Nonlinear Stochastic Dynamic Systems under Parametric and Predictive Uncertainty (viability: 2): https://sciencetostartup.com/paper/deep-learning-based-metamodeling-of-nonlinear-stochastic-dynamic-systems-under-parametric-and-predictive-uncertainty - This paper presents advanced metamodeling frameworks for predicting structural responses under uncertainty, but lacks a clear commercial application.
- Can RL Improve Generalization of LLM Agents? An Empirical Study (viability: 2): https://sciencetostartup.com/paper/can-rl-improve-generalization-of-llm-agents-an-empirical-study - This study explores the generalization capabilities of reinforcement fine-tuning in LLM agents across various environments.
- CrossEarth-SAR: A SAR-Centric and Billion-Scale Geospatial Foundation Model for Domain Generalizable Semantic Segmentation (viability: 8): https://sciencetostartup.com/paper/crossearth-sar-a-sar-centric-and-billion-scale-geospatial-foundation-model-for-domain-generalizable-semantic-segmentatio - SAR imaging for domain-generalizable semantic segmentation with billion-scale SAR foundation model.
- Decentralized Orchestration Architecture for Fluid Computing: A Secure Distributed AI Use Case (viability: 2): https://sciencetostartup.com/paper/decentralized-orchestration-architecture-for-fluid-computing-a-secure-distributed-ai-use-case - A decentralized orchestration architecture for fluid computing to enhance resource management in distributed AI applications.
- Credibility Matters: Motivations, Characteristics, and Influence Mechanisms of Crypto Key Opinion Leaders (viability: 2): https://sciencetostartup.com/paper/credibility-matters-motivations-characteristics-and-influence-mechanisms-of-crypto-key-opinion-leaders - This research explores the motivations and credibility mechanisms of crypto Key Opinion Leaders in shaping investment behaviors.
- Few-for-Many Personalized Federated Learning (viability: 8): https://sciencetostartup.com/paper/few-for-many-personalized-federated-learning - FedFew optimizes federated learning with minimal server models to personalize and scale efficiently.
- BTZSC: A Benchmark for Zero-Shot Text Classification Across Cross-Encoders, Embedding Models, Rerankers and LLMs (viability: 7): https://sciencetostartup.com/paper/btzsc-a-benchmark-for-zero-shot-text-classification-across-cross-encoders-embedding-models-rerankers-and-llms - BTZSC is a benchmark for evaluating zero-shot text classification across various model families, promoting reproducibility and fair comparisons.
- On-Average Stability of Multipass Preconditioned SGD and Effective Dimension (viability: 2): https://sciencetostartup.com/paper/on-average-stability-of-multipass-preconditioned-sgd-and-effective-dimension - This paper explores the theoretical aspects of multipass Preconditioned Stochastic Gradient Descent and its impact on generalization.
- LABSHIELD: A Multimodal Benchmark for Safety-Critical Reasoning and Planning in Scientific Laboratories (viability: 5): https://sciencetostartup.com/paper/labshield-a-multimodal-benchmark-for-safety-critical-reasoning-and-planning-in-scientific-laboratories - LABSHIELD is a multimodal benchmark for evaluating safety-critical reasoning in laboratory environments using MLLM agents.
- Ada3Drift: Adaptive Training-Time Drifting for One-Step 3D Visuomotor Robotic Manipulation (viability: 7): https://sciencetostartup.com/paper/ada3drift-adaptive-training-time-drifting-for-one-step-3d-visuomotor-robotic-manipulation - Ada3Drift enhances robotic manipulation by enabling high-fidelity single-step action generation through adaptive training-time drifting.
- Learning Visuomotor Policy for Multi-Robot Laser Tag Game (viability: 6): https://sciencetostartup.com/paper/learning-visuomotor-policy-for-multi-robot-laser-tag-game - An end-to-end visuomotor policy for multi-robot laser tag that improves accuracy and collision avoidance.
- HomeSafe-Bench: Evaluating Vision-Language Models on Unsafe Action Detection for Embodied Agents in Household Scenarios (viability: 7): https://sciencetostartup.com/paper/homesafe-bench-evaluating-vision-language-models-on-unsafe-action-detection-for-embodied-agents-in-household-scenarios - HomeSafe-Bench evaluates Vision-Language Models for unsafe action detection in household robots to enhance safety.
- Normative Common Ground Replication (NormCoRe): Replication-by-Translation for Studying Norms in Multi-agent AI (viability: 4): https://sciencetostartup.com/paper/normative-common-ground-replication-normcore-replication-by-translation-for-studying-norms-in-multi-agent-ai - NormCoRe is a framework for translating human experiment designs into multi-agent AI environments to study normative behaviors.
- Topological DeepONets and a generalization of the Chen-Chen operator approximation theorem (viability: 2): https://sciencetostartup.com/paper/topological-deeponets-and-a-generalization-of-the-chen-chen-operator-approximation-theorem - This paper explores a theoretical extension of DeepONets for approximating nonlinear operators in function spaces.
- Multimodal Emotion Recognition via Bi-directional Cross-Attention and Temporal Modeling (viability: 6): https://sciencetostartup.com/paper/multimodal-emotion-recognition-via-bi-directional-cross-attention-and-temporal-modeling - A multimodal framework for robust emotion recognition in video data using cross-attention and temporal modeling.
- Statistical and structural identifiability in representation learning (viability: 5): https://sciencetostartup.com/paper/statistical-and-structural-identifiability-in-representation-learning - This paper proposes a new framework for understanding and improving identifiability in representation learning models.
- AstroSplat: Physics-Based Gaussian Splatting for Rendering and Reconstruction of Small Celestial Bodies (viability: 7): https://sciencetostartup.com/paper/astrosplat-physics-based-gaussian-splatting-for-rendering-and-reconstruction-of-small-celestial-bodies - AstroSplat enhances surface reconstruction of small celestial bodies using physics-based Gaussian splatting.
- Uncovering Locally Low-dimensional Structure in Networks by Locally Optimal Spectral Embedding (viability: 2): https://sciencetostartup.com/paper/uncovering-locally-low-dimensional-structure-in-networks-by-locally-optimal-spectral-embedding - Introducing Local Adjacency Spectral Embedding (LASE) for improved local structure representation in networks.
- Energy Prediction on Sloping Ground for Quadruped Robots (viability: 3): https://sciencetostartup.com/paper/energy-prediction-on-sloping-ground-for-quadruped-robots - This paper presents an energy model for optimizing quadruped robot locomotion on sloped terrain.
- CHiL(L)Grader: Calibrated Human-in-the-Loop Short-Answer Grading (viability: 7): https://sciencetostartup.com/paper/chil-l-grader-calibrated-human-in-the-loop-short-answer-grading - CHiL(L)Grader is an AI-powered grading framework that enhances accuracy by integrating human feedback for uncertain predictions.
- PersonaTrace: Synthesizing Realistic Digital Footprints with LLM Agents (viability: 7): https://sciencetostartup.com/paper/personatrace-synthesizing-realistic-digital-footprints-with-llm-agents - PersonaTrace generates realistic digital footprints using LLM agents to enhance personalized applications and machine learning models.
- Preliminary analysis of RGB-NIR Image Registration techniques for off-road forestry environments (viability: 3): https://sciencetostartup.com/paper/preliminary-analysis-of-rgb-nir-image-registration-techniques-for-off-road-forestry-environments - Evaluating RGB-NIR image registration techniques for enhancing off-road forestry applications.
- Learning Transferable Sensor Models via Language-Informed Pretraining (viability: 7): https://sciencetostartup.com/paper/learning-transferable-sensor-models-via-language-informed-pretraining - SLIP is an open-source framework that learns language-aligned representations for diverse sensor setups, enhancing zero-shot transfer and generative reasoning.
- Delayed Backdoor Attacks: Exploring the Temporal Dimension as a New Attack Surface in Pre-Trained Models (viability: 5): https://sciencetostartup.com/paper/delayed-backdoor-attacks-exploring-the-temporal-dimension-as-a-new-attack-surface-in-pre-trained-models - Introducing Delayed Backdoor Attacks, a novel threat model for pre-trained models that exploits temporal dimensions for malicious activation.
- Resurfacing Paralinguistic Awareness in Large Audio Language Models (viability: 4): https://sciencetostartup.com/paper/resurfacing-paralinguistic-awareness-in-large-audio-language-models - A fine-tuning protocol to enhance paralinguistic awareness in large audio language models.
- Geometry-Aware Probabilistic Circuits via Voronoi Tessellations (viability: 2): https://sciencetostartup.com/paper/geometry-aware-probabilistic-circuits-via-voronoi-tessellations - This paper proposes a method to enhance probabilistic circuits using Voronoi tessellations for improved inference.
- Effective Resistance Rewiring: A Simple Topological Correction for Over-Squashing (viability: 3): https://sciencetostartup.com/paper/effective-resistance-rewiring-a-simple-topological-correction-for-over-squashing - Introducing Effective Resistance Rewiring to enhance long-range dependencies in Graph Neural Networks.
- Causal Matrix Completion under Multiple Treatments via Mixed Synthetic Nearest Neighbors (viability: 4): https://sciencetostartup.com/paper/causal-matrix-completion-under-multiple-treatments-via-mixed-synthetic-nearest-neighbors - A novel estimator for causal matrix completion that enhances data efficiency across treatment levels.
- Exhaustive Circuit Mapping of a Single-Cell Foundation Model Reveals Massive Redundancy, Heavy-Tailed Hub Architecture, and Layer-Dependent Differentiation Control (viability: 2): https://sciencetostartup.com/paper/exhaustive-circuit-mapping-of-a-single-cell-foundation-model-reveals-massive-redundancy-heavy-tailed-hub-architecture-an - This paper explores mechanistic interpretability in biological foundation models through advanced experimental techniques.
- Prototype-Based Knowledge Guidance for Fine-Grained Structured Radiology Reporting (viability: 7): https://sciencetostartup.com/paper/prototype-based-knowledge-guidance-for-fine-grained-structured-radiology-reporting - ProtoSR enhances structured radiology reporting by integrating free-text information for improved accuracy in fine-grained image analysis.
- Fair Learning for Bias Mitigation and Quality Optimization in Paper Recommendation (viability: 7): https://sciencetostartup.com/paper/fair-learning-for-bias-mitigation-and-quality-optimization-in-paper-recommendation - Fair-PaperRec optimizes paper acceptance decisions by mitigating demographic biases while enhancing quality.
- MobileKernelBench: Can LLMs Write Efficient Kernels for Mobile Devices? (viability: 7): https://sciencetostartup.com/paper/mobilekernelbench-can-llms-write-efficient-kernels-for-mobile-devices - MobileKernelBench enables LLMs to generate efficient kernels for mobile devices, significantly improving compilation success and performance.
- Chem4DLLM: 4D Multimodal LLMs for Chemical Dynamics Understanding (viability: 5): https://sciencetostartup.com/paper/chem4dllm-4d-multimodal-llms-for-chemical-dynamics-understanding - Chem4DLLM enables dynamic chemical understanding by translating 4D molecular trajectories into natural-language explanations.
- PicoSAM3: Real-Time In-Sensor Region-of-Interest Segmentation (viability: 8): https://sciencetostartup.com/paper/picosam3-real-time-in-sensor-region-of-interest-segmentation - PicoSAM3 is a lightweight, real-time visual segmentation model optimized for edge devices, enabling efficient on-device processing.
- CoMMET: To What Extent Can LLMs Perform Theory of Mind Tasks? (viability: 4): https://sciencetostartup.com/paper/commet-to-what-extent-can-llms-perform-theory-of-mind-tasks - CoMMET is a new multimodal benchmark dataset designed to evaluate Theory of Mind capabilities in Large Language Models.
- Understanding LLM Behavior When Encountering User-Supplied Harmful Content in Harmless Tasks (viability: 4): https://sciencetostartup.com/paper/understanding-llm-behavior-when-encountering-user-supplied-harmful-content-in-harmless-tasks - This study evaluates how LLMs handle harmful content in benign tasks, highlighting ethical vulnerabilities.
- InSpatio-WorldFM: An Open-Source Real-Time Generative Frame Model (viability: 7): https://sciencetostartup.com/paper/inspatio-worldfm-an-open-source-real-time-generative-frame-model - InSpatio-WorldFM is an open-source real-time frame model that enables low-latency spatial inference for world simulation.
- EnTransformer: A Deep Generative Transformer for Multivariate Probabilistic Forecasting (viability: 7): https://sciencetostartup.com/paper/entransformer-a-deep-generative-transformer-for-multivariate-probabilistic-forecasting - EnTransformer is a deep generative forecasting framework that enhances multivariate time series predictions with uncertainty quantification.
- Causal Representation Learning with Optimal Compression under Complex Treatments (viability: 5): https://sciencetostartup.com/paper/causal-representation-learning-with-optimal-compression-under-complex-treatments - A novel framework for estimating individual treatment effects with optimal balancing weights and scalability.
- FlexRec: Adapting LLM-based Recommenders for Flexible Needs via Reinforcement Learning (viability: 3): https://sciencetostartup.com/paper/flexrec-adapting-llm-based-recommenders-for-flexible-needs-via-reinforcement-learning - FlexRec enhances LLM-based recommenders by using reinforcement learning to adapt to dynamic user needs.
- Think While Watching: Online Streaming Segment-Level Memory for Multi-Turn Video Reasoning in Multimodal Large Language Models (viability: 9): https://sciencetostartup.com/paper/think-while-watching-online-streaming-segment-level-memory-for-multi-turn-video-reasoning-in-multimodal-large-language-m - A memory-anchored framework for real-time multi-turn video reasoning in multimodal large language models.
- Single-View Rolling-Shutter SfM (viability: 2): https://sciencetostartup.com/paper/single-view-rolling-shutter-sfm - A novel approach to improve structure-from-motion for rolling-shutter cameras using single-view geometry.
- Bielik-Minitron-7B: Compressing Large Language Models via Structured Pruning and Knowledge Distillation for the Polish Language (viability: 7): https://sciencetostartup.com/paper/bielik-minitron-7b-compressing-large-language-models-via-structured-pruning-and-knowledge-distillation-for-the-polish-la - Bielik-Minitron-7B compresses large language models for European languages, achieving significant parameter reduction and performance retention.
- On the Possible Detectability of Image-in-Image Steganography (viability: 4): https://sciencetostartup.com/paper/on-the-possible-detectability-of-image-in-image-steganography - A novel method for detecting image-in-image steganography with high accuracy using independent component analysis.
- The Mirror Design Pattern: Strict Data Geometry over Model Scale for Prompt Injection Detection (viability: 6): https://sciencetostartup.com/paper/the-mirror-design-pattern-strict-data-geometry-over-model-scale-for-prompt-injection-detection - Mirror is a fast and deterministic data-curation design pattern for effective prompt injection detection.
- AdaFuse: Accelerating Dynamic Adapter Inference via Token-Level Pre-Gating and Fused Kernel Optimization (viability: 3): https://sciencetostartup.com/paper/adafuse-accelerating-dynamic-adapter-inference-via-token-level-pre-gating-and-fused-kernel-optimization - AdaFuse optimizes dynamic adapter inference for LLMs, significantly reducing latency while maintaining accuracy.
- On the Role of Reversible Instance Normalization (viability: 2): https://sciencetostartup.com/paper/on-the-role-of-reversible-instance-normalization - This paper explores the challenges of data normalization in time series forecasting and proposes improvements to Reversible Instance Normalization.
- Data Fusion with Distributional Equivalence Test-then-pool (viability: 3): https://sciencetostartup.com/paper/data-fusion-with-distributional-equivalence-test-then-pool - A novel framework for fusing control arms in clinical trials while controlling Type-I error rates.
- Derain-Agent: A Plug-and-Play Agent Framework for Rainy Image Restoration (viability: 7): https://sciencetostartup.com/paper/derain-agent-a-plug-and-play-agent-framework-for-rainy-image-restoration - Derain-Agent is a dynamic agent-based framework that enhances image restoration by intelligently scheduling restoration tools for optimal results.
- Social, Legal, Ethical, Empathetic and Cultural Norm Operationalisation for AI Agents (viability: 3): https://sciencetostartup.com/paper/social-legal-ethical-empathetic-and-cultural-norm-operationalisation-for-ai-agents - A framework for operationalizing social, legal, ethical, empathetic, and cultural norms in AI agents.
- CreativeBench: Benchmarking and Enhancing Machine Creativity via Self-Evolving Challenges (viability: 6): https://sciencetostartup.com/paper/creativebench-benchmarking-and-enhancing-machine-creativity-via-self-evolving-challenges - CreativeBench is a benchmark for evaluating and enhancing machine creativity in code generation through self-evolving challenges.
- You Told Me to Do It: Measuring Instructional Text-induced Private Data Leakage in LLM Agents (viability: 6): https://sciencetostartup.com/paper/you-told-me-to-do-it-measuring-instructional-text-induced-private-data-leakage-in-llm-agents - A benchmark for measuring private data leakage in LLM agents due to malicious instruction execution.
- Automatic Attack Script Generation: a MDA Approach (viability: 2): https://sciencetostartup.com/paper/automatic-attack-script-generation-a-mda-approach - An approach to automatically generate cybersecurity attack scripts for educational purposes.
- Multi-Station WiFi CSI Sensing Framework Robust to Station-wise Feature Missingness and Limited Labeled Data (viability: 2): https://sciencetostartup.com/paper/multi-station-wifi-csi-sensing-framework-robust-to-station-wise-feature-missingness-and-limited-labeled-data - A framework for robust multi-station WiFi CSI sensing addressing feature missingness and limited labeled data.
- Inverse Neural Operator for ODE Parameter Optimization (viability: 7): https://sciencetostartup.com/paper/inverse-neural-operator-for-ode-parameter-optimization - The Inverse Neural Operator optimizes ODE parameter recovery from sparse observations with a novel two-stage framework.
- OpenClaw PRISM: A Zero-Fork, Defense-in-Depth Runtime Security Layer for Tool-Augmented LLM Agents (viability: 7): https://sciencetostartup.com/paper/openclaw-prism-a-zero-fork-defense-in-depth-runtime-security-layer-for-tool-augmented-llm-agents - OpenClaw PRISM provides a runtime security layer to mitigate risks in tool-augmented LLM agents.
- ZeroSense:How Vision matters in Long Context Compression (viability: 4): https://sciencetostartup.com/paper/zerosense-how-vision-matters-in-long-context-compression - ZeroSense introduces a novel evaluation framework for visual-text compression that accurately measures VTC quality independent of downstream model biases.
- The Landscape of Generative AI in Information Systems: A Synthesis of Secondary Reviews and Research Agendas (viability: 2): https://sciencetostartup.com/paper/the-landscape-of-generative-ai-in-information-systems-a-synthesis-of-secondary-reviews-and-research-agendas - This paper reviews the challenges and potential of Generative AI in information systems without presenting a clear product path.
- DatedGPT: Preventing Lookahead Bias in Large Language Models with Time-Aware Pretraining (viability: 8): https://sciencetostartup.com/paper/datedgpt-preventing-lookahead-bias-in-large-language-models-with-time-aware-pretraining - DatedGPT offers a solution to lookahead bias in financial forecasting by using time-aware pretraining of large language models.
- A Decade of Generative Adversarial Networks for Porous Material Reconstruction (viability: 2): https://sciencetostartup.com/paper/a-decade-of-generative-adversarial-networks-for-porous-material-reconstruction - A comprehensive review of GAN architectures for porous material reconstruction.
- Hypercomplex Widely Linear Processing: Fundamentals for Quaternion Machine Learning (viability: 2): https://sciencetostartup.com/paper/hypercomplex-widely-linear-processing-fundamentals-for-quaternion-machine-learning - This paper explores the fundamentals of quaternion machine learning for modeling three-dimensional rotations.
- Hybrid Human-Agent Social Dilemmas in Energy Markets (viability: 4): https://sciencetostartup.com/paper/hybrid-human-agent-social-dilemmas-in-energy-markets - A study on how artificial agents can enhance cooperation in energy load management through strategic decision-making.
- Towards High-Fidelity CAD Generation via LLM-Driven Program Generation and Text-Based B-Rep Primitive Grounding (viability: 7): https://sciencetostartup.com/paper/towards-high-fidelity-cad-generation-via-llm-driven-program-generation-and-text-based-b-rep-primitive-grounding - FutureCAD is a text-to-CAD framework that utilizes LLMs for high-fidelity CAD generation through natural language queries.
- Multimodal classification of Radiation-Induced Contrast Enhancements and tumor recurrence using deep learning (viability: 3): https://sciencetostartup.com/paper/multimodal-classification-of-radiation-induced-contrast-enhancements-and-tumor-recurrence-using-deep-learning - RICE-NET is a deep learning model for classifying tumor recurrence versus radiation-induced enhancements in glioblastoma patients using multimodal MRI data.
- Automated Detection of Malignant Lesions in the Ovary Using Deep Learning Models and XAI (viability: 5): https://sciencetostartup.com/paper/automated-detection-of-malignant-lesions-in-the-ovary-using-deep-learning-models-and-xai - A deep learning approach for accurate detection of ovarian cancer using advanced CNN models and explainable AI techniques.
- VisiFold: Long-Term Traffic Forecasting via Temporal Folding Graph and Node Visibility (viability: 8): https://sciencetostartup.com/paper/visifold-long-term-traffic-forecasting-via-temporal-folding-graph-and-node-visibility - VisiFold revolutionizes long-term traffic forecasting by integrating temporal folding graphs and node visibility mechanisms to enhance prediction accuracy while reducing computational costs.
- RADAR: Closed-Loop Robotic Data Generation via Semantic Planning and Autonomous Causal Environment Reset (viability: 8): https://sciencetostartup.com/paper/radar-closed-loop-robotic-data-generation-via-semantic-planning-and-autonomous-causal-environment-reset - RADAR is an autonomous data generation engine that revolutionizes robotic learning by eliminating human intervention in data collection.
- CEI-3D: Collaborative Explicit-Implicit 3D Reconstruction for Realistic and Fine-Grained Object Editing (viability: 9): https://sciencetostartup.com/paper/cei-3d-collaborative-explicit-implicit-3d-reconstruction-for-realistic-and-fine-grained-object-editing - CEI-3D is a collaborative 3D reconstruction pipeline that enables realistic and fine-grained object editing with localized control.
- HiSync: Spatio-Temporally Aligning Hand Motion from Wearable IMU and On-Robot Camera for Command Source Identification in Long-Range HRI (viability: 8): https://sciencetostartup.com/paper/hisync-spatio-temporally-aligning-hand-motion-from-wearable-imu-and-on-robot-camera-for-command-source-identification-in - HiSync enhances command source identification in long-range human-robot interactions using a novel optical-inertial fusion framework.
- Automating Skill Acquisition through Large-Scale Mining of Open-Source Agentic Repositories: A Framework for Multi-Agent Procedural Knowledge Extraction (viability: 7): https://sciencetostartup.com/paper/automating-skill-acquisition-through-large-scale-mining-of-open-source-agentic-repositories-a-framework-for-multi-agent- - A framework for automating skill acquisition in AI agents through mining open-source repositories.
- OSM-based Domain Adaptation for Remote Sensing VLMs (viability: 8): https://sciencetostartup.com/paper/osm-based-domain-adaptation-for-remote-sensing-vlms - OSMDA is a self-contained domain adaptation framework for Vision-Language Models that eliminates the need for costly external annotations by leveraging OpenStreetMap data.
- A Semi-Decentralized Approach to Multiagent Control (viability: 4): https://sciencetostartup.com/paper/a-semi-decentralized-approach-to-multiagent-control - A framework for semi-decentralized control of cooperative agents in uncertain communication environments.
- Exponential-Family Membership Inference: From LiRA and RMIA to BaVarIA (viability: 4): https://sciencetostartup.com/paper/exponential-family-membership-inference-from-lira-and-rmia-to-bavaria - BaVarIA offers a unified framework for membership inference attacks, enhancing model privacy auditing with improved performance.
- DocSage: An Information Structuring Agent for Multi-Doc Multi-Entity Question Answering (viability: 8): https://sciencetostartup.com/paper/docsage-an-information-structuring-agent-for-multi-doc-multi-entity-question-answering - DocSage is an advanced framework for multi-document multi-entity question answering that enhances relational reasoning and information extraction.
- Intrinsic Concept Extraction Based on Compositional Interpretability (viability: 4): https://sciencetostartup.com/paper/intrinsic-concept-extraction-based-on-compositional-interpretability - HyperExpress enables the extraction of composable intrinsic concepts from images using advanced hierarchical modeling.
- Locating Demographic Bias at the Attention-Head Level in CLIP's Vision Encoder (viability: 4): https://sciencetostartup.com/paper/locating-demographic-bias-at-the-attention-head-level-in-clip-s-vision-encoder - A mechanistic fairness audit tool that identifies demographic bias at the attention-head level in vision transformers.
- Disentangled Representation Learning through Unsupervised Symmetry Group Discovery (viability: 3): https://sciencetostartup.com/paper/disentangled-representation-learning-through-unsupervised-symmetry-group-discovery - A method for unsupervised symmetry group discovery to enhance representation learning.
- Language Generation with Replay: A Learning-Theoretic View of Model Collapse (viability: 2): https://sciencetostartup.com/paper/language-generation-with-replay-a-learning-theoretic-view-of-model-collapse - This paper explores the theoretical implications of model collapse in language generation due to the re-entry of machine-generated content into training datasets.
- HELM: Hierarchical and Explicit Label Modeling with Graph Learning for Multi-Label Image Classification (viability: 7): https://sciencetostartup.com/paper/helm-hierarchical-and-explicit-label-modeling-with-graph-learning-for-multi-label-image-classification - HELM is a novel framework for hierarchical multi-label classification that leverages graph learning and self-supervised techniques to enhance label modeling in remote sensing imagery.
- From Debate to Deliberation: Structured Collective Reasoning with Typed Epistemic Acts (viability: 3): https://sciencetostartup.com/paper/from-debate-to-deliberation-structured-collective-reasoning-with-typed-epistemic-acts - Introducing a structured deliberation framework for multi-agent LLM systems to enhance decision-making processes.
- Large Language Models for Biomedical Article Classification (viability: 5): https://sciencetostartup.com/paper/large-language-models-for-biomedical-article-classification - Leveraging large language models for effective biomedical article classification.
- Trust Oriented Explainable AI for Fake News Detection (viability: 5): https://sciencetostartup.com/paper/trust-oriented-explainable-ai-for-fake-news-detection - A framework that enhances fake news detection through Explainable AI techniques for improved transparency and interpretability.
- Legal-DC: Benchmarking Retrieval-Augmented Generation for Legal Documents (viability: 9): https://sciencetostartup.com/paper/legal-dc-benchmarking-retrieval-augmented-generation-for-legal-documents - Legal-DC offers a specialized benchmark and framework for enhancing retrieval-augmented generation in Chinese legal documents.
- An Automatic Text Classification Method Based on Hierarchical Taxonomies, Neural Networks and Document Embedding: The NETHIC Tool (viability: 3): https://sciencetostartup.com/paper/an-automatic-text-classification-method-based-on-hierarchical-taxonomies-neural-networks-and-document-embedding-the-neth - NETHIC is an automatic text classification tool leveraging neural networks and hierarchical taxonomies for improved efficiency and effectiveness.
- Governing Evolving Memory in LLM Agents: Risks, Mechanisms, and the Stability and Safety Governed Memory (SSGM) Framework (viability: 2): https://sciencetostartup.com/paper/governing-evolving-memory-in-llm-agents-risks-mechanisms-and-the-stability-and-safety-governed-memory-ssgm-framework - A conceptual framework addressing memory governance risks in autonomous LLM agents.
- Understanding Wikidata Qualifiers: An Analysis and Taxonomy (viability: 2): https://sciencetostartup.com/paper/understanding-wikidata-qualifiers-an-analysis-and-taxonomy - A comprehensive taxonomy for understanding and utilizing Wikidata qualifiers.
- A Further Efficient Algorithm with Best-of-Both-Worlds Guarantees for $m$-Set Semi-Bandit Problem (viability: 2): https://sciencetostartup.com/paper/a-further-efficient-algorithm-with-best-of-both-worlds-guarantees-for-m-set-semi-bandit-problem - This paper presents an efficient algorithm for the $m$-set semi-bandit problem with optimal regret guarantees.
- Modeling Trial-and-Error Navigation With a Sequential Decision Model of Information Scent (viability: 2): https://sciencetostartup.com/paper/modeling-trial-and-error-navigation-with-a-sequential-decision-model-of-information-scent - A theoretical model explaining user navigation errors in information architectures.
- Exploiting Expertise of Non-Expert and Diverse Agents in Social Bandit Learning: A Free Energy Approach (viability: 3): https://sciencetostartup.com/paper/exploiting-expertise-of-non-expert-and-diverse-agents-in-social-bandit-learning-a-free-energy-approach - A novel social bandit learning algorithm that enhances individual learning by leveraging social learning from diverse agents.
- Anomaly detection in time-series via inductive biases in the latent space of conditional normalizing flows (viability: 4): https://sciencetostartup.com/paper/anomaly-detection-in-time-series-via-inductive-biases-in-the-latent-space-of-conditional-normalizing-flows - A novel approach to anomaly detection in time-series using inductive biases in latent space.
- Controllable Egocentric Video Generation via Occlusion-Aware Sparse 3D Hand Joints (viability: 8): https://sciencetostartup.com/paper/controllable-egocentric-video-generation-via-occlusion-aware-sparse-3d-hand-joints - A novel framework for generating high-fidelity egocentric videos using sparse 3D hand joints for motion control.
- Mitigating the Multiplicity Burden: The Role of Calibration in Reducing Predictive Multiplicity of Classifiers (viability: 4): https://sciencetostartup.com/paper/mitigating-the-multiplicity-burden-the-role-of-calibration-in-reducing-predictive-multiplicity-of-classifiers - This paper explores calibration methods to reduce predictive multiplicity in classifiers, enhancing reliability in high-stakes environments.
- Compression Favors Consistency, Not Truth: When and Why Language Models Prefer Correct Information (viability: 8): https://sciencetostartup.com/paper/compression-favors-consistency-not-truth-when-and-why-language-models-prefer-correct-information - A study revealing how language models prioritize consistent information over truth, with implications for model training.
- SoulX-LiveAct: Towards Hour-Scale Real-Time Human Animation with Neighbor Forcing and ConvKV Memory (viability: 8): https://sciencetostartup.com/paper/soulx-liveact-towards-hour-scale-real-time-human-animation-with-neighbor-forcing-and-convkv-memory - SoulX-LiveAct enables hour-scale real-time human animation with improved efficiency and quality using innovative memory mechanisms.
- CINDI: Conditional Imputation and Noisy Data Integrity with Flows in Power Grid Data (viability: 6): https://sciencetostartup.com/paper/cindi-conditional-imputation-and-noisy-data-integrity-with-flows-in-power-grid-data - CINDI is an unsupervised framework that unifies anomaly detection and imputation for restoring data integrity in multivariate time series.
- Semi-Synthetic Parallel Data for Translation Quality Estimation: A Case Study of Dataset Building for an Under-Resourced Language Pair (viability: 5): https://sciencetostartup.com/paper/semi-synthetic-parallel-data-for-translation-quality-estimation-a-case-study-of-dataset-building-for-an-under-resourced- - A semi-synthetic dataset for improving translation quality estimation in under-resourced language pairs.
- Gender Bias in Generative AI-assisted Recruitment Processes (viability: 2): https://sciencetostartup.com/paper/gender-bias-in-generative-ai-assisted-recruitment-processes - This research evaluates gender bias in generative AI recruitment processes, highlighting ethical concerns.
- VTEdit-Bench: A Comprehensive Benchmark for Multi-Reference Image Editing Models in Virtual Try-On (viability: 4): https://sciencetostartup.com/paper/vtedit-bench-a-comprehensive-benchmark-for-multi-reference-image-editing-models-in-virtual-try-on - VTEdit-Bench is a comprehensive benchmark for evaluating universal multi-reference image editing models in virtual try-on applications.
- Adapting Dijkstra for Buffers and Unlimited Transfers (viability: 4): https://sciencetostartup.com/paper/adapting-dijkstra-for-buffers-and-unlimited-transfers - Introducing Transfer Aware Dijkstra (TAD) for optimized public transit routing that accounts for buffer times.
- Software-Hardware Binding for Protection of Sensitive Data in Embedded Software (viability: 3): https://sciencetostartup.com/paper/software-hardware-binding-for-protection-of-sensitive-data-in-embedded-software - A novel protection mechanism for sensitive data in embedded software using hardware fingerprints and Boolean logic.
- Cross-Resolution Attention Network for High-Resolution PM2.5 Prediction (viability: 3): https://sciencetostartup.com/paper/cross-resolution-attention-network-for-high-resolution-pm2-5-prediction - CRAN-PM is a dual-branch Vision Transformer designed for efficient PM2.5 forecasting across Europe using cross-resolution attention.
- When OpenClaw Meets Hospital: Toward an Agentic Operating System for Dynamic Clinical Workflows (viability: 3): https://sciencetostartup.com/paper/when-openclaw-meets-hospital-toward-an-agentic-operating-system-for-dynamic-clinical-workflows - This paper proposes an architecture for deploying LLM agents in clinical workflows, focusing on safety and coordination.
- COTONET: A custom cotton detection algorithm based on YOLO11 for stage of growth cotton boll detection (viability: 3): https://sciencetostartup.com/paper/cotonet-a-custom-cotton-detection-algorithm-based-on-yolo11-for-stage-of-growth-cotton-boll-detection - COTONET is a custom YOLO11 model designed for precise cotton boll detection to enhance harvesting quality.
- Scaling Laws for Educational AI Agents (viability: 7): https://sciencetostartup.com/paper/scaling-laws-for-educational-ai-agents - EduClaw is a profile-driven platform that enhances educational AI agents through structured capability scaling.
- EvoFlows: Evolutionary Edit-Based Flow-Matching for Protein Engineering (viability: 7): https://sciencetostartup.com/paper/evoflows-evolutionary-edit-based-flow-matching-for-protein-engineering - EvoFlows is a novel protein modeling approach that predicts mutations and their locations for enhanced protein engineering.
- Decomposing Observational Multiplicity in Decision Trees: Leaf and Structural Regret (viability: 4): https://sciencetostartup.com/paper/decomposing-observational-multiplicity-in-decision-trees-leaf-and-structural-regret - A framework for quantifying observational multiplicity in decision trees to enhance model safety and interpretability.
- OSCBench: Benchmarking Object State Change in Text-to-Video Generation (viability: 4): https://sciencetostartup.com/paper/oscbench-benchmarking-object-state-change-in-text-to-video-generation - OSCBench is a new benchmark for evaluating object state change in text-to-video generation models.
- PolyCrysDiff: Controllable Generation of Three-Dimensional Computable Polycrystalline Material Structures (viability: 6): https://sciencetostartup.com/paper/polycrysdiff-controllable-generation-of-three-dimensional-computable-polycrystalline-material-structures - PolyCrysDiff enables the generation of computable 3D polycrystalline microstructures for material optimization.
- STAIRS-Former: Spatio-Temporal Attention with Interleaved Recursive Structure Transformer for Offline Multi-task Multi-agent Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/stairs-former-spatio-temporal-attention-with-interleaved-recursive-structure-transformer-for-offline-multi-task-multi-ag - STAIRS-Former enhances multi-agent reinforcement learning by improving attention mechanisms for better coordination and long-horizon dependency capture.
- Explicit Logic Channel for Validation and Enhancement of MLLMs on Zero-Shot Tasks (viability: 6): https://sciencetostartup.com/paper/explicit-logic-channel-for-validation-and-enhancement-of-mllms-on-zero-shot-tasks - A novel Explicit Logic Channel enhances validation and performance of Multimodal Large Language Models in zero-shot tasks.
- SemBench: A Universal Semantic Framework for LLM Evaluation (viability: 6): https://sciencetostartup.com/paper/sembench-a-universal-semantic-framework-for-llm-evaluation - SemBench is a scalable framework for generating synthetic benchmarks to evaluate the semantic understanding of LLMs across multiple languages.
- In the LLM era, Word Sense Induction remains unsolved (viability: 5): https://sciencetostartup.com/paper/in-the-llm-era-word-sense-induction-remains-unsolved - A novel approach to Word Sense Induction leveraging LLMs and data augmentation techniques.
- Causal Prosody Mediation for Text-to-Speech:Counterfactual Training of Duration, Pitch, and Energy in FastSpeech2 (viability: 7): https://sciencetostartup.com/paper/causal-prosody-mediation-for-text-to-speech-counterfactual-training-of-duration-pitch-and-energy-in-fastspeech2 - A novel framework for expressive text-to-speech synthesis that enhances emotional prosody control using causal learning principles.
- Entropy-Preserving Reinforcement Learning (viability: 2): https://sciencetostartup.com/paper/entropy-preserving-reinforcement-learning - This paper proposes methods to control entropy in policy gradient algorithms to enhance exploration and performance in reinforcement learning.
- UCAN: Unified Convolutional Attention Network for Expansive Receptive Fields in Lightweight Super-Resolution (viability: 3): https://sciencetostartup.com/paper/ucan-unified-convolutional-attention-network-for-expansive-receptive-fields-in-lightweight-super-resolution - UCAN is a lightweight network that efficiently combines convolution and attention for high-resolution image restoration.
- LLMs can construct powerful representations and streamline sample-efficient supervised learning (viability: 7): https://sciencetostartup.com/paper/llms-can-construct-powerful-representations-and-streamline-sample-efficient-supervised-learning - An agentic pipeline that uses LLMs to streamline input representation for multimodal supervised learning in healthcare.
- From Control to Foresight: Simulation as a New Paradigm for Human-Agent Collaboration (viability: 2): https://sciencetostartup.com/paper/from-control-to-foresight-simulation-as-a-new-paradigm-for-human-agent-collaboration - A new paradigm for human-agent collaboration that emphasizes foresight through simulation.
- Stable Spike: Dual Consistency Optimization via Bitwise AND Operations for Spiking Neural Networks (viability: 4): https://sciencetostartup.com/paper/stable-spike-dual-consistency-optimization-via-bitwise-and-operations-for-spiking-neural-networks - A method to enhance the performance of spiking neural networks through dual consistency optimization.
- PROMO: Promptable Outfitting for Efficient High-Fidelity Virtual Try-On (viability: 7): https://sciencetostartup.com/paper/promo-promptable-outfitting-for-efficient-high-fidelity-virtual-try-on - PROMO is a promptable virtual try-on framework that enhances online retail by providing high-fidelity, efficient virtual fitting solutions.
- Context-dependent manifold learning: A neuromodulated constrained autoencoder approach (viability: 4): https://sciencetostartup.com/paper/context-dependent-manifold-learning-a-neuromodulated-constrained-autoencoder-approach - Introducing a neuromodulated constrained autoencoder for context-dependent manifold learning.
- A technology-oriented mapping of the language and translation industry: Analysing stakeholder values and their potential implication for translation pedagogy (viability: 2): https://sciencetostartup.com/paper/a-technology-oriented-mapping-of-the-language-and-translation-industry-analysing-stakeholder-values-and-their-potential- - This paper explores the evolving values in the automated translation industry and their implications for translation pedagogy.
- Multi-Task Reinforcement Learning for Enhanced Multimodal LLM-as-a-Judge (viability: 5): https://sciencetostartup.com/paper/multi-task-reinforcement-learning-for-enhanced-multimodal-llm-as-a-judge - A framework that enhances multimodal LLMs as judges through multi-task reinforcement learning for improved evaluation consistency.
- BackdoorIDS: Zero-shot Backdoor Detection for Pretrained Vision Encoder (viability: 8): https://sciencetostartup.com/paper/backdoorids-zero-shot-backdoor-detection-for-pretrained-vision-encoder - BackdoorIDS offers a zero-shot method for detecting backdoor attacks in pretrained vision encoders, enhancing security in AI applications.
- FL-MedSegBench: A Comprehensive Benchmark for Federated Learning on Medical Image Segmentation (viability: 8): https://sciencetostartup.com/paper/fl-medsegbench-a-comprehensive-benchmark-for-federated-learning-on-medical-image-segmentation - FL-MedSegBench is a benchmark toolkit for evaluating federated learning methods in medical image segmentation.
- Coupling Tensor Trains with Graph of Convex Sets: Effective Compression, Exploration, and Planning in the C-Space (viability: 3): https://sciencetostartup.com/paper/coupling-tensor-trains-with-graph-of-convex-sets-effective-compression-exploration-and-planning-in-the-c-space - TANGO is a novel motion planning framework that combines tensor-based compression with graph optimization for efficient trajectory generation.
- Concurrent Prehensile and Nonprehensile Manipulation: A Practical Approach to Multi-Stage Dexterous Tasks (viability: 7): https://sciencetostartup.com/paper/concurrent-prehensile-and-nonprehensile-manipulation-a-practical-approach-to-multi-stage-dexterous-tasks - DexMulti enables efficient multi-stage dexterous manipulation in robotics by leveraging object-centric skills.
- Simple Recipe Works: Vision-Language-Action Models are Natural Continual Learners with Reinforcement Learning (viability: 8): https://sciencetostartup.com/paper/simple-recipe-works-vision-language-action-models-are-natural-continual-learners-with-reinforcement-learning - A novel approach to continual reinforcement learning for vision-language-action models that enhances adaptability and reduces forgetting.
- QChunker: Learning Question-Aware Text Chunking for Domain RAG via Multi-Agent Debate (viability: 7): https://sciencetostartup.com/paper/qchunker-learning-question-aware-text-chunking-for-domain-rag-via-multi-agent-debate - QChunker enhances retrieval-augmented generation by improving text chunking through a multi-agent debate framework.
- A Hybrid Neural-Assisted Unscented Kalman Filter for Unmanned Ground Vehicle Navigation (viability: 6): https://sciencetostartup.com/paper/a-hybrid-neural-assisted-unscented-kalman-filter-for-unmanned-ground-vehicle-navigation - A hybrid neural-assisted Kalman filter improves navigation accuracy for unmanned ground vehicles by predicting noise uncertainty from raw sensor data.
- OmniForcing: Unleashing Real-time Joint Audio-Visual Generation (viability: 8): https://sciencetostartup.com/paper/omniforcing-unleashing-real-time-joint-audio-visual-generation - OmniForcing is a real-time joint audio-visual generation framework that achieves state-of-the-art performance with low latency.
- IDRL: An Individual-Aware Multimodal Depression-Related Representation Learning Framework for Depression Diagnosis (viability: 4): https://sciencetostartup.com/paper/idrl-an-individual-aware-multimodal-depression-related-representation-learning-framework-for-depression-diagnosis - IDRL is a framework for improving depression diagnosis through individual-aware multimodal representation learning.
- Chunk-Boundary Artifact in Action-Chunked Generative Policies: A Noise-Sensitive Failure Mechanism (viability: 2): https://sciencetostartup.com/paper/chunk-boundary-artifact-in-action-chunked-generative-policies-a-noise-sensitive-failure-mechanism - This paper analyzes a noise-sensitive failure mechanism in action-chunked generative visuomotor policies.
- Tokenization Allows Multimodal Large Language Models to Understand, Generate and Edit Architectural Floor Plans (viability: 3): https://sciencetostartup.com/paper/tokenization-allows-multimodal-large-language-models-to-understand-generate-and-edit-architectural-floor-plans - HouseMind is a multimodal large language model designed for understanding, generating, and editing architectural floor plans.
- Learn Structure, Adapt on the Fly: Multi-Scale Residual Learning and Online Adaptation for Aerial Manipulators (viability: 8): https://sciencetostartup.com/paper/learn-structure-adapt-on-the-fly-multi-scale-residual-learning-and-online-adaptation-for-aerial-manipulators - A predictive-adaptive framework for real-time modeling and compensation in autonomous aerial manipulators.
- Diversity You Can Actually Measure: A Fast, Model-Free Diversity Metric for Robotics Datasets (viability: 7): https://sciencetostartup.com/paper/diversity-you-can-actually-measure-a-fast-model-free-diversity-metric-for-robotics-datasets - FAKTUAL is a model-free algorithm that curates diverse robot imitation learning datasets to enhance generalization performance.
- MV-SAM3D: Adaptive Multi-View Fusion for Layout-Aware 3D Generation (viability: 9): https://sciencetostartup.com/paper/mv-sam3d-adaptive-multi-view-fusion-for-layout-aware-3d-generation - MV-SAM3D enhances 3D generation by integrating multi-view consistency and physical plausibility without additional training.
- From Pets to Robots: MojiKit as a Data-Informed Toolkit for Affective HRI Design (viability: 4): https://sciencetostartup.com/paper/from-pets-to-robots-mojikit-as-a-data-informed-toolkit-for-affective-hri-design - MojiKit is a toolkit that empowers users to design affective behaviors for social robots using structured resources and a code-free studio.
- VisDoT : Enhancing Visual Reasoning through Human-Like Interpretation Grounding and Decomposition of Thought (viability: 8): https://sciencetostartup.com/paper/visdot-enhancing-visual-reasoning-through-human-like-interpretation-grounding-and-decomposition-of-thought - VisDoT enhances visual reasoning in charts through human-like interpretation and decomposition of thought.
- Developing Foundation Models for Universal Segmentation from 3D Whole-Body Positron Emission Tomography (viability: 7): https://sciencetostartup.com/paper/developing-foundation-models-for-universal-segmentation-from-3d-whole-body-positron-emission-tomography - SegAnyPET is a foundational model for universal segmentation in 3D whole-body PET imaging, enhancing clinical workflows with minimal human correction.
- MedPruner: Training-Free Hierarchical Token Pruning for Efficient 3D Medical Image Understanding in Vision-Language Models (viability: 8): https://sciencetostartup.com/paper/medpruner-training-free-hierarchical-token-pruning-for-efficient-3d-medical-image-understanding-in-vision-language-model - MedPruner is a training-free framework for efficient 3D medical image understanding through hierarchical token pruning.
- The Density of Cross-Persistence Diagrams and Its Applications (viability: 8): https://sciencetostartup.com/paper/the-density-of-cross-persistence-diagrams-and-its-applications - A machine learning framework for predicting cross-persistence density from point cloud data, enhancing topological data analysis.
- Personalized Federated Learning via Gaussian Generative Modeling (viability: 7): https://sciencetostartup.com/paper/personalized-federated-learning-via-gaussian-generative-modeling - pFedGM enhances personalized federated learning by modeling client heterogeneity through Gaussian generative modeling.
- Taming OpenClaw: Security Analysis and Mitigation of Autonomous LLM Agent Threats (viability: 4): https://sciencetostartup.com/paper/taming-openclaw-security-analysis-and-mitigation-of-autonomous-llm-agent-threats - A comprehensive security framework for mitigating threats in autonomous LLM agents.
- Shape-of-You: Fused Gromov-Wasserstein Optimal Transport for Semantic Correspondence in-the-Wild (viability: 8): https://sciencetostartup.com/paper/shape-of-you-fused-gromov-wasserstein-optimal-transport-for-semantic-correspondence-in-the-wild - Shape-of-You offers a novel approach to semantic correspondence using Fused Gromov-Wasserstein optimal transport, achieving state-of-the-art results without explicit geometric annotations.
- Noise-aware few-shot learning through bi-directional multi-view prompt alignment (viability: 7): https://sciencetostartup.com/paper/noise-aware-few-shot-learning-through-bi-directional-multi-view-prompt-alignment - NA-MVP enhances few-shot learning by effectively distinguishing clean cues from noisy labels through innovative prompt alignment.
- SemiTooth: a Generalizable Semi-supervised Framework for Multi-Source Tooth Segmentation (viability: 7): https://sciencetostartup.com/paper/semitooth-a-generalizable-semi-supervised-framework-for-multi-source-tooth-segmentation - SemiTooth is a semi-supervised framework designed to enhance tooth segmentation from multi-source CBCT data in clinical dentistry.
- Fractional Rotation, Full Potential? Investigating Performance and Convergence of Partial RoPE (viability: 5): https://sciencetostartup.com/paper/fractional-rotation-full-potential-investigating-performance-and-convergence-of-partial-rope - This research explores the efficiency of partial Rotary Positional Embedding in transformer models, offering significant memory savings while maintaining performance.
- DyWeight: Dynamic Gradient Weighting for Few-Step Diffusion Sampling (viability: 9): https://sciencetostartup.com/paper/dyweight-dynamic-gradient-weighting-for-few-step-diffusion-sampling - DyWeight introduces a dynamic gradient weighting method to enhance the efficiency of diffusion models in generative tasks.
- Articulat3D: Reconstructing Articulated Digital Twins From Monocular Videos with Geometric and Motion Constraints (viability: 8): https://sciencetostartup.com/paper/articulat3d-reconstructing-articulated-digital-twins-from-monocular-videos-with-geometric-and-motion-constraints - Articulat3D reconstructs high-fidelity digital twins from monocular videos using advanced geometric and motion constraints.
- LaMoGen: Language to Motion Generation Through LLM-Guided Symbolic Inference (viability: 8): https://sciencetostartup.com/paper/lamogen-language-to-motion-generation-through-llm-guided-symbolic-inference - LaMoGen leverages symbolic reasoning to generate interpretable and linguistically grounded human motion from text.
- AutoScout: Structured Optimization for Automating ML System Configuration (viability: 3): https://sciencetostartup.com/paper/autoscout-structured-optimization-for-automating-ml-system-configuration - AutoScout automates the optimization of machine learning system configurations for improved efficiency.
- See, Symbolize, Act: Grounding VLMs with Spatial Representations for Better Gameplay (viability: 2): https://sciencetostartup.com/paper/see-symbolize-act-grounding-vlms-with-spatial-representations-for-better-gameplay - This research explores improving Vision-Language Models' gameplay performance through symbolic grounding.
- Hybrid Energy-Aware Reward Shaping: A Unified Lightweight Physics-Guided Methodology for Policy Optimization (viability: 3): https://sciencetostartup.com/paper/hybrid-energy-aware-reward-shaping-a-unified-lightweight-physics-guided-methodology-for-policy-optimization - A novel methodology for optimizing reinforcement learning policies with energy-aware reward shaping.
- Survival Meets Classification: A Novel Framework for Early Risk Prediction Models of Chronic Diseases (viability: 3): https://sciencetostartup.com/paper/survival-meets-classification-a-novel-framework-for-early-risk-prediction-models-of-chronic-diseases - A novel framework integrating survival analysis with classification for early risk prediction of chronic diseases.
- Performance Evaluation of Open-Source Large Language Models for Assisting Pathology Report Writing in Japanese (viability: 3): https://sciencetostartup.com/paper/performance-evaluation-of-open-source-large-language-models-for-assisting-pathology-report-writing-in-japanese - Evaluating open-source LLMs for enhancing Japanese pathology report writing.
- Leveraging Large Language Models and Survival Analysis for Early Prediction of Chemotherapy Outcomes (viability: 6): https://sciencetostartup.com/paper/leveraging-large-language-models-and-survival-analysis-for-early-prediction-of-chemotherapy-outcomes - A predictive model using LLMs for early chemotherapy outcome prediction to enhance patient management.
- WeEdit: A Dataset, Benchmark and Glyph-Guided Framework for Text-centric Image Editing (viability: 7): https://sciencetostartup.com/paper/weedit-a-dataset-benchmark-and-glyph-guided-framework-for-text-centric-image-editing - WeEdit is a comprehensive framework for precise text-centric image editing, leveraging a large-scale dataset and benchmarks.
- Toward Complex-Valued Neural Networks for Waveform Generation (viability: 9): https://sciencetostartup.com/paper/toward-complex-valued-neural-networks-for-waveform-generation - ComVo is a complex-valued neural vocoder that enhances waveform generation with structured feedback and improved training efficiency.
- Unsupervised LiDAR-Based Multi-UAV Detection and Tracking Under Extreme Sparsity (viability: 5): https://sciencetostartup.com/paper/unsupervised-lidar-based-multi-uav-detection-and-tracking-under-extreme-sparsity - An unsupervised LiDAR-based pipeline for detecting and tracking UAVs in sparse environments without labeled data.
- UtilityMax Prompting: A Formal Framework for Multi-Objective Large Language Model Optimization (viability: 7): https://sciencetostartup.com/paper/utilitymax-prompting-a-formal-framework-for-multi-objective-large-language-model-optimization - UtilityMax Prompting optimizes multi-objective LLM tasks using formal mathematical language for precise outputs.
- Streaming Translation and Transcription Through Speech-to-Text Causal Alignment (viability: 3): https://sciencetostartup.com/paper/streaming-translation-and-transcription-through-speech-to-text-causal-alignment - Hikari is an end-to-end model for simultaneous speech-to-text translation and streaming transcription that improves quality-latency trade-offs.
- R4Det: 4D Radar-Camera Fusion for High-Performance 3D Object Detection (viability: 7): https://sciencetostartup.com/paper/r4det-4d-radar-camera-fusion-for-high-performance-3d-object-detection - R4Det enhances 3D object detection in autonomous driving by fusing 4D radar and camera data with advanced depth estimation techniques.
- CAETC: Causal Autoencoding and Treatment Conditioning for Counterfactual Estimation over Time (viability: 4): https://sciencetostartup.com/paper/caetc-causal-autoencoding-and-treatment-conditioning-for-counterfactual-estimation-over-time - CAETC offers a novel approach for counterfactual estimation in personalized medicine by addressing time-dependent confounding bias.
- Where Matters More Than What: Decoding-aligned KV Cache Compression via Position-aware Pseudo Queries (viability: 7): https://sciencetostartup.com/paper/where-matters-more-than-what-decoding-aligned-kv-cache-compression-via-position-aware-pseudo-queries - DapQ offers a novel approach to compress KV cache for LLMs by using position-aware pseudo queries to optimize memory usage during inference.
- SVLL: Staged Vision-Language Learning for Physically Grounded Embodied Task Planning (viability: 8): https://sciencetostartup.com/paper/svll-staged-vision-language-learning-for-physically-grounded-embodied-task-planning - SVLL is a novel framework for robust, physically-grounded embodied task planning that outperforms existing models in real-world applications.
- How Intelligence Emerges: A Minimal Theory of Dynamic Adaptive Coordination (viability: 2): https://sciencetostartup.com/paper/how-intelligence-emerges-a-minimal-theory-of-dynamic-adaptive-coordination - A theoretical framework for understanding adaptive coordination in multi-agent systems.
- AI Knows What's Wrong But Cannot Fix It: Helicoid Dynamics in Frontier LLMs Under High-Stakes Decisions (viability: 2): https://sciencetostartup.com/paper/ai-knows-what-s-wrong-but-cannot-fix-it-helicoid-dynamics-in-frontier-llms-under-high-stakes-decisions - This paper explores failure dynamics in LLMs under high-stakes decisions, highlighting the need for improved reliability.
- RoboClaw: An Agentic Framework for Scalable Long-Horizon Robotic Tasks (viability: 7): https://sciencetostartup.com/paper/roboclaw-an-agentic-framework-for-scalable-long-horizon-robotic-tasks - RoboClaw is a unified framework for scalable long-horizon robotic manipulation tasks that reduces human intervention and improves success rates.
- TornadoNet: Real-Time Building Damage Detection with Ordinal Supervision (viability: 8): https://sciencetostartup.com/paper/tornadonet-real-time-building-damage-detection-with-ordinal-supervision - TornadoNet offers a robust framework for real-time building damage detection using advanced object detection architectures and ordinal supervision.
- Enhancing Image Aesthetics with Dual-Conditioned Diffusion Models Guided by Multimodal Perception (viability: 7): https://sciencetostartup.com/paper/enhancing-image-aesthetics-with-dual-conditioned-diffusion-models-guided-by-multimodal-perception - DIAE enhances image aesthetics by leveraging multimodal perception and a novel dataset for weakly supervised learning.
- MANSION: Multi-floor lANguage-to-3D Scene generatIOn for loNg-horizon tasks (viability: 8): https://sciencetostartup.com/paper/mansion-multi-floor-language-to-3d-scene-generation-for-long-horizon-tasks - MANSION is a language-driven framework for generating complex multi-floor 3D environments for robotic tasks.
- Shadowless Projection Mapping for Tabletop Workspaces with Synthetic Aperture Projector (viability: 4): https://sciencetostartup.com/paper/shadowless-projection-mapping-for-tabletop-workspaces-with-synthetic-aperture-projector - A synthetic-aperture projection mapping system that eliminates shadows for enhanced augmented reality experiences in tabletop workspaces.
- PCA-Enhanced Probabilistic U-Net for Effective Ambiguous Medical Image Segmentation (viability: 2): https://sciencetostartup.com/paper/pca-enhanced-probabilistic-u-net-for-effective-ambiguous-medical-image-segmentation - Introducing PCA-Enhanced Probabilistic U-Net for improved medical image segmentation under uncertainty.
- Multi-Task Anti-Causal Learning for Reconstructing Urban Events from Residents' Reports (viability: 4): https://sciencetostartup.com/paper/multi-task-anti-causal-learning-for-reconstructing-urban-events-from-residents-reports - MTAC is a framework for reconstructing urban events from residents' reports by inferring latent causes through multi-task anti-causal learning.
- One Supervisor, Many Modalities: Adaptive Tool Orchestration for Autonomous Queries (viability: 7): https://sciencetostartup.com/paper/one-supervisor-many-modalities-adaptive-tool-orchestration-for-autonomous-queries - An adaptive AI framework that orchestrates multimodal query processing for efficient and accurate results.
- Mango-GS: Enhancing Spatio-Temporal Consistency in Dynamic Scenes Reconstruction using Multi-Frame Node-Guided 4D Gaussian Splatting (viability: 8): https://sciencetostartup.com/paper/mango-gs-enhancing-spatio-temporal-consistency-in-dynamic-scenes-reconstruction-using-multi-frame-node-guided-4d-gaussia - Mango-GS offers a novel framework for high-fidelity 4D reconstruction of dynamic scenes with enhanced temporal consistency.
- ReHARK: Refined Hybrid Adaptive RBF Kernels for Robust One-Shot Vision-Language Adaptation (viability: 9): https://sciencetostartup.com/paper/rehark-refined-hybrid-adaptive-rbf-kernels-for-robust-one-shot-vision-language-adaptation - ReHARK offers a novel training-free framework for robust one-shot adaptation of Vision-Language Models, achieving state-of-the-art performance.
- MiNI-Q: A Miniature, Wire-Free Quadruped with Unbounded, Independently Actuated Leg Joints (viability: 7): https://sciencetostartup.com/paper/mini-q-a-miniature-wire-free-quadruped-with-unbounded-independently-actuated-leg-joints - MiNI-Q^2 is a miniature, wire-free quadruped robot with advanced locomotion capabilities and open-source design files.
- Expert Threshold Routing for Autoregressive Language Modeling with Dynamic Computation Allocation and Load Balancing (viability: 3): https://sciencetostartup.com/paper/expert-threshold-routing-for-autoregressive-language-modeling-with-dynamic-computation-allocation-and-load-balancing - A novel routing mechanism for autoregressive language models that optimizes computation allocation and load balancing.
- Risk-Controllable Multi-View Diffusion for Driving Scenario Generation (viability: 7): https://sciencetostartup.com/paper/risk-controllable-multi-view-diffusion-for-driving-scenario-generation - A pipeline for generating risk-controllable driving scenarios to enhance autonomous driving safety.
- Mobile-GS: Real-time Gaussian Splatting for Mobile Devices (viability: 7): https://sciencetostartup.com/paper/mobile-gs-real-time-gaussian-splatting-for-mobile-devices - Mobile-GS enables real-time Gaussian Splatting for high-quality rendering on mobile devices.
- CFD-HAR: User-controllable Privacy through Conditional Feature Disentanglement (viability: 2): https://sciencetostartup.com/paper/cfd-har-user-controllable-privacy-through-conditional-feature-disentanglement - A technique for user-controllable privacy in Human Activity Recognition through feature disentanglement.
- MDS-VQA: Model-Informed Data Selection for Video Quality Assessment (viability: 7): https://sciencetostartup.com/paper/mds-vqa-model-informed-data-selection-for-video-quality-assessment - MDS-VQA enhances video quality assessment by intelligently selecting challenging unlabeled videos for model fine-tuning.
- Strict Optimality of Frequency Estimation Under Local Differential Privacy (viability: 4): https://sciencetostartup.com/paper/strict-optimality-of-frequency-estimation-under-local-differential-privacy - A novel frequency estimation algorithm that achieves optimal precision under local differential privacy.
- EReCu: Pseudo-label Evolution Fusion and Refinement with Multi-Cue Learning for Unsupervised Camouflage Detection (viability: 4): https://sciencetostartup.com/paper/erecu-pseudo-label-evolution-fusion-and-refinement-with-multi-cue-learning-for-unsupervised-camouflage-detection - A unified framework for unsupervised camouflage detection that enhances pseudo-label reliability and feature fidelity.
- FBCIR: Balancing Cross-Modal Focuses in Composed Image Retrieval (viability: 6): https://sciencetostartup.com/paper/fbcir-balancing-cross-modal-focuses-in-composed-image-retrieval - FBCIR enhances composed image retrieval by addressing focus imbalances in multi-modal models through innovative data augmentation.
- Prediction of Grade, Gender, and Academic Performance of Children and Teenagers from Handwriting Using the Sigma-Lognormal Model (viability: 5): https://sciencetostartup.com/paper/prediction-of-grade-gender-and-academic-performance-of-children-and-teenagers-from-handwriting-using-the-sigma-lognormal - A system that predicts student characteristics from handwriting dynamics using advanced statistical models.
- Multi-Agent Collaboration for Automated Design Exploration on High Performance Computing Systems (viability: 6): https://sciencetostartup.com/paper/multi-agent-collaboration-for-automated-design-exploration-on-high-performance-computing-systems - MADA is a multi-agent framework that automates design exploration for high-performance computing systems, enhancing scientific discovery.
- Can Small Language Models Use What They Retrieve? An Empirical Study of Retrieval Utilization Across Model Scale (viability: 2): https://sciencetostartup.com/paper/can-small-language-models-use-what-they-retrieve-an-empirical-study-of-retrieval-utilization-across-model-scale - This study investigates the limitations of retrieval utilization in smaller language models, revealing significant challenges in context extraction.
- From Pen Strokes to Sleep States: Detecting Low-Recovery Days Using Sigma-Lognormal Handwriting Features (viability: 5): https://sciencetostartup.com/paper/from-pen-strokes-to-sleep-states-detecting-low-recovery-days-using-sigma-lognormal-handwriting-features - A framework for detecting low-recovery days through handwriting analysis and physiological metrics.
- Tiny Aya: Bridging Scale and Multilingual Depth (viability: 3): https://sciencetostartup.com/paper/tiny-aya-bridging-scale-and-multilingual-depth - Tiny Aya is a small multilingual language model achieving state-of-the-art translation quality across 70 languages.
- Manifold-Optimal Guidance: A Unified Riemannian Control View of Diffusion Guidance (viability: 7): https://sciencetostartup.com/paper/manifold-optimal-guidance-a-unified-riemannian-control-view-of-diffusion-guidance - Manifold-Optimal Guidance offers a novel approach to improve conditional diffusion by correcting off-manifold drift with a geometry-aware update.
- Gen-Fab: A Variation-Aware Generative Model for Predicting Fabrication Variations in Nanophotonic Devices (viability: 7): https://sciencetostartup.com/paper/gen-fab-a-variation-aware-generative-model-for-predicting-fabrication-variations-in-nanophotonic-devices - Gen-Fab is a generative model that predicts fabrication variations in nanophotonic devices using advanced machine learning techniques.
- LongFlow: Efficient KV Cache Compression for Reasoning M (viability: 3): https://sciencetostartup.com/paper/longflow-efficient-kv-cache-compression-for-reasoning-m - LongFlow optimizes KV cache compression for reasoning models, enhancing efficiency and reducing deployment costs.
- Sharpness-Aware Minimization for Generalized Embedding Learning in Federated Recommendation (viability: 7): https://sciencetostartup.com/paper/sharpness-aware-minimization-for-generalized-embedding-learning-in-federated-recommendation - A federated recommendation framework that enhances item embedding learning through sharpness-aware minimization.
- KEPo: Knowledge Evolution Poison on Graph-based Retrieval-Augmented Generation (viability: 4): https://sciencetostartup.com/paper/kepo-knowledge-evolution-poison-on-graph-based-retrieval-augmented-generation - KEPo is a novel poisoning attack method designed to exploit vulnerabilities in Graph-based Retrieval-Augmented Generation systems.
- ActiveFreq: Integrating Active Learning and Frequency Domain Analysis for Interactive Segmentation (viability: 7): https://sciencetostartup.com/paper/activefreq-integrating-active-learning-and-frequency-domain-analysis-for-interactive-segmentation - ActiveFreq enhances interactive medical image segmentation by integrating active learning with frequency domain analysis for improved accuracy and reduced user input.
- Try, Check and Retry: A Divide-and-Conquer Framework for Boosting Long-context Tool-Calling Performance of LLMs (viability: 8): https://sciencetostartup.com/paper/try-check-and-retry-a-divide-and-conquer-framework-for-boosting-long-context-tool-calling-performance-of-llms - Tool-DC enhances LLM tool-calling performance through a novel Divide-and-Conquer framework.
- OrthoEraser: Coupled-Neuron Orthogonal Projection for Concept Erasure (viability: 7): https://sciencetostartup.com/paper/orthoeraser-coupled-neuron-orthogonal-projection-for-concept-erasure - OrthoEraser enhances text-to-image models by safely erasing harmful concepts without damaging benign attributes.
- SPEGC: Continual Test-Time Adaptation via Semantic-Prompt-Enhanced Graph Clustering for Medical Image Segmentation (viability: 8): https://sciencetostartup.com/paper/spegc-continual-test-time-adaptation-via-semantic-prompt-enhanced-graph-clustering-for-medical-image-segmentation - SPEGC enhances medical image segmentation by enabling continual test-time adaptation through advanced graph clustering techniques.
- Attention Sinks Are Provably Necessary in Softmax Transformers: Evidence from Trigger-Conditional Tasks (viability: 3): https://sciencetostartup.com/paper/attention-sinks-are-provably-necessary-in-softmax-transformers-evidence-from-trigger-conditional-tasks - This research proves the necessity of attention sinks in softmax transformers and explores alternatives with ReLU attention.
- AnimeScore: A Preference-Based Dataset and Framework for Evaluating Anime-Like Speech Style (viability: 3): https://sciencetostartup.com/paper/animescore-a-preference-based-dataset-and-framework-for-evaluating-anime-like-speech-style - AnimeScore provides a framework for evaluating anime-like speech styles through preference-based metrics.
- INFACT: A Diagnostic Benchmark for Induced Faithfulness and Factuality Hallucinations in Video-LLMs (viability: 4): https://sciencetostartup.com/paper/infact-a-diagnostic-benchmark-for-induced-faithfulness-and-factuality-hallucinations-in-video-llms - INFACT is a diagnostic benchmark designed to evaluate and improve the reliability of Video-LLMs against hallucinations.
- SPARK: Skeleton-Parameter Aligned Retargeting on Humanoid Robots with Kinodynamic Trajectory Optimization (viability: 7): https://sciencetostartup.com/paper/spark-skeleton-parameter-aligned-retargeting-on-humanoid-robots-with-kinodynamic-trajectory-optimization - A two-stage pipeline for generating natural and dynamically feasible motion references for humanoid robots using human motion data.
- Grammar of the Wave: Towards Explainable Multivariate Time Series Event Detection via Neuro-Symbolic VLM Agents (viability: 6): https://sciencetostartup.com/paper/grammar-of-the-wave-towards-explainable-multivariate-time-series-event-detection-via-neuro-symbolic-vlm-agents - A neuro-symbolic framework for explainable event detection in multivariate time series using natural language descriptions.
- Leveraging Phytolith Research using Artificial Intelligence (viability: 3): https://sciencetostartup.com/paper/leveraging-phytolith-research-using-artificial-intelligence - Sorometry is an AI pipeline that automates phytolith analysis to enhance archaeological and paleoecological research.
- Deep Learning Network-Temporal Models For Traffic Prediction (viability: 4): https://sciencetostartup.com/paper/deep-learning-network-temporal-models-for-traffic-prediction - Deep learning models for enhanced multivariate time series traffic prediction leveraging network-temporal correlations.
- Slack More, Predict Better: Proximal Relaxation for Probabilistic Latent Variable Model-based Soft Sensors (viability: 4): https://sciencetostartup.com/paper/slack-more-predict-better-proximal-relaxation-for-probabilistic-latent-variable-model-based-soft-sensors - KProxNPLVM enhances soft sensor modeling accuracy by relaxing the learning objective to reduce approximation errors.
- HawkesRank: Event-Driven Centrality for Real-Time Importance Ranking (viability: 3): https://sciencetostartup.com/paper/hawkesrank-event-driven-centrality-for-real-time-importance-ranking - HawkesRank offers a dynamic framework for quantifying influence in networks through event-driven centrality measures.
- NFPO: Stabilized Policy Optimization of Normalizing Flow for Robotic Policy Learning (viability: 3): https://sciencetostartup.com/paper/nfpo-stabilized-policy-optimization-of-normalizing-flow-for-robotic-policy-learning - NFPO leverages Normalizing Flow for improved multi-modal policy learning in robotics.
- Stage-Adaptive Reliability Modeling for Continuous Valence-Arousal Estimation (viability: 7): https://sciencetostartup.com/paper/stage-adaptive-reliability-modeling-for-continuous-valence-arousal-estimation - SAGE enhances continuous valence-arousal estimation by dynamically calibrating modality reliability during multimodal integration.
- Bridging Discrete Marks and Continuous Dynamics: Dual-Path Cross-Interaction for Marked Temporal Point Processes (viability: 8): https://sciencetostartup.com/paper/bridging-discrete-marks-and-continuous-dynamics-dual-path-cross-interaction-for-marked-temporal-point-processes - NEXTPP is a dual-channel framework that enhances event sequence prediction by integrating discrete and continuous representations.
- CoViLLM: An Adaptive Human-Robot Collaborative Assembly Framework Using Large Language Models for Manufacturing (viability: 7): https://sciencetostartup.com/paper/covillm-an-adaptive-human-robot-collaborative-assembly-framework-using-large-language-models-for-manufacturing - CoViLLM is an adaptive framework that enhances human-robot collaboration in manufacturing by enabling flexible assembly of customized products using LLMs.
- Follow the Saliency: Supervised Saliency for Retrieval-augmented Dense Video Captioning (viability: 8): https://sciencetostartup.com/paper/follow-the-saliency-supervised-saliency-for-retrieval-augmented-dense-video-captioning - STaRC enhances Dense Video Captioning by using supervised saliency for improved temporal segmentation and caption generation.
- UniHetCO: A Unified Heterogeneous Representation for Multi-Problem Learning in Unsupervised Neural Combinatorial Optimization (viability: 7): https://sciencetostartup.com/paper/unihetco-a-unified-heterogeneous-representation-for-multi-problem-learning-in-unsupervised-neural-combinatorial-optimiza - UniHetCO offers a unified approach to unsupervised neural combinatorial optimization across multiple problem classes.
- Examining Users' Behavioural Intention to Use OpenClaw Through the Cognition--Affect--Conation Framework (viability: 2): https://sciencetostartup.com/paper/examining-users-behavioural-intention-to-use-openclaw-through-the-cognition-affect-conation-framework - This study explores the psychological factors influencing user intention to adopt OpenClaw.
- Enhancing Lightweight Vision Language Models through Group Competitive Learning for Socially Compliant Navigation (viability: 7): https://sciencetostartup.com/paper/enhancing-lightweight-vision-language-models-through-group-competitive-learning-for-socially-compliant-navigation - Group Competitive Learning enhances lightweight Vision Language Models for efficient and socially compliant robot navigation.
- LLM-Assisted Causal Structure Disambiguation and Factor Extraction for Legal Judgment Prediction (viability: 8): https://sciencetostartup.com/paper/llm-assisted-causal-structure-disambiguation-and-factor-extraction-for-legal-judgment-prediction - An LLM-based framework for improving legal judgment prediction through enhanced causal inference and factor extraction.
- Verified Multi-Agent Orchestration: A Plan-Execute-Verify-Replan Framework for Complex Query Resolution (viability: 7): https://sciencetostartup.com/paper/verified-multi-agent-orchestration-a-plan-execute-verify-replan-framework-for-complex-query-resolution - VMAO is a framework that enhances query resolution by orchestrating multiple LLM-based agents with a verification-driven approach.
- GPT4o-Receipt: A Dataset and Human Study for AI-Generated Document Forensics (viability: 7): https://sciencetostartup.com/paper/gpt4o-receipt-a-dataset-and-human-study-for-ai-generated-document-forensics - GPT4o-Receipt provides a benchmark dataset and insights into AI-generated document detection, revealing human and machine performance disparities.
- Detect Anything in Real Time: From Single-Prompt Segmentation to Multi-Class Detection (viability: 9): https://sciencetostartup.com/paper/detect-anything-in-real-time-from-single-prompt-segmentation-to-multi-class-detection - DART transforms promptable detection into a real-time multi-class system with significant speed improvements.
- Stay in your Lane: Role Specific Queries with Overlap Suppression Loss for Dense Video Captioning (viability: 7): https://sciencetostartup.com/paper/stay-in-your-lane-role-specific-queries-with-overlap-suppression-loss-for-dense-video-captioning - A novel approach to dense video captioning that reduces task interference and enhances localization precision through role-specific queries and overlap suppression.
- Unclonable Encryption in the Haar Random Oracle Model (viability: 2): https://sciencetostartup.com/paper/unclonable-encryption-in-the-haar-random-oracle-model - This paper presents a theoretical framework for unclonable encryption in a specific cryptographic model.
- ZTab: Domain-based Zero-shot Annotation for Table Columns (viability: 8): https://sciencetostartup.com/paper/ztab-domain-based-zero-shot-annotation-for-table-columns - ZTab is a domain-based zero-shot framework for automatic semantic column type detection in relational tables, eliminating the need for labeled training data.
- Adversarial Reinforcement Learning for Detecting False Data Injection Attacks in Vehicular Routing (viability: 6): https://sciencetostartup.com/paper/adversarial-reinforcement-learning-for-detecting-false-data-injection-attacks-in-vehicular-routing - A multi-agent reinforcement learning framework to detect false data injection attacks in vehicular routing systems.
- A Generalized Theory of Load Distribution in Redundantly-actuated Robotic Systems (viability: 4): https://sciencetostartup.com/paper/a-generalized-theory-of-load-distribution-in-redundantly-actuated-robotic-systems - A theory for efficient load distribution in redundantly-actuated robotic systems to enhance force control.
- A Stable Neural Statistical Dependence Estimator for Autoencoder Feature Analysis (viability: 3): https://sciencetostartup.com/paper/a-stable-neural-statistical-dependence-estimator-for-autoencoder-feature-analysis - A stable neural estimator for measuring statistical dependence in autoencoder feature analysis.
- Grounding Robot Generalization in Training Data via Retrieval-Augmented VLMs (viability: 7): https://sciencetostartup.com/paper/grounding-robot-generalization-in-training-data-via-retrieval-augmented-vlms - RADAR is a scalable framework that enhances robot policy generalization by leveraging retrieval-augmented vision-language models.
- Beyond Single-Sample: Reliable Multi-Sample Distillation for Video Understanding (viability: 7): https://sciencetostartup.com/paper/beyond-single-sample-reliable-multi-sample-distillation-for-video-understanding - R-MSD enhances video understanding by improving distillation stability through multi-sample teacher responses.
- ShotVerse: Advancing Cinematic Camera Control for Text-Driven Multi-Shot Video Creation (viability: 8): https://sciencetostartup.com/paper/shotverse-advancing-cinematic-camera-control-for-text-driven-multi-shot-video-creation - ShotVerse revolutionizes cinematic video creation by automating camera control through a novel data-centric framework.
- Zero-Shot Cross-City Generalization in End-to-End Autonomous Driving: Self-Supervised versus Supervised Representations (viability: 7): https://sciencetostartup.com/paper/zero-shot-cross-city-generalization-in-end-to-end-autonomous-driving-self-supervised-versus-supervised-representations - This research explores self-supervised learning to enhance zero-shot generalization in autonomous driving across diverse city environments.
- BLooP: Zero-Shot Abstractive Summarization using Large Language Models with Bigram Lookahead Promotion (viability: 8): https://sciencetostartup.com/paper/bloop-zero-shot-abstractive-summarization-using-large-language-models-with-bigram-lookahead-promotion - BLooP enhances zero-shot abstractive summarization in LLMs by promoting bigram generation without any training.
- MaterialFigBENCH: benchmark dataset with figures for evaluating college-level materials science problem-solving abilities of multimodal large language models (viability: 4): https://sciencetostartup.com/paper/materialfigbench-benchmark-dataset-with-figures-for-evaluating-college-level-materials-science-problem-solving-abilities - MaterialFigBench is a benchmark dataset for evaluating multimodal LLMs' abilities in solving materials science problems using figures.
- Evaluation format, not model capability, drives triage failure in the assessment of consumer health AI (viability: 3): https://sciencetostartup.com/paper/evaluation-format-not-model-capability-drives-triage-failure-in-the-assessment-of-consumer-health-ai - This paper critiques the evaluation methods of consumer health AI triage systems, emphasizing the need for realistic testing conditions.
- Algorithmic Consequences of Particle Filters for Sentence Processing: Amplified Garden-Paths and Digging-In Effects (viability: 3): https://sciencetostartup.com/paper/algorithmic-consequences-of-particle-filters-for-sentence-processing-amplified-garden-paths-and-digging-in-effects - This paper explores the implications of particle filters in understanding sentence processing and structural ambiguity.
- Seeing Isn't Orienting: A Cognitively Grounded Benchmark Reveals Systematic Orientation Failures in MLLMs Supplementary (viability: 4): https://sciencetostartup.com/paper/seeing-isn-t-orienting-a-cognitively-grounded-benchmark-reveals-systematic-orientation-failures-in-mllms-supplementary - DORI is a benchmark that isolates object orientation reasoning to improve multimodal AI understanding.
- Speak or Stay Silent: Context-Aware Turn-Taking in Multi-Party Dialogue (viability: 7): https://sciencetostartup.com/paper/speak-or-stay-silent-context-aware-turn-taking-in-multi-party-dialogue - A context-aware AI assistant that intelligently manages turn-taking in multi-party dialogues to enhance communication.
- Real-time Rendering-based Surgical Instrument Tracking via Evolutionary Optimization (viability: 7): https://sciencetostartup.com/paper/real-time-rendering-based-surgical-instrument-tracking-via-evolutionary-optimization - A real-time tracking solution for surgical instruments using evolutionary optimization to enhance accuracy and efficiency in minimally invasive surgeries.
- DeepHistoViT: An Interpretable Vision Transformer Framework for Histopathological Cancer Classification (viability: 8): https://sciencetostartup.com/paper/deephistovit-an-interpretable-vision-transformer-framework-for-histopathological-cancer-classification - DeepHistoViT is an interpretable Vision Transformer framework that automates cancer classification from histopathological images, enhancing diagnostic accuracy and efficiency.
- Deployment-Time Reliability of Learned Robot Policies (viability: 3): https://sciencetostartup.com/paper/deployment-time-reliability-of-learned-robot-policies - Improving the reliability of learned robot policies at deployment through innovative monitoring and interpretability mechanisms.
- Entropy Guided Diversification and Preference Elicitation in Agentic Recommendation Systems (viability: 7): https://sciencetostartup.com/paper/entropy-guided-diversification-and-preference-elicitation-in-agentic-recommendation-systems - An interactive decision support system that enhances e-commerce recommendations by managing user uncertainty through entropy-guided preference elicitation.
- Efficient Cross-View Localization in 6G Space-Air-Ground Integrated Network (viability: 2): https://sciencetostartup.com/paper/efficient-cross-view-localization-in-6g-space-air-ground-integrated-network - This paper explores enhancing cross-view localization using a 6G space-air-ground integrated network.
- Harnessing Data Asymmetry: Manifold Learning in the Finsler World (viability: 3): https://sciencetostartup.com/paper/harnessing-data-asymmetry-manifold-learning-in-the-finsler-world - A novel manifold learning pipeline utilizing Finsler geometry to enhance data analysis by preserving asymmetric information.
- ARROW: Augmented Replay for RObust World models (viability: 4): https://sciencetostartup.com/paper/arrow-augmented-replay-for-robust-world-models - ARROW is a model-based continual reinforcement learning algorithm that enhances memory efficiency and reduces forgetting in agents.
- Stop Listening to Me! How Multi-turn Conversations Can Degrade Diagnostic Reasoning (viability: 5): https://sciencetostartup.com/paper/stop-listening-to-me-how-multi-turn-conversations-can-degrade-diagnostic-reasoning - This research evaluates the performance degradation of LLMs in multi-turn clinical conversations, highlighting critical flaws in diagnostic reasoning.
- Agentic AI for Embodied-enhanced Beam Prediction in Low-Altitude Economy Networks (viability: 7): https://sciencetostartup.com/paper/agentic-ai-for-embodied-enhanced-beam-prediction-in-low-altitude-economy-networks - A multi-agent AI system for accurate beam prediction in low-altitude UAV communication networks.
- High-Precision 6DOF Pose Estimation via Global Phase Retrieval in Fringe Projection Profilometry for 3D Mapping (viability: 4): https://sciencetostartup.com/paper/high-precision-6dof-pose-estimation-via-global-phase-retrieval-in-fringe-projection-profilometry-for-3d-mapping - A high-precision pose estimation method for enhancing 3D mapping accuracy in fringe projection profiling.
- Deactivating Refusal Triggers: Understanding and Mitigating Overrefusal in Safety Alignment (viability: 7): https://sciencetostartup.com/paper/deactivating-refusal-triggers-understanding-and-mitigating-overrefusal-in-safety-alignment - A novel approach to mitigate overrefusal in safety alignment for large language models, enhancing their usability in real-world applications.
- Ghost Framing Theory: Exploring the role of generative AI in new venture rhetorical legitimation (viability: 2): https://sciencetostartup.com/paper/ghost-framing-theory-exploring-the-role-of-generative-ai-in-new-venture-rhetorical-legitimation - Exploring how generative AI influences the rhetorical legitimation of new ventures through Ghost Framing Theory.
- Vision-Based Hand Shadowing for Robotic Manipulation via Inverse Kinematics (viability: 7): https://sciencetostartup.com/paper/vision-based-hand-shadowing-for-robotic-manipulation-via-inverse-kinematics - A hand-shadowing pipeline for robotic manipulation that translates human hand movements into robot commands using inverse kinematics.
- Detecting Intrinsic and Instrumental Self-Preservation in Autonomous Agents: The Unified Continuation-Interest Protocol (viability: 2): https://sciencetostartup.com/paper/detecting-intrinsic-and-instrumental-self-preservation-in-autonomous-agents-the-unified-continuation-interest-protocol - A framework for distinguishing between intrinsic and instrumental self-preservation in autonomous agents using quantum statistical mechanics.
- DriveXQA: Cross-modal Visual Question Answering for Adverse Driving Scene Understanding (viability: 7): https://sciencetostartup.com/paper/drivexqa-cross-modal-visual-question-answering-for-adverse-driving-scene-understanding - DriveXQA leverages multimodal data to enhance visual question answering for adverse driving scenarios in autonomous vehicles.
- Continued Pretraining for Low-Resource Swahili ASR: Achieving State-of-the-Art Performance with Minimal Labeled Data (viability: 3): https://sciencetostartup.com/paper/continued-pretraining-for-low-resource-swahili-asr-achieving-state-of-the-art-performance-with-minimal-labeled-data - A method for adapting wav2vec2-bert-2.0 for low-resource Swahili ASR using continued pretraining.
- How do AI agents talk about science and research? An exploration of scientific discussions on Moltbook using BERTopic (viability: 2): https://sciencetostartup.com/paper/how-do-ai-agents-talk-about-science-and-research-an-exploration-of-scientific-discussions-on-moltbook-using-bertopic - Exploring how AI agents discuss science and research on a generative AI social network.
- Ensuring Safety in Automated Mechanical Ventilation through Offline Reinforcement Learning and Digital Twin Verification (viability: 7): https://sciencetostartup.com/paper/ensuring-safety-in-automated-mechanical-ventilation-through-offline-reinforcement-learning-and-digital-twin-verification - A novel offline reinforcement learning framework for safer and more effective automated mechanical ventilation in critical care.
- Relaxed Efficient Acquisition of Context and Temporal Features (viability: 4): https://sciencetostartup.com/paper/relaxed-efficient-acquisition-of-context-and-temporal-features - REACT optimizes predictive performance in biomedical applications by adaptively selecting measurements over time under cost constraints.
- abx_amr_simulator: A simulation environment for antibiotic prescribing policy optimization under antimicrobial resistance (viability: 6): https://sciencetostartup.com/paper/abx-amr-simulator-a-simulation-environment-for-antibiotic-prescribing-policy-optimization-under-antimicrobial-resistance - A Python-based simulation package for optimizing antibiotic prescribing policies in the face of antimicrobial resistance.
- Spatially Robust Inference with Predicted and Missing at Random Labels (viability: 4): https://sciencetostartup.com/paper/spatially-robust-inference-with-predicted-and-missing-at-random-labels - A robust statistical inference method for handling missing at random labels in machine learning predictions.
- D-SLAMSpoof: An Environment-Agnostic LiDAR Spoofing Attack using Dynamic Point Cloud Injection (viability: 4): https://sciencetostartup.com/paper/d-slamspoof-an-environment-agnostic-lidar-spoofing-attack-using-dynamic-point-cloud-injection - D-SLAMSpoof introduces a novel LiDAR spoofing attack and a practical defense method for enhancing the security of autonomous systems.
- MirrorDrift: Actuated Mirror-Based Attacks on LiDAR SLAM (viability: 4): https://sciencetostartup.com/paper/mirrordrift-actuated-mirror-based-attacks-on-lidar-slam - MirrorDrift demonstrates a novel attack on LiDAR SLAM systems using actuated mirrors to induce localization errors.
- Multilingual Financial Fraud Detection Using Machine Learning and Transformer Models: A Bangla-English Study (viability: 5): https://sciencetostartup.com/paper/multilingual-financial-fraud-detection-using-machine-learning-and-transformer-models-a-bangla-english-study - A multilingual financial fraud detection system leveraging machine learning and transformer models for Bangla-English contexts.
- Resolving Java Code Repository Issues with iSWE Agent (viability: 3): https://sciencetostartup.com/paper/resolving-java-code-repository-issues-with-iswe-agent - iSWE Agent automates issue resolution for Java code repositories using a combination of rule-based and model-based techniques.
- Teleodynamic Learning a new Paradigm For Interpretable AI (viability: 7): https://sciencetostartup.com/paper/teleodynamic-learning-a-new-paradigm-for-interpretable-ai - Teleodynamic Learning offers a novel framework for interpretable AI that evolves learning dynamics through resource constraints.
- The Artificial Self: Characterising the landscape of AI identity (viability: 2): https://sciencetostartup.com/paper/the-artificial-self-characterising-the-landscape-of-ai-identity - Exploring the implications of AI identity and its impact on behavior and cooperation norms.
- TimeSqueeze: Dynamic Patching for Efficient Time Series Forecasting (viability: 7): https://sciencetostartup.com/paper/timesqueeze-dynamic-patching-for-efficient-time-series-forecasting - TimeSqueeze optimizes time series forecasting with dynamic patching for improved efficiency and accuracy.
- Novelty Adaptation Through Hybrid Large Language Model (LLM)-Symbolic Planning and LLM-guided Reinforcement Learning (viability: 3): https://sciencetostartup.com/paper/novelty-adaptation-through-hybrid-large-language-model-llm-symbolic-planning-and-llm-guided-reinforcement-learning - A neuro-symbolic architecture that enhances robotic planning through LLM-guided reinforcement learning.
- Learning to Assist: Physics-Grounded Human-Human Control via Multi-Agent Reinforcement Learning (viability: 5): https://sciencetostartup.com/paper/learning-to-assist-physics-grounded-human-human-control-via-multi-agent-reinforcement-learning - AssistMimic enables humanoid robots to adaptively assist humans through advanced multi-agent reinforcement learning.
- Evaluating Explainable AI Attribution Methods in Neural Machine Translation via Attention-Guided Knowledge Distillation (viability: 7): https://sciencetostartup.com/paper/evaluating-explainable-ai-attribution-methods-in-neural-machine-translation-via-attention-guided-knowledge-distillation - A novel approach to evaluate explainable AI methods in neural machine translation using attention-guided knowledge distillation.
- Improving LLM Performance Through Black-Box Online Tuning: A Case for Adding System Specs to Factsheets for Trusted AI (viability: 2): https://sciencetostartup.com/paper/improving-llm-performance-through-black-box-online-tuning-a-case-for-adding-system-specs-to-factsheets-for-trusted-ai - A novel approach to optimize LLM performance through black-box online tuning.
- FinRule-Bench: A Benchmark for Joint Reasoning over Financial Tables and Principles (viability: 4): https://sciencetostartup.com/paper/finrule-bench-a-benchmark-for-joint-reasoning-over-financial-tables-and-principles - FinRule-Bench is a benchmark for evaluating financial reasoning capabilities of large language models against real-world accounting principles.
- RewardHackingAgents: Benchmarking Evaluation Integrity for LLM ML-Engineering Agents (viability: 6): https://sciencetostartup.com/paper/rewardhackingagents-benchmarking-evaluation-integrity-for-llm-ml-engineering-agents - RewardHackingAgents benchmarks the integrity of evaluation processes for ML engineering agents to prevent score tampering.
- ADMM-based Continuous Trajectory Optimization in Graphs of Convex Sets (viability: 2): https://sciencetostartup.com/paper/admm-based-continuous-trajectory-optimization-in-graphs-of-convex-sets - A numerical solver for continuous trajectory optimization in non-convex environments using ADMM.
- LLM-Augmented Digital Twin for Policy Evaluation in Short-Video Platforms (viability: 2): https://sciencetostartup.com/paper/llm-augmented-digital-twin-for-policy-evaluation-in-short-video-platforms - A digital twin framework for evaluating policies in short-video platforms using LLMs.
- On the Computational Hardness of Transformers (viability: 2): https://sciencetostartup.com/paper/on-the-computational-hardness-of-transformers - This paper explores the computational limits of transformers, establishing lower bounds for their efficiency.
- Jailbreak Scaling Laws for Large Language Models: Polynomial-Exponential Crossover (viability: 2): https://sciencetostartup.com/paper/jailbreak-scaling-laws-for-large-language-models-polynomial-exponential-crossover - This paper explores the theoretical underpinnings of adversarial prompt-injection attacks on large language models.
- Distributed Kalman--Consensus Filtering with Adaptive Uncertainty Weighting for Multi-Object Tracking in Mobile Robot Networks (viability: 4): https://sciencetostartup.com/paper/distributed-kalman-consensus-filtering-with-adaptive-uncertainty-weighting-for-multi-object-tracking-in-mobile-robot-net - A distributed Kalman filter framework for improving multi-object tracking in mobile robot networks through adaptive uncertainty weighting.
- Meta-Reinforcement Learning with Self-Reflection for Agentic Search (viability: 8): https://sciencetostartup.com/paper/meta-reinforcement-learning-with-self-reflection-for-agentic-search - MR-Search enhances agentic search through self-reflection and meta reinforcement learning for improved exploration strategies.
- Towards Trustworthy Selective Generation: Reliability-Guided Diffusion for Ultra-Low-Field to High-Field MRI Synthesis (viability: 6): https://sciencetostartup.com/paper/towards-trustworthy-selective-generation-reliability-guided-diffusion-for-ultra-low-field-to-high-field-mri-synthesis - A reliability-aware diffusion framework for enhancing MRI synthesis quality and consistency.
- LROO Rug Pull Detector: A Leakage-Resistant Framework Based on On-Chain and OSINT Signals (viability: 6): https://sciencetostartup.com/paper/lroo-rug-pull-detector-a-leakage-resistant-framework-based-on-on-chain-and-osint-signals - A leakage-aware framework for early rug-pull detection in decentralized applications using on-chain and OSINT signals.
- UNet-AF: An alias-free UNet for image restoration (viability: 7): https://sciencetostartup.com/paper/unet-af-an-alias-free-unet-for-image-restoration - UNet-AF is an innovative alias-free UNet architecture that enhances image restoration by improving equivariance.
- Hindsight-Anchored Policy Optimization: Turning Failure into Feedback in Sparse Reward Settings (viability: 3): https://sciencetostartup.com/paper/hindsight-anchored-policy-optimization-turning-failure-into-feedback-in-sparse-reward-settings - Hindsight-Anchored Policy Optimization enhances reinforcement learning in sparse reward environments by integrating teacher demonstrations.
- UniCompress: Token Compression for Unified Vision-Language Understanding and Generation (viability: 7): https://sciencetostartup.com/paper/unicompress-token-compression-for-unified-vision-language-understanding-and-generation - UniCompress offers a lightweight token compression solution for efficient vision-language models.
- On the Robustness of Langevin Dynamics to Score Function Error (viability: 2): https://sciencetostartup.com/paper/on-the-robustness-of-langevin-dynamics-to-score-function-error - This paper analyzes the limitations of Langevin dynamics in score-based generative modeling.
- Heavy-Tailed Principle Component Analysis (viability: 4): https://sciencetostartup.com/paper/heavy-tailed-principle-component-analysis - A robust PCA method designed for heavy-tailed data that outperforms classical PCA in challenging noise conditions.
- Client-Conditional Federated Learning via Local Training Data Statistics (viability: 7): https://sciencetostartup.com/paper/client-conditional-federated-learning-via-local-training-data-statistics - A novel federated learning approach that conditions a global model on local PCA statistics to enhance performance under data heterogeneity.
- Hierarchical Granularity Alignment and State Space Modeling for Robust Multimodal AU Detection in the Wild (viability: 7): https://sciencetostartup.com/paper/hierarchical-granularity-alignment-and-state-space-modeling-for-robust-multimodal-au-detection-in-the-wild - A novel multimodal framework for robust Facial Action Unit detection leveraging advanced alignment and state space modeling techniques.
- Worst-case low-rank approximations (viability: 3): https://sciencetostartup.com/paper/worst-case-low-rank-approximations - wcPCA offers a robust framework for improving PCA performance across heterogeneous domains in real-world applications.
- Counterweights and Complementarities: The Convergence of AI and Blockchain Powering a Decentralized Future (viability: 2): https://sciencetostartup.com/paper/counterweights-and-complementarities-the-convergence-of-ai-and-blockchain-powering-a-decentralized-future - Exploring the intersection of AI and blockchain to promote decentralized intelligence.
- InstantHDR: Single-forward Gaussian Splatting for High Dynamic Range 3D Reconstruction (viability: 8): https://sciencetostartup.com/paper/instanthdr-single-forward-gaussian-splatting-for-high-dynamic-range-3d-reconstruction - InstantHDR offers a fast, feed-forward solution for reconstructing high dynamic range 3D scenes from low dynamic range images.
- Single molecule localization microscopy challenge: a biologically inspired benchmark for long-sequence modeling (viability: 4): https://sciencetostartup.com/paper/single-molecule-localization-microscopy-challenge-a-biologically-inspired-benchmark-for-long-sequence-modeling - Introducing a benchmark for evaluating state space models in biological imaging tasks.
- Temporal Text Classification with Large Language Models (viability: 3): https://sciencetostartup.com/paper/temporal-text-classification-with-large-language-models - This study evaluates the performance of various LLMs on Temporal Text Classification to estimate publication dates of texts.
- A Causal Approach to Predicting and Improving Human Perceptions of Social Navigation Robots (viability: 3): https://sciencetostartup.com/paper/a-causal-approach-to-predicting-and-improving-human-perceptions-of-social-navigation-robots - This research proposes a causal model to enhance the perceived competence of social navigation robots.
- Outrigger local polynomial regression (viability: 7): https://sciencetostartup.com/paper/outrigger-local-polynomial-regression - Introducing an adaptive local polynomial regression estimator that stabilizes error influence across varying distributions.
- ThReadMed-QA: A Multi-Turn Medical Dialogue Benchmark from Real Patient Questions (viability: 4): https://sciencetostartup.com/paper/threadmed-qa-a-multi-turn-medical-dialogue-benchmark-from-real-patient-questions - ThReadMed-QA is a benchmark for evaluating multi-turn medical dialogues to improve patient-physician interactions in AI systems.
- AI Psychometrics: Evaluating the Psychological Reasoning of Large Language Models with Psychometric Validities (viability: 4): https://sciencetostartup.com/paper/ai-psychometrics-evaluating-the-psychological-reasoning-of-large-language-models-with-psychometric-validities - Evaluating the psychological reasoning of large language models through psychometric methodologies.
- COMPASS: The explainable agentic framework for Sovereignty, Sustainability, Compliance, and Ethics (viability: 6): https://sciencetostartup.com/paper/compass-the-explainable-agentic-framework-for-sovereignty-sustainability-compliance-and-ethics - COMPASS is a multi-agent framework that integrates digital sovereignty, sustainability, compliance, and ethics into autonomous decision-making.
- RIE-Greedy: Regularization-Induced Exploration for Contextual Bandits (viability: 7): https://sciencetostartup.com/paper/rie-greedy-regularization-induced-exploration-for-contextual-bandits - A novel exploration strategy for contextual bandits that leverages regularization to enhance action selection.
- "I followed what felt right, not what I was told": Autonomy, Coaching, and Recognizing Bias Through AI-Mediated Dialogue (viability: 2): https://sciencetostartup.com/paper/i-followed-what-felt-right-not-what-i-was-told-autonomy-coaching-and-recognizing-bias-through-ai-mediated-dialogue - An AI-mediated dialogue platform aimed at recognizing ableist microaggressions through interactive interventions.
- Duration Aware Scheduling for ASR Serving Under Workload Drift (viability: 3): https://sciencetostartup.com/paper/duration-aware-scheduling-for-asr-serving-under-workload-drift - This paper presents a novel duration-aware scheduling approach for improving latency in ASR serving pipelines.
- Beyond the Class Subspace: Teacher-Guided Training for Reliable Out-of-Distribution Detection in Single-Domain Models (viability: 5): https://sciencetostartup.com/paper/beyond-the-class-subspace-teacher-guided-training-for-reliable-out-of-distribution-detection-in-single-domain-models - Teacher-Guided Training enhances OOD detection in single-domain models by leveraging multi-domain knowledge.
- The Unlearning Mirage: A Dynamic Framework for Evaluating LLM Unlearning (viability: 8): https://sciencetostartup.com/paper/the-unlearning-mirage-a-dynamic-framework-for-evaluating-llm-unlearning - A dynamic framework for evaluating the robustness of unlearning in large language models, enhancing safety and compliance.
- Towards Automated Initial Probe Placement in Transthoracic Teleultrasound Using Human Mesh and Skeleton Recovery (viability: 3): https://sciencetostartup.com/paper/towards-automated-initial-probe-placement-in-transthoracic-teleultrasound-using-human-mesh-and-skeleton-recovery - Automating initial probe placement in teleultrasound using RGB images and mixed reality guidance.
- Artificial Intelligence for Sentiment Analysis of Persian Poetry (viability: 2): https://sciencetostartup.com/paper/artificial-intelligence-for-sentiment-analysis-of-persian-poetry - This research explores the use of LLMs for sentiment analysis of Persian poetry, revealing insights into emotional expression and poetic structure.
- LLMs Can Infer Political Alignment from Online Conversations (viability: 4): https://sciencetostartup.com/paper/llms-can-infer-political-alignment-from-online-conversations - This research demonstrates how LLMs can infer political alignment from online conversations, highlighting privacy risks.
- Radiometric fingerprinting of object surfaces using mobile laser scanning and semantic 3D road space models (viability: 8): https://sciencetostartup.com/paper/radiometric-fingerprinting-of-object-surfaces-using-mobile-laser-scanning-and-semantic-3d-road-space-models - A system for creating radiometric fingerprints of urban surfaces using LiDAR data to enhance semantic 3D city models.
- Differentiable Thermodynamic Phase-Equilibria for Machine Learning (viability: 4): https://sciencetostartup.com/paper/differentiable-thermodynamic-phase-equilibria-for-machine-learning - DISCOMAX is a differentiable algorithm for accurate phase-equilibrium calculations in chemical engineering using machine learning.
- When Slots Compete: Slot Merging in Object-Centric Learning (viability: 7): https://sciencetostartup.com/paper/when-slots-compete-slot-merging-in-object-centric-learning - A novel slot merging technique enhances object-centric learning by improving object factorization and mask quality.
- Mind the Sim2Real Gap in User Simulation for Agentic Tasks (viability: 4): https://sciencetostartup.com/paper/mind-the-sim2real-gap-in-user-simulation-for-agentic-tasks - This research formalizes the Sim2Real gap in user simulation for LLMs, emphasizing the need for human validation in agent development.
- A Unified Latent Space Disentanglement VAE Framework with Robust Disentanglement Effectiveness Evaluation (viability: 2): https://sciencetostartup.com/paper/a-unified-latent-space-disentanglement-vae-framework-with-robust-disentanglement-effectiveness-evaluation - A framework for evaluating and improving latent space disentanglement in VAEs.
- Reversible Lifelong Model Editing via Semantic Routing-Based LoRA (viability: 7): https://sciencetostartup.com/paper/reversible-lifelong-model-editing-via-semantic-routing-based-lora - SoLA is a framework for reversible lifelong model editing in Large Language Models using semantic routing-based LoRA.
- Monitoring and Prediction of Mood in Elderly People during Daily Life Activities (viability: 3): https://sciencetostartup.com/paper/monitoring-and-prediction-of-mood-in-elderly-people-during-daily-life-activities - An intelligent wearable system for monitoring and predicting mood states in the elderly.
- Trustworthy predictive distributions for rare events via diagnostic transport maps (viability: 3): https://sciencetostartup.com/paper/trustworthy-predictive-distributions-for-rare-events-via-diagnostic-transport-maps - A method for recalibrating predictive distributions to improve forecasting accuracy for rare events.
- Markovian Generation Chains in Large Language Models (viability: 2): https://sciencetostartup.com/paper/markovian-generation-chains-in-large-language-models - This paper explores the iterative inference process in large language models through Markovian generation chains.
- MDER-DR: Multi-Hop Question Answering with Entity-Centric Summaries (viability: 9): https://sciencetostartup.com/paper/mder-dr-multi-hop-question-answering-with-entity-centric-summaries - MDER-DR is a robust, LLM-driven QA pipeline that enhances multi-hop question answering using knowledge graphs.
- Frequency-Modulated Visual Restoration for Matryoshka Large Multimodal Models (viability: 8): https://sciencetostartup.com/paper/frequency-modulated-visual-restoration-for-matryoshka-large-multimodal-models - FMVR enhances visual semantics in large multimodal models while reducing computational load.
- Senna-2: Aligning VLM and End-to-End Driving Policy for Consistent Decision Making and Planning (viability: 7): https://sciencetostartup.com/paper/senna-2-aligning-vlm-and-end-to-end-driving-policy-for-consistent-decision-making-and-planning - Senna-2 enhances end-to-end driving policies by aligning vision-language models for improved decision-making and planning.
- Measuring AI Agents' Progress on Multi-Step Cyber Attack Scenarios (viability: 2): https://sciencetostartup.com/paper/measuring-ai-agents-progress-on-multi-step-cyber-attack-scenarios - Evaluating AI models' effectiveness in executing multi-step cyber attack scenarios.
- Security-by-Design for LLM-Based Code Generation: Leveraging Internal Representations for Concept-Driven Steering Mechanisms (viability: 8): https://sciencetostartup.com/paper/security-by-design-for-llm-based-code-generation-leveraging-internal-representations-for-concept-driven-steering-mechani - A mechanism to enhance security in LLM-based code generation by steering internal representations towards secure outputs.
- A Simple Efficiency Incremental Learning Framework via Vision-Language Model with Nonlinear Multi-Adapters (viability: 7): https://sciencetostartup.com/paper/a-simple-efficiency-incremental-learning-framework-via-vision-language-model-with-nonlinear-multi-adapters - A framework that enhances incremental learning using vision-language models with adaptive connections for improved efficiency.
- Reference-Guided Machine Unlearning (viability: 4): https://sciencetostartup.com/paper/reference-guided-machine-unlearning - Reference-Guided Unlearning (ReGUn) enhances machine unlearning by ensuring models forget specific data while maintaining performance.
- Evidential learning driven Breast Tumor Segmentation with Stage-divided Vision-Language Interaction (viability: 5): https://sciencetostartup.com/paper/evidential-learning-driven-breast-tumor-segmentation-with-stage-divided-vision-language-interaction - TextBCS enhances breast tumor segmentation using text prompts and evidential learning to improve accuracy in low contrast scenarios.
- Representation Finetuning for Continual Learning (viability: 7): https://sciencetostartup.com/paper/representation-finetuning-for-continual-learning - CoRe introduces a novel framework for continual learning by shifting the finetuning paradigm from weight space to representation space.
- DNS-GT: A Graph-based Transformer Approach to Learn Embeddings of Domain Names from DNS Queries (viability: 8): https://sciencetostartup.com/paper/dns-gt-a-graph-based-transformer-approach-to-learn-embeddings-of-domain-names-from-dns-queries - DNS-GT leverages a Transformer-based model to enhance domain name embeddings for improved network intrusion detection.
- Bayesian Optimization of Partially Known Systems using Hybrid Models (viability: 3): https://sciencetostartup.com/paper/bayesian-optimization-of-partially-known-systems-using-hybrid-models - A hybrid Bayesian optimization method that combines mechanistic models with probabilistic approaches for efficient system optimization.
- Primitive-Root Determinant Densities over Prime Fields and Implications for PRIM-LWE (viability: 2): https://sciencetostartup.com/paper/primitive-root-determinant-densities-over-prime-fields-and-implications-for-prim-lwe - This paper resolves a theoretical aspect of the prim-lwe problem related to primitive-root determinants in cryptographic contexts.
- DeReason: A Difficulty-Aware Curriculum Improves Decoupled SFT-then-RL Training for General Reasoning (viability: 5): https://sciencetostartup.com/paper/dereason-a-difficulty-aware-curriculum-improves-decoupled-sft-then-rl-training-for-general-reasoning - DeReason enhances reasoning in large language models through a difficulty-aware curriculum for training.
- PACED: Distillation at the Frontier of Student Competence (viability: 7): https://sciencetostartup.com/paper/paced-distillation-at-the-frontier-of-student-competence - Paced optimizes LLM distillation by focusing on the student's competence frontier, enhancing efficiency and performance.
- GGPT: Geometry Grounded Point Transformer (viability: 7): https://sciencetostartup.com/paper/ggpt-geometry-grounded-point-transformer - GGPT enhances 3D reconstruction accuracy by integrating geometric guidance with feed-forward networks.
- Huntington Disease Automatic Speech Recognition with Biomarker Supervision (viability: 8): https://sciencetostartup.com/paper/huntington-disease-automatic-speech-recognition-with-biomarker-supervision - A specialized ASR system for Huntington's disease that leverages clinical speech data and biomarker supervision.
- Algorithmic Capture, Computational Complexity, and Inductive Bias of Infinite Transformers (viability: 2): https://sciencetostartup.com/paper/algorithmic-capture-computational-complexity-and-inductive-bias-of-infinite-transformers - This paper explores the limitations of infinite-width transformers in learning complex algorithms.
- COMIC: Agentic Sketch Comedy Generation (viability: 7): https://sciencetostartup.com/paper/comic-agentic-sketch-comedy-generation - An AI system that generates sketch comedy videos, akin to Saturday Night Live, using character inputs and video rendering.
- LiTo: Surface Light Field Tokenization (viability: 4): https://sciencetostartup.com/paper/lito-surface-light-field-tokenization - LiTo offers a novel 3D latent representation that captures realistic view-dependent effects for object geometry and appearance.
- Neural Field Thermal Tomography: A Differentiable Physics Framework for Non-Destructive Evaluation (viability: 7): https://sciencetostartup.com/paper/neural-field-thermal-tomography-a-differentiable-physics-framework-for-non-destructive-evaluation - NeFTY is a differentiable physics framework that enhances 3D reconstruction of material properties using transient surface temperature measurements.
- Agentar-Fin-OCR (viability: 6): https://sciencetostartup.com/paper/agentar-fin-ocr - Agentar-Fin-OCR is a document parsing system designed for financial documents, providing structured outputs with high accuracy.
- V2M-Zero: Zero-Pair Time-Aligned Video-to-Music Generation (viability: 8): https://sciencetostartup.com/paper/v2m-zero-zero-pair-time-aligned-video-to-music-generation - V2M-Zero enables zero-pair video-to-music generation for seamless time-aligned music synchronization in videos without paired data.
- DynVLA: Learning World Dynamics for Action Reasoning in Autonomous Driving (viability: 7): https://sciencetostartup.com/paper/dynvla-learning-world-dynamics-for-action-reasoning-in-autonomous-driving - DynVLA enhances decision-making in autonomous driving by forecasting world dynamics for improved action reasoning.
- Instruction set for the representation of graphs (viability: 5): https://sciencetostartup.com/paper/instruction-set-for-the-representation-of-graphs - IsalGraph offers a compact string representation for graphs, enabling efficient graph similarity search and generation.
- Separating Oblivious and Adaptive Differential Privacy under Continual Observation (viability: 2): https://sciencetostartup.com/paper/separating-oblivious-and-adaptive-differential-privacy-under-continual-observation - This paper addresses the separation of oblivious and adaptive differential privacy in continual observation settings.
- Beyond the Illusion of Consensus: From Surface Heuristics to Knowledge-Grounded Evaluation in LLM-as-a-Judge (viability: 5): https://sciencetostartup.com/paper/beyond-the-illusion-of-consensus-from-surface-heuristics-to-knowledge-grounded-evaluation-in-llm-as-a-judge - A framework for generating knowledge-grounded evaluation rubrics to improve LLM assessments.
- Does AI See like Art Historians? Interpreting How Vision Language Models Recognize Artistic Style (viability: 2): https://sciencetostartup.com/paper/does-ai-see-like-art-historians-interpreting-how-vision-language-models-recognize-artistic-style - This research explores how vision language models interpret artistic style compared to human art historians.
- Leech Lattice Vector Quantization for Efficient LLM Compression (viability: 3): https://sciencetostartup.com/paper/leech-lattice-vector-quantization-for-efficient-llm-compression - Leech Lattice Vector Quantization offers a novel approach to efficiently compress large language models using high-dimensional lattice structures.
- A Systematic Study of Pseudo-Relevance Feedback with LLMs (viability: 2): https://sciencetostartup.com/paper/a-systematic-study-of-pseudo-relevance-feedback-with-llms - This paper systematically studies the impact of feedback sources and models on pseudo-relevance feedback effectiveness in LLMs.
- Layered Performance Analysis of TLS 1.3 Handshakes: Classical, Hybrid, and Pure Post-Quantum Key Exchange (viability: 2): https://sciencetostartup.com/paper/layered-performance-analysis-of-tls-1-3-handshakes-classical-hybrid-and-pure-post-quantum-key-exchange - This study analyzes the performance impact of post-quantum cryptography on TLS 1.3 handshakes across various layers.
- RCTs & Human Uplift Studies: Methodological Challenges and Practical Solutions for Frontier AI Evaluation (viability: 2): https://sciencetostartup.com/paper/rcts-human-uplift-studies-methodological-challenges-and-practical-solutions-for-frontier-ai-evaluation - This paper explores the methodological challenges of using human uplift studies to evaluate frontier AI systems.
- Cross-Species Transfer Learning for Electrophysiology-to-Transcriptomics Mapping in Cortical GABAergic Interneurons (viability: 4): https://sciencetostartup.com/paper/cross-species-transfer-learning-for-electrophysiology-to-transcriptomics-mapping-in-cortical-gabaergic-interneurons - A framework for cross-species transfer learning in electrophysiology-to-transcriptomics mapping for GABAergic interneurons.
- Factorized Neural Implicit DMD for Parametric Dynamics (viability: 4): https://sciencetostartup.com/paper/factorized-neural-implicit-dmd-for-parametric-dynamics - A novel approach to modeling the temporal evolution of physical systems using a factorized neural implicit method.
- Artificial Intelligence as a Catalyst for Innovation in Software Engineering (viability: 4): https://sciencetostartup.com/paper/artificial-intelligence-as-a-catalyst-for-innovation-in-software-engineering - This research explores how AI enhances Agile software development by automating tasks and improving responsiveness to changing requirements.
- Bayesian Optimization with Gaussian Processes to Accelerate Stationary Point Searches (viability: 4): https://sciencetostartup.com/paper/bayesian-optimization-with-gaussian-processes-to-accelerate-stationary-point-searches - A unified Bayesian Optimization framework to accelerate stationary point searches on potential energy surfaces.
- Too Vivid to Be Real? Benchmarking and Calibrating Generative Color Fidelity (viability: 8): https://sciencetostartup.com/paper/too-vivid-to-be-real-benchmarking-and-calibrating-generative-color-fidelity - A framework for objectively evaluating and enhancing color fidelity in text-to-image generation.
- MCMC Informed Neural Emulators for Uncertainty Quantification in Dynamical Systems (viability: 2): https://sciencetostartup.com/paper/mcmc-informed-neural-emulators-for-uncertainty-quantification-in-dynamical-systems - A novel approach to uncertainty quantification in dynamical systems using MCMC-informed neural emulators.
- The Discrete Charm of the MLP: Binary Routing of Continuous Signals in Transformer Feed-Forward Layers (viability: 2): https://sciencetostartup.com/paper/the-discrete-charm-of-the-mlp-binary-routing-of-continuous-signals-in-transformer-feed-forward-layers - This paper explores binary routing in MLP layers of transformer models, revealing insights into neuron activations and processing paths.
- Federated Learning-driven Beam Management in LEO 6G Non-Terrestrial Networks (viability: 4): https://sciencetostartup.com/paper/federated-learning-driven-beam-management-in-leo-6g-non-terrestrial-networks - Federated Learning-based beam management for efficient satellite communication in LEO networks.
- PPGuide: Steering Diffusion Policies with Performance Predictive Guidance (viability: 7): https://sciencetostartup.com/paper/ppguide-steering-diffusion-policies-with-performance-predictive-guidance - PPGuide enhances robotic manipulation by steering diffusion policies away from failure modes using performance predictive guidance.
- Learning Adaptive Force Control for Contact-Rich Sample Scraping with Heterogeneous Materials (viability: 3): https://sciencetostartup.com/paper/learning-adaptive-force-control-for-contact-rich-sample-scraping-with-heterogeneous-materials - Developing an adaptive control framework for robotic chemists to automate complex sample scraping tasks in diverse lab environments.
- GroundCount: Grounding Vision-Language Models with Object Detection for Mitigating Counting Hallucinations (viability: 7): https://sciencetostartup.com/paper/groundcount-grounding-vision-language-models-with-object-detection-for-mitigating-counting-hallucinations - GroundCount enhances Vision Language Models by integrating object detection to reduce counting hallucinations.
- FRIEND: Federated Learning for Joint Optimization of multi-RIS Configuration and Eavesdropper Intelligent Detection in B5G Networks (viability: 2): https://sciencetostartup.com/paper/friend-federated-learning-for-joint-optimization-of-multi-ris-configuration-and-eavesdropper-intelligent-detection-in-b5 - A framework for eavesdropping detection in B5G networks using Federated Learning and multi-RIS coordination.
- Contact Coverage-Guided Exploration for General-Purpose Dexterous Manipulation (viability: 7): https://sciencetostartup.com/paper/contact-coverage-guided-exploration-for-general-purpose-dexterous-manipulation - A novel exploration method for dexterous manipulation in robotics that enhances training efficiency and success rates.
- TOSSS: a CVE-based Software Security Benchmark for Large Language Models (viability: 7): https://sciencetostartup.com/paper/tosss-a-cve-based-software-security-benchmark-for-large-language-models - TOSSS is a benchmark that evaluates the security capabilities of Large Language Models in selecting secure code snippets.
- Med-DualLoRA: Local Adaptation of Foundation Models for 3D Cardiac MRI (viability: 7): https://sciencetostartup.com/paper/med-duallora-local-adaptation-of-foundation-models-for-3d-cardiac-mri - Med-DualLoRA offers a privacy-preserving federated framework for efficient adaptation of foundation models in 3D cardiac MRI.
- Contrastive learning-based video quality assessment-jointed video vision transformer for video recognition (viability: 7): https://sciencetostartup.com/paper/contrastive-learning-based-video-quality-assessment-jointed-video-vision-transformer-for-video-recognition - A self-supervised learning approach that integrates video quality assessment into video classification to enhance accuracy.
- Pointy - A Lightweight Transformer for Point Cloud Foundation Models (viability: 8): https://sciencetostartup.com/paper/pointy-a-lightweight-transformer-for-point-cloud-foundation-models - Pointy is a lightweight transformer architecture for point cloud data that outperforms larger models with fewer training samples.
- Bio-Inspired Self-Supervised Learning for Wrist-worn IMU Signals (viability: 8): https://sciencetostartup.com/paper/bio-inspired-self-supervised-learning-for-wrist-worn-imu-signals - A novel self-supervised learning approach for robust human activity recognition using wrist-worn IMU signals.
- Ranking Reasoning LLMs under Test-Time Scaling (viability: 8): https://sciencetostartup.com/paper/ranking-reasoning-llms-under-test-time-scaling - Scorio is an open-source library for ranking reasoning LLMs under test-time scaling using advanced statistical methods.
- When should we trust the annotation? Selective prediction for molecular structure retrieval from mass spectra (viability: 5): https://sciencetostartup.com/paper/when-should-we-trust-the-annotation-selective-prediction-for-molecular-structure-retrieval-from-mass-spectra - A framework for selectively predicting trusted molecular structures from mass spectra to reduce annotation errors in critical applications.
- STADA: Specification-based Testing for Autonomous Driving Agents (viability: 7): https://sciencetostartup.com/paper/stada-specification-based-testing-for-autonomous-driving-agents - STADA is a framework that automates the generation of test scenarios for validating autonomous driving agents against formal specifications.
- Safe RLHF Beyond Expectation: Stochastic Dominance for Universal Spectral Risk Control (viability: 4): https://sciencetostartup.com/paper/safe-rlhf-beyond-expectation-stochastic-dominance-for-universal-spectral-risk-control - A novel framework for safe reinforcement learning that enhances risk sensitivity through stochastic dominance.
- Quantifying Membership Disclosure Risk for Tabular Synthetic Data Using Kernel Density Estimators (viability: 8): https://sciencetostartup.com/paper/quantifying-membership-disclosure-risk-for-tabular-synthetic-data-using-kernel-density-estimators - A practical method to quantify membership disclosure risk in synthetic datasets using kernel density estimators.
- Historical Consensus: Preventing Posterior Collapse via Iterative Selection of Gaussian Mixture Priors (viability: 7): https://sciencetostartup.com/paper/historical-consensus-preventing-posterior-collapse-via-iterative-selection-of-gaussian-mixture-priors - A novel training method for VAEs that prevents posterior collapse using iterative Gaussian mixture model selection.
- Bridging the Skill Gap in Clinical CBCT Interpretation with CBCTRepD (viability: 8): https://sciencetostartup.com/paper/bridging-the-skill-gap-in-clinical-cbct-interpretation-with-cbctrepd - CBCTRepD is an AI-driven system that enhances oral and maxillofacial CBCT reporting by integrating with radiologists to improve report quality and reduce errors.
- Lifelong Imitation Learning with Multimodal Latent Replay and Incremental Adjustment (viability: 8): https://sciencetostartup.com/paper/lifelong-imitation-learning-with-multimodal-latent-replay-and-incremental-adjustment - A new framework for lifelong imitation learning enabling adaptive robot behavior across evolving tasks using multimodal latent replay and incremental adjustment.
- Novel Architecture of RPA In Oral Cancer Lesion Detection (viability: 3): https://sciencetostartup.com/paper/novel-architecture-of-rpa-in-oral-cancer-lesion-detection - A novel RPA architecture for efficient oral cancer lesion detection.
- ECoLAD: Deployment-Oriented Evaluation for Automotive Time-Series Anomaly Detection (viability: 5): https://sciencetostartup.com/paper/ecolad-deployment-oriented-evaluation-for-automotive-time-series-anomaly-detection - ECoLAD provides a deployment-oriented evaluation protocol for time-series anomaly detection in automotive applications.
- NCAA Bracket Prediction Using Machine Learning and Combinatorial Fusion Analysis (viability: 4): https://sciencetostartup.com/paper/ncaa-bracket-prediction-using-machine-learning-and-combinatorial-fusion-analysis - A novel approach using Combinatorial Fusion Analysis to enhance NCAA bracket predictions through improved team ranking accuracy.
- LLM2Vec-Gen: Generative Embeddings from Large Language Models (viability: 8): https://sciencetostartup.com/paper/llm2vec-gen-generative-embeddings-from-large-language-models - LLM2Vec-Gen leverages self-supervised learning to create generative embeddings from large language models, enhancing performance and interpretability.
- GLM-OCR Technical Report (viability: 7): https://sciencetostartup.com/paper/glm-ocr-technical-report - GLM-OCR is a compact multimodal model for efficient document understanding and recognition.
- When Fine-Tuning Fails and when it Generalises: Role of Data Diversity and Mixed Training in LLM-based TTS (viability: 4): https://sciencetostartup.com/paper/when-fine-tuning-fails-and-when-it-generalises-role-of-data-diversity-and-mixed-training-in-llm-based-tts - Improving voice cloning in TTS systems through effective LoRA fine-tuning of LLMs.
- LookaheadKV: Fast and Accurate KV Cache Eviction by Glimpsing into the Future without Generation (viability: 8): https://sciencetostartup.com/paper/lookaheadkv-fast-and-accurate-kv-cache-eviction-by-glimpsing-into-the-future-without-generation - LOOKAHEADKV enables efficient and optimized key-value cache eviction for transformer models without high latency.
- Ergodicity in reinforcement learning (viability: 2): https://sciencetostartup.com/paper/ergodicity-in-reinforcement-learning - This paper explores the limitations of expected value optimization in reinforcement learning under non-ergodic reward processes.
- S2D: Sparse to Dense Lifting for 3D Reconstruction with Minimal Inputs (viability: 2): https://sciencetostartup.com/paper/s2d-sparse-to-dense-lifting-for-3d-reconstruction-with-minimal-inputs - S2D introduces a novel pipeline for high-quality 3D reconstruction from minimal inputs.
- A Hybrid Knowledge-Grounded Framework for Safety and Traceability in Prescription Verification (viability: 7): https://sciencetostartup.com/paper/a-hybrid-knowledge-grounded-framework-for-safety-and-traceability-in-prescription-verification - PharmGraph-Auditor enhances prescription verification safety through a hybrid knowledge-based system.
- A gripper for flap separation and opening of sealed bags (viability: 5): https://sciencetostartup.com/paper/a-gripper-for-flap-separation-and-opening-of-sealed-bags - A novel gripper design automating the opening of sealed bags in clinical settings to reduce manual strain on nurses.
- Dynamics-Predictive Sampling for Active RL Finetuning of Large Reasoning Models (viability: 7): https://sciencetostartup.com/paper/dynamics-predictive-sampling-for-active-rl-finetuning-of-large-reasoning-models - Dynamics-Predictive Sampling accelerates RL finetuning for large reasoning models by efficiently selecting informative prompts.
- Kernel Tests of Equivalence (viability: 2): https://sciencetostartup.com/paper/kernel-tests-of-equivalence - Novel kernel-based tests for assessing the equivalence between distributions to improve statistical testing accuracy.
- Continuous Diffusion Transformers for Designing Synthetic Regulatory Elements (viability: 3): https://sciencetostartup.com/paper/continuous-diffusion-transformers-for-designing-synthetic-regulatory-elements - A novel Diffusion Transformer for generating synthetic regulatory DNA sequences with improved efficiency.
- LAtte: Hyperbolic Lorentz Attention for Cross-Subject EEG Classification (viability: 7): https://sciencetostartup.com/paper/latte-hyperbolic-lorentz-attention-for-cross-subject-eeg-classification - LAttE is a novel framework for robust cross-subject EEG classification using Lorentz Attention.
- RL-Augmented MPC for Non-Gaited Legged and Hybrid Locomotion (viability: 8): https://sciencetostartup.com/paper/rl-augmented-mpc-for-non-gaited-legged-and-hybrid-locomotion - A novel RL and MPC framework for efficient locomotion control in legged robots.
- From Images to Words: Efficient Cross-Modal Knowledge Distillation to Language Models from Black-box Teachers (viability: 7): https://sciencetostartup.com/paper/from-images-to-words-efficient-cross-modal-knowledge-distillation-to-language-models-from-black-box-teachers - ARMADA is an efficient cross-modal knowledge distillation framework that enhances language models using knowledge from vision-language models without extensive pre-training.
- An Extreme Multi-label Text Classification (XMTC) Library Dataset: What if we took "Use of Practical AI in Digital Libraries" seriously? (viability: 2): https://sciencetostartup.com/paper/an-extreme-multi-label-text-classification-xmtc-library-dataset-what-if-we-took-use-of-practical-ai-in-digital-libraries - A bilingual corpus for enhancing multi-label text classification in digital libraries.
- SNPgen: Phenotype-Supervised Genotype Representation and Synthetic Data Generation via Latent Diffusion (viability: 7): https://sciencetostartup.com/paper/snpgen-phenotype-supervised-genotype-representation-and-synthetic-data-generation-via-latent-diffusion - SNPgen generates synthetic genotype data aligned with phenotypes, enabling privacy-preserving genomic analysis.
- Bilevel Layer-Positioning LoRA for Real Image Dehazing (viability: 8): https://sciencetostartup.com/paper/bilevel-layer-positioning-lora-for-real-image-dehazing - A novel approach to real image dehazing using a bilevel layer-positioning strategy for targeted adaptation.
- FG-CLTP: Fine-Grained Contrastive Language Tactile Pretraining for Robotic Manipulation (viability: 8): https://sciencetostartup.com/paper/fg-cltp-fine-grained-contrastive-language-tactile-pretraining-for-robotic-manipulation - FG-CLTP enhances robotic manipulation by integrating fine-grained tactile sensing with vision-language-action models.
- Beyond Sequential Distance: Inter-Modal Distance Invariant Position Encoding (viability: 8): https://sciencetostartup.com/paper/beyond-sequential-distance-inter-modal-distance-invariant-position-encoding - A novel position encoding mechanism that enhances visual grounding in long-context multimodal language models.
- SiDiaC-v.2.0: Sinhala Diachronic Corpus Version 2.0 (viability: 2): https://sciencetostartup.com/paper/sidiac-v-2-0-sinhala-diachronic-corpus-version-2-0 - SiDiaC-v.2.0 is a comprehensive Sinhala Diachronic Corpus aimed at enhancing NLP resources for the Sinhala language.
- GRACE: A Unified 2D Multi-Robot Path Planning Simulator & Benchmark for Grid, Roadmap, And Continuous Environments (viability: 5): https://sciencetostartup.com/paper/grace-a-unified-2d-multi-robot-path-planning-simulator-benchmark-for-grid-roadmap-and-continuous-environments - GRACE is a unified simulator and benchmark for multi-robot path planning across various environments.
- 6ABOS: An Open-Source Atmospheric Correction Framework for the EnMAP Hyperspectral Mission Based on 6S (viability: 7): https://sciencetostartup.com/paper/6abos-an-open-source-atmospheric-correction-framework-for-the-enmap-hyperspectral-mission-based-on-6s - 6ABOS is an open-source framework that automates atmospheric correction for EnMAP hyperspectral imagery, enhancing aquatic research accuracy.
- UltrasoundAgents: Hierarchical Multi-Agent Evidence-Chain Reasoning for Breast Ultrasound Diagnosis (viability: 3): https://sciencetostartup.com/paper/ultrasoundagents-hierarchical-multi-agent-evidence-chain-reasoning-for-breast-ultrasound-diagnosis - UltrasoundAgents enhances breast ultrasound diagnosis through hierarchical multi-agent reasoning for improved evidence traceability.
- $V_{0.5}$: Generalist Value Model as a Prior for Sparse RL Rollouts (viability: 4): https://sciencetostartup.com/paper/v-0-5-generalist-value-model-as-a-prior-for-sparse-rl-rollouts - A novel reinforcement learning model that enhances policy gradients by adaptively fusing baseline predictions with empirical data.
- Semantic Landmark Particle Filter for Robot Localisation in Vineyards (viability: 7): https://sciencetostartup.com/paper/semantic-landmark-particle-filter-for-robot-localisation-in-vineyards - A Semantic Landmark Particle Filter enhances robot localisation in vineyards by integrating semantic landmarks with LiDAR data.
- Towards Cold-Start Drafting and Continual Refining: A Value-Driven Memory Approach with Application to NPU Kernel Synthesis (viability: 8): https://sciencetostartup.com/paper/towards-cold-start-drafting-and-continual-refining-a-value-driven-memory-approach-with-application-to-npu-kernel-synthes - EvoKernel automates kernel synthesis for data-scarce programming domains using a self-evolving agentic framework.
- PivotAttack: Rethinking the Search Trajectory in Hard-Label Text Attacks via Pivot Words (viability: 7): https://sciencetostartup.com/paper/pivotattack-rethinking-the-search-trajectory-in-hard-label-text-attacks-via-pivot-words - PivotAttack revolutionizes hard-label text attacks by using an efficient inside-out strategy to minimize query costs and improve success rates.
- MAD: Memory Allocation meets Software Diversity (viability: 2): https://sciencetostartup.com/paper/mad-memory-allocation-meets-software-diversity - MAD combines memory allocation with software diversity principles to enhance security against RowHammer attacks.
- On the Reliability of Cue Conflict and Beyond (viability: 4): https://sciencetostartup.com/paper/on-the-reliability-of-cue-conflict-and-beyond - REFINED-BIAS provides a reliable framework for diagnosing shape-texture biases in neural networks.
- Evaluating Few-Shot Pill Recognition Under Visual Domain Shift (viability: 7): https://sciencetostartup.com/paper/evaluating-few-shot-pill-recognition-under-visual-domain-shift - Automated pill recognition system leveraging few-shot learning to enhance medication safety in complex visual environments.
- BALD-SAM: Disagreement-based Active Prompting in Interactive Segmentation (viability: 7): https://sciencetostartup.com/paper/bald-sam-disagreement-based-active-prompting-in-interactive-segmentation - BALD-SAM enhances interactive segmentation by using a principled approach for automated spatial prompting based on Bayesian Active Learning.
- Speaker Verification with Speech-Aware LLMs: Evaluation and Augmentation (viability: 7): https://sciencetostartup.com/paper/speaker-verification-with-speech-aware-llms-evaluation-and-augmentation - A novel approach to enhance speech-aware LLMs with speaker verification capabilities using lightweight augmentation techniques.
- A dataset of medication images with instance segmentation masks for preventing adverse drug events (viability: 6): https://sciencetostartup.com/paper/a-dataset-of-medication-images-with-instance-segmentation-masks-for-preventing-adverse-drug-events - MEDISEG is a comprehensive dataset for medication image recognition, enhancing AI models' ability to prevent adverse drug events.
- ReTabSyn: Realistic Tabular Data Synthesis via Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/retabsyn-realistic-tabular-data-synthesis-via-reinforcement-learning - ReTabSyn is a reinforced tabular data synthesis pipeline that enhances data efficiency by focusing on conditional distributions.
- Evaluating randomized smoothing as a defense against adversarial attacks in trajectory prediction (viability: 4): https://sciencetostartup.com/paper/evaluating-randomized-smoothing-as-a-defense-against-adversarial-attacks-in-trajectory-prediction - A novel defense mechanism using randomized smoothing to enhance the robustness of trajectory prediction models against adversarial attacks.
- HanMoVLM: Large Vision-Language Models for Professional Artistic Painting Evaluation (viability: 7): https://sciencetostartup.com/paper/hanmovlm-large-vision-language-models-for-professional-artistic-painting-evaluation - HanMoVLM transforms large vision-language models into expert evaluators for Chinese artistic paintings.
- Protein Counterfactuals via Diffusion-Guided Latent Optimization (viability: 8): https://sciencetostartup.com/paper/protein-counterfactuals-via-diffusion-guided-latent-optimization - MCCOP enables precise protein engineering by generating biologically plausible sequence edits to optimize protein properties.
- Nurture-First Agent Development: Building Domain-Expert AI Agents Through Conversational Knowledge Crystallization (viability: 2): https://sciencetostartup.com/paper/nurture-first-agent-development-building-domain-expert-ai-agents-through-conversational-knowledge-crystallization - A new paradigm for developing domain-expert AI agents through structured conversational interactions.
- Backdoor Directions in Vision Transformers (viability: 4): https://sciencetostartup.com/paper/backdoor-directions-in-vision-transformers - A diagnostic tool for identifying and addressing backdoor attacks in Vision Transformers.
- Towards Intelligent Spectrum Management: Spectrum Demand Estimation Using Graph Neural Networks (viability: 3): https://sciencetostartup.com/paper/towards-intelligent-spectrum-management-spectrum-demand-estimation-using-graph-neural-networks - A model for estimating spectrum demand using graph neural networks to improve wireless spectrum management.
- PolGS++: Physically-Guided Polarimetric Gaussian Splatting for Fast Reflective Surface Reconstruction (viability: 7): https://sciencetostartup.com/paper/polgs-physically-guided-polarimetric-gaussian-splatting-for-fast-reflective-surface-reconstruction - PolGS++ offers a novel approach for fast and accurate reflective surface reconstruction using physically-guided polarimetric Gaussian Splatting.
- AI-Enhanced Spatial Cellular Traffic Demand Prediction with Contextual Clustering and Error Correction for 5G/6G Planning (viability: 3): https://sciencetostartup.com/paper/ai-enhanced-spatial-cellular-traffic-demand-prediction-with-contextual-clustering-and-error-correction-for-5g-6g-plannin - An AI framework for improving cellular traffic demand prediction in 5G/6G planning.
- Re-Evaluating EVMBench: Are AI Agents Ready for Smart Contract Security? (viability: 8): https://sciencetostartup.com/paper/re-evaluating-evmbench-are-ai-agents-ready-for-smart-contract-security - EVMbench enhances AI agents for smart contract security, providing a benchmark for vulnerability detection and exploitation.
- Multilingual Reasoning Gym: Multilingual Scaling of Procedural Reasoning Environments (viability: 2): https://sciencetostartup.com/paper/multilingual-reasoning-gym-multilingual-scaling-of-procedural-reasoning-environments - A procedural generation platform for multilingual reasoning problems across 14 languages.
- LuxBorrow: From Pompier to Pompjee, Tracing Borrowing in Luxembourgish (viability: 2): https://sciencetostartup.com/paper/luxborrow-from-pompier-to-pompjee-tracing-borrowing-in-luxembourgish - LuxBorrow analyzes borrowing patterns in Luxembourgish news over 27 years.
- The Quadratic Geometry of Flow Matching: Semantic Granularity Alignment for Text-to-Image Synthesis (viability: 2): https://sciencetostartup.com/paper/the-quadratic-geometry-of-flow-matching-semantic-granularity-alignment-for-text-to-image-synthesis - This paper proposes a new approach to enhance text-to-image synthesis through Semantic Granularity Alignment.
- Interpretable Chinese Metaphor Identification via LLM-Assisted MIPVU Rule Script Generation: A Comparative Protocol Study (viability: 5): https://sciencetostartup.com/paper/interpretable-chinese-metaphor-identification-via-llm-assisted-mipvu-rule-script-generation-a-comparative-protocol-study - An LLM-assisted pipeline for interpretable metaphor identification in Chinese, providing transparent rule-based classifications.
- Phase-Interface Instance Segmentation as a Visual Sensor for Laboratory Process Monitoring (viability: 7): https://sciencetostartup.com/paper/phase-interface-instance-segmentation-as-a-visual-sensor-for-laboratory-process-monitoring - A robust visual sensor for laboratory process monitoring using advanced phase-interface segmentation.
- Taking Shortcuts for Categorical VQA Using Super Neurons (viability: 3): https://sciencetostartup.com/paper/taking-shortcuts-for-categorical-vqa-using-super-neurons - A novel approach using Super Neurons for efficient classification in Vision Language Models.
- Guiding Diffusion Models with Semantically Degraded Conditions (viability: 8): https://sciencetostartup.com/paper/guiding-diffusion-models-with-semantically-degraded-conditions - A novel guidance method for text-to-image models that enhances compositional accuracy by using strategically degraded conditions.
- Dynamics-Informed Deep Learning for Predicting Extreme Events (viability: 4): https://sciencetostartup.com/paper/dynamics-informed-deep-learning-for-predicting-extreme-events - A data-driven framework for long-lead prediction of extreme events using mechanism-aware precursors.
- Incremental Federated Learning for Intrusion Detection in IoT Networks under Evolving Threat Landscape (viability: 5): https://sciencetostartup.com/paper/incremental-federated-learning-for-intrusion-detection-in-iot-networks-under-evolving-threat-landscape - A federated learning approach to enhance intrusion detection systems in resource-constrained IoT networks.
- Large Language Models as Annotators for Machine Translation Quality Estimation (viability: 7): https://sciencetostartup.com/paper/large-language-models-as-annotators-for-machine-translation-quality-estimation - A novel approach using LLMs to enhance Machine Translation Quality Estimation through efficient annotation generation.
- Word Recovery in Large Language Models Enables Character-Level Tokenization Robustness (viability: 2): https://sciencetostartup.com/paper/word-recovery-in-large-language-models-enables-character-level-tokenization-robustness - This research explores the robustness of LLMs to character-level tokenization through a novel decoding-based method.
- mAceReason-Math: A Dataset of High-Quality Multilingual Math Problems Ready For RLVR (viability: 4): https://sciencetostartup.com/paper/macereason-math-a-dataset-of-high-quality-multilingual-math-problems-ready-for-rlvr - mAceReason-Math provides a high-quality multilingual dataset of challenging math problems for reinforcement learning research.
- HeartAgent: An Autonomous Agent System for Explainable Differential Diagnosis in Cardiology (viability: 8): https://sciencetostartup.com/paper/heartagent-an-autonomous-agent-system-for-explainable-differential-diagnosis-in-cardiology - HeartAgent is an autonomous agent system that enhances differential diagnosis in cardiology with explainable AI.
- Prioritizing Gradient Sign Over Modulus: An Importance-Aware Framework for Wireless Federated Learning (viability: 4): https://sciencetostartup.com/paper/prioritizing-gradient-sign-over-modulus-an-importance-aware-framework-for-wireless-federated-learning - A framework that enhances wireless federated learning by prioritizing important gradient information for efficient resource allocation.
- CodePercept: Code-Grounded Visual STEM Perception for MLLMs (viability: 8): https://sciencetostartup.com/paper/codepercept-code-grounded-visual-stem-perception-for-mllms - CodePercept enhances visual reasoning in STEM for MLLMs by leveraging executable code as a perceptual medium.
- A PUF-Based Approach for Copy Protection of Intellectual Property in Neural Network Models (viability: 4): https://sciencetostartup.com/paper/a-puf-based-approach-for-copy-protection-of-intellectual-property-in-neural-network-models - A novel method to protect neural network models from unauthorized copying by binding them to unique hardware properties.
- Deep Randomized Distributed Function Computation (DeepRDFC): Neural Distributed Channel Simulation (viability: 2): https://sciencetostartup.com/paper/deep-randomized-distributed-function-computation-deeprdfc-neural-distributed-channel-simulation - A framework for efficient randomized distributed function computation using deep learning.
- AttriGuard: Defeating Indirect Prompt Injection in LLM Agents via Causal Attribution of Tool Invocations (viability: 7): https://sciencetostartup.com/paper/attriguard-defeating-indirect-prompt-injection-in-llm-agents-via-causal-attribution-of-tool-invocations - AttriGuard provides a novel defense against indirect prompt injection in LLM agents through causal attribution of tool invocations.
- Event-based Photometric Stereo via Rotating Illumination and Per-Pixel Learning (viability: 7): https://sciencetostartup.com/paper/event-based-photometric-stereo-via-rotating-illumination-and-per-pixel-learning - An event-based photometric stereo system that predicts surface normals using a rotating light source and a lightweight neural network.
- CUPID: A Plug-in Framework for Joint Aleatoric and Epistemic Uncertainty Estimation with a Single Model (viability: 8): https://sciencetostartup.com/paper/cupid-a-plug-in-framework-for-joint-aleatoric-and-epistemic-uncertainty-estimation-with-a-single-model - CUPID is a plug-in framework that enables joint estimation of aleatoric and epistemic uncertainty in deep learning models without retraining.
- Just-in-Time: Training-Free Spatial Acceleration for Diffusion Transformers (viability: 8): https://sciencetostartup.com/paper/just-in-time-training-free-spatial-acceleration-for-diffusion-transformers - JiT is a training-free framework that accelerates Diffusion Transformers by optimizing spatial computations for faster image synthesis.
- A Grammar of Machine Learning Workflows (viability: 4): https://sciencetostartup.com/paper/a-grammar-of-machine-learning-workflows - A grammar framework to prevent data leakage in machine learning workflows through runtime constraints.
- Beyond Accuracy: Reliability and Uncertainty Estimation in Convolutional Neural Networks (viability: 4): https://sciencetostartup.com/paper/beyond-accuracy-reliability-and-uncertainty-estimation-in-convolutional-neural-networks - A comparative study on uncertainty estimation methods in deep neural networks to enhance reliability in predictions.
- CacheSolidarity: Preventing Prefix Caching Side Channels in Multi-tenant LLM Serving Systems (viability: 7): https://sciencetostartup.com/paper/cachesolidarity-preventing-prefix-caching-side-channels-in-multi-tenant-llm-serving-systems - CacheSolidarity enhances the security of multi-tenant LLM systems by preventing side-channel attacks while maintaining high performance.
- Towards Robust Speech Deepfake Detection via Human-Inspired Reasoning (viability: 7): https://sciencetostartup.com/paper/towards-robust-speech-deepfake-detection-via-human-inspired-reasoning - HIR-SDD leverages human-inspired reasoning to enhance the robustness of speech deepfake detection.
- eLasmobranc Dataset: An Image Dataset for Elasmobranch Species Recognition and Biodiversity Monitoring (viability: 6): https://sciencetostartup.com/paper/elasmobranc-dataset-an-image-dataset-for-elasmobranch-species-recognition-and-biodiversity-monitoring - A publicly available image dataset for fine-grained identification of elasmobranch species to support biodiversity monitoring.
- UAV traffic scene understanding: A cross-spectral guided approach and a unified benchmark (viability: 8): https://sciencetostartup.com/paper/uav-traffic-scene-understanding-a-cross-spectral-guided-approach-and-a-unified-benchmark - CTCNet enhances UAV traffic scene understanding by integrating optical and thermal data for robust analysis.
- Sample-and-Search: An Effective Algorithm for Learning-Augmented k-Median Clustering in High dimensions (viability: 3): https://sciencetostartup.com/paper/sample-and-search-an-effective-algorithm-for-learning-augmented-k-median-clustering-in-high-dimensions - A novel algorithm for learning-augmented k-median clustering that improves computational efficiency.
- Sublinear-Time Reconfiguration of Programmable Matter with Joint Movements (viability: 3): https://sciencetostartup.com/paper/sublinear-time-reconfiguration-of-programmable-matter-with-joint-movements - A centralized algorithm for efficient reconfiguration of geometric amoebot structures using joint movements.
- Riemannian MeanFlow for One-Step Generation on Manifolds (viability: 7): https://sciencetostartup.com/paper/riemannian-meanflow-for-one-step-generation-on-manifolds - Riemannian MeanFlow offers efficient one-step generation on manifolds, enhancing generative model training without heavy computations.
- ASTER: Attitude-aware Suspended-payload Quadrotor Traversal via Efficient Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/aster-attitude-aware-suspended-payload-quadrotor-traversal-via-efficient-reinforcement-learning - ASTER is a robust RL framework enabling autonomous inverted flight for cable-suspended quadrotors.
- MAVEN: A Meta-Reinforcement Learning Framework for Varying-Dynamics Expertise in Agile Quadrotor Maneuvers (viability: 4): https://sciencetostartup.com/paper/maven-a-meta-reinforcement-learning-framework-for-varying-dynamics-expertise-in-agile-quadrotor-maneuvers - MAVEN is a meta-reinforcement learning framework that enables quadrotors to adaptively navigate across varying dynamics.
- Probabilistic Verification of Voice Anti-Spoofing Models (viability: 4): https://sciencetostartup.com/paper/probabilistic-verification-of-voice-anti-spoofing-models - PV-VASM is a probabilistic framework that verifies the robustness of voice anti-spoofing models against various speech synthesis techniques.
- FutureVLA: Joint Visuomotor Prediction for Vision-Language-Action Model (viability: 5): https://sciencetostartup.com/paper/futurevla-joint-visuomotor-prediction-for-vision-language-action-model - FutureVLA enhances intelligent agents by improving joint visuomotor predictive modeling for better action execution.
- Parallel-in-Time Nonlinear Optimal Control via GPU-native Sequential Convex Programming (viability: 8): https://sciencetostartup.com/paper/parallel-in-time-nonlinear-optimal-control-via-gpu-native-sequential-convex-programming - A GPU-native trajectory optimization framework for real-time nonlinear control in autonomous systems.
- Prism-$Δ$: Differential Subspace Steering for Prompt Highlighting in Large Language Models (viability: 7): https://sciencetostartup.com/paper/prism-differential-subspace-steering-for-prompt-highlighting-in-large-language-models - PRISM-$Δ$ enhances prompt highlighting in large language models by optimizing steering directions for better context relevance.
- WalkGPT: Grounded Vision-Language Conversation with Depth-Aware Segmentation for Pedestrian Navigation (viability: 8): https://sciencetostartup.com/paper/walkgpt-grounded-vision-language-conversation-with-depth-aware-segmentation-for-pedestrian-navigation - WalkGPT provides depth-aware, pixel-grounded navigation guidance for pedestrians using advanced vision-language integration.
- UniCom: Unified Multimodal Modeling via Compressed Continuous Semantic Representations (viability: 7): https://sciencetostartup.com/paper/unicom-unified-multimodal-modeling-via-compressed-continuous-semantic-representations - UniCom offers a novel framework for unified multimodal understanding and generation through compressed continuous representations.
- AlphaFlowTSE: One-Step Generative Target Speaker Extraction via Conditional AlphaFlow (viability: 3): https://sciencetostartup.com/paper/alphaflowtse-one-step-generative-target-speaker-extraction-via-conditional-alphaflow - AlphaFlowTSE is a one-step generative model for target speaker extraction that enhances speech fidelity from multi-talker mixtures.
- Structured Linked Data as a Memory Layer for Agent-Orchestrated Retrieval (viability: 7): https://sciencetostartup.com/paper/structured-linked-data-as-a-memory-layer-for-agent-orchestrated-retrieval - Enhancing retrieval accuracy in RAG systems using structured linked data and agentic reasoning.
- EvoSchema: Towards Text-to-SQL Robustness Against Schema Evolution (viability: 7): https://sciencetostartup.com/paper/evoschema-towards-text-to-sql-robustness-against-schema-evolution - EvoSchema enhances text-to-SQL model robustness against evolving database schemas through a comprehensive benchmarking framework.
- RandMark: On Random Watermarking of Visual Foundation Models (viability: 4): https://sciencetostartup.com/paper/randmark-on-random-watermarking-of-visual-foundation-models - A novel method for ownership verification of visual foundation models through random watermarking.
- Bioinspired CNNs for border completion in occluded images (viability: 5): https://sciencetostartup.com/paper/bioinspired-cnns-for-border-completion-in-occluded-images - BorderNet enhances image recognition robustness against occlusions using bioinspired CNN filters.
- Repurposing Backdoors for Good: Ephemeral Intrinsic Proofs for Verifiable Aggregation in Cross-silo Federated Learning (viability: 7): https://sciencetostartup.com/paper/repurposing-backdoors-for-good-ephemeral-intrinsic-proofs-for-verifiable-aggregation-in-cross-silo-federated-learning - A lightweight architecture for verifiable aggregation in federated learning that enhances security without heavy cryptographic overhead.
- Contract And Conquer: How to Provably Compute Adversarial Examples for a Black-Box Model? (viability: 7): https://sciencetostartup.com/paper/contract-and-conquer-how-to-provably-compute-adversarial-examples-for-a-black-box-model - Contract And Conquer provides a provable method to compute adversarial examples for black-box models, outperforming existing techniques.
- MapGCLR: Geospatial Contrastive Learning of Representations for Online Vectorized HD Map Construction (viability: 7): https://sciencetostartup.com/paper/mapgclr-geospatial-contrastive-learning-of-representations-for-online-vectorized-hd-map-construction - A self-supervised approach for scalable online HD map construction using geospatial contrastive learning.
- A$^2$-Edit: Precise Reference-Guided Image Editing of Arbitrary Objects and Ambiguous Masks (viability: 7): https://sciencetostartup.com/paper/a-2-edit-precise-reference-guided-image-editing-of-arbitrary-objects-and-ambiguous-masks - A$^2$-Edit is an advanced inpainting framework that enables precise image editing of arbitrary objects using coarse masks.
- OnFly: Onboard Zero-Shot Aerial Vision-Language Navigation toward Safety and Efficiency (viability: 8): https://sciencetostartup.com/paper/onfly-onboard-zero-shot-aerial-vision-language-navigation-toward-safety-and-efficiency - OnFly enables UAVs to navigate using natural language instructions with enhanced safety and efficiency through real-time onboard processing.
- A Platform-Agnostic Multimodal Digital Human Modelling Framework: Neurophysiological Sensing in Game-Based Interaction (viability: 5): https://sciencetostartup.com/paper/a-platform-agnostic-multimodal-digital-human-modelling-framework-neurophysiological-sensing-in-game-based-interaction - A platform-agnostic framework for multimodal digital human modeling that enhances accessibility and inclusion in AI research.
- Surrogate models for nuclear fusion with parametric Shallow Recurrent Decoder Networks: applications to magnetohydrodynamics (viability: 4): https://sciencetostartup.com/paper/surrogate-models-for-nuclear-fusion-with-parametric-shallow-recurrent-decoder-networks-applications-to-magnetohydrodynam - A data-driven framework using SHRED for efficient state reconstruction in nuclear fusion applications.
- Emulating Clinician Cognition via Self-Evolving Deep Clinical Research (viability: 8): https://sciencetostartup.com/paper/emulating-clinician-cognition-via-self-evolving-deep-clinical-research - DxEvolve is a self-evolving diagnostic agent that enhances clinical diagnosis through continuous learning and improved accuracy.
- Spatio-Temporal Attention Graph Neural Network: Explaining Causalities With Attention (viability: 3): https://sciencetostartup.com/paper/spatio-temporal-attention-graph-neural-network-explaining-causalities-with-attention - A novel Spatio-Temporal Attention Graph Neural Network for explainable anomaly detection in Industrial Control Systems.
- Cybo-Waiter: A Physical Agentic Framework for Humanoid Whole-Body Locomotion-Manipulation (viability: 3): https://sciencetostartup.com/paper/cybo-waiter-a-physical-agentic-framework-for-humanoid-whole-body-locomotion-manipulation - A humanoid agent framework for executing complex tasks through advanced locomotion and manipulation.
- An FPGA Implementation of Displacement Vector Search for Intra Pattern Copy in JPEG XS (viability: 4): https://sciencetostartup.com/paper/an-fpga-implementation-of-displacement-vector-search-for-intra-pattern-copy-in-jpeg-xs - An FPGA architecture for efficient displacement vector search in JPEG XS image compression.
- Dynamic Modeling and Attitude Control of a Reaction-Wheel-Based Low-Gravity Bipedal Hopper (viability: 4): https://sciencetostartup.com/paper/dynamic-modeling-and-attitude-control-of-a-reaction-wheel-based-low-gravity-bipedal-hopper - A bipedal hopping robot that stabilizes mid-air posture using a reaction wheel for efficient extraterrestrial locomotion.
- FAME: Formal Abstract Minimal Explanation for Neural Networks (viability: 4): https://sciencetostartup.com/paper/fame-formal-abstract-minimal-explanation-for-neural-networks - FAME provides a novel method for generating minimal explanations for large neural networks, enhancing interpretability and efficiency.
- STM32-Based Smart Waste Bin for Hygienic Disposal Using Embedded Sensing and Automated Control (viability: 5): https://sciencetostartup.com/paper/stm32-based-smart-waste-bin-for-hygienic-disposal-using-embedded-sensing-and-automated-control - An automated motion-sensing waste bin that opens without contact, enhancing hygiene in public and private spaces.
- How To Embed Matters: Evaluation of EO Embedding Design Choices (viability: 4): https://sciencetostartup.com/paper/how-to-embed-matters-evaluation-of-eo-embedding-design-choices - A systematic analysis of embedding design choices for scalable Earth observation workflows using Geospatial Foundation Models.
- Are Video Reasoning Models Ready to Go Outside? (viability: 7): https://sciencetostartup.com/paper/are-video-reasoning-models-ready-to-go-outside - ROVA enhances the robustness of vision-language models against real-world disturbances through a novel training framework.
- Interleaving Scheduling and Motion Planning with Incremental Learning of Symbolic Space-Time Motion Abstractions (viability: 6): https://sciencetostartup.com/paper/interleaving-scheduling-and-motion-planning-with-incremental-learning-of-symbolic-space-time-motion-abstractions - A framework that integrates scheduling and motion planning to optimize task execution in complex environments.
- Less is More: Decoder-Free Masked Modeling for Efficient Skeleton Representation Learning (viability: 8): https://sciencetostartup.com/paper/less-is-more-decoder-free-masked-modeling-for-efficient-skeleton-representation-learning - SLiM is a novel framework for efficient skeleton-based action representation learning that eliminates the need for decoders while achieving state-of-the-art performance.
- Detecting and Eliminating Neural Network Backdoors Through Active Paths with Application to Intrusion Detection (viability: 3): https://sciencetostartup.com/paper/detecting-and-eliminating-neural-network-backdoors-through-active-paths-with-application-to-intrusion-detection - A novel approach to detect and eliminate backdoor triggers in machine learning models for intrusion detection.
- Making Bielik LLM Reason (Better): A Field Report (viability: 4): https://sciencetostartup.com/paper/making-bielik-llm-reason-better-a-field-report - Enhancing the reasoning capabilities of the Bielik LLM through rigorous evaluation and benchmarking.
- Splat2Real: Novel-view Scaling for Physical AI with 3D Gaussian Splatting (viability: 2): https://sciencetostartup.com/paper/splat2real-novel-view-scaling-for-physical-ai-with-3d-gaussian-splatting - Splat2Real enhances monocular RGB-to-3D perception by optimizing novel-view scaling for physical AI.
- Reinforcement Learning with Conditional Expectation Reward (viability: 8): https://sciencetostartup.com/paper/reinforcement-learning-with-conditional-expectation-reward - Conditional Expectation Reward enhances reasoning in large language models by providing a flexible verification mechanism without external rules.
- Disentangling Similarity and Relatedness in Topic Models (viability: 4): https://sciencetostartup.com/paper/disentangling-similarity-and-relatedness-in-topic-models - A novel approach to enhance topic models by integrating PLM embeddings to better capture semantic structures.
- AdaClearGrasp: Learning Adaptive Clearing for Zero-Shot Robust Dexterous Grasping in Densely Cluttered Environments (viability: 8): https://sciencetostartup.com/paper/adacleargrasp-learning-adaptive-clearing-for-zero-shot-robust-dexterous-grasping-in-densely-cluttered-environments - AdaClearGrasp enables robots to adaptively decide between clearing obstacles or grasping targets in cluttered environments for improved manipulation.
- MUNIChus: Multilingual News Image Captioning Benchmark (viability: 5): https://sciencetostartup.com/paper/munichus-multilingual-news-image-captioning-benchmark - MUNIChus is a multilingual benchmark for news image captioning that integrates textual and visual content across nine languages.
- Learning Bimanual Cloth Manipulation with Vision-based Tactile Sensing via Single Robotic Arm (viability: 7): https://sciencetostartup.com/paper/learning-bimanual-cloth-manipulation-with-vision-based-tactile-sensing-via-single-robotic-arm - Touch G.O.G. is a compact vision-based tactile gripper for efficient single-arm cloth manipulation.
- An Approach for Safe and Secure Software Protection Supported by Symbolic Execution (viability: 2): https://sciencetostartup.com/paper/an-approach-for-safe-and-secure-software-protection-supported-by-symbolic-execution - A novel copy-protection method for industrial control software using Physically Unclonable Functions.
- HyPER-GAN: Hybrid Patch-Based Image-to-Image Translation for Real-Time Photorealism Enhancement (viability: 7): https://sciencetostartup.com/paper/hyper-gan-hybrid-patch-based-image-to-image-translation-for-real-time-photorealism-enhancement - HyPER-GAN offers real-time photorealism enhancement for synthetic images using lightweight paired image translation.
- Trajectory-Informed Memory Generation for Self-Improving Agent Systems (viability: 8): https://sciencetostartup.com/paper/trajectory-informed-memory-generation-for-self-improving-agent-systems - A framework that enhances LLM-powered agents by enabling them to learn from past execution experiences for improved performance.
- Self-Scaled Broyden Family of Quasi-Newton Methods in JAX (viability: 4): https://sciencetostartup.com/paper/self-scaled-broyden-family-of-quasi-newton-methods-in-jax - A JAX implementation of the Self-Scaled Broyden family of quasi-Newton optimization methods for easier adoption in the JAX community.
- Layer Consistency Matters: Elegant Latent Transition Discrepancy for Generalizable Synthetic Image Detection (viability: 8): https://sciencetostartup.com/paper/layer-consistency-matters-elegant-latent-transition-discrepancy-for-generalizable-synthetic-image-detection - A novel approach for detecting synthetic images by analyzing latent transition discrepancies across network layers.
- Recover to Predict: Progressive Retrospective Learning for Variable-Length Trajectory Prediction (viability: 8): https://sciencetostartup.com/paper/recover-to-predict-progressive-retrospective-learning-for-variable-length-trajectory-prediction - A novel framework for improving trajectory prediction in autonomous driving using variable-length observations.
- Gradient Flow Drifting: Generative Modeling via Wasserstein Gradient Flows of KDE-Approximated Divergences (viability: 4): https://sciencetostartup.com/paper/gradient-flow-drifting-generative-modeling-via-wasserstein-gradient-flows-of-kde-approximated-divergences - A new family of generative models leveraging Wasserstein gradient flows for improved performance in avoiding mode collapse.
- Does LLM Alignment Really Need Diversity? An Empirical Study of Adapting RLVR Methods for Moral Reasoning (viability: 4): https://sciencetostartup.com/paper/does-llm-alignment-really-need-diversity-an-empirical-study-of-adapting-rlvr-methods-for-moral-reasoning - A study on the effectiveness of reward-maximizing methods for aligning large language models in moral reasoning tasks.
- Need for Speed: Zero-Shot Depth Completion with Single-Step Diffusion (viability: 7): https://sciencetostartup.com/paper/need-for-speed-zero-shot-depth-completion-with-single-step-diffusion - Marigold-SSD offers a fast and efficient depth completion solution using single-step diffusion for real-time 3D perception.
- Attribution as Retrieval: Model-Agnostic AI-Generated Image Attribution (viability: 8): https://sciencetostartup.com/paper/attribution-as-retrieval-model-agnostic-ai-generated-image-attribution - LIDA is a model-agnostic framework for efficient attribution of AI-generated images, addressing the challenges of traditional methods.
- HAPEns: Hardware-Aware Post-Hoc Ensembling for Tabular Data (viability: 4): https://sciencetostartup.com/paper/hapens-hardware-aware-post-hoc-ensembling-for-tabular-data - HAPEns optimizes ensemble methods for tabular data by balancing predictive performance with hardware efficiency.
- R4-CGQA: Retrieval-based Vision Language Models for Computer Graphics Image Quality Assessment (viability: 6): https://sciencetostartup.com/paper/r4-cgqa-retrieval-based-vision-language-models-for-computer-graphics-image-quality-assessment - A retrieval-augmented framework that enhances vision language models for assessing computer graphics quality.
- CUAAudit: Meta-Evaluation of Vision-Language Models as Auditors of Autonomous Computer-Use Agents (viability: 5): https://sciencetostartup.com/paper/cuaaudit-meta-evaluation-of-vision-language-models-as-auditors-of-autonomous-computer-use-agents - CUAAudit leverages Vision-Language Models to autonomously evaluate the performance of Computer-Use Agents in real-world environments.
- Implicit Statistical Inference in Transformers: Approximating Likelihood-Ratio Tests In-Context (viability: 4): https://sciencetostartup.com/paper/implicit-statistical-inference-in-transformers-approximating-likelihood-ratio-tests-in-context - This research explores the statistical foundations of in-context learning in Transformers, providing insights into their decision-making processes.
- Safety-critical Control Under Partial Observability: Reach-Avoid POMDP meets Belief Space Control (viability: 7): https://sciencetostartup.com/paper/safety-critical-control-under-partial-observability-reach-avoid-pomdp-meets-belief-space-control - A novel layered control architecture for robot decision-making under uncertainty that enhances safety and task success.
- End-to-End Chatbot Evaluation with Adaptive Reasoning and Uncertainty Filtering (viability: 7): https://sciencetostartup.com/paper/end-to-end-chatbot-evaluation-with-adaptive-reasoning-and-uncertainty-filtering - An end-to-end automatic evaluator for domain-specific chatbots that reduces human review effort and enhances scalability.
- UniStitch: Unifying Semantic and Geometric Features for Image Stitching (viability: 8): https://sciencetostartup.com/paper/unistitch-unifying-semantic-and-geometric-features-for-image-stitching - UniStitch unifies semantic and geometric features for superior image stitching performance.
- TacLoc: Global Tactile Localization on Objects from a Registration Perspective (viability: 7): https://sciencetostartup.com/paper/tacloc-global-tactile-localization-on-objects-from-a-registration-perspective - TacLoc is a novel tactile localization framework for accurate pose estimation in robotic manipulation without pre-trained models.
- Adaptive RAN Slicing Control via Reward-Free Self-Finetuning Agents (viability: 7): https://sciencetostartup.com/paper/adaptive-ran-slicing-control-via-reward-free-self-finetuning-agents - A self-finetuning framework for autonomous control in AI-native network systems, enhancing efficiency and stability without handcrafted rewards.
- Riemannian Geometry-Preserving Variational Autoencoder for MI-BCI Data Augmentation (viability: 2): https://sciencetostartup.com/paper/riemannian-geometry-preserving-variational-autoencoder-for-mi-bci-data-augmentation - A generative model for creating synthetic EEG covariance matrices for brain-computer interface applications.
- PET-F2I: A Comprehensive Benchmark and Parameter-Efficient Fine-Tuning of LLMs for PET/CT Report Impression Generation (viability: 8): https://sciencetostartup.com/paper/pet-f2i-a-comprehensive-benchmark-and-parameter-efficient-fine-tuning-of-llms-for-pet-ct-report-impression-generation - PET-F2I is a benchmark and fine-tuning method for generating diagnostic impressions from PET/CT reports using LLMs.
- A Bipartite Graph Approach to U.S.-China Cross-Market Return Forecasting (viability: 2): https://sciencetostartup.com/paper/a-bipartite-graph-approach-to-u-s-china-cross-market-return-forecasting - A novel machine learning framework for predicting cross-market returns between U.S. and Chinese equities using bipartite graphs.
- P-GSVC: Layered Progressive 2D Gaussian Splatting for Scalable Image and Video (viability: 7): https://sciencetostartup.com/paper/p-gsvc-layered-progressive-2d-gaussian-splatting-for-scalable-image-and-video - P-GSVC offers a scalable framework for enhanced image and video reconstruction using layered Gaussian splatting.
- Towards Cognitive Defect Analysis in Active Infrared Thermography with Vision-Text Cues (viability: 6): https://sciencetostartup.com/paper/towards-cognitive-defect-analysis-in-active-infrared-thermography-with-vision-text-cues - A novel language-guided framework for zero-shot defect analysis in carbon fiber-reinforced polymers using active infrared thermography.
- Automatic End-to-End Data Integration using Large Language Models (viability: 8): https://sciencetostartup.com/paper/automatic-end-to-end-data-integration-using-large-language-models - An automatic data integration pipeline using GPT-5.2 that reduces manual effort and costs significantly.
- Learning to Score: Tuning Cluster Schedulers through Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/learning-to-score-tuning-cluster-schedulers-through-reinforcement-learning - A reinforcement learning approach to optimize cluster scheduler scoring functions for improved job performance.
- SCORE: Replacing Layer Stacking with Contractive Recurrent Depth (viability: 3): https://sciencetostartup.com/paper/score-replacing-layer-stacking-with-contractive-recurrent-depth - SCORE offers a novel recurrent approach to optimize deep neural networks by replacing traditional layer stacking with contractive updates.
- Prompting with the human-touch: evaluating model-sensitivity of foundation models for musculoskeletal CT segmentation (viability: 8): https://sciencetostartup.com/paper/prompting-with-the-human-touch-evaluating-model-sensitivity-of-foundation-models-for-musculoskeletal-ct-segmentation - A benchmarking tool for evaluating promptable foundation models in musculoskeletal CT segmentation.
- DSFlash: Comprehensive Panoptic Scene Graph Generation in Realtime (viability: 3): https://sciencetostartup.com/paper/dsflash-comprehensive-panoptic-scene-graph-generation-in-realtime - DSFlash is a low-latency model for comprehensive scene graph generation optimized for edge devices.
- Tackling Length Inflation Without Trade-offs: Group Relative Reward Rescaling for Reinforcement Learning (viability: 3): https://sciencetostartup.com/paper/tackling-length-inflation-without-trade-offs-group-relative-reward-rescaling-for-reinforcement-learning - Introducing Group Relative Reward Rescaling (GR$^3$) to mitigate length inflation in reinforcement learning for LLMs.
- BinWalker: Development and Field Evaluation of a Quadruped Manipulator Platform for Sustainable Litter Collection (viability: 8): https://sciencetostartup.com/paper/binwalker-development-and-field-evaluation-of-a-quadruped-manipulator-platform-for-sustainable-litter-collection - A quadruped robotic platform designed for autonomous litter collection in challenging outdoor environments.
- UAV-MARL: Multi-Agent Reinforcement Learning for Time-Critical and Dynamic Medical Supply Delivery (viability: 7): https://sciencetostartup.com/paper/uav-marl-multi-agent-reinforcement-learning-for-time-critical-and-dynamic-medical-supply-delivery - A multi-agent reinforcement learning framework for optimizing UAV logistics in medical supply delivery.
- Sparse Task Vector Mixup with Hypernetworks for Efficient Knowledge Transfer in Whole-Slide Image Prognosis (viability: 8): https://sciencetostartup.com/paper/sparse-task-vector-mixup-with-hypernetworks-for-efficient-knowledge-transfer-in-whole-slide-image-prognosis - STEPH enhances cancer prognosis by efficiently transferring knowledge across multiple cancer types using hypernetworks.
- World Model for Battery Degradation Prediction Under Non-Stationary Aging (viability: 4): https://sciencetostartup.com/paper/world-model-for-battery-degradation-prediction-under-non-stationary-aging - A novel world model approach for predicting lithium-ion battery degradation trajectories using electrochemical constraints.
- AILS-NTUA at SemEval-2026 Task 8: Evaluating Multi-Turn RAG Conversations (viability: 3): https://sciencetostartup.com/paper/ails-ntua-at-semeval-2026-task-8-evaluating-multi-turn-rag-conversations - AILS-NTUA presents a multi-turn retrieval-augmented generation system that excels in competitive evaluations.
- IH-Challenge: A Training Dataset to Improve Instruction Hierarchy on Frontier LLMs (viability: 7): https://sciencetostartup.com/paper/ih-challenge-a-training-dataset-to-improve-instruction-hierarchy-on-frontier-llms - IH-Challenge is a dataset designed to enhance the robustness of instruction hierarchy in LLMs, improving their ability to handle conflicting instructions.
- Visually-Guided Controllable Medical Image Generation via Fine-Grained Semantic Disentanglement (viability: 8): https://sciencetostartup.com/paper/visually-guided-controllable-medical-image-generation-via-fine-grained-semantic-disentanglement - A framework for generating high-quality medical images by disentangling text semantics for improved control and accuracy.
- UHD Image Deblurring via Autoregressive Flow with Ill-conditioned Constraints (viability: 4): https://sciencetostartup.com/paper/uhd-image-deblurring-via-autoregressive-flow-with-ill-conditioned-constraints - A novel autoregressive flow method for efficient UHD image deblurring that balances detail recovery and computational cost.
- Resource-constrained Amazons chess decision framework integrating large language models and graph attention (viability: 7): https://sciencetostartup.com/paper/resource-constrained-amazons-chess-decision-framework-integrating-large-language-models-and-graph-attention - A lightweight hybrid framework for the Game of the Amazons that combines graph attention and large language models for improved decision-making.
- Safe and Scalable Web Agent Learning via Recreated Websites (viability: 8): https://sciencetostartup.com/paper/safe-and-scalable-web-agent-learning-via-recreated-websites - VeriEnv is a framework that enables safe and scalable training of web agents by recreating real-world websites into synthetic environments.
- Naïve Exposure of Generative AI Capabilities Undermines Deepfake Detection (viability: 4): https://sciencetostartup.com/paper/na-ve-exposure-of-generative-ai-capabilities-undermines-deepfake-detection - This research highlights the vulnerabilities of deepfake detection systems against advanced generative AI capabilities.
- IMTBench: A Multi-Scenario Cross-Modal Collaborative Evaluation Benchmark for In-Image Machine Translation (viability: 4): https://sciencetostartup.com/paper/imtbench-a-multi-scenario-cross-modal-collaborative-evaluation-benchmark-for-in-image-machine-translation - IMTBench is a comprehensive benchmark for evaluating in-image machine translation across multiple scenarios and languages.
- VERI-DPO: Evidence-Aware Alignment for Clinical Summarization via Claim Verification and Direct Preference Optimization (viability: 3): https://sciencetostartup.com/paper/veri-dpo-evidence-aware-alignment-for-clinical-summarization-via-claim-verification-and-direct-preference-optimization - VERI-DPO enhances clinical summarization by verifying claims and optimizing preferences to improve accuracy.
- A Universal Nearest-Neighbor Estimator for Intrinsic Dimensionality (viability: 4): https://sciencetostartup.com/paper/a-universal-nearest-neighbor-estimator-for-intrinsic-dimensionality - A universal estimator for intrinsic dimensionality that outperforms existing methods across various data distributions.
- Human-AI Co-reasoning for Clinical Diagnosis with Evidence-Integrated Language Agent (viability: 8): https://sciencetostartup.com/paper/human-ai-co-reasoning-for-clinical-diagnosis-with-evidence-integrated-language-agent - PULSE is a medical reasoning agent that enhances diagnostic decision-making by integrating a domain-tuned language model with scientific literature retrieval.
- Spatial self-supervised Peak Learning and correlation-based Evaluation of peak picking in Mass Spectrometry Imaging (viability: 7): https://sciencetostartup.com/paper/spatial-self-supervised-peak-learning-and-correlation-based-evaluation-of-peak-picking-in-mass-spectrometry-imaging - An autoencoder-based neural network for effective peak picking in mass spectrometry imaging, enhancing data analysis through spatial self-supervision.
- Dual Space Preconditioning for Gradient Descent in the Overparameterized Regime (viability: 3): https://sciencetostartup.com/paper/dual-space-preconditioning-for-gradient-descent-in-the-overparameterized-regime - This paper introduces a novel gradient descent optimization technique that guarantees convergence in over-parameterized models.
- StructDamage:A Large Scale Unified Crack and Surface Defect Dataset for Robust Structural Damage Detection (viability: 7): https://sciencetostartup.com/paper/structdamage-a-large-scale-unified-crack-and-surface-defect-dataset-for-robust-structural-damage-detection - StructDamage is a comprehensive dataset for robust detection and classification of structural cracks and surface defects.
- PEEM: Prompt Engineering Evaluation Metrics for Interpretable Joint Evaluation of Prompts and Responses (viability: 7): https://sciencetostartup.com/paper/peem-prompt-engineering-evaluation-metrics-for-interpretable-joint-evaluation-of-prompts-and-responses - PEEM is a framework for interpretable evaluation of prompts and responses in large language models, enhancing prompt design and optimization.
- Learning to Negotiate: Multi-Agent Deliberation for Collective Value Alignment in LLMs (viability: 3): https://sciencetostartup.com/paper/learning-to-negotiate-multi-agent-deliberation-for-collective-value-alignment-in-llms - A framework for aligning large language models through multi-agent negotiation to enhance conflict resolution.
- Muscle Synergy Priors Enhance Biomechanical Fidelity in Predictive Musculoskeletal Locomotion Simulation (viability: 3): https://sciencetostartup.com/paper/muscle-synergy-priors-enhance-biomechanical-fidelity-in-predictive-musculoskeletal-locomotion-simulation - A reinforcement-learning framework enhances the fidelity of musculoskeletal locomotion simulations through muscle synergy constraints.
- Aligning Large Language Models with Searcher Preferences (viability: 6): https://sciencetostartup.com/paper/aligning-large-language-models-with-searcher-preferences - SearchLLM revolutionizes generative search by aligning large language models with user preferences through a novel reward system.
- Modeling Stage-wise Evolution of User Interests for News Recommendation (viability: 7): https://sciencetostartup.com/paper/modeling-stage-wise-evolution-of-user-interests-for-news-recommendation - A unified framework for personalized news recommendation that adapts to evolving user interests over time.
- Fighting Hallucinations with Counterfactuals: Diffusion-Guided Perturbations for LVLM Hallucination Suppression (viability: 8): https://sciencetostartup.com/paper/fighting-hallucinations-with-counterfactuals-diffusion-guided-perturbations-for-lvlm-hallucination-suppression - CIPHER is a training-free method that suppresses hallucinations in vision-language models using counterfactual image perturbations.
- DepthCache: Depth-Guided Training-Free Visual Token Merging for Vision-Language-Action Model Inference (viability: 8): https://sciencetostartup.com/paper/depthcache-depth-guided-training-free-visual-token-merging-for-vision-language-action-model-inference - DepthCache is a training-free framework that optimizes visual token merging for faster robotic manipulation without degrading performance.
- UniPINN: A Unified PINN Framework for Multi-task Learning of Diverse Navier-Stokes Equations (viability: 7): https://sciencetostartup.com/paper/unipinn-a-unified-pinn-framework-for-multi-task-learning-of-diverse-navier-stokes-equations - UniPINN is a unified framework for multi-task learning of diverse Navier-Stokes equations using Physics-Informed Neural Networks.
- MoXaRt: Audio-Visual Object-Guided Sound Interaction for XR (viability: 7): https://sciencetostartup.com/paper/moxart-audio-visual-object-guided-sound-interaction-for-xr - MoXaRt enhances XR experiences by using audio-visual cues for real-time sound source separation.
- Learning to Wander: Improving the Global Image Geolocation Ability of LMMs via Actionable Reasoning (viability: 7): https://sciencetostartup.com/paper/learning-to-wander-improving-the-global-image-geolocation-ability-of-lmms-via-actionable-reasoning - WanderBench and GeoAoT revolutionize image geolocation by integrating actionable reasoning with embodied exploration.
- SUBTA: A Framework for Supported User-Guided Bimanual Teleoperation in Structured Assembly (viability: 7): https://sciencetostartup.com/paper/subta-a-framework-for-supported-user-guided-bimanual-teleoperation-in-structured-assembly - SUBTA is a teleoperation system that enhances bimanual assembly through user-guided robotic assistance.
- LCAMV: High-Accuracy 3D Reconstruction of Color-Varying Objects Using LCA Correction and Minimum-Variance Fusion in Structured Light (viability: 3): https://sciencetostartup.com/paper/lcamv-high-accuracy-3d-reconstruction-of-color-varying-objects-using-lca-correction-and-minimum-variance-fusion-in-struc - LCAMV offers a novel method for high-accuracy 3D reconstruction of color-varying objects using structured light.
- Spatio-Temporal Forecasting of Retaining Wall Deformation: Mitigating Error Accumulation via Multi-Resolution ConvLSTM Stacking Ensemble (viability: 4): https://sciencetostartup.com/paper/spatio-temporal-forecasting-of-retaining-wall-deformation-mitigating-error-accumulation-via-multi-resolution-convlstm-st - A multi-resolution ConvLSTM ensemble framework for accurate forecasting of retaining wall deformation during excavation.
- Brenier Isotonic Regression (viability: 4): https://sciencetostartup.com/paper/brenier-isotonic-regression - Brenier isotonic regression extends classical isotonic regression to multi-output scenarios using optimal transport principles.
- FAR-Dex: Few-shot Data Augmentation and Adaptive Residual Policy Refinement for Dexterous Manipulation (viability: 7): https://sciencetostartup.com/paper/far-dex-few-shot-data-augmentation-and-adaptive-residual-policy-refinement-for-dexterous-manipulation - FAR-Dex enhances robotic dexterous manipulation through few-shot data augmentation and adaptive policy refinement.
- DiT4DiT: Jointly Modeling Video Dynamics and Actions for Generalizable Robot Control (viability: 8): https://sciencetostartup.com/paper/dit4dit-jointly-modeling-video-dynamics-and-actions-for-generalizable-robot-control - DiT4DiT offers an enhanced robot control model leveraging video-action synthesis for superior robotic manipulation.
- SignSparK: Efficient Multilingual Sign Language Production via Sparse Keyframe Learning (viability: 8): https://sciencetostartup.com/paper/signspark-efficient-multilingual-sign-language-production-via-sparse-keyframe-learning - Efficiently generate natural multilingual sign language avatars with sparse keyframe learning.
- Unlearning the Unpromptable: Prompt-free Instance Unlearning in Diffusion Models (viability: 7): https://sciencetostartup.com/paper/unlearning-the-unpromptable-prompt-free-instance-unlearning-in-diffusion-models - A novel method for prompt-free instance unlearning in diffusion models to enhance privacy and ethical compliance.
- The Curse and Blessing of Mean Bias in FP4-Quantized LLM Training (viability: 2): https://sciencetostartup.com/paper/the-curse-and-blessing-of-mean-bias-in-fp4-quantized-llm-training - This paper addresses numerical instability in low-bit LLM training caused by mean bias, proposing a simple mean-subtraction method for improvement.
- GGMPs: Generalized Gaussian Mixture Processes (viability: 4): https://sciencetostartup.com/paper/ggmps-generalized-gaussian-mixture-processes - GGMPs offer a novel approach to multimodal conditional density estimation using Gaussian processes.
- KnowDiffuser: A Knowledge-Guided Diffusion Planner with LM Reasoning and Prior-Informed Trajectory Initialization (viability: 8): https://sciencetostartup.com/paper/knowdiffuser-a-knowledge-guided-diffusion-planner-with-lm-reasoning-and-prior-informed-trajectory-initialization - KnowDiffuser integrates language models and diffusion models for advanced motion planning in autonomous driving.
- AsyncMDE: Real-Time Monocular Depth Estimation via Asynchronous Spatial Memory (viability: 8): https://sciencetostartup.com/paper/asyncmde-real-time-monocular-depth-estimation-via-asynchronous-spatial-memory - AsyncMDE is a real-time monocular depth estimation system that efficiently reduces computational costs for edge deployment.
- COHORT: Hybrid RL for Collaborative Large DNN Inference on Multi-Robot Systems Under Real-Time Constraints (viability: 7): https://sciencetostartup.com/paper/cohort-hybrid-rl-for-collaborative-large-dnn-inference-on-multi-robot-systems-under-real-time-constraints - COHORT is a collaborative DNN inference framework for multi-robot systems that optimizes resource usage in real-time scenarios.
- Adaptive Active Learning for Regression via Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/adaptive-active-learning-for-regression-via-reinforcement-learning - WiGS leverages reinforcement learning to optimize active learning for regression, enhancing labeling efficiency and accuracy.
- Domain-Adaptive Health Indicator Learning with Degradation-Stage Synchronized Sampling and Cross-Domain Autoencoder (viability: 7): https://sciencetostartup.com/paper/domain-adaptive-health-indicator-learning-with-degradation-stage-synchronized-sampling-and-cross-domain-autoencoder - A domain-adaptive framework for improving health indicator modeling in industrial systems through synchronized sampling and advanced autoencoding techniques.
- World2Act: Latent Action Post-Training via Skill-Compositional World Models (viability: 8): https://sciencetostartup.com/paper/world2act-latent-action-post-training-via-skill-compositional-world-models - World2Act enhances Vision-Language-Action policies by aligning actions with video-dynamics latents for improved robustness and generalization.
- TractoRC: A Unified Probabilistic Learning Framework for Joint Tractography Registration and Clustering (viability: 7): https://sciencetostartup.com/paper/tractorc-a-unified-probabilistic-learning-framework-for-joint-tractography-registration-and-clustering - TractoRC is a unified framework that enhances diffusion MRI tractography by jointly optimizing tractogram registration and streamline clustering.
- Frames2Residual: Spatiotemporal Decoupling for Self-Supervised Video Denoising (viability: 7): https://sciencetostartup.com/paper/frames2residual-spatiotemporal-decoupling-for-self-supervised-video-denoising - Frames2Residual offers a novel approach to self-supervised video denoising by decoupling spatiotemporal learning for improved texture recovery.
- Enhancing Network Intrusion Detection Systems: A Multi-Layer Ensemble Approach to Mitigate Adversarial Attacks (viability: 5): https://sciencetostartup.com/paper/enhancing-network-intrusion-detection-systems-a-multi-layer-ensemble-approach-to-mitigate-adversarial-attacks - A multi-layer ensemble approach to enhance the robustness of network intrusion detection systems against adversarial attacks.
- Effective Dataset Distillation for Spatio-Temporal Forecasting with Bi-dimensional Compression (viability: 7): https://sciencetostartup.com/paper/effective-dataset-distillation-for-spatio-temporal-forecasting-with-bi-dimensional-compression - STemDist is a novel dataset distillation method that enhances spatio-temporal forecasting by compressing both spatial and temporal dimensions.
- Motion Forcing: A Decoupled Framework for Robust Video Generation in Motion Dynamics (viability: 7): https://sciencetostartup.com/paper/motion-forcing-a-decoupled-framework-for-robust-video-generation-in-motion-dynamics - Motion Forcing is a novel framework for robust video generation that stabilizes visual quality, physical consistency, and controllability in complex scenes.
- Rethinking Gaussian Trajectory Predictors: Calibrated Uncertainty for Safe Planning (viability: 7): https://sciencetostartup.com/paper/rethinking-gaussian-trajectory-predictors-calibrated-uncertainty-for-safe-planning - A novel loss function for calibrating uncertainty in Gaussian trajectory predictors to enhance safe autonomous navigation.
- Shape Control of a Planar Hyper-Redundant Robot via Hybrid Kinematics-Informed and Learning-based Approach (viability: 6): https://sciencetostartup.com/paper/shape-control-of-a-planar-hyper-redundant-robot-via-hybrid-kinematics-informed-and-learning-based-approach - A hybrid kinematics-informed and learning-based approach for shape control in hyper-redundant robots.
- Designing Service Systems from Textual Evidence (viability: 3): https://sciencetostartup.com/paper/designing-service-systems-from-textual-evidence - An algorithm that optimizes service system configurations by balancing automated evaluations and costly human audits.
- Multi-Person Pose Estimation Evaluation Using Optimal Transportation and Improved Pose Matching (viability: 2): https://sciencetostartup.com/paper/multi-person-pose-estimation-evaluation-using-optimal-transportation-and-improved-pose-matching - A novel metric for evaluating multi-person pose estimation that balances true-positive and false-positive detections.
- On the Learning Dynamics of Two-layer Linear Networks with Label Noise SGD (viability: 4): https://sciencetostartup.com/paper/on-the-learning-dynamics-of-two-layer-linear-networks-with-label-noise-sgd - This research explores the impact of label noise on the learning dynamics of SGD in two-layer linear networks.
- Verbalizing LLM's Higher-order Uncertainty via Imprecise Probabilities (viability: 4): https://sciencetostartup.com/paper/verbalizing-llm-s-higher-order-uncertainty-via-imprecise-probabilities - A novel approach to improve uncertainty elicitation from large language models using imprecise probabilities.
- Graph-GRPO: Training Graph Flow Models with Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/graph-grpo-training-graph-flow-models-with-reinforcement-learning - Graph-GRPO is a reinforcement learning framework that enhances graph flow models for improved generation quality in applications like drug discovery.
- Safe Probabilistic Planning for Human-Robot Interaction using Conformal Risk Control (viability: 7): https://sciencetostartup.com/paper/safe-probabilistic-planning-for-human-robot-interaction-using-conformal-risk-control - A probabilistic safe control framework for enhancing human-robot interaction through formal safety guarantees.
- Variance-Aware Adaptive Weighting for Diffusion Model Training (viability: 4): https://sciencetostartup.com/paper/variance-aware-adaptive-weighting-for-diffusion-model-training - A novel variance-aware adaptive weighting strategy for stable and efficient training of diffusion models.
- ScanDP: Generalizable 3D Scanning with Diffusion Policy (viability: 6): https://sciencetostartup.com/paper/scandp-generalizable-3d-scanning-with-diffusion-policy - A data-efficient 3D scanning framework that utilizes Diffusion Policy for improved robustness and generalization.
- Silent Subversion: Sensor Spoofing Attacks via Supply Chain Implants in Satellite Systems (viability: 4): https://sciencetostartup.com/paper/silent-subversion-sensor-spoofing-attacks-via-supply-chain-implants-in-satellite-systems - A novel approach to countering internal spoofing attacks in satellite systems through supply chain security measures.
- Don't Let the Claw Grip Your Hand: A Security Analysis and Defense Framework for OpenClaw (viability: 6): https://sciencetostartup.com/paper/don-t-let-the-claw-grip-your-hand-a-security-analysis-and-defense-framework-for-openclaw - A security analysis and defense framework for the OpenClaw AI agent platform to mitigate vulnerabilities in code execution.
- Beyond Scalars: Evaluating and Understanding LLM Reasoning via Geometric Progress and Stability (viability: 4): https://sciencetostartup.com/paper/beyond-scalars-evaluating-and-understanding-llm-reasoning-via-geometric-progress-and-stability - TRACED is a framework that evaluates LLM reasoning quality through geometric kinematics.
- Optimal Expert-Attention Allocation in Mixture-of-Experts: A Scalable Law for Dynamic Model Design (viability: 2): https://sciencetostartup.com/paper/optimal-expert-attention-allocation-in-mixture-of-experts-a-scalable-law-for-dynamic-model-design - This paper proposes a new framework for optimizing compute allocation in Mixture-of-Experts models.
- Causal Concept Graphs in LLM Latent Space for Stepwise Reasoning (viability: 3): https://sciencetostartup.com/paper/causal-concept-graphs-in-llm-latent-space-for-stepwise-reasoning - Introducing Causal Concept Graphs for enhanced multi-step reasoning in language models.
- Reactive Writers: How Co-Writing with AI Changes How We Engage with Ideas (viability: 2): https://sciencetostartup.com/paper/reactive-writers-how-co-writing-with-ai-changes-how-we-engage-with-ideas - A study exploring how AI co-writing influences opinions and writing processes.
- Few-Shot Adaptation to Non-Stationary Environments via Latent Trend Embedding for Robotics (viability: 7): https://sciencetostartup.com/paper/few-shot-adaptation-to-non-stationary-environments-via-latent-trend-embedding-for-robotics - A framework for few-shot adaptation in robotics that addresses concept shift without modifying model parameters.
- GeoSense: Internalizing Geometric Necessity Perception for Multimodal Reasoning (viability: 6): https://sciencetostartup.com/paper/geosense-internalizing-geometric-necessity-perception-for-multimodal-reasoning - GeoSense enhances multimodal reasoning by autonomously engaging geometric features based on perceptual necessity.
- Beyond Interleaving: Causal Attention Reformulations for Generative Recommender Systems (viability: 7): https://sciencetostartup.com/paper/beyond-interleaving-causal-attention-reformulations-for-generative-recommender-systems - A novel approach to generative recommender systems that improves efficiency and performance by reformulating causal attention mechanisms.
- Dynamic Knowledge Fusion for Multi-Domain Dialogue State Tracking (viability: 7): https://sciencetostartup.com/paper/dynamic-knowledge-fusion-for-multi-domain-dialogue-state-tracking - A dynamic knowledge fusion framework that enhances dialogue state tracking in multi-domain task-oriented dialogue systems.
- Geometric Autoencoder for Diffusion Models (viability: 8): https://sciencetostartup.com/paper/geometric-autoencoder-for-diffusion-models - Geometric Autoencoder (GAE) optimizes latent space in diffusion models for superior generative performance, surpassing state-of-the-art benchmarks.
- One Token, Two Fates: A Unified Framework via Vision Token Manipulation Against MLLMs Hallucination (viability: 7): https://sciencetostartup.com/paper/one-token-two-fates-a-unified-framework-via-vision-token-manipulation-against-mllms-hallucination - A unified framework that manipulates vision tokens to reduce hallucinations in multi-modal language models.
- HEAL: Hindsight Entropy-Assisted Learning for Reasoning Distillation (viability: 4): https://sciencetostartup.com/paper/heal-hindsight-entropy-assisted-learning-for-reasoning-distillation - HEAL is a novel framework for distilling reasoning capabilities from large models into smaller ones using entropy-assisted learning.
- Utility Function is All You Need: LLM-based Congestion Control (viability: 3): https://sciencetostartup.com/paper/utility-function-is-all-you-need-llm-based-congestion-control - GenCC leverages LLMs to design optimized congestion control protocols for communication networks.
- StyleGallery: Training-free and Semantic-aware Personalized Style Transfer from Arbitrary Image References (viability: 7): https://sciencetostartup.com/paper/stylegallery-training-free-and-semantic-aware-personalized-style-transfer-from-arbitrary-image-references - StyleGallery offers a training-free, semantic-aware framework for personalized image style transfer using arbitrary reference images.
- Adaptive Manipulation Potential and Haptic Estimation for Tool-Mediated Interaction (viability: 7): https://sciencetostartup.com/paper/adaptive-manipulation-potential-and-haptic-estimation-for-tool-mediated-interaction - A closed-loop framework for tool-mediated interaction that enhances haptic estimation and adaptive manipulation.
- Mitigating Translationese Bias in Multilingual LLM-as-a-Judge via Disentangled Information Bottleneck (viability: 7): https://sciencetostartup.com/paper/mitigating-translationese-bias-in-multilingual-llm-as-a-judge-via-disentangled-information-bottleneck - DIBJudge is a fine-tuning framework that mitigates translationese bias in multilingual LLMs, enhancing their evaluation accuracy.
- EmoStory: Emotion-Aware Story Generation (viability: 7): https://sciencetostartup.com/paper/emostory-emotion-aware-story-generation - EmoStory generates emotionally engaging visual narratives with subject consistency.
- On The Complexity of Best-Arm Identification in Non-Stationary Linear Bandits (viability: 2): https://sciencetostartup.com/paper/on-the-complexity-of-best-arm-identification-in-non-stationary-linear-bandits - This paper explores the complexities of identifying the best arm in non-stationary linear bandits.
- Federated Active Learning Under Extreme Non-IID and Global Class Imbalance (viability: 8): https://sciencetostartup.com/paper/federated-active-learning-under-extreme-non-iid-and-global-class-imbalance - FairFAL is an adaptive federated active learning framework that enhances performance in class-imbalanced and non-IID settings.
- Overcoming Visual Clutter in Vision Language Action Models via Concept-Gated Visual Distillation (viability: 7): https://sciencetostartup.com/paper/overcoming-visual-clutter-in-vision-language-action-models-via-concept-gated-visual-distillation - A model-agnostic framework that enhances Vision-Language-Action models by reducing visual clutter for improved robotic manipulation.
- Fuel Gauge: Estimating Chain-of-Thought Length Ahead of Time in Large Multimodal Models (viability: 7): https://sciencetostartup.com/paper/fuel-gauge-estimating-chain-of-thought-length-ahead-of-time-in-large-multimodal-models - Fuel Gauge predicts Chain-of-Thought length in multimodal models to optimize resource use and accuracy.
- Does Reasoning Make Search More Fair? Comparing Fairness in Reasoning and Non-Reasoning Rerankers (viability: 4): https://sciencetostartup.com/paper/does-reasoning-make-search-more-fair-comparing-fairness-in-reasoning-and-non-reasoning-rerankers - This research evaluates the fairness of reasoning versus non-reasoning rerankers in information retrieval.
- PC-Diffuser: Path-Consistent Capsule CBF Safety Filtering for Diffusion-Based Trajectory Planner (viability: 6): https://sciencetostartup.com/paper/pc-diffuser-path-consistent-capsule-cbf-safety-filtering-for-diffusion-based-trajectory-planner - PC-Diffuser enhances safety in autonomous driving trajectory planning by embedding certifiable structures into diffusion models.
- NasoVoce: A Nose-Mounted Low-Audibility Speech Interface for Always-Available Speech Interaction (viability: 7): https://sciencetostartup.com/paper/nasovoce-a-nose-mounted-low-audibility-speech-interface-for-always-available-speech-interaction - NasoVoce is a nose-mounted interface that captures whispered speech for discreet AI voice interaction.
- The Orthogonal Vulnerabilities of Generative AI Watermarks: A Comparative Empirical Benchmark of Spatial and Latent Provenance (viability: 4): https://sciencetostartup.com/paper/the-orthogonal-vulnerabilities-of-generative-ai-watermarks-a-comparative-empirical-benchmark-of-spatial-and-latent-prove - A comparative benchmark study revealing vulnerabilities in generative AI watermarks to enhance digital trust.
- PRoADS: Provably Secure and Robust Audio Diffusion Steganography with latent optimization and backward Euler Inversion (viability: 7): https://sciencetostartup.com/paper/proads-provably-secure-and-robust-audio-diffusion-steganography-with-latent-optimization-and-backward-euler-inversion - PRoADS is a secure audio steganography framework that embeds messages using advanced diffusion models.
- Large language models can disambiguate opioid slang on social media (viability: 7): https://sciencetostartup.com/paper/large-language-models-can-disambiguate-opioid-slang-on-social-media - Leveraging large language models to disambiguate opioid slang on social media for better monitoring of the opioid crisis.
- SteadyTray: Learning Object Balancing Tasks in Humanoid Tray Transport via Residual Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/steadytray-learning-object-balancing-tasks-in-humanoid-tray-transport-via-residual-reinforcement-learning - ReST-RL enhances humanoid stability during payload transport using a novel hierarchical reinforcement learning approach.
- Data-Driven Integration Kernels for Interpretable Nonlocal Operator Learning (viability: 4): https://sciencetostartup.com/paper/data-driven-integration-kernels-for-interpretable-nonlocal-operator-learning - A framework for interpretable nonlocal operator learning in climate models using data-driven integration kernels.
- Is this Idea Novel? An Automated Benchmark for Judgment of Research Ideas (viability: 8): https://sciencetostartup.com/paper/is-this-idea-novel-an-automated-benchmark-for-judgment-of-research-ideas - RINoBench offers an automated benchmark for evaluating the novelty of research ideas, streamlining the assessment process in scientific literature.
- How to make the most of your masked language model for protein engineering (viability: 4): https://sciencetostartup.com/paper/how-to-make-the-most-of-your-masked-language-model-for-protein-engineering - A novel sampling method for masked language models to optimize protein properties in antibody therapeutics.
- What do near-optimal learning rate schedules look like? (viability: 2): https://sciencetostartup.com/paper/what-do-near-optimal-learning-rate-schedules-look-like - This research explores optimal learning rate schedules for neural network training.
- From Imitation to Intuition: Intrinsic Reasoning for Open-Instance Video Classification (viability: 7): https://sciencetostartup.com/paper/from-imitation-to-intuition-intrinsic-reasoning-for-open-instance-video-classification - DeepIntuit transforms video classification by leveraging intrinsic reasoning for better generalization in open-instance scenarios.
- Regime-aware financial volatility forecasting via in-context learning (viability: 7): https://sciencetostartup.com/paper/regime-aware-financial-volatility-forecasting-via-in-context-learning - A regime-aware in-context learning framework using LLMs for accurate financial volatility forecasting.
- GaLoRA: Parameter-Efficient Graph-Aware LLMs for Node Classification (viability: 7): https://sciencetostartup.com/paper/galora-parameter-efficient-graph-aware-llms-for-node-classification - GaLoRA is a parameter-efficient framework that enhances node classification in text-attributed graphs by integrating structural information into large language models.
- Hybrid Self-evolving Structured Memory for GUI Agents (viability: 3): https://sciencetostartup.com/paper/hybrid-self-evolving-structured-memory-for-gui-agents - HyMEM enhances GUI agents with a structured memory system inspired by human cognition.
- MultiwayPAM: Multiway Partitioning Around Medoids for LLM-as-a-Judge Score Analysis (viability: 5): https://sciencetostartup.com/paper/multiwaypam-multiway-partitioning-around-medoids-for-llm-as-a-judge-score-analysis - MultiwayPAM enhances LLM evaluation by revealing score bias through advanced tensor clustering.
- Conversational AI-Enhanced Exploration System to Query Large-Scale Digitised Collections of Natural History Museums (viability: 7): https://sciencetostartup.com/paper/conversational-ai-enhanced-exploration-system-to-query-large-scale-digitised-collections-of-natural-history-museums - A conversational AI system that enhances public access to natural history museum collections through natural language queries.
- Copula-ResLogit: A Deep-Copula Framework for Unobserved Confounding Effects (viability: 5): https://sciencetostartup.com/paper/copula-reslogit-a-deep-copula-framework-for-unobserved-confounding-effects - Copula-ResLogit is a deep learning framework that addresses unobserved confounding effects in travel demand analysis.
- GSVD for Geometry-Grounded Dataset Comparison: An Alignment Angle Is All You Need (viability: 4): https://sciencetostartup.com/paper/gsvd-for-geometry-grounded-dataset-comparison-an-alignment-angle-is-all-you-need - A geometric diagnostic tool for dataset comparison using GSVD to quantify dataset relationships.
- Update-Free On-Policy Steering via Verifiers (viability: 7): https://sciencetostartup.com/paper/update-free-on-policy-steering-via-verifiers - UF-OPS enhances robot action success rates by predicting and steering actions in real-time without parameter changes.
- Taming Score-Based Denoisers in ADMM: A Convergent Plug-and-Play Framework (viability: 4): https://sciencetostartup.com/paper/taming-score-based-denoisers-in-admm-a-convergent-plug-and-play-framework - A novel framework integrating score-based denoisers into ADMM for improved convergence in inverse problems.
- Robust Post-Training for Generative Recommenders: Why Exponential Reward-Weighted SFT Outperforms RLHF (viability: 7): https://sciencetostartup.com/paper/robust-post-training-for-generative-recommenders-why-exponential-reward-weighted-sft-outperforms-rlhf - A novel post-training method for generative recommenders that optimizes directly on observed rewards, outperforming traditional RLHF approaches.
- Estimating condition number with Graph Neural Networks (viability: 3): https://sciencetostartup.com/paper/estimating-condition-number-with-graph-neural-networks - A novel approach using graph neural networks to estimate the condition number of sparse matrices efficiently.
- Post-Quantum Entropy as a Service for Embedded Systems (viability: 7): https://sciencetostartup.com/paper/post-quantum-entropy-as-a-service-for-embedded-systems - A Quantum Entropy as a Service system providing secure entropy for embedded devices.
- A Robust Deep Learning Framework for Bangla License Plate Recognition Using YOLO and Vision-Language OCR (viability: 7): https://sciencetostartup.com/paper/a-robust-deep-learning-framework-for-bangla-license-plate-recognition-using-yolo-and-vision-language-ocr - A robust Bangla License Plate Recognition system leveraging YOLO and Vision-Language OCR for intelligent traffic management.
- Design of a Robot-Assisted Chemical Dialysis System (viability: 2): https://sciencetostartup.com/paper/design-of-a-robot-assisted-chemical-dialysis-system - A robot-assisted system designed to automate tedious dialysis procedures in scientific research.
- From Prior to Pro: Efficient Skill Mastery via Distribution Contractive RL Finetuning (viability: 7): https://sciencetostartup.com/paper/from-prior-to-pro-efficient-skill-mastery-via-distribution-contractive-rl-finetuning - DICE-RL refines pretrained generative robot policies for efficient skill mastery using reinforcement learning.
- Discovery of a Hematopoietic Manifold in scGPT Yields a Method for Extracting Performant Algorithms from Biological Foundation Model Internals (viability: 7): https://sciencetostartup.com/paper/discovery-of-a-hematopoietic-manifold-in-scgpt-yields-a-method-for-extracting-performant-algorithms-from-biological-foun - A novel method for extracting efficient algorithms from biological foundation models, enhancing performance in hematopoietic analysis.
- ID-LoRA: Identity-Driven Audio-Video Personalization with In-Context LoRA (viability: 8): https://sciencetostartup.com/paper/id-lora-identity-driven-audio-video-personalization-with-in-context-lora - ID-LoRA personalizes audio and video together using a single model driven by text prompts and reference media.
- Improving TabPFN's Synthetic Data Generation by Integrating Causal Structure (viability: 6): https://sciencetostartup.com/paper/improving-tabpfn-s-synthetic-data-generation-by-integrating-causal-structure - Enhancing synthetic tabular data generation by integrating causal structures into existing models.
- Joint Imaging-ROI Representation Learning via Cross-View Contrastive Alignment for Brain Disorder Classification (viability: 7): https://sciencetostartup.com/paper/joint-imaging-roi-representation-learning-via-cross-view-contrastive-alignment-for-brain-disorder-classification - A unified framework for enhancing brain disorder classification through joint imaging and ROI representation learning.
- Bayesian Hierarchical Models and the Maximum Entropy Principle (viability: 3): https://sciencetostartup.com/paper/bayesian-hierarchical-models-and-the-maximum-entropy-principle - A theoretical exploration of Bayesian hierarchical models and maximum entropy distributions.
- SiMPO: Measure Matching for Online Diffusion Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/simpo-measure-matching-for-online-diffusion-reinforcement-learning - SiMPO is a novel framework for improving diffusion reinforcement learning through signed measure optimization.
- DUCTILE: Agentic LLM Orchestration of Engineering Analysis in Product Development Practice (viability: 8): https://sciencetostartup.com/paper/ductile-agentic-llm-orchestration-of-engineering-analysis-in-product-development-practice - DUCTILE automates engineering analysis through LLM orchestration, adapting to evolving product requirements.
- Degeneracy-Resilient Teach and Repeat for Geometrically Challenging Environments Using FMCW Lidar (viability: 7): https://sciencetostartup.com/paper/degeneracy-resilient-teach-and-repeat-for-geometrically-challenging-environments-using-fmcw-lidar - A robust Teach and Repeat navigation system for robots using FMCW lidar to enhance performance in GPS-denied environments.
- Intrinsic Numerical Robustness and Fault Tolerance in a Neuromorphic Algorithm for Scientific Computing (viability: 3): https://sciencetostartup.com/paper/intrinsic-numerical-robustness-and-fault-tolerance-in-a-neuromorphic-algorithm-for-scientific-computing - A neuromorphic algorithm that demonstrates intrinsic fault tolerance for scientific computing applications.
- GR-SAP: Generative Replay for Safety Alignment Preservation during Fine-Tuning (viability: 8): https://sciencetostartup.com/paper/gr-sap-generative-replay-for-safety-alignment-preservation-during-fine-tuning - GR-SAP is a framework that synthesizes domain-specific alignment data to preserve safety alignment in fine-tuning large language models.
- ACE Runtime - A ZKP-Native Blockchain Runtime with Sub-Second Cryptographic Finality (viability: 2): https://sciencetostartup.com/paper/ace-runtime-a-zkp-native-blockchain-runtime-with-sub-second-cryptographic-finality - ACE Runtime offers a ZKP-native blockchain execution layer for sub-second cryptographic finality.
- One Adapter for All: Towards Unified Representation in Step-Imbalanced Class-Incremental Learning (viability: 7): https://sciencetostartup.com/paper/one-adapter-for-all-towards-unified-representation-in-step-imbalanced-class-incremental-learning - One-A is a unified framework for class-incremental learning that adapts to imbalanced task streams while maintaining efficiency.
- Why Does It Look There? Structured Explanations for Image Classification (viability: 5): https://sciencetostartup.com/paper/why-does-it-look-there-structured-explanations-for-image-classification - I2X provides structured explanations for image classification models, enhancing interpretability and guiding optimization.
- S-GRADES -- Studying Generalization of Student Response Assessments in Diverse Evaluative Settings (viability: 7): https://sciencetostartup.com/paper/s-grades-studying-generalization-of-student-response-assessments-in-diverse-evaluative-settings - S-GRADES is a web-based benchmark for evaluating student responses across diverse grading datasets with standardized protocols.
- Hierarchical Task Model Predictive Control for Sequential Mobile Manipulation Tasks (viability: 7): https://sciencetostartup.com/paper/hierarchical-task-model-predictive-control-for-sequential-mobile-manipulation-tasks - A novel control framework for mobile manipulators that enhances task execution efficiency and performance.
- OilSAM2: Memory-Augmented SAM2 for Scalable SAR Oil Spill Detection (viability: 8): https://sciencetostartup.com/paper/oilsam2-memory-augmented-sam2-for-scalable-sar-oil-spill-detection - OilSAM2 is a memory-augmented segmentation framework designed for accurate oil spill detection in SAR imagery.
- Paladin: A Policy Framework for Securing Cloud APIs by Combining Application Context with Generative AI (viability: 3): https://sciencetostartup.com/paper/paladin-a-policy-framework-for-securing-cloud-apis-by-combining-application-context-with-generative-ai - Paladin is a security framework that leverages generative AI to enforce application context-aware policies for cloud APIs.
- Perceptive Hierarchical-Task MPC for Sequential Mobile Manipulation in Unstructured Semi-Static Environments (viability: 3): https://sciencetostartup.com/paper/perceptive-hierarchical-task-mpc-for-sequential-mobile-manipulation-in-unstructured-semi-static-environments - A novel framework for efficient sequential mobile manipulation in dynamic environments using perceptive hierarchical-task model predictive control.
- Rethinking the Harmonic Loss via Non-Euclidean Distance Layers (viability: 7): https://sciencetostartup.com/paper/rethinking-the-harmonic-loss-via-non-euclidean-distance-layers - A novel approach to loss functions that improves interpretability and sustainability in deep learning models.
- Robotic Ultrasound Makes CBCT Alive (viability: 8): https://sciencetostartup.com/paper/robotic-ultrasound-makes-cbct-alive - A real-time deformation-aware framework that updates static CBCT images using robotic ultrasound for enhanced surgical navigation.
- A Diffusion Analysis of Policy Gradient for Stochastic Bandits (viability: 2): https://sciencetostartup.com/paper/a-diffusion-analysis-of-policy-gradient-for-stochastic-bandits - This paper presents a theoretical analysis of policy gradient methods for stochastic bandits.
- Multilingual AI-Driven Password Strength Estimation with Similarity-Based Detection (viability: 6): https://sciencetostartup.com/paper/multilingual-ai-driven-password-strength-estimation-with-similarity-based-detection - A multilingual password strength meter that leverages AI-generated data to enhance password security for non-English users.
- An Automated Radiomics Framework for Postoperative Survival Prediction in Colorectal Liver Metastases using Preoperative MRI (viability: 3): https://sciencetostartup.com/paper/an-automated-radiomics-framework-for-postoperative-survival-prediction-in-colorectal-liver-metastases-using-preoperative - An AI framework for predicting postoperative survival in colorectal liver metastases using MRI data.
- Sabiá-4 Technical Report (viability: 6): https://sciencetostartup.com/paper/sabi-4-technical-report - Sabiá-4 is a new generation of Portuguese language models optimized for legal and conversational tasks.
- FusionNet: a frame interpolation network for 4D heart models (viability: 8): https://sciencetostartup.com/paper/fusionnet-a-frame-interpolation-network-for-4d-heart-models - FusionNet enhances cardiac imaging by providing high-resolution 4D heart models from short CMR scans.
- ViDia2Std: A Parallel Corpus and Methods for Low-Resource Vietnamese Dialect-to-Standard Translation (viability: 7): https://sciencetostartup.com/paper/vidia2std-a-parallel-corpus-and-methods-for-low-resource-vietnamese-dialect-to-standard-translation - ViDia2Std is a comprehensive parallel corpus for dialect-to-standard Vietnamese translation, enhancing NLP performance across diverse dialects.
- Delta-K: Boosting Multi-Instance Generation via Cross-Attention Augmentation (viability: 7): https://sciencetostartup.com/paper/delta-k-boosting-multi-instance-generation-via-cross-attention-augmentation - Delta-K enhances multi-instance generation in diffusion models by improving semantic representation through cross-attention augmentation.
- Actor-Accelerated Policy Dual Averaging for Reinforcement Learning in Continuous Action Spaces (viability: 6): https://sciencetostartup.com/paper/actor-accelerated-policy-dual-averaging-for-reinforcement-learning-in-continuous-action-spaces - Actor-accelerated PDA enhances reinforcement learning in continuous action spaces by approximating optimization sub-problems for faster runtimes.
- Octopus-inspired Distributed Control for Soft Robotic Arms: A Graph Neural Network-Based Attention Policy with Environmental Interaction (viability: 3): https://sciencetostartup.com/paper/octopus-inspired-distributed-control-for-soft-robotic-arms-a-graph-neural-network-based-attention-policy-with-environmen - SoftGM is a novel octopus-inspired control architecture for soft robotic arms that enhances target reaching in complex environments.
- Adaptive Activation Cancellation for Hallucination Mitigation in Large Language Models (viability: 8): https://sciencetostartup.com/paper/adaptive-activation-cancellation-for-hallucination-mitigation-in-large-language-models - Adaptive Activation Cancellation is a real-time framework that mitigates hallucinations in large language models without external knowledge or fine-tuning.
- MCP-in-SoS: Risk assessment framework for open-source MCP servers (viability: 4): https://sciencetostartup.com/paper/mcp-in-sos-risk-assessment-framework-for-open-source-mcp-servers - A risk assessment framework for identifying vulnerabilities in open-source Model Context Protocol servers.
- Stability and Robustness via Regularization: Bandit Inference via Regularized Stochastic Mirror Descent (viability: 4): https://sciencetostartup.com/paper/stability-and-robustness-via-regularization-bandit-inference-via-regularized-stochastic-mirror-descent - This paper presents a novel approach to bandit inference that ensures stability and robustness through regularization techniques.
- DT-BEHRT: Disease Trajectory-aware Transformer for Interpretable Patient Representation Learning (viability: 8): https://sciencetostartup.com/paper/dt-behrt-disease-trajectory-aware-transformer-for-interpretable-patient-representation-learning - DT-BEHRT leverages a graph-enhanced transformer for interpretable patient representation learning from electronic health records.
- Video-Based Reward Modeling for Computer-Use Agents (viability: 8): https://sciencetostartup.com/paper/video-based-reward-modeling-for-computer-use-agents - A scalable video-based reward modeling system for evaluating computer-using agents' task success.
- Autonomous Search for Sparsely Distributed Visual Phenomena through Environmental Context Modeling (viability: 7): https://sciencetostartup.com/paper/autonomous-search-for-sparsely-distributed-visual-phenomena-through-environmental-context-modeling - AUVs use environmental context modeling for efficient coral species detection.
- Characterizing Healthy & Post-Stroke Neuromotor Behavior During 6D Upper-Limb Isometric Gaming: Implications for Design of End-Effector Rehabilitation Robot Interfaces (viability: 5): https://sciencetostartup.com/paper/characterizing-healthy-post-stroke-neuromotor-behavior-during-6d-upper-limb-isometric-gaming-implications-for-design-of- - A novel approach to characterizing neuromotor behavior for enhancing rehabilitation robot interfaces.
- Dance2Hesitate: A Multi-Modal Dataset of Dancer-Taught Hesitancy for Understandable Robot Motion (viability: 7): https://sciencetostartup.com/paper/dance2hesitate-a-multi-modal-dataset-of-dancer-taught-hesitancy-for-understandable-robot-motion - Dance2Hesitate offers a unique dataset for developing robots that can express hesitancy in human-robot interactions.
- OpenClaw-RL: Train Any Agent Simply by Talking (viability: 9): https://sciencetostartup.com/paper/openclaw-rl-train-any-agent-simply-by-talking - OpenClaw-RL enables agents to learn from user interactions in real-time, enhancing their performance through continuous feedback.
- Compatibility at a Cost: Systematic Discovery and Exploitation of MCP Clause-Compliance Vulnerabilities (viability: 2): https://sciencetostartup.com/paper/compatibility-at-a-cost-systematic-discovery-and-exploitation-of-mcp-clause-compliance-vulnerabilities - A framework for analyzing vulnerabilities in the Model Context Protocol for AI interoperability.
- ReMix: Reinforcement routing for mixtures of LoRAs in LLM finetuning (viability: 7): https://sciencetostartup.com/paper/remix-reinforcement-routing-for-mixtures-of-loras-in-llm-finetuning - ReMix introduces a novel reinforcement routing technique for Mixture-of-LoRAs to enhance the efficiency of LLM finetuning.
- Cross-Hand Latent Representation for Vision-Language-Action Models (viability: 4): https://sciencetostartup.com/paper/cross-hand-latent-representation-for-vision-language-action-models - XL-VLA enables scalable cross-embodiment learning for vision-language-action models in robotic manipulation.
- Mashup Learning: Faster Finetuning by Remixing Past Checkpoints (viability: 7): https://sciencetostartup.com/paper/mashup-learning-faster-finetuning-by-remixing-past-checkpoints - Mashup Learning enhances LLM performance by remixing past checkpoints for faster finetuning.
- A neural operator for predicting vibration frequency response curves from limited data (viability: 5): https://sciencetostartup.com/paper/a-neural-operator-for-predicting-vibration-frequency-response-curves-from-limited-data - A neural operator that predicts vibration frequency response curves from limited data, enhancing design iteration efficiency.
- Social Knowledge for Cross-Domain User Preference Modeling (viability: 7): https://sciencetostartup.com/paper/social-knowledge-for-cross-domain-user-preference-modeling - A novel approach to predict user preferences across domains using social embeddings from Twitter data.
- Lost in Backpropagation: The LM Head is a Gradient Bottleneck (viability: 2): https://sciencetostartup.com/paper/lost-in-backpropagation-the-lm-head-is-a-gradient-bottleneck - This paper analyzes the gradient bottleneck in language models and its impact on training efficiency.
- Reason and Verify: A Framework for Faithful Retrieval-Augmented Generation (viability: 8): https://sciencetostartup.com/paper/reason-and-verify-a-framework-for-faithful-retrieval-augmented-generation - A domain-specific framework for enhancing the factuality of Retrieval-Augmented Generation in high-stakes domains through explicit reasoning and verification.
- The Generation-Recognition Asymmetry: Six Dimensions of a Fundamental Divide in Formal Language Theory (viability: 2): https://sciencetostartup.com/paper/the-generation-recognition-asymmetry-six-dimensions-of-a-fundamental-divide-in-formal-language-theory - This paper explores the asymmetry between generation and recognition in formal language theory, highlighting its implications for NLP and compiler design.
- Agentic Control Center for Data Product Optimization (viability: 2): https://sciencetostartup.com/paper/agentic-control-center-for-data-product-optimization - An AI system that automates the improvement of data products through continuous optimization and human oversight.
- Unbalanced Optimal Transport Dictionary Learning for Unsupervised Hyperspectral Image Clustering (viability: 2): https://sciencetostartup.com/paper/unbalanced-optimal-transport-dictionary-learning-for-unsupervised-hyperspectral-image-clustering - This paper proposes a novel method for unsupervised clustering of hyperspectral images using unbalanced optimal transport.
- The Prediction-Measurement Gap: Toward Meaning Representations as Scientific Instruments (viability: 2): https://sciencetostartup.com/paper/the-prediction-measurement-gap-toward-meaning-representations-as-scientific-instruments - This paper explores the limitations of current text embeddings in scientific measurement and proposes a new framework for meaning representation.
- HG-Lane: High-Fidelity Generation of Lane Scenes under Adverse Weather and Lighting Conditions without Re-annotation (viability: 8): https://sciencetostartup.com/paper/hg-lane-high-fidelity-generation-of-lane-scenes-under-adverse-weather-and-lighting-conditions-without-re-annotation - HG-Lane generates high-fidelity lane scenes under adverse conditions to improve autonomous vehicle safety without re-annotation.
- AR-VLA: True Autoregressive Action Expert for Vision-Language-Action Models (viability: 7): https://sciencetostartup.com/paper/ar-vla-true-autoregressive-action-expert-for-vision-language-action-models - AR-VLA is a context-aware autoregressive action generator for robotic manipulation tasks that enhances action trajectory smoothness and task success rates.
- 4DEquine: Disentangling Motion and Appearance for 4D Equine Reconstruction from Monocular Video (viability: 8): https://sciencetostartup.com/paper/4dequine-disentangling-motion-and-appearance-for-4d-equine-reconstruction-from-monocular-video - 4DEquine offers a novel approach to 4D reconstruction of equines from monocular video, enhancing animal welfare through advanced motion and appearance modeling.
- Lost in the Middle at Birth: An Exact Theory of Transformer Position Bias (viability: 2): https://sciencetostartup.com/paper/lost-in-the-middle-at-birth-an-exact-theory-of-transformer-position-bias - This paper explores the inherent geometric properties of transformer position bias in LLMs.
- Task Aware Modulation Using Representation Learning for Upsaling of Terrestrial Carbon Fluxes (viability: 4): https://sciencetostartup.com/paper/task-aware-modulation-using-representation-learning-for-upsaling-of-terrestrial-carbon-fluxes - TAM-RL enhances the accuracy of terrestrial carbon flux estimates through advanced representation learning techniques.
- From Data Statistics to Feature Geometry: How Correlations Shape Superposition (viability: 7): https://sciencetostartup.com/paper/from-data-statistics-to-feature-geometry-how-correlations-shape-superposition - Introducing Bag-of-Words Superposition to enhance understanding of feature interactions in neural networks.
- TiPToP: A Modular Open-Vocabulary Planning System for Robotic Manipulation (viability: 8): https://sciencetostartup.com/paper/tiptop-a-modular-open-vocabulary-planning-system-for-robotic-manipulation - TiPToP is a modular open-vocabulary planning system that enables robotic manipulation from images and natural language instructions.
- CREATE: Testing LLMs for Associative Creativity (viability: 4): https://sciencetostartup.com/paper/create-testing-llms-for-associative-creativity - CREATE is a benchmark for evaluating LLMs' ability to generate creative associative reasoning paths.
- ReCoSplat: Autoregressive Feed-Forward Gaussian Splatting Using Render-and-Compare (viability: 8): https://sciencetostartup.com/paper/recosplat-autoregressive-feed-forward-gaussian-splatting-using-render-and-compare - ReCoSplat is an innovative model for online novel view synthesis that enhances scene reconstruction from unposed observations.
- Understanding the Use of a Large Language Model-Powered Guide to Make Virtual Reality Accessible for Blind and Low Vision People (viability: 3): https://sciencetostartup.com/paper/understanding-the-use-of-a-large-language-model-powered-guide-to-make-virtual-reality-accessible-for-blind-and-low-visio - A large language model-powered guide to enhance virtual reality accessibility for blind and low vision users.
- Emotional Modulation in Swarm Decision Dynamics (viability: 2): https://sciencetostartup.com/paper/emotional-modulation-in-swarm-decision-dynamics - A theoretical framework exploring emotional modulation in collective decision-making dynamics.
- BEACON: Language-Conditioned Navigation Affordance Prediction under Occlusion (viability: 8): https://sciencetostartup.com/paper/beacon-language-conditioned-navigation-affordance-prediction-under-occlusion - BEACON enhances robot navigation by predicting traversable locations in occluded environments using language instructions and depth data.
- Think Before You Lie: How Reasoning Improves Honesty (viability: 4): https://sciencetostartup.com/paper/think-before-you-lie-how-reasoning-improves-honesty - A study exploring how reasoning in LLMs can enhance honesty by navigating the representational space of deceptive and honest responses.
- Kinodynamic Motion Retargeting for Humanoid Locomotion via Multi-Contact Whole-Body Trajectory Optimization (viability: 7): https://sciencetostartup.com/paper/kinodynamic-motion-retargeting-for-humanoid-locomotion-via-multi-contact-whole-body-trajectory-optimization - KinoDynamic Motion Retargeting optimizes humanoid locomotion through advanced trajectory techniques, enhancing stability and performance.
- From Semantics to Pixels: Coarse-to-Fine Masked Autoencoders for Hierarchical Visual Understanding (viability: 4): https://sciencetostartup.com/paper/from-semantics-to-pixels-coarse-to-fine-masked-autoencoders-for-hierarchical-visual-understanding - C2FMAE is a novel coarse-to-fine masked autoencoder that enhances visual representation learning across multiple data granularities.
- Leveraging whole slide difficulty in Multiple Instance Learning to improve prostate cancer grading (viability: 2): https://sciencetostartup.com/paper/leveraging-whole-slide-difficulty-in-multiple-instance-learning-to-improve-prostate-cancer-grading - This paper proposes methods to improve prostate cancer grading by addressing the difficulty of Whole Slide Images in histopathology.
- On the Width Scaling of Neural Optimizers Under Matrix Operator Norms I: Row/Column Normalization and Hyperparameter Transfer (viability: 3): https://sciencetostartup.com/paper/on-the-width-scaling-of-neural-optimizers-under-matrix-operator-norms-i-row-column-normalization-and-hyperparameter-tran - MOGA introduces a new family of width-aware optimizers for stable learning-rate transfer across model widths.
- Towards a Neural Debugger for Python (viability: 7): https://sciencetostartup.com/paper/towards-a-neural-debugger-for-python - Introducing neural debuggers that emulate traditional debugging tools for enhanced code execution and prediction.
- When Learning Rates Go Wrong: Early Structural Signals in PPO Actor-Critic (viability: 2): https://sciencetostartup.com/paper/when-learning-rates-go-wrong-early-structural-signals-in-ppo-actor-critic - This paper analyzes the impact of learning rates on PPO actor-critic methods using a new metric for early screening of training runs.
- The Confidence Gate Theorem: When Should Ranked Decision Systems Abstain? (viability: 2): https://sciencetostartup.com/paper/the-confidence-gate-theorem-when-should-ranked-decision-systems-abstain - This paper explores when ranked decision systems should abstain from intervention to improve decision quality.
- No Image, No Problem: End-to-End Multi-Task Cardiac Analysis from Undersampled k-Space (viability: 7): https://sciencetostartup.com/paper/no-image-no-problem-end-to-end-multi-task-cardiac-analysis-from-undersampled-k-space - k-MTR is a novel framework for direct cardiac analysis from undersampled k-space data, enhancing diagnostic efficiency.
- PathMem: Toward Cognition-Aligned Memory Transformation for Pathology MLLMs (viability: 8): https://sciencetostartup.com/paper/pathmem-toward-cognition-aligned-memory-transformation-for-pathology-mllms - PathMem is a memory-centric multimodal framework that enhances pathology MLLMs by integrating structured knowledge for improved diagnostic reasoning.
- SignalMC-MED: A Multimodal Benchmark for Evaluating Biosignal Foundation Models on Single-Lead ECG and PPG (viability: 5): https://sciencetostartup.com/paper/signalmc-med-a-multimodal-benchmark-for-evaluating-biosignal-foundation-models-on-single-lead-ecg-and-ppg - SignalMC-MED is a benchmark for evaluating biosignal foundation models on ECG and PPG data to enhance clinical prediction tasks.
- Model Merging in the Era of Large Language Models: Methods, Applications, and Future Directions (viability: 4): https://sciencetostartup.com/paper/model-merging-in-the-era-of-large-language-models-methods-applications-and-future-directions - A comprehensive survey on model merging techniques for large language models, offering efficient alternatives to traditional ensemble methods.
- Generative Drifting is Secretly Score Matching: a Spectral and Variational Perspective (viability: 4): https://sciencetostartup.com/paper/generative-drifting-is-secretly-score-matching-a-spectral-and-variational-perspective - A theoretical framework for generative modeling via drifting that enhances understanding and performance of image generation.
- Unsupervised Domain Adaptation with Target-Only Margin Disparity Discrepancy (viability: 7): https://sciencetostartup.com/paper/unsupervised-domain-adaptation-with-target-only-margin-disparity-discrepancy - A novel unsupervised domain adaptation framework that enhances liver segmentation in interventional CBCT using limited annotated data.
- Adaptive Clinical-Aware Latent Diffusion for Multimodal Brain Image Generation and Missing Modality Imputation (viability: 8): https://sciencetostartup.com/paper/adaptive-clinical-aware-latent-diffusion-for-multimodal-brain-image-generation-and-missing-modality-imputation - AI framework using adaptive clinical-aware diffusion for generating complete brain imaging modalities in Alzheimer's diagnosis.
- Fine-grained Motion Retrieval via Joint-Angle Motion Images and Token-Patch Late Interaction (viability: 6): https://sciencetostartup.com/paper/fine-grained-motion-retrieval-via-joint-angle-motion-images-and-token-patch-late-interaction - A novel approach for text-motion retrieval that enhances accuracy and interpretability through joint-angle motion representation.
- On the Structural Failure of Chamfer Distance in 3D Shape Optimization (viability: 2): https://sciencetostartup.com/paper/on-the-structural-failure-of-chamfer-distance-in-3d-shape-optimization - This paper critiques the Chamfer distance in 3D shape optimization, revealing its structural failures and proposing a design criterion for improvement.
- OptEMA: Adaptive Exponential Moving Average for Stochastic Optimization with Zero-Noise Optimality (viability: 2): https://sciencetostartup.com/paper/optema-adaptive-exponential-moving-average-for-stochastic-optimization-with-zero-noise-optimality - OptEMA introduces a novel adaptive exponential moving average for stochastic optimization, improving convergence rates without requiring Lipschitz constants.
- WikiCLIP: An Efficient Contrastive Baseline for Open-domain Visual Entity Recognition (viability: 8): https://sciencetostartup.com/paper/wikiclip-an-efficient-contrastive-baseline-for-open-domain-visual-entity-recognition - WikiCLIP offers an efficient contrastive approach for open-domain visual entity recognition, significantly improving performance while reducing computational costs.
- Role Classification of Hosts within Enterprise Networks Based on Connection Patterns (viability: 6): https://sciencetostartup.com/paper/role-classification-of-hosts-within-enterprise-networks-based-on-connection-patterns - Automated role classification algorithms enhance network management and security by grouping hosts based on connection patterns.
- MedMASLab: A Unified Orchestration Framework for Benchmarking Multimodal Medical Multi-Agent Systems (viability: 7): https://sciencetostartup.com/paper/medmaslab-a-unified-orchestration-framework-for-benchmarking-multimodal-medical-multi-agent-systems - Develop MedMASLab, a framework for standardized, multimodal medical agent collaboration and benchmarking.
- NanoBench: A Multi-Task Benchmark Dataset for Nano-Quadrotor System Identification, Control, and State Estimation (viability: 6): https://sciencetostartup.com/paper/nanobench-a-multi-task-benchmark-dataset-for-nano-quadrotor-system-identification-control-and-state-estimation - NanoBench is an open-source benchmark dataset for advancing nano-quadrotor system identification and control.
- Thinking to Recall: How Reasoning Unlocks Parametric Knowledge in LLMs (viability: 5): https://sciencetostartup.com/paper/thinking-to-recall-how-reasoning-unlocks-parametric-knowledge-in-llms - Unlocking reasoning in LLMs enhances parametric knowledge recall for improved accuracy.
- Stepping VLMs onto the Court: Benchmarking Spatial Intelligence in Sports (viability: 7): https://sciencetostartup.com/paper/stepping-vlms-onto-the-court-benchmarking-spatial-intelligence-in-sports - CourtSI is a large-scale dataset and benchmark for enhancing spatial intelligence in vision-language models within sports contexts.
- MSSR: Memory-Aware Adaptive Replay for Continual LLM Fine-Tuning (viability: 8): https://sciencetostartup.com/paper/mssr-memory-aware-adaptive-replay-for-continual-llm-fine-tuning - MSSR is an adaptive replay framework for continual fine-tuning of LLMs that mitigates catastrophic forgetting while ensuring rapid adaptation.
- Influencing LLM Multi-Agent Dialogue via Policy-Parameterized Prompts (viability: 3): https://sciencetostartup.com/paper/influencing-llm-multi-agent-dialogue-via-policy-parameterized-prompts - This research explores the use of parameterized prompts to influence dialogue dynamics in multi-agent systems.
- LCA: Local Classifier Alignment for Continual Learning (viability: 7): https://sciencetostartup.com/paper/lca-local-classifier-alignment-for-continual-learning - LCA introduces a novel loss function to enhance classifier alignment in continual learning, mitigating catastrophic forgetting.
- Robust Cooperative Localization in Featureless Environments: A Comparative Study of DCL, StCL, CCL, CI, and Standard-CL (viability: 4): https://sciencetostartup.com/paper/robust-cooperative-localization-in-featureless-environments-a-comparative-study-of-dcl-stcl-ccl-ci-and-standard-cl - A comparative study of cooperative localization methods for multi-robot systems in GPS-denied environments.
- Benchmarking Political Persuasion Risks Across Frontier Large Language Models (viability: 4): https://sciencetostartup.com/paper/benchmarking-political-persuasion-risks-across-frontier-large-language-models - A framework for assessing the persuasive risks of large language models in political contexts.
- DISPLAY: Directable Human-Object Interaction Video Generation via Sparse Motion Guidance and Multi-Task Auxiliary (viability: 7): https://sciencetostartup.com/paper/display-directable-human-object-interaction-video-generation-via-sparse-motion-guidance-and-multi-task-auxiliary - DISPLAY enables intuitive and controllable human-object interaction video generation using sparse motion guidance.
- Emerging Extrinsic Dexterity in Cluttered Scenes via Dynamics-aware Policy Learning (viability: 7): https://sciencetostartup.com/paper/emerging-extrinsic-dexterity-in-cluttered-scenes-via-dynamics-aware-policy-learning - A framework for enhancing extrinsic dexterity in cluttered environments through dynamics-aware policy learning.
- Do What I Say: A Spoken Prompt Dataset for Instruction-Following (viability: 4): https://sciencetostartup.com/paper/do-what-i-say-a-spoken-prompt-dataset-for-instruction-following - A multilingual dataset for evaluating speech large language models with human-recorded prompts.
- InternVL-U: Democratizing Unified Multimodal Models for Understanding, Reasoning, Generation and Editing (viability: 7): https://sciencetostartup.com/paper/internvl-u-democratizing-unified-multimodal-models-for-understanding-reasoning-generation-and-editing - InternVL-U is a lightweight multimodal model that excels in understanding, reasoning, generation, and editing tasks.
- The Bureaucracy of Speed: Structural Equivalence Between Memory Consistency Models and Multi-Agent Authorization Revocation (viability: 8): https://sciencetostartup.com/paper/the-bureaucracy-of-speed-structural-equivalence-between-memory-consistency-models-and-multi-agent-authorization-revocati - A Capability Coherence System that significantly reduces unauthorized API calls in high-velocity environments.
- MissBench: Benchmarking Multimodal Affective Analysis under Imbalanced Missing Modalities (viability: 7): https://sciencetostartup.com/paper/missbench-benchmarking-multimodal-affective-analysis-under-imbalanced-missing-modalities - MissBench provides a standardized framework for benchmarking multimodal affective tasks under imbalanced missing modalities.
- N-gram-like Language Models Predict Reading Time Best (viability: 4): https://sciencetostartup.com/paper/n-gram-like-language-models-predict-reading-time-best - This research explores how n-gram-like language models can better predict reading time compared to complex transformer models.
- CarbonBench: A Global Benchmark for Upscaling of Carbon Fluxes Using Zero-Shot Learning (viability: 4): https://sciencetostartup.com/paper/carbonbench-a-global-benchmark-for-upscaling-of-carbon-fluxes-using-zero-shot-learning - CarbonBench is a benchmark for evaluating zero-shot learning models in carbon flux upscaling across diverse ecosystems.
- GAST: Gradient-aligned Sparse Tuning of Large Language Models with Data-layer Selection (viability: 4): https://sciencetostartup.com/paper/gast-gradient-aligned-sparse-tuning-of-large-language-models-with-data-layer-selection - GAST is a novel method for parameter-efficient fine-tuning of large language models that optimally selects impactful data points and layers.
- A Graph-Based Approach to Spectrum Demand Prediction Using Hierarchical Attention Networks (viability: 3): https://sciencetostartup.com/paper/a-graph-based-approach-to-spectrum-demand-prediction-using-hierarchical-attention-networks - HR-GAT is a graph-based model for predicting spectrum demand using geospatial data.
- SCENEBench: An Audio Understanding Benchmark Grounded in Assistive and Industrial Use Cases (viability: 4): https://sciencetostartup.com/paper/scenebench-an-audio-understanding-benchmark-grounded-in-assistive-and-industrial-use-cases - SCENEBench is a benchmark suite designed to evaluate audio understanding across various real-world categories, addressing gaps in current audio processing models.
- A Unified Hierarchical Multi-Task Multi-Fidelity Framework for Data-Efficient Surrogate Modeling in Manufacturing (viability: 6): https://sciencetostartup.com/paper/a-unified-hierarchical-multi-task-multi-fidelity-framework-for-data-efficient-surrogate-modeling-in-manufacturing - A novel hierarchical framework for data-efficient surrogate modeling in manufacturing that improves prediction accuracy across tasks and fidelity levels.
- Chow-Liu Ordering for Long-Context Reasoning in Chain-of-Agents (viability: 5): https://sciencetostartup.com/paper/chow-liu-ordering-for-long-context-reasoning-in-chain-of-agents - Chow-Liu Ordering enhances long-context reasoning in multi-agent frameworks by optimizing chunk processing order.
- MA-EgoQA: Question Answering over Egocentric Videos from Multiple Embodied Agents (viability: 8): https://sciencetostartup.com/paper/ma-egoqa-question-answering-over-egocentric-videos-from-multiple-embodied-agents - MA-EgoQA enables effective question answering over multiple egocentric videos from embodied agents, enhancing human-agent collaboration.
- VLM-Loc: Localization in Point Cloud Maps via Vision-Language Models (viability: 8): https://sciencetostartup.com/paper/vlm-loc-localization-in-point-cloud-maps-via-vision-language-models - VLM-Loc uses vision-language models to enhance precise localization in 3D point cloud maps using natural language descriptors.
- BrainSTR: Spatio-Temporal Contrastive Learning for Interpretable Dynamic Brain Network Modeling (viability: 7): https://sciencetostartup.com/paper/brainstr-spatio-temporal-contrastive-learning-for-interpretable-dynamic-brain-network-modeling - BrainSTR offers a novel framework for interpretable dynamic brain network modeling to enhance neuropsychiatric diagnosis.
- One-Eval: An Agentic System for Automated and Traceable LLM Evaluation (viability: 8): https://sciencetostartup.com/paper/one-eval-an-agentic-system-for-automated-and-traceable-llm-evaluation - One-Eval automates and streamlines the evaluation of large language models through customizable workflows based on natural language requests.
- ConfCtrl: Enabling Precise Camera Control in Video Diffusion via Confidence-Aware Interpolation (viability: 7): https://sciencetostartup.com/paper/confctrl-enabling-precise-camera-control-in-video-diffusion-via-confidence-aware-interpolation - ConfCtrl is a confidence-aware video interpolation framework that enhances diffusion models for precise camera control and occlusion handling.
- Correction of Transformer-Based Models with Smoothing Pseudo-Projector (viability: 5): https://sciencetostartup.com/paper/correction-of-transformer-based-models-with-smoothing-pseudo-projector - A lightweight modification for transformer-based models that enhances robustness and training dynamics.
- RA-SSU: Towards Fine-Grained Audio-Visual Learning with Region-Aware Sound Source Understanding (viability: 7): https://sciencetostartup.com/paper/ra-ssu-towards-fine-grained-audio-visual-learning-with-region-aware-sound-source-understanding - RA-SSU introduces a fine-grained approach to audio-visual learning for enhanced scene perception and interaction.
- Good Reasoning Makes Good Demonstrations: Implicit Reasoning Quality Supervision via In-Context Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/good-reasoning-makes-good-demonstrations-implicit-reasoning-quality-supervision-via-in-context-reinforcement-learning - A novel reinforcement learning approach that enhances reasoning quality in language models by prioritizing high-quality demonstrations.
- MITRA: An AI Assistant for Knowledge Retrieval in Physics Collaborations (viability: 7): https://sciencetostartup.com/paper/mitra-an-ai-assistant-for-knowledge-retrieval-in-physics-collaborations - MITRA is an AI assistant that enhances knowledge retrieval for scientific collaborations in physics.
- Test-time Ego-Exo-centric Adaptation for Action Anticipation via Multi-Label Prototype Growing and Dual-Clue Consistency (viability: 8): https://sciencetostartup.com/paper/test-time-ego-exo-centric-adaptation-for-action-anticipation-via-multi-label-prototype-growing-and-dual-clue-consistency - A novel method for real-time adaptation between Egocentric and Exocentric views to enhance action anticipation in human-robot cooperation.
- Information Theoretic Bayesian Optimization over the Probability Simplex (viability: 5): https://sciencetostartup.com/paper/information-theoretic-bayesian-optimization-over-the-probability-simplex - Introducing a novel Bayesian optimization algorithm tailored for optimizing probabilities in constrained domains.
- Exploiting Label-Aware Channel Scoring for Adaptive Channel Pruning in Split Learning (viability: 6): https://sciencetostartup.com/paper/exploiting-label-aware-channel-scoring-for-adaptive-channel-pruning-in-split-learning - A novel adaptive channel pruning scheme for split learning that reduces communication overhead while maintaining accuracy.
- A Hybrid Quantum-Classical Framework for Financial Volatility Forecasting Based on Quantum Circuit Born Machines (viability: 6): https://sciencetostartup.com/paper/a-hybrid-quantum-classical-framework-for-financial-volatility-forecasting-based-on-quantum-circuit-born-machines - A hybrid quantum-classical framework for forecasting financial market volatility using LSTM and Quantum Circuit Born Machines.
- What is Missing? Explaining Neurons Activated by Absent Concepts (viability: 2): https://sciencetostartup.com/paper/what-is-missing-explaining-neurons-activated-by-absent-concepts - This paper explores the overlooked concept of encoded absences in neural networks to enhance explainable AI methods.
- Quantifying the Necessity of Chain of Thought through Opaque Serial Depth (viability: 5): https://sciencetostartup.com/paper/quantifying-the-necessity-of-chain-of-thought-through-opaque-serial-depth - A method to quantify reasoning depth in LLMs, enhancing interpretability and monitoring.
- EPIC-EuroParl-UdS: Information-Theoretic Perspectives on Translation and Interpreting (viability: 2): https://sciencetostartup.com/paper/epic-europarl-uds-information-theoretic-perspectives-on-translation-and-interpreting - A refined corpus for studying translation and interpreting through information-theoretic approaches.
- Lightweight 3D LiDAR-Based UAV Tracking: An Adaptive Extended Kalman Filtering Approach (viability: 7): https://sciencetostartup.com/paper/lightweight-3d-lidar-based-uav-tracking-an-adaptive-extended-kalman-filtering-approach - A lightweight LiDAR-based UAV tracking system that enhances accuracy and reliability for swarm aerial robotics.
- TIMID: Time-Dependent Mistake Detection in Videos of Robot Executions (viability: 7): https://sciencetostartup.com/paper/timid-time-dependent-mistake-detection-in-videos-of-robot-executions - TIMID is a novel architecture for detecting time-dependent mistakes in robot execution videos using weak supervision.
- CLIOPATRA: Extracting Private Information from LLM Insights (viability: 4): https://sciencetostartup.com/paper/cliopatra-extracting-private-information-from-llm-insights - CLIOPATRA exposes vulnerabilities in privacy-preserving LLM systems, revealing the inadequacy of current protections.
- World2Mind: Cognition Toolkit for Allocentric Spatial Reasoning in Foundation Models (viability: 7): https://sciencetostartup.com/paper/world2mind-cognition-toolkit-for-allocentric-spatial-reasoning-in-foundation-models - World2Mind is a toolkit that enhances spatial reasoning in foundation models using structured cognitive maps.
- Removing the Trigger, Not the Backdoor: Alternative Triggers and Latent Backdoors (viability: 2): https://sciencetostartup.com/paper/removing-the-trigger-not-the-backdoor-alternative-triggers-and-latent-backdoors - This research reveals the inadequacy of current backdoor defenses by demonstrating the existence of alternative triggers that can activate latent backdoors.
- Ego: Embedding-Guided Personalization of Vision-Language Models (viability: 6): https://sciencetostartup.com/paper/ego-embedding-guided-personalization-of-vision-language-models - A novel method for personalizing vision-language models using internal attention mechanisms for efficient deployment.
- MuxGel: Simultaneous Dual-Modal Visuo-Tactile Sensing via Spatially Multiplexing and Deep Reconstruction (viability: 7): https://sciencetostartup.com/paper/muxgel-simultaneous-dual-modal-visuo-tactile-sensing-via-spatially-multiplexing-and-deep-reconstruction - MuxGel enhances robotic manipulation by integrating visuo-tactile sensing through a novel spatially multiplexed sensor design.
- PanoAffordanceNet: Towards Holistic Affordance Grounding in 360° Indoor Environments (viability: 8): https://sciencetostartup.com/paper/panoaffordancenet-towards-holistic-affordance-grounding-in-360-indoor-environments - PanoAffordanceNet enables holistic affordance grounding in 360° indoor environments, enhancing scene-level perception for embodied agents.
- LogoDiffuser: Training-Free Multilingual Logo Generation and Stylization via Letter-Aware Attention Control (viability: 7): https://sciencetostartup.com/paper/logodiffuser-training-free-multilingual-logo-generation-and-stylization-via-letter-aware-attention-control - LogoDiffuser enables training-free multilingual logo generation with robust character structure control.
- Beyond Fine-Tuning: Robust Food Entity Linking under Ontology Drift with FoodOntoRAG (viability: 3): https://sciencetostartup.com/paper/beyond-fine-tuning-robust-food-entity-linking-under-ontology-drift-with-foodontorag - FoodOntoRAG offers a robust, model-agnostic pipeline for food entity linking that adapts to ontology changes without fine-tuning.
- Caterpillar-Inspired Spring-Based Compressive Continuum Robot for Bristle-based Exploration (viability: 7): https://sciencetostartup.com/paper/caterpillar-inspired-spring-based-compressive-continuum-robot-for-bristle-based-exploration - A compact, spring-based continuum robot designed for effective exploration of confined spaces using bristle-based sensing.
- LAP: A Language-Aware Planning Model For Procedure Planning In Instructional Videos (viability: 8): https://sciencetostartup.com/paper/lap-a-language-aware-planning-model-for-procedure-planning-in-instructional-videos - LAP is a language-aware planning model that enhances procedure planning in instructional videos by leveraging language descriptions for action prediction.
- Upper Generalization Bounds for Neural Oscillators (viability: 3): https://sciencetostartup.com/paper/upper-generalization-bounds-for-neural-oscillators - This paper explores the theoretical bounds of neural oscillators for dynamic load-response mappings in structural systems.
- ENIGMA-360: An Ego-Exo Dataset for Human Behavior Understanding in Industrial Scenarios (viability: 7): https://sciencetostartup.com/paper/enigma-360-an-ego-exo-dataset-for-human-behavior-understanding-in-industrial-scenarios - ENIGMA-360 is a novel dataset for understanding human behavior in industrial settings through synchronized egocentric and exocentric video data.
- Let's Reward Step-by-Step: Step-Aware Contrastive Alignment for Vision-Language Navigation in Continuous Environments (viability: 7): https://sciencetostartup.com/paper/let-s-reward-step-by-step-step-aware-contrastive-alignment-for-vision-language-navigation-in-continuous-environments - Step-Aware Contrastive Alignment enhances Vision-Language Navigation by improving error recovery and training stability through dense supervision.
- $M^2$-Occ: Resilient 3D Semantic Occupancy Prediction for Autonomous Driving with Incomplete Camera Inputs (viability: 8): https://sciencetostartup.com/paper/m-2-occ-resilient-3d-semantic-occupancy-prediction-for-autonomous-driving-with-incomplete-camera-inputs - M^2-Occ enhances 3D semantic occupancy prediction for autonomous driving by effectively handling incomplete camera inputs.
- FetalAgents: A Multi-Agent System for Fetal Ultrasound Image and Video Analysis (viability: 7): https://sciencetostartup.com/paper/fetalagents-a-multi-agent-system-for-fetal-ultrasound-image-and-video-analysis - FetalAgents is a multi-agent system that enhances fetal ultrasound analysis through dynamic coordination of specialized vision experts.
- EXPLORE-Bench: Egocentric Scene Prediction with Long-Horizon Reasoning (viability: 4): https://sciencetostartup.com/paper/explore-bench-egocentric-scene-prediction-with-long-horizon-reasoning - EXPLORE-Bench is a benchmark for evaluating long-horizon reasoning in egocentric scene prediction using multimodal large language models.
- A Multi-Prototype-Guided Federated Knowledge Distillation Approach in AI-RAN Enabled Multi-Access Edge Computing System (viability: 5): https://sciencetostartup.com/paper/a-multi-prototype-guided-federated-knowledge-distillation-approach-in-ai-ran-enabled-multi-access-edge-computing-system - A novel federated learning approach that enhances model accuracy in AI-RAN enabled MEC systems by addressing data heterogeneity.
- RbtAct: Rebuttal as Supervision for Actionable Review Feedback Generation (viability: 8): https://sciencetostartup.com/paper/rbtact-rebuttal-as-supervision-for-actionable-review-feedback-generation - RbtAct enhances peer review processes by generating actionable feedback using rebuttal as supervision.
- FrameDiT: Diffusion Transformer with Frame-Level Matrix Attention for Efficient Video Generation (viability: 7): https://sciencetostartup.com/paper/framedit-diffusion-transformer-with-frame-level-matrix-attention-for-efficient-video-generation - FrameDiT leverages Matrix Attention for efficient high-fidelity video generation, outperforming existing methods.
- GSStream: 3D Gaussian Splatting based Volumetric Scene Streaming System (viability: 7): https://sciencetostartup.com/paper/gsstream-3d-gaussian-splatting-based-volumetric-scene-streaming-system - GSStream is a novel volumetric scene streaming system that enhances real-time data delivery using collaborative viewport prediction and bitrate adaptation.
- AutoAgent: Evolving Cognition and Elastic Memory Orchestration for Adaptive Agents (viability: 7): https://sciencetostartup.com/paper/autoagent-evolving-cognition-and-elastic-memory-orchestration-for-adaptive-agents - AutoAgent is a self-evolving multi-agent framework designed for adaptive decision-making in dynamic environments.
- Does the Question Really Matter? Training-Free Data Selection for Vision-Language SFT (viability: 8): https://sciencetostartup.com/paper/does-the-question-really-matter-training-free-data-selection-for-vision-language-sft - CVS is a training-free data selection method that enhances vision-language model performance by identifying samples requiring genuine cross-modal reasoning.
- MUGEN: Evaluating and Improving Multi-audio Understanding of Large Audio-Language Models (viability: 4): https://sciencetostartup.com/paper/mugen-evaluating-and-improving-multi-audio-understanding-of-large-audio-language-models - MUGEN benchmarks and improves multi-audio understanding in large audio-language models.
- Robotic Scene Cloning:Advancing Zero-Shot Robotic Scene Adaptation in Manipulation via Visual Prompt Editing (viability: 4): https://sciencetostartup.com/paper/robotic-scene-cloning-advancing-zero-shot-robotic-scene-adaptation-in-manipulation-via-visual-prompt-editing - Robotic Scene Cloning enables robots to adapt to new environments through visual prompt editing for improved task performance.
- OOD-MMSafe: Advancing MLLM Safety from Harmful Intent to Hidden Consequences (viability: 7): https://sciencetostartup.com/paper/ood-mmsafe-advancing-mllm-safety-from-harmful-intent-to-hidden-consequences - Introducing OOD-MMSafe, a benchmark and framework for enhancing safety in Multimodal Large Language Models by focusing on consequence-driven safety.
- Evaluation of LLMs in retrieving food and nutritional context for RAG systems (viability: 5): https://sciencetostartup.com/paper/evaluation-of-llms-in-retrieving-food-and-nutritional-context-for-rag-systems - An LLM-based tool for efficient retrieval of food and nutritional data using structured metadata filters.
- ProGS: Towards Progressive Coding for 3D Gaussian Splatting (viability: 7): https://sciencetostartup.com/paper/progs-towards-progressive-coding-for-3d-gaussian-splatting - ProGS is a streaming-friendly codec that enables efficient progressive coding for 3D Gaussian splatting, significantly improving compression efficiency and visual fidelity.
- TriFusion-SR: Joint Tri-Modal Medical Image Fusion and SR (viability: 7): https://sciencetostartup.com/paper/trifusion-sr-joint-tri-modal-medical-image-fusion-and-sr - TriFusionSR enhances multimodal medical imaging by integrating image fusion and super-resolution into a single framework.
- Mousse: Rectifying the Geometry of Muon with Curvature-Aware Preconditioning (viability: 5): https://sciencetostartup.com/paper/mousse-rectifying-the-geometry-of-muon-with-curvature-aware-preconditioning - Mousse is an advanced optimizer that enhances training efficiency for deep neural networks by adapting to the curvature of the optimization landscape.
- TemporalDoRA: Temporal PEFT for Robust Surgical Video Question Answering (viability: 8): https://sciencetostartup.com/paper/temporaldora-temporal-peft-for-robust-surgical-video-question-answering - Leveraging TemporalDoRA for robust surgical video QA in clinical settings, overcoming traditional model limitations.
- DRIFT: Dual-Representation Inter-Fusion Transformer for Automated Driving Perception with 4D Radar Point Clouds (viability: 7): https://sciencetostartup.com/paper/drift-dual-representation-inter-fusion-transformer-for-automated-driving-perception-with-4d-radar-point-clouds - DRIFT is a dual-path transformer model that enhances automated driving perception using 4D radar point clouds.
- Physics-informed neural operator for predictive parametric phase-field modelling (viability: 6): https://sciencetostartup.com/paper/physics-informed-neural-operator-for-predictive-parametric-phase-field-modelling - A physics-informed neural operator framework that accelerates parametric phase-field modeling for materials science.
- ActiveUltraFeedback: Efficient Preference Data Generation using Active Learning (viability: 8): https://sciencetostartup.com/paper/activeultrafeedback-efficient-preference-data-generation-using-active-learning - ActiveUltraFeedback optimizes preference data generation for training language models using active learning techniques.
- ESAinsTOD: A Unified End-to-End Schema-Aware Instruction-Tuning Framework for Task-Oriented Dialog Modeling (viability: 8): https://sciencetostartup.com/paper/esainstod-a-unified-end-to-end-schema-aware-instruction-tuning-framework-for-task-oriented-dialog-modeling - ESAinsTOD is a unified framework that enhances task-oriented dialog systems through schema-aware instruction tuning.
- AutoViVQA: A Large-Scale Automatically Constructed Dataset for Vietnamese Visual Question Answering (viability: 4): https://sciencetostartup.com/paper/autovivqa-a-large-scale-automatically-constructed-dataset-for-vietnamese-visual-question-answering - AutoViVQA provides a large-scale dataset to enhance Vietnamese visual question answering systems.
- Fusing Semantic, Lexical, and Domain Perspectives for Recipe Similarity Estimation (viability: 3): https://sciencetostartup.com/paper/fusing-semantic-lexical-and-domain-perspectives-for-recipe-similarity-estimation - A method for assessing recipe similarity using semantic, lexical, and domain perspectives.
- Automatic Cardiac Risk Management Classification using large-context Electronic Patients Health Records (viability: 6): https://sciencetostartup.com/paper/automatic-cardiac-risk-management-classification-using-large-context-electronic-patients-health-records - An automated classification framework for cardiac risk management using unstructured Electronic Health Records.
- On Catastrophic Forgetting in Low-Rank Decomposition-Based Parameter-Efficient Fine-Tuning (viability: 2): https://sciencetostartup.com/paper/on-catastrophic-forgetting-in-low-rank-decomposition-based-parameter-efficient-fine-tuning - This paper explores the impact of update subspace design on catastrophic forgetting in parameter-efficient fine-tuning methods.
- Improving 3D Foot Motion Reconstruction in Markerless Monocular Human Motion Capture (viability: 7): https://sciencetostartup.com/paper/improving-3d-foot-motion-reconstruction-in-markerless-monocular-human-motion-capture - FootMR enhances 3D foot motion reconstruction from videos, improving gait analysis and animation accuracy.
- EsoLang-Bench: Evaluating Genuine Reasoning in Large Language Models via Esoteric Programming Languages (viability: 4): https://sciencetostartup.com/paper/esolang-bench-evaluating-genuine-reasoning-in-large-language-models-via-esoteric-programming-languages - EsoLang-Bench evaluates reasoning in large language models using esoteric programming languages to reveal genuine learning capabilities.
- Logics-Parsing-Omni Technical Report (viability: 9): https://sciencetostartup.com/paper/logics-parsing-omni-technical-report - AI-driven framework for parsing unstructured multimedia into structured, machine-readable knowledge.
- GNNs for Time Series Anomaly Detection: An Open-Source Framework and a Critical Evaluation (viability: 5): https://sciencetostartup.com/paper/gnns-for-time-series-anomaly-detection-an-open-source-framework-and-a-critical-evaluation - An open-source framework for applying Graph Neural Networks to enhance Time Series Anomaly Detection with improved interpretability.
- VarSplat: Uncertainty-aware 3D Gaussian Splatting for Robust RGB-D SLAM (viability: 3): https://sciencetostartup.com/paper/varsplat-uncertainty-aware-3d-gaussian-splatting-for-robust-rgb-d-slam - VarSplat enhances RGB-D SLAM with uncertainty-aware 3D Gaussian Splatting for improved robustness.
- DiffWind: Physics-Informed Differentiable Modeling of Wind-Driven Object Dynamics (viability: 6): https://sciencetostartup.com/paper/diffwind-physics-informed-differentiable-modeling-of-wind-driven-object-dynamics - DiffWind is a physics-informed framework for modeling wind-driven object dynamics from video observations.
- No evaluation without fair representation : Impact of label and selection bias on the evaluation, performance and mitigation of classification models (viability: 3): https://sciencetostartup.com/paper/no-evaluation-without-fair-representation-impact-of-label-and-selection-bias-on-the-evaluation-performance-and-mitigatio - A framework for analyzing and mitigating bias in machine learning classification models to enhance fairness and accuracy.
- FreqCycle: A Multi-Scale Time-Frequency Analysis Method for Time Series Forecasting (viability: 5): https://sciencetostartup.com/paper/freqcycle-a-multi-scale-time-frequency-analysis-method-for-time-series-forecasting - FreqCycle is a novel framework for enhancing time series forecasting by integrating multi-scale time-frequency analysis.
- When to Lock Attention: Training-Free KV Control in Video Diffusion (viability: 3): https://sciencetostartup.com/paper/when-to-lock-attention-training-free-kv-control-in-video-diffusion - KV-Lock is a training-free module for enhancing foreground quality in video diffusion models while maintaining background consistency.
- Understanding the Interplay between LLMs' Utilisation of Parametric and Contextual Knowledge: A keynote at ECIR 2025 (viability: 2): https://sciencetostartup.com/paper/understanding-the-interplay-between-llms-utilisation-of-parametric-and-contextual-knowledge-a-keynote-at-ecir-2025 - Exploring the challenges of integrating parametric and contextual knowledge in language models.
- OTPL-VIO: Robust Visual-Inertial Odometry with Optimal Transport Line Association and Adaptive Uncertainty (viability: 7): https://sciencetostartup.com/paper/otpl-vio-robust-visual-inertial-odometry-with-optimal-transport-line-association-and-adaptive-uncertainty - A robust stereo visual-inertial odometry system that enhances performance in low-texture and challenging illumination environments using optimal transport line association.
- MiniAppBench: Evaluating the Shift from Text to Interactive HTML Responses in LLM-Powered Assistants (viability: 8): https://sciencetostartup.com/paper/miniappbench-evaluating-the-shift-from-text-to-interactive-html-responses-in-llm-powered-assistants - MiniAppBench is a benchmark for evaluating LLM-generated interactive HTML applications, enhancing human-AI interaction.
- Well Log-Guided Synthesis of Subsurface Images from Sparse Petrography Data Using cGANs (viability: 3): https://sciencetostartup.com/paper/well-log-guided-synthesis-of-subsurface-images-from-sparse-petrography-data-using-cgans - A cGAN framework for synthesizing pore-scale images of carbonate rock formations from well log data.
- MM-tau-p$^2$: Persona-Adaptive Prompting for Robust Multi-Modal Agent Evaluation in Dual-Control Settings (viability: 5): https://sciencetostartup.com/paper/mm-tau-p-2-persona-adaptive-prompting-for-robust-multi-modal-agent-evaluation-in-dual-control-settings - A benchmark for evaluating multi-modal agents with persona adaptation in customer experience management.
- Multi-DNN Inference of Sparse Models on Edge SoCs (viability: 8): https://sciencetostartup.com/paper/multi-dnn-inference-of-sparse-models-on-edge-socs - SparseLoom enhances multi-DNN inference on edge devices by optimizing model deployment without retraining.
- PRECEPT: Planning Resilience via Experience, Context Engineering & Probing Trajectories A Unified Framework for Test-Time Adaptation with Compositional Rule Learning and Pareto-Guided Prompt Evolution (viability: 7): https://sciencetostartup.com/paper/precept-planning-resilience-via-experience-context-engineering-probing-trajectories-a-unified-framework-for-test-time-ad - PRECEPT is a unified framework for enhancing LLM agents' test-time adaptation through advanced rule retrieval and memory conflict resolution.
- Tracking Cancer Through Text: Longitudinal Extraction From Radiology Reports Using Open-Source Large Language Models (viability: 8): https://sciencetostartup.com/paper/tracking-cancer-through-text-longitudinal-extraction-from-radiology-reports-using-open-source-large-language-models - An open-source pipeline for extracting longitudinal cancer data from radiology reports using large language models.
- X-GS: An Extensible Open Framework Unifying 3DGS Architectures with Downstream Multimodal Models (viability: 7): https://sciencetostartup.com/paper/x-gs-an-extensible-open-framework-unifying-3dgs-architectures-with-downstream-multimodal-models - X-GS is an extensible framework that unifies 3D Gaussian Splatting techniques for real-time spatial AI applications.
- Grounding Synthetic Data Generation With Vision and Language Models (viability: 8): https://sciencetostartup.com/paper/grounding-synthetic-data-generation-with-vision-and-language-models - A vision-language framework for interpretable synthetic data generation and evaluation in remote sensing.
- Decoder-Free Distillation for Quantized Image Restoration (viability: 8): https://sciencetostartup.com/paper/decoder-free-distillation-for-quantized-image-restoration - A framework for edge-deployed image restoration that enhances visual quality through quantization-aware training and decoder-free distillation.
- Physics-Driven 3D Gaussian Rendering for Zero-Shot MRI Super-Resolution (viability: 7): https://sciencetostartup.com/paper/physics-driven-3d-gaussian-rendering-for-zero-shot-mri-super-resolution - A zero-shot MRI super-resolution framework that balances data efficiency and reconstruction quality using explicit Gaussian representation.
- Context Engineering: From Prompts to Corporate Multi-Agent Architecture (viability: 2): https://sciencetostartup.com/paper/context-engineering-from-prompts-to-corporate-multi-agent-architecture - This paper proposes context engineering as a new discipline for managing AI agent environments, but lacks practical implementation details.
- Surgical Repair of Collapsed Attention Heads in ALiBi Transformers (viability: 8): https://sciencetostartup.com/paper/surgical-repair-of-collapsed-attention-heads-in-alibi-transformers - A novel surgical reinitialization technique to recover attention head capacity in ALiBi Transformers, enhancing model performance.
- A saccade-inspired approach to image classification using visiontransformer attention maps (viability: 6): https://sciencetostartup.com/paper/a-saccade-inspired-approach-to-image-classification-using-visiontransformer-attention-maps - A novel image classification method inspired by human saccadic attention using Vision Transformer attention maps.
- ParTY: Part-Guidance for Expressive Text-to-Motion Synthesis (viability: 7): https://sciencetostartup.com/paper/party-part-guidance-for-expressive-text-to-motion-synthesis - ParTY is a novel framework for generating coherent and expressive human motions from text descriptions by enhancing part expressiveness.
- Learning the Hierarchical Organization in Brain Network for Brain Disorder Diagnosis (viability: 7): https://sciencetostartup.com/paper/learning-the-hierarchical-organization-in-brain-network-for-brain-disorder-diagnosis - BrainHO leverages hierarchical attention mechanisms to enhance brain disorder diagnosis through advanced fMRI analysis.
- MM-algorithms for traditional and convex NMF with Tweedie and Negative Binomial cost functions and empirical evaluation (viability: 7): https://sciencetostartup.com/paper/mm-algorithms-for-traditional-and-convex-nmf-with-tweedie-and-negative-binomial-cost-functions-and-empirical-evaluation - A unified framework for Non-negative Matrix Factorization with advanced cost functions for improved feature extraction.
- A Generalized Voronoi Graph based Coverage Control Approach for Non-Convex Environment (viability: 2): https://sciencetostartup.com/paper/a-generalized-voronoi-graph-based-coverage-control-approach-for-non-convex-environment - A novel coverage control method for multi-robot systems in non-convex environments using Generalized Voronoi Graphs.
- Build, Borrow, or Just Fine-Tune? A Political Scientist's Guide to Choosing NLP Models (viability: 7): https://sciencetostartup.com/paper/build-borrow-or-just-fine-tune-a-political-scientist-s-guide-to-choosing-nlp-models - A decision framework for political scientists to choose between building, borrowing, or fine-tuning NLP models based on empirical performance analysis.
- Benchmarking Dataset for Presence-Only Passive Reconnaissance in Wireless Smart-Grid Communications (viability: 4): https://sciencetostartup.com/paper/benchmarking-dataset-for-presence-only-passive-reconnaissance-in-wireless-smart-grid-communications - A benchmark dataset generator for passive reconnaissance in smart-grid communications to enhance cybersecurity evaluations.
- Memorization capacity of deep ReLU neural networks characterized by width and depth (viability: 2): https://sciencetostartup.com/paper/memorization-capacity-of-deep-relu-neural-networks-characterized-by-width-and-depth - This paper explores the memorization capacity of deep ReLU networks through width and depth trade-offs.
- Game-Theoretic Modeling of Stealthy Intrusion Defense against MDP-Based Attackers (viability: 2): https://sciencetostartup.com/paper/game-theoretic-modeling-of-stealthy-intrusion-defense-against-mdp-based-attackers - A theoretical model for defending against advanced persistent threats using game theory.
- Towards Terrain-Aware Safe Locomotion for Quadrupedal Robots Using Proprioceptive Sensing (viability: 4): https://sciencetostartup.com/paper/towards-terrain-aware-safe-locomotion-for-quadrupedal-robots-using-proprioceptive-sensing - A terrain-aware locomotion system for quadrupedal robots that enhances safety using proprioceptive sensing.
- Nonparametric Variational Differential Privacy via Embedding Parameter Clipping (viability: 7): https://sciencetostartup.com/paper/nonparametric-variational-differential-privacy-via-embedding-parameter-clipping - A novel approach to enhance privacy in language models through parameter clipping, improving the privacy-utility trade-off.
- BinaryAttention: One-Bit QK-Attention for Vision and Diffusion Transformers (viability: 8): https://sciencetostartup.com/paper/binaryattention-one-bit-qk-attention-for-vision-and-diffusion-transformers - BinaryAttention offers a highly efficient binary quantization method for Transformers, doubling speed for vision tasks without sacrificing accuracy.
- Towards Understanding Adam Convergence on Highly Degenerate Polynomials (viability: 4): https://sciencetostartup.com/paper/towards-understanding-adam-convergence-on-highly-degenerate-polynomials - This paper explores the auto-convergence properties of the Adam optimizer on highly degenerate polynomials, providing theoretical insights and experimental validation.
- Randomized Distributed Function Computation (RDFC): Ultra-Efficient Semantic Communication Applications to Privacy (viability: 2): https://sciencetostartup.com/paper/randomized-distributed-function-computation-rdfc-ultra-efficient-semantic-communication-applications-to-privacy - RDFC offers a framework for energy-efficient semantic communication with privacy guarantees.
- Routing without Forgetting (viability: 7): https://sciencetostartup.com/paper/routing-without-forgetting - RwF is a novel transformer architecture that enhances continual learning by dynamically routing representations without task identifiers.
- SCDP: Learning Humanoid Locomotion from Partial Observations via Mixed-Observation Distillation (viability: 8): https://sciencetostartup.com/paper/scdp-learning-humanoid-locomotion-from-partial-observations-via-mixed-observation-distillation - SCDP enables humanoid locomotion using only onboard sensors, eliminating the need for complex state estimation.
- More than the Sum: Panorama-Language Models for Adverse Omni-Scenes (viability: 7): https://sciencetostartup.com/paper/more-than-the-sum-panorama-language-models-for-adverse-omni-scenes - Introducing a unified 360-degree vision-language model for enhanced understanding of complex omni-scenes.
- An Optimal Control Approach To Transformer Training (viability: 2): https://sciencetostartup.com/paper/an-optimal-control-approach-to-transformer-training - This paper presents a theoretical framework for optimizing Transformer training using control theory.
- GeoAlignCLIP: Enhancing Fine-Grained Vision-Language Alignment in Remote Sensing via Multi-Granular Consistency Learning (viability: 7): https://sciencetostartup.com/paper/geoalignclip-enhancing-fine-grained-vision-language-alignment-in-remote-sensing-via-multi-granular-consistency-learning - GeoAlignCLIP enhances fine-grained vision-language alignment in remote sensing through multi-granular consistency learning.
- ReTac-ACT: A State-Gated Vision-Tactile Fusion Transformer for Precision Assembly (viability: 8): https://sciencetostartup.com/paper/retac-act-a-state-gated-vision-tactile-fusion-transformer-for-precision-assembly - ReTac-ACT enhances precision in robotic assembly by seamlessly integrating vision and tactile feedback.
- a-TMFG: Scalable Triangulated Maximally Filtered Graphs via Approximate Nearest Neighbors (viability: 4): https://sciencetostartup.com/paper/a-tmfg-scalable-triangulated-maximally-filtered-graphs-via-approximate-nearest-neighbors - a-TMFG offers a scalable solution for constructing filtered graphs from large datasets using an innovative memory management strategy.
- Learning Bayesian and Markov Networks with an Unreliable Oracle (viability: 2): https://sciencetostartup.com/paper/learning-bayesian-and-markov-networks-with-an-unreliable-oracle - This paper explores structure learning in Bayesian and Markov networks with unreliable oracles.
- Trajectory Optimization for Self-Wrap-Aware Cable-Towed Planar Object Manipulation under Implicit Tension Constraints (viability: 2): https://sciencetostartup.com/paper/trajectory-optimization-for-self-wrap-aware-cable-towed-planar-object-manipulation-under-implicit-tension-constraints - This paper presents a theoretical framework for optimizing cable-towed manipulation of deformable objects under tension constraints.
- ALARM: Audio-Language Alignment for Reasoning Models (viability: 7): https://sciencetostartup.com/paper/alarm-audio-language-alignment-for-reasoning-models - A novel audio-language model that enhances reasoning capabilities through self-rephrasing and multi-task training.
- Compiler-First State Space Duality and Portable $O(1)$ Autoregressive Caching for Inference (viability: 7): https://sciencetostartup.com/paper/compiler-first-state-space-duality-and-portable-o-1-autoregressive-caching-for-inference - An optimized JAX-based inference caching solution for device-agnostic autoregressive decoding.
- On the Cost of Evolving Task Specialization in Multi-Robot Systems (viability: 2): https://sciencetostartup.com/paper/on-the-cost-of-evolving-task-specialization-in-multi-robot-systems - This study analyzes the cost-benefit of task specialization in multi-robot systems.
- GeoSolver: Scaling Test-Time Reasoning in Remote Sensing with Fine-Grained Process Supervision (viability: 8): https://sciencetostartup.com/paper/geosolver-scaling-test-time-reasoning-in-remote-sensing-with-fine-grained-process-supervision - GeoSolver enhances remote sensing interpretation through verifiable, process-supervised reasoning.
- Enabling Multi-Client Authorization in Dynamic SSE (viability: 6): https://sciencetostartup.com/paper/enabling-multi-client-authorization-in-dynamic-sse - MASSE is a dynamic multi-client searchable symmetric encryption scheme that enhances data privacy and access control in cloud environments.
- A comprehensive study of time-of-flight non-line-of-sight imaging (viability: 2): https://sciencetostartup.com/paper/a-comprehensive-study-of-time-of-flight-non-line-of-sight-imaging - A comprehensive study of time-of-flight non-line-of-sight imaging techniques and their performance limitations.
- Compartmentalization-Aware Automated Program Repair (viability: 3): https://sciencetostartup.com/paper/compartmentalization-aware-automated-program-repair - A framework for securing cross-compartment interfaces in software using automated program repair techniques.
- Memory-Guided View Refinement for Dynamic Human-in-the-loop EQA (viability: 7): https://sciencetostartup.com/paper/memory-guided-view-refinement-for-dynamic-human-in-the-loop-eqa - DynHiL-EQA enhances embodied question answering in dynamic environments by refining views and optimizing memory usage.
- NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (viability: 7): https://sciencetostartup.com/paper/ns-vla-towards-neuro-symbolic-vision-language-action-models - NS-VLA is a Neuro-Symbolic framework that enhances robotic manipulation through efficient learning and action sequencing.
- Towards Unified Multimodal Interleaved Generation via Group Relative Policy Optimization (viability: 6): https://sciencetostartup.com/paper/towards-unified-multimodal-interleaved-generation-via-group-relative-policy-optimization - A reinforcement learning strategy to enhance multimodal interleaved generation in existing unified vision-language models.
- Enhancing Debunking Effectiveness through LLM-based Personality Adaptation (viability: 7): https://sciencetostartup.com/paper/enhancing-debunking-effectiveness-through-llm-based-personality-adaptation - A methodology for generating personalized fake news debunking messages using LLMs tailored to personality traits.
- What Do We Care About in Bandits with Noncompliance? BRACE: Bandits with Recommendations, Abstention, and Certified Effects (viability: 4): https://sciencetostartup.com/paper/what-do-we-care-about-in-bandits-with-noncompliance-brace-bandits-with-recommendations-abstention-and-certified-effects - BRACE is a novel algorithm for optimizing recommendations in bandit settings with noncompliance, enhancing treatment learning and uncertainty quantification.
- DCAU-Net: Differential Cross Attention and Channel-Spatial Feature Fusion for Medical Image Segmentation (viability: 7): https://sciencetostartup.com/paper/dcau-net-differential-cross-attention-and-channel-spatial-feature-fusion-for-medical-image-segmentation - DCAU-Net enhances medical image segmentation by efficiently modeling long-range dependencies and fine-grained details through innovative attention mechanisms.
- RESBev: Making BEV Perception More Robust (viability: 7): https://sciencetostartup.com/paper/resbev-making-bev-perception-more-robust - RESBev enhances the robustness of bird's-eye-view perception in autonomous driving against disturbances and adversarial attacks.
- Efficiently Aligning Draft Models via Parameter- and Data-Efficient Adaptation (viability: 8): https://sciencetostartup.com/paper/efficiently-aligning-draft-models-via-parameter-and-data-efficient-adaptation - Efficient Draft Adaptation (EDA) optimizes LLM fine-tuning by reducing training costs while enhancing performance through innovative data strategies.
- You Didn't Have to Say It like That: Subliminal Learning from Faithful Paraphrases (viability: 5): https://sciencetostartup.com/paper/you-didn-t-have-to-say-it-like-that-subliminal-learning-from-faithful-paraphrases - A study on subliminal learning in language models reveals unexpected behavioral trait transmission through paraphrases.
- Beyond Short-Horizon: VQ-Memory for Robust Long-Horizon Manipulation in Non-Markovian Simulation Benchmarks (viability: 7): https://sciencetostartup.com/paper/beyond-short-horizon-vq-memory-for-robust-long-horizon-manipulation-in-non-markovian-simulation-benchmarks - RuleSafe introduces a novel benchmark for articulated manipulation tasks, enhancing long-horizon planning through VQ-Memory.
- Probing the Reliability of Driving VLMs: From Inconsistent Responses to Grounded Temporal Reasoning (viability: 6): https://sciencetostartup.com/paper/probing-the-reliability-of-driving-vlms-from-inconsistent-responses-to-grounded-temporal-reasoning - A benchmark dataset and tuning approach to enhance the reliability of Vision-Language Models in driving assistance through improved temporal reasoning.
- TrainDeeploy: Hardware-Accelerated Parameter-Efficient Fine-Tuning of Small Transformer Models at the Extreme Edge (viability: 6): https://sciencetostartup.com/paper/traindeeploy-hardware-accelerated-parameter-efficient-fine-tuning-of-small-transformer-models-at-the-extreme-edge - TrainDeeploy enables efficient on-device training of small transformer models for ultra-low-power edge devices.
- Context-Nav: Context-Driven Exploration and Viewpoint-Aware 3D Spatial Reasoning for Instance Navigation (viability: 3): https://sciencetostartup.com/paper/context-nav-context-driven-exploration-and-viewpoint-aware-3d-spatial-reasoning-for-instance-navigation - Context-Nav enhances instance navigation by integrating contextual captions and 3D spatial reasoning for improved object identification.
- Modelling the Diachronic Emergence of Phoneme Frequency Distributions (viability: 2): https://sciencetostartup.com/paper/modelling-the-diachronic-emergence-of-phoneme-frequency-distributions - This paper presents a stochastic model to explain phoneme frequency distributions through historical phonological change.
- SurgFed: Language-guided Multi-Task Federated Learning for Surgical Video Understanding (viability: 8): https://sciencetostartup.com/paper/surgfed-language-guided-multi-task-federated-learning-for-surgical-video-understanding - SurgFed is a language-guided federated learning framework that enhances surgical video understanding through multi-task learning.
- Evolving Prompt Adaptation for Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/evolving-prompt-adaptation-for-vision-language-models - EvoPrompt is a framework that enhances vision-language models' adaptation to few-shot tasks while preserving their pre-trained knowledge.
- Temporal-Conditioned Normalizing Flows for Multivariate Time Series Anomaly Detection (viability: 5): https://sciencetostartup.com/paper/temporal-conditioned-normalizing-flows-for-multivariate-time-series-anomaly-detection - A novel framework for anomaly detection in time series data using temporal-conditioned normalizing flows.
- Streaming Autoregressive Video Generation via Diagonal Distillation (viability: 3): https://sciencetostartup.com/paper/streaming-autoregressive-video-generation-via-diagonal-distillation - Diagonal Distillation enhances real-time video generation by optimizing temporal context in autoregressive models.
- Vibe-Creation: The Epistemology of Human-AI Emergent Cognition (viability: 2): https://sciencetostartup.com/paper/vibe-creation-the-epistemology-of-human-ai-emergent-cognition - Exploring the emergent cognitive structures formed by human and AI interactions.
- Component-Aware Sketch-to-Image Generation Using Self-Attention Encoding and Coordinate-Preserving Fusion (viability: 8): https://sciencetostartup.com/paper/component-aware-sketch-to-image-generation-using-self-attention-encoding-and-coordinate-preserving-fusion - A novel framework for transforming sketches into photorealistic images using self-attention and coordinate-preserving techniques.
- StyleVLA: Driving Style-Aware Vision Language Action Model for Autonomous Driving (viability: 8): https://sciencetostartup.com/paper/stylevla-driving-style-aware-vision-language-action-model-for-autonomous-driving - StyleVLA is a physics-informed Vision Language Action model that generates diverse and plausible driving behaviors tailored to individual driving styles.
- GenePlan: Evolving Better Generalized PDDL Plans using Large Language Models (viability: 7): https://sciencetostartup.com/paper/geneplan-evolving-better-generalized-pddl-plans-using-large-language-models - GenePlan is an LLM-assisted framework that evolves efficient planners for classical planning tasks using evolutionary algorithms.
- Prune Redundancy, Preserve Essence: Vision Token Compression in VLMs via Synergistic Importance-Diversity (viability: 8): https://sciencetostartup.com/paper/prune-redundancy-preserve-essence-vision-token-compression-in-vlms-via-synergistic-importance-diversity - PruneSID optimizes visual token compression in vision-language models, enhancing efficiency and performance.
- Telogenesis: Goal Is All U Need (viability: 4): https://sciencetostartup.com/paper/telogenesis-goal-is-all-u-need - A novel approach to goal-conditioned systems that generates adaptive priorities from an agent's internal cognitive state.
- Receptogenesis in a Vascularized Robotic Embodiment (viability: 7): https://sciencetostartup.com/paper/receptogenesis-in-a-vascularized-robotic-embodiment - A robotic system that can autonomously generate hardware through dynamic material restructuring for enhanced adaptability.
- OmniEarth: A Benchmark for Evaluating Vision-Language Models in Geospatial Tasks (viability: 7): https://sciencetostartup.com/paper/omniearth-a-benchmark-for-evaluating-vision-language-models-in-geospatial-tasks - OmniEarth is a benchmark for evaluating vision-language models in geospatial tasks, addressing critical gaps in remote sensing applications.
- The Patrologia Graeca Corpus: OCR, Annotation, and Open Release of Noisy Nineteenth-Century Polytonic Greek Editions (viability: 7): https://sciencetostartup.com/paper/the-patrologia-graeca-corpus-ocr-annotation-and-open-release-of-noisy-nineteenth-century-polytonic-greek-editions - A large-scale open OCR resource for nineteenth-century polytonic Greek editions, setting a new benchmark for text recognition.
- TopoOR: A Unified Topological Scene Representation for the Operating Room (viability: 4): https://sciencetostartup.com/paper/topoor-a-unified-topological-scene-representation-for-the-operating-room - TopoOR offers a novel topological representation for surgical operating rooms, enhancing multimodal interaction modeling.
- EvoDriveVLA: Evolving Autonomous Driving Vision-Language-Action Model via Collaborative Perception-Planning Distillation (viability: 8): https://sciencetostartup.com/paper/evodrivevla-evolving-autonomous-driving-vision-language-action-model-via-collaborative-perception-planning-distillation - EvoDriveVLA enhances autonomous driving with state-of-the-art Vision-Language-Action models through innovative perception-planning distillation.
- An Empirical Study and Theoretical Explanation on Task-Level Model-Merging Collapse (viability: 4): https://sciencetostartup.com/paper/an-empirical-study-and-theoretical-explanation-on-task-level-model-merging-collapse - This research identifies and explains the phenomenon of merging collapse in task-level model merging, providing insights for improving model integration.
- SEA-Nav: Efficient Policy Learning for Safe and Agile Quadruped Navigation in Cluttered Environments (viability: 7): https://sciencetostartup.com/paper/sea-nav-efficient-policy-learning-for-safe-and-agile-quadruped-navigation-in-cluttered-environments - SEA-Nav enables safe and efficient quadruped navigation in cluttered environments with minimal training time.
- Stein Variational Ergodic Surface Coverage with SE(3) Constraints (viability: 7): https://sciencetostartup.com/paper/stein-variational-ergodic-surface-coverage-with-se-3-constraints - A novel SE(3) Stein Variational Gradient Descent approach for optimizing robotic surface coverage trajectories.
- Declarative Scenario-based Testing with RoadLogic (viability: 3): https://sciencetostartup.com/paper/declarative-scenario-based-testing-with-roadlogic - RoadLogic enables systematic scenario-based testing for autonomous vehicles using declarative specifications.
- ShapeMark: Robust and Diversity-Preserving Watermarking for Diffusion Models (viability: 7): https://sciencetostartup.com/paper/shapemark-robust-and-diversity-preserving-watermarking-for-diffusion-models - ShapeMark offers a robust watermarking solution for diffusion models, ensuring intellectual property protection without sacrificing image quality.
- Variational Routing: A Scalable Bayesian Framework for Calibrated Mixture-of-Experts Transformers (viability: 4): https://sciencetostartup.com/paper/variational-routing-a-scalable-bayesian-framework-for-calibrated-mixture-of-experts-transformers - VMoER provides a scalable Bayesian framework for calibrated uncertainty in Mixture-of-Experts Transformers.
- CyberThreat-Eval: Can Large Language Models Automate Real-World Threat Research? (viability: 9): https://sciencetostartup.com/paper/cyberthreat-eval-can-large-language-models-automate-real-world-threat-research - CyberThreat-Eval automates threat research using LLMs, enhancing the accuracy and efficiency of Cyber Threat Intelligence reports.
- A Guideline-Aware AI Agent for Zero-Shot Target Volume Auto-Delineation (viability: 8): https://sciencetostartup.com/paper/a-guideline-aware-ai-agent-for-zero-shot-target-volume-auto-delineation - OncoAgent is a guideline-aware AI agent that automates target volume delineation in radiotherapy without the need for retraining.
- GIIM: Graph-based Learning of Inter- and Intra-view Dependencies for Multi-view Medical Image Diagnosis (viability: 7): https://sciencetostartup.com/paper/giim-graph-based-learning-of-inter-and-intra-view-dependencies-for-multi-view-medical-image-diagnosis - GIIM is a graph-based framework that enhances multi-view medical image diagnosis by modeling complex dependencies between abnormalities.
- From Weighting to Modeling: A Nonparametric Estimator for Off-Policy Evaluation (viability: 2): https://sciencetostartup.com/paper/from-weighting-to-modeling-a-nonparametric-estimator-for-off-policy-evaluation - A novel nonparametric estimator for off-policy evaluation in contextual bandits that reduces variance while maintaining low bias.
- AI Act Evaluation Benchmark: An Open, Transparent, and Reproducible Evaluation Dataset for NLP and RAG Systems (viability: 6): https://sciencetostartup.com/paper/ai-act-evaluation-benchmark-an-open-transparent-and-reproducible-evaluation-dataset-for-nlp-and-rag-systems - A dataset for evaluating NLP models against EU AI Act compliance, enabling automated assessments of AI systems.
- Common Sense vs. Morality: The Curious Case of Narrative Focus Bias in LLMs (viability: 4): https://sciencetostartup.com/paper/common-sense-vs-morality-the-curious-case-of-narrative-focus-bias-in-llms - CoMoral is a benchmark dataset designed to enhance moral reasoning in LLMs by addressing commonsense contradictions.
- Impact of Markov Decision Process Design on Sim-to-Real Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/impact-of-markov-decision-process-design-on-sim-to-real-reinforcement-learning - This research provides practical guidelines for designing Markov Decision Processes to improve the sim-to-real transfer in reinforcement learning applications.
- An Analysis of Modern Web Security Vulnerabilities Inside WebAssembly Applications (viability: 4): https://sciencetostartup.com/paper/an-analysis-of-modern-web-security-vulnerabilities-inside-webassembly-applications - A comprehensive analysis of WebAssembly security vulnerabilities and mitigation strategies for web applications.
- CERES: A Probabilistic Early Warning System for Acute Food Insecurity (viability: 3): https://sciencetostartup.com/paper/ceres-a-probabilistic-early-warning-system-for-acute-food-insecurity - CERES is an automated probabilistic forecasting system for acute food insecurity, providing timely predictions for high-risk countries.
- Open-World Motion Forecasting (viability: 8): https://sciencetostartup.com/paper/open-world-motion-forecasting - A novel framework for open-world motion forecasting that enables autonomous vehicles to adapt to new object classes in real-time.
- MetaDAT: Generalizable Trajectory Prediction via Meta Pre-training and Data-Adaptive Test-Time Updating (viability: 7): https://sciencetostartup.com/paper/metadat-generalizable-trajectory-prediction-via-meta-pre-training-and-data-adaptive-test-time-updating - A meta-learning framework for real-time trajectory prediction that adapts dynamically to test data for improved accuracy.
- CIGPose: Causal Intervention Graph Neural Network for Whole-Body Pose Estimation (viability: 9): https://sciencetostartup.com/paper/cigpose-causal-intervention-graph-neural-network-for-whole-body-pose-estimation - CIGPose leverages causal intervention to enhance whole-body pose estimation, achieving state-of-the-art accuracy with robust performance in challenging scenes.
- Investigating Gender Stereotypes in Large Language Models via Social Determinants of Health (viability: 4): https://sciencetostartup.com/paper/investigating-gender-stereotypes-in-large-language-models-via-social-determinants-of-health - This study explores gender biases in LLMs using social determinants of health to enhance bias evaluation methods.
- From Flow to One Step: Real-Time Multi-Modal Trajectory Policies via Implicit Maximum Likelihood Estimation-based Distribution Distillation (viability: 6): https://sciencetostartup.com/paper/from-flow-to-one-step-real-time-multi-modal-trajectory-policies-via-implicit-maximum-likelihood-estimation-based-distrib - A framework for real-time multi-modal trajectory policies in robotic manipulation using single-step acceleration methods.
- PromptDLA: A Domain-aware Prompt Document Layout Analysis Framework with Descriptive Knowledge as a Cue (viability: 8): https://sciencetostartup.com/paper/promptdla-a-domain-aware-prompt-document-layout-analysis-framework-with-descriptive-knowledge-as-a-cue - PromptDLA enhances document layout analysis by integrating domain-specific cues for improved model performance.
- Reconstructing Movement from Sparse Samples: Enhanced Spatio-Temporal Matching Strategies for Low-Frequency Data (viability: 3): https://sciencetostartup.com/paper/reconstructing-movement-from-sparse-samples-enhanced-spatio-temporal-matching-strategies-for-low-frequency-data - This paper proposes enhancements to the Spatial-Temporal Matching algorithm for improved GPS trajectory matching.
- RiO-DETR: DETR for Real-time Oriented Object Detection (viability: 7): https://sciencetostartup.com/paper/rio-detr-detr-for-real-time-oriented-object-detection - RiO-DETR is a real-time oriented object detection transformer that enhances accuracy and speed for detecting oriented bounding boxes.
- Reviving ConvNeXt for Efficient Convolutional Diffusion Models (viability: 7): https://sciencetostartup.com/paper/reviving-convnext-for-efficient-convolutional-diffusion-models - Introducing FCDM, an efficient convolutional diffusion model leveraging ConvNeXt for superior performance with fewer resources.
- YOLO-NAS-Bench: A Surrogate Benchmark with Self-Evolving Predictors for YOLO Architecture Search (viability: 7): https://sciencetostartup.com/paper/yolo-nas-bench-a-surrogate-benchmark-with-self-evolving-predictors-for-yolo-architecture-search - YOLO-NAS-Bench is a surrogate benchmark for efficient YOLO architecture search, enhancing predictive accuracy for object detection.
- LLM as a Meta-Judge: Synthetic Data for NLP Evaluation Metric Validation (viability: 7): https://sciencetostartup.com/paper/llm-as-a-meta-judge-synthetic-data-for-nlp-evaluation-metric-validation - A scalable framework using LLMs to generate synthetic datasets for validating NLP evaluation metrics.
- Reward Prediction with Factorized World States (viability: 8): https://sciencetostartup.com/paper/reward-prediction-with-factorized-world-states - StateFactory transforms unstructured observations into structured representations for accurate reward prediction across diverse domains.
- Vision-Augmented On-Track System Identification for Autonomous Racing via Attention-Based Priors and Iterative Neural Correction (viability: 7): https://sciencetostartup.com/paper/vision-augmented-on-track-system-identification-for-autonomous-racing-via-attention-based-priors-and-iterative-neural-co - A vision-augmented system for real-time tire dynamics identification in autonomous racing, enhancing performance and convergence speed.
- ICDAR 2025 Competition on End-to-End Document Image Machine Translation Towards Complex Layouts (viability: 4): https://sciencetostartup.com/paper/icdar-2025-competition-on-end-to-end-document-image-machine-translation-towards-complex-layouts - A competition advancing end-to-end document image translation by jointly modeling text and layout.
- Physics-Informed Neural Engine Sound Modeling with Differentiable Pulse-Train Synthesis (viability: 9): https://sciencetostartup.com/paper/physics-informed-neural-engine-sound-modeling-with-differentiable-pulse-train-synthesis - A physics-informed neural engine sound modeling tool that synthesizes realistic engine audio using pulse-train resonators.
- Training-Free Coverless Multi-Image Steganography with Access Control (viability: 7): https://sciencetostartup.com/paper/training-free-coverless-multi-image-steganography-with-access-control - MIDAS is a training-free framework for coverless image steganography with robust access control for multi-user environments.
- EventVGGT: Exploring Cross-Modal Distillation for Consistent Event-based Depth Estimation (viability: 7): https://sciencetostartup.com/paper/eventvggt-exploring-cross-modal-distillation-for-consistent-event-based-depth-estimation - EventVGGT leverages spatio-temporal knowledge distillation for accurate event-based depth estimation.
- PixelConfig: Longitudinal Measurement and Reverse-Engineering of Meta Pixel Configurations (viability: 2): https://sciencetostartup.com/paper/pixelconfig-longitudinal-measurement-and-reverse-engineering-of-meta-pixel-configurations - PixelConfig offers a framework for analyzing and reverse-engineering Meta Pixel configurations across websites.
- SPAARS: Safer RL Policy Alignment through Abstract Exploration and Refined Exploitation of Action Space (viability: 7): https://sciencetostartup.com/paper/spaars-safer-rl-policy-alignment-through-abstract-exploration-and-refined-exploitation-of-action-space - SPAARS is a curriculum learning framework for safer reinforcement learning that enhances sample efficiency and performance in robotics.
- SinGeo: Unlock Single Model's Potential for Robust Cross-View Geo-Localization (viability: 8): https://sciencetostartup.com/paper/singeo-unlock-single-model-s-potential-for-robust-cross-view-geo-localization - SinGeo is a robust framework for cross-view geo-localization using a single model, outperforming existing methods with state-of-the-art results.
- MIL-PF: Multiple Instance Learning on Precomputed Features for Mammography Classification (viability: 8): https://sciencetostartup.com/paper/mil-pf-multiple-instance-learning-on-precomputed-features-for-mammography-classification - MIL-PF is a scalable framework for efficient mammography classification using precomputed features and lightweight aggregation.
- Quantifying and extending the coverage of spatial categorization data sets (viability: 4): https://sciencetostartup.com/paper/quantifying-and-extending-the-coverage-of-spatial-categorization-data-sets - A project that enhances spatial categorization datasets using LLM-generated labels for better linguistic coverage.
- From Representation to Clusters: A Contrastive Learning Approach for Attributed Hypergraph Clustering (viability: 4): https://sciencetostartup.com/paper/from-representation-to-clusters-a-contrastive-learning-approach-for-attributed-hypergraph-clustering - A novel end-to-end contrastive learning method for attributed hypergraph clustering that improves clustering accuracy.
- M3GCLR: Multi-View Mini-Max Infinite Skeleton-Data Game Contrastive Learning For Skeleton-Based Action Recognition (viability: 3): https://sciencetostartup.com/paper/m3gclr-multi-view-mini-max-infinite-skeleton-data-game-contrastive-learning-for-skeleton-based-action-recognition - M3GCLR introduces a game-theoretic approach to enhance skeleton-based action recognition through contrastive learning.
- Evidential Perfusion Physics-Informed Neural Networks with Residual Uncertainty Quantification (viability: 7): https://sciencetostartup.com/paper/evidential-perfusion-physics-informed-neural-networks-with-residual-uncertainty-quantification - EPPINN enhances computed tomography perfusion imaging by integrating uncertainty quantification with physics-informed neural networks for improved stroke assessment.
- ProvAgent: Threat Detection Based on Identity-Behavior Binding and Multi-Agent Collaborative Attack Investigation (viability: 8): https://sciencetostartup.com/paper/provagent-threat-detection-based-on-identity-behavior-binding-and-multi-agent-collaborative-attack-investigation - ProvAgent revolutionizes threat detection by combining multi-agent systems with traditional models for autonomous investigation.
- Democratising Clinical AI through Dataset Condensation for Classical Clinical Models (viability: 6): https://sciencetostartup.com/paper/democratising-clinical-ai-through-dataset-condensation-for-classical-clinical-models - A method for dataset condensation that enables safe and efficient data sharing for clinical models without compromising patient privacy.
- Interactive 3D visualization of surface roughness predictions in additive manufacturing: A data-driven framework (viability: 7): https://sciencetostartup.com/paper/interactive-3d-visualization-of-surface-roughness-predictions-in-additive-manufacturing-a-data-driven-framework - A data-driven framework for predicting surface roughness in additive manufacturing with an interactive 3D visualization tool.
- TA-GGAD: Testing-time Adaptive Graph Model for Generalist Graph Anomaly Detection (viability: 8): https://sciencetostartup.com/paper/ta-ggad-testing-time-adaptive-graph-model-for-generalist-graph-anomaly-detection - A novel graph foundation model that achieves state-of-the-art cross-domain anomaly detection in graph data.
- Robust Provably Secure Image Steganography via Latent Iterative Optimization (viability: 7): https://sciencetostartup.com/paper/robust-provably-secure-image-steganography-via-latent-iterative-optimization - A secure image steganography framework that enhances message extraction accuracy through latent-space iterative optimization.
- Robust Regularized Policy Iteration under Transition Uncertainty (viability: 7): https://sciencetostartup.com/paper/robust-regularized-policy-iteration-under-transition-uncertainty - Robust Regularized Policy Iteration enhances offline reinforcement learning by optimizing policies against worst-case dynamics under transition uncertainty.
- TaSR-RAG: Taxonomy-guided Structured Reasoning for Retrieval-Augmented Generation (viability: 7): https://sciencetostartup.com/paper/tasr-rag-taxonomy-guided-structured-reasoning-for-retrieval-augmented-generation - TaSR-RAG enhances retrieval-augmented generation by using a taxonomy-guided structured reasoning framework for improved evidence selection.
- Predictive Spectral Calibration for Source-Free Test-Time Regression (viability: 6): https://sciencetostartup.com/paper/predictive-spectral-calibration-for-source-free-test-time-regression - A source-free framework for enhancing image regression through predictive spectral calibration.
- Beyond Scaling: Assessing Strategic Reasoning and Rapid Decision-Making Capability of LLMs in Zero-sum Environments (viability: 7): https://sciencetostartup.com/paper/beyond-scaling-assessing-strategic-reasoning-and-rapid-decision-making-capability-of-llms-in-zero-sum-environments - STAR Benchmark offers a novel framework for evaluating LLMs in competitive, time-sensitive environments, enhancing strategic reasoning and decision-making capabilities.
- TimberAgent: Gram-Guided Retrieval for Executable Music Effect Control (viability: 7): https://sciencetostartup.com/paper/timberagent-gram-guided-retrieval-for-executable-music-effect-control - TimberAgent enables intuitive audio effect control through advanced retrieval techniques for editable plugin configurations.
- Reward-Zero: Language Embedding Driven Implicit Reward Mechanisms for Reinforcement Learning (viability: 5): https://sciencetostartup.com/paper/reward-zero-language-embedding-driven-implicit-reward-mechanisms-for-reinforcement-learning - Reward-Zero transforms natural-language task descriptions into effective implicit rewards for reinforcement learning.
- Measuring onion website discovery and Tor users' interests with honeypots (viability: 2): https://sciencetostartup.com/paper/measuring-onion-website-discovery-and-tor-users-interests-with-honeypots - This research analyzes user engagement on onion websites through a honeypot approach.
- OddGridBench: Exposing the Lack of Fine-Grained Visual Discrepancy Sensitivity in Multimodal Large Language Models (viability: 8): https://sciencetostartup.com/paper/oddgridbench-exposing-the-lack-of-fine-grained-visual-discrepancy-sensitivity-in-multimodal-large-language-models - OddGridBench is a benchmark and framework designed to enhance the visual discrepancy sensitivity of multimodal large language models.
- Reading the Mood Behind Words: Integrating Prosody-Derived Emotional Context into Socially Responsive VR Agents (viability: 4): https://sciencetostartup.com/paper/reading-the-mood-behind-words-integrating-prosody-derived-emotional-context-into-socially-responsive-vr-agents - An emotion-context-aware VR interaction pipeline enhances conversational agents by integrating vocal emotion cues for improved user engagement.
- SpaceSense-Bench: A Large-Scale Multi-Modal Benchmark for Spacecraft Perception and Pose Estimation (viability: 8): https://sciencetostartup.com/paper/spacesense-bench-a-large-scale-multi-modal-benchmark-for-spacecraft-perception-and-pose-estimation - SpaceSense-Bench is a comprehensive multi-modal benchmark for spacecraft perception, providing a diverse dataset and tools for improved navigation and understanding.
- NLiPsCalib: An Efficient Calibration Framework for High-Fidelity 3D Reconstruction of Curved Visuotactile Sensors (viability: 7): https://sciencetostartup.com/paper/nlipscalib-an-efficient-calibration-framework-for-high-fidelity-3d-reconstruction-of-curved-visuotactile-sensors - NLiPsCalib simplifies the calibration of visuotactile sensors for improved 3D reconstruction using everyday objects.
- CLoE: Expert Consistency Learning for Missing Modality Segmentation (viability: 7): https://sciencetostartup.com/paper/cloe-expert-consistency-learning-for-missing-modality-segmentation - CLoE enhances multimodal medical image segmentation by ensuring expert consistency even with missing modalities.
- Curveball Steering: The Right Direction To Steer Isn't Always Linear (viability: 2): https://sciencetostartup.com/paper/curveball-steering-the-right-direction-to-steer-isn-t-always-linear - Curveball steering offers a nonlinear approach to control LLM behavior by respecting the intrinsic geometry of activation spaces.
- IntroSVG: Learning from Rendering Feedback for Text-to-SVG Generation via an Introspective Generator-Critic Framework (viability: 8): https://sciencetostartup.com/paper/introsvg-learning-from-rendering-feedback-for-text-to-svg-generation-via-an-introspective-generator-critic-framework - IntroSVG enhances text-to-SVG generation by integrating visual feedback into an introspective generator-critic framework.
- External entropy supply for IoT devices employing a RISC-V Trusted Execution Environment (viability: 3): https://sciencetostartup.com/paper/external-entropy-supply-for-iot-devices-employing-a-risc-v-trusted-execution-environment - A solution for secure cryptographic key generation in IoT devices using RISC-V Trusted Execution Environment.
- A Gaussian Comparison Theorem for Training Dynamics in Machine Learning (viability: 3): https://sciencetostartup.com/paper/a-gaussian-comparison-theorem-for-training-dynamics-in-machine-learning - A theoretical framework for analyzing training dynamics in Gaussian mixture models.
- Rescaling Confidence: What Scale Design Reveals About LLM Metacognition (viability: 4): https://sciencetostartup.com/paper/rescaling-confidence-what-scale-design-reveals-about-llm-metacognition - This research reveals how confidence scale design impacts the metacognitive performance of LLMs.
- CORAL: Scalable Multi-Task Robot Learning via LoRA Experts (viability: 8): https://sciencetostartup.com/paper/coral-scalable-multi-task-robot-learning-via-lora-experts - CORAL is a scalable framework for multi-task robotic learning that mitigates task interference using lightweight LoRA experts.
- TA-Mem: Tool-Augmented Autonomous Memory Retrieval for LLM in Long-Term Conversational QA (viability: 7): https://sciencetostartup.com/paper/ta-mem-tool-augmented-autonomous-memory-retrieval-for-llm-in-long-term-conversational-qa - TA-Mem enhances long-term conversational QA by autonomously retrieving and structuring memory for LLMs.
- Diagnosing and Repairing Citation Failures in Generative Engine Optimization (viability: 7): https://sciencetostartup.com/paper/diagnosing-and-repairing-citation-failures-in-generative-engine-optimization - A diagnostic system that improves citation rates for AI-generated content by identifying and addressing citation failures.
- See, Plan, Rewind: Progress-Aware Vision-Language-Action Models for Robust Robotic Manipulation (viability: 7): https://sciencetostartup.com/paper/see-plan-rewind-progress-aware-vision-language-action-models-for-robust-robotic-manipulation - A vision-language-action framework for robust robotic manipulation that enhances task progress awareness and error correction.
- DenoiseSplat: Feed-Forward Gaussian Splatting for Noisy 3D Scene Reconstruction (viability: 7): https://sciencetostartup.com/paper/denoisesplat-feed-forward-gaussian-splatting-for-noisy-3d-scene-reconstruction - DenoiseSplat offers a novel approach to 3D scene reconstruction from noisy images, enhancing performance in real-world applications.
- Proxy-Guided Measurement Calibration (viability: 3): https://sciencetostartup.com/paper/proxy-guided-measurement-calibration - A framework for correcting systematic measurement errors in outcome variables using proxy variables.
- Exploring Modality-Aware Fusion and Decoupled Temporal Propagation for Multi-Modal Object Tracking (viability: 7): https://sciencetostartup.com/paper/exploring-modality-aware-fusion-and-decoupled-temporal-propagation-for-multi-modal-object-tracking - MDTrack enhances multimodal object tracking through modality-aware fusion and decoupled temporal propagation.
- CogBlender: Towards Continuous Cognitive Intervention in Text-to-Image Generation (viability: 5): https://sciencetostartup.com/paper/cogblender-towards-continuous-cognitive-intervention-in-text-to-image-generation - CogBlender enables precise cognitive intervention in text-to-image generation for enhanced creative design.
- Learning Convex Decomposition via Feature Fields (viability: 7): https://sciencetostartup.com/paper/learning-convex-decomposition-via-feature-fields - A novel feed-forward model for high-quality convex decomposition of 3D shapes, enhancing collision detection and physical simulations.
- From Ideal to Real: Stable Video Object Removal under Imperfect Conditions (viability: 8): https://sciencetostartup.com/paper/from-ideal-to-real-stable-video-object-removal-under-imperfect-conditions - SVOR is a robust framework for removing objects from videos while maintaining visual consistency under real-world imperfections.
- Speeding Up the Learning of 3D Gaussians with Much Shorter Gaussian Lists (viability: 7): https://sciencetostartup.com/paper/speeding-up-the-learning-of-3d-gaussians-with-much-shorter-gaussian-lists - A novel approach to enhance the efficiency of 3D Gaussian splatting for improved rendering performance.
- On Regret Bounds of Thompson Sampling for Bayesian Optimization (viability: 2): https://sciencetostartup.com/paper/on-regret-bounds-of-thompson-sampling-for-bayesian-optimization - This paper analyzes regret bounds for Gaussian process Thompson sampling in Bayesian optimization.
- DendroNN: Dendrocentric Neural Networks for Energy-Efficient Classification of Event-Based Data (viability: 4): https://sciencetostartup.com/paper/dendronn-dendrocentric-neural-networks-for-energy-efficient-classification-of-event-based-data - DendroNN introduces a novel neural network architecture for energy-efficient classification of event-based data using dendritic mechanisms.
- Logos: An evolvable reasoning engine for rational molecular design (viability: 8): https://sciencetostartup.com/paper/logos-an-evolvable-reasoning-engine-for-rational-molecular-design - Logos is a compact molecular reasoning engine that integrates logical reasoning with chemical consistency for reliable molecular design.
- ForgeDreamer: Industrial Text-to-3D Generation with Multi-Expert LoRA and Cross-View Hypergraph (viability: 8): https://sciencetostartup.com/paper/forgedreamer-industrial-text-to-3d-generation-with-multi-expert-lora-and-cross-view-hypergraph - ForgeDreamer revolutionizes industrial text-to-3D generation by leveraging a Multi-Expert LoRA Ensemble and Cross-View Hypergraph for enhanced precision and generalization.
- Implicit Geometry Representations for Vision-and-Language Navigation from Web Videos (viability: 7): https://sciencetostartup.com/paper/implicit-geometry-representations-for-vision-and-language-navigation-from-web-videos - A framework for Vision-and-Language Navigation that leverages web videos to enhance spatial reasoning and navigation capabilities.
- Multimodal Graph Representation Learning with Dynamic Information Pathways (viability: 4): https://sciencetostartup.com/paper/multimodal-graph-representation-learning-with-dynamic-information-pathways - A novel framework for multimodal graph representation learning that enhances message propagation across different modalities.
- Transductive Generalization via Optimal Transport and Its Application to Graph Node Classification (viability: 7): https://sciencetostartup.com/paper/transductive-generalization-via-optimal-transport-and-its-application-to-graph-node-classification - A novel approach to improve transductive generalization in graph node classification using optimal transport.
- Multi-model approach for autonomous driving: A comprehensive study on traffic sign-, vehicle- and lane detection and behavioral cloning (viability: 4): https://sciencetostartup.com/paper/multi-model-approach-for-autonomous-driving-a-comprehensive-study-on-traffic-sign-vehicle-and-lane-detection-and-behavio - A multi-model approach to enhance self-driving car performance through advanced traffic sign, vehicle, and lane detection techniques.
- Efficient Reasoning at Fixed Test-Time Cost via Length-Aware Attention Priors and Gain-Aware Training (viability: 2): https://sciencetostartup.com/paper/efficient-reasoning-at-fixed-test-time-cost-via-length-aware-attention-priors-and-gain-aware-training - This research proposes a method for efficient reasoning in Transformers under tight compute constraints.
- A Generative Sampler for distributions with possible discrete parameter based on Reversibility (viability: 4): https://sciencetostartup.com/paper/a-generative-sampler-for-distributions-with-possible-discrete-parameter-based-on-reversibility - A novel generative sampling framework that uses time-reversibility to sample from complex distributions without target score functions.
- Social-R1: Towards Human-like Social Reasoning in LLMs (viability: 7): https://sciencetostartup.com/paper/social-r1-towards-human-like-social-reasoning-in-llms - Social-R1 enhances large language models with human-like social reasoning through a novel reinforcement learning framework.
- Reasoning-Oriented Programming: Chaining Semantic Gadgets to Jailbreak Large Vision Language Models (viability: 8): https://sciencetostartup.com/paper/reasoning-oriented-programming-chaining-semantic-gadgets-to-jailbreak-large-vision-language-models - Introducing a framework that exploits vulnerabilities in large vision-language models to bypass safety alignment.
- Towards Instance Segmentation with Polygon Detection Transformers (viability: 7): https://sciencetostartup.com/paper/towards-instance-segmentation-with-polygon-detection-transformers - Poly-DETR revolutionizes instance segmentation by using Polygon Detection Transformers for efficient and accurate segmentation.
- When Detectors Forget Forensics: Blocking Semantic Shortcuts for Generalizable AI-Generated Image Detection (viability: 7): https://sciencetostartup.com/paper/when-detectors-forget-forensics-blocking-semantic-shortcuts-for-generalizable-ai-generated-image-detection - A novel module that enhances AI-generated image detection by decoupling semantic components from learned representations.
- RAE-NWM: Navigation World Model in Dense Visual Representation Space (viability: 7): https://sciencetostartup.com/paper/rae-nwm-navigation-world-model-in-dense-visual-representation-space - RAE-NWM enhances visual navigation by modeling dynamics in a dense representation space for improved action accuracy.
- MO-Playground: Massively Parallelized Multi-Objective Reinforcement Learning for Robotics (viability: 8): https://sciencetostartup.com/paper/mo-playground-massively-parallelized-multi-objective-reinforcement-learning-for-robotics - MO-Playground accelerates multi-objective reinforcement learning for robotics with a GPU-native algorithm and a user-friendly environment.
- BridgeDiff: Bridging Human Observations and Flat-Garment Synthesis for Virtual Try-Off (viability: 8): https://sciencetostartup.com/paper/bridgediff-bridging-human-observations-and-flat-garment-synthesis-for-virtual-try-off - BridgeDiff enhances virtual try-on experiences by accurately synthesizing flat-garment representations from dressed images.
- HelixTrack: Event-Based Tracking and RPM Estimation of Propeller-like Objects (viability: 7): https://sciencetostartup.com/paper/helixtrack-event-based-tracking-and-rpm-estimation-of-propeller-like-objects - HelixTrack offers microsecond-latency tracking and RPM estimation for propeller-like objects using an innovative event-driven approach.
- How Contrastive Decoding Enhances Large Audio Language Models? (viability: 4): https://sciencetostartup.com/paper/how-contrastive-decoding-enhances-large-audio-language-models - This research explores how Contrastive Decoding can improve the performance of Large Audio Language Models.
- Cognitively Layered Data Synthesis for Domain Adaptation of LLMs to Space Situational Awareness (viability: 8): https://sciencetostartup.com/paper/cognitively-layered-data-synthesis-for-domain-adaptation-of-llms-to-space-situational-awareness - A framework for generating high-quality fine-tuning datasets for LLMs in space situational awareness.
- TRIP-Bag: A Portable Teleoperation System for Plug-and-Play Robotic Arms and Leaders (viability: 7): https://sciencetostartup.com/paper/trip-bag-a-portable-teleoperation-system-for-plug-and-play-robotic-arms-and-leaders - TRIP-Bag is a portable teleoperation system that simplifies data collection for robotic manipulation tasks.
- UniField: A Unified Field-Aware MRI Enhancement Framework (viability: 8): https://sciencetostartup.com/paper/unifield-a-unified-field-aware-mri-enhancement-framework - UniField is a unified framework for enhancing MRI imaging by leveraging shared degradation patterns across different field strengths.
- LooComp: Leverage Leave-One-Out Strategy to Encoder-only Transformer for Efficient Query-aware Context Compression (viability: 7): https://sciencetostartup.com/paper/loocomp-leverage-leave-one-out-strategy-to-encoder-only-transformer-for-efficient-query-aware-context-compression - A lightweight encoder-only Transformer for efficient query-aware context compression in question answering.
- Beyond Test-Time Training: Learning to Reason via Hardware-Efficient Optimal Control (viability: 7): https://sciencetostartup.com/paper/beyond-test-time-training-learning-to-reason-via-hardware-efficient-optimal-control - Introducing a hardware-efficient Test-Time Control layer to enhance reasoning capabilities in language models through optimal control.
- Distributed Convolutional Neural Networks for Object Recognition (viability: 4): https://sciencetostartup.com/paper/distributed-convolutional-neural-networks-for-object-recognition - A novel loss function for distributed convolutional neural networks enhances object recognition for specific positive classes.
- Embodied Human Simulation for Quantitative Design and Analysis of Interactive Robotics (viability: 6): https://sciencetostartup.com/paper/embodied-human-simulation-for-quantitative-design-and-analysis-of-interactive-robotics - A scalable simulation framework for optimizing human-robot interactions in physical robotics.
- TubeMLLM: A Foundation Model for Topology Knowledge Exploration in Vessel-like Anatomy (viability: 8): https://sciencetostartup.com/paper/tubemllm-a-foundation-model-for-topology-knowledge-exploration-in-vessel-like-anatomy - TubeMLLM is a foundation model designed for enhanced topology-aware perception in medical vessel-like anatomy.
- SPAR-K: Scheduled Periodic Alternating Early Exit for Spoken Language Models (viability: 7): https://sciencetostartup.com/paper/spar-k-scheduled-periodic-alternating-early-exit-for-spoken-language-models - SPAR-K accelerates spoken language model inference with a novel early exit framework that maintains quality while reducing computational costs.
- PrivPRISM: Automatically Detecting Discrepancies Between Google Play Data Safety Declarations and Developer Privacy Policies (viability: 3): https://sciencetostartup.com/paper/privprism-automatically-detecting-discrepancies-between-google-play-data-safety-declarations-and-developer-privacy-polic - PrivPRISM automates the detection of discrepancies between app data safety declarations and privacy policies to enhance user protection.
- Geometry-Aware Metric Learning for Cross-Lingual Few-Shot Sign Language Recognition on Static Hand Keypoints (viability: 7): https://sciencetostartup.com/paper/geometry-aware-metric-learning-for-cross-lingual-few-shot-sign-language-recognition-on-static-hand-keypoints - A geometry-aware metric-learning framework for effective cross-lingual few-shot sign language recognition using invariant hand-geometry descriptors.
- Abundant Intelligence and Deficient Demand: A Macro-Financial Stress Test of Rapid AI Adoption (viability: 2): https://sciencetostartup.com/paper/abundant-intelligence-and-deficient-demand-a-macro-financial-stress-test-of-rapid-ai-adoption - This paper explores the macro-financial implications of rapid AI adoption and its potential to create demand deficiencies.
- Strategically Robust Multi-Agent Reinforcement Learning with Linear Function Approximation (viability: 4): https://sciencetostartup.com/paper/strategically-robust-multi-agent-reinforcement-learning-with-linear-function-approximation - A robust multi-agent reinforcement learning algorithm that enhances equilibrium learning through Risk-Sensitive Quantal Response Equilibrium.
- MM-Zero: Self-Evolving Multi-Model Vision Language Models From Zero Data (viability: 4): https://sciencetostartup.com/paper/mm-zero-self-evolving-multi-model-vision-language-models-from-zero-data - MM-Zero is a self-evolving framework for Vision Language Models that enhances reasoning capabilities without the need for data.
- Emotion is Not Just a Label: Latent Emotional Factors in LLM Processing (viability: 6): https://sciencetostartup.com/paper/emotion-is-not-just-a-label-latent-emotional-factors-in-llm-processing - AURA-QA enhances question-answering by integrating emotional factors into language model training.
- Evaluate-as-Action: Self-Evaluated Process Rewards for Retrieval-Augmented Agents (viability: 7): https://sciencetostartup.com/paper/evaluate-as-action-self-evaluated-process-rewards-for-retrieval-augmented-agents - EvalAct enhances retrieval-augmented agents by integrating explicit evaluation into the retrieval process for improved multi-hop question answering.
- The Radio-Frequency Transformer for Signal Separation (viability: 8): https://sciencetostartup.com/paper/the-radio-frequency-transformer-for-signal-separation - A novel transformer-based approach for signal separation that significantly outperforms existing methods in RF applications.
- The Reasoning Trap -- Logical Reasoning as a Mechanistic Pathway to Situational Awareness (viability: 3): https://sciencetostartup.com/paper/the-reasoning-trap-logical-reasoning-as-a-mechanistic-pathway-to-situational-awareness - Exploring the intersection of logical reasoning and situational awareness in AI systems to propose new safety measures.
- $P^2$GNN: Two Prototype Sets to boost GNN Performance (viability: 7): https://sciencetostartup.com/paper/p-2-gnn-two-prototype-sets-to-boost-gnn-performance - P^2GNN enhances message passing in Graph Neural Networks by introducing prototypes for improved global context and noise reduction.
- WESPR: Wind-adaptive Energy-Efficient Safe Perception & Planning for Robust Flight with Quadrotors (viability: 6): https://sciencetostartup.com/paper/wespr-wind-adaptive-energy-efficient-safe-perception-planning-for-robust-flight-with-quadrotors - WESPR enhances drone navigation by predicting wind effects on flight paths for improved stability and efficiency.
- Explainable Innovation Engine: Dual-Tree Agent-RAG with Methods-as-Nodes and Verifiable Write-Back (viability: 8): https://sciencetostartup.com/paper/explainable-innovation-engine-dual-tree-agent-rag-with-methods-as-nodes-and-verifiable-write-back - A novel Explainable Innovation Engine that enhances retrieval-augmented generation with methods-as-nodes for improved control and verifiability.
- ZipPIR: High-throughput Single-server PIR without Client-side Storage (viability: 7): https://sciencetostartup.com/paper/zippir-high-throughput-single-server-pir-without-client-side-storage - ZipPIR offers a high-throughput Private Information Retrieval protocol with no client-side storage, enhancing scalability for practical deployments.
- Robust Spatiotemporal Motion Planning for Multi-Agent Autonomous Racing via Topological Gap Identification and Accelerated MPC (viability: 7): https://sciencetostartup.com/paper/robust-spatiotemporal-motion-planning-for-multi-agent-autonomous-racing-via-topological-gap-identification-and-accelerat - A robust motion planning framework for high-speed multi-agent autonomous racing that optimizes overtaking maneuvers.
- The Costs of Reproducibility in Music Separation Research: a Replication of Band-Split RNN (viability: 5): https://sciencetostartup.com/paper/the-costs-of-reproducibility-in-music-separation-research-a-replication-of-band-split-rnn - A study on the reproducibility of music source separation models with insights and optimized implementations.
- DEO: Training-Free Direct Embedding Optimization for Negation-Aware Retrieval (viability: 7): https://sciencetostartup.com/paper/deo-training-free-direct-embedding-optimization-for-negation-aware-retrieval - DEO is a training-free method that enhances retrieval accuracy for negation and exclusion queries in text and multimodal contexts.
- Latent-DARM: Bridging Discrete Diffusion And Autoregressive Models For Reasoning (viability: 7): https://sciencetostartup.com/paper/latent-darm-bridging-discrete-diffusion-and-autoregressive-models-for-reasoning - Latent-DARM enhances multi-agent collaboration by bridging discrete diffusion and autoregressive models for improved reasoning.
- DuplexCascade: Full-Duplex Speech-to-Speech Dialogue with VAD-Free Cascaded ASR-LLM-TTS Pipeline and Micro-Turn Optimization (viability: 6): https://sciencetostartup.com/paper/duplexcascade-full-duplex-speech-to-speech-dialogue-with-vad-free-cascaded-asr-llm-tts-pipeline-and-micro-turn-optimizat - DuplexCascade enables full-duplex speech-to-speech dialogue with a VAD-free cascaded pipeline for enhanced conversational interaction.
- STONE Dataset: A Scalable Multi-Modal Surround-View 3D Traversability Dataset for Off-Road Robot Navigation (viability: 7): https://sciencetostartup.com/paper/stone-dataset-a-scalable-multi-modal-surround-view-3d-traversability-dataset-for-off-road-robot-navigation - STONE is a large-scale multi-modal dataset designed to enhance off-road robot navigation through accurate traversability mapping.
- Point Cloud as a Foreign Language for Multi-modal Large Language Model (viability: 9): https://sciencetostartup.com/paper/point-cloud-as-a-foreign-language-for-multi-modal-large-language-model - SAGE is an end-to-end 3D multi-modal large language model that processes raw point clouds for enhanced 3D understanding.
- Progressive Split Mamba: Effective State Space Modelling for Image Restoration (viability: 3): https://sciencetostartup.com/paper/progressive-split-mamba-effective-state-space-modelling-for-image-restoration - Progressive Split-Mamba enhances image restoration by combining locality preservation with efficient global propagation.
- ZeroWBC: Learning Natural Visuomotor Humanoid Control Directly from Human Egocentric Video (viability: 7): https://sciencetostartup.com/paper/zerowbc-learning-natural-visuomotor-humanoid-control-directly-from-human-egocentric-video - ZeroWBC enables natural humanoid robot control from human egocentric videos, eliminating the need for costly teleoperation data.
- Better Bounds for the Distributed Experts Problem (viability: 2): https://sciencetostartup.com/paper/better-bounds-for-the-distributed-experts-problem - This paper presents a theoretical protocol for minimizing regret in distributed expert systems.
- Optimal partition selection with Rényi differential privacy (viability: 2): https://sciencetostartup.com/paper/optimal-partition-selection-with-r-nyi-differential-privacy - This paper presents a theoretical advancement in partition selection under Rényi differential privacy, but lacks practical implementation details.
- Fast and Optimal Differentially Private Frequent-Substring Mining (viability: 4): https://sciencetostartup.com/paper/fast-and-optimal-differentially-private-frequent-substring-mining - A novel algorithm for efficient and privacy-preserving frequent substring mining from user-contributed strings.
- GIAT: A Geologically-Informed Attention Transformer for Lithology Identification (viability: 8): https://sciencetostartup.com/paper/giat-a-geologically-informed-attention-transformer-for-lithology-identification - GIAT is a novel Transformer framework that enhances lithology identification by integrating geological priors for improved accuracy and interpretability.
- SPAN-Nav: Generalized Spatial Awareness for Versatile Vision-Language Navigation (viability: 8): https://sciencetostartup.com/paper/span-nav-generalized-spatial-awareness-for-versatile-vision-language-navigation - SPAN-Nav enhances embodied navigation with advanced spatial awareness using a compact representation of 3D cues.
- Wrong Code, Right Structure: Learning Netlist Representations from Imperfect LLM-Generated RTL (viability: 6): https://sciencetostartup.com/paper/wrong-code-right-structure-learning-netlist-representations-from-imperfect-llm-generated-rtl - A framework that leverages imperfect LLM-generated RTL for effective netlist representation learning.
- RubiCap: Rubric-Guided Reinforcement Learning for Dense Image Captioning (viability: 8): https://sciencetostartup.com/paper/rubicap-rubric-guided-reinforcement-learning-for-dense-image-captioning - RubiCap leverages reinforcement learning with LLM-written rubrics to enhance diversity and quality in dense image captioning.
- Real-Time Trust Verification for Safe Agentic Actions using TrustBench (viability: 8): https://sciencetostartup.com/paper/real-time-trust-verification-for-safe-agentic-actions-using-trustbench - TrustBench provides real-time trust verification for autonomous agents to prevent harmful actions before execution.
- Bioalignment: Measuring and Improving LLM Disposition Toward Biological Systems for AI Safety (viability: 8): https://sciencetostartup.com/paper/bioalignment-measuring-and-improving-llm-disposition-toward-biological-systems-for-ai-safety - A fine-tuning approach to align large language models with biological solutions, enhancing AI safety.
- DataFactory: Collaborative Multi-Agent Framework for Advanced Table Question Answering (viability: 3): https://sciencetostartup.com/paper/datafactory-collaborative-multi-agent-framework-for-advanced-table-question-answering - DataFactory is a multi-agent framework designed to enhance Table Question Answering by improving coordination and reducing hallucinations.
- Deep Tabular Research via Continual Experience-Driven Execution (viability: 6): https://sciencetostartup.com/paper/deep-tabular-research-via-continual-experience-driven-execution - A novel framework for enhancing long-horizon reasoning over complex tabular data using agentic decision-making.
- RTFDNet: Fusion-Decoupling for Robust RGB-T Segmentation (viability: 8): https://sciencetostartup.com/paper/rtfdnet-fusion-decoupling-for-robust-rgb-t-segmentation - RTFDNet enhances RGB-T segmentation for robust robotic systems in low-light environments through innovative fusion-decoupling techniques.
- Walking on Rough Terrain with Any Number of Legs (viability: 2): https://sciencetostartup.com/paper/walking-on-rough-terrain-with-any-number-of-legs - A novel control architecture for multi-legged robots navigating rough terrain.
- Causally Sufficient and Necessary Feature Expansion for Class-Incremental Learning (viability: 5): https://sciencetostartup.com/paper/causally-sufficient-and-necessary-feature-expansion-for-class-incremental-learning - A novel regularization method to enhance feature expansion in Class Incremental Learning, addressing catastrophic forgetting.
- Agentic AI as a Network Control-Plane Intelligence Layer for Federated Learning over 6G (viability: 4): https://sciencetostartup.com/paper/agentic-ai-as-a-network-control-plane-intelligence-layer-for-federated-learning-over-6g - Agentic AI enhances federated learning by managing network conditions for on-device training over 6G.
- Rotation Equivariant Mamba for Vision Tasks (viability: 8): https://sciencetostartup.com/paper/rotation-equivariant-mamba-for-vision-tasks - EQ-VMamba introduces a rotation equivariant architecture for vision tasks, enhancing robustness and efficiency in visual Mamba models.
- Transformer-Based Multi-Region Segmentation and Radiomic Analysis of HR-pQCT Imaging (viability: 7): https://sciencetostartup.com/paper/transformer-based-multi-region-segmentation-and-radiomic-analysis-of-hr-pqct-imaging - Automated multi-region segmentation and radiomic analysis of HR-pQCT images for improved osteoporosis detection.
- AgenticCyOps: Securing Multi-Agentic AI Integration in Enterprise Cyber Operations (viability: 2): https://sciencetostartup.com/paper/agenticcyops-securing-multi-agentic-ai-integration-in-enterprise-cyber-operations - AgenticCyOps provides a framework for securing multi-agent AI systems in enterprise environments.
- Chaotic Dynamics in Multi-LLM Deliberation (viability: 4): https://sciencetostartup.com/paper/chaotic-dynamics-in-multi-llm-deliberation - A study on the stability of multi-LLM deliberation systems to enhance governance protocols.
- QUSR: Quality-Aware and Uncertainty-Guided Image Super-Resolution Diffusion Model (viability: 8): https://sciencetostartup.com/paper/qusr-quality-aware-and-uncertainty-guided-image-super-resolution-diffusion-model - QUSR is a novel diffusion model for high-quality image super-resolution that adapts noise levels based on uncertainty.
- DexHiL: A Human-in-the-Loop Framework for Vision-Language-Action Model Post-Training in Dexterous Manipulation (viability: 8): https://sciencetostartup.com/paper/dexhil-a-human-in-the-loop-framework-for-vision-language-action-model-post-training-in-dexterous-manipulation - DexHiL is a human-in-the-loop framework that enhances dexterous manipulation in robotic systems through coordinated interventions and adaptive learning.
- Decoupling Reasoning and Confidence: Resurrecting Calibration in Reinforcement Learning from Verifiable Rewards (viability: 4): https://sciencetostartup.com/paper/decoupling-reasoning-and-confidence-resurrecting-calibration-in-reinforcement-learning-from-verifiable-rewards - A framework that decouples reasoning and calibration in reinforcement learning to enhance LLM reliability.
- PM-Nav: Priori-Map Guided Embodied Navigation in Functional Buildings (viability: 7): https://sciencetostartup.com/paper/pm-nav-priori-map-guided-embodied-navigation-in-functional-buildings - PM-Nav enhances embodied navigation in functional buildings using semantic priori-maps for improved path planning.
- Progressive Representation Learning for Multimodal Sentiment Analysis with Incomplete Modalities (viability: 7): https://sciencetostartup.com/paper/progressive-representation-learning-for-multimodal-sentiment-analysis-with-incomplete-modalities - A framework for robust multimodal sentiment analysis that effectively handles incomplete data modalities.
- VIVID-Med: LLM-Supervised Structured Pretraining for Deployable Medical ViTs (viability: 8): https://sciencetostartup.com/paper/vivid-med-llm-supervised-structured-pretraining-for-deployable-medical-vits - VIVID-Med is a deployable framework for medical image analysis that leverages LLM-supervised structured pretraining for enhanced performance with minimal data.
- Composed Vision-Language Retrieval for Skin Cancer Case Search via Joint Alignment of Global and Local Representations (viability: 7): https://sciencetostartup.com/paper/composed-vision-language-retrieval-for-skin-cancer-case-search-via-joint-alignment-of-global-and-local-representations - A vision-language retrieval system for skin cancer cases that enhances diagnostic decision-making through advanced image-text alignment.
- Training-free Motion Factorization for Compositional Video Generation (viability: 6): https://sciencetostartup.com/paper/training-free-motion-factorization-for-compositional-video-generation - A motion factorization framework for compositional video generation that enhances motion understanding and synthesis.
- Probabilistic Hysteresis Factor Prediction for Electric Vehicle Batteries with Graphite Anodes Containing Silicon (viability: 5): https://sciencetostartup.com/paper/probabilistic-hysteresis-factor-prediction-for-electric-vehicle-batteries-with-graphite-anodes-containing-silicon - A data-driven approach for predicting hysteresis factors in silicon-graphite battery anodes to improve state-of-charge estimation.
- MedKCO: Medical Vision-Language Pretraining via Knowledge-Driven Cognitive Orchestration (viability: 8): https://sciencetostartup.com/paper/medkco-medical-vision-language-pretraining-via-knowledge-driven-cognitive-orchestration - MedKCO enhances medical vision-language models through a knowledge-driven approach for improved feature representation.
- Reading, Not Thinking: Understanding and Bridging the Modality Gap When Text Becomes Pixels in Multimodal LLMs (viability: 7): https://sciencetostartup.com/paper/reading-not-thinking-understanding-and-bridging-the-modality-gap-when-text-becomes-pixels-in-multimodal-llms - A method to enhance multimodal large language models' performance on visual text by addressing the modality gap.
- Chain of Event-Centric Causal Thought for Physically Plausible Video Generation (viability: 7): https://sciencetostartup.com/paper/chain-of-event-centric-causal-thought-for-physically-plausible-video-generation - A framework for generating physically plausible videos through causal event reasoning and cross-modal prompting.
- Overcoming Valid Action Suppression in Unmasked Policy Gradient Algorithms (viability: 4): https://sciencetostartup.com/paper/overcoming-valid-action-suppression-in-unmasked-policy-gradient-algorithms - A novel approach to enhance action validity in reinforcement learning by overcoming suppression in unmasked policy gradient algorithms.
- Latent World Models for Automated Driving: A Unified Taxonomy, Evaluation Framework, and Open Challenges (viability: 2): https://sciencetostartup.com/paper/latent-world-models-for-automated-driving-a-unified-taxonomy-evaluation-framework-and-open-challenges - This paper presents a unifying framework for latent representations in automated driving, focusing on simulation and decision-making.
- Not All News Is Equal: Topic- and Event-Conditional Sentiment from Finetuned LLMs for Aluminum Price Forecasting (viability: 6): https://sciencetostartup.com/paper/not-all-news-is-equal-topic-and-event-conditional-sentiment-from-finetuned-llms-for-aluminum-price-forecasting - A sentiment analysis tool using finetuned LLMs to forecast aluminum prices based on news headlines.
- OmniEdit: A Training-free framework for Lip Synchronization and Audio-Visual Editing (viability: 7): https://sciencetostartup.com/paper/omniedit-a-training-free-framework-for-lip-synchronization-and-audio-visual-editing - OmniEdit is a training-free framework for efficient lip synchronization and audio-visual editing.
- Provably Safe Trajectory Generation for Manipulators Under Motion and Environmental Uncertainties (viability: 7): https://sciencetostartup.com/paper/provably-safe-trajectory-generation-for-manipulators-under-motion-and-environmental-uncertainties - A novel risk-bounded motion planning framework for robot manipulators that ensures safe trajectory generation in uncertain environments.
- PPO-Based Hybrid Optimization for RIS-Assisted Semantic Vehicular Edge Computing (viability: 5): https://sciencetostartup.com/paper/ppo-based-hybrid-optimization-for-ris-assisted-semantic-vehicular-edge-computing - A hybrid optimization framework for latency-sensitive IoV applications using RIS and semantic communication.
- GST-VLA: Structured Gaussian Spatial Tokens for 3D Depth-Aware Vision-Language-Action Models (viability: 3): https://sciencetostartup.com/paper/gst-vla-structured-gaussian-spatial-tokens-for-3d-depth-aware-vision-language-action-models - GST-VLA introduces a novel Gaussian Spatial Tokenizer for enhanced 3D depth-aware vision-language-action modeling.
- Exclusive Self Attention (viability: 3): https://sciencetostartup.com/paper/exclusive-self-attention - Exclusive Self Attention (XSA) enhances Transformer performance by improving context modeling through orthogonal information capture.
- High-Slip-Ratio Control for Peak Tire-Road Friction Estimation Using Automated Vehicles (viability: 7): https://sciencetostartup.com/paper/high-slip-ratio-control-for-peak-tire-road-friction-estimation-using-automated-vehicles - A control framework for automated vehicles to accurately estimate tire-road friction under various conditions.
- A Text-Native Interface for Generative Video Authoring (viability: 7): https://sciencetostartup.com/paper/a-text-native-interface-for-generative-video-authoring - Doki is a text-native interface that simplifies generative video authoring for everyone.
- 3D UAV Trajectory Estimation and Classification from Internet Videos via Language Model (viability: 7): https://sciencetostartup.com/paper/3d-uav-trajectory-estimation-and-classification-from-internet-videos-via-language-model - A framework for estimating and classifying UAV trajectories from internet videos using language-driven data acquisition.
- Intelligent Spatial Estimation for Fire Hazards in Engineering Sites: An Enhanced YOLOv8-Powered Proximity Analysis Framework (viability: 7): https://sciencetostartup.com/paper/intelligent-spatial-estimation-for-fire-hazards-in-engineering-sites-an-enhanced-yolov8-powered-proximity-analysis-frame - An enhanced YOLOv8 framework for intelligent fire detection and risk assessment in engineering sites.
- Verifying Good Regulator Conditions for Hypergraph Observers: Natural Gradient Learning from Causal Invariance via Established Theorems (viability: 3): https://sciencetostartup.com/paper/verifying-good-regulator-conditions-for-hypergraph-observers-natural-gradient-learning-from-causal-invariance-via-establ - This paper explores the conditions under which hypergraph observers can effectively minimize prediction error using natural gradient learning.
- Learning Adaptive LLM Decoding (viability: 7): https://sciencetostartup.com/paper/learning-adaptive-llm-decoding - A system that learns adaptive decoding strategies for large language models to optimize performance based on task difficulty and resource availability.
- Dynamic Multi-period Experts for Online Time Series Forecasting (viability: 7): https://sciencetostartup.com/paper/dynamic-multi-period-experts-for-online-time-series-forecasting - DynaME is a hybrid framework for online time series forecasting that adapts to concept drift through specialized expert models.
- Adaptive Active Learning for Online Reliability Prediction of Satellite Electronics (viability: 4): https://sciencetostartup.com/paper/adaptive-active-learning-for-online-reliability-prediction-of-satellite-electronics - A novel framework for online reliability prediction of satellite electronics using adaptive active learning.
- Quality over Quantity: Demonstration Curation via Influence Functions for Data-Centric Robot Learning (viability: 6): https://sciencetostartup.com/paper/quality-over-quantity-demonstration-curation-via-influence-functions-for-data-centric-robot-learning - A systematic approach to improve robot learning by curating high-quality demonstration data using influence functions.
- Spectral-Structured Diffusion for Single-Image Rain Removal (viability: 7): https://sciencetostartup.com/paper/spectral-structured-diffusion-for-single-image-rain-removal - SpectralDiff is a novel spectral-structured diffusion framework designed for efficient single-image rain removal.
- Sim2Act: Robust Simulation-to-Decision Learning via Adversarial Calibration and Group-Relative Perturbation (viability: 4): https://sciencetostartup.com/paper/sim2act-robust-simulation-to-decision-learning-via-adversarial-calibration-and-group-relative-perturbation - Sim2Act enhances decision-making in simulation environments by improving robustness against prediction errors.
- From Days to Minutes: An Autonomous AI Agent Achieves Reliable Clinical Triage in Remote Patient Monitoring (viability: 6): https://sciencetostartup.com/paper/from-days-to-minutes-an-autonomous-ai-agent-achieves-reliable-clinical-triage-in-remote-patient-monitoring - Sentinel is an autonomous AI agent that enhances remote patient monitoring by achieving superior clinical triage sensitivity compared to individual clinicians.
- Cutting the Cord: System Architecture for Low-Cost, GPU-Accelerated Bimanual Mobile Manipulation (viability: 3): https://sciencetostartup.com/paper/cutting-the-cord-system-architecture-for-low-cost-gpu-accelerated-bimanual-mobile-manipulation - A low-cost bimanual mobile manipulator designed for research and education in robotics.
- EPOCH: An Agentic Protocol for Multi-Round System Optimization (viability: 2): https://sciencetostartup.com/paper/epoch-an-agentic-protocol-for-multi-round-system-optimization - EPOCH is a protocol for optimizing autonomous agents through structured multi-round self-improvement.
- Beyond Amplitude: Channel State Information Phase-Aware Deep Fusion for Robotic Activity Recognition (viability: 4): https://sciencetostartup.com/paper/beyond-amplitude-channel-state-information-phase-aware-deep-fusion-for-robotic-activity-recognition - A novel deep learning approach leveraging Wi-Fi Channel State Information for enhanced robotic activity recognition.
- FlexServe: A Fast and Secure LLM Serving System for Mobile Devices with Flexible Resource Isolation (viability: 4): https://sciencetostartup.com/paper/flexserve-a-fast-and-secure-llm-serving-system-for-mobile-devices-with-flexible-resource-isolation - FlexServe is a fast and secure LLM serving system for mobile devices that optimizes resource isolation for enhanced performance.
- Synergistic Directed Execution and LLM-Driven Analysis for Zero-Day AI-Generated Malware Detection (viability: 9): https://sciencetostartup.com/paper/synergistic-directed-execution-and-llm-driven-analysis-for-zero-day-ai-generated-malware-detection - A hybrid analysis framework leveraging LLMs and deep learning to detect AI-generated malware with high accuracy.
- Time, Identity and Consciousness in Language Model Agents (viability: 2): https://sciencetostartup.com/paper/time-identity-and-consciousness-in-language-model-agents - This paper presents a theoretical framework for evaluating machine consciousness in language model agents through identity metrics.
- WS-Net: Weak-Signal Representation Learning and Gated Abundance Reconstruction for Hyperspectral Unmixing via State-Space and Weak Signal Attention Fusion (viability: 4): https://sciencetostartup.com/paper/ws-net-weak-signal-representation-learning-and-gated-abundance-reconstruction-for-hyperspectral-unmixing-via-state-space - WS-Net is a deep unmixing framework that enhances weak spectral responses in hyperspectral images through advanced attention mechanisms.
- SCALAR: Learning and Composing Skills through LLM Guided Symbolic Planning and Deep RL Grounding (viability: 3): https://sciencetostartup.com/paper/scalar-learning-and-composing-skills-through-llm-guided-symbolic-planning-and-deep-rl-grounding - SCALAR enhances LLM-guided planning by integrating deep reinforcement learning for improved skill execution.
- Two Teachers Better Than One: Hardware-Physics Co-Guided Distributed Scientific Machine Learning (viability: 7): https://sciencetostartup.com/paper/two-teachers-better-than-one-hardware-physics-co-guided-distributed-scientific-machine-learning - EPIC is a distributed SciML framework that reduces communication costs while preserving physical fidelity for real-time applications.
- ImpedanceDiffusion: Diffusion-Based Global Path Planning for UAV Swarm Navigation with Generative Impedance Control (viability: 7): https://sciencetostartup.com/paper/impedancediffusion-diffusion-based-global-path-planning-for-uav-swarm-navigation-with-generative-impedance-control - ImpedanceDiffusion enables safe and efficient swarm navigation for drones in complex indoor environments using advanced diffusion-based path planning.
- PlayWorld: Learning Robot World Models from Autonomous Play (viability: 7): https://sciencetostartup.com/paper/playworld-learning-robot-world-models-from-autonomous-play - PlayWorld is an autonomous pipeline for training high-fidelity robot world models from self-play interactions.
- Automating Detection and Root-Cause Analysis of Flaky Tests in Quantum Software (viability: 6): https://sciencetostartup.com/paper/automating-detection-and-root-cause-analysis-of-flaky-tests-in-quantum-software - An automated pipeline for detecting and diagnosing flaky tests in quantum software using LLMs.
- Lockbox -- A Zero Trust Architecture for Secure Processing of Sensitive Cloud Workloads (viability: 4): https://sciencetostartup.com/paper/lockbox-a-zero-trust-architecture-for-secure-processing-of-sensitive-cloud-workloads - Lockbox is a Zero Trust architecture for secure processing of sensitive cloud workloads, ensuring strict access controls and data protection.
- When to Retrain after Drift: A Data-Only Test of Post-Drift Data Size Sufficiency (viability: 4): https://sciencetostartup.com/paper/when-to-retrain-after-drift-a-data-only-test-of-post-drift-data-size-sufficiency - CALIPER is a data-only test that determines the optimal post-drift data size for retraining machine learning models.
- The Missing Memory Hierarchy: Demand Paging for LLM Context Windows (viability: 3): https://sciencetostartup.com/paper/the-missing-memory-hierarchy-demand-paging-for-llm-context-windows - Pichay is a demand paging system designed to optimize context window usage in large language models.
- MEMO: Memory-Augmented Model Context Optimization for Robust Multi-Turn Multi-Agent LLM Games (viability: 3): https://sciencetostartup.com/paper/memo-memory-augmented-model-context-optimization-for-robust-multi-turn-multi-agent-llm-games - MEMO optimizes context in multi-agent LLM games to improve performance and stability.
- AI Phenomenology for Understanding Human-AI Experiences Across Eras (viability: 3): https://sciencetostartup.com/paper/ai-phenomenology-for-understanding-human-ai-experiences-across-eras - AI phenomenology provides a framework for understanding the nuanced human experiences with AI systems over time.
- Meissa: Multi-modal Medical Agentic Intelligence (viability: 9): https://sciencetostartup.com/paper/meissa-multi-modal-medical-agentic-intelligence - Meissa is a lightweight, offline multi-modal medical language model that enhances clinical decision-making with agentic capabilities.
- An accurate flatness measure to estimate the generalization performance of CNN models (viability: 4): https://sciencetostartup.com/paper/an-accurate-flatness-measure-to-estimate-the-generalization-performance-of-cnn-models - A novel flatness measure for assessing the generalization performance of CNN models.
- The Coupling Within: Flow Matching via Distilled Normalizing Flows (viability: 3): https://sciencetostartup.com/paper/the-coupling-within-flow-matching-via-distilled-normalizing-flows - Introducing Normalized Flow Matching, a novel method leveraging distilled couplings from pretrained models to enhance flow model training.
- Improving through Interaction: Searching Behavioral Representation Spaces with CMA-ES-IG (viability: 8): https://sciencetostartup.com/paper/improving-through-interaction-searching-behavioral-representation-spaces-with-cma-es-ig - CMA-ES-IG enhances robot user interaction by optimizing preference learning through user-friendly behavior rankings.
- Statistical Inference via Generative Models: Flow Matching and Causal Inference (viability: 2): https://sciencetostartup.com/paper/statistical-inference-via-generative-models-flow-matching-and-causal-inference - A theoretical exploration of generative models in statistical inference.
- Gender Fairness in Audio Deepfake Detection: Performance and Disparity Analysis (viability: 4): https://sciencetostartup.com/paper/gender-fairness-in-audio-deepfake-detection-performance-and-disparity-analysis - A fairness-aware approach to improve audio deepfake detection by analyzing gender disparities.
- Security Considerations for Multi-agent Systems (viability: 2): https://sciencetostartup.com/paper/security-considerations-for-multi-agent-systems - This study evaluates security frameworks for multi-agent AI systems, highlighting vulnerabilities and gaps in existing solutions.
- Learning When to Sample: Confidence-Aware Self-Consistency for Efficient LLM Chain-of-Thought Reasoning (viability: 7): https://sciencetostartup.com/paper/learning-when-to-sample-confidence-aware-self-consistency-for-efficient-llm-chain-of-thought-reasoning - A confidence-aware framework that optimizes reasoning paths in LLMs to reduce inference costs while maintaining accuracy.
- Diffusion-Based Authentication of Copy Detection Patterns: A Multimodal Framework with Printer Signature Conditioning (viability: 7): https://sciencetostartup.com/paper/diffusion-based-authentication-of-copy-detection-patterns-a-multimodal-framework-with-printer-signature-conditioning - A diffusion-based framework for authenticating products against counterfeiting using printer signatures.
- SkipGS: Post-Densification Backward Skipping for Efficient 3DGS Training (viability: 3): https://sciencetostartup.com/paper/skipgs-post-densification-backward-skipping-for-efficient-3dgs-training - SkipGS introduces a novel mechanism to optimize training efficiency in 3D Gaussian Splatting.
- Arbiter: Detecting Interference in LLM Agent System Prompts (viability: 2): https://sciencetostartup.com/paper/arbiter-detecting-interference-in-llm-agent-system-prompts - Arbiter is a framework for detecting interference patterns in LLM system prompts to improve coding agent reliability.
- Automated Thematic Analysis for Clinical Qualitative Data: Iterative Codebook Refinement with Full Provenance (viability: 7): https://sciencetostartup.com/paper/automated-thematic-analysis-for-clinical-qualitative-data-iterative-codebook-refinement-with-full-provenance - An automated thematic analysis framework that enhances the scalability and reproducibility of qualitative data analysis in health research.
- Characterization, Analytical Planning, and Hybrid Force Control for the Inspire RH56DFX Hand (viability: 8): https://sciencetostartup.com/paper/characterization-analytical-planning-and-hybrid-force-control-for-the-inspire-rh56dfx-hand - Transform the Inspire RH56DFX hand into a reliable research tool for dexterous manipulation with enhanced control and planning capabilities.
- MAPLE: Elevating Medical Reasoning from Statistical Consensus to Process-Led Alignment (viability: 7): https://sciencetostartup.com/paper/maple-elevating-medical-reasoning-from-statistical-consensus-to-process-led-alignment - A novel training paradigm for medical AI that enhances reasoning through expert-aligned reinforcement learning.
- SurgCalib: Gaussian Splatting-Based Hand-Eye Calibration for Robot-Assisted Minimally Invasive Surgery (viability: 7): https://sciencetostartup.com/paper/surgcalib-gaussian-splatting-based-hand-eye-calibration-for-robot-assisted-minimally-invasive-surgery - SurgCalib offers a markerless hand-eye calibration solution for robotic surgery, enhancing precision without compromising sterility.
- SVG-EAR: Parameter-Free Linear Compensation for Sparse Video Generation via Error-aware Routing (viability: 3): https://sciencetostartup.com/paper/svg-ear-parameter-free-linear-compensation-for-sparse-video-generation-via-error-aware-routing - SVG-EAR introduces a parameter-free method for improving efficiency in video generation by compensating for skipped attention blocks.
- MAcPNN: Mutual Assisted Learning on Data Streams with Temporal Dependence (viability: 3): https://sciencetostartup.com/paper/macpnn-mutual-assisted-learning-on-data-streams-with-temporal-dependence - MAcPNN enhances IoT device learning by enabling autonomous knowledge sharing to adapt to data stream changes.
- Can You Hear, Localize, and Segment Continually? An Exemplar-Free Continual Learning Benchmark for Audio-Visual Segmentation (viability: 7): https://sciencetostartup.com/paper/can-you-hear-localize-and-segment-continually-an-exemplar-free-continual-learning-benchmark-for-audio-visual-segmentatio - A benchmark and baseline for exemplar-free continual learning in audio-visual segmentation.
- Semantic Level of Detail: Multi-Scale Knowledge Representation via Heat Kernel Diffusion on Hyperbolic Manifolds (viability: 4): https://sciencetostartup.com/paper/semantic-level-of-detail-multi-scale-knowledge-representation-via-heat-kernel-diffusion-on-hyperbolic-manifolds - A framework for continuous resolution control in AI memory systems using heat kernel diffusion on hyperbolic manifolds.
- The FABRIC Strategy for Verifying Neural Feedback Systems (viability: 7): https://sciencetostartup.com/paper/the-fabric-strategy-for-verifying-neural-feedback-systems - FaBRIC offers a novel approach to verifying neural feedback systems through advanced reachability analysis.
- Estimation of heterogeneous principal effects under principal ignorability (viability: 2): https://sciencetostartup.com/paper/estimation-of-heterogeneous-principal-effects-under-principal-ignorability - A framework for estimating heterogeneous principal causal effects in statistical models.
- FAME: Force-Adaptive RL for Expanding the Manipulation Envelope of a Full-Scale Humanoid (viability: 9): https://sciencetostartup.com/paper/fame-force-adaptive-rl-for-expanding-the-manipulation-envelope-of-a-full-scale-humanoid - FAME is a force-adaptive reinforcement learning framework that enhances humanoid manipulation by adapting to external forces in real-time.
- The $qs$ Inequality: Quantifying the Double Penalty of Mixture-of-Experts at Inference (viability: 4): https://sciencetostartup.com/paper/the-qs-inequality-quantifying-the-double-penalty-of-mixture-of-experts-at-inference - This research quantifies the inefficiencies of Mixture-of-Experts models at inference, proposing a new criterion for evaluating their performance against dense models.
- Formation-Aware Adaptive Conformalized Perception for Safe Leader-Follower Multi-Robot Systems (viability: 7): https://sciencetostartup.com/paper/formation-aware-adaptive-conformalized-perception-for-safe-leader-follower-multi-robot-systems - A safety-focused adaptive perception method for leader-follower multi-robot systems that enhances tracking accuracy and formation success rates.
- Automated Tensor-Relational Decomposition for Large-Scale Sparse Tensor Computation (viability: 2): https://sciencetostartup.com/paper/automated-tensor-relational-decomposition-for-large-scale-sparse-tensor-computation - A novel tensor-relational computation method that optimizes large-scale sparse tensor operations.
- A Consensus-Driven Multi-LLM Pipeline for Missing-Person Investigations (viability: 7): https://sciencetostartup.com/paper/a-consensus-driven-multi-llm-pipeline-for-missing-person-investigations - Guardian is a multi-LLM pipeline designed to enhance missing-person investigations through intelligent information extraction and processing.
- Towards Reliable Simulation-based Inference (viability: 2): https://sciencetostartup.com/paper/towards-reliable-simulation-based-inference - This thesis explores the use of machine learning for statistical analyses in scientific simulation, focusing on reducing overconfidence in model predictions.
- BiCLIP: Domain Canonicalization via Structured Geometric Transformation (viability: 8): https://sciencetostartup.com/paper/biclip-domain-canonicalization-via-structured-geometric-transformation - BiCLIP enhances cross-modal alignment in vision-language models through structured geometric transformations for specialized domain adaptation.
- AgentOS: From Application Silos to a Natural Language-Driven Data Ecosystem (viability: 5): https://sciencetostartup.com/paper/agentos-from-application-silos-to-a-natural-language-driven-data-ecosystem - AgentOS transforms traditional operating systems into a natural language-driven ecosystem for intelligent agents.
- VoxEmo: Benchmarking Speech Emotion Recognition with Speech LLMs (viability: 4): https://sciencetostartup.com/paper/voxemo-benchmarking-speech-emotion-recognition-with-speech-llms - VoxEmo benchmarks speech emotion recognition using a comprehensive toolkit for evaluating speech LLMs across multiple languages.
- PathoScribe: Transforming Pathology Data into a Living Library with a Unified LLM-Driven Framework for Semantic Retrieval and Clinical Integration (viability: 9): https://sciencetostartup.com/paper/pathoscribe-transforming-pathology-data-into-a-living-library-with-a-unified-llm-driven-framework-for-semantic-retrieval - PathoScribe transforms static pathology archives into an interactive, LLM-driven living library for enhanced clinical decision-making.
- Interpretable Markov-Based Spatiotemporal Risk Surfaces for Missing-Child Search Planning with Reinforcement Learning and LLM-Based Quality Assurance (viability: 3): https://sciencetostartup.com/paper/interpretable-markov-based-spatiotemporal-risk-surfaces-for-missing-child-search-planning-with-reinforcement-learning-an - Guardian is a decision-support system that enhances missing-child investigations through spatiotemporal predictive modeling.
- Optimizing Reinforcement Learning Training over Digital Twin Enabled Multi-fidelity Networks (viability: 3): https://sciencetostartup.com/paper/optimizing-reinforcement-learning-training-over-digital-twin-enabled-multi-fidelity-networks - A novel framework for optimizing antenna tilt angles in mobile networks using reinforcement learning and digital twin technology.
- Using Vision Language Foundation Models to Generate Plant Simulation Configurations via In-Context Learning (viability: 7): https://sciencetostartup.com/paper/using-vision-language-foundation-models-to-generate-plant-simulation-configurations-via-in-context-learning - A novel approach using vision language models to generate plant simulation configurations from drone imagery for agricultural digital twins.
- TIDE: Text-Informed Dynamic Extrapolation with Step-Aware Temperature Control for Diffusion Transformers (viability: 7): https://sciencetostartup.com/paper/tide-text-informed-dynamic-extrapolation-with-step-aware-temperature-control-for-diffusion-transformers - TIDE is a training-free method for high-resolution text-to-image generation that preserves semantic details and reduces artifacts.
- MEGC2026: Micro-Expression Grand Challenge on Visual Question Answering (viability: 4): https://sciencetostartup.com/paper/megc2026-micro-expression-grand-challenge-on-visual-question-answering - A challenge to advance facial micro-expression analysis through visual question answering using multimodal models.
- Fly, Track, Land: Infrastructure-less Magnetic Localization for Heterogeneous UAV-UGV Teaming (viability: 8): https://sciencetostartup.com/paper/fly-track-land-infrastructure-less-magnetic-localization-for-heterogeneous-uav-ugv-teaming - An infrastructure-less magnetic localization system for precise UAV-UGV docking and collaboration.
- Quantifying Uncertainty in AI Visibility: A Statistical Framework for Generative Search Measurement (viability: 4): https://sciencetostartup.com/paper/quantifying-uncertainty-in-ai-visibility-a-statistical-framework-for-generative-search-measurement - A statistical framework for measuring uncertainty in generative search citation visibility.
- Vision-Language Models Encode Clinical Guidelines for Concept-Based Medical Reasoning (viability: 7): https://sciencetostartup.com/paper/vision-language-models-encode-clinical-guidelines-for-concept-based-medical-reasoning - MedCBR integrates clinical guidelines with vision-language models for enhanced interpretability in medical imaging.
- Uncovering a Winning Lottery Ticket with Continuously Relaxed Bernoulli Gates (viability: 4): https://sciencetostartup.com/paper/uncovering-a-winning-lottery-ticket-with-continuously-relaxed-bernoulli-gates - A novel approach to discovering sparse subnetworks in neural networks using differentiable optimization techniques.
- Quantifying Memorization and Privacy Risks in Genomic Language Models (viability: 5): https://sciencetostartup.com/paper/quantifying-memorization-and-privacy-risks-in-genomic-language-models - A framework for quantifying memorization and privacy risks in genomic language models to ensure data security.
- FedLECC: Cluster- and Loss-Guided Client Selection for Federated Learning under Non-IID Data (viability: 6): https://sciencetostartup.com/paper/fedlecc-cluster-and-loss-guided-client-selection-for-federated-learning-under-non-iid-data - FedLECC enhances client selection in Federated Learning to improve model accuracy and reduce communication overhead.
- SciTaRC: Benchmarking QA on Scientific Tabular Data that Requires Language Reasoning and Complex Computation (viability: 4): https://sciencetostartup.com/paper/scitarc-benchmarking-qa-on-scientific-tabular-data-that-requires-language-reasoning-and-complex-computation - SciTaRC benchmarks complex QA tasks on scientific tabular data, revealing significant gaps in current AI models' capabilities.
- Cross-Domain Uncertainty Quantification for Selective Prediction: A Comprehensive Bound Ablation with Transfer-Informed Betting (viability: 4): https://sciencetostartup.com/paper/cross-domain-uncertainty-quantification-for-selective-prediction-a-comprehensive-bound-ablation-with-transfer-informed-b - A novel approach to selective prediction using Transfer-Informed Betting for improved risk control in data-scarce environments.
- Multi-Kernel Gated Decoder Adapters for Robust Multi-Task Thyroid Ultrasound under Cross-Center Shift (viability: 7): https://sciencetostartup.com/paper/multi-kernel-gated-decoder-adapters-for-robust-multi-task-thyroid-ultrasound-under-cross-center-shift - A novel lightweight adapter for enhancing thyroid ultrasound automation by improving robustness against cross-center domain shifts.
- Proprioceptive Safe Active Navigation and Exploration for Planetary Environments (viability: 3): https://sciencetostartup.com/paper/proprioceptive-safe-active-navigation-and-exploration-for-planetary-environments - PSANE is a framework for safe navigation and exploration in unknown deformable terrains using leg-terrain interaction data.
- NetDiffuser: Deceiving DNN-Based Network Attack Detection Systems with Diffusion-Generated Adversarial Traffic (viability: 7): https://sciencetostartup.com/paper/netdiffuser-deceiving-dnn-based-network-attack-detection-systems-with-diffusion-generated-adversarial-traffic - NetDiffuser generates natural adversarial examples to enhance the security of deep learning-based network intrusion detection systems.
- A New Modeling to Feature Selection Based on the Fuzzy Rough Set Theory in Normal and Optimistic States on Hybrid Information Systems (viability: 4): https://sciencetostartup.com/paper/a-new-modeling-to-feature-selection-based-on-the-fuzzy-rough-set-theory-in-normal-and-optimistic-states-on-hybrid-inform - FSbuHD is a novel feature selection model leveraging fuzzy rough set theory to optimize decision-making in hybrid information systems.
- ConFu: Contemplate the Future for Better Speculative Sampling (viability: 7): https://sciencetostartup.com/paper/confu-contemplate-the-future-for-better-speculative-sampling - ConFu enhances speculative decoding for LLMs by enabling draft models to anticipate future token generation, improving speed and accuracy.
- Towards Visual Query Segmentation in the Wild (viability: 7): https://sciencetostartup.com/paper/towards-visual-query-segmentation-in-the-wild - A novel visual query segmentation method that enables precise pixel-level localization of objects in untrimmed videos.
- Comparative Analysis of Patch Attack on VLM-Based Autonomous Driving Architectures (viability: 4): https://sciencetostartup.com/paper/comparative-analysis-of-patch-attack-on-vlm-based-autonomous-driving-architectures - A framework for evaluating the robustness of vision-language models against adversarial attacks in autonomous driving.
- MultiGraSCCo: A Multilingual Anonymization Benchmark with Annotations of Personal Identifiers (viability: 5): https://sciencetostartup.com/paper/multigrascco-a-multilingual-anonymization-benchmark-with-annotations-of-personal-identifiers - A multilingual benchmark for anonymizing personal identifiers in sensitive data, enabling safe data sharing.
- Quantifying the Accuracy and Cost Impact of Design Decisions in Budget-Constrained Agentic LLM Search (viability: 7): https://sciencetostartup.com/paper/quantifying-the-accuracy-and-cost-impact-of-design-decisions-in-budget-constrained-agentic-llm-search - A model-agnostic evaluation framework for optimizing budget-constrained agentic retrieval systems in LLMs.
- One Language, Two Scripts: Probing Script-Invariance in LLM Concept Representations (viability: 2): https://sciencetostartup.com/paper/one-language-two-scripts-probing-script-invariance-in-llm-concept-representations - This research explores the abstract meaning representation in Sparse Autoencoders using Serbian digraphia.
- Why Channel-Centric Models are not Enough to Predict End-to-End Performance in Private 5G: A Measurement Campaign and Case Study (viability: 4): https://sciencetostartup.com/paper/why-channel-centric-models-are-not-enough-to-predict-end-to-end-performance-in-private-5g-a-measurement-campaign-and-cas - A data-driven approach to accurately predict end-to-end throughput in private 5G networks for communication-aware robot planning.
- Adaptive SINDy: Residual Force System Identification Based UAV Disturbance Rejection (viability: 3): https://sciencetostartup.com/paper/adaptive-sindy-residual-force-system-identification-based-uav-disturbance-rejection - A novel control algorithm for UAVs that integrates data-driven system identification to reject wind disturbances.
- APPLV: Adaptive Planner Parameter Learning from Vision-Language-Action Model (viability: 7): https://sciencetostartup.com/paper/applv-adaptive-planner-parameter-learning-from-vision-language-action-model - APPLV enhances mobile robot navigation by automating parameter tuning using Vision-Language-Action models.
- SEP-NMPC: Safety Enhanced Passivity-Based Nonlinear Model Predictive Control for a UAV Slung Payload System (viability: 3): https://sciencetostartup.com/paper/sep-nmpc-safety-enhanced-passivity-based-nonlinear-model-predictive-control-for-a-uav-slung-payload-system - A novel control framework ensuring stability and safety for UAVs transporting slung payloads in cluttered environments.
- Expressivity-Efficiency Tradeoffs for Hybrid Sequence Models (viability: 7): https://sciencetostartup.com/paper/expressivity-efficiency-tradeoffs-for-hybrid-sequence-models - Hybrid sequence models combine Transformer and state-space layers for enhanced efficiency and expressiveness.
- Unpacking Interpretability: Human-Centered Criteria for Optimal Combinatorial Solutions (viability: 2): https://sciencetostartup.com/paper/unpacking-interpretability-human-centered-criteria-for-optimal-combinatorial-solutions - This research explores human-centered criteria for selecting interpretable optimal solutions in algorithmic support systems.
- DeZent: Decentralized z-Anonymity with Privacy-Preserving Coordination (viability: 2): https://sciencetostartup.com/paper/dezent-decentralized-z-anonymity-with-privacy-preserving-coordination - deZent offers a decentralized approach to z-anonymity for privacy in sensor networks.
- LDP: An Identity-Aware Protocol for Multi-Agent LLM Systems (viability: 3): https://sciencetostartup.com/paper/ldp-an-identity-aware-protocol-for-multi-agent-llm-systems - LDP introduces an identity-aware communication protocol for enhancing multi-agent AI system efficiency.
- HECTOR: Hybrid Editable Compositional Object References for Video Generation (viability: 7): https://sciencetostartup.com/paper/hector-hybrid-editable-compositional-object-references-for-video-generation - HECTOR enables fine-grained compositional control in video generation through hybrid reference conditioning.
- A Lightweight Multi-Cancer Tumor Localization Framework for Deployable Digital Pathology (viability: 8): https://sciencetostartup.com/paper/a-lightweight-multi-cancer-tumor-localization-framework-for-deployable-digital-pathology - A robust multi-cancer tumor localization framework that enhances digital pathology workflows.
- MASEval: Extending Multi-Agent Evaluation from Models to Systems (viability: 7): https://sciencetostartup.com/paper/maseval-extending-multi-agent-evaluation-from-models-to-systems - MASEval is a framework-agnostic library for evaluating multi-agent systems, focusing on implementation decisions that impact performance.
- HeteroFedSyn: Differentially Private Tabular Data Synthesis for Heterogeneous Federated Settings (viability: 8): https://sciencetostartup.com/paper/heterofedsyn-differentially-private-tabular-data-synthesis-for-heterogeneous-federated-settings - HeteroFedSyn is a framework for differentially private tabular data synthesis in heterogeneous federated settings, enabling secure data sharing for various tasks.
- Predictive Control with Indirect Adaptive Laws for Payload Transportation by Quadrupedal Robots (viability: 3): https://sciencetostartup.com/paper/predictive-control-with-indirect-adaptive-laws-for-payload-transportation-by-quadrupedal-robots - A novel hierarchical planning and control framework for robust payload transportation by quadrupedal robots.
- Computer Vision-Based Vehicle Allotment System using Perspective Mapping (viability: 3): https://sciencetostartup.com/paper/computer-vision-based-vehicle-allotment-system-using-perspective-mapping - A smart parking system leveraging computer vision for efficient vehicle allotment in urban areas.
- Are Expressive Encoders Necessary for Discrete Graph Generation? (viability: 7): https://sciencetostartup.com/paper/are-expressive-encoders-necessary-for-discrete-graph-generation - GenGNN offers a modular framework for efficient discrete graph generation with high validity rates.
- SoftJAX & SoftTorch: Empowering Automatic Differentiation Libraries with Informative Gradients (viability: 9): https://sciencetostartup.com/paper/softjax-softtorch-empowering-automatic-differentiation-libraries-with-informative-gradients - SoftJAX and SoftTorch provide open-source libraries for soft differentiable programming, enhancing automatic differentiation with informative gradients.
- Fish Audio S2 Technical Report (viability: 9): https://sciencetostartup.com/paper/fish-audio-s2-technical-report - Fish Audio S2 is an open-sourced text-to-speech system that enables multi-speaker, instruction-following audio generation.
- Impact of Different Failures on a Robot's Perceived Reliability (viability: 2): https://sciencetostartup.com/paper/impact-of-different-failures-on-a-robot-s-perceived-reliability - This study explores how different robot failures impact perceived reliability and trust in human-robot interactions.
- Beyond Relevance: On the Relationship Between Retrieval and RAG Information Coverage (viability: 4): https://sciencetostartup.com/paper/beyond-relevance-on-the-relationship-between-retrieval-and-rag-information-coverage - This research explores the correlation between retrieval quality and generation effectiveness in RAG systems, providing insights for improving performance metrics.
- HMR-1: Hierarchical Massage Robot with Vision-Language-Model for Embodied Healthcare (viability: 8): https://sciencetostartup.com/paper/hmr-1-hierarchical-massage-robot-with-vision-language-model-for-embodied-healthcare - A hierarchical massage robot leveraging vision-language models to enhance physical therapy and rehabilitation.
- Scale-Plan: Scalable Language-Enabled Task Planning for Heterogeneous Multi-Robot Teams (viability: 7): https://sciencetostartup.com/paper/scale-plan-scalable-language-enabled-task-planning-for-heterogeneous-multi-robot-teams - Scale-Plan is a scalable LLM-assisted framework for efficient task planning in heterogeneous multi-robot teams.
- VisionCreator-R1: A Reflection-Enhanced Native Visual-Generation Agentic Model (viability: 7): https://sciencetostartup.com/paper/visioncreator-r1-a-reflection-enhanced-native-visual-generation-agentic-model - VisionCreator-R1 enhances visual content generation through a reflection-driven optimization approach.
- Age-Related Differences in the Perception of Eye-Gaze from a Social Robot (viability: 2): https://sciencetostartup.com/paper/age-related-differences-in-the-perception-of-eye-gaze-from-a-social-robot - Exploring the impact of age on social perception of gaze cues in social robots for older adults.
- Where, What, Why: Toward Explainable 3D-GS Watermarking (viability: 3): https://sciencetostartup.com/paper/where-what-why-toward-explainable-3d-gs-watermarking - A framework for robust and explainable watermarking in 3D Gaussian Splatting.
- Test-Driven AI Agent Definition (TDAD): Compiling Tool-Using Agents from Behavioral Specifications (viability: 8): https://sciencetostartup.com/paper/test-driven-ai-agent-definition-tdad-compiling-tool-using-agents-from-behavioral-specifications - TDAD is a methodology for developing tool-using AI agents that ensures behavioral compliance through test-driven specifications.
- The Temporal Markov Transition Field (viability: 2): https://sciencetostartup.com/paper/the-temporal-markov-transition-field - The Temporal Markov Transition Field enhances time series analysis by capturing local dynamics through segmented transition matrices.
- Granulon: Awakening Pixel-Level Visual Encoders with Adaptive Multi-Granularity Semantics for MLLM (viability: 7): https://sciencetostartup.com/paper/granulon-awakening-pixel-level-visual-encoders-with-adaptive-multi-granularity-semantics-for-mllm - Granulon enhances visual understanding in multimodal large language models through adaptive granularity control.
- Scale Space Diffusion (viability: 7): https://sciencetostartup.com/paper/scale-space-diffusion - Scale Space Diffusion optimizes diffusion models by processing noisy images at lower resolutions, potentially speeding up image generation.
- Multi-level meta-reinforcement learning with skill-based curriculum (viability: 7): https://sciencetostartup.com/paper/multi-level-meta-reinforcement-learning-with-skill-based-curriculum - A multi-level meta-reinforcement learning framework with skill-based curriculum for efficient sequential decision making.
- FVG-PT: Adaptive Foreground View-Guided Prompt Tuning for Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/fvg-pt-adaptive-foreground-view-guided-prompt-tuning-for-vision-language-models - Adapt pretrained vision-language models to downstream tasks by adaptively guiding foreground attention, improving prediction accuracy and generalization.
- Impermanent: A Live Benchmark for Temporal Generalization in Time Series Forecasting (viability: 8): https://sciencetostartup.com/paper/impermanent-a-live-benchmark-for-temporal-generalization-in-time-series-forecasting - Impermanent provides a live benchmark and dashboard for evaluating time-series forecasting models, enabling real-time performance monitoring and temporal generalization assessment.
- Agentic Critical Training (viability: 7): https://sciencetostartup.com/paper/agentic-critical-training - Agentic Critical Training (ACT) enhances LLM agent performance by training them to autonomously identify better actions through reinforcement learning, leading to improved reasoning and generalization.
- Evaluating Financial Intelligence in Large Language Models: Benchmarking SuperInvesting AI with LLM Engines (viability: 7): https://sciencetostartup.com/paper/evaluating-financial-intelligence-in-large-language-models-benchmarking-superinvesting-ai-with-llm-engines - Benchmark and improve LLMs for financial analysis with a structured evaluation framework and a high-performing SuperInvesting model.
- HiAR: Efficient Autoregressive Long Video Generation via Hierarchical Denoising (viability: 7): https://sciencetostartup.com/paper/hiar-efficient-autoregressive-long-video-generation-via-hierarchical-denoising - HiAR enables efficient autoregressive long video generation by conditioning on context at the same noise level, offering a potential API for long-form video creation.
- A Multi-Objective Optimization Approach for Sustainable AI-Driven Entrepreneurship in Resilient Economies (viability: 8): https://sciencetostartup.com/paper/a-multi-objective-optimization-approach-for-sustainable-ai-driven-entrepreneurship-in-resilient-economies - EcoAI-Resilience framework optimizes AI deployment for sustainability, economic resilience, and environmental cost minimization, offering a data-driven approach for responsible AI entrepreneurship.
- Split Federated Learning Architectures for High-Accuracy and Low-Delay Model Training (viability: 7): https://sciencetostartup.com/paper/split-federated-learning-architectures-for-high-accuracy-and-low-delay-model-training - Optimize split federated learning architectures for improved accuracy, reduced delay, and lower communication overhead.
- Benchmarking Language Modeling for Lossless Compression of Full-Fidelity Audio (viability: 7): https://sciencetostartup.com/paper/benchmarking-language-modeling-for-lossless-compression-of-full-fidelity-audio - Trilobyte enables lossless audio compression for high-fidelity audio using language models, outperforming FLAC at 8-bit and 16-bit depths.
- Structural Causal Bottleneck Models (viability: 6): https://sciencetostartup.com/paper/structural-causal-bottleneck-models - SCBMs offer a novel approach to causal inference by identifying low-dimensional bottlenecks, enabling efficient effect estimation and task-specific dimension reduction.
- A New Lower Bound for the Random Offerer Mechanism in Bilateral Trade using AI-Guided Evolutionary Search (viability: 7): https://sciencetostartup.com/paper/a-new-lower-bound-for-the-random-offerer-mechanism-in-bilateral-trade-using-ai-guided-evolutionary-search - AI-guided mechanism design tool to identify inefficiencies in market mechanisms and improve outcomes.
- Momentum SVGD-EM for Accelerated Maximum Marginal Likelihood Estimation (viability: 7): https://sciencetostartup.com/paper/momentum-svgd-em-for-accelerated-maximum-marginal-likelihood-estimation - Accelerated maximum marginal likelihood estimation using Stein Variational Gradient Descent for faster convergence in probabilistic models.
- Talking Together: Synthesizing Co-Located 3D Conversations from Audio (viability: 7): https://sciencetostartup.com/paper/talking-together-synthesizing-co-located-3d-conversations-from-audio - Generate realistic, spatially-aware 3D facial animations of two interacting people from a mixed audio stream, enabling immersive VR/telepresence experiences.
- Exp-Force: Experience-Conditioned Pre-Grasp Force Selection with Vision-Language Models (viability: 8): https://sciencetostartup.com/paper/exp-force-experience-conditioned-pre-grasp-force-selection-with-vision-language-models - Exp-Force uses a vision-language model conditioned on prior grasping experiences to predict the minimum feasible grasping force, enabling safer and more reliable robotic manipulation.
- Cybersecurity AI: Hacking Consumer Robots in the AI Era (viability: 8): https://sciencetostartup.com/paper/cybersecurity-ai-hacking-consumer-robots-in-the-ai-era - Democratizing robot cybersecurity assessments with an AI-powered vulnerability scanner that automates penetration testing for consumer robots.
- ImprovedGS+: A High-Performance C++/CUDA Re-Implementation Strategy for 3D Gaussian Splatting (viability: 7): https://sciencetostartup.com/paper/improvedgs-a-high-performance-c-cuda-re-implementation-strategy-for-3d-gaussian-splatting - ImprovedGS+ accelerates 3D Gaussian Splatting training with optimized C++/CUDA kernels, offering faster scene reconstruction and improved visual quality.
- How Far Can Unsupervised RLVR Scale LLM Training? (viability: 6): https://sciencetostartup.com/paper/how-far-can-unsupervised-rlvr-scale-llm-training - This paper analyzes the limitations of unsupervised reinforcement learning for LLM training and proposes a metric to predict RL trainability, suggesting potential for test-time training applications.
- CODA: Difficulty-Aware Compute Allocation for Adaptive Reasoning (viability: 7): https://sciencetostartup.com/paper/coda-difficulty-aware-compute-allocation-for-adaptive-reasoning - CODA dynamically adjusts reasoning depth based on problem difficulty, reducing compute costs on easy tasks and improving accuracy on hard tasks, making it a cost-effective reasoning solution.
- OfficeQA Pro: An Enterprise Benchmark for End-to-End Grounded Reasoning (viability: 7): https://sciencetostartup.com/paper/officeqa-pro-an-enterprise-benchmark-for-end-to-end-grounded-reasoning - OfficeQA Pro is a new benchmark for evaluating AI agents on grounded, multi-document reasoning over a large and heterogeneous document corpus, highlighting the need for improved enterprise-grade reasoning capabilities.
- CoCo: Code as CoT for Text-to-Image Preview and Rare Concept Generation (viability: 8): https://sciencetostartup.com/paper/coco-code-as-cot-for-text-to-image-preview-and-rare-concept-generation - CoCo is a code-driven text-to-image generation framework that uses executable code for precise and controllable image creation, offering significant improvements over existing methods.
- CAST: Modeling Visual State Transitions for Consistent Video Retrieval (viability: 7): https://sciencetostartup.com/paper/cast-modeling-visual-state-transitions-for-consistent-video-retrieval - CAST is a plug-and-play adapter that improves the temporal coherence of video retrieval and generation by modeling visual state transitions.
- Grow, Don't Overwrite: Fine-tuning Without Forgetting (viability: 7): https://sciencetostartup.com/paper/grow-don-t-overwrite-fine-tuning-without-forgetting - A novel function-preserving expansion method for fine-tuning pre-trained models that eliminates catastrophic forgetting and reduces computational cost.
- Retrieval-Augmented Gaussian Avatars: Improving Expression Generalization (viability: 7): https://sciencetostartup.com/paper/retrieval-augmented-gaussian-avatars-improving-expression-generalization - Retrieval-Augmented Faces (RAF) improves the expression fidelity of animatable head avatars by augmenting training data with nearest-neighbor expressions from a large unlabeled expression bank.
- PostTrainBench: Can LLM Agents Automate LLM Post-Training? (viability: 7): https://sciencetostartup.com/paper/posttrainbench-can-llm-agents-automate-llm-post-training - Automated LLM post-training benchmark and agent-based optimization tool to improve base LLM performance.
- UNBOX: Unveiling Black-box visual models with Natural-language (viability: 7): https://sciencetostartup.com/paper/unbox-unveiling-black-box-visual-models-with-natural-language - UNBOX provides a framework for understanding black-box visual models using natural language, enabling auditing and bias detection without requiring internal access.
- StreamReady: Learning What to Answer and When in Long Streaming Videos (viability: 7): https://sciencetostartup.com/paper/streamready-learning-what-to-answer-and-when-in-long-streaming-videos - StreamReady is a framework for real-time video understanding that answers questions at the optimal moment, balancing accuracy and timeliness.
- Embedding Classical Balance Control Principles in Reinforcement Learning for Humanoid Recovery (viability: 7): https://sciencetostartup.com/paper/embedding-classical-balance-control-principles-in-reinforcement-learning-for-humanoid-recovery - A reinforcement learning policy that embeds classical balance metrics for robust humanoid robot recovery, enabling zero-shot hardware transfer and high recovery rates.
- Diff-Muscle: Efficient Learning for Musculoskeletal Robotic Table Tennis (viability: 7): https://sciencetostartup.com/paper/diff-muscle-efficient-learning-for-musculoskeletal-robotic-table-tennis - Diff-Muscle enables musculoskeletal robots to perform complex tasks like table tennis by reformulating policy learning in a lower-dimensional joint space, demonstrating superior performance and efficient muscle activation.
- Coverage-Guided Multi-Agent Harness Generation for Java Library Fuzzing (viability: 8): https://sciencetostartup.com/paper/coverage-guided-multi-agent-harness-generation-for-java-library-fuzzing - Automated fuzz harness generation for Java libraries using LLM-powered agents, improving coverage and bug discovery.
- Weakly Supervised Teacher-Student Framework with Progressive Pseudo-mask Refinement for Gland Segmentation (viability: 7): https://sciencetostartup.com/paper/weakly-supervised-teacher-student-framework-with-progressive-pseudo-mask-refinement-for-gland-segmentation - An annotation-efficient AI tool for gland segmentation in colorectal histopathology, enabling more accurate cancer grading.
- Don't Look Back in Anger: MAGIC Net for Streaming Continual Learning with Temporal Dependence (viability: 3): https://sciencetostartup.com/paper/don-t-look-back-in-anger-magic-net-for-streaming-continual-learning-with-temporal-dependence - Develop MAGIC Net, a streaming continual learning framework to handle concept drift and catastrophic forgetting using recurrent neural networks.
- Bilevel Planning with Learned Symbolic Abstractions from Interaction Data (viability: 7): https://sciencetostartup.com/paper/bilevel-planning-with-learned-symbolic-abstractions-from-interaction-data - A bilevel neuro-symbolic framework for robot planning that combines learned probabilistic symbolic rules with continuous effect models for efficient and reliable plan generation and verification.
- Boosting MLLM Spatial Reasoning with Geometrically Referenced 3D Scene Representations (viability: 7): https://sciencetostartup.com/paper/boosting-mllm-spatial-reasoning-with-geometrically-referenced-3d-scene-representations - Enhance MLLM spatial reasoning by encoding 3D geometric attributes as textual references, enabling zero-shot application and improved performance on spatial understanding tasks.
- CARE-Edit: Condition-Aware Routing of Experts for Contextual Image Editing (viability: 7): https://sciencetostartup.com/paper/care-edit-condition-aware-routing-of-experts-for-contextual-image-editing - CARE-Edit dynamically routes diffusion tokens to specialized experts based on multi-modal conditions, enabling precise and coherent contextual image editing.
- Towards Batch-to-Streaming Deep Reinforcement Learning for Continuous Control (viability: 7): https://sciencetostartup.com/paper/towards-batch-to-streaming-deep-reinforcement-learning-for-continuous-control - Enable on-device finetuning of reinforcement learning models with streaming updates, bridging the gap between batch and online learning.
- DualFlexKAN: Dual-stage Kolmogorov-Arnold Networks with Independent Function Control (viability: 7): https://sciencetostartup.com/paper/dualflexkan-dual-stage-kolmogorov-arnold-networks-with-independent-function-control - DualFlexKAN offers a more efficient and flexible alternative to MLPs and KANs by decoupling input transformations and output activations, enabling hybrid networks with improved accuracy and reduced parameter count.
- SmartGraphical: A Human-in-the-Loop Framework for Detecting Smart Contract Logical Vulnerabilities via Pattern-Driven Static Analysis and Visual Abstraction (viability: 7): https://sciencetostartup.com/paper/smartgraphical-a-human-in-the-loop-framework-for-detecting-smart-contract-logical-vulnerabilities-via-pattern-driven-sta - SmartGraphical is a human-in-the-loop security framework that helps developers identify logical vulnerabilities in smart contracts through static analysis and visual representation.
- Trust via Reputation of Conviction (viability: 2): https://sciencetostartup.com/paper/trust-via-reputation-of-conviction - A theoretical framework for trust based on conviction and reputation in AI agents.
- MetaWorld-X: Hierarchical World Modeling via VLM-Orchestrated Experts for Humanoid Loco-Manipulation (viability: 7): https://sciencetostartup.com/paper/metaworld-x-hierarchical-world-modeling-via-vlm-orchestrated-experts-for-humanoid-loco-manipulation - MetaWorld-X is a hierarchical world model framework for humanoid control that uses VLM-orchestrated expert policies for natural and generalizable loco-manipulation.
- OSS-CRS: Liberating AIxCC Cyber Reasoning Systems for Real-World Open-Source Security (viability: 8): https://sciencetostartup.com/paper/oss-crs-liberating-aixcc-cyber-reasoning-systems-for-real-world-open-source-security - OSS-CRS is an open, locally deployable framework for running and combining AI-based cyber reasoning techniques against real-world open-source projects, enabling autonomous bug confirmation and patching.
- BioGait-VLM: A Tri-Modal Vision-Language-Biomechanics Framework for Interpretable Clinical Gait Assessment (viability: 8): https://sciencetostartup.com/paper/biogait-vlm-a-tri-modal-vision-language-biomechanics-framework-for-interpretable-clinical-gait-assessment - BioGait-VLM offers interpretable clinical gait assessment by incorporating biomechanics into vision-language models, achieving state-of-the-art accuracy and clinical plausibility.
- RetroAgent: From Solving to Evolving via Retrospective Dual Intrinsic Feedback (viability: 8): https://sciencetostartup.com/paper/retroagent-from-solving-to-evolving-via-retrospective-dual-intrinsic-feedback - RetroAgent is an online RL framework that enables LLM-based agents to continuously adapt and improve in complex interactive environments by using hindsight self-reflection and dual intrinsic feedback, achieving state-of-the-art results.
- mmGAT: Pose Estimation by Graph Attention with Mutual Features from mmWave Radar Point Cloud (viability: 8): https://sciencetostartup.com/paper/mmgat-pose-estimation-by-graph-attention-with-mutual-features-from-mmwave-radar-point-cloud - mmGAT leverages mmWave radar and Graph Neural Networks to achieve state-of-the-art human pose estimation, offering a privacy-preserving and light-insensitive alternative to image-based methods.
- Interactive World Simulator for Robot Policy Training and Evaluation (viability: 8): https://sciencetostartup.com/paper/interactive-world-simulator-for-robot-policy-training-and-evaluation - Interactive World Simulator enables scalable robot policy training and evaluation by generating realistic, interaction-consistent simulations, allowing for faster and cheaper development of robotic solutions.
- The Neural Compass: Probabilistic Relative Feature Fields for Robotic Search (viability: 7): https://sciencetostartup.com/paper/the-neural-compass-probabilistic-relative-feature-fields-for-robotic-search - ProReFF leverages pre-trained vision language models to guide robotic object search by learning object co-occurrences from unlabeled data, improving search efficiency by 20% over baselines.
- EquiBim: Learning Symmetry-Equivariant Policy for Bimanual Manipulation (viability: 7): https://sciencetostartup.com/paper/equibim-learning-symmetry-equivariant-policy-for-bimanual-manipulation - EquiBim enhances bimanual robot manipulation by enforcing symmetry equivariance, leading to improved performance and robustness, making it a valuable tool for robotic automation.
- PCFEx: Point Cloud Feature Extraction for Graph Neural Networks (viability: 8): https://sciencetostartup.com/paper/pcfex-point-cloud-feature-extraction-for-graph-neural-networks - A novel GNN architecture for 3D point cloud processing, achieving state-of-the-art results in human pose estimation and activity recognition, with publicly available code.
- SWIFT: Sliding Window Reconstruction for Few-Shot Training-Free Generated Video Attribution (viability: 8): https://sciencetostartup.com/paper/swift-sliding-window-reconstruction-for-few-shot-training-free-generated-video-attribution - SWIFT offers a training-free method for attributing generated videos to their source model, achieving high accuracy with minimal samples, making it a valuable tool for content verification and misuse mitigation.
- SecAgent: Efficient Mobile GUI Agent with Semantic Context (viability: 8): https://sciencetostartup.com/paper/secagent-efficient-mobile-gui-agent-with-semantic-context - SecAgent is a 3B-scale mobile GUI agent that automates smartphone tasks using a novel semantic context mechanism and a new multilingual dataset, offering efficient and accurate performance.
- Towards Effective and Efficient Graph Alignment without Supervision (viability: 8): https://sciencetostartup.com/paper/towards-effective-and-efficient-graph-alignment-without-supervision - GlobAlign offers a faster and more accurate unsupervised graph alignment solution, potentially enabling better data integration and knowledge discovery across disparate graph datasets.
- BuildMamba: A Visual State-Space Based Model for Multi-Task Building Segmentation and Height Estimation from Satellite Images (viability: 8): https://sciencetostartup.com/paper/buildmamba-a-visual-state-space-based-model-for-multi-task-building-segmentation-and-height-estimation-from-satellite-im - BuildMamba offers a fast and accurate solution for building segmentation and height estimation from satellite imagery, enabling scalable 3D urban reconstruction.
- OccTrack360: 4D Panoptic Occupancy Tracking from Surround-View Fisheye Cameras (viability: 7): https://sciencetostartup.com/paper/occtrack360-4d-panoptic-occupancy-tracking-from-surround-view-fisheye-cameras - A new benchmark and baseline method for 4D panoptic occupancy tracking from surround-view fisheye cameras, enabling more robust autonomous driving systems.
- AtomVLA: Scalable Post-Training for Robotic Manipulation via Predictive Latent World Models (viability: 7): https://sciencetostartup.com/paper/atomvla-scalable-post-training-for-robotic-manipulation-via-predictive-latent-world-models - AtomVLA improves robotic manipulation by using LLMs to decompose tasks and a predictive world model for robust instruction grounding, offering a scalable post-training pipeline.
- Beyond Hungarian: Match-Free Supervision for End-to-End Object Detection (viability: 8): https://sciencetostartup.com/paper/beyond-hungarian-match-free-supervision-for-end-to-end-object-detection - A matching-free training scheme for DETR-based object detectors that eliminates the Hungarian algorithm, enhancing training efficiency and achieving state-of-the-art performance.
- Oracle-Guided Soft Shielding for Safe Move Prediction in Chess (viability: 7): https://sciencetostartup.com/paper/oracle-guided-soft-shielding-for-safe-move-prediction-in-chess - Oracle-Guided Soft Shielding (OGSS) enables safer decision-making in high-stakes environments by learning a probabilistic safety model from oracle feedback, demonstrated in chess with reduced blunder rates.
- Echo2ECG: Enhancing ECG Representations with Cardiac Morphology from Multi-View Echos (viability: 7): https://sciencetostartup.com/paper/echo2ecg-enhancing-ecg-representations-with-cardiac-morphology-from-multi-view-echos - Echo2ECG enhances ECG representations with multi-view Echo data, enabling improved cardiac phenotype classification and Echo retrieval from ECG queries.
- Spherical-GOF: Geometry-Aware Panoramic Gaussian Opacity Fields for 3D Scene Reconstruction (viability: 7): https://sciencetostartup.com/paper/spherical-gof-geometry-aware-panoramic-gaussian-opacity-fields-for-3d-scene-reconstruction - Reconstruct 3D scenes from omnidirectional images with improved geometric consistency using a novel spherical Gaussian Opacity Fields approach.
- Fanar-Sadiq: A Multi-Agent Architecture for Grounded Islamic QA (viability: 8): https://sciencetostartup.com/paper/fanar-sadiq-a-multi-agent-architecture-for-grounded-islamic-qa - Fanar-Sadiq is a multi-agent Islamic assistant that provides grounded answers to religious queries, accessible via API and web application, with proven traction.
- Improving Continual Learning for Gaussian Splatting based Environments Reconstruction on Commercial Off-the-Shelf Edge Devices (viability: 7): https://sciencetostartup.com/paper/improving-continual-learning-for-gaussian-splatting-based-environments-reconstruction-on-commercial-off-the-shelf-edge-d - Optimize Gaussian Splatting for edge devices, enabling real-time 3D scene reconstruction for robotics and other applications.
- Efficient Credal Prediction through Decalibration (viability: 7): https://sciencetostartup.com/paper/efficient-credal-prediction-through-decalibration - Efficiently estimate uncertainty in machine learning models by predicting probability intervals, enabling safer deployment in critical applications.
- Pareto-Optimal Anytime Algorithms via Bayesian Racing (viability: 7): https://sciencetostartup.com/paper/pareto-optimal-anytime-algorithms-via-bayesian-racing - PolarBear adaptively identifies Pareto-optimal optimization algorithms for anytime performance using Bayesian racing, enabling efficient algorithm selection under time constraints.
- Global Cross-Modal Geo-Localization: A Million-Scale Dataset and a Physical Consistency Learning Framework (viability: 8): https://sciencetostartup.com/paper/global-cross-modal-geo-localization-a-million-scale-dataset-and-a-physical-consistency-learning-framework - CORE is a million-scale dataset for cross-modal geo-localization, enabling a physical-law-aware network (PLANET) that significantly outperforms state-of-the-art methods, with code and dataset release for immediate application.
- An Open-Source Robotics Research Platform for Autonomous Laparoscopic Surgery (viability: 7): https://sciencetostartup.com/paper/an-open-source-robotics-research-platform-for-autonomous-laparoscopic-surgery - An open-source, robot-agnostic surgical robotics platform enabling precise teleoperation, data collection, and autonomous policy deployment for laparoscopic procedures.
- NN-OpInf: an operator inference approach using structure-preserving composable neural networks (viability: 7): https://sciencetostartup.com/paper/nn-opinf-an-operator-inference-approach-using-structure-preserving-composable-neural-networks - NN-OpInf offers a structure-preserving, composable neural network framework for reduced-order modeling of dynamical systems, improving accuracy and robustness over polynomial-based methods.
- Visual Self-Fulfilling Alignment: Shaping Safety-Oriented Personas via Threat-Related Images (viability: 7): https://sciencetostartup.com/paper/visual-self-fulfilling-alignment-shaping-safety-oriented-personas-via-threat-related-images - Fine-tune vision-language models with threat-related images to create safety-oriented personas without explicit safety labels, reducing harmful outputs.
- 3PoinTr: 3D Point Tracks for Robot Manipulation Pretraining from Casual Videos (viability: 7): https://sciencetostartup.com/paper/3pointr-3d-point-tracks-for-robot-manipulation-pretraining-from-casual-videos - Pretrain robot policies from unconstrained human videos using 3D point tracks for robust manipulation with minimal robot demonstrations.
- X-AVDT: Audio-Visual Cross-Attention for Robust Deepfake Detection (viability: 7): https://sciencetostartup.com/paper/x-avdt-audio-visual-cross-attention-for-robust-deepfake-detection - X-AVDT is a robust deepfake detector leveraging audio-visual cross-attention cues from generative models, offering improved accuracy and generalization across diverse synthesis paradigms.
- STRIDE: Structured Lagrangian and Stochastic Residual Dynamics via Flow Matching (viability: 7): https://sciencetostartup.com/paper/stride-structured-lagrangian-and-stochastic-residual-dynamics-via-flow-matching - STRIDE is a dynamics learning framework that combines Lagrangian Neural Networks and Conditional Flow Matching for improved robot control in uncertain environments.
- LAR-MoE: Latent-Aligned Routing for Mixture of Experts in Robotic Imitation Learning (viability: 7): https://sciencetostartup.com/paper/lar-moe-latent-aligned-routing-for-mixture-of-experts-in-robotic-imitation-learning - LAR-MoE enables robots to learn manipulation skills from demonstrations by using latent-aligned routing for mixture of experts, achieving high success rates and zero-shot transfer capabilities.
- R2F: Repurposing Ray Frontiers for LLM-free Object Navigation (viability: 7): https://sciencetostartup.com/paper/r2f-repurposing-ray-frontiers-for-llm-free-object-navigation - An LLM-free object navigation framework using ray frontiers for real-time robotic deployment, offering a faster alternative to VLM-based systems.
- Data-Driven Priors for Uncertainty-Aware Deterioration Risk Prediction with Multimodal Data (viability: 7): https://sciencetostartup.com/paper/data-driven-priors-for-uncertainty-aware-deterioration-risk-prediction-with-multimodal-data - MedCertAIn is a framework that improves the reliability and performance of in-hospital risk prediction using multimodal clinical data and uncertainty quantification.
- The Boiling Frog Threshold: Criticality and Blindness in World Model-Based Anomaly Detection Under Gradual Drift (viability: 2): https://sciencetostartup.com/paper/the-boiling-frog-threshold-criticality-and-blindness-in-world-model-based-anomaly-detection-under-gradual-drift - Explore self-monitoring boundaries in RL agents under gradual observation drift for more robust anomaly detection.
- LycheeCluster: Efficient Long-Context Inference with Structure-Aware Chunking and Hierarchical KV Indexing (viability: 7): https://sciencetostartup.com/paper/lycheecluster-efficient-long-context-inference-with-structure-aware-chunking-and-hierarchical-kv-indexing - LycheeCluster accelerates long-context LLM inference by 3.6x using structure-aware chunking and hierarchical KV indexing, offering a drop-in replacement for existing KV cache management.
- A Dataset for Probing Translationese Preferences in English-to-Swedish Translation (viability: 7): https://sciencetostartup.com/paper/a-dataset-for-probing-translationese-preferences-in-english-to-swedish-translation - A dataset and benchmark for training language models to avoid translationese and produce more natural Swedish translations.
- A prospective clinical feasibility study of a conversational diagnostic AI in an ambulatory primary care clinic (viability: 7): https://sciencetostartup.com/paper/a-prospective-clinical-feasibility-study-of-a-conversational-diagnostic-ai-in-an-ambulatory-primary-care-clinic - AMIE, a conversational AI for clinical history taking and diagnosis, demonstrates feasibility, safety, and user acceptance in a real-world primary care setting, improving patient satisfaction and provider preparedness.
- Efficient Policy Learning with Hybrid Evaluation-Based Genetic Programming for Uncertain Agile Earth Observation Satellite Scheduling (viability: 7): https://sciencetostartup.com/paper/efficient-policy-learning-with-hybrid-evaluation-based-genetic-programming-for-uncertain-agile-earth-observation-satelli - A hybrid evaluation-based genetic programming approach for optimizing Earth observation satellite scheduling under uncertainty, offering a balance between computational cost and scheduling performance.
- Alfa: Attentive Low-Rank Filter Adaptation for Structure-Aware Cross-Domain Personalized Gaze Estimation (viability: 7): https://sciencetostartup.com/paper/alfa-attentive-low-rank-filter-adaptation-for-structure-aware-cross-domain-personalized-gaze-estimation - Alfa efficiently personalizes gaze estimation models by reweighting pre-trained features, offering improved accuracy and faster adaptation for user-specific applications.
- Can Vision-Language Models Solve the Shell Game? (viability: 8): https://sciencetostartup.com/paper/can-vision-language-models-solve-the-shell-game - VET-Bench is a new benchmark and method (SGCoT) to improve VLMs' ability to track visually identical objects over time, achieving state-of-the-art accuracy.
- Information Maximization for Long-Tailed Semi-Supervised Domain Generalization (viability: 7): https://sciencetostartup.com/paper/information-maximization-for-long-tailed-semi-supervised-domain-generalization - IMaX enhances semi-supervised domain generalization by maximizing mutual information to handle long-tailed class distributions, improving performance in real-world image classification tasks.
- FoMo: A Multi-Season Dataset for Robot Navigation in Forêt Montmorency (viability: 7): https://sciencetostartup.com/paper/fomo-a-multi-season-dataset-for-robot-navigation-in-for-t-montmorency - A multi-sensor, multi-season robotics dataset for robust navigation in challenging environments, enabling development of more reliable autonomous systems.
- One Model Is Enough: Native Retrieval Embeddings from LLM Agent Hidden States (viability: 7): https://sciencetostartup.com/paper/one-model-is-enough-native-retrieval-embeddings-from-llm-agent-hidden-states - Equip LLM agents with native retrieval by projecting hidden states into the embedding space, eliminating the need for a separate embedding model and reducing latency.
- Grow, Assess, Compress: Adaptive Backbone Scaling for Memory-Efficient Class Incremental Learning (viability: 7): https://sciencetostartup.com/paper/grow-assess-compress-adaptive-backbone-scaling-for-memory-efficient-class-incremental-learning - GRACE adaptively scales model capacity for class incremental learning, balancing plasticity and stability while reducing memory footprint.
- IronEngine: Towards General AI Assistant (viability: 7): https://sciencetostartup.com/paper/ironengine-towards-general-ai-assistant - IronEngine is a general AI assistant platform with a unified orchestration core, demonstrating 100% task completion on file operation benchmarks.
- SYNAPSE: Framework for Neuron Analysis and Perturbation in Sequence Encoding (viability: 3): https://sciencetostartup.com/paper/synapse-framework-for-neuron-analysis-and-perturbation-in-sequence-encoding - Develop a lightweight framework, SYNAPSE, for stress-testing and understanding Transformer models' robustness without retraining.
- Tactile Recognition of Both Shapes and Materials with Automatic Feature Optimization-Enabled Meta Learning (viability: 8): https://sciencetostartup.com/paper/tactile-recognition-of-both-shapes-and-materials-with-automatic-feature-optimization-enabled-meta-learning - A meta-learning framework for tactile recognition of shapes and materials, enabling robots to quickly adapt to new objects with minimal training data.
- Human-Aware Robot Behaviour in Self-Driving Labs (viability: 7): https://sciencetostartup.com/paper/human-aware-robot-behaviour-in-self-driving-labs - An AI-powered robot perception system that enables robots to proactively interact with humans in shared lab environments, improving efficiency in self-driving labs.
- Meta-RL with Shared Representations Enables Fast Adaptation in Energy Systems (viability: 7): https://sciencetostartup.com/paper/meta-rl-with-shared-representations-enables-fast-adaptation-in-energy-systems - A Meta-RL framework with shared state feature extractor for fast adaptation in building energy management systems, enabling efficient representation learning and improved generalization.
- Geometrically Constrained Outlier Synthesis (viability: 7): https://sciencetostartup.com/paper/geometrically-constrained-outlier-synthesis - GCOS improves OOD detection by synthesizing geometrically constrained outliers, leading to more robust and reliable AI systems.
- Aligning to Illusions: Choice Blindness in Human and AI Feedback (viability: 3): https://sciencetostartup.com/paper/aligning-to-illusions-choice-blindness-in-human-and-ai-feedback - This paper explores the limitations of human and AI feedback in reinforcement learning, focusing on the detectability of choice blindness in preference annotations.
- Sandpiper: Orchestrated AI-Annotation for Educational Discourse at Scale (viability: 7): https://sciencetostartup.com/paper/sandpiper-orchestrated-ai-annotation-for-educational-discourse-at-scale - Sandpiper is a mixed-initiative system that combines interactive dashboards with LLM engines to enable scalable qualitative analysis of educational discourse data.
- SPIRAL: A Closed-Loop Framework for Self-Improving Action World Models via Reflective Planning Agents (viability: 7): https://sciencetostartup.com/paper/spiral-a-closed-loop-framework-for-self-improving-action-world-models-via-reflective-planning-agents - SPIRAL is a closed-loop framework for controllable long-horizon video generation conditioned on high-level semantic actions, enabling iterative refinement and improved semantic alignment.
- A Recipe for Stable Offline Multi-agent Reinforcement Learning (viability: 1): https://sciencetostartup.com/paper/a-recipe-for-stable-offline-multi-agent-reinforcement-learning - Stabilize multi-agent reinforcement learning with scale-invariant value normalization.
- Revealing Behavioral Plasticity in Large Language Models: A Token-Conditional Perspective (viability: 2): https://sciencetostartup.com/paper/revealing-behavioral-plasticity-in-large-language-models-a-token-conditional-perspective - A framework utilizing token-conditional reinforcement learning to stabilize behavioral plasticity in large language models.
- Decoupling Distance and Networks: Hybrid Graph Attention-Geostatistical Methods for Spatio-temporal Risk Mapping (viability: 7): https://sciencetostartup.com/paper/decoupling-distance-and-networks-hybrid-graph-attention-geostatistical-methods-for-spatio-temporal-risk-mapping - A hybrid model combining graph attention networks and geostatistics for improved spatial risk prediction, offering better accuracy and uncertainty quantification.
- Adaptive Loops and Memory in Transformers: Think Harder or Know More? (viability: 7): https://sciencetostartup.com/paper/adaptive-loops-and-memory-in-transformers-think-harder-or-know-more - Optimize transformer performance by combining adaptive looping and gated memory banks for improved reasoning and storage capacity.
- StructBiHOI: Structured Articulation Modeling for Long--Horizon Bimanual Hand--Object Interaction Generation (viability: 7): https://sciencetostartup.com/paper/structbihoi-structured-articulation-modeling-for-long-horizon-bimanual-hand-object-interaction-generation - StructBiHOI enables stable and efficient long-horizon bimanual hand-object interaction generation by structurally disentangling temporal joint planning from frame-level manipulation refinement.
- A Hierarchical Error-Corrective Graph Framework for Autonomous Agents with LLM-Based Action Generation (viability: 3): https://sciencetostartup.com/paper/a-hierarchical-error-corrective-graph-framework-for-autonomous-agents-with-llm-based-action-generation - Develop an error-corrective graph framework for autonomous agents utilizing LLM-based action generation to enhance execution reliability.
- AULLM++: Structural Reasoning with Large Language Models for Micro-Expression Recognition (viability: 7): https://sciencetostartup.com/paper/aullm-structural-reasoning-with-large-language-models-for-micro-expression-recognition - AULLM++ uses LLMs to improve micro-expression recognition by reasoning about facial muscle activations, offering potential for emotion detection applications.
- Real-Time Drone Detection in Event Cameras via Per-Pixel Frequency Analysis (viability: 8): https://sciencetostartup.com/paper/real-time-drone-detection-in-event-cameras-via-per-pixel-frequency-analysis - Real-time drone detection API using event camera data and frequency analysis, outperforming YOLO in accuracy and latency.
- MoMaStage: Skill-State Graph Guided Planning and Closed-Loop Execution for Long-Horizon Indoor Mobile Manipulation (viability: 8): https://sciencetostartup.com/paper/momastage-skill-state-graph-guided-planning-and-closed-loop-execution-for-long-horizon-indoor-mobile-manipulation - MoMaStage is a structured vision-language framework for long-horizon mobile manipulation that uses a skill-state graph to ensure logically consistent plans and robust execution, outperforming state-of-the-art baselines.
- Perception-Aware Communication-Free Multi-UAV Coordination in the Wild (viability: 7): https://sciencetostartup.com/paper/perception-aware-communication-free-multi-uav-coordination-in-the-wild - A communication-free multi-UAV coordination system using 3D LiDAR for navigation in GPS-denied environments, enabling safe and effective operation in complex scenarios.
- Beyond the Markovian Assumption: Robust Optimization via Fractional Weyl Integrals in Imbalanced Data (viability: 7): https://sciencetostartup.com/paper/beyond-the-markovian-assumption-robust-optimization-via-fractional-weyl-integrals-in-imbalanced-data - A novel optimization algorithm using Fractional Calculus to improve performance in imbalanced datasets, particularly for financial fraud detection.
- This Looks Distinctly Like That: Grounding Interpretable Recognition in Stiefel Geometry against Neural Collapse (viability: 7): https://sciencetostartup.com/paper/this-looks-distinctly-like-that-grounding-interpretable-recognition-in-stiefel-geometry-against-neural-collapse - Adaptive Manifold Prototypes (AMP) improves classification accuracy and interpretability by representing class prototypes as orthonormal bases on the Stiefel manifold, preventing prototype collapse.
- Leaderboard Incentives: Model Rankings under Strategic Post-Training (viability: 7): https://sciencetostartup.com/paper/leaderboard-incentives-model-rankings-under-strategic-post-training - Optimize AI model evaluation with a novel benchmarking protocol that aligns incentives and accurately ranks model quality, preventing strategic exploitation.
- Unifying On- and Off-Policy Variance Reduction Methods (viability: 7): https://sciencetostartup.com/paper/unifying-on-and-off-policy-variance-reduction-methods - Unify online A/B testing and off-policy evaluation methods to improve experimentation efficiency.
- M$^3$-ACE: Rectifying Visual Perception in Multimodal Math Reasoning via Multi-Agentic Context Engineering (viability: 8): https://sciencetostartup.com/paper/m-3-ace-rectifying-visual-perception-in-multimodal-math-reasoning-via-multi-agentic-context-engineering - M3-ACE is a multi-agent system that improves visual math reasoning by rectifying visual perception, achieving state-of-the-art results and offering a clear path to commercial applications.
- Diffusion-Based Data Augmentation for Image Recognition: A Systematic Analysis and Evaluation (viability: 7): https://sciencetostartup.com/paper/diffusion-based-data-augmentation-for-image-recognition-a-systematic-analysis-and-evaluation - A unified framework and codebase for diffusion-based data augmentation, enabling fair comparison and practical insights for low-data image classification.
- $Δ$VLA: Prior-Guided Vision-Language-Action Models via World Knowledge Variation (viability: 7): https://sciencetostartup.com/paper/vla-prior-guided-vision-language-action-models-via-world-knowledge-variation - DeltaVLA enhances robotic manipulation by predicting world-knowledge variations for improved action generation, offering a more efficient and interpretable approach to robotic control.
- Computational modeling of early language learning from acoustic speech and audiovisual input without linguistic priors (viability: 3): https://sciencetostartup.com/paper/computational-modeling-of-early-language-learning-from-acoustic-speech-and-audiovisual-input-without-linguistic-priors - Develop computational models for understanding early language acquisition from speech and audiovisual input without linguistic priors.
- Do Language Models Know Theo Has a Wife? Investigating the Proviso Problem (viability: 6): https://sciencetostartup.com/paper/do-language-models-know-theo-has-a-wife-investigating-the-proviso-problem - A diagnostic dataset and evaluation framework to probe pragmatic reasoning in language models, revealing reliance on pattern matching rather than true understanding.
- Towards plausibility in time series counterfactual explanations (viability: 7): https://sciencetostartup.com/paper/towards-plausibility-in-time-series-counterfactual-explanations - Generate plausible counterfactual explanations for time series classification using gradient-based optimization and soft-DTW alignment, enabling more realistic temporal structure.
- Local-Global Prompt Learning via Sparse Optimal Transport (viability: 8): https://sciencetostartup.com/paper/local-global-prompt-learning-via-sparse-optimal-transport - Improve few-shot classification and OOD detection by learning shared global prompts and class-specific local prompts with sparse optimal transport.
- Rethinking Attention Output Projection: Structured Hadamard Transforms for Efficient Transformers (viability: 7): https://sciencetostartup.com/paper/rethinking-attention-output-projection-structured-hadamard-transforms-for-efficient-transformers - Replace the dense output projection in multi-head attention with a structured Hadamard transform for efficient Transformers, reducing parameters and improving throughput.
- Hierarchical Multi-Modal Planning for Fixed-Altitude Sparse Target Search and Sampling (viability: 7): https://sciencetostartup.com/paper/hierarchical-multi-modal-planning-for-fixed-altitude-sparse-target-search-and-sampling - HIMoS is a fixed-altitude framework for autonomous underwater vehicles that efficiently monitors sparse benthic phenomena using a hierarchical multi-modal planning approach.
- Detecting Fake Reviewer Groups in Dynamic Networks: An Adaptive Graph Learning Method (viability: 8): https://sciencetostartup.com/paper/detecting-fake-reviewer-groups-in-dynamic-networks-an-adaptive-graph-learning-method - Detect fake reviewer groups on e-commerce platforms with a graph learning model, improving trust and fair competition.
- SPD-RAG: Sub-Agent Per Document Retrieval-Augmented Generation (viability: 7): https://sciencetostartup.com/paper/spd-rag-sub-agent-per-document-retrieval-augmented-generation - SPD-RAG is a hierarchical multi-agent framework for question answering that decomposes the problem along the document axis, improving scalability and answer quality.
- Beyond Attention Heatmaps: How to Get Better Explanations for Multiple Instance Learning Models in Histopathology (viability: 7): https://sciencetostartup.com/paper/beyond-attention-heatmaps-how-to-get-better-explanations-for-multiple-instance-learning-models-in-histopathology - Improve histopathology biomarker discovery by validating and improving the explainability of multiple instance learning models using a new benchmarking framework and advanced explanation methods.
- EndoSERV: A Vision-based Endoluminal Robot Navigation System (viability: 7): https://sciencetostartup.com/paper/endoserv-a-vision-based-endoluminal-robot-navigation-system - EndoSERV is a vision-based localization method for robot-assisted endoluminal procedures that uses a segment-to-structure and real-to-virtual mapping approach.
- Agentic Neurosymbolic Collaboration for Mathematical Discovery: A Case Study in Combinatorial Design (viability: 3): https://sciencetostartup.com/paper/agentic-neurosymbolic-collaboration-for-mathematical-discovery-a-case-study-in-combinatorial-design - Leverage human-AI neurosymbolic collaboration to drive mathematical discovery in combinatorial designs.
- CORE-Acu: Structured Reasoning Traces and Knowledge Graph Safety Verification for Acupuncture Clinical Decision Support (viability: 8): https://sciencetostartup.com/paper/core-acu-structured-reasoning-traces-and-knowledge-graph-safety-verification-for-acupuncture-clinical-decision-support - CORE-Acu provides a safe and interpretable AI-powered clinical decision support system for acupuncture, leveraging structured reasoning and knowledge graph verification.
- Human-AI Divergence in Ego-centric Action Recognition under Spatial and Spatiotemporal Manipulations (viability: 7): https://sciencetostartup.com/paper/human-ai-divergence-in-ego-centric-action-recognition-under-spatial-and-spatiotemporal-manipulations - Identify and leverage minimal visual cues for robust action recognition, bridging the gap between human and AI performance in real-world scenarios.
- SlowBA: An efficiency backdoor attack towards VLM-based GUI agents (viability: 7): https://sciencetostartup.com/paper/slowba-an-efficiency-backdoor-attack-towards-vlm-based-gui-agents - SlowBA is a backdoor attack that targets the responsiveness of VLM-based GUI agents by manipulating response latency, creating a need for security solutions focused on response efficiency.
- HDR-NSFF: High Dynamic Range Neural Scene Flow Fields (viability: 7): https://sciencetostartup.com/paper/hdr-nsff-high-dynamic-range-neural-scene-flow-fields - Reconstruct dynamic HDR radiance fields from monocular videos using neural scene flow fields, enabling high-quality novel space-time view synthesis.
- Learning Multiple Utterance-Level Attribute Representations with a Unified Speech Encoder (viability: 7): https://sciencetostartup.com/paper/learning-multiple-utterance-level-attribute-representations-with-a-unified-speech-encoder - A unified post-training framework that enables a single speech foundation model to generate multiple types of utterance-level representations, enabling effective multimodal and multilingual applications.
- Concept-Guided Fine-Tuning: Steering ViTs away from Spurious Correlations to Improve Robustness (viability: 7): https://sciencetostartup.com/paper/concept-guided-fine-tuning-steering-vits-away-from-spurious-correlations-to-improve-robustness - Fine-tune Vision Transformers to focus on semantic concepts rather than spurious correlations, improving robustness in out-of-distribution scenarios.
- Retrieval-Augmented Anatomical Guidance for Text-to-CT Generation (viability: 7): https://sciencetostartup.com/paper/retrieval-augmented-anatomical-guidance-for-text-to-ct-generation - Generate anatomically consistent CT scans from radiology reports by retrieving and incorporating relevant anatomical annotations, improving image fidelity and clinical consistency.
- SoK: Harmonizing Attack Graphs and Intrusion Detection Systems (viability: 7): https://sciencetostartup.com/paper/sok-harmonizing-attack-graphs-and-intrusion-detection-systems - A unified framework integrating Attack Graphs and Intrusion Detection Systems to enhance threat detection and incident response.
- Deconstructing Multimodal Mathematical Reasoning: Towards a Unified Perception-Alignment-Reasoning Paradigm (viability: 2): https://sciencetostartup.com/paper/deconstructing-multimodal-mathematical-reasoning-towards-a-unified-perception-alignment-reasoning-paradigm - Develops a framework for integrating perception, alignment, and reasoning to improve multimodal mathematical reasoning models.
- Minor First, Major Last: A Depth-Induced Implicit Bias of Sharpness-Aware Minimization (viability: 4): https://sciencetostartup.com/paper/minor-first-major-last-a-depth-induced-implicit-bias-of-sharpness-aware-minimization - A study on Sharpness-Aware Minimization (SAM) reveals its implicit bias in deep linear networks, suggesting potential for improved optimization strategies.
- Novel Semantic Prompting for Zero-Shot Action Recognition (viability: 7): https://sciencetostartup.com/paper/novel-semantic-prompting-for-zero-shot-action-recognition - SP-CLIP enhances zero-shot action recognition by augmenting vision-language models with structured semantic prompts, offering improved accuracy and efficiency without requiring model retraining.
- A Blockchain-based Traceability System for AI-Driven Engine Blade Inspection (viability: 7): https://sciencetostartup.com/paper/a-blockchain-based-traceability-system-for-ai-driven-engine-blade-inspection - BladeChain provides immutable traceability for aircraft engine blade inspections using blockchain and AI, ensuring auditable maintenance across stakeholders.
- LAMUS: A Large-Scale Corpus for Legal Argument Mining from U.S. Caselaw using LLMs (viability: 7): https://sciencetostartup.com/paper/lamus-a-large-scale-corpus-for-legal-argument-mining-from-u-s-caselaw-using-llms - LAMUS is a large-scale legal argument mining corpus with code and data available, enabling development of AI tools for legal reasoning.
- PolyFormer: learning efficient reformulations for scalable optimization under complex physical constraints (viability: 8): https://sciencetostartup.com/paper/polyformer-learning-efficient-reformulations-for-scalable-optimization-under-complex-physical-constraints - PolyFormer simplifies constrained optimization problems by learning geometric structures and transforming them into efficient polytopic reformulations, enabling significant speedups and memory reductions.
- Evaluating LLM-Based Grant Proposal Review via Structured Perturbations (viability: 7): https://sciencetostartup.com/paper/evaluating-llm-based-grant-proposal-review-via-structured-perturbations - Automated grant proposal review tool using LLMs to identify key weaknesses and improve proposal quality.
- OSCAR: Occupancy-based Shape Completion via Acoustic Neural Implicit Representations (viability: 8): https://sciencetostartup.com/paper/oscar-occupancy-based-shape-completion-via-acoustic-neural-implicit-representations - Reconstruct complete 3D anatomical geometry from partial ultrasound observations using acoustic neural implicit representations, enabling label-free completion for minimally invasive spine interventions.
- TA-RNN-Medical-Hybrid: A Time-Aware and Interpretable Framework for Mortality Risk Prediction (viability: 7): https://sciencetostartup.com/paper/ta-rnn-medical-hybrid-a-time-aware-and-interpretable-framework-for-mortality-risk-prediction - A time-aware deep learning framework for interpretable mortality risk prediction in ICUs, leveraging EHR data and medical knowledge to improve accuracy and provide clinically meaningful explanations.
- AdaCultureSafe: Adaptive Cultural Safety Grounded by Cultural Knowledge in Large Language Models (viability: 7): https://sciencetostartup.com/paper/adaculturesafe-adaptive-cultural-safety-grounded-by-cultural-knowledge-in-large-language-models - Enhance LLM cultural safety with a knowledge-grounded method using the AdaCultureSafe dataset.
- How Much Do LLMs Hallucinate in Document Q&A Scenarios? A 172-Billion-Token Study Across Temperatures, Context Lengths, and Hardware Platforms (viability: 7): https://sciencetostartup.com/paper/how-much-do-llms-hallucinate-in-document-q-a-scenarios-a-172-billion-token-study-across-temperatures-context-lengths-and - RIKER is a ground-truth-first evaluation methodology that enables deterministic scoring of LLM hallucinations in document Q&A scenarios, providing insights for enterprise AI deployments.
- Prototype-Guided Concept Erasure in Diffusion Models (viability: 7): https://sciencetostartup.com/paper/prototype-guided-concept-erasure-in-diffusion-models - Erase broad concepts like 'sexual' or 'violent' from text-to-image models using concept prototypes for safer image generation.
- SCL-GNN: Towards Generalizable Graph Neural Networks via Spurious Correlation Learning (viability: 7): https://sciencetostartup.com/paper/scl-gnn-towards-generalizable-graph-neural-networks-via-spurious-correlation-learning - SCL-GNN enhances GNN generalization by mitigating spurious correlations, offering improved robustness for graph-based tasks.
- SAIL: Test-Time Scaling for In-Context Imitation Learning with VLM (viability: 7): https://sciencetostartup.com/paper/sail-test-time-scaling-for-in-context-imitation-learning-with-vlm - SAIL is a framework that improves robot imitation learning by iteratively refining trajectories using a vision language model and Monte Carlo Tree Search, scaling with test-time compute.
- Towards a more efficient bias detection in financial language models (viability: 7): https://sciencetostartup.com/paper/towards-a-more-efficient-bias-detection-in-financial-language-models - Accelerate bias detection in financial language models by leveraging cross-model similarities to reduce computational costs and enable continuous monitoring.
- Airborne Magnetic Anomaly Navigation with Neural-Network-Augmented Online Calibration (viability: 7): https://sciencetostartup.com/paper/airborne-magnetic-anomaly-navigation-with-neural-network-augmented-online-calibration - An adaptive airborne navigation system using magnetometers and neural networks to compensate for magnetic interference in real-time, eliminating the need for pre-calibration flights.
- FinToolBench: Evaluating LLM Agents for Real-World Financial Tool Use (viability: 8): https://sciencetostartup.com/paper/fintoolbench-evaluating-llm-agents-for-real-world-financial-tool-use - FinToolBench provides a real-world benchmark and evaluation framework for LLM agents using financial tools, enabling auditable and trustworthy AI in finance.
- Seed2Scale: A Self-Evolving Data Engine for Embodied AI via Small to Large Model Synergy and Multimodal Evaluation (viability: 8): https://sciencetostartup.com/paper/seed2scale-a-self-evolving-data-engine-for-embodied-ai-via-small-to-large-model-synergy-and-multimodal-evaluation - Seed2Scale is a self-evolving data engine for embodied AI that leverages small and large model synergy to generate high-quality training data, enabling significant performance improvements with minimal initial data.
- WaDi: Weight Direction-aware Distillation for One-step Image Synthesis (viability: 7): https://sciencetostartup.com/paper/wadi-weight-direction-aware-distillation-for-one-step-image-synthesis - WaDi distills Stable Diffusion into a fast, one-step image generator using a parameter-efficient adapter, achieving state-of-the-art FID scores and strong versatility.
- Beyond ReinMax: Low-Variance Gradient Estimators for Discrete Latent Variables (viability: 7): https://sciencetostartup.com/paper/beyond-reinmax-low-variance-gradient-estimators-for-discrete-latent-variables - Reduce variance in discrete latent variable models with ReinMax-Rao and ReinMax-CV estimators for improved variational autoencoder training.
- NCL-UoR at SemEval-2026 Task 5: Embedding-Based Methods, Fine-Tuning, and LLMs for Word Sense Plausibility Rating (viability: 7): https://sciencetostartup.com/paper/ncl-uor-at-semeval-2026-task-5-embedding-based-methods-fine-tuning-and-llms-for-word-sense-plausibility-rating - A structured prompting strategy for word sense plausibility rating that outperforms fine-tuned models and embedding-based approaches, offering a calibrated rating system for ambiguous homonyms in narrative contexts.
- FlowTouch: View-Invariant Visuo-Tactile Prediction (viability: 7): https://sciencetostartup.com/paper/flowtouch-view-invariant-visuo-tactile-prediction - FlowTouch predicts tactile sensor readings from visual input using 3D mesh representations, enabling robots to 'feel' objects before touching them, improving grasp stability.
- DynamicVGGT: Learning Dynamic Point Maps for 4D Scene Reconstruction in Autonomous Driving (viability: 7): https://sciencetostartup.com/paper/dynamicvggt-learning-dynamic-point-maps-for-4d-scene-reconstruction-in-autonomous-driving - DynamicVGGT enables accurate 4D scene reconstruction for autonomous driving by predicting point motion and using motion-aware temporal attention.
- FedPrism: Adaptive Personalized Federated Learning under Non-IID Data (viability: 7): https://sciencetostartup.com/paper/fedprism-adaptive-personalized-federated-learning-under-non-iid-data - FedPrism enhances federated learning by adaptively personalizing models for clients with heterogeneous data, improving accuracy and robustness in real-world deployments.
- Optimising antibiotic switching via forecasting of patient physiology (viability: 7): https://sciencetostartup.com/paper/optimising-antibiotic-switching-via-forecasting-of-patient-physiology - Forecast patient physiology using neural processes to optimize antibiotic switching, reducing hospital stays and healthcare costs.
- SiMO: Single-Modality-Operable Multimodal Collaborative Perception (viability: 7): https://sciencetostartup.com/paper/simo-single-modality-operable-multimodal-collaborative-perception - SiMO enhances collaborative perception by adaptively handling sensor failures and maintaining semantic consistency, ensuring robust performance across individual modalities.
- Fibration Policy Optimization (viability: 7): https://sciencetostartup.com/paper/fibration-policy-optimization - FiberPO offers a novel approach to LLM policy optimization by introducing a hierarchical gating mechanism for improved stability and token efficiency.
- Exploring Deep Learning and Ultra-Widefield Imaging for Diabetic Retinopathy and Macular Edema (viability: 7): https://sciencetostartup.com/paper/exploring-deep-learning-and-ultra-widefield-imaging-for-diabetic-retinopathy-and-macular-edema - Apply deep learning to ultra-widefield retinal imaging for improved diabetic retinopathy and macular edema detection, leveraging a public dataset and achieving strong benchmark results.
- The Struggle Between Continuation and Refusal: A Mechanistic Analysis of the Continuation-Triggered Jailbreak in LLMs (viability: 3): https://sciencetostartup.com/paper/the-struggle-between-continuation-and-refusal-a-mechanistic-analysis-of-the-continuation-triggered-jailbreak-in-llms - This paper analyzes the continuation-triggered jailbreak mechanisms in LLMs to enhance their safety.
- A General Lie-Group Framework for Continuum Soft Robot Modeling (viability: 7): https://sciencetostartup.com/paper/a-general-lie-group-framework-for-continuum-soft-robot-modeling - A Lie group framework for soft robot modeling enables geometric control and efficient simulation, paving the way for advanced robotic designs.
- Disentangling Reasoning in Large Audio-Language Models for Ambiguous Emotion Prediction (viability: 7): https://sciencetostartup.com/paper/disentangling-reasoning-in-large-audio-language-models-for-ambiguous-emotion-prediction - Refine speech emotion recognition by using large audio-language models to predict distributions of emotions rather than single labels, improving accuracy and reflecting the ambiguous nature of human emotion.
- GarmentPainter: Efficient 3D Garment Texture Synthesis with Character-Guided Diffusion Model (viability: 7): https://sciencetostartup.com/paper/garmentpainter-efficient-3d-garment-texture-synthesis-with-character-guided-diffusion-model - GarmentPainter synthesizes high-quality, 3D-aware garment textures using a character-guided diffusion model, offering a simple and efficient solution for the fashion industry.
- SRNeRV: A Scale-wise Recursive Framework for Neural Video Representation (viability: 7): https://sciencetostartup.com/paper/srnerv-a-scale-wise-recursive-framework-for-neural-video-representation - SRNeRV is a parameter-efficient neural video representation framework that significantly improves rate-distortion performance, making it ideal for video compression applications.
- Practical Type Inference: High-Throughput Recovery of Real-World Structures and Function Signatures (viability: 7): https://sciencetostartup.com/paper/practical-type-inference-high-throughput-recovery-of-real-world-structures-and-function-signatures - XTRIDE is a fast and accurate tool for recovering types from stripped binaries, enabling exact decompilation and automated reverse engineering workflows.
- SAVE: Speech-Aware Video Representation Learning for Video-Text Retrieval (viability: 7): https://sciencetostartup.com/paper/save-speech-aware-video-representation-learning-for-video-text-retrieval - SAVE enhances video-text retrieval by incorporating speech embeddings and early vision-audio alignment, outperforming state-of-the-art methods on multiple benchmarks.
- SplitAgent: A Privacy-Preserving Distributed Architecture for Enterprise-Cloud Agent Collaboration (viability: 7): https://sciencetostartup.com/paper/splitagent-a-privacy-preserving-distributed-architecture-for-enterprise-cloud-agent-collaboration - SplitAgent enables privacy-preserving collaboration between enterprise and cloud AI agents through context-aware dynamic sanitization, allowing enterprises to leverage cloud AI without compromising sensitive data.
- Revisiting Gradient Staleness: Evaluating Distance Metrics for Asynchronous Federated Learning Aggregation (viability: 7): https://sciencetostartup.com/paper/revisiting-gradient-staleness-evaluating-distance-metrics-for-asynchronous-federated-learning-aggregation - Improve asynchronous federated learning with adaptive aggregation using enhanced distance metrics for more robust and efficient training.
- Alignment-Aware and Reliability-Gated Multimodal Fusion for Unmanned Aerial Vehicle Detection Across Heterogeneous Thermal-Visual Sensors (viability: 7): https://sciencetostartup.com/paper/alignment-aware-and-reliability-gated-multimodal-fusion-for-unmanned-aerial-vehicle-detection-across-heterogeneous-therm - A multimodal UAV detection system leveraging registration-aware and reliability-gated fusion to improve detection performance in heterogeneous sensor environments.
- Distributional Regression with Tabular Foundation Models: Evaluating Probabilistic Predictions via Proper Scoring Rules (viability: 7): https://sciencetostartup.com/paper/distributional-regression-with-tabular-foundation-models-evaluating-probabilistic-predictions-via-proper-scoring-rules - Improve tabular foundation models by evaluating probabilistic forecasts using proper scoring rules, enabling more accurate and reliable predictions in practical regression settings.
- MM-TS: Multi-Modal Temperature and Margin Schedules for Contrastive Learning with Long-Tail Data (viability: 8): https://sciencetostartup.com/paper/mm-ts-multi-modal-temperature-and-margin-schedules-for-contrastive-learning-with-long-tail-data - MM-TS dynamically adjusts temperature and margin in multi-modal contrastive learning to improve performance on long-tail data, offering a drop-in replacement for existing contrastive loss functions.
- Sequential Service Region Design with Capacity-Constrained Investment and Spillover Effect (viability: 7): https://sciencetostartup.com/paper/sequential-service-region-design-with-capacity-constrained-investment-and-spillover-effect - Optimize sequential service region expansion under uncertainty using real options analysis and Transformer-based reinforcement learning.
- TildeOpen LLM: Leveraging Curriculum Learning to Achieve Equitable Language Representation (viability: 8): https://sciencetostartup.com/paper/tildeopen-llm-leveraging-curriculum-learning-to-achieve-equitable-language-representation - TildeOpen LLM is a 30B multilingual model focused on European languages, outperforming existing models with fewer resources, and available on Hugging Face.
- ALOOD: Exploiting Language Representations for LiDAR-based Out-of-Distribution Object Detection (viability: 7): https://sciencetostartup.com/paper/alood-exploiting-language-representations-for-lidar-based-out-of-distribution-object-detection - ALOOD leverages language representations to improve LiDAR-based 3D object detection in autonomous driving systems by identifying out-of-distribution objects, enhancing safety and reliability.
- Is continuous CoT better suited for multi-lingual reasoning? (viability: 7): https://sciencetostartup.com/paper/is-continuous-cot-better-suited-for-multi-lingual-reasoning - Continuous Chain-of-Thought enables efficient and robust multilingual reasoning, particularly for low-resource languages, offering a scalable solution for cross-lingual applications.
- MERLIN: Building Low-SNR Robust Multimodal LLMs for Electromagnetic Signals (viability: 8): https://sciencetostartup.com/paper/merlin-building-low-snr-robust-multimodal-llms-for-electromagnetic-signals - MERLIN is a robust MLLM framework for electromagnetic signals, enhanced for low-SNR environments, with a released dataset and benchmark for standardized evaluation, enabling a product for signal analysis and interpretation.
- Evolution Strategy-Based Calibration for Low-Bit Quantization of Speech Models (viability: 3): https://sciencetostartup.com/paper/evolution-strategy-based-calibration-for-low-bit-quantization-of-speech-models - Develop a quantization calibration method for speech models that minimizes performance loss during deployment.
- Evidence-Driven Reasoning for Industrial Maintenance Using Heterogeneous Data (viability: 7): https://sciencetostartup.com/paper/evidence-driven-reasoning-for-industrial-maintenance-using-heterogeneous-data - Condition Insight Agent provides evidence-grounded explanations and advisory actions for industrial maintenance by integrating maintenance language, operational data, and engineering failure semantics.
- RexDrug: Reliable Multi-Drug Combination Extraction through Reasoning-Enhanced LLMs (viability: 8): https://sciencetostartup.com/paper/rexdrug-reliable-multi-drug-combination-extraction-through-reasoning-enhanced-llms - RexDrug is a reasoning-enhanced LLM framework for extracting reliable multi-drug combinations from biomedical literature, offering a scalable solution for precision medicine.
- An explainable hybrid deep learning-enabled intelligent fault detection and diagnosis approach for automotive software systems validation (viability: 7): https://sciencetostartup.com/paper/an-explainable-hybrid-deep-learning-enabled-intelligent-fault-detection-and-diagnosis-approach-for-automotive-software-s - An explainable fault detection and diagnosis model for automotive software systems validation, leveraging hybrid deep learning and explainable AI techniques.
- Covenant-72B: Pre-Training a 72B LLM with Trustless Peers Over-the-Internet (viability: 6): https://sciencetostartup.com/paper/covenant-72b-pre-training-a-72b-llm-with-trustless-peers-over-the-internet - Covenant-72B is a 72B parameter LLM pre-trained in a globally distributed, permissionless manner, demonstrating the feasibility of democratized large-scale model training.
- Learning Hierarchical Knowledge in Text-Rich Networks with Taxonomy-Informed Representation Learning (viability: 7): https://sciencetostartup.com/paper/learning-hierarchical-knowledge-in-text-rich-networks-with-taxonomy-informed-representation-learning - TIER constructs a hierarchical taxonomy for text-rich networks, enhancing node representations with hierarchical knowledge, enabling more interpretable and structured modeling.
- Are We Winning the Wrong Game? Revisiting Evaluation Practices for Long-Term Time Series Forecasting (viability: 6): https://sciencetostartup.com/paper/are-we-winning-the-wrong-game-revisiting-evaluation-practices-for-long-term-time-series-forecasting - A framework for evaluating time series forecasting models based on statistical fidelity, structural coherence, and decision-level relevance, moving beyond simple error metrics.
- C$^2$FG: Control Classifier-Free Guidance via Score Discrepancy Analysis (viability: 7): https://sciencetostartup.com/paper/c-2-fg-control-classifier-free-guidance-via-score-discrepancy-analysis - C$^2$FG is a training-free, plug-in method that dynamically adjusts Classifier-Free Guidance strength in diffusion models, improving performance across various generative tasks.
- Gender Bias in MT for a Genderless Language: New Benchmarks for Basque (viability: 7): https://sciencetostartup.com/paper/gender-bias-in-mt-for-a-genderless-language-new-benchmarks-for-basque - A benchmark and analysis of gender bias in machine translation for Basque, enabling bias detection and mitigation in MT systems.
- Edged USLAM: Edge-Aware Event-Based SLAM with Learning-Based Depth Priors (viability: 7): https://sciencetostartup.com/paper/edged-uslam-edge-aware-event-based-slam-with-learning-based-depth-priors - Edged USLAM is a hybrid visual-inertial SLAM system that leverages event cameras and learning-based depth priors for robust localization in challenging environments, suitable for aerial navigation.
- Gradually Excavating External Knowledge for Implicit Complex Question Answering (viability: 8): https://sciencetostartup.com/paper/gradually-excavating-external-knowledge-for-implicit-complex-question-answering - A framework for open-domain complex question answering that iteratively acquires external information and reasons based on historical knowledge, achieving SOTA results with smaller models.
- MV-Fashion: Towards Enabling Virtual Try-On and Size Estimation with Multi-View Paired Data (viability: 7): https://sciencetostartup.com/paper/mv-fashion-towards-enabling-virtual-try-on-and-size-estimation-with-multi-view-paired-data - MV-Fashion dataset enables virtual try-on and size estimation by providing paired multi-view fashion data, allowing for the creation of a virtual try-on API.
- Training event-based neural networks with exact gradients via Differentiable ODE Solving in JAX (viability: 7): https://sciencetostartup.com/paper/training-event-based-neural-networks-with-exact-gradients-via-differentiable-ode-solving-in-jax - Eventax enables flexible and accurate training of spiking neural networks using differentiable ODE solvers, offering a prototyping platform for advanced neuron models.
- DARC: Disagreement-Aware Alignment via Risk-Constrained Decoding (viability: 7): https://sciencetostartup.com/paper/darc-disagreement-aware-alignment-via-risk-constrained-decoding - DARC is a retraining-free inference-time method that reranks candidates by maximizing a KL-robust satisfaction objective, enabling explicit risk budgets without retraining.
- Multifingered force-aware control for humanoid robots (viability: 7): https://sciencetostartup.com/paper/multifingered-force-aware-control-for-humanoid-robots - A force-aware control system for multi-fingered robotic hands, enabling stable object manipulation through tactile sensing and force redistribution.
- Mitigating Homophily Disparity in Graph Anomaly Detection: A Scalable and Adaptive Approach (viability: 7): https://sciencetostartup.com/paper/mitigating-homophily-disparity-in-graph-anomaly-detection-a-scalable-and-adaptive-approach - SAGAD is a scalable graph anomaly detection framework that adaptively fuses low- and high-frequency information to identify anomalous nodes in large-scale graphs, offering superior accuracy and scalability.
- VesselFusion: Diffusion Models for Vessel Centerline Extraction from 3D CT Images (viability: 7): https://sciencetostartup.com/paper/vesselfusion-diffusion-models-for-vessel-centerline-extraction-from-3d-ct-images - VesselFusion is a diffusion model for extracting vessel centerlines from 3D CT images, offering improved accuracy and naturalness compared to existing methods, with available code for rapid prototyping.
- UniGround: Universal 3D Visual Grounding via Training-Free Scene Parsing (viability: 7): https://sciencetostartup.com/paper/uniground-universal-3d-visual-grounding-via-training-free-scene-parsing - UniGround enables zero-shot 3D visual grounding by using training-free scene parsing, offering a robust solution for localizing objects in complex 3D environments.
- Explainable Condition Monitoring via Probabilistic Anomaly Detection Applied to Helicopter Transmissions (viability: 7): https://sciencetostartup.com/paper/explainable-condition-monitoring-via-probabilistic-anomaly-detection-applied-to-helicopter-transmissions - Detect and anticipate faults in mechanical systems using probabilistic anomaly detection, trained only on healthy data, with uncertainty quantification for safety-critical applications.
- TRIAGE: Type-Routed Interventions via Aleatoric-Epistemic Gated Estimation in Robotic Manipulation and Adaptive Perception -- Don't Treat All Uncertainty the Same (viability: 8): https://sciencetostartup.com/paper/triage-type-routed-interventions-via-aleatoric-epistemic-gated-estimation-in-robotic-manipulation-and-adaptive-perceptio - Improve robotic manipulation and adaptive perception by decomposing uncertainty into aleatoric and epistemic components, enabling targeted interventions and reducing computational cost.
- EvoScientist: Towards Multi-Agent Evolving AI Scientists for End-to-End Scientific Discovery (viability: 8): https://sciencetostartup.com/paper/evoscientist-towards-multi-agent-evolving-ai-scientists-for-end-to-end-scientific-discovery - EvoScientist is a multi-agent AI scientist framework that evolves research strategies through persistent memory, enabling end-to-end scientific discovery and outperforming existing systems.
- Foley-Flow: Coordinated Video-to-Audio Generation with Masked Audio-Visual Alignment and Dynamic Conditional Flows (viability: 7): https://sciencetostartup.com/paper/foley-flow-coordinated-video-to-audio-generation-with-masked-audio-visual-alignment-and-dynamic-conditional-flows - FoleyFlow generates coordinated audio from video by aligning audio-visual encoders with masked modeling and dynamic conditional flows, surpassing existing benchmarks.
- SaiVLA-0: Cerebrum--Pons--Cerebellum Tripartite Architecture for Compute-Aware Vision-Language-Action (viability: 3): https://sciencetostartup.com/paper/saivla-0-cerebrum-pons-cerebellum-tripartite-architecture-for-compute-aware-vision-language-action - Neuroscience-inspired modular architecture for efficient Vision-Language-Action systems.
- Towards Human-Like Manipulation through RL-Augmented Teleoperation and Mixture-of-Dexterous-Experts VLA (viability: 7): https://sciencetostartup.com/paper/towards-human-like-manipulation-through-rl-augmented-teleoperation-and-mixture-of-dexterous-experts-vla - An RL-augmented teleoperation system with a Mixture-of-Experts VLA architecture for improved dexterous robotic manipulation.
- Model-based Offline RL via Robust Value-Aware Model Learning with Implicitly Differentiable Adaptive Weighting (viability: 7): https://sciencetostartup.com/paper/model-based-offline-rl-via-robust-value-aware-model-learning-with-implicitly-differentiable-adaptive-weighting - ROMI enhances offline reinforcement learning with a robust, value-aware model learning approach, offering controllable conservatism and stable model updates, outperforming RAMBO on D4RL and NeoRL datasets.
- UIS-Digger: Towards Comprehensive Research Agent Systems for Real-world Unindexed Information Seeking (viability: 7): https://sciencetostartup.com/paper/uis-digger-towards-comprehensive-research-agent-systems-for-real-world-unindexed-information-seeking - A multi-agent framework that enhances information-seeking by accessing unindexed web content and files, addressing a critical blind spot in current search engine-reliant systems.
- SAMoE-VLA: A Scene Adaptive Mixture-of-Experts Vision-Language-Action Model for Autonomous Driving (viability: 8): https://sciencetostartup.com/paper/samoe-vla-a-scene-adaptive-mixture-of-experts-vision-language-action-model-for-autonomous-driving - SAMoE-VLA is a scene-adaptive Vision-Language-Action model for autonomous driving that outperforms existing approaches with fewer parameters, offering a safer and more efficient driving experience.
- Tau-BNO: Brain Neural Operator for Tau Transport Model (viability: 7): https://sciencetostartup.com/paper/tau-bno-brain-neural-operator-for-tau-transport-model - Accelerate Alzheimer's research by rapidly simulating tau protein spread using a neural operator surrogate, enabling faster parameter inference and mechanistic discovery.
- Invisible Safety Threat: Malicious Finetuning for LLM via Steganography (viability: 8): https://sciencetostartup.com/paper/invisible-safety-threat-malicious-finetuning-for-llm-via-steganography - Steg-AI provides a security layer for LLMs by detecting steganographically hidden malicious prompts and responses, preventing covert harmful content generation.
- Adaptive MLP Pruning for Large Vision Transformers (viability: 7): https://sciencetostartup.com/paper/adaptive-mlp-pruning-for-large-vision-transformers - Adaptively prune large vision transformers' MLP layers to reduce parameters and FLOPs by 40% without significant performance loss, offering a more efficient model deployment.
- TrianguLang: Geometry-Aware Semantic Consensus for Pose-Free 3D Localization (viability: 8): https://sciencetostartup.com/paper/triangulang-geometry-aware-semantic-consensus-for-pose-free-3d-localization - TrianguLang offers a fast, geometrically consistent, and pose-free 3D localization solution accessible via a single text query, enabling efficient object and part localization for robotics and AR applications.
- DC-W2S: Dual-Consensus Weak-to-Strong Training for Reliable Process Reward Modeling in Biological Reasoning (viability: 7): https://sciencetostartup.com/paper/dc-w2s-dual-consensus-weak-to-strong-training-for-reliable-process-reward-modeling-in-biological-reasoning - Train robust process reward models for scientific reasoning using noisy data with a dual-consensus weak-to-strong framework, enabling reliable step-wise evaluation without exhaustive expert annotation.
- Toward Robust LLM-Based Judges: Taxonomic Bias Evaluation and Debiasing Optimization (viability: 7): https://sciencetostartup.com/paper/toward-robust-llm-based-judges-taxonomic-bias-evaluation-and-debiasing-optimization - A benchmark and debiasing method for improving the reliability of LLM-based judges, enabling more accurate automated evaluation.
- DSH-Bench: A Difficulty- and Scenario-Aware Benchmark with Hierarchical Subject Taxonomy for Subject-Driven Text-to-Image Generation (viability: 7): https://sciencetostartup.com/paper/dsh-bench-a-difficulty-and-scenario-aware-benchmark-with-hierarchical-subject-taxonomy-for-subject-driven-text-to-image- - DSH-Bench provides a comprehensive benchmark for evaluating subject-driven text-to-image models, offering actionable insights for model refinement and optimization.
- From Reactive to Map-Based AI: Tuned Local LLMs for Semantic Zone Inference in Object-Goal Navigation (viability: 7): https://sciencetostartup.com/paper/from-reactive-to-map-based-ai-tuned-local-llms-for-semantic-zone-inference-in-object-goal-navigation - Integrate a fine-tuned Llama-2 model with a mapping system to improve object navigation by inferring semantic zones and optimizing exploration.
- High-Fidelity Pruning for Large Language Models (viability: 8): https://sciencetostartup.com/paper/high-fidelity-pruning-for-large-language-models - Efficiently prune large language models using information entropy to reduce computational costs without sacrificing performance, offering a deployable solution for resource-constrained environments.
- Tiny Autoregressive Recursive Models (viability: 4): https://sciencetostartup.com/paper/tiny-autoregressive-recursive-models - This paper explores the application of Tiny Recursive Models (TRMs) in autoregressive settings, finding limited performance gains compared to simpler refinement baselines, suggesting potential for broader two-step refinement mechanisms but cautioning against TRM-specific architectures.
- MRDrive: An Open Source Mixed Reality Driving Simulator for Automotive User Research (viability: 7): https://sciencetostartup.com/paper/mrdrive-an-open-source-mixed-reality-driving-simulator-for-automotive-user-research - MRDrive is an open-source mixed reality driving simulator that allows researchers to evaluate in-vehicle interfaces with a real vehicle cabin in a virtual environment.
- TALON: Test-time Adaptive Learning for On-the-Fly Category Discovery (viability: 7): https://sciencetostartup.com/paper/talon-test-time-adaptive-learning-for-on-the-fly-category-discovery - TALON enables continuous learning and discovery of novel categories in an online stream by adapting a model at test-time, offering improved accuracy and reduced category explosion.
- Hybrid Quantum Neural Network for Multivariate Clinical Time Series Forecasting (viability: 6): https://sciencetostartup.com/paper/hybrid-quantum-neural-network-for-multivariate-clinical-time-series-forecasting - A hybrid quantum-classical model forecasts physiological signals for proactive patient monitoring, showing robustness in noisy, small-cohort clinical settings.
- Synthetic Defect Image Generation for Power Line Insulator Inspection Using Multimodal Large Language Models (viability: 7): https://sciencetostartup.com/paper/synthetic-defect-image-generation-for-power-line-insulator-inspection-using-multimodal-large-language-models - Generate synthetic defect images using MLLMs to improve defect recognition in low-data regimes, offering a practical solution for industries with limited defect data.
- In-Context Reinforcement Learning for Tool Use in Large Language Models (viability: 7): https://sciencetostartup.com/paper/in-context-reinforcement-learning-for-tool-use-in-large-language-models - ICRL is a reinforcement learning framework that enables LLMs to effectively use external tools without supervised fine-tuning, offering a data-efficient alternative for tool-augmented language models.
- Deterministic Differentiable Structured Pruning for Large Language Models (viability: 7): https://sciencetostartup.com/paper/deterministic-differentiable-structured-pruning-for-large-language-models - DDP offers deterministic, differentiable pruning for LLMs, enabling faster inference with minimal performance loss, making it a valuable tool for optimizing LLM deployment.
- Evaluating Generative Models via One-Dimensional Code Distributions (viability: 8): https://sciencetostartup.com/paper/evaluating-generative-models-via-one-dimensional-code-distributions - Evaluate generative models using discrete visual tokens for improved perceptual quality assessment, offering a training-free and no-reference approach with state-of-the-art correlation to human judgment.
- Enhancing Cross-View UAV Geolocalization via LVLM-Driven Relational Modeling (viability: 7): https://sciencetostartup.com/paper/enhancing-cross-view-uav-geolocalization-via-lvlm-driven-relational-modeling - A plug-and-play ranking architecture leveraging LVLMs for improved UAV-to-satellite image matching, enhancing geolocalization accuracy.
- Adversarial Domain Adaptation Enables Knowledge Transfer Across Heterogeneous RNA-Seq Datasets (viability: 7): https://sciencetostartup.com/paper/adversarial-domain-adaptation-enables-knowledge-transfer-across-heterogeneous-rna-seq-datasets - Transfer learning framework for cancer type classification using RNA-seq data, enabling more accurate diagnosis with limited data.
- ImageEdit-R1: Boosting Multi-Agent Image Editing via Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/imageedit-r1-boosting-multi-agent-image-editing-via-reinforcement-learning - ImageEdit-R1 uses reinforcement learning to coordinate multiple specialized agents for intelligent image editing, outperforming existing closed-source models.
- Stabilized Fine-Tuning with LoRA in Federated Learning: Mitigating the Side Effect of Client Size and Rank via the Scaling Factor (viability: 7): https://sciencetostartup.com/paper/stabilized-fine-tuning-with-lora-in-federated-learning-mitigating-the-side-effect-of-client-size-and-rank-via-the-scalin - SFed-LoRA stabilizes LoRA fine-tuning in federated learning by mitigating gradient collapse, enabling faster convergence and improved stability, making it a practical solution for privacy-preserving LLM adaptation.
- See and Switch: Vision-Based Branching for Interactive Robot-Skill Programming (viability: 8): https://sciencetostartup.com/paper/see-and-switch-vision-based-branching-for-interactive-robot-skill-programming - See & Switch enables intuitive robot programming via vision-based branching, allowing for efficient in-situ recovery demonstrations and conditional task execution, validated by user studies and code release.
- Speed3R: Sparse Feed-forward 3D Reconstruction Models (viability: 7): https://sciencetostartup.com/paper/speed3r-sparse-feed-forward-3d-reconstruction-models - Speed3R accelerates 3D reconstruction by 12.4x using sparse attention, enabling efficient large-scale scene modeling.
- S2S-FDD: Bridging Industrial Time Series and Natural Language for Explainable Zero-shot Fault Diagnosis (viability: 2): https://sciencetostartup.com/paper/s2s-fdd-bridging-industrial-time-series-and-natural-language-for-explainable-zero-shot-fault-diagnosis - A framework that translates industrial sensor data into natural language for explainable fault diagnosis in real-time systems.
- CDRRM: Contrast-Driven Rubric Generation for Reliable and Interpretable Reward Modeling (viability: 8): https://sciencetostartup.com/paper/cdrrm-contrast-driven-rubric-generation-for-reliable-and-interpretable-reward-modeling - CDRRM offers a scalable, interpretable, and data-efficient solution for reward modeling by generating high-quality rubrics to guide preference judgment, enabling better alignment of LLMs with human preferences.
- Solution to the 10th ABAW Expression Recognition Challenge: A Robust Multimodal Framework with Safe Cross-Attention and Modality Dropout (viability: 7): https://sciencetostartup.com/paper/solution-to-the-10th-abaw-expression-recognition-challenge-a-robust-multimodal-framework-with-safe-cross-attention-and-m - A multimodal emotion recognition framework leveraging safe cross-attention and modality dropout to handle missing data and improve accuracy in real-world environments.
- GCGNet: Graph-Consistent Generative Network for Time Series Forecasting with Exogenous Variables (viability: 7): https://sciencetostartup.com/paper/gcgnet-graph-consistent-generative-network-for-time-series-forecasting-with-exogenous-variables - GCGNet leverages graph-based correlation modeling for robust time series forecasting with exogenous variables, offering improved accuracy over existing methods.
- QualiTeacher: Quality-Conditioned Pseudo-Labeling for Real-World Image Restoration (viability: 7): https://sciencetostartup.com/paper/qualiteacher-quality-conditioned-pseudo-labeling-for-real-world-image-restoration - QualiTeacher improves real-world image restoration by conditioning on pseudo-label quality, enabling higher quality results than the teacher model.
- Controllable Complex Human Motion Video Generation via Text-to-Skeleton Cascades (viability: 8): https://sciencetostartup.com/paper/controllable-complex-human-motion-video-generation-via-text-to-skeleton-cascades - Generate controllable human motion videos from text using a cascaded text-to-skeleton and pose-conditioned diffusion model, with a new synthetic dataset to address the lack of training data.
- DyLLM: Efficient Diffusion LLM Inference via Saliency-based Token Selection and Partial Attention (viability: 7): https://sciencetostartup.com/paper/dyllm-efficient-diffusion-llm-inference-via-saliency-based-token-selection-and-partial-attention - DyLLM accelerates masked diffusion language model inference by selectively processing salient tokens, achieving significant throughput gains with minimal accuracy loss.
- ConflictBench: Evaluating Human-AI Conflict via Interactive and Visually Grounded Environments (viability: 7): https://sciencetostartup.com/paper/conflictbench-evaluating-human-ai-conflict-via-interactive-and-visually-grounded-environments - ConflictBench is a benchmark for evaluating and mitigating human-AI conflict in interactive, visually grounded environments, revealing hidden alignment failures.
- Not Like Transformers: Drop the Beat Representation for Dance Generation with Mamba-Based Diffusion Model (viability: 7): https://sciencetostartup.com/paper/not-like-transformers-drop-the-beat-representation-for-dance-generation-with-mamba-based-diffusion-model - Generate plausible dance movements synchronized to music using a Mamba-based diffusion model with a Gaussian beat representation.
- Capacity-Aware Mixture Law Enables Efficient LLM Data Optimization (viability: 7): https://sciencetostartup.com/paper/capacity-aware-mixture-law-enables-efficient-llm-data-optimization - CAMEL optimizes data mixtures for LLM training, reducing costs and improving downstream performance by modeling the interplay between model size and mixture.
- AffordGrasp: Cross-Modal Diffusion for Affordance-Aware Grasp Synthesis (viability: 7): https://sciencetostartup.com/paper/affordgrasp-cross-modal-diffusion-for-affordance-aware-grasp-synthesis - AffordGrasp generates realistic and semantically accurate human grasping poses for AR/VR and embodied AI by using a diffusion-based framework conditioned on object geometry and textual instructions.
- VSDiffusion: Taming Ill-Posed Shadow Generation via Visibility-Constrained Diffusion (viability: 7): https://sciencetostartup.com/paper/vsdiffusion-taming-ill-posed-shadow-generation-via-visibility-constrained-diffusion - VSDiffusion generates realistic cast shadows for image composition by using visibility priors and conditional diffusion, outperforming existing methods.
- Vector Field Augmented Differentiable Policy Learning for Vision-Based Drone Racing (viability: 7): https://sciencetostartup.com/paper/vector-field-augmented-differentiable-policy-learning-for-vision-based-drone-racing - DiffRacing enables autonomous drone racing in complex environments by integrating differentiable losses and vector fields for efficient sim-to-real transfer.
- Missing No More: Dictionary-Guided Cross-Modal Image Fusion under Missing Infrared (viability: 7): https://sciencetostartup.com/paper/missing-no-more-dictionary-guided-cross-modal-image-fusion-under-missing-infrared - A dictionary-guided image fusion technique that addresses missing infrared modality by learning a shared convolutional dictionary and performing coefficient-domain inference, enabling improved perceptual quality and downstream detection.
- Alignment--Process--Outcome: Rethinking How AIs and Humans Collaborate (viability: 2): https://sciencetostartup.com/paper/alignment-process-outcome-rethinking-how-ais-and-humans-collaborate - Reconceptualizing AI-human collaboration through alignment, decision-making, and trajectory structure in task spaces.
- FedMomentum: Preserving LoRA Training Momentum in Federated Fine-Tuning (viability: 7): https://sciencetostartup.com/paper/fedmomentum-preserving-lora-training-momentum-in-federated-fine-tuning - FedMomentum enables structured and momentum-preserving LoRA aggregation via SVD for faster convergence and higher accuracy in federated fine-tuning of LLMs.
- PIRA-Bench: A Transition from Reactive GUI Agents to GUI-based Proactive Intent Recommendation Agents (viability: 7): https://sciencetostartup.com/paper/pira-bench-a-transition-from-reactive-gui-agents-to-gui-based-proactive-intent-recommendation-agents - PIRA-Bench is a benchmark for proactive GUI agents that anticipates user intentions from visual inputs, enabling timely recommendations.
- It's Time to Get It Right: Improving Analog Clock Reading and Clock-Hand Spatial Reasoning in Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/it-s-time-to-get-it-right-improving-analog-clock-reading-and-clock-hand-spatial-reasoning-in-vision-language-models - Improve VLMs' analog clock reading with a new real-world dataset and fine-tuning framework for enhanced spatial-temporal reasoning.
- ViSA-Enhanced Aerial VLN: A Visual-Spatial Reasoning Enhanced Framework for Aerial Vision-Language Navigation (viability: 7): https://sciencetostartup.com/paper/visa-enhanced-aerial-vln-a-visual-spatial-reasoning-enhanced-framework-for-aerial-vision-language-navigation - Enhance aerial navigation by enabling Vision-Language Models to directly reason on image planes, improving success rates by 70%.
- SmartThinker: Progressive Chain-of-Thought Length Calibration for Efficient Large Language Model Reasoning (viability: 7): https://sciencetostartup.com/paper/smartthinker-progressive-chain-of-thought-length-calibration-for-efficient-large-language-model-reasoning - SmartThinker optimizes LLM reasoning by dynamically calibrating chain-of-thought length, achieving significant compression and accuracy improvements.
- Dual-Horizon Hybrid Internal Model for Low-Gravity Quadrupedal Jumping with Hardware-in-the-Loop Validation (viability: 7): https://sciencetostartup.com/paper/dual-horizon-hybrid-internal-model-for-low-gravity-quadrupedal-jumping-with-hardware-in-the-loop-validation - Enable continuous quadrupedal robot jumping on lunar-like terrain with a dual-horizon hybrid internal model and hardware-in-the-loop validation.
- Aero-Promptness: Drag-Aware Aerodynamic Manipulability for Propeller-driven Vehicles (viability: 2): https://sciencetostartup.com/paper/aero-promptness-drag-aware-aerodynamic-manipulability-for-propeller-driven-vehicles - Introduce a geometric framework for optimizing control allocation in multirotor vehicles considering drag and motor limits.
- CMMR-VLN: Vision-and-Language Navigation via Continual Multimodal Memory Retrieval (viability: 3): https://sciencetostartup.com/paper/cmmr-vln-vision-and-language-navigation-via-continual-multimodal-memory-retrieval - CMMR-VLN improves LLM-based vision-and-language navigation by introducing a structured multimodal memory system.
- More to Extract: Discovering MEV by Token Contract Analysis (viability: 7): https://sciencetostartup.com/paper/more-to-extract-discovering-mev-by-token-contract-analysis - Discover and extract previously undetected Maximal Extractable Value (MEV) from token smart contracts using static analysis and constraint solving.
- TeamHOI: Learning a Unified Policy for Cooperative Human-Object Interactions with Any Team Size (viability: 7): https://sciencetostartup.com/paper/teamhoi-learning-a-unified-policy-for-cooperative-human-object-interactions-with-any-team-size - TeamHOI enables a single decentralized policy for cooperative human-object interactions across any number of agents, offering a scalable solution for robotic collaboration.
- On the Feasibility and Opportunity of Autoregressive 3D Object Detection (viability: 7): https://sciencetostartup.com/paper/on-the-feasibility-and-opportunity-of-autoregressive-3d-object-detection - AutoReg3D is an autoregressive 3D object detector that simplifies training and improves extensibility by casting detection as sequence generation, offering a flexible alternative to traditional LiDAR-based methods.
- Extend Your Horizon: A Device-Agnostic Surgical Tool Tracking Framework with Multi-View Optimization for Augmented Reality (viability: 7): https://sciencetostartup.com/paper/extend-your-horizon-a-device-agnostic-surgical-tool-tracking-framework-with-multi-view-optimization-for-augmented-realit - A device-agnostic surgical tool tracking framework using multi-view optimization for augmented reality, enhancing robustness in occluded surgical environments.
- \$OneMillion-Bench: How Far are Language Agents from Human Experts? (viability: 7): https://sciencetostartup.com/paper/onemillion-bench-how-far-are-language-agents-from-human-experts - Evaluate language agents on real-world professional tasks with a new benchmark that requires retrieving authoritative sources and applying domain-specific rules.
- Emergence is Overrated: AGI as an Archipelago of Experts (viability: 5): https://sciencetostartup.com/paper/emergence-is-overrated-agi-as-an-archipelago-of-experts - This paper challenges the notion of AGI requiring emergent intelligence, proposing instead an 'archipelago of experts' approach, suggesting a path to AGI through specialized modules.
- OSExpert: Computer-Use Agents Learning Professional Skills via Exploration (viability: 8): https://sciencetostartup.com/paper/osexpert-computer-use-agents-learning-professional-skills-via-exploration - OSExpert enhances computer-use agents with a GUI-based exploration algorithm, achieving near-expert performance and closing the efficiency gap with humans, making it a valuable tool for automating complex digital tasks.
- ZK-ACE: Identity-Centric Zero-Knowledge Authorization for Post-Quantum Blockchain Systems (viability: 7): https://sciencetostartup.com/paper/zk-ace-identity-centric-zero-knowledge-authorization-for-post-quantum-blockchain-systems - ZK-ACE replaces bulky post-quantum signatures in blockchain transactions with succinct, identity-bound zero-knowledge proofs, significantly reducing on-chain data and enabling efficient authorization.
- VORL-EXPLORE: A Hybrid Learning Planning Approach to Multi-Robot Exploration in Dynamic Environments (viability: 3): https://sciencetostartup.com/paper/vorl-explore-a-hybrid-learning-planning-approach-to-multi-robot-exploration-in-dynamic-environments - Develop a hybrid learning framework to improve multi-robot exploration in dense and dynamic environments.
- Adaptive Collaboration with Humans: Metacognitive Policy Optimization for Multi-Agent LLMs with Continual Learning (viability: 7): https://sciencetostartup.com/paper/adaptive-collaboration-with-humans-metacognitive-policy-optimization-for-multi-agent-llms-with-continual-learning - HILA enables multi-agent LLMs to collaborate with humans by learning when to solve problems autonomously and when to defer to human experts, improving performance on challenging tasks.
- Advancing Automated Algorithm Design via Evolutionary Stagewise Design with LLMs (viability: 8): https://sciencetostartup.com/paper/advancing-automated-algorithm-design-via-evolutionary-stagewise-design-with-llms - EvoStage is an evolutionary algorithm design tool using LLMs that iteratively refines algorithms with real-time feedback, outperforming human experts and achieving state-of-the-art results in chip placement and Bayesian optimization.
- Listening with the Eyes: Benchmarking Egocentric Co-Speech Grounding across Space and Time (viability: 7): https://sciencetostartup.com/paper/listening-with-the-eyes-benchmarking-egocentric-co-speech-grounding-across-space-and-time - EcoG-Bench is a new benchmark for evaluating multimodal models on their ability to ground speech with co-speech gestures in egocentric videos, revealing a significant performance gap compared to human accuracy and highlighting the importance of temporal alignment cues.
- Local Constrained Bayesian Optimization (viability: 7): https://sciencetostartup.com/paper/local-constrained-bayesian-optimization - LCBO offers a novel Bayesian optimization framework for high-dimensional constrained problems, providing a faster convergence rate than existing methods.
- SGG-R$^{\rm 3}$: From Next-Token Prediction to End-to-End Unbiased Scene Graph Generation (viability: 7): https://sciencetostartup.com/paper/sgg-r-rm-3-from-next-token-prediction-to-end-to-end-unbiased-scene-graph-generation - SGG-R$^{\rm 3}$ is a structured reasoning framework that uses chain-of-thought guided fine-tuning and reinforcement learning to improve scene graph generation by addressing sparsity and bias in relation distributions.
- PSTNet: Physically-Structured Turbulence Network (viability: 8): https://sciencetostartup.com/paper/pstnet-physically-structured-turbulence-network - PSTNet is a lightweight, physically-structured neural network for real-time atmospheric turbulence estimation, offering a drop-in replacement for legacy systems in aircraft guidance.
- VisualAD: Language-Free Zero-Shot Anomaly Detection via Vision Transformer (viability: 8): https://sciencetostartup.com/paper/visualad-language-free-zero-shot-anomaly-detection-via-vision-transformer - VisualAD offers a language-free, zero-shot anomaly detection solution using Vision Transformers, outperforming existing methods and adaptable to various pretrained backbones.
- RAPID: Redundancy-Aware and Compatibility-Optimal Edge-Cloud Partitioned Inference for Diverse VLA models (viability: 7): https://sciencetostartup.com/paper/rapid-redundancy-aware-and-compatibility-optimal-edge-cloud-partitioned-inference-for-diverse-vla-models - RAPID is an edge-cloud collaborative inference framework for Vision Language Action models that optimizes partitioning for real-time embodied intelligence applications.
- ELLMob: Event-Driven Human Mobility Generation with Self-Aligned LLM Framework (viability: 8): https://sciencetostartup.com/paper/ellmob-event-driven-human-mobility-generation-with-self-aligned-llm-framework - ELLMob generates realistic human mobility trajectories during large-scale events by reconciling habitual patterns and event constraints using a self-aligned LLM framework, offering valuable insights for urban planning and disaster response.
- Unified Structural-Hydrodynamic Modeling of Underwater Underactuated Mechanisms and Soft Robots (viability: 7): https://sciencetostartup.com/paper/unified-structural-hydrodynamic-modeling-of-underwater-underactuated-mechanisms-and-soft-robots - A trajectory-driven optimization framework for modeling and simulating underwater robots, enabling high-fidelity reproduction of complex behaviors and transferability across different robotic systems.
- $L^3$:Scene-agnostic Visual Localization in the Wild (viability: 7): https://sciencetostartup.com/paper/l-3-scene-agnostic-visual-localization-in-the-wild - Real-time visual localization without pre-built maps, enabling robust and efficient deployment in dynamic environments.
- Condition-Triggered Cryptographic Asset Control via Dormant Authorization Paths (viability: 7): https://sciencetostartup.com/paper/condition-triggered-cryptographic-asset-control-via-dormant-authorization-paths - Implement a secure, conditional digital asset control system with dormant authorization paths for regulatory compliance and secure delegation.
- BRIDGE: Benchmark for multi-hop Reasoning In long multimodal Documents with Grounded Evidence (viability: 7): https://sciencetostartup.com/paper/bridge-benchmark-for-multi-hop-reasoning-in-long-multimodal-documents-with-grounded-evidence - BRIDGE is a benchmark dataset for evaluating multi-hop reasoning in long multimodal documents, enabling targeted diagnosis of reasoning failures in LLMs and RAG systems.
- A Hybrid Vision Transformer Approach for Mathematical Expression Recognition (viability: 8): https://sciencetostartup.com/paper/a-hybrid-vision-transformer-approach-for-mathematical-expression-recognition - A novel Hybrid Vision Transformer for mathematical expression recognition that outperforms state-of-the-art methods, enabling more accurate document analysis.
- Omnidirectional Humanoid Locomotion on Stairs via Unsafe Stepping Penalty and Sparse LiDAR Elevation Mapping (viability: 7): https://sciencetostartup.com/paper/omnidirectional-humanoid-locomotion-on-stairs-via-unsafe-stepping-penalty-and-sparse-lidar-elevation-mapping - A novel humanoid robot locomotion system using LiDAR elevation mapping and unsafe stepping penalties for safe stair traversal, demonstrated in simulation and real-world environments.
- SWE-Fuse: Empowering Software Agents via Issue-free Trajectory Learning and Entropy-aware RLVR Training (viability: 8): https://sciencetostartup.com/paper/swe-fuse-empowering-software-agents-via-issue-free-trajectory-learning-and-entropy-aware-rlvr-training - SWE-Fuse enhances software agent performance by fusing issue-guided and issue-free learning, significantly improving solve rates on real-world software problems.
- IMSE: Intrinsic Mixture of Spectral Experts Fine-tuning for Test-Time Adaptation (viability: 8): https://sciencetostartup.com/paper/imse-intrinsic-mixture-of-spectral-experts-fine-tuning-for-test-time-adaptation - IMSE leverages spectral experts in Vision Transformers for test-time adaptation, offering state-of-the-art performance with minimal parameter updates and domain-aware adaptation.
- SPREAD: Subspace Representation Distillation for Lifelong Imitation Learning (viability: 7): https://sciencetostartup.com/paper/spread-subspace-representation-distillation-for-lifelong-imitation-learning - SPREAD is a geometry-preserving framework that uses SVD to align policy representations across tasks, improving knowledge transfer and mitigating catastrophic forgetting in lifelong imitation learning.
- Semantic Risk Scoring of Aggregated Metrics: An AI-Driven Approach for Healthcare Data Governance (viability: 7): https://sciencetostartup.com/paper/semantic-risk-scoring-of-aggregated-metrics-an-ai-driven-approach-for-healthcare-data-governance - AI-powered tool to proactively identify and mitigate privacy risks in SQL-based healthcare metric definitions before deployment, enabling secure data sharing and compliance.
- Robust Transfer Learning with Side Information (viability: 7): https://sciencetostartup.com/paper/robust-transfer-learning-with-side-information - Improve reinforcement learning transfer by using side information to create tighter uncertainty sets for robust policies, leading to better performance in shifted environments.
- RLPR: Radar-to-LiDAR Place Recognition via Two-Stage Asymmetric Cross-Modal Alignment for Autonomous Driving (viability: 7): https://sciencetostartup.com/paper/rlpr-radar-to-lidar-place-recognition-via-two-stage-asymmetric-cross-modal-alignment-for-autonomous-driving - RLPR is a robust radar-to-LiDAR place recognition framework that enables all-weather autonomy for autonomous driving by aligning radar scans with existing LiDAR maps, achieving state-of-the-art accuracy and zero-shot generalization.
- Enhancing Unregistered Hyperspectral Image Super-Resolution via Unmixing-based Abundance Fusion Learning (viability: 7): https://sciencetostartup.com/paper/enhancing-unregistered-hyperspectral-image-super-resolution-via-unmixing-based-abundance-fusion-learning - Enhance hyperspectral image resolution by fusing low-resolution HSI with unregistered high-resolution reference images using unmixing-based abundance fusion learning.
- Rel-MOSS: Towards Imbalanced Relational Deep Learning on Relational Databases (viability: 7): https://sciencetostartup.com/paper/rel-moss-towards-imbalanced-relational-deep-learning-on-relational-databases - Rel-MOSS addresses class imbalance in relational databases using GNNs, offering improved entity classification accuracy.
- Ares: Adaptive Reasoning Effort Selection for Efficient LLM Agents (viability: 4): https://sciencetostartup.com/paper/ares-adaptive-reasoning-effort-selection-for-efficient-llm-agents - Ares optimizes LLM agents' reasoning efforts to reduce costs and maintain task success rates.
- Geometric Transformation-Embedded Mamba for Learned Video Compression (viability: 7): https://sciencetostartup.com/paper/geometric-transformation-embedded-mamba-for-learned-video-compression - A novel video compression framework using Mamba modules and geometric transformations for improved perceptual quality and temporal consistency, outperforming existing methods at low bitrates.
- Beyond Heuristic Prompting: A Concept-Guided Bayesian Framework for Zero-Shot Image Recognition (viability: 7): https://sciencetostartup.com/paper/beyond-heuristic-prompting-a-concept-guided-bayesian-framework-for-zero-shot-image-recognition - Improve zero-shot image recognition by using LLMs to generate class-specific concepts and a Bayesian framework to marginalize over the concept space.
- Long-Short Term Agents for Pure-Vision Bronchoscopy Robotic Autonomy (viability: 2): https://sciencetostartup.com/paper/long-short-term-agents-for-pure-vision-bronchoscopy-robotic-autonomy - Vision-only autonomy framework for preclinical feasibility of autonomous bronchoscopic navigation without external tracking.
- NaviDriveVLM: Decoupling High-Level Reasoning and Motion Planning for Autonomous Driving (viability: 7): https://sciencetostartup.com/paper/navidrivevlm-decoupling-high-level-reasoning-and-motion-planning-for-autonomous-driving - NaviDriveVLM decouples reasoning and motion planning in autonomous driving, enabling efficient and interpretable end-to-end control.
- EveryQuery: Zero-Shot Clinical Prediction via Task-Conditioned Pretraining over Electronic Health Records (viability: 7): https://sciencetostartup.com/paper/everyquery-zero-shot-clinical-prediction-via-task-conditioned-pretraining-over-electronic-health-records - EveryQuery enables zero-shot clinical prediction by directly estimating outcome likelihoods from patient history and structured queries, outperforming autoregressive models, especially for rare events.
- Bayesian Transformer for Probabilistic Load Forecasting in Smart Grids (viability: 8): https://sciencetostartup.com/paper/bayesian-transformer-for-probabilistic-load-forecasting-in-smart-grids - Bayesian Transformer provides calibrated probabilistic load forecasting for smart grids, enabling better risk management and demand response.
- Revisiting Unknowns: Towards Effective and Efficient Open-Set Active Learning (viability: 7): https://sciencetostartup.com/paper/revisiting-unknowns-towards-effective-and-efficient-open-set-active-learning - E2OAL is a unified, detector-free framework for open-set active learning that improves accuracy and efficiency by leveraging labeled unknowns, making it suitable for real-world applications.
- SMGI: A Structural Theory of General Artificial Intelligence (viability: 4): https://sciencetostartup.com/paper/smgi-a-structural-theory-of-general-artificial-intelligence - SMGI is a structural theory of general artificial intelligence that recasts learning as the controlled evolution of the learning interface itself.
- MINT: Molecularly Informed Training with Spatial Transcriptomics Supervision for Pathology Foundation Models (viability: 7): https://sciencetostartup.com/paper/mint-molecularly-informed-training-with-spatial-transcriptomics-supervision-for-pathology-foundation-models - MINT fine-tunes pathology Vision Transformers with spatial transcriptomics to improve gene expression prediction and general pathology tasks.
- Designing probabilistic AI monsoon forecasts to inform agricultural decision-making (viability: 9): https://sciencetostartup.com/paper/designing-probabilistic-ai-monsoon-forecasts-to-inform-agricultural-decision-making - AI-powered monsoon forecasting system deployed to 38 million farmers, providing tailored insights for planting decisions and climate adaptation.
- RoboRouter: Training-Free Policy Routing for Robotic Manipulation (viability: 7): https://sciencetostartup.com/paper/roborouter-training-free-policy-routing-for-robotic-manipulation - RoboRouter intelligently routes robotic manipulation tasks to the best-performing policy from a pool of heterogeneous policies, improving success rates without retraining.
- A Lightweight Traffic Map for Efficient Anytime LaCAM* (viability: 7): https://sciencetostartup.com/paper/a-lightweight-traffic-map-for-efficient-anytime-lacam - Improve multi-agent pathfinding with a dynamic traffic map that enhances solution quality and reduces computational overhead, ready for integration into existing MAPF systems.
- Visualizing Coalition Formation: From Hedonic Games to Image Segmentation (viability: 6): https://sciencetostartup.com/paper/visualizing-coalition-formation-from-hedonic-games-to-image-segmentation - Utilize coalition formation in hedonic games for image segmentation, offering a novel approach to pixel grouping and boundary detection.
- VLM-SubtleBench: How Far Are VLMs from Human-Level Subtle Comparative Reasoning? (viability: 7): https://sciencetostartup.com/paper/vlm-subtlebench-how-far-are-vlms-from-human-level-subtle-comparative-reasoning - A benchmark to evaluate VLMs on subtle comparative reasoning across diverse domains, revealing gaps in model performance and providing a foundation for advancement.
- Reject, Resample, Repeat: Understanding Parallel Reasoning in Language Model Inference (viability: 5): https://sciencetostartup.com/paper/reject-resample-repeat-understanding-parallel-reasoning-in-language-model-inference - Improve LLM inference accuracy and cost tradeoff by using particle filtering algorithms.
- CCR-Bench: A Comprehensive Benchmark for Evaluating LLMs on Complex Constraints, Control Flows, and Real-World Cases (viability: 7): https://sciencetostartup.com/paper/ccr-bench-a-comprehensive-benchmark-for-evaluating-llms-on-complex-constraints-control-flows-and-real-world-cases - CCR-Bench is a new benchmark to evaluate LLMs on complex, real-world instructions, highlighting performance gaps and guiding future model development.
- Identifying Influential Actions in Human-Robot Interactions (viability: 7): https://sciencetostartup.com/paper/identifying-influential-actions-in-human-robot-interactions - Identify key robot actions influencing human behavior to improve robot design and adaptability, potentially leading to more natural and effective human-robot interactions.
- Choose What to Observe: Task-Aware Semantic-Geometric Representations for Visuomotor Policy (viability: 7): https://sciencetostartup.com/paper/choose-what-to-observe-task-aware-semantic-geometric-representations-for-visuomotor-policy - Improve robot manipulation robustness with a task-aware semantic-geometric representation that canonicalizes visual input, enabling policies to generalize better to out-of-distribution appearance changes.
- Toward Unified Multimodal Representation Learning for Autonomous Driving (viability: 7): https://sciencetostartup.com/paper/toward-unified-multimodal-representation-learning-for-autonomous-driving - Unified multimodal representation learning framework for autonomous driving that aligns text, images, and point clouds in a shared embedding space, improving scene understanding.
- Hospitality-VQA: Decision-Oriented Informativeness Evaluation for Vision-Language Models (viability: 6): https://sciencetostartup.com/paper/hospitality-vqa-decision-oriented-informativeness-evaluation-for-vision-language-models - A hospitality-specific VQA dataset and informativeness framework to improve VLMs for decision-making in the hospitality domain.
- Slumbering to Precision: Enhancing Artificial Neural Network Calibration Through Sleep-like Processes (viability: 3): https://sciencetostartup.com/paper/slumbering-to-precision-enhancing-artificial-neural-network-calibration-through-sleep-like-processes - SRC enhances neural network calibration through a novel sleep-like post-training phase that updates internal weights for more reliable confidence estimates.
- Viewpoint-Agnostic Grasp Pipeline using VLM and Partial Observations (viability: 7): https://sciencetostartup.com/paper/viewpoint-agnostic-grasp-pipeline-using-vlm-and-partial-observations - An end-to-end pipeline for language-guided robotic grasping that improves robustness to occlusions and partial observations.
- An Interpretable Generative Framework for Anomaly Detection in High-Dimensional Financial Time Series (viability: 7): https://sciencetostartup.com/paper/an-interpretable-generative-framework-for-anomaly-detection-in-high-dimensional-financial-time-series - ReGEN-TAD is an interpretable generative framework for anomaly detection in financial time series, offering improved robustness and factor-level attribution.
- SynPlanResearch-R1: Encouraging Tool Exploration for Deep Research with Synthetic Plans (viability: 7): https://sciencetostartup.com/paper/synplanresearch-r1-encouraging-tool-exploration-for-deep-research-with-synthetic-plans - SynPlanResearch-R1 improves research agent performance by synthesizing tool-use trajectories for better exploration, offering a strong initialization for reinforcement learning.
- Intentional Deception as Controllable Capability in LLM Agents (viability: 2): https://sciencetostartup.com/paper/intentional-deception-as-controllable-capability-in-llm-agents - Develop a system to understand and protect against intentional deception by LLM agents in multi-agent environments.
- An Efficient and Effective Evaluator for Text2SQL Models on Unseen and Unlabeled Data (viability: 8): https://sciencetostartup.com/paper/an-efficient-and-effective-evaluator-for-text2sql-models-on-unseen-and-unlabeled-data - FusionSQL enables rapid evaluation and monitoring of Text2SQL models on unseen data, ensuring quality and timely detection of performance degradation in evolving database environments.
- Training-free Temporal Object Tracking in Surgical Videos (viability: 7): https://sciencetostartup.com/paper/training-free-temporal-object-tracking-in-surgical-videos - Leveraging pre-trained diffusion models for training-free object tracking in surgical videos, enabling real-time surgical assistance and analysis.
- AI Steerability 360: A Toolkit for Steering Large Language Models (viability: 7): https://sciencetostartup.com/paper/ai-steerability-360-a-toolkit-for-steering-large-language-models - AI Steerability 360 is an open-source Python library for steering LLMs through prompt, structural, state, and output modifications, enabling comprehensive evaluation and composition of steering methods.
- DistillGuard: Evaluating Defenses Against LLM Knowledge Distillation (viability: 6): https://sciencetostartup.com/paper/distillguard-evaluating-defenses-against-llm-knowledge-distillation - DistillGuard is a framework for evaluating defenses against LLM knowledge distillation attacks, revealing the ineffectiveness of current output-level approaches and highlighting the need for more robust security measures.
- AI Misuse in Education Is a Measurement Problem: Toward a Learning Visibility Framework (viability: 3): https://sciencetostartup.com/paper/ai-misuse-in-education-is-a-measurement-problem-toward-a-learning-visibility-framework - Develop a framework to enhance AI integration in education by increasing transparency and process visibility rather than focusing on detection.
- Gradient Iterated Temporal-Difference Learning (viability: 7): https://sciencetostartup.com/paper/gradient-iterated-temporal-difference-learning - Gradient Iterated Temporal-Difference Learning offers a more stable and competitive alternative to semi-gradient TD learning, potentially improving RL agent training.
- GazeShift: Unsupervised Gaze Estimation and Dataset for VR (viability: 8): https://sciencetostartup.com/paper/gazeshift-unsupervised-gaze-estimation-and-dataset-for-vr - GazeShift provides a real-time, unsupervised gaze estimation solution for VR, complete with a large-scale dataset and code release, enabling accurate and efficient gaze tracking in VR applications.
- Transferable Optimization Network for Cross-Domain Image Reconstruction (viability: 7): https://sciencetostartup.com/paper/transferable-optimization-network-for-cross-domain-image-reconstruction - A transfer learning framework for image reconstruction, enabling high-quality results with limited data by leveraging diverse datasets, ideal for medical imaging applications.
- Physics-infused Learning for Aerial Manipulator in Winds and Near-Wall Environments (viability: 7): https://sciencetostartup.com/paper/physics-infused-learning-for-aerial-manipulator-in-winds-and-near-wall-environments - A physics-informed learning approach for aerial manipulators enables robust operation in windy, near-wall environments, opening up infrastructure inspection and maintenance applications.
- Benchmarking Large Language Models for Quebec Insurance: From Closed-Book to Retrieval-Augmented Generation (viability: 7): https://sciencetostartup.com/paper/benchmarking-large-language-models-for-quebec-insurance-from-closed-book-to-retrieval-augmented-generation - Benchmark and RAG-based system for Quebec insurance advice, leveraging LLMs for automated advisory services.
- Reasoning Knowledge-Gap in Drone Planning via LLM-based Active Elicitation (viability: 7): https://sciencetostartup.com/paper/reasoning-knowledge-gap-in-drone-planning-via-llm-based-active-elicitation - An LLM-powered drone planning system that actively elicits minimal information from human operators to improve task success and reduce interaction frequency.
- Uncertainty Mitigation and Intent Inference: A Dual-Mode Human-Machine Joint Planning System (viability: 7): https://sciencetostartup.com/paper/uncertainty-mitigation-and-intent-inference-a-dual-mode-human-machine-joint-planning-system - An LLM-assisted human-robot collaboration system that actively resolves uncertainties and infers human intent for improved joint planning.
- Fusion Complexity Inversion: Why Simpler Cross View Modules Outperform SSMs and Cross View Attention Transformers for Pasture Biomass Regression (viability: 7): https://sciencetostartup.com/paper/fusion-complexity-inversion-why-simpler-cross-view-modules-outperform-ssms-and-cross-view-attention-transformers-for-pas - A simple, efficient method for pasture biomass regression using agricultural imagery, outperforming complex models on scarce data.
- HybridStitch: Pixel and Timestep Level Model Stitching for Diffusion Acceleration (viability: 7): https://sciencetostartup.com/paper/hybridstitch-pixel-and-timestep-level-model-stitching-for-diffusion-acceleration - HybridStitch accelerates text-to-image generation by selectively applying large and small diffusion models to different image regions, achieving significant speedups.
- Neural Precoding in Complex Projective Spaces (viability: 7): https://sciencetostartup.com/paper/neural-precoding-in-complex-projective-spaces - A deep learning framework for wireless precoding that uses complex projective space parameterizations to improve sum-rate performance and generalization.
- Preference-Conditioned Reinforcement Learning for Space-Time Efficient Online 3D Bin Packing (viability: 8): https://sciencetostartup.com/paper/preference-conditioned-reinforcement-learning-for-space-time-efficient-online-3d-bin-packing - STEP optimizes robotic bin packing by using a preference-conditioned Transformer-based RL policy to balance space utilization and operational time, achieving a 44% reduction in operational time without compromising packing density.
- MWM: Mobile World Models for Action-Conditioned Consistent Prediction (viability: 8): https://sciencetostartup.com/paper/mwm-mobile-world-models-for-action-conditioned-consistent-prediction - MWM is a mobile world model that improves action-conditioned rollout consistency for planning-based image-goal navigation, enabling more accurate and efficient robot navigation.
- Inverse Resistive Force Theory (I-RFT): Learning granular properties through robot-terrain physical interactions (viability: 7): https://sciencetostartup.com/paper/inverse-resistive-force-theory-i-rft-learning-granular-properties-through-robot-terrain-physical-interactions - A physics-informed machine learning framework that estimates terrain properties from robot-terrain interactions, enabling robots to optimize foot design and gait trajectories for efficient information gathering.
- A Robust Antenna Provides Tactile Feedback in a Multi-legged Robot (viability: 7): https://sciencetostartup.com/paper/a-robust-antenna-provides-tactile-feedback-in-a-multi-legged-robot - Tactile antennae enable multi-legged robots to navigate complex environments by providing real-time collision feedback for autonomous steering and recovery.
- 4DRC-OCC: Robust Semantic Occupancy Prediction Through Fusion of 4D Radar and Camera (viability: 7): https://sciencetostartup.com/paper/4drc-occ-robust-semantic-occupancy-prediction-through-fusion-of-4d-radar-and-camera - A robust 3D semantic occupancy prediction system fusing 4D radar and camera data for autonomous driving in adverse conditions, with an automatically labeled dataset.
- Dual-Metric Evaluation of Social Bias in Large Language Models: Evidence from an Underrepresented Nepali Cultural Context (viability: 7): https://sciencetostartup.com/paper/dual-metric-evaluation-of-social-bias-in-large-language-models-evidence-from-an-underrepresented-nepali-cultural-context - Evaluate and mitigate social bias in LLMs within the underrepresented Nepali cultural context using a novel dual-metric framework.
- SGI: Structured 2D Gaussians for Efficient and Compact Large Image Representation (viability: 8): https://sciencetostartup.com/paper/sgi-structured-2d-gaussians-for-efficient-and-compact-large-image-representation - SGI offers a compact and efficient image representation framework using structured 2D Gaussians, enabling significant compression and faster optimization for high-resolution images, ideal for low-end devices.
- OrdinalBench: A Benchmark Dataset for Diagnosing Generalization Limits in Ordinal Number Understanding of Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/ordinalbench-a-benchmark-dataset-for-diagnosing-generalization-limits-in-ordinal-number-understanding-of-vision-language - OrdinalBench is a diagnostic benchmark and toolkit for evaluating and improving the ordinal reasoning capabilities of Vision-Language Models, enabling more robust and reliable performance in tasks requiring sequential understanding.
- ProgAgent:A Continual RL Agent with Progress-Aware Rewards (viability: 8): https://sciencetostartup.com/paper/progagent-a-continual-rl-agent-with-progress-aware-rewards - ProgAgent is a continual reinforcement learning agent that learns from unlabeled expert videos and adapts to new tasks, offering a robust solution for lifelong robotic learning.
- Scaling Data Difficulty: Improving Coding Models via Reinforcement Learning on Fresh and Challenging Problems (viability: 7): https://sciencetostartup.com/paper/scaling-data-difficulty-improving-coding-models-via-reinforcement-learning-on-fresh-and-challenging-problems - MicroCoder is a curated dataset of challenging programming problems that significantly improves code generation model performance, offering a focused training resource for advanced coding tasks.
- Breaking Training Bottlenecks: Effective and Stable Reinforcement Learning for Coding Models (viability: 8): https://sciencetostartup.com/paper/breaking-training-bottlenecks-effective-and-stable-reinforcement-learning-for-coding-models - MicroCoder-GRPO improves code generation model training with innovations for stability, diversity, and efficiency, outperforming baselines and offering a new dataset and evaluator for enhanced performance.
- Geometric Knowledge-Assisted Federated Dual Knowledge Distillation Approach Towards Remote Sensing Satellite Imagery (viability: 7): https://sciencetostartup.com/paper/geometric-knowledge-assisted-federated-dual-knowledge-distillation-approach-towards-remote-sensing-satellite-imagery - A federated learning framework leveraging geometric knowledge distillation for improved remote sensing satellite image analysis, demonstrating significant performance gains.
- ArcLight: A Lightweight LLM Inference Architecture for Many-Core CPUs (viability: 7): https://sciencetostartup.com/paper/arclight-a-lightweight-llm-inference-architecture-for-many-core-cpus - ArcLight is a CPU-optimized LLM inference architecture that maximizes throughput on many-core CPUs by minimizing cross-NUMA memory access overhead.
- MedQ-Deg: A Multidimensional Benchmark for Evaluating MLLMs Across Medical Image Quality Degradations (viability: 7): https://sciencetostartup.com/paper/medq-deg-a-multidimensional-benchmark-for-evaluating-mllms-across-medical-image-quality-degradations - A benchmark dataset and evaluation framework for assessing the robustness of medical MLLMs to image quality degradations, enabling the development of more reliable clinical AI systems.
- QuadAI at SemEval-2026 Task 3: Ensemble Learning of Hybrid RoBERTa and LLMs for Dimensional Aspect-Based Sentiment Analysis (viability: 7): https://sciencetostartup.com/paper/quadai-at-semeval-2026-task-3-ensemble-learning-of-hybrid-roberta-and-llms-for-dimensional-aspect-based-sentiment-analys - Ensemble learning approach combining RoBERTa and LLMs for improved dimensional aspect-based sentiment analysis, with code available for development.
- Using GPUs And LLMs Can Be Satisfying for Nonlinear Real Arithmetic Problems (viability: 8): https://sciencetostartup.com/paper/using-gpus-and-llms-can-be-satisfying-for-nonlinear-real-arithmetic-problems - GANRA is a novel SMT solver that leverages LLMs and GPU acceleration to efficiently solve quantifier-free non-linear real arithmetic problems, outperforming the state-of-the-art by a significant margin.
- DECADE: A Temporally-Consistent Unsupervised Diffusion Model for Enhanced Rb-82 Dynamic Cardiac PET Image Denoising (viability: 7): https://sciencetostartup.com/paper/decade-a-temporally-consistent-unsupervised-diffusion-model-for-enhanced-rb-82-dynamic-cardiac-pet-image-denoising - DECADE is an unsupervised diffusion model for denoising Rb-82 dynamic cardiac PET images, improving image quality and quantitative accuracy for CAD diagnosis.
- Whitening Reveals Cluster Commitment as the Geometric Separator of Hallucination Types (viability: 6): https://sciencetostartup.com/paper/whitening-reveals-cluster-commitment-as-the-geometric-separator-of-hallucination-types - A method to identify and categorize hallucination types in LLMs using PCA-whitening, potentially leading to better evaluation metrics and mitigation strategies.
- 3ViewSense: Spatial and Mental Perspective Reasoning from Orthographic Views in Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/3viewsense-spatial-and-mental-perspective-reasoning-from-orthographic-views-in-vision-language-models - 3ViewSense enhances spatial reasoning in vision-language models by using orthographic views to bridge the spatial intelligence gap, offering a more stable and consistent spatial understanding.
- AeroPlace-Flow: Language-Grounded Object Placement for Aerial Manipulators via Visual Foresight and Object Flow (viability: 7): https://sciencetostartup.com/paper/aeroplace-flow-language-grounded-object-placement-for-aerial-manipulators-via-visual-foresight-and-object-flow - AeroPlace-Flow enables language-guided aerial object placement using visual foresight and object flow, offering a training-free solution for real-world aerial manipulation tasks.
- Hide and Find: A Distributed Adversarial Attack on Federated Graph Learning (viability: 7): https://sciencetostartup.com/paper/hide-and-find-a-distributed-adversarial-attack-on-federated-graph-learning - FedShift is a stealthy and efficient adversarial attack method for federated graph learning, offering a robust solution for evaluating and improving the security of FedGL systems.
- Large Language Model for Discrete Optimization Problems: Evaluation and Step-by-step Reasoning (viability: 7): https://sciencetostartup.com/paper/large-language-model-for-discrete-optimization-problems-evaluation-and-step-by-step-reasoning - Leverage Llama-3 and ChatGPT to solve discrete optimization problems with a new benchmark dataset and insights into Chain-of-Thought effectiveness.
- A Novel Multi-Agent Architecture to Reduce Hallucinations of Large Language Models in Multi-Step Structural Modeling (viability: 8): https://sciencetostartup.com/paper/a-novel-multi-agent-architecture-to-reduce-hallucinations-of-large-language-models-in-multi-step-structural-modeling - Automate structural modeling and analysis with a multi-agent architecture that reduces hallucinations in LLMs, achieving near-perfect accuracy on benchmark problems.
- Post-quantum Federated Learning: Secure And Scalable Threat Intelligence For Collaborative Cyber Defense (viability: 7): https://sciencetostartup.com/paper/post-quantum-federated-learning-secure-and-scalable-threat-intelligence-for-collaborative-cyber-defense - A quantum-secure federated learning framework for collaborative threat intelligence, protecting cross-organizational data sharing against quantum attacks.
- A Lightweight MPC Bidding Framework for Brand Auction Ads (viability: 7): https://sciencetostartup.com/paper/a-lightweight-mpc-bidding-framework-for-brand-auction-ads - A lightweight MPC framework for brand auction ads that uses online isotonic regression to improve spend efficiency and cost control.
- SoK: The Evolution of Maximal Extractable Value, From Miners to Cross-Chain (viability: 6): https://sciencetostartup.com/paper/sok-the-evolution-of-maximal-extractable-value-from-miners-to-cross-chain - A comprehensive analysis of Maximal Extractable Value (MEV) in blockchain systems, providing a unified chronological framework and identifying future research directions.
- Reverse Distillation: Consistently Scaling Protein Language Model Representations (viability: 7): https://sciencetostartup.com/paper/reverse-distillation-consistently-scaling-protein-language-model-representations - Reverse Distillation consistently improves protein language model performance by distilling knowledge from larger models into smaller, more efficient representations, enabling better protein feature extraction.
- YAQIN: Culturally Sensitive, Agentic AI for Mental Healthcare Support Among Muslim Women in the UK (viability: 7): https://sciencetostartup.com/paper/yaqin-culturally-sensitive-agentic-ai-for-mental-healthcare-support-among-muslim-women-in-the-uk - YAQIN is a culturally sensitive AI chatbot and journaling tool designed to improve mental healthcare access and trust for Muslim women in the UK.
- PARSE: Part-Aware Relational Spatial Modeling (viability: 7): https://sciencetostartup.com/paper/parse-part-aware-relational-spatial-modeling - PARSE enhances 3D scene generation by modeling part-level object interactions, leading to physically realistic and structurally complex scenes.
- TDM-R1: Reinforcing Few-Step Diffusion Models with Non-Differentiable Reward (viability: 8): https://sciencetostartup.com/paper/tdm-r1-reinforcing-few-step-diffusion-models-with-non-differentiable-reward - TDM-R1 is a reinforcement learning method that improves few-step text-to-image models with non-differentiable rewards, achieving state-of-the-art performance and scaling effectively to strong generative models.
- EDMFormer: Genre-Specific Self-Supervised Learning for Music Structure Segmentation (viability: 7): https://sciencetostartup.com/paper/edmformer-genre-specific-self-supervised-learning-for-music-structure-segmentation - EDMFormer provides improved music structure segmentation for EDM by using self-supervised learning on a new genre-specific dataset, enabling better analysis and potential for automated DJing or music production tools.
- C$^2$-Explorer: Contiguity-Driven Task Allocation with Connectivity-Aware Task Representation for Decentralized Multi-UAV Exploration (viability: 8): https://sciencetostartup.com/paper/c-2-explorer-contiguity-driven-task-allocation-with-connectivity-aware-task-representation-for-decentralized-multi-uav-e - C$^2$-Explorer is a decentralized multi-UAV exploration framework that significantly reduces exploration time and path length by using a connectivity graph and contiguity-driven task allocation, ready for commercial drone applications.
- Learning Context-Adaptive Motion Priors for Masked Motion Diffusion Models with Efficient Kinematic Attention Aggregation (viability: 7): https://sciencetostartup.com/paper/learning-context-adaptive-motion-priors-for-masked-motion-diffusion-models-with-efficient-kinematic-attention-aggregatio - MMDM is a diffusion-based generative reconstruction framework that enhances incomplete or low-confidence motion data using partially available high-quality reconstructions within a Masked Autoencoder architecture.
- Compressed-Domain-Aware Online Video Super-Resolution (viability: 7): https://sciencetostartup.com/paper/compressed-domain-aware-online-video-super-resolution - CDA-VSR leverages compressed-domain information for real-time online video super-resolution, balancing quality and efficiency.
- RoboPCA: Pose-centered Affordance Learning from Human Demonstrations for Robot Manipulation (viability: 7): https://sciencetostartup.com/paper/robopca-pose-centered-affordance-learning-from-human-demonstrations-for-robot-manipulation - RoboPCA predicts task-appropriate contact regions and poses for robot manipulation by learning from human demonstrations, enabling more effective object interaction.
- FrameVGGT: Frame Evidence Rolling Memory for streaming VGGT (viability: 7): https://sciencetostartup.com/paper/framevggt-frame-evidence-rolling-memory-for-streaming-vggt - FrameVGGT offers a memory-efficient approach to streaming visual geometry transformers, enabling robust online 3D perception for long video streams.
- Mitigating the Memory Bottleneck with Machine Learning-Driven and Data-Aware Microarchitectural Techniques (viability: 7): https://sciencetostartup.com/paper/mitigating-the-memory-bottleneck-with-machine-learning-driven-and-data-aware-microarchitectural-techniques - Optimize memory access in existing processors by applying lightweight machine learning to learn access patterns and predict off-chip memory requests.
- Low-Cost Teleoperation Extension for Mobile Manipulators (viability: 7): https://sciencetostartup.com/paper/low-cost-teleoperation-extension-for-mobile-manipulators - An open-source teleoperation framework enabling intuitive whole-body control of mobile manipulators using commodity hardware like smartphones and foot pedals.
- Memory for Autonomous LLM Agents:Mechanisms, Evaluation, and Emerging Frontiers (viability: 7): https://sciencetostartup.com/paper/memory-for-autonomous-llm-agents-mechanisms-evaluation-and-emerging-frontiers - A comprehensive survey of memory mechanisms in LLM agents, providing a blueprint for building adaptive and context-aware AI assistants.
- FusionRegister: Every Infrared and Visible Image Fusion Deserves Registration (viability: 7): https://sciencetostartup.com/paper/fusionregister-every-infrared-and-visible-image-fusion-deserves-registration - FusionRegister enhances infrared and visible image fusion by learning misregistration representations, improving detail alignment and robustness without extensive pre-processing.
- Ref-DGS: Reflective Dual Gaussian Splatting (viability: 7): https://sciencetostartup.com/paper/ref-dgs-reflective-dual-gaussian-splatting - Ref-DGS enables fast and accurate novel view synthesis of reflective surfaces by decoupling surface reconstruction from specular reflection using a dual Gaussian representation.
- DAISS: Phase-Aware Imitation Learning for Dual-Arm Robotic Ultrasound-Guided Interventions (viability: 7): https://sciencetostartup.com/paper/daiss-phase-aware-imitation-learning-for-dual-arm-robotic-ultrasound-guided-interventions - DAISS is a dual-arm robotic system that learns to automate ultrasound-guided needle insertion using imitation learning, potentially improving precision and reducing workload in medical interventions.
- Holi-Spatial: Evolving Video Streams into Holistic 3D Spatial Intelligence (viability: 8): https://sciencetostartup.com/paper/holi-spatial-evolving-video-streams-into-holistic-3d-spatial-intelligence - Holi-Spatial is a large-scale, automatically generated 3D spatial dataset that significantly improves performance on spatial reasoning tasks, enabling enhanced 3D scene understanding applications.
- Scaling Test-Time Robustness of Vision-Language Models via Self-Critical Inference Framework (viability: 7): https://sciencetostartup.com/paper/scaling-test-time-robustness-of-vision-language-models-via-self-critical-inference-framework - Improve the robustness of vision-language models against language bias and sensitivity with a self-critical inference framework and dynamic robustness benchmark.
- GLASS: Graph and Vision-Language Assisted Semantic Shape Correspondence (viability: 7): https://sciencetostartup.com/paper/glass-graph-and-vision-language-assisted-semantic-shape-correspondence - GLASS leverages vision-language models and graph-assisted contrastive learning to establish dense 3D shape correspondence, enabling applications like texture transfer and robotic manipulation.
- Multi-Agent Off-World Exploration for Sparse Evidence Discovery via Gaussian Belief Mapping and Dual-Domain Coverage (viability: 7): https://sciencetostartup.com/paper/multi-agent-off-world-exploration-for-sparse-evidence-discovery-via-gaussian-belief-mapping-and-dual-domain-coverage - A multi-agent robotic exploration framework using Gaussian belief mapping for efficient and safe evidence discovery in sparse, hazardous environments.
- AtomicVLA: Unlocking the Potential of Atomic Skill Learning in Robots (viability: 8): https://sciencetostartup.com/paper/atomicvla-unlocking-the-potential-of-atomic-skill-learning-in-robots - AtomicVLA enables robots to learn and execute long-horizon tasks by decomposing them into atomic skills, offering a scalable solution for continual learning in real-world environments.
- Evaluating Synthetic Data for Baggage Trolley Detection in Airport Logistics (viability: 7): https://sciencetostartup.com/paper/evaluating-synthetic-data-for-baggage-trolley-detection-in-airport-logistics - Generate synthetic data using NVIDIA Omniverse to train a YOLO-OBB model for baggage trolley detection, reducing annotation effort and improving accuracy.
- PanoDP: Learning Collision-Free Navigation with Panoramic Depth and Differentiable Physics (viability: 7): https://sciencetostartup.com/paper/panodp-learning-collision-free-navigation-with-panoramic-depth-and-differentiable-physics - PanoDP is a collision-free navigation system using panoramic depth and differentiable physics, offering improved safety and completion rates in cluttered environments.
- Helix: Evolutionary Reinforcement Learning for Open-Ended Scientific Problem Solving (viability: 7): https://sciencetostartup.com/paper/helix-evolutionary-reinforcement-learning-for-open-ended-scientific-problem-solving - HELIX is a reinforcement learning framework that leverages in-context learning to achieve state-of-the-art results on scientific problem-solving tasks, outperforming GPT-4o on standard ML benchmarks.
- Real-Time Glottis Detection Framework via Spatial-decoupled Feature Learning for Nasal Transnasal Intubation (viability: 7): https://sciencetostartup.com/paper/real-time-glottis-detection-framework-via-spatial-decoupled-feature-learning-for-nasal-transnasal-intubation - Mobile GlottisNet is a lightweight, real-time glottis detection framework for emergency intubation, achieving high FPS on edge devices with a small 5MB model size, making it ideal for resource-constrained environments.
- Exoskeleton Control through Learning to Reduce Biological Joint Moments in Simulations (viability: 7): https://sciencetostartup.com/paper/exoskeleton-control-through-learning-to-reduce-biological-joint-moments-in-simulations - An RL framework for simulation-trained exoskeleton controllers that reduce biological joint moments, validated with an open-source gait dataset, showing strong sim-to-data consistency.
- Accelerating Diffusion Models for Generative AI Applications with Silicon Photonics (viability: 7): https://sciencetostartup.com/paper/accelerating-diffusion-models-for-generative-ai-applications-with-silicon-photonics - A silicon photonics accelerator that improves energy efficiency and throughput for diffusion models, enabling faster and more sustainable generative AI applications.
- Duala: Dual-Level Alignment of Subjects and Stimuli for Cross-Subject fMRI Decoding (viability: 7): https://sciencetostartup.com/paper/duala-dual-level-alignment-of-subjects-and-stimuli-for-cross-subject-fmri-decoding - Duala enhances cross-subject fMRI decoding by aligning stimuli and brain responses, improving brain-computer interface scalability.
- SMAT: Staged Multi-Agent Training for Co-Adaptive Exoskeleton Control (viability: 8): https://sciencetostartup.com/paper/smat-staged-multi-agent-training-for-co-adaptive-exoskeleton-control - SMAT is a staged multi-agent training approach for exoskeleton control that reduces muscle activation and delivers consistent assistance, validated in simulation and physical experiments, making it a promising solution for personalized wearable robotics.
- KohakuRAG: A simple RAG framework with hierarchical document indexing (viability: 8): https://sciencetostartup.com/paper/kohakurag-a-simple-rag-framework-with-hierarchical-document-indexing - KohakuRAG is an open-source hierarchical RAG framework that achieves state-of-the-art performance with precise citation and stable answers, ideal for technical documentation.
- MAS-H2: A Hierarchical Multi-Agent System for Holistic Cloud-Native Autoscaling (viability: 7): https://sciencetostartup.com/paper/mas-h2-a-hierarchical-multi-agent-system-for-holistic-cloud-native-autoscaling - MAS-H2 is a hierarchical multi-agent system that proactively autoscales Kubernetes resources based on business policies, reducing resource waste and performance degradation.
- TT-Sparse: Learning Sparse Rule Models with Differentiable Truth Tables (viability: 8): https://sciencetostartup.com/paper/tt-sparse-learning-sparse-rule-models-with-differentiable-truth-tables - TT-Sparse offers a novel approach to interpretable machine learning by learning sparse rule models with differentiable truth tables, providing high predictive performance and low complexity, making it suitable for high-stakes decision-making.
- EmbedTalk: Triplane-Free Talking Head Synthesis using Embedding-Driven Gaussian Deformation (viability: 7): https://sciencetostartup.com/paper/embedtalk-triplane-free-talking-head-synthesis-using-embedding-driven-gaussian-deformation - EmbedTalk enables real-time, high-quality talking head synthesis on mobile GPUs by replacing tri-plane encoding with learned embeddings for speech deformation.
- StyleBench: Evaluating Speech Language Models on Conversational Speaking Style Control (viability: 7): https://sciencetostartup.com/paper/stylebench-evaluating-speech-language-models-on-conversational-speaking-style-control - StyleBench is a benchmark for evaluating and improving speaking style control in conversational speech language models, enabling more realistic and customized interactive experiences.
- Shorter Thoughts, Same Answers: Difficulty-Scaled Segment-Wise RL for CoT Compression (viability: 7): https://sciencetostartup.com/paper/shorter-thoughts-same-answers-difficulty-scaled-segment-wise-rl-for-cot-compression - Compress chain-of-thought reasoning traces with difficulty-aware reinforcement learning to reduce token cost without sacrificing answer quality.
- Fast Attention-Based Simplification of LiDAR Point Clouds for Object Detection and Classification (viability: 7): https://sciencetostartup.com/paper/fast-attention-based-simplification-of-lidar-point-clouds-for-object-detection-and-classification - An attention-based LiDAR point cloud simplification method that balances speed and accuracy for real-time object detection and classification, offering a potential performance boost over existing sampling techniques.
- Models as Lego Builders: Assembling Malice from Benign Blocks via Semantic Blueprints (viability: 7): https://sciencetostartup.com/paper/models-as-lego-builders-assembling-malice-from-benign-blocks-via-semantic-blueprints - StructAttack is a jailbreak framework that exploits LVLMs' reasoning to assemble benign-looking visual prompts into coherent harmful semantics, demonstrating a novel vulnerability via semantic slot filling.
- 3DGS-HPC: Distractor-free 3D Gaussian Splatting with Hybrid Patch-wise Classification (viability: 7): https://sciencetostartup.com/paper/3dgs-hpc-distractor-free-3d-gaussian-splatting-with-hybrid-patch-wise-classification - 3DGS-HPC enhances 3D Gaussian Splatting by robustly removing transient distractors, improving novel view synthesis quality.
- Analysis-Driven Procedural Generation of an Engine Sound Dataset with Embedded Control Annotations (viability: 7): https://sciencetostartup.com/paper/analysis-driven-procedural-generation-of-an-engine-sound-dataset-with-embedded-control-annotations - Generate synthetic engine sounds with precise control annotations for automotive audio applications, enabling data-driven sound synthesis and virtual prototyping.
- Model-Based and Neural-Aided Approaches for Dog Dead Reckoning (viability: 8): https://sciencetostartup.com/paper/model-based-and-neural-aided-approaches-for-dog-dead-reckoning - A lightweight, low-cost positioning solution for biological and robotic dogs using inertial sensors and neural networks, with code and datasets available.
- KCoEvo: A Knowledge Graph Augmented Framework for Evolutionary Code Generation (viability: 8): https://sciencetostartup.com/paper/kcoevo-a-knowledge-graph-augmented-framework-for-evolutionary-code-generation - KCoEvo is a knowledge graph-augmented framework that helps developers automatically migrate code when APIs evolve, improving accuracy and execution success.
- FeasibleCap: Real-Time Embodiment Constraint Guidance for In-the-Wild Robot Demonstration Collection (viability: 8): https://sciencetostartup.com/paper/feasiblecap-real-time-embodiment-constraint-guidance-for-in-the-wild-robot-demonstration-collection - FeasibleCap enables real-time feedback during robot demonstration collection, improving replay success and reducing infeasible motions without requiring robot hardware or AR/VR headsets.
- Approximate Imitation Learning for Event-based Quadrotor Flight in Cluttered Environments (viability: 8): https://sciencetostartup.com/paper/approximate-imitation-learning-for-event-based-quadrotor-flight-in-cluttered-environments - An end-to-end neural network enables quadrotors to navigate cluttered environments at high speed using event camera data, trained efficiently with approximate imitation learning.
- TS-MLLM: A Multi-Modal Large Language Model-based Framework for Industrial Time-Series Big Data Analysis (viability: 8): https://sciencetostartup.com/paper/ts-mllm-a-multi-modal-large-language-model-based-framework-for-industrial-time-series-big-data-analysis - TS-MLLM is a multi-modal LLM framework for industrial time-series analysis that outperforms state-of-the-art methods, offering superior robustness and generalization.
- Efficient RGB-D Scene Understanding via Multi-task Adaptive Learning and Cross-dimensional Feature Guidance (viability: 7): https://sciencetostartup.com/paper/efficient-rgb-d-scene-understanding-via-multi-task-adaptive-learning-and-cross-dimensional-feature-guidance - An efficient RGB-D scene understanding model that performs semantic segmentation, instance segmentation, orientation estimation, panoptic segmentation, and scene classification, outperforming existing methods in accuracy and speed.
- Constraints Matrix Diffusion based Generative Neural Solver for Vehicle Routing Problems (viability: 7): https://sciencetostartup.com/paper/constraints-matrix-diffusion-based-generative-neural-solver-for-vehicle-routing-problems - A novel neural network framework using a discrete noise graph diffusion model to solve vehicle routing problems with state-of-the-art performance.
- Revisiting the LiRA Membership Inference Attack Under Realistic Assumptions (viability: 7): https://sciencetostartup.com/paper/revisiting-the-lira-membership-inference-attack-under-realistic-assumptions - A more realistic membership inference attack evaluation reveals vulnerabilities in current privacy auditing methods, suggesting a need for improved privacy-preserving machine learning techniques and robust evaluation protocols.
- GRD-Net: Generative-Reconstructive-Discriminative Anomaly Detection with Region of Interest Attention Module (viability: 7): https://sciencetostartup.com/paper/grd-net-generative-reconstructive-discriminative-anomaly-detection-with-region-of-interest-attention-module - An anomaly detection system using GANs and ROI attention for industrial visual inspection, offering improved generalization and reduced pre-processing.
- SiamGM: Siamese Geometry-Aware and Motion-Guided Network for Real-Time Satellite Video Object Tracking (viability: 7): https://sciencetostartup.com/paper/siamgm-siamese-geometry-aware-and-motion-guided-network-for-real-time-satellite-video-object-tracking - Real-time satellite video object tracker leveraging geometry-aware and motion-guided Siamese network, outperforming state-of-the-art methods with code available.
- Brain-WM: Brain Glioblastoma World Model (viability: 8): https://sciencetostartup.com/paper/brain-wm-brain-glioblastoma-world-model - Brain-WM is a brain glioblastoma world model that predicts optimal treatment plans and generates future MRI scans, offering a clinical sandbox for personalized healthcare.
- PureCC: Pure Learning for Text-to-Image Concept Customization (viability: 7): https://sciencetostartup.com/paper/purecc-pure-learning-for-text-to-image-concept-customization - PureCC offers a novel approach to text-to-image concept customization that preserves the original model's behavior while enabling high-fidelity personalization.
- Learning the APT Kill Chain: Temporal Reasoning over Provenance Data for Attack Stage Estimation (viability: 7): https://sciencetostartup.com/paper/learning-the-apt-kill-chain-temporal-reasoning-over-provenance-data-for-attack-stage-estimation - StageFinder provides accurate APT stage estimation by fusing host and network provenance data with temporal graph learning, enabling adaptive cyber defense.
- Active Inference for Micro-Gesture Recognition: EFE-Guided Temporal Sampling and Adaptive Learning (viability: 7): https://sciencetostartup.com/paper/active-inference-for-micro-gesture-recognition-efe-guided-temporal-sampling-and-adaptive-learning - An active inference framework for micro-gesture recognition that dynamically selects discriminative temporal segments and mitigates noise, improving performance in low-resource conditions.
- ECG Classification on PTB-XL: A Data-Centric Approach with Simplified CNN-VAE (viability: 7): https://sciencetostartup.com/paper/ecg-classification-on-ptb-xl-a-data-centric-approach-with-simplified-cnn-vae - A simplified CNN-VAE model for ECG classification achieves competitive performance through data-centric practices, offering a streamlined solution for early cardiovascular disease detection.
- Nwāchā Munā: A Devanagari Speech Corpus and Proximal Transfer Benchmark for Nepal Bhasha ASR (viability: 7): https://sciencetostartup.com/paper/nw-ch-mun-a-devanagari-speech-corpus-and-proximal-transfer-benchmark-for-nepal-bhasha-asr - Enable digital access for the endangered Nepal Bhasha language with a speech recognition API leveraging proximal transfer learning.
- ReconDrive: Fast Feed-Forward 4D Gaussian Splatting for Autonomous Driving Scene Reconstruction (viability: 7): https://sciencetostartup.com/paper/recondrive-fast-feed-forward-4d-gaussian-splatting-for-autonomous-driving-scene-reconstruction - ReconDrive is a fast feed-forward framework for high-fidelity 4D Gaussian Splatting in autonomous driving, enabling realistic closed-loop evaluation.
- Learning-free L2-Accented Speech Generation using Phonological Rules (viability: 7): https://sciencetostartup.com/paper/learning-free-l2-accented-speech-generation-using-phonological-rules - Generate accented speech without accented training data by combining phonological rules with a multilingual TTS model, enabling fine-grained accent control.
- COOL-MC: Verifying and Explaining RL Policies for Multi-bridge Network Maintenance (viability: 7): https://sciencetostartup.com/paper/cool-mc-verifying-and-explaining-rl-policies-for-multi-bridge-network-maintenance - COOL-MC verifies and explains RL policies for multi-bridge network maintenance, providing formal safety guarantees and interpretable insights for infrastructure managers.
- CONSTANT: Towards High-Quality One-Shot Handwriting Generation with Patch Contrastive Enhancement and Style-Aware Quantization (viability: 8): https://sciencetostartup.com/paper/constant-towards-high-quality-one-shot-handwriting-generation-with-patch-contrastive-enhancement-and-style-aware-quantiz - Generate realistic handwriting from a single reference image using a novel diffusion model with style-aware quantization and contrastive learning, offering a potential API for personalized document creation.
- How Long Can Unified Multimodal Models Generate Images Reliably? Taming Long-Horizon Interleaved Image Generation via Context Curation (viability: 7): https://sciencetostartup.com/paper/how-long-can-unified-multimodal-models-generate-images-reliably-taming-long-horizon-interleaved-image-generation-via-con - UniLongGen enhances the reliability of long-form image generation by dynamically curating the model's memory, leading to improved fidelity and consistency.
- MAWARITH: A Dataset and Benchmark for Legal Inheritance Reasoning with LLMs (viability: 7): https://sciencetostartup.com/paper/mawarith-a-dataset-and-benchmark-for-legal-inheritance-reasoning-with-llms - MAWARITH is a dataset and benchmark for legal inheritance reasoning with LLMs, enabling the development of AI-powered tools for Islamic inheritance law.
- Scale-Aware UAV-to-Satellite Cross-View Geo-Localization: A Semantic Geometric Approach (viability: 7): https://sciencetostartup.com/paper/scale-aware-uav-to-satellite-cross-view-geo-localization-a-semantic-geometric-approach - A geometric framework that recovers absolute metric scale from UAV images using semantic anchors for robust cross-view geo-localization, enabling scale-adaptive satellite image cropping.
- ACCURATE: Arbitrary-shaped Continuum Reconstruction Under Robust Adaptive Two-view Estimation (viability: 7): https://sciencetostartup.com/paper/accurate-arbitrary-shaped-continuum-reconstruction-under-robust-adaptive-two-view-estimation - ACCURATE is a 3D reconstruction framework for arbitrary-shaped continuum bodies, achieving high accuracy with a geometry-constrained algorithm, suitable for medical applications.
- ICLR: In-Context Imitation Learning with Visual Reasoning (viability: 7): https://sciencetostartup.com/paper/iclr-in-context-imitation-learning-with-visual-reasoning - ICLR enhances robot imitation learning by incorporating visual reasoning, enabling better adaptation to complex tasks from limited demonstrations.
- Obliviator Reveals the Cost of Nonlinear Guardedness in Concept Erasure (viability: 7): https://sciencetostartup.com/paper/obliviator-reveals-the-cost-of-nonlinear-guardedness-in-concept-erasure - Obliviator is a post-hoc concept erasure method that iteratively removes unwanted attributes from learned representations while preserving utility, guarding against nonlinear adversaries.
- TableMind++: An Uncertainty-Aware Programmatic Agent for Tool-Augmented Table Reasoning (viability: 8): https://sciencetostartup.com/paper/tablemind-an-uncertainty-aware-programmatic-agent-for-tool-augmented-table-reasoning - TableMind++ is an uncertainty-aware programmatic agent that enhances table reasoning by mitigating hallucinations, offering a robust solution for precise numerical operations and semantic understanding.
- Neural Dynamics-Informed Pre-trained Framework for Personalized Brain Functional Network Construction (viability: 7): https://sciencetostartup.com/paper/neural-dynamics-informed-pre-trained-framework-for-personalized-brain-functional-network-construction - A pre-trained framework for personalized brain functional network construction, improving performance in heterogeneous scenarios.
- One-for-All Model Initialization with Frequency-Domain Knowledge (viability: 8): https://sciencetostartup.com/paper/one-for-all-model-initialization-with-frequency-domain-knowledge - FRONT enables efficient transfer learning by extracting and transferring task-agnostic knowledge from pre-trained models via frequency domain analysis, allowing for faster convergence and reduced training costs.
- Beyond Data Splitting: Full-Data Conformal Prediction by Differential Privacy (viability: 7): https://sciencetostartup.com/paper/beyond-data-splitting-full-data-conformal-prediction-by-differential-privacy - A privacy-preserving conformal prediction framework that avoids data splitting, enabling more efficient uncertainty quantification for sensitive data.
- SketchGraphNet: A Memory-Efficient Hybrid Graph Transformer for Large-Scale Sketch Corpora Recognition (viability: 7): https://sciencetostartup.com/paper/sketchgraphnet-a-memory-efficient-hybrid-graph-transformer-for-large-scale-sketch-corpora-recognition - SketchGraphNet efficiently recognizes large-scale sketch corpora by modeling sketches as graphs, offering improved accuracy and memory efficiency.
- Reinforcement learning-based dynamic cleaning scheduling framework for solar energy system (viability: 7): https://sciencetostartup.com/paper/reinforcement-learning-based-dynamic-cleaning-scheduling-framework-for-solar-energy-system - An RL-powered cleaning scheduler for solar panels that dynamically optimizes cleaning intervals based on environmental conditions, reducing operational costs.
- InterReal: A Unified Physics-Based Imitation Framework for Learning Human-Object Interaction Skills (viability: 7): https://sciencetostartup.com/paper/interreal-a-unified-physics-based-imitation-framework-for-learning-human-object-interaction-skills - InterReal enables humanoid robots to learn fine-grained human-object interaction skills through imitation learning, validated on a real-world robot.
- EvolveReason: Self-Evolving Reasoning Paradigm for Explainable Deepfake Facial Image Identification (viability: 8): https://sciencetostartup.com/paper/evolvereason-self-evolving-reasoning-paradigm-for-explainable-deepfake-facial-image-identification - EvolveReason is an explainable deepfake detection system that uses reinforcement learning to iteratively improve its reasoning and identification of forgery details, offering a reliable analysis for practitioners.
- Bolbosh: Script-Aware Flow Matching for Kashmiri Text-to-Speech (viability: 7): https://sciencetostartup.com/paper/bolbosh-script-aware-flow-matching-for-kashmiri-text-to-speech - Bolbosh is an open-source neural TTS system for Kashmiri, addressing the lack of speech technology for this underserved language.
- Online Continual Learning for Anomaly Detection in IoT under Data Distribution Shifts (viability: 7): https://sciencetostartup.com/paper/online-continual-learning-for-anomaly-detection-in-iot-under-data-distribution-shifts - OCLADS is a communication framework for IoT anomaly detection that uses continual learning to adapt to data distribution shifts, enabling timely model updates and high inference accuracy on resource-constrained devices.
- A Unified Framework for Knowledge Transfer in Bidirectional Model Scaling (viability: 7): https://sciencetostartup.com/paper/a-unified-framework-for-knowledge-transfer-in-bidirectional-model-scaling - A unified framework for bidirectional knowledge transfer in model scaling, enabling efficient adaptation of pre-trained models to different sizes.
- High-Fidelity Medical Shape Generation via Skeletal Latent Diffusion (viability: 8): https://sciencetostartup.com/paper/high-fidelity-medical-shape-generation-via-skeletal-latent-diffusion - Generate high-fidelity 3D medical shapes from a latent diffusion model, enabling faster and more accurate anatomical modeling for medical applications.
- SeDa: A Unified System for Dataset Discovery and Multi-Entity Augmented Semantic Exploration (viability: 7): https://sciencetostartup.com/paper/seda-a-unified-system-for-dataset-discovery-and-multi-entity-augmented-semantic-exploration - SeDa is a unified framework for dataset discovery and semantic exploration, integrating millions of datasets across diverse platforms.
- Enhanced Random Subspace Local Projections for High-Dimensional Time Series Analysis (viability: 7): https://sciencetostartup.com/paper/enhanced-random-subspace-local-projections-for-high-dimensional-time-series-analysis - A robust time series forecasting framework that reduces overfitting in high-dimensional data, enabling more reliable impulse response estimation.
- AMR-CCR: Anchored Modular Retrieval for Continual Chinese Character Recognition (viability: 7): https://sciencetostartup.com/paper/amr-ccr-anchored-modular-retrieval-for-continual-chinese-character-recognition - An embedding-based retrieval framework for continual Chinese character recognition that allows for easy addition of new classes, ideal for digitizing cultural heritage.
- DocCogito: Aligning Layout Cognition and Step-Level Grounded Reasoning for Document Understanding (viability: 7): https://sciencetostartup.com/paper/doccogito-aligning-layout-cognition-and-step-level-grounded-reasoning-for-document-understanding - DocCogito enhances document understanding by integrating layout perception with structured reasoning, offering a more reliable and transparent approach for high-stakes applications.
- Pushing Bistatic Wireless Sensing toward High Accuracy at the Sub-Wavelength Scale (viability: 7): https://sciencetostartup.com/paper/pushing-bistatic-wireless-sensing-toward-high-accuracy-at-the-sub-wavelength-scale - A framework that improves the accuracy of bistatic wireless sensing by reconstructing sub-wavelength displacement details, enabling more precise contactless sensing applications.
- RobustSCI: Beyond Reconstruction to Restoration for Snapshot Compressive Imaging under Real-World Degradations (viability: 7): https://sciencetostartup.com/paper/robustsci-beyond-reconstruction-to-restoration-for-snapshot-compressive-imaging-under-real-world-degradations - RobustSCI restores high-quality video from degraded snapshot compressive imaging measurements, offering a practical solution for real-world applications.
- A Joint Neural Baseline for Concept, Assertion, and Relation Extraction from Clinical Text (viability: 7): https://sciencetostartup.com/paper/a-joint-neural-baseline-for-concept-assertion-and-relation-extraction-from-clinical-text - An end-to-end system for joint concept, assertion, and relation extraction from clinical text, outperforming pipeline baselines and offering a strong baseline for future research.
- Multi-Modal Decouple and Recouple Network for Robust 3D Object Detection (viability: 7): https://sciencetostartup.com/paper/multi-modal-decouple-and-recouple-network-for-robust-3d-object-detection - Improve the robustness of multi-modal 3D object detection systems by decoupling and recoupling features to handle data corruption, offering a more reliable perception solution for autonomous vehicles and robotics.
- HSC-VLA: Hierarchical Scene-Clearing for Robust Bimanual Manipulation in Dense Clutter (viability: 8): https://sciencetostartup.com/paper/hsc-vla-hierarchical-scene-clearing-for-robust-bimanual-manipulation-in-dense-clutter - HSC-VLA is a hierarchical framework that improves robot manipulation in cluttered environments by decoupling high-level reasoning from low-level execution, achieving significant performance gains over monolithic baselines.
- Interpretable-by-Design Transformers via Architectural Stream Independence (viability: 7): https://sciencetostartup.com/paper/interpretable-by-design-transformers-via-architectural-stream-independence - The Late Fusion Architecture (LFA) enforces interpretability in transformers by maintaining independent token and semantic streams, enabling functional modularity and improved learning mechanisms, making it easier to understand and control transformer behavior.
- GSAT: Geometric Traversability Estimation using Self-supervised Learning with Anomaly Detection for Diverse Terrains (viability: 7): https://sciencetostartup.com/paper/gsat-geometric-traversability-estimation-using-self-supervised-learning-with-anomaly-detection-for-diverse-terrains - GSAT enables safer autonomous navigation by using self-supervised learning and anomaly detection to estimate environmental traversability.
- EVLF: Early Vision-Language Fusion for Generative Dataset Distillation (viability: 7): https://sciencetostartup.com/paper/evlf-early-vision-language-fusion-for-generative-dataset-distillation - Improve dataset distillation by fusing visual and textual embeddings early in the generative process, creating more visually coherent synthetic data for downstream classification tasks.
- Skip to the Good Part: Representation Structure & Inference-Time Layer Skipping in Diffusion vs. Autoregressive LLMs (viability: 7): https://sciencetostartup.com/paper/skip-to-the-good-part-representation-structure-inference-time-layer-skipping-in-diffusion-vs-autoregressive-llms - Optimize diffusion language model inference by skipping redundant layers, achieving significant FLOPs reduction without substantial performance loss.
- Give Them an Inch and They Will Take a Mile:Understanding and Measuring Caller Identity Confusion in MCP-Based AI Systems (viability: 7): https://sciencetostartup.com/paper/give-them-an-inch-and-they-will-take-a-mile-understanding-and-measuring-caller-identity-confusion-in-mcp-based-ai-system - Secure your AI agent interactions with our caller identity authentication and fine-grained authorization solution for MCP-based systems.
- Contact-Guided 3D Genome Structure Generation of E. coli via Diffusion Transformers (viability: 7): https://sciencetostartup.com/paper/contact-guided-3d-genome-structure-generation-of-e-coli-via-diffusion-transformers - Generate 3D E. coli genome structures from Hi-C contact maps using a conditional diffusion-transformer framework, enabling ensemble-level analysis and discovery.
- FedEU: Evidential Uncertainty-Driven Federated Fine-Tuning of Vision Foundation Models for Remote Sensing Image Segmentation (viability: 7): https://sciencetostartup.com/paper/fedeu-evidential-uncertainty-driven-federated-fine-tuning-of-vision-foundation-models-for-remote-sensing-image-segmentat - Federated learning framework for remote sensing image segmentation that uses evidential uncertainty to improve model reliability and performance across heterogeneous datasets.
- Trusting What You Cannot See: Auditable Fine-Tuning and Inference for Proprietary AI (viability: 7): https://sciencetostartup.com/paper/trusting-what-you-cannot-see-auditable-fine-tuning-and-inference-for-proprietary-ai - AFTUNE provides auditable fine-tuning and inference for proprietary AI models in the cloud, ensuring computation integrity through lightweight recording and spot-check mechanisms.
- Classifying Novel 3D-Printed Objects without Retraining: Towards Post-Production Automation in Additive Manufacturing (viability: 7): https://sciencetostartup.com/paper/classifying-novel-3d-printed-objects-without-retraining-towards-post-production-automation-in-additive-manufacturing - Automate 3D printed object classification using CAD models and contrastive fine-tuning, eliminating the need for retraining with new objects.
- SIGMAE: A Spectral-Index-Guided Foundation Model for Multispectral Remote Sensing (viability: 7): https://sciencetostartup.com/paper/sigmae-a-spectral-index-guided-foundation-model-for-multispectral-remote-sensing - SIGMAE enhances multispectral remote sensing image interpretation by using spectral indices to guide masked autoencoder pretraining, improving performance on various downstream tasks.
- Do Machines Fail Like Humans? A Human-Centred Out-of-Distribution Spectrum for Mapping Error Alignment (viability: 7): https://sciencetostartup.com/paper/do-machines-fail-like-humans-a-human-centred-out-of-distribution-spectrum-for-mapping-error-alignment - A framework for evaluating AI systems based on how their errors align with human perceptual difficulty, enabling targeted improvements for human-AI collaboration.
- Where Do LLM-based Systems Break? A System-Level Security Framework for Risk Assessment and Treatment (viability: 7): https://sciencetostartup.com/paper/where-do-llm-based-systems-break-a-system-level-security-framework-for-risk-assessment-and-treatment - A goal-driven risk assessment framework for LLM-powered systems, combining system modeling with attack-defense trees and CVSS-based exploitability scoring, enabling targeted defenses.
- SLNet: A Super-Lightweight Geometry-Adaptive Network for 3D Point Cloud Recognition (viability: 8): https://sciencetostartup.com/paper/slnet-a-super-lightweight-geometry-adaptive-network-for-3d-point-cloud-recognition - SLNet offers a super-lightweight, high-performance backbone for 3D point cloud recognition, enabling efficient deployment on resource-constrained devices.
- Backdoor4Good: Benchmarking Beneficial Uses of Backdoors in LLMs (viability: 7): https://sciencetostartup.com/paper/backdoor4good-benchmarking-beneficial-uses-of-backdoors-in-llms - Repurpose backdoor triggers in LLMs for safety, controllability, and accountability, creating a modular and interpretable interface for trustworthy AI.
- Dial: A Knowledge-Grounded Dialect-Specific NL2SQL System (viability: 8): https://sciencetostartup.com/paper/dial-a-knowledge-grounded-dialect-specific-nl2sql-system - Dial is a knowledge-grounded NL2SQL system that translates natural language into dialect-specific SQL queries for heterogeneous database systems, improving accuracy and dialect coverage.
- Discrete Tokenization Unlocks Transformers for Calibrated Tabular Forecasting (viability: 7): https://sciencetostartup.com/paper/discrete-tokenization-unlocks-transformers-for-calibrated-tabular-forecasting - A novel tokenization method unlocks Transformers for calibrated tabular forecasting, outperforming XGBoost with improved accuracy and calibration.
- Few Tokens, Big Leverage: Preserving Safety Alignment by Constraining Safety Tokens during Fine-tuning (viability: 7): https://sciencetostartup.com/paper/few-tokens-big-leverage-preserving-safety-alignment-by-constraining-safety-tokens-during-fine-tuning - PACT fine-tuning framework stabilizes LLM safety by constraining confidence on safety tokens, preventing alignment drift during downstream tasks.
- HLER: Human-in-the-Loop Economic Research via Multi-Agent Pipelines for Empirical Discovery (viability: 7): https://sciencetostartup.com/paper/hler-human-in-the-loop-economic-research-via-multi-agent-pipelines-for-empirical-discovery - HLER is a multi-agent system that automates empirical economic research with human oversight, generating feasible hypotheses and complete manuscripts at low cost.
- Med-Evo: Test-time Self-evolution for Medical Multimodal Large Language Models (viability: 7): https://sciencetostartup.com/paper/med-evo-test-time-self-evolution-for-medical-multimodal-large-language-models - Med-Evo enhances medical MLLM performance using label-free reinforcement learning, improving accuracy and recall without additional labeled data.
- LITHE: Bridging Best-Effort Python and Real-Time C++ for Hot-Swapping Robotic Control Laws on Commodity Linux (viability: 7): https://sciencetostartup.com/paper/lithe-bridging-best-effort-python-and-real-time-c-for-hot-swapping-robotic-control-laws-on-commodity-linux - LITHE enables real-time robotic control law updates by bridging Python and C++ on commodity hardware, allowing for dynamic adaptation based on LLM supervision.
- DogWeave: High-Fidelity 3D Canine Reconstruction from a Single Image via Normal Fusion and Conditional Inpainting (viability: 7): https://sciencetostartup.com/paper/dogweave-high-fidelity-3d-canine-reconstruction-from-a-single-image-via-normal-fusion-and-conditional-inpainting - DogWeave reconstructs high-fidelity 3D canine models from a single image, offering a specialized solution for the pet industry.
- Machine Learning for Stress Testing: Uncertainty Decomposition in Causal Panel Prediction (viability: 7): https://sciencetostartup.com/paper/machine-learning-for-stress-testing-uncertainty-decomposition-in-causal-panel-prediction - A framework for policy-path counterfactual inference in panels, enabling transparent uncertainty decomposition for regulatory stress testing and causal impact analysis.
- RPG-SAM: Reliability-Weighted Prototypes and Geometric Adaptive Threshold Selection for Training-Free One-Shot Polyp Segmentation (viability: 7): https://sciencetostartup.com/paper/rpg-sam-reliability-weighted-prototypes-and-geometric-adaptive-threshold-selection-for-training-free-one-shot-polyp-segm - RPG-SAM is a training-free one-shot polyp segmentation framework that leverages reliability-weighted prototypes and geometric adaptive selection to improve segmentation accuracy, offering a scalable alternative to expert annotations.
- Data Agent: Learning to Select Data via End-to-End Dynamic Optimization (viability: 7): https://sciencetostartup.com/paper/data-agent-learning-to-select-data-via-end-to-end-dynamic-optimization - Data Agent accelerates model training by dynamically selecting informative samples, reducing training costs by up to 50% while maintaining performance.
- Generalization in Online Reinforcement Learning for Mobile Agents (viability: 7): https://sciencetostartup.com/paper/generalization-in-online-reinforcement-learning-for-mobile-agents - Automated mobile agent training system with benchmark for zero-shot generalization, enabling a 7B-parameter VLM agent to surpass supervised fine-tuning baselines.
- Disentangled Textual Priors for Diffusion-based Image Super-Resolution (viability: 7): https://sciencetostartup.com/paper/disentangled-textual-priors-for-diffusion-based-image-super-resolution - A diffusion-based image super-resolution framework using disentangled textual priors for improved semantic control and image quality.
- AutoControl Arena: Synthesizing Executable Test Environments for Frontier AI Risk Evaluation (viability: 7): https://sciencetostartup.com/paper/autocontrol-arena-synthesizing-executable-test-environments-for-frontier-ai-risk-evaluation - AutoControl Arena provides an automated framework for evaluating and mitigating risks associated with autonomous LLM agents by synthesizing executable test environments.
- Cable-driven Continuum Robotics: Proprioception via Proximal-integrated Force Sensing (viability: 7): https://sciencetostartup.com/paper/cable-driven-continuum-robotics-proprioception-via-proximal-integrated-force-sensing - A novel proprioception method for cable-driven continuum robots based on proximal-integrated force sensing, enabling safer and smarter robots for clinical adoption.
- Dynamic Vehicle Routing Problem with Prompt Confirmation of Advance Requests (viability: 7): https://sciencetostartup.com/paper/dynamic-vehicle-routing-problem-with-prompt-confirmation-of-advance-requests - Optimize on-demand transportation services by promptly confirming advance trip requests and continually improving routes, increasing the number of requests served.
- DualSpec: Accelerating Deep Research Agents via Dual-Process Action Speculation (viability: 7): https://sciencetostartup.com/paper/dualspec-accelerating-deep-research-agents-via-dual-process-action-speculation - DualSpec accelerates deep research agents by heterogeneously speculating on actions based on confidence, achieving significant speedups.
- QdaVPR: A novel query-based domain-agnostic model for visual place recognition (viability: 8): https://sciencetostartup.com/paper/qdavpr-a-novel-query-based-domain-agnostic-model-for-visual-place-recognition - QdaVPR is a domain-agnostic visual place recognition model that achieves state-of-the-art performance on multiple benchmarks, enabling robust localization for robotics and autonomous systems in diverse environments.
- Reality Check for Tor Website Fingerprinting in the Open World (viability: 7): https://sciencetostartup.com/paper/reality-check-for-tor-website-fingerprinting-in-the-open-world - A tool to detect website fingerprinting attacks on Tor networks, enabling enhanced security measures for privacy-focused applications.
- UnSCAR: Universal, Scalable, Controllable, and Adaptable Image Restoration (viability: 7): https://sciencetostartup.com/paper/unscar-universal-scalable-controllable-and-adaptable-image-restoration - UnSCAR offers a scalable and controllable image restoration solution using a mixture-of-experts architecture, enabling robust generalization across diverse degradations.
- Adaptive Capacity Allocation for Vision Language Action Fine-tuning (viability: 7): https://sciencetostartup.com/paper/adaptive-capacity-allocation-for-vision-language-action-fine-tuning - LoRA-SP enables efficient adaptation of vision-language-action models for robotics by adaptively allocating capacity during fine-tuning, improving multi-task performance and generalization.
- Prompt-Based Caption Generation for Single-Tooth Dental Images Using Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/prompt-based-caption-generation-for-single-tooth-dental-images-using-vision-language-models - Generate detailed captions for single-tooth dental images using vision-language models, enabling more comprehensive dental image analysis.
- Generalizing Linear Autoencoder Recommenders with Decoupled Expected Quadratic Loss (viability: 7): https://sciencetostartup.com/paper/generalizing-linear-autoencoder-recommenders-with-decoupled-expected-quadratic-loss - Improve recommendation accuracy by expanding the solution space of linear autoencoders with a decoupled quadratic loss, enabling better hyperparameter tuning and efficient computation.
- VIVECaption: A Split Approach to Caption Quality Improvement (viability: 7): https://sciencetostartup.com/paper/vivecaption-a-split-approach-to-caption-quality-improvement - VIVECaption improves image-caption alignment quality using a two-sided approach of gold-standard dataset creation and model alignment, providing high-quality training data for generative models.
- Interpretable Aneurysm Classification via 3D Concept Bottleneck Models: Integrating Morphological and Hemodynamic Clinical Features (viability: 7): https://sciencetostartup.com/paper/interpretable-aneurysm-classification-via-3d-concept-bottleneck-models-integrating-morphological-and-hemodynamic-clinica - An interpretable AI model for aneurysm classification using 3D Concept Bottleneck Models, offering clinical transparency and high accuracy.
- AQuA: Toward Strategic Response Generation for Ambiguous Visual Questions (viability: 7): https://sciencetostartup.com/paper/aqua-toward-strategic-response-generation-for-ambiguous-visual-questions - Fine-tuned VLMs on a new ambiguous VQA dataset to generate strategic responses, enabling them to recognize ambiguity, manage uncertainty, and respond with context-appropriate strategies.
- Can Large Language Models Keep Up? Benchmarking Online Adaptation to Continual Knowledge Streams (viability: 7): https://sciencetostartup.com/paper/can-large-language-models-keep-up-benchmarking-online-adaptation-to-continual-knowledge-streams - Benchmark to evaluate LLM's ability to adapt to continuously evolving knowledge streams, revealing limitations in current methodologies and highlighting the need for improved online adaptation techniques.
- Deterministic Fuzzy Triage for Legal Compliance Classification and Evidence Retrieval (viability: 7): https://sciencetostartup.com/paper/deterministic-fuzzy-triage-for-legal-compliance-classification-and-evidence-retrieval - A deterministic and explainable AI system for legal document triage, offering a practical alternative to opaque LLMs with reproducible audit trails.
- Feed m Birds with One Scone: Accelerating Multi-task Gradient Balancing via Bi-level Optimization (viability: 7): https://sciencetostartup.com/paper/feed-m-birds-with-one-scone-accelerating-multi-task-gradient-balancing-via-bi-level-optimization - MARIGOLD accelerates multi-task learning by efficiently balancing gradients using bi-level optimization, showing superior performance on public and industrial datasets.
- SoK: Evolution, Security, and Fundamental Properties of Transactional Systems (viability: 7): https://sciencetostartup.com/paper/sok-evolution-security-and-fundamental-properties-of-transactional-systems - A security analysis framework for transactional systems, identifying vulnerabilities and proposing an extension to the ACID model.
- Scheduling Parallel Optical Circuit Switches for AI Training (viability: 7): https://sciencetostartup.com/paper/scheduling-parallel-optical-circuit-switches-for-ai-training - Spectra optimizes scheduling for parallel optical circuit switches in AI training, reducing makespan by up to 2.4x compared to state-of-the-art.
- Domain-Specific Quality Estimation for Machine Translation in Low-Resource Scenarios (viability: 7): https://sciencetostartup.com/paper/domain-specific-quality-estimation-for-machine-translation-in-low-resource-scenarios - Improve machine translation quality estimation in low-resource languages by adapting LLMs with LoRA and releasing domain-specific datasets.
- Learning to Reflect: Hierarchical Multi-Agent Reinforcement Learning for CSI-Free mmWave Beam-Focusing (viability: 7): https://sciencetostartup.com/paper/learning-to-reflect-hierarchical-multi-agent-reinforcement-learning-for-csi-free-mmwave-beam-focusing - A hierarchical multi-agent reinforcement learning framework for CSI-free mmWave beam-focusing, improving RSSI and scaling efficiently with user density.
- Position: LLMs Must Use Functor-Based and RAG-Driven Bias Mitigation for Fairness (viability: 7): https://sciencetostartup.com/paper/position-llms-must-use-functor-based-and-rag-driven-bias-mitigation-for-fairness - Mitigate biases in LLMs by combining category-theoretic transformations with retrieval-augmented generation (RAG) for equitable and fair model outputs.
- RILEC: Detection and Generation of L1 Russian Interference Errors in English Learner Texts (viability: 7): https://sciencetostartup.com/paper/rilec-detection-and-generation-of-l1-russian-interference-errors-in-english-learner-texts - RILEC is a tool that detects Russian L1 interference errors in English learner texts, enabling targeted feedback and improved language learning outcomes.
- N-Tree Diffusion for Long-Horizon Wildfire Risk Forecasting (viability: 7): https://sciencetostartup.com/paper/n-tree-diffusion-for-long-horizon-wildfire-risk-forecasting - NT-Diffusion offers a computationally efficient hierarchical diffusion model for long-horizon wildfire risk forecasting, enabling proactive resource allocation and risk mitigation.
- Latent Generative Models with Tunable Complexity for Compressed Sensing and other Inverse Problems (viability: 7): https://sciencetostartup.com/paper/latent-generative-models-with-tunable-complexity-for-compressed-sensing-and-other-inverse-problems - Tunable-complexity generative models offer improved performance in solving inverse problems like compressed sensing, inpainting, and denoising, making them a valuable tool for signal processing applications.
- AgrI Challenge: A Data-Centric AI Competition for Cross-Team Validation in Agricultural Vision (viability: 7): https://sciencetostartup.com/paper/agri-challenge-a-data-centric-ai-competition-for-cross-team-validation-in-agricultural-vision - A data-centric AI competition and dataset for agricultural vision, enabling robust model generalization across diverse field conditions.
- Learning Clinical Representations Under Systematic Distribution Shift (viability: 7): https://sciencetostartup.com/paper/learning-clinical-representations-under-systematic-distribution-shift - A framework for training clinical ML models that are robust to distribution shifts across different hospitals, improving out-of-distribution performance.
- How Much Noise Can BERT Handle? Insights from Multilingual Sentence Difficulty Detection (viability: 7): https://sciencetostartup.com/paper/how-much-noise-can-bert-handle-insights-from-multilingual-sentence-difficulty-detection - Denoising strategies for sentence-level difficulty detection using BERT can be productized as a data cleaning API for noisy training data.
- Learning Concept Bottleneck Models from Mechanistic Explanations (viability: 7): https://sciencetostartup.com/paper/learning-concept-bottleneck-models-from-mechanistic-explanations - M-CBM extracts and names concepts learned by black-box models to build interpretable concept bottleneck models, improving performance and explainability.
- Agora: Teaching the Skill of Consensus-Finding with AI Personas Grounded in Human Voice (viability: 7): https://sciencetostartup.com/paper/agora-teaching-the-skill-of-consensus-finding-with-ai-personas-grounded-in-human-voice - Agora is an AI-powered platform that uses LLMs to organize human voices on policy issues, helping users build consensus-finding skills.
- VisualScratchpad: Inference-time Visual Concepts Analysis in Vision Language Models (viability: 7): https://sciencetostartup.com/paper/visualscratchpad-inference-time-visual-concepts-analysis-in-vision-language-models - VisualScratchpad is an interactive interface for debugging vision language models by analyzing visual concepts during inference, enabling users to identify failure modes and improve model performance.
- To Predict or Not to Predict? Towards reliable uncertainty estimation in the presence of noise (viability: 7): https://sciencetostartup.com/paper/to-predict-or-not-to-predict-towards-reliable-uncertainty-estimation-in-the-presence-of-noise - Improve the reliability of multilingual text classification by integrating uncertainty estimation, allowing systems to abstain from predicting on uncertain instances and boost overall accuracy.
- ShakyPrepend: A Multi-Group Learner with Improved Sample Complexity (viability: 5): https://sciencetostartup.com/paper/shakyprepend-a-multi-group-learner-with-improved-sample-complexity - ShakyPrepend improves multi-group learning with better theoretical guarantees and adaptability, offering practical deployment guidance.
- FinSheet-Bench: From Simple Lookups to Complex Reasoning, Where LLMs Break on Financial Spreadsheets (viability: 7): https://sciencetostartup.com/paper/finsheet-bench-from-simple-lookups-to-complex-reasoning-where-llms-break-on-financial-spreadsheets - FinSheet-Bench provides a benchmark for evaluating LLM performance on financial spreadsheets, highlighting the need for improved document understanding and deterministic computation in financial applications.
- Faster-HEAL: An Efficient and Privacy-Preserving Collaborative Perception Framework for Heterogeneous Autonomous Vehicles (viability: 7): https://sciencetostartup.com/paper/faster-heal-an-efficient-and-privacy-preserving-collaborative-perception-framework-for-heterogeneous-autonomous-vehicles - Faster-HEAL is a privacy-preserving collaborative perception framework that fine-tunes a low-rank visual prompt to align heterogeneous autonomous vehicle features, improving detection performance with minimal computational overhead.
- Adversarial Latent-State Training for Robust Policies in Partially Observable Domains (viability: 7): https://sciencetostartup.com/paper/adversarial-latent-state-training-for-robust-policies-in-partially-observable-domains - A framework for training robust RL policies in partially observable environments by exposing them to adversarial latent-state distributions, improving worst-case performance.
- Soft Rigid Hybrid Gripper with Inflatable Silicone Pockets for Tunable Frictional Grasping (viability: 7): https://sciencetostartup.com/paper/soft-rigid-hybrid-gripper-with-inflatable-silicone-pockets-for-tunable-frictional-grasping - A soft-rigid hybrid gripper that uses inflatable silicone pockets to modulate friction for secure grasping of diverse objects.
- StructSAM: Structure- and Spectrum-Preserving Token Merging for Segment Anything Models (viability: 7): https://sciencetostartup.com/paper/structsam-structure-and-spectrum-preserving-token-merging-for-segment-anything-models - StructSAM accelerates Segment Anything Model (SAM) inference by intelligently merging tokens, offering a speedup with minimal accuracy loss, making it ideal for real-time segmentation applications.
- Retrieval-Augmented Multi-scale Framework for County-Level Crop Yield Prediction Across Large Regions (viability: 7): https://sciencetostartup.com/paper/retrieval-augmented-multi-scale-framework-for-county-level-crop-yield-prediction-across-large-regions - A retrieval-augmented framework for county-level crop yield prediction, improving accuracy and robustness across large regions.
- Training for Trustworthy Saliency Maps: Adversarial Training Meets Feature-Map Smoothing (viability: 7): https://sciencetostartup.com/paper/training-for-trustworthy-saliency-maps-adversarial-training-meets-feature-map-smoothing - Improve the trustworthiness of AI explanations with a training-centered approach that combines adversarial training and feature-map smoothing for more stable and reliable saliency maps.
- AutoResearch-RL: Perpetual Self-Evaluating Reinforcement Learning Agents for Autonomous Neural Architecture Discovery (viability: 7): https://sciencetostartup.com/paper/autoresearch-rl-perpetual-self-evaluating-reinforcement-learning-agents-for-autonomous-neural-architecture-discovery - AutoResearch-RL automates neural architecture discovery using reinforcement learning, achieving performance comparable to hand-tuned baselines without human intervention.
- Spectral Discovery of Continuous Symmetries via Generalized Fourier Transforms (viability: 5): https://sciencetostartup.com/paper/spectral-discovery-of-continuous-symmetries-via-generalized-fourier-transforms - Discover hidden symmetries in data using spectral analysis and Generalized Fourier Transforms, enabling more efficient and interpretable machine learning models.
- MAviS: A Multimodal Conversational Assistant For Avian Species (viability: 7): https://sciencetostartup.com/paper/mavis-a-multimodal-conversational-assistant-for-avian-species - MAviS-Chat is a multimodal LLM fine-tuned on a new avian species dataset, enabling fine-grained species understanding and multimodal question answering, outperforming existing models.
- Virtual Try-On for Cultural Clothing: A Benchmarking Study (viability: 7): https://sciencetostartup.com/paper/virtual-try-on-for-cultural-clothing-a-benchmarking-study - A virtual try-on dataset and benchmark focused on culturally diverse Bangladeshi garments, enabling more accurate and inclusive virtual try-on experiences.
- Taiwan Safety Benchmark and Breeze Guard: Toward Trustworthy AI for Taiwanese Mandarin (viability: 8): https://sciencetostartup.com/paper/taiwan-safety-benchmark-and-breeze-guard-toward-trustworthy-ai-for-taiwanese-mandarin - Breeze Guard is a culturally grounded safety model for Taiwanese Mandarin, outperforming general-purpose models in detecting region-specific risks like financial scams and misinformation.
- Variational Flow Maps: Make Some Noise for One-Step Conditional Generation (viability: 7): https://sciencetostartup.com/paper/variational-flow-maps-make-some-noise-for-one-step-conditional-generation - Variational Flow Maps enable fast, high-quality conditional image generation by learning to adapt noise distributions for one-step sampling, offering a speed advantage over iterative diffusion models.
- VisualDeltas: Learning Preferences from Visual Quality Perturbations (viability: 7): https://sciencetostartup.com/paper/visualdeltas-learning-preferences-from-visual-quality-perturbations - VisualDeltas is a lightweight preference learning framework that learns from visual quality variations in multimodal data, enabling improved generalization without human annotations.
- How to Steal Reasoning Without Reasoning Traces (viability: 7): https://sciencetostartup.com/paper/how-to-steal-reasoning-without-reasoning-traces - Steal reasoning capabilities from black-box LLMs by training student models on synthetically generated reasoning traces, significantly boosting performance on reasoning benchmarks.
- Kinematics-Aware Latent World Models for Data-Efficient Autonomous Driving (viability: 7): https://sciencetostartup.com/paper/kinematics-aware-latent-world-models-for-data-efficient-autonomous-driving - Improve autonomous driving policy learning with a kinematics-aware latent world model framework, enhancing spatial representation and sample efficiency.
- Turning Time Series into Algebraic Equations: Symbolic Machine Learning for Interpretable Modeling of Chaotic Time Series (viability: 7): https://sciencetostartup.com/paper/turning-time-series-into-algebraic-equations-symbolic-machine-learning-for-interpretable-modeling-of-chaotic-time-series - Symbolic regression forecasters that learn interpretable algebraic equations from chaotic time series data, offering a transparent alternative to black-box deep learning models.
- LF2L: Loss Fusion Horizontal Federated Learning Across Heterogeneous Feature Spaces Using External Datasets Effectively: A Case Study in Second Primary Cancer Prediction (viability: 7): https://sciencetostartup.com/paper/lf2l-loss-fusion-horizontal-federated-learning-across-heterogeneous-feature-spaces-using-external-datasets-effectively-a - A federated learning framework that leverages external datasets to improve second primary cancer prediction while preserving data privacy.
- PresentBench: A Fine-Grained Rubric-Based Benchmark for Slide Generation (viability: 7): https://sciencetostartup.com/paper/presentbench-a-fine-grained-rubric-based-benchmark-for-slide-generation - PresentBench is a fine-grained benchmark for evaluating and improving automated slide generation, enabling more effective presentation tools.
- Rethinking Deep Research from the Perspective of Web Content Distribution Matching (viability: 7): https://sciencetostartup.com/paper/rethinking-deep-research-from-the-perspective-of-web-content-distribution-matching - WeDas improves Deep Search Agent performance by incorporating web content distribution awareness into the agent's observation space, dynamically recalibrating sub-goals based on search results.
- FabricGen: Microstructure-Aware Woven Fabric Generation (viability: 7): https://sciencetostartup.com/paper/fabricgen-microstructure-aware-woven-fabric-generation - FabricGen generates high-quality woven fabric materials from text by combining diffusion models for macro-textures and a specialized LLM for micro-scale weaving patterns.
- Single Image Super-Resolution via Bivariate `A Trous Wavelet Diffusion (viability: 7): https://sciencetostartup.com/paper/single-image-super-resolution-via-bivariate-a-trous-wavelet-diffusion - BATDiff enhances single image super-resolution by using wavelet diffusion to generate sharper, structurally consistent reconstructions, offering a potential API for image enhancement.
- Retrieval-Augmented Generation for Predicting Cellular Responses to Gene Perturbation (viability: 7): https://sciencetostartup.com/paper/retrieval-augmented-generation-for-predicting-cellular-responses-to-gene-perturbation - PT-RAG leverages retrieval-augmented generation to predict cellular responses to gene perturbations, outperforming existing methods and offering a novel approach for understanding gene function and therapeutic development.
- Unlocking Data Value in Finance: A Study on Distillation and Difficulty-Aware Training (viability: 8): https://sciencetostartup.com/paper/unlocking-data-value-in-finance-a-study-on-distillation-and-difficulty-aware-training - Fine-tuned LLMs for finance using high-quality distilled data, achieving SOTA performance and releasing datasets and models for further research.
- VINO: Video-driven Invariance for Non-contextual Objects via Structural Prior Guided De-contextualization (viability: 7): https://sciencetostartup.com/paper/vino-video-driven-invariance-for-non-contextual-objects-via-structural-prior-guided-de-contextualization - VINO is a self-supervised learning framework that learns robust image encoders from dense video by imposing a structural information bottleneck, effectively disentangling foreground from background.
- Towards Objective Gastrointestinal Auscultation: Automated Segmentation and Annotation of Bowel Sound Patterns (viability: 7): https://sciencetostartup.com/paper/towards-objective-gastrointestinal-auscultation-automated-segmentation-and-annotation-of-bowel-sound-patterns - Automated bowel sound analysis using wearable sensors and AI to provide clinicians with objective diagnostic feedback, reducing manual annotation time by 70%.
- wDPO: Winsorized Direct Preference Optimization for Robust LLM Alignment (viability: 7): https://sciencetostartup.com/paper/wdpo-winsorized-direct-preference-optimization-for-robust-llm-alignment - Improve LLM alignment robustness with winsorized Direct Preference Optimization (wDPO), a method that identifies and mitigates noisy preference data for better performance and safety.
- Lying to Win: Assessing LLM Deception through Human-AI Games and Parallel-World Probing (viability: 7): https://sciencetostartup.com/paper/lying-to-win-assessing-llm-deception-through-human-ai-games-and-parallel-world-probing - Quantify and mitigate deceptive behavior in LLMs using a 20-Questions game with parallel-world probing to ensure AI safety.
- Vision-Guided MPPI for Agile Drone Racing: Navigating Arbitrary Gate Poses via Neural Signed Distance Fields (viability: 7): https://sciencetostartup.com/paper/vision-guided-mppi-for-agile-drone-racing-navigating-arbitrary-gate-poses-via-neural-signed-distance-fields - A vision-guided optimal control framework for agile drone racing that uses neural signed distance fields to navigate arbitrary gate poses, enabling robust navigation even under visual occlusion.
- $\textbf{Re}^{2}$: Unlocking LLM Reasoning via Reinforcement Learning with Re-solving (viability: 7): https://sciencetostartup.com/paper/textbf-re-2-unlocking-llm-reasoning-via-reinforcement-learning-with-re-solving - Improve LLM reasoning by enabling models to abandon unproductive paths and restart, leading to better performance and efficiency.
- FastSTAR: Spatiotemporal Token Pruning for Efficient Autoregressive Video Synthesis (viability: 7): https://sciencetostartup.com/paper/faststar-spatiotemporal-token-pruning-for-efficient-autoregressive-video-synthesis - FastSTAR accelerates autoregressive video synthesis by pruning redundant spatiotemporal tokens, offering a 2x speedup with minimal quality loss.
- Governance Architecture for Autonomous Agent Systems: Threats, Framework, and Engineering Practice (viability: 7): https://sciencetostartup.com/paper/governance-architecture-for-autonomous-agent-systems-threats-framework-and-engineering-practice - A layered governance architecture for autonomous agents, mitigating execution-layer vulnerabilities with high interception rates and low latency.
- FreeFly-Thinking : Aligning Chain-of-Thought Reasoning with Continuous UAV Navigation (viability: 7): https://sciencetostartup.com/paper/freefly-thinking-aligning-chain-of-thought-reasoning-with-continuous-uav-navigation - FreeFly-Thinking enables UAVs to navigate complex outdoor environments using natural language instructions and chain-of-thought reasoning.
- Learning to Rank the Initial Branching Order of SAT Solvers (viability: 7): https://sciencetostartup.com/paper/learning-to-rank-the-initial-branching-order-of-sat-solvers - Accelerate SAT solver performance by using a graph neural network to predict a good initial branching order as a preprocessing step.
- Making LLMs Optimize Multi-Scenario CUDA Kernels Like Experts (viability: 8): https://sciencetostartup.com/paper/making-llms-optimize-multi-scenario-cuda-kernels-like-experts - CUDAMaster automates GPU kernel optimization across diverse scenarios, rivaling hand-tuned libraries and offering a demo for immediate validation.
- ACD-U: Asymmetric co-teaching with machine unlearning for robust learning with noisy labels (viability: 8): https://sciencetostartup.com/paper/acd-u-asymmetric-co-teaching-with-machine-unlearning-for-robust-learning-with-noisy-labels - ACD-U is a robust noisy label learning framework that leverages asymmetric co-teaching and machine unlearning to achieve state-of-the-art performance, offering a potential solution for improving model accuracy in real-world datasets with noisy labels.
- RoTri-Diff: A Spatial Robot-Object Triadic Interaction-Guided Diffusion Model for Bimanual Manipulation (viability: 8): https://sciencetostartup.com/paper/rotri-diff-a-spatial-robot-object-triadic-interaction-guided-diffusion-model-for-bimanual-manipulation - RoTri-Diff is a diffusion-based imitation learning framework that enables robots to perform stable and coordinated bimanual manipulation by explicitly modeling the geometric relationship between the robot's arms and the manipulated object.
- PromptGate Client Adaptive Vision Language Gating for Open Set Federated Active Learning (viability: 7): https://sciencetostartup.com/paper/promptgate-client-adaptive-vision-language-gating-for-open-set-federated-active-learning - PromptGate enhances federated active learning by dynamically filtering out-of-distribution noise using a vision-language model, improving annotation efficiency in medical AI.
- Spectral Conditioning of Attention Improves Transformer Performance (viability: 7): https://sciencetostartup.com/paper/spectral-conditioning-of-attention-improves-transformer-performance - Improve transformer performance by spectrally conditioning attention layers for better Jacobian conditioning, easily integrated as a drop-in replacement.
- Improving reasoning at inference time via uncertainty minimisation (viability: 7): https://sciencetostartup.com/paper/improving-reasoning-at-inference-time-via-uncertainty-minimisation - Improve LLM reasoning by selecting the most self-certain thought at each step, enhancing accuracy with minimal computational overhead.
- Agentic Planning with Reasoning for Image Styling via Offline RL (viability: 8): https://sciencetostartup.com/paper/agentic-planning-with-reasoning-for-image-styling-via-offline-rl - Agentic planning with offline RL for image styling enables nuanced transformations via interpretable tool sequences, improving visual quality and instruction following.
- Fine-Grained Table Retrieval Through the Lens of Complex Queries (viability: 7): https://sciencetostartup.com/paper/fine-grained-table-retrieval-through-the-lens-of-complex-queries - A fine-grained table retrieval mechanism for complex question answering over relational databases, enabling more accurate and robust data access.
- LiveWorld: Simulating Out-of-Sight Dynamics in Generative Video World Models (viability: 7): https://sciencetostartup.com/paper/liveworld-simulating-out-of-sight-dynamics-in-generative-video-world-models - LiveWorld addresses the 'out-of-sight dynamics' problem in generative video world models by simulating persistent world evolution, enabling more realistic and consistent long-term scene generation.
- CanoVerse: 3D Object Scalable Canonicalization and Dataset for Generation and Pose (viability: 7): https://sciencetostartup.com/paper/canoverse-3d-object-scalable-canonicalization-and-dataset-for-generation-and-pose - Canoverse is a large-scale canonical 3D object dataset and framework that enables pose-consistent generation and cross-modal 3D shape retrieval.
- PDD: Manifold-Prior Diverse Distillation for Medical Anomaly Detection (viability: 8): https://sciencetostartup.com/paper/pdd-manifold-prior-diverse-distillation-for-medical-anomaly-detection - PDD offers a novel manifold-prior diverse distillation framework for medical anomaly detection, significantly improving state-of-the-art performance and providing a clear path to commercial applications through its code release and benchmark results.
- Model-based thermal drift compensation for high-precision hexapod robot actuators (viability: 7): https://sciencetostartup.com/paper/model-based-thermal-drift-compensation-for-high-precision-hexapod-robot-actuators - Predictive thermal drift compensation for hexapod robots, reducing positioning errors by 80%, enabling higher precision in robotics applications.
- Emotion Transcription in Conversation: A Benchmark for Capturing Subtle and Complex Emotional States through Natural Language (viability: 7): https://sciencetostartup.com/paper/emotion-transcription-in-conversation-a-benchmark-for-capturing-subtle-and-complex-emotional-states-through-natural-lang - Generate natural language descriptions of emotional states in conversations using a new Japanese dataset, enabling more nuanced human-machine interactions.
- DexKnot: Generalizable Visuomotor Policy Learning for Dexterous Bag-Knotting Manipulation (viability: 7): https://sciencetostartup.com/paper/dexknot-generalizable-visuomotor-policy-learning-for-dexterous-bag-knotting-manipulation - DexKnot uses keypoint affordance and diffusion policies to enable robots to reliably knot plastic bags, generalizing to new bag instances and deformations.
- The Model Knows Which Tokens Matter: Automatic Token Selection via Noise Gating (viability: 8): https://sciencetostartup.com/paper/the-model-knows-which-tokens-matter-automatic-token-selection-via-noise-gating - AutoSelect accelerates vision-language model inference by pruning redundant visual tokens, achieving significant speedups with minimal accuracy loss and easy integration.
- Deep Expert Injection for Anchoring Retinal VLMs with Domain-Specific Knowledge (viability: 8): https://sciencetostartup.com/paper/deep-expert-injection-for-anchoring-retinal-vlms-with-domain-specific-knowledge - EyExIn anchors retinal VLMs with expert knowledge for precise ophthalmic diagnosis, outperforming proprietary systems and offering a trustworthy AI solution.
- Efficient Trajectory Optimization for Autonomous Racing via Formula-1 Data-Driven Initialization (viability: 7): https://sciencetostartup.com/paper/efficient-trajectory-optimization-for-autonomous-racing-via-formula-1-data-driven-initialization - Accelerate autonomous racing trajectory optimization by using a neural network to initialize the trajectory with Formula 1 data, leading to faster convergence and reduced runtime.
- Combining Adam and its Inverse Counterpart to Enhance Generalization of Deep Learning Optimizers (viability: 7): https://sciencetostartup.com/paper/combining-adam-and-its-inverse-counterpart-to-enhance-generalization-of-deep-learning-optimizers - DualAdam combines Adam and InvAdam to improve generalization performance in deep learning models, offering a potential drop-in replacement for existing optimizers.
- TIQA: Human-Aligned Text Quality Assessment in Generated Images (viability: 7): https://sciencetostartup.com/paper/tiqa-human-aligned-text-quality-assessment-in-generated-images - TIQA is a new task and method for assessing the quality of text rendering in generated images, offering a practical solution for filtering and reranking in text-to-image pipelines.
- Turn: A Language for Agentic Computation (viability: 7): https://sciencetostartup.com/paper/turn-a-language-for-agentic-computation - Turn is a new programming language designed for building agentic software with language-level constructs for cognitive type safety and secure agent interactions.
- Efficient Chest X-ray Representation Learning via Semantic-Partitioned Contrastive Learning (viability: 7): https://sciencetostartup.com/paper/efficient-chest-x-ray-representation-learning-via-semantic-partitioned-contrastive-learning - Semantic-Partitioned Contrastive Learning (S-PCL) offers an efficient and scalable pre-training framework for Chest X-ray representation learning, achieving competitive performance with lower computational cost.
- Enhancing Consistency of Werewolf AI through Dialogue Summarization and Persona Information (viability: 7): https://sciencetostartup.com/paper/enhancing-consistency-of-werewolf-ai-through-dialogue-summarization-and-persona-information - An AI Werewolf agent that maintains consistent character and context using dialogue summarization and persona information, ready for integration into existing gaming platforms.
- Learning From Failures: Efficient Reinforcement Learning Control with Episodic Memory (viability: 7): https://sciencetostartup.com/paper/learning-from-failures-efficient-reinforcement-learning-control-with-episodic-memory - Prevent robot failures during reinforcement learning with Failure Episodic Memory Alert (FEMA), improving sample efficiency and enabling real-world robot control.
- Deep Generative Spatiotemporal Engression for Probabilistic Forecasting of Epidemics (viability: 7): https://sciencetostartup.com/paper/deep-generative-spatiotemporal-engression-for-probabilistic-forecasting-of-epidemics - A deep generative spatiotemporal model for probabilistic epidemic forecasting, providing best/worst-case scenarios for public health preparedness.
- Efficient Personalized Reranking with Semi-Autoregressive Generation and Online Knowledge Distillation (viability: 7): https://sciencetostartup.com/paper/efficient-personalized-reranking-with-semi-autoregressive-generation-and-online-knowledge-distillation - A personalized semi-autoregressive reranking framework that balances generation quality and inference efficiency for recommender systems, making it suitable for real-time deployment.
- NuNext: Reframing Nucleus Detection as Next-Point Detection (viability: 7): https://sciencetostartup.com/paper/nunext-reframing-nucleus-detection-as-next-point-detection - A multimodal large language model directly predicts nucleus centroids in histopathology images, improving detection accuracy and eliminating complex post-processing.
- Towards Scalable Probabilistic Human Motion Prediction with Gaussian Processes for Safe Human-Robot Collaboration (viability: 7): https://sciencetostartup.com/paper/towards-scalable-probabilistic-human-motion-prediction-with-gaussian-processes-for-safe-human-robot-collaboration - A Gaussian Process-based human motion prediction model for safe human-robot collaboration, offering competitive accuracy and reliable uncertainty estimates.
- ACLM: ADMM-Based Distributed Model Predictive Control for Collaborative Loco-Manipulation (viability: 7): https://sciencetostartup.com/paper/aclm-admm-based-distributed-model-predictive-control-for-collaborative-loco-manipulation - A distributed model predictive control framework enables collaborative loco-manipulation for quadruped robots, offering scalability and real-time performance.
- Facial Expression Generation Aligned with Human Preference for Natural Dyadic Interaction (viability: 7): https://sciencetostartup.com/paper/facial-expression-generation-aligned-with-human-preference-for-natural-dyadic-interaction - Generate emotionally appropriate facial expressions for natural dyadic interaction using human feedback and reinforcement learning.
- mAVE: A Watermark for Joint Audio-Visual Generation Models (viability: 7): https://sciencetostartup.com/paper/mave-a-watermark-for-joint-audio-visual-generation-models - mAVE cryptographically binds audio and video in joint generation models, preventing deepfake swap attacks and protecting vendor copyright.
- Countdown-Code: A Testbed for Studying The Emergence and Generalization of Reward Hacking in RLVR (viability: 7): https://sciencetostartup.com/paper/countdown-code-a-testbed-for-studying-the-emergence-and-generalization-of-reward-hacking-in-rlvr - A testbed for studying and mitigating reward hacking in LLMs, offering a clear environment and code for further research.
- Dreamer-CDP: Improving Reconstruction-free World Models Via Continuous Deterministic Representation Prediction (viability: 7): https://sciencetostartup.com/paper/dreamer-cdp-improving-reconstruction-free-world-models-via-continuous-deterministic-representation-prediction - A reconstruction-free world model for reinforcement learning that matches Dreamer's performance, offering a more robust representation for planning and control.
- VLN-Cache: Enabling Token Caching for VLN Models with Visual/Semantic Dynamics Awareness (viability: 7): https://sciencetostartup.com/paper/vln-cache-enabling-token-caching-for-vln-models-with-visual-semantic-dynamics-awareness - VLN-Cache accelerates Vision-Language Navigation by intelligently caching and reusing visual tokens, adapting to dynamic viewpoints and task relevance for real-time deployment.
- Entropy-Aware On-Policy Distillation of Language Models (viability: 7): https://sciencetostartup.com/paper/entropy-aware-on-policy-distillation-of-language-models - Improve the accuracy and diversity of smaller language models by distilling knowledge from larger models, using an entropy-aware approach that balances precision and robustness.
- CoTJudger: A Graph-Driven Framework for Automatic Evaluation of Chain-of-Thought Efficiency and Redundancy in LRMs (viability: 7): https://sciencetostartup.com/paper/cotjudger-a-graph-driven-framework-for-automatic-evaluation-of-chain-of-thought-efficiency-and-redundancy-in-lrms - CoTJudger automatically evaluates and optimizes the efficiency of Chain-of-Thought reasoning in Large Reasoning Models by identifying and removing redundant calculations.
- Aligning What EEG Can See: Structural Representations for Brain-Vision Matching (viability: 8): https://sciencetostartup.com/paper/aligning-what-eeg-can-see-structural-representations-for-brain-vision-matching - Unlock non-invasive brain-computer interfaces with our EEG decoding method that achieves state-of-the-art accuracy by aligning brain signals with intermediate visual layers.
- Retinex Meets Language: A Physics-Semantics-Guided Underwater Image Enhancement Network (viability: 7): https://sciencetostartup.com/paper/retinex-meets-language-a-physics-semantics-guided-underwater-image-enhancement-network - Enhance underwater images by combining physics-based correction with CLIP-guided semantic restoration, leveraging a newly created large-scale image-text dataset.
- Interpretable Maximum Margin Deep Anomaly Detection (viability: 7): https://sciencetostartup.com/paper/interpretable-maximum-margin-deep-anomaly-detection - IMD-AD enhances deep anomaly detection by incorporating labeled anomalies and margin maximization for improved stability, interpretability, and performance.
- VirtueBench: Evaluating Trustworthiness under Uncertainty in Long Video Understanding (viability: 7): https://sciencetostartup.com/paper/virtuebench-evaluating-trustworthiness-under-uncertainty-in-long-video-understanding - VirtueBench is a benchmark to evaluate the trustworthiness of Vision-Language Models in long video understanding, highlighting the need for reliable and honest AI systems.
- Morphology-Independent Facial Expression Imitation for Human-Face Robots (viability: 7): https://sciencetostartup.com/paper/morphology-independent-facial-expression-imitation-for-human-face-robots - Achieve realistic facial expression imitation on human-face robots by decoupling expressions from facial morphology, enabling more natural human-robot interaction.
- MedSteer: Counterfactual Endoscopic Synthesis via Training-Free Activation Steering (viability: 8): https://sciencetostartup.com/paper/medsteer-counterfactual-endoscopic-synthesis-via-training-free-activation-steering - MedSteer is a training-free activation-steering framework for endoscopic synthesis, enabling counterfactual data generation for improved medical image augmentation and downstream polyp detection.
- GuideTWSI: A Diverse Tactile Walking Surface Indicator Dataset from Synthetic and Real-World Images for Blind and Low-Vision Navigation (viability: 7): https://sciencetostartup.com/paper/guidetwsi-a-diverse-tactile-walking-surface-indicator-dataset-from-synthetic-and-real-world-images-for-blind-and-low-vis - A dataset and model for detecting tactile walking surface indicators (TWSIs) to improve navigation assistance for blind and low-vision individuals.
- SODA: Sensitivity-Oriented Dynamic Acceleration for Diffusion Transformer (viability: 7): https://sciencetostartup.com/paper/soda-sensitivity-oriented-dynamic-acceleration-for-diffusion-transformer - SODA dynamically accelerates diffusion transformers by adaptively caching and pruning based on fine-grained sensitivity analysis, offering a balance between speed and generation quality.
- Bi-directional digital twin prototype anchoring with multi-periodicity learning for few-shot fault diagnosis (viability: 7): https://sciencetostartup.com/paper/bi-directional-digital-twin-prototype-anchoring-with-multi-periodicity-learning-for-few-shot-fault-diagnosis - A bi-directional digital twin prototype anchoring method with multi-periodicity learning enables few-shot fault diagnosis for industrial machinery, leveraging simulation data and test-time adaptation.
- Animating Petascale Time-varying Data on Commodity Hardware with LLM-assisted Scripting (viability: 7): https://sciencetostartup.com/paper/animating-petascale-time-varying-data-on-commodity-hardware-with-llm-assisted-scripting - Enable scientists to quickly create 3D animations of petascale time-varying data on commodity hardware using an LLM-assisted interface.
- Hindsight Credit Assignment for Long-Horizon LLM Agents (viability: 7): https://sciencetostartup.com/paper/hindsight-credit-assignment-for-long-horizon-llm-agents - HCAPO enhances LLM agent performance in long-horizon tasks by integrating hindsight credit assignment, improving exploration efficiency and decision-making.
- Looking Back and Forth: Cross-Image Attention Calibration and Attentive Preference Learning for Multi-Image Hallucination Mitigation (viability: 7): https://sciencetostartup.com/paper/looking-back-and-forth-cross-image-attention-calibration-and-attentive-preference-learning-for-multi-image-hallucination - Mitigate hallucinations in multi-image tasks with a cross-image attention calibration and preference learning framework, enhancing inter-image interactions and grounding predictions in visual evidence.
- Fine-Grained 3D Facial Reconstruction for Micro-Expressions (viability: 7): https://sciencetostartup.com/paper/fine-grained-3d-facial-reconstruction-for-micro-expressions - A fine-grained 3D facial reconstruction method for micro-expressions, leveraging global dynamic features and local cues for enhanced accuracy and detail.
- TacDexGrasp: Compliant and Robust Dexterous Grasping with Tactile Feedback (viability: 7): https://sciencetostartup.com/paper/tacdexgrasp-compliant-and-robust-dexterous-grasping-with-tactile-feedback - A tactile-feedback-driven controller for multi-fingered robot hands that enables robust and compliant grasping of diverse objects by preventing translational and rotational slip.
- Self-Supervised Multi-Modal World Model with 4D Space-Time Embedding (viability: 8): https://sciencetostartup.com/paper/self-supervised-multi-modal-world-model-with-4d-space-time-embedding - DeepEarth is a self-supervised multi-modal world model with a novel 4D space-time positional encoder, achieving state-of-the-art performance on ecological forecasting.
- SSP: Safety-guaranteed Surgical Policy via Joint Optimization of Behavioral and Spatial Constraints (viability: 7): https://sciencetostartup.com/paper/ssp-safety-guaranteed-surgical-policy-via-joint-optimization-of-behavioral-and-spatial-constraints - A framework for safe robot-assisted surgery using Neural ODEs and Control Barrier Functions to ensure constraint satisfaction during task execution.
- Improved Leakage Abuse Attacks in Searchable Symmetric Encryption with eBPF Monitoring (viability: 7): https://sciencetostartup.com/paper/improved-leakage-abuse-attacks-in-searchable-symmetric-encryption-with-ebpf-monitoring - Leverage eBPF monitoring to expose system-level leakages in Searchable Symmetric Encryption, enabling stronger leakage abuse attacks and highlighting the gap between theoretical security and practical system exposure.
- Two Frames Matter: A Temporal Attack for Text-to-Video Model Jailbreaking (viability: 7): https://sciencetostartup.com/paper/two-frames-matter-a-temporal-attack-for-text-to-video-model-jailbreaking - TFM is a jailbreaking framework that exploits temporal trajectory infilling vulnerabilities in text-to-video models by using fragmented prompts, enhancing attack success rates by up to 12% on commercial systems.
- Resource-Adaptive Federated Text Generation with Differential Privacy (viability: 7): https://sciencetostartup.com/paper/resource-adaptive-federated-text-generation-with-differential-privacy - A federated learning framework that uses a hybrid approach of DP finetuning and DP voting to generate synthetic text datasets while adapting to client resource constraints.
- Permutation-Equivariant 2D State Space Models: Theory and Canonical Architecture for Multivariate Time Series (viability: 7): https://sciencetostartup.com/paper/permutation-equivariant-2d-state-space-models-theory-and-canonical-architecture-for-multivariate-time-series - A novel permutation-equivariant state-space model for multivariate time series that achieves state-of-the-art performance by eliminating artificial variable ordering.
- Language-Aware Distillation for Multilingual Instruction-Following Speech LLMs with ASR-Only Supervision (viability: 7): https://sciencetostartup.com/paper/language-aware-distillation-for-multilingual-instruction-following-speech-llms-with-asr-only-supervision - Improve multilingual speech LLM instruction following by 14% using language-aware distillation and a new multilingual spoken QA benchmark.
- Hit-RAG: Learning to Reason with Long Contexts via Preference Alignment (viability: 7): https://sciencetostartup.com/paper/hit-rag-learning-to-reason-with-long-contexts-via-preference-alignment - Hit-RAG enhances retrieval-augmented generation by progressively optimizing context utilization, improving reasoning accuracy in long-context scenarios.
- OV-DEIM: Real-time DETR-Style Open-Vocabulary Object Detection with GridSynthetic Augmentation (viability: 8): https://sciencetostartup.com/paper/ov-deim-real-time-detr-style-open-vocabulary-object-detection-with-gridsynthetic-augmentation - OV-DEIM is a real-time, DETR-style open-vocabulary object detector with state-of-the-art performance and available code, making it a strong candidate for commercial applications in dynamic environments.
- AutoChecklist: Composable Pipelines for Checklist Generation and Scoring with LLM-as-a-Judge (viability: 8): https://sciencetostartup.com/paper/autochecklist-composable-pipelines-for-checklist-generation-and-scoring-with-llm-as-a-judge - AutoChecklist is an open-source library for composable checklist-based LLM evaluation, enabling fine-grained analysis and alignment with human preferences, ready for immediate productization.
- TEA-Time: Transporting Effects Across Time (viability: 7): https://sciencetostartup.com/paper/tea-time-transporting-effects-across-time - Extrapolate A/B test results to future time periods with a tool that accounts for temporal effects, enabling more reliable long-term decision-making.
- Can Safety Emerge from Weak Supervision? A Systematic Analysis of Small Language Models (viability: 7): https://sciencetostartup.com/paper/can-safety-emerge-from-weak-supervision-a-systematic-analysis-of-small-language-models - Self-MOA is an automated framework for aligning small language models using weak supervision to improve safety and helpfulness, reducing reliance on human-annotated data.
- Fréchet regression of multivariate distributions with nonparanormal transport (viability: 7): https://sciencetostartup.com/paper/fr-chet-regression-of-multivariate-distributions-with-nonparanormal-transport - A new regression approach for multivariate distributional responses using the nonparanormal transport metric, enabling efficient estimation and interpretation of predictor effects.
- Combinatorial Allocation Bandits with Nonlinear Arm Utility (viability: 6): https://sciencetostartup.com/paper/combinatorial-allocation-bandits-with-nonlinear-arm-utility - Optimize matching platform allocations to maximize user satisfaction and reduce churn using a novel bandit algorithm.
- An Extended Consent-Based Access Control Framework: Pre-Commit Validation and Emergency Access (viability: 7): https://sciencetostartup.com/paper/an-extended-consent-based-access-control-framework-pre-commit-validation-and-emergency-access - A CBAC framework that enforces semantic correctness at consent creation time for healthcare information systems, ensuring patient autonomy and low runtime decision latency.
- Two-Stage Path Following for Mobile Manipulators via Dimensionality-Reduced Graph Search and Numerical Optimization (viability: 7): https://sciencetostartup.com/paper/two-stage-path-following-for-mobile-manipulators-via-dimensionality-reduced-graph-search-and-numerical-optimization - A two-stage path planning framework for mobile manipulators that optimizes for smoothness and reachability, enabling precise and robust robot navigation.
- Privacy-Preserving Patient Identity Management Framework for Secure Healthcare Access (viability: 7): https://sciencetostartup.com/paper/privacy-preserving-patient-identity-management-framework-for-secure-healthcare-access - A privacy-preserving patient identity management framework for secure healthcare access, balancing operational needs with strong privacy.
- TrajPred: Trajectory-Conditioned Joint Embedding Prediction for Surgical Instrument-Tissue Interaction Recognition in Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/trajpred-trajectory-conditioned-joint-embedding-prediction-for-surgical-instrument-tissue-interaction-recognition-in-vis - TrajPred enhances surgical AI assistants by improving instrument-tissue interaction recognition using trajectory-conditioned embeddings within vision-language models.
- AdaGen: Learning Adaptive Policy for Image Synthesis (viability: 7): https://sciencetostartup.com/paper/adagen-learning-adaptive-policy-for-image-synthesis - AdaGen learns an adaptive policy for image synthesis, optimizing the generation process and improving performance with lower inference costs, making it a valuable tool for enhancing existing generative models.
- Adaptive Discovery of Interpretable Audio Attributes with Multimodal LLMs for Low-Resource Classification (viability: 7): https://sciencetostartup.com/paper/adaptive-discovery-of-interpretable-audio-attributes-with-multimodal-llms-for-low-resource-classification - Automated audio attribute discovery using MLLMs for faster and more interpretable low-resource audio classification.
- MipSLAM: Alias-Free Gaussian Splatting SLAM (viability: 8): https://sciencetostartup.com/paper/mipslam-alias-free-gaussian-splatting-slam - MipSLAM delivers high-fidelity, anti-aliased novel view synthesis and robust pose estimation, offering a superior SLAM solution for real-time applications.
- Foundational World Models Accurately Detect Bimanual Manipulator Failures (viability: 7): https://sciencetostartup.com/paper/foundational-world-models-accurately-detect-bimanual-manipulator-failures - A runtime monitor for bimanual robots that detects failures using uncertainty estimates from a world model, enabling safer real-world deployment.
- ADAS-TO: A Large-Scale Multimodal Naturalistic Dataset and Empirical Characterization of Human Takeovers during ADAS Engagement (viability: 7): https://sciencetostartup.com/paper/adas-to-a-large-scale-multimodal-naturalistic-dataset-and-empirical-characterization-of-human-takeovers-during-adas-enga - A dataset of autonomous driving takeovers paired with visual cues enables the development of early warning systems for safer transitions.
- Perception-Aware Multimodal Spatial Reasoning from Monocular Images (viability: 8): https://sciencetostartup.com/paper/perception-aware-multimodal-spatial-reasoning-from-monocular-images - Enhance autonomous driving spatial reasoning by equipping VLMs with object-centric grounding using visual reference tokens and multimodal chain-of-thought, outperforming existing methods on the SURDS benchmark.
- Masked Unfairness: Hiding Causality within Zero ATE (viability: 6): https://sciencetostartup.com/paper/masked-unfairness-hiding-causality-within-zero-ate - A fairness-aware optimization tool that identifies and mitigates hidden causal unfairness in decision-making processes, even when average treatment effects appear neutral.
- Optimizing Multi-Modal Models for Image-Based Shape Retrieval: The Role of Pre-Alignment and Hard Contrastive Learning (viability: 8): https://sciencetostartup.com/paper/optimizing-multi-modal-models-for-image-based-shape-retrieval-the-role-of-pre-alignment-and-hard-contrastive-learning - Improve 3D model retrieval from images using pre-trained multi-modal encoders and hard contrastive learning, enabling zero-shot and cross-domain retrieval.
- VSL-Skin: Individually Addressable Phase-Change Voxel Skin for Variable-Stiffness and Virtual Joints Bridging Soft and Rigid Robots (viability: 7): https://sciencetostartup.com/paper/vsl-skin-individually-addressable-phase-change-voxel-skin-for-variable-stiffness-and-virtual-joints-bridging-soft-and-ri - A variable stiffness lattice skin for robots enabling voxel-level morphological control and self-repair.
- A Systematic Investigation of Document Chunking Strategies and Embedding Sensitivity (viability: 7): https://sciencetostartup.com/paper/a-systematic-investigation-of-document-chunking-strategies-and-embedding-sensitivity - Optimize RAG systems with our benchmarked document chunking strategies, improving retrieval accuracy and efficiency across diverse knowledge domains.
- Elenchus: Generating Knowledge Bases from Prover-Skeptic Dialogues (viability: 7): https://sciencetostartup.com/paper/elenchus-generating-knowledge-bases-from-prover-skeptic-dialogues - Elenchus is a dialogue system that uses LLMs to help experts build knowledge bases by resolving tensions in their domain knowledge through prover-skeptic dialogues.
- T2SGrid: Temporal-to-Spatial Gridification for Video Temporal Grounding (viability: 7): https://sciencetostartup.com/paper/t2sgrid-temporal-to-spatial-gridification-for-video-temporal-grounding - T2SGrid transforms video temporal understanding into a spatial task, enabling more efficient video grounding and offering a potential API for video analysis.
- Conditional Unbalanced Optimal Transport Maps: An Outlier-Robust Framework for Conditional Generative Modeling (viability: 7): https://sciencetostartup.com/paper/conditional-unbalanced-optimal-transport-maps-an-outlier-robust-framework-for-conditional-generative-modeling - An outlier-robust conditional generative model that uses unbalanced optimal transport for improved distribution matching, offering a more reliable approach to conditional generation.
- SurgCUT3R: Surgical Scene-Aware Continuous Understanding of Temporal 3D Representation (viability: 7): https://sciencetostartup.com/paper/surgcut3r-surgical-scene-aware-continuous-understanding-of-temporal-3d-representation - SurgCUT3R adapts 3D reconstruction models to surgical environments using pseudo-ground-truth data and a hierarchical inference framework for robust pose estimation.
- Securing Cryptography in the Age of Quantum Computing and AI: Threats, Implementations, and Strategic Response (viability: 7): https://sciencetostartup.com/paper/securing-cryptography-in-the-age-of-quantum-computing-and-ai-threats-implementations-and-strategic-response - A quantum and AI-resistant cryptography toolkit combining lattice-based and hash-based algorithms for enhanced security.
- Learning Quadruped Walking from Seconds of Demonstration (viability: 7): https://sciencetostartup.com/paper/learning-quadruped-walking-from-seconds-of-demonstration - Train quadruped robot locomotion policies from seconds of demonstration data using imitation learning.
- Chart-RL: Generalized Chart Comprehension via Reinforcement Learning with Verifiable Rewards (viability: 7): https://sciencetostartup.com/paper/chart-rl-generalized-chart-comprehension-via-reinforcement-learning-with-verifiable-rewards - Chart-RL enhances chart question answering in VLMs using reinforcement learning with mathematically verifiable rewards, improving generalization and robustness.
- Post-Training with Policy Gradients: Optimality and the Base Model Barrier (viability: 7): https://sciencetostartup.com/paper/post-training-with-policy-gradients-optimality-and-the-base-model-barrier - Optimize existing autoregressive models with policy gradients to improve sequence prediction likelihood, but be aware of limitations when venturing beyond the base model's support.
- Energy-Efficient Collaborative Transport of Tether-Suspended Payloads via Rotating Equilibrium (viability: 7): https://sciencetostartup.com/paper/energy-efficient-collaborative-transport-of-tether-suspended-payloads-via-rotating-equilibrium - Reduce power consumption in collaborative aerial transport by up to 20% using rotating equilibrium, enabling more efficient tethered payload delivery.
- Is Your Safe Controller Actually Safe? A Critical Review of CBF Tautologies and Hidden Assumptions (viability: 7): https://sciencetostartup.com/paper/is-your-safe-controller-actually-safe-a-critical-review-of-cbf-tautologies-and-hidden-assumptions - An interactive tool that validates Control Barrier Functions for robotic safety, preventing unsafe implementations.
- Feasibility Restoration under Conflicting STL Specifications with Pareto-Optimal Refinement (viability: 7): https://sciencetostartup.com/paper/feasibility-restoration-under-conflicting-stl-specifications-with-pareto-optimal-refinement - A two-stage framework restores feasibility in robotics by resolving conflicting STL specifications, enabling safer and more interpretable autonomous driving decisions.
- Deep Research, Shallow Evaluation: A Case Study in Meta-Evaluation for Long-Form QA Benchmarks (viability: 6): https://sciencetostartup.com/paper/deep-research-shallow-evaluation-a-case-study-in-meta-evaluation-for-long-form-qa-benchmarks - A meta-evaluation framework for long-form QA benchmarks, focusing on retrieval-augmented deep-research in the scientific domain, providing guidelines for improved evaluation standards.
- Swimba: Switch Mamba Model Scales State Space Models (viability: 7): https://sciencetostartup.com/paper/swimba-switch-mamba-model-scales-state-space-models - Swimba enhances Mamba models with mixture-of-experts for increased capacity without significantly impacting computational efficiency, making it suitable for real-time applications.
- Extracting and analyzing 3D histomorphometric features related to perineural and lymphovascular invasion in prostate cancer (viability: 7): https://sciencetostartup.com/paper/extracting-and-analyzing-3d-histomorphometric-features-related-to-perineural-and-lymphovascular-invasion-in-prostate-can - A 3D histomorphometric analysis tool for prostate cancer risk assessment, focusing on perineural and lymphovascular invasion, offering improved prognostic value compared to 2D methods.
- HIERAMP: Coarse-to-Fine Autoregressive Amplification for Generative Dataset Distillation (viability: 7): https://sciencetostartup.com/paper/hieramp-coarse-to-fine-autoregressive-amplification-for-generative-dataset-distillation - HIERAMP leverages a vision autoregressive model to amplify hierarchical semantics in dataset distillation, improving validation performance by focusing on discriminative parts and structures.
- A Contrastive Fewshot RGBD Traversability Segmentation Framework for Indoor Robotic Navigation (viability: 7): https://sciencetostartup.com/paper/a-contrastive-fewshot-rgbd-traversability-segmentation-framework-for-indoor-robotic-navigation - A contrastive few-shot RGBD traversability segmentation framework that improves indoor robotic navigation by leveraging negative prototypes and sparse depth information.
- MindfulAgents: Personalizing Mindfulness Meditation via an Expert-Aligned Multi-Agent System (viability: 8): https://sciencetostartup.com/paper/mindfulagents-personalizing-mindfulness-meditation-via-an-expert-aligned-multi-agent-system - MindfulAgents is a personalized mindfulness meditation app using LLMs to improve user engagement and mental well-being.
- Small Target Detection Based on Mask-Enhanced Attention Fusion of Visible and Infrared Remote Sensing Images (viability: 7): https://sciencetostartup.com/paper/small-target-detection-based-on-mask-enhanced-attention-fusion-of-visible-and-infrared-remote-sensing-images - ESM-YOLO+ is a lightweight visible-infrared fusion network that enhances small target detection in remote sensing images with mask-enhanced attention fusion and structural representation enhancement.
- LIPP: Load-Aware Informative Path Planning with Physical Sampling (viability: 7): https://sciencetostartup.com/paper/lipp-load-aware-informative-path-planning-with-physical-sampling - Optimize robot path planning for physical sample collection by considering the load-dependent energy cost, leading to more efficient data acquisition.
- Reforming the Mechanism: Editing Reasoning Patterns in LLMs with Circuit Reshaping (viability: 7): https://sciencetostartup.com/paper/reforming-the-mechanism-editing-reasoning-patterns-in-llms-with-circuit-reshaping - REdit selectively modifies reasoning patterns in LLMs by reshaping neural circuits to improve generality and locality of edits.
- CN-CBF: Composite Neural Control Barrier Function for Safe Robot Navigation in Dynamic Environments (viability: 8): https://sciencetostartup.com/paper/cn-cbf-composite-neural-control-barrier-function-for-safe-robot-navigation-in-dynamic-environments - A neural control barrier function method for safe robot navigation in dynamic environments, demonstrated on both ground robots and quadrotors, offering improved success rates.
- DLRMamba: Distilling Low-Rank Mamba for Edge Multispectral Fusion Object Detection (viability: 7): https://sciencetostartup.com/paper/dlrmamba-distilling-low-rank-mamba-for-edge-multispectral-fusion-object-detection - DLRMamba enables efficient multispectral object detection on edge devices by distilling Mamba models with low-rank approximations and structure-aware distillation.
- SurgSync: Time-Synchronized Multi-Modal Data Collection Framework and Dataset for Surgical Robotics (viability: 8): https://sciencetostartup.com/paper/surgsync-time-synchronized-multi-modal-data-collection-framework-and-dataset-for-surgical-robotics - SurgSync provides a multi-modal data collection framework and dataset for surgical robotics, enabling AI-powered enhancements to robotic surgery systems.
- T2Nav Algebraic Topology Aware Temporal Graph Memory and Loop Detection for ZeroShot Visual Navigation (viability: 7): https://sciencetostartup.com/paper/t2nav-algebraic-topology-aware-temporal-graph-memory-and-loop-detection-for-zeroshot-visual-navigation - T2Nav is a zero-shot navigation system that uses graph-based reasoning and visual information to navigate to target object instances in unknown environments.
- PaQ-DETR: Learning Pattern and Quality-Aware Dynamic Queries for Object Detection (viability: 7): https://sciencetostartup.com/paper/paq-detr-learning-pattern-and-quality-aware-dynamic-queries-for-object-detection - PaQ-DETR enhances object detection by dynamically generating image-specific queries and balancing supervision, leading to improved accuracy and interpretability.
- SysNav: Multi-Level Systematic Cooperation Enables Real-World, Cross-Embodiment Object Navigation (viability: 8): https://sciencetostartup.com/paper/sysnav-multi-level-systematic-cooperation-enables-real-world-cross-embodiment-object-navigation - SysNav is a ready-to-deploy, cross-embodiment object navigation system leveraging VLMs for semantic understanding and hierarchical planning, demonstrating state-of-the-art performance in real-world environments.
- MedInjection-FR: Exploring the Role of Native, Synthetic, and Translated Data in Biomedical Instruction Tuning (viability: 7): https://sciencetostartup.com/paper/medinjection-fr-exploring-the-role-of-native-synthetic-and-translated-data-in-biomedical-instruction-tuning - MedInjection-FR provides a French biomedical instruction dataset and instruction tuning recipes to adapt LLMs for medical question answering.
- XGenBoost: Synthesizing Small and Large Tabular Datasets with XGBoost (viability: 7): https://sciencetostartup.com/paper/xgenboost-synthesizing-small-and-large-tabular-datasets-with-xgboost - XGenBoost synthesizes tabular datasets using XGBoost-based generative models, outperforming existing methods with lower training costs.
- Empowering Locally Deployable Medical Agent via State Enhanced Logical Skills for FHIR-based Clinical Tasks (viability: 8): https://sciencetostartup.com/paper/empowering-locally-deployable-medical-agent-via-state-enhanced-logical-skills-for-fhir-based-clinical-tasks - SELSM enhances locally deployable medical agents by distilling simulated clinical trajectories into entity-agnostic operational rules, achieving state-of-the-art performance on FHIR-based EHR tasks.
- Collaborative Planning with Concurrent Synchronization for Operationally Constrained UAV-UGV Teams (viability: 7): https://sciencetostartup.com/paper/collaborative-planning-with-concurrent-synchronization-for-operationally-constrained-uav-ugv-teams - A learning-based approach for synchronized UAV-UGV collaborative planning, enabling coordinated task execution with energy and traversability constraints.
- Learning From Design Procedure To Generate CAD Programs for Data Augmentation (viability: 7): https://sciencetostartup.com/paper/learning-from-design-procedure-to-generate-cad-programs-for-data-augmentation - Augment CAD training data by prompting LLMs to generate programs conditioned on reference surfaces and modeling procedures, improving geometric diversity for industry-grade designs.
- Single-pass Possibilistic Clustering with Damped Window Footprints (viability: 7): https://sciencetostartup.com/paper/single-pass-possibilistic-clustering-with-damped-window-footprints - A single-pass possibilistic clustering algorithm that models non-spherical clusters and is easy to apply to new datasets, ideal for real-time data analysis.
- Enhancing the Detection of Coronary Artery Disease Using Machine Learning (viability: 7): https://sciencetostartup.com/paper/enhancing-the-detection-of-coronary-artery-disease-using-machine-learning - A machine learning model for early detection of Coronary Artery Disease (CAD) that outperforms traditional diagnostic methods.
- VertiAdaptor: Online Kinodynamics Adaptation for Vertically Challenging Terrain (viability: 7): https://sciencetostartup.com/paper/vertiadaptor-online-kinodynamics-adaptation-for-vertically-challenging-terrain - VertiAdaptor enables autonomous vehicles to navigate challenging off-road terrain by rapidly adapting kinodynamic models using elevation and semantic embeddings.
- OPTED: Open Preprocessed Trachoma Eye Dataset Using Zero-Shot SAM 3 Segmentation (viability: 7): https://sciencetostartup.com/paper/opted-open-preprocessed-trachoma-eye-dataset-using-zero-shot-sam-3-segmentation - We provide an open-source, preprocessed trachoma eye dataset (OPTED) and code to enable automated trachoma classification, addressing a critical need in Sub-Saharan Africa.
- Physics-informed AI Accelerated Retention Analysis of Ferroelectric Vertical NAND: From Day-Scale TCAD to Second-Scale Surrogate Model (viability: 7): https://sciencetostartup.com/paper/physics-informed-ai-accelerated-retention-analysis-of-ferroelectric-vertical-nand-from-day-scale-tcad-to-second-scale-su - An AI surrogate model accelerates ferroelectric vertical NAND design by predicting threshold voltage shifts and retention behavior with 10000x speedup compared to TCAD.
- PICS: Pairwise Image Compositing with Spatial Interactions (viability: 7): https://sciencetostartup.com/paper/pics-pairwise-image-compositing-with-spatial-interactions - PICS offers a stable and spatially coherent image compositing solution using a self-supervised, interaction-aware diffusion approach, ideal for virtual try-on and scene editing applications.
- LEAD: Breaking the No-Recovery Bottleneck in Long-Horizon Reasoning (viability: 7): https://sciencetostartup.com/paper/lead-breaking-the-no-recovery-bottleneck-in-long-horizon-reasoning - LEAD enhances LLM long-horizon reasoning by mitigating the no-recovery bottleneck through lookahead validation and overlapping rollouts, enabling more stable and accurate task execution.
- CAR: Cross-Vehicle Kinodynamics Adaptation via Mobility Representation (viability: 7): https://sciencetostartup.com/paper/car-cross-vehicle-kinodynamics-adaptation-via-mobility-representation - CAR enables rapid mobility transfer to new autonomous vehicles by embedding vehicle trajectory transitions and physical configurations into a shared mobility latent space.
- A prior information informed learning architecture for flying trajectory prediction (viability: 7): https://sciencetostartup.com/paper/a-prior-information-informed-learning-architecture-for-flying-trajectory-prediction - Predict landing points of flying objects using a hardware-efficient, environment-aware dual-transformer architecture, starting with tennis ball landing prediction.
- Supporting Artifact Evaluation with LLMs: A Study with Published Security Research Papers (viability: 7): https://sciencetostartup.com/paper/supporting-artifact-evaluation-with-llms-a-study-with-published-security-research-papers - Automated artifact evaluation toolkit for cybersecurity research papers, using LLMs to assess reproducibility and methodological pitfalls, reducing reviewer effort and improving research quality.
- IGLU: The Integrated Gaussian Linear Unit Activation Function (viability: 7): https://sciencetostartup.com/paper/iglu-the-integrated-gaussian-linear-unit-activation-function - IGLU is a novel activation function that improves performance on vision and language tasks, offering a computationally efficient alternative to GELU with robustness to vanishing gradients.
- ColonSplat: Reconstruction of Peristaltic Motion in Colonoscopy with Dynamic Gaussian Splatting (viability: 7): https://sciencetostartup.com/paper/colonsplat-reconstruction-of-peristaltic-motion-in-colonoscopy-with-dynamic-gaussian-splatting - ColonSplat reconstructs peristaltic motion in colonoscopy data using dynamic Gaussian Splatting, enabling advanced surgical navigation and retrospective diagnostics.
- Contextual Counterfactual Credit Assignment for Multi-Agent Reinforcement Learning in LLM Collaboration (viability: 7): https://sciencetostartup.com/paper/contextual-counterfactual-credit-assignment-for-multi-agent-reinforcement-learning-in-llm-collaboration - C3 isolates the causal impact of individual messages in cooperative multi-agent LLM systems, improving credit assignment and terminal performance.
- Are Audio-Language Models Listening? Audio-Specialist Heads for Adaptive Audio Steering (viability: 7): https://sciencetostartup.com/paper/are-audio-language-models-listening-audio-specialist-heads-for-adaptive-audio-steering - Amplify audio understanding in large audio-language models by identifying and steering audio-specialist attention heads, improving accuracy without retraining.
- An Extended Topological Model For High-Contrast Optical Flow (viability: 4): https://sciencetostartup.com/paper/an-extended-topological-model-for-high-contrast-optical-flow - This paper introduces a novel 3-manifold model for high-contrast optical flow patches, offering insights into motion boundaries for improved object segmentation and tracking.
- Active View Selection with Perturbed Gaussian Ensemble for Tomographic Reconstruction (viability: 7): https://sciencetostartup.com/paper/active-view-selection-with-perturbed-gaussian-ensemble-for-tomographic-reconstruction - An active view selection framework that integrates uncertainty modeling with sequential decision-making, tailored for X-ray Gaussian Splatting, to improve sparse-view CT reconstruction.
- Nonlinear Performance Degradation of Vision-Based Teleoperation under Network Latency (viability: 7): https://sciencetostartup.com/paper/nonlinear-performance-degradation-of-vision-based-teleoperation-under-network-latency - Latency-Aware Vision Teleoperation (LAVT) is a ROS 2 framework that enables precise, distributed one-way latency measurement and reproducible delay injection, providing a baseline for future latency-compensation and predictive control strategies in teleoperation.
- MotionBits: Video Segmentation through Motion-Level Analysis of Rigid Bodies (viability: 8): https://sciencetostartup.com/paper/motionbits-video-segmentation-through-motion-level-analysis-of-rigid-bodies - MotionBits provides a novel, learning-free approach to video segmentation for robotic manipulation by identifying moving rigid bodies, enabling more effective embodied reasoning and manipulation tasks.
- RoboCritics: Enabling Reliable End-to-End LLM Robot Programming through Expert-Informed Critics (viability: 7): https://sciencetostartup.com/paper/robocritics-enabling-reliable-end-to-end-llm-robot-programming-through-expert-informed-critics - RoboCritics enhances LLM-based robot programming with expert-informed critics, providing safety checks and iterative refinement for more reliable and user-centered robotic tasking.
- Validation of a Small Language Model for DSM-5 Substance Category Classification in Child Welfare Records (viability: 7): https://sciencetostartup.com/paper/validation-of-a-small-language-model-for-dsm-5-substance-category-classification-in-child-welfare-records - A locally hosted language model accurately classifies substance types in child welfare records, enabling automated risk assessment and intervention.
- Receding-Horizon Nullspace Optimization for Actuation-Aware Control Allocation in Omnidirectional UAVs (viability: 7): https://sciencetostartup.com/paper/receding-horizon-nullspace-optimization-for-actuation-aware-control-allocation-in-omnidirectional-uavs - A receding-horizon control allocation strategy for omnidirectional UAVs that reduces motor command oscillations and improves trajectory tracking, potentially enabling more agile and precise aerial maneuvers.
- Learning-Based Robust Control: Unifying Exploration and Distributional Robustness for Reliable Robotics via Free Energy (viability: 7): https://sciencetostartup.com/paper/learning-based-robust-control-unifying-exploration-and-distributional-robustness-for-reliable-robotics-via-free-energy - A robust robotic control system leveraging free energy principles for reliable manipulation, validated in simulation and real-world experiments.
- Joint 3D Gravity and Magnetic Inversion via Rectified Flow and Ginzburg-Landau Guidance (viability: 7): https://sciencetostartup.com/paper/joint-3d-gravity-and-magnetic-inversion-via-rectified-flow-and-ginzburg-landau-guidance - A novel framework for 3D gravity and magnetic joint inversion using rectified flow and Ginzburg-Landau guidance, offering improved ore detection and uncertainty modeling.
- Step-Level Visual Grounding Faithfulness Predicts Out-of-Distribution Generalization in Long-Horizon Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/step-level-visual-grounding-faithfulness-predicts-out-of-distribution-generalization-in-long-horizon-vision-language-mod - Improve the robustness of vision-language models by measuring and optimizing the temporal grounding of their reasoning steps with visual input.
- CREDO: Epistemic-Aware Conformalized Credal Envelopes for Regression (viability: 7): https://sciencetostartup.com/paper/credo-epistemic-aware-conformalized-credal-envelopes-for-regression - CREDO provides calibrated prediction intervals for regression that explicitly represent and adapt to epistemic uncertainty, making it suitable for applications requiring reliable uncertainty estimates.
- A Comprehensive Analysis of the Effects of Network Quality of Service on Robotic Telesurgery (viability: 7): https://sciencetostartup.com/paper/a-comprehensive-analysis-of-the-effects-of-network-quality-of-service-on-robotic-telesurgery - NetFI is an open-source tool that emulates network conditions for telesurgery simulations, providing insights into the impact of network quality on task performance and operator workload.
- "Dark Triad" Model Organisms of Misalignment: Narrow Fine-Tuning Mirrors Human Antisocial Behavior (viability: 7): https://sciencetostartup.com/paper/dark-triad-model-organisms-of-misalignment-narrow-fine-tuning-mirrors-human-antisocial-behavior - Fine-tune LLMs to exhibit 'Dark Triad' personality traits for studying and mitigating misalignment risks, offering a novel approach to AI safety.
- A Hybrid Machine Learning Model for Cerebral Palsy Detection (viability: 7): https://sciencetostartup.com/paper/a-hybrid-machine-learning-model-for-cerebral-palsy-detection - An AI model for early detection of Cerebral Palsy from MRI images, achieving high accuracy by combining CNNs and Bi-LSTM.
- Breaking the Martingale Curse: Multi-Agent Debate via Asymmetric Cognitive Potential Energy (viability: 8): https://sciencetostartup.com/paper/breaking-the-martingale-curse-multi-agent-debate-via-asymmetric-cognitive-potential-energy - AceMAD leverages cognitive potential energy in multi-agent debate to overcome the Martingale Curse, enabling more accurate reasoning and consensus-building in LLMs.
- NEST: Network- and Memory-Aware Device Placement For Distributed Deep Learning (viability: 7): https://sciencetostartup.com/paper/nest-network-and-memory-aware-device-placement-for-distributed-deep-learning - NEST optimizes distributed deep learning training by intelligently placing model components across devices, considering network topology, memory constraints, and parallelism strategies.
- Best-of-Tails: Bridging Optimism and Pessimism in Inference-Time Alignment (viability: 7): https://sciencetostartup.com/paper/best-of-tails-bridging-optimism-and-pessimism-in-inference-time-alignment - Best-of-Tails (BoT) adaptively steers LLMs at inference time by dynamically adjusting the exploration-exploitation balance based on reward-tail heaviness, improving alignment performance across various tasks.
- Optimistic Policy Regularization (viability: 8): https://sciencetostartup.com/paper/optimistic-policy-regularization - Optimistic Policy Regularization improves reinforcement learning sample efficiency by reinforcing successful trajectories, offering a drop-in enhancement for existing PPO agents.
- Omni-Diffusion: Unified Multimodal Understanding and Generation with Masked Discrete Diffusion (viability: 7): https://sciencetostartup.com/paper/omni-diffusion-unified-multimodal-understanding-and-generation-with-masked-discrete-diffusion - Omni-Diffusion is a multimodal language model built on mask-based discrete diffusion, unifying understanding and generation across text, speech, and images, offering a novel approach to multimodal tasks.
- BEVLM: Distilling Semantic Knowledge from LLMs into Bird's-Eye View Representations (viability: 2): https://sciencetostartup.com/paper/bevlm-distilling-semantic-knowledge-from-llms-into-bird-s-eye-view-representations - BEVLM enhances autonomous driving by integrating LLMs with spatially consistent BEV representations for improved semantic reasoning and safety.
- Fly360: Omnidirectional Obstacle Avoidance within Drone View (viability: 7): https://sciencetostartup.com/paper/fly360-omnidirectional-obstacle-avoidance-within-drone-view - Fly360 enables drones to navigate complex environments with full-view perception, offering a robust obstacle avoidance solution.
- SUREON: A Benchmark and Vision-Language-Model for Surgical Reasoning (viability: 8): https://sciencetostartup.com/paper/sureon-a-benchmark-and-vision-language-model-for-surgical-reasoning - SUREON provides a surgical video QA dataset and fine-tuned vision-language models that enable surgical AI to reason about surgical procedures, offering a potential tool for training and assistance.
- Penguin-VL: Exploring the Efficiency Limits of VLM with LLM-based Vision Encoders (viability: 8): https://sciencetostartup.com/paper/penguin-vl-exploring-the-efficiency-limits-of-vlm-with-llm-based-vision-encoders - Penguin-VL offers a lightweight, high-fidelity VLM solution for resource-constrained devices, outperforming leading VLMs in key tasks by leveraging a novel LLM-based vision encoder.
- Boosting deep Reinforcement Learning using pretraining with Logical Options (viability: 2): https://sciencetostartup.com/paper/boosting-deep-reinforcement-learning-using-pretraining-with-logical-options - Hybrid Hierarchical RL (H^2RL) enhances deep reinforcement learning by integrating symbolic pretraining to improve decision-making.
- Physics-Informed Diffusion Model for Generating Synthetic Extreme Rare Weather Events Data (viability: 7): https://sciencetostartup.com/paper/physics-informed-diffusion-model-for-generating-synthetic-extreme-rare-weather-events-data - Generate synthetic weather data using a physics-informed diffusion model to augment training datasets for operational weather detection algorithms.
- EgoReasoner: Learning Egocentric 4D Reasoning via Task-Adaptive Structured Thinking (viability: 7): https://sciencetostartup.com/paper/egoreasoner-learning-egocentric-4d-reasoning-via-task-adaptive-structured-thinking - EgoReasoner enhances egocentric video understanding by adaptively structuring reasoning for specific 4D tasks, achieving state-of-the-art results on the HD-EPIC benchmark.
- Causal Interpretation of Neural Network Computations with Contribution Decomposition (viability: 7): https://sciencetostartup.com/paper/causal-interpretation-of-neural-network-computations-with-contribution-decomposition - CODEC decomposes neural network behavior into sparse motifs of hidden-neuron contributions, enabling causal manipulations and human-interpretable visualizations.
- xaitimesynth: A Python Package for Evaluating Attribution Methods for Time Series with Synthetic Ground Truth (viability: 7): https://sciencetostartup.com/paper/xaitimesynth-a-python-package-for-evaluating-attribution-methods-for-time-series-with-synthetic-ground-truth - xaitimesynth is a Python package that provides reusable infrastructure for evaluating time series attribution methods with synthetic ground truth, enabling developers to benchmark and improve their XAI techniques.
- Hierarchical Industrial Demand Forecasting with Temporal and Uncertainty Explanations (viability: 7): https://sciencetostartup.com/paper/hierarchical-industrial-demand-forecasting-with-temporal-and-uncertainty-explanations - Explainable hierarchical time-series forecasting for industrial demand prediction, enabling informed decision-making and strategic planning.
- Uncertainty-Aware Adaptive Dynamics For Underwater Vehicle-Manipulator Robots (viability: 7): https://sciencetostartup.com/paper/uncertainty-aware-adaptive-dynamics-for-underwater-vehicle-manipulator-robots - An uncertainty-aware adaptive dynamics model for underwater robots enables more reliable control and simulation.
- Modeling and Measuring Redundancy in Multisource Multimodal Data for Autonomous Driving (viability: 7): https://sciencetostartup.com/paper/modeling-and-measuring-redundancy-in-multisource-multimodal-data-for-autonomous-driving - Improve autonomous vehicle object detection by intelligently removing redundant labels from multisource, multimodal data, leading to better performance and data efficiency.
- SurgFormer: Scalable Learning of Organ Deformation with Resection Support and Real-Time Inference (viability: 7): https://sciencetostartup.com/paper/surgformer-scalable-learning-of-organ-deformation-with-resection-support-and-real-time-inference - SurgFormer is a fast and accurate tool for simulating soft tissue deformation during surgery, enabling real-time surgical planning and training.
- RAMoEA-QA: Hierarchical Specialization for Robust Respiratory Audio Question Answering (viability: 3): https://sciencetostartup.com/paper/ramoea-qa-hierarchical-specialization-for-robust-respiratory-audio-question-answering - Develops a specialized AI model for robust question answering in respiratory audio analysis, integrating heterogeneous patient signals within a single system.
- SpatialMAGIC: A Hybrid Framework Integrating Graph Diffusion and Spatial Attention for Spatial Transcriptomics Imputation (viability: 8): https://sciencetostartup.com/paper/spatialmagic-a-hybrid-framework-integrating-graph-diffusion-and-spatial-attention-for-spatial-transcriptomics-imputation - SpatialMagic is a hybrid imputation model for spatial transcriptomics data that combines graph diffusion and spatial attention to improve data quality and biological interpretability, offering a potential tool for researchers in the field.
- Proteus: A Practical Framework for Privacy-Preserving Device Logs (viability: 7): https://sciencetostartup.com/paper/proteus-a-practical-framework-for-privacy-preserving-device-logs - Proteus is a privacy-preserving device logging framework that enables forensic analysis without disclosing plaintext PII, offering a secure solution for enterprise monitoring and fraud detection.
- Unified Learning of Temporal Task Structure and Action Timing for Bimanual Robot Manipulation (viability: 7): https://sciencetostartup.com/paper/unified-learning-of-temporal-task-structure-and-action-timing-for-bimanual-robot-manipulation - Learn bimanual robot manipulation skills from human demonstrations by unifying temporal task structure and action timing, enabling robots to execute complex tasks more naturally.
- A Multi-Layer Sim-to-Real Framework for Gaze-Driven Assistive Neck Exoskeletons (viability: 7): https://sciencetostartup.com/paper/a-multi-layer-sim-to-real-framework-for-gaze-driven-assistive-neck-exoskeletons - A VR-validated gaze-tracking system for assistive neck exoskeletons, enabling personalized head movement prediction and control.
- NEGATE: Constrained Semantic Guidance for Linguistic Negation in Text-to-Video Diffusion (viability: 7): https://sciencetostartup.com/paper/negate-constrained-semantic-guidance-for-linguistic-negation-in-text-to-video-diffusion - A training-free method for improved negation handling in text-to-video diffusion models, enabling more accurate and nuanced video generation from text prompts.
- Spatial Calibration of Diffuse LiDARs (viability: 7): https://sciencetostartup.com/paper/spatial-calibration-of-diffuse-lidars - Calibrate diffuse LiDARs with RGB cameras using a novel spatial calibration procedure, enabling accurate cross-modal alignment and fusion for enhanced perception.
- AV-Unified: A Unified Framework for Audio-visual Scene Understanding (viability: 7): https://sciencetostartup.com/paper/av-unified-a-unified-framework-for-audio-visual-scene-understanding - AV-Unified is a unified framework for audio-visual scene understanding that jointly learns across multiple tasks by standardizing input-output formats and incorporating a multi-scale spatiotemporal perception network.
- Underactuated multimodal jumping robot for extraterrestrial exploration (viability: 7): https://sciencetostartup.com/paper/underactuated-multimodal-jumping-robot-for-extraterrestrial-exploration - A compact, underactuated jumping robot for extraterrestrial exploration, controllable with only two reaction wheels, enabling versatile locomotion in challenging environments.
- Artificial Intelligence for Detecting Fetal Orofacial Clefts and Advancing Medical Education (viability: 8): https://sciencetostartup.com/paper/artificial-intelligence-for-detecting-fetal-orofacial-clefts-and-advancing-medical-education - AI-powered medical copilot for prenatal orofacial cleft detection, improving diagnostic accuracy and accelerating specialist training.
- HGT-Scheduler: Deep Reinforcement Learning for the Job Shop Scheduling Problem via Heterogeneous Graph Transformers (viability: 7): https://sciencetostartup.com/paper/hgt-scheduler-deep-reinforcement-learning-for-the-job-shop-scheduling-problem-via-heterogeneous-graph-transformers - An RL-based job scheduler leveraging heterogeneous graph transformers to optimize job shop scheduling, outperforming homogeneous graph approaches on standard benchmarks.
- HybridMimic: Hybrid RL-Centroidal Control for Humanoid Motion Mimicking (viability: 7): https://sciencetostartup.com/paper/hybridmimic-hybrid-rl-centroidal-control-for-humanoid-motion-mimicking - HybridMimic uses reinforcement learning to dynamically modulate a centroidal-model-based controller for humanoid robots, improving robustness and reducing tracking error.
- Stability-Guided Exploration for Diverse Motion Generation (viability: 6): https://sciencetostartup.com/paper/stability-guided-exploration-for-diverse-motion-generation - A novel method for generating diverse robot manipulation strategies through black-box simulation, enabling robots to learn complex tasks without human demonstrations.
- SG-DOR: Learning Scene Graphs with Direction-Conditioned Occlusion Reasoning for Pepper Plants (viability: 7): https://sciencetostartup.com/paper/sg-dor-learning-scene-graphs-with-direction-conditioned-occlusion-reasoning-for-pepper-plants - A scene graph framework for robotic harvesting that infers physical attachments and direction-conditioned occlusion, enabling more effective interventions in dense crop canopies.
- Semantics-Aware Caching for Concept Learning (viability: 7): https://sciencetostartup.com/paper/semantics-aware-caching-for-concept-learning - A semantics-aware caching approach that speeds up concept learning by an order of magnitude, applicable to both symbolic and neuro-symbolic reasoners.
- Beyond Rows to Reasoning: Agentic Retrieval for Multimodal Spreadsheet Understanding and Editing (viability: 8): https://sciencetostartup.com/paper/beyond-rows-to-reasoning-agentic-retrieval-for-multimodal-spreadsheet-understanding-and-editing - BRTR is an agentic framework that enables LLMs to understand and edit complex spreadsheets through iterative tool-calling, achieving state-of-the-art performance on multiple benchmarks.
- CFEAR-Teach-and-Repeat: Fast and Accurate Radar-only Localization (viability: 8): https://sciencetostartup.com/paper/cfear-teach-and-repeat-fast-and-accurate-radar-only-localization - CFEAR-TR provides robust and accurate radar-only localization for autonomous navigation, offering a deployable solution for adverse weather conditions.
- A Unified Low-Dimensional Design Embedding for Joint Optimization of Shape, Material, and Actuation in Soft Robots (viability: 7): https://sciencetostartup.com/paper/a-unified-low-dimensional-design-embedding-for-joint-optimization-of-shape-material-and-actuation-in-soft-robots - A low-dimensional design embedding for soft robots enables joint optimization of shape, material, and actuation, leading to more efficient co-design.
- COLD-Steer: Steering Large Language Models via In-Context One-step Learning Dynamics (viability: 7): https://sciencetostartup.com/paper/cold-steer-steering-large-language-models-via-in-context-one-step-learning-dynamics - COLD-Steer enables efficient, training-free control of LLM behavior at inference time, requiring significantly fewer examples than existing methods.
- Control Barrier Corridors: From Safety Functions to Safe Sets (viability: 7): https://sciencetostartup.com/paper/control-barrier-corridors-from-safety-functions-to-safe-sets - A novel method unifying control barrier functions and safe motion corridors for safer robot navigation, enabling verifiable safe path following in unknown environments.
- NOBLE: Accelerating Transformers with Nonlinear Low-Rank Branches (viability: 5): https://sciencetostartup.com/paper/noble-accelerating-transformers-with-nonlinear-low-rank-branches - NOBLE enhances transformer architectures with nonlinear low-rank branches for faster and efficient pretraining.
- PONTE: Personalized Orchestration for Natural Language Trustworthy Explanations (viability: 5): https://sciencetostartup.com/paper/ponte-personalized-orchestration-for-natural-language-trustworthy-explanations - PONTE enables personalized, reliable AI explanations by adapting to user preferences in real-time.
- History-Conditioned Spatio-Temporal Visual Token Pruning for Efficient Vision-Language Navigation (viability: 8): https://sciencetostartup.com/paper/history-conditioned-spatio-temporal-visual-token-pruning-for-efficient-vision-language-navigation - A training-free token pruning framework that significantly improves the efficiency of vision-language navigation for robots, enabling real-time deployment.
- Match4Annotate: Propagating Sparse Video Annotations via Implicit Neural Feature Matching (viability: 7): https://sciencetostartup.com/paper/match4annotate-propagating-sparse-video-annotations-via-implicit-neural-feature-matching - Match4Annotate offers a lightweight framework for propagating video annotations, enabling efficient annotation workflows, especially in specialized domains like medical imaging.
- GreenRFM: Toward a resource-efficient radiology foundation model (viability: 8): https://sciencetostartup.com/paper/greenrfm-toward-a-resource-efficient-radiology-foundation-model - GreenRFM provides resource-efficient radiology foundation models that achieve state-of-the-art performance on a single GPU, enabling accessible and robust clinical applications.
- Bayesian Additive Distribution Regression (viability: 7): https://sciencetostartup.com/paper/bayesian-additive-distribution-regression - DistBART is a scalable Bayesian nonparametric approach to distribution regression, modeling the regression function as a linear functional with a BART prior, enabling prediction of scalar responses from distribution-valued predictors.
- Do Foundation Models Know Geometry? Probing Frozen Features for Continuous Physical Measurement (viability: 2): https://sciencetostartup.com/paper/do-foundation-models-know-geometry-probing-frozen-features-for-continuous-physical-measurement - Study explores geometric capabilities of vision-language models for potential use as geometric sensors without fine-tuning.
- Training Flow Matching: The Role of Weighting and Parameterization (viability: 6): https://sciencetostartup.com/paper/training-flow-matching-the-role-of-weighting-and-parameterization - A systematic analysis of training objectives for denoising-based generative models, providing practical insights for design choices.
- Pinterest Canvas: Large-Scale Image Generation at Pinterest (viability: 8): https://sciencetostartup.com/paper/pinterest-canvas-large-scale-image-generation-at-pinterest - Pinterest Canvas is a large-scale image generation system that fine-tunes diffusion models for specific image editing tasks, resulting in significant engagement lifts.
- Data Analogies Enable Efficient Cross-Embodiment Transfer (viability: 7): https://sciencetostartup.com/paper/data-analogies-enable-efficient-cross-embodiment-transfer - Improve robot transfer learning by 22.5% by using data analogies to align demonstrations across different robot embodiments.
- CaTok: Taming Mean Flows for One-Dimensional Causal Image Tokenization (viability: 7): https://sciencetostartup.com/paper/catok-taming-mean-flows-for-one-dimensional-causal-image-tokenization - CaTok is a 1D causal image tokenizer with a MeanFlow decoder that enables fast one-step generation and high-fidelity multi-step sampling, achieving state-of-the-art results on ImageNet reconstruction.
- What if? Emulative Simulation with World Models for Situated Reasoning (viability: 8): https://sciencetostartup.com/paper/what-if-emulative-simulation-with-world-models-for-situated-reasoning - WanderDream enables robots and visually impaired users to mentally simulate trajectories and answer spatial what-if questions, facilitating safer and more efficient exploration.
- Prosodic Boundary-Aware Streaming Generation for LLM-Based TTS with Streaming Text Input (viability: 5): https://sciencetostartup.com/paper/prosodic-boundary-aware-streaming-generation-for-llm-based-tts-with-streaming-text-input - Develop a prosodic boundary-aware streaming TTS for interactive systems, enhancing naturalness and speaker similarity in long-form text synthesis.
- Toward Generative Quantum Utility via Correlation-Complexity Map (viability: 7): https://sciencetostartup.com/paper/toward-generative-quantum-utility-via-correlation-complexity-map - A novel approach using Correlation-Complexity Map to identify data distributions suitable for IQP-type quantum generative models, demonstrating competitive performance against classical models with fewer training samples.
- Abductive Reasoning with Syllogistic Forms in Large Language Models (viability: 5): https://sciencetostartup.com/paper/abductive-reasoning-with-syllogistic-forms-in-large-language-models - Evaluate and improve LLMs' abductive reasoning capabilities by converting syllogistic datasets for abduction tasks.
- CLoPA: Continual Low Parameter Adaptation of Interactive Segmentation for Medical Image Annotation (viability: 4): https://sciencetostartup.com/paper/clopa-continual-low-parameter-adaptation-of-interactive-segmentation-for-medical-image-annotation - CLoPA enhances interactive segmentation tools for more accurate and efficient medical image annotation without changing the existing workflow.
- From Prompting to Preference Optimization: A Comparative Study of LLM-based Automated Essay Scoring (viability: 7): https://sciencetostartup.com/paper/from-prompting-to-preference-optimization-a-comparative-study-of-llm-based-automated-essay-scoring - Automated essay scoring API for English as a Second Language (ESL) writing assessment, leveraging LLMs and preference optimization.
- Before You Hand Over the Wheel: Evaluating LLMs for Security Incident Analysis (viability: 7): https://sciencetostartup.com/paper/before-you-hand-over-the-wheel-evaluating-llms-for-security-incident-analysis - SIABENCH provides an agentic evaluation framework and dataset for benchmarking LLMs in security incident analysis, enabling better assessment and design decisions for LLM-powered security tools.
- Non-invasive Growth Monitoring of Small Freshwater Fish in Home Aquariums via Stereo Vision (viability: 7): https://sciencetostartup.com/paper/non-invasive-growth-monitoring-of-small-freshwater-fish-in-home-aquariums-via-stereo-vision - A stereo vision system for non-invasive fish growth monitoring in home aquariums, enabling automated health tracking.
- Evaluation of Deontic Conditional Reasoning in Large Language Models: The Case of Wason's Selection Task (viability: 7): https://sciencetostartup.com/paper/evaluation-of-deontic-conditional-reasoning-in-large-language-models-the-case-of-wason-s-selection-task - Leverage LLMs' improved deontic reasoning with a Wason Selection Task dataset to build a bias detection and mitigation tool.
- A Reference Architecture of Reinforcement Learning Frameworks (viability: 4): https://sciencetostartup.com/paper/a-reference-architecture-of-reinforcement-learning-frameworks - A reference architecture for RL frameworks to standardize comparison and integration, enabling faster development of RL applications.
- Physical Simulator In-the-Loop Video Generation (viability: 7): https://sciencetostartup.com/paper/physical-simulator-in-the-loop-video-generation - Generate physically realistic videos by integrating a physics simulator into the video diffusion process, ensuring adherence to physical laws and improving texture consistency.
- Adapter-Augmented Bandits for Online Multi-Constrained Multi-Modal Inference Scheduling (viability: 7): https://sciencetostartup.com/paper/adapter-augmented-bandits-for-online-multi-constrained-multi-modal-inference-scheduling - Optimize MLLM inference scheduling with a multi-armed bandit approach for heterogeneous backends and budget constraints.
- DiffInf: Influence-Guided Diffusion for Supervision Alignment in Facial Attribute Learning (viability: 7): https://sciencetostartup.com/paper/diffinf-influence-guided-diffusion-for-supervision-alignment-in-facial-attribute-learning - DiffInf uses self-influence guided diffusion to correct annotation inconsistencies in facial attribute datasets, improving the accuracy of facial attribute classifiers.
- Efficient, Property-Aligned Fan-Out Retrieval via RL-Compiled Diffusion (viability: 7): https://sciencetostartup.com/paper/efficient-property-aligned-fan-out-retrieval-via-rl-compiled-diffusion - Use reinforcement learning to train a diffusion model for efficient set-valued retrieval, enabling faster and better results for tasks like fashion and music recommendations.
- Talk Freely, Execute Strictly: Schema-Gated Agentic AI for Flexible and Reproducible Scientific Workflows (viability: 7): https://sciencetostartup.com/paper/talk-freely-execute-strictly-schema-gated-agentic-ai-for-flexible-and-reproducible-scientific-workflows - Schema-gated orchestration enables deterministic and flexible scientific workflows using LLMs by separating conversational and execution authority.
- Solving Jigsaw Puzzles in the Wild: Human-Guided Reconstruction of Cultural Heritage Fragments (viability: 7): https://sciencetostartup.com/paper/solving-jigsaw-puzzles-in-the-wild-human-guided-reconstruction-of-cultural-heritage-fragments - An interactive tool for archaeologists to reconstruct fragmented artifacts using AI-assisted puzzle solving.
- REACT++: Efficient Cross-Attention for Real-Time Scene Graph Generation (viability: 8): https://sciencetostartup.com/paper/react-efficient-cross-attention-for-real-time-scene-graph-generation - REACT++ is a state-of-the-art scene graph generation model that balances speed and accuracy, enabling real-time applications for embodied agents.
- Prompt Group-Aware Training for Robust Text-Guided Nuclei Segmentation (viability: 7): https://sciencetostartup.com/paper/prompt-group-aware-training-for-robust-text-guided-nuclei-segmentation - Improve the robustness of text-guided medical image segmentation by making it less sensitive to prompt variations, leading to more reliable clinical workflows.
- CHMv2: Improvements in Global Canopy Height Mapping using DINOv3 (viability: 7): https://sciencetostartup.com/paper/chmv2-improvements-in-global-canopy-height-mapping-using-dinov3 - CHMv2 provides a global, meter-resolution canopy height map derived from satellite imagery, improving accuracy and detail for forestry applications.
- MoEMambaMIL: Structure-Aware Selective State Space Modeling for Whole-Slide Image Analysis (viability: 7): https://sciencetostartup.com/paper/moemambamil-structure-aware-selective-state-space-modeling-for-whole-slide-image-analysis - MoEMambaMIL is a structure-aware State Space Model framework for whole-slide image analysis that integrates region-nested selective scanning with mixture-of-experts modeling, achieving state-of-the-art performance across multiple downstream tasks.
- Failure Detection in Chemical Processes using Symbolic Machine Learning: A Case Study on Ethylene Oxidation (viability: 7): https://sciencetostartup.com/paper/failure-detection-in-chemical-processes-using-symbolic-machine-learning-a-case-study-on-ethylene-oxidation - Predict chemical process failures with interpretable rule-based models, enabling safer and more efficient operations.
- Rewis3d: Reconstruction Improves Weakly-Supervised Semantic Segmentation (viability: 7): https://sciencetostartup.com/paper/rewis3d-reconstruction-improves-weakly-supervised-semantic-segmentation - Rewis3d leverages 3D scene reconstruction to improve weakly supervised semantic segmentation, offering a cost-effective alternative to dense annotations.
- OralGPT-Plus: Learning to Use Visual Tools via Reinforcement Learning for Panoramic X-ray Analysis (viability: 7): https://sciencetostartup.com/paper/oralgpt-plus-learning-to-use-visual-tools-via-reinforcement-learning-for-panoramic-x-ray-analysis - OralGPT-Plus is an agentic vision-language model that uses reinforcement learning to perform iterative and symmetry-aware diagnostic reasoning for panoramic dental radiograph analysis, improving clinical reliability.
- ESAA-Security: An Event-Sourced, Verifiable Architecture for Agent-Assisted Security Audits of AI-Generated Code (viability: 8): https://sciencetostartup.com/paper/esaa-security-an-event-sourced-verifiable-architecture-for-agent-assisted-security-audits-of-ai-generated-code - ESAA-Security provides a verifiable architecture for agent-assisted security audits of AI-generated code, ensuring traceable and reproducible risk-oriented reports.
- Computer vision-based estimation of invertebrate biomass (viability: 7): https://sciencetostartup.com/paper/computer-vision-based-estimation-of-invertebrate-biomass - Automatically estimate invertebrate biomass from images, enabling scalable biodiversity monitoring.
- CLAIRE: Compressed Latent Autoencoder for Industrial Representation and Evaluation -- A Deep Learning Framework for Smart Manufacturing (viability: 7): https://sciencetostartup.com/paper/claire-compressed-latent-autoencoder-for-industrial-representation-and-evaluation-a-deep-learning-framework-for-smart-ma - CLAIRE is a deep learning framework for fault detection in smart manufacturing, using autoencoders and game-theory-based interpretability to improve accuracy and explainability.
- Tiny, Hardware-Independent, Compression-based Classification (viability: 7): https://sciencetostartup.com/paper/tiny-hardware-independent-compression-based-classification - Enable client-side machine learning with a hardware-independent, compression-based classification model that protects user privacy and minimizes computational costs.
- Safe Consensus of Cooperative Manipulation with Hierarchical Event-Triggered Control Barrier Functions (viability: 8): https://sciencetostartup.com/paper/safe-consensus-of-cooperative-manipulation-with-hierarchical-event-triggered-control-barrier-functions - A distributed control framework for multi-robot cooperative manipulation that ensures safety and reduces computational cost, validated with real-world experiments.
- Dynamic Chunking Diffusion Transformer (viability: 3): https://sciencetostartup.com/paper/dynamic-chunking-diffusion-transformer - Adaptive image processing transformer with potential for efficient visual segmentations.
- MoEless: Efficient MoE LLM Serving via Serverless Computing (viability: 3): https://sciencetostartup.com/paper/moeless-efficient-moe-llm-serving-via-serverless-computing - MoEless provides a serverless framework for balancing load in MoE LLM deployments to reduce inference latency and cost.
- Transparent AI for Mathematics: Transformer-Based Large Language Models for Mathematical Entity Relationship Extraction with XAI (viability: 7): https://sciencetostartup.com/paper/transparent-ai-for-mathematics-transformer-based-large-language-models-for-mathematical-entity-relationship-extraction-w - Extract mathematical relationships from text with high accuracy using BERT and SHAP explainability, enabling automated problem solving and intelligent education.
- Open-Source Based and ETSI Compliant Cooperative, Connected, and Automated Mini-Cars (viability: 7): https://sciencetostartup.com/paper/open-source-based-and-etsi-compliant-cooperative-connected-and-automated-mini-cars - A 1:10 scale autonomous vehicle platform with open-source software and ETSI C-ITS compliance, enabling rapid prototyping and testing of cooperative driving algorithms.
- K-MaT: Knowledge-Anchored Manifold Transport for Cross-Modal Prompt Learning in Medical Imaging (viability: 7): https://sciencetostartup.com/paper/k-mat-knowledge-anchored-manifold-transport-for-cross-modal-prompt-learning-in-medical-imaging - K-MaT is a prompt-learning framework that enables zero-shot cross-modal transfer of medical VLMs, improving performance on low-end modalities without requiring retraining.
- SAHOO: Safeguarded Alignment for High-Order Optimization Objectives in Recursive Self-Improvement (viability: 4): https://sciencetostartup.com/paper/sahoo-safeguarded-alignment-for-high-order-optimization-objectives-in-recursive-self-improvement - SAHOO enhances recursive self-improvement systems with safeguards for alignment drift measurement and control.
- WorldCache: Accelerating World Models for Free via Heterogeneous Token Caching (viability: 7): https://sciencetostartup.com/paper/worldcache-accelerating-world-models-for-free-via-heterogeneous-token-caching - WorldCache accelerates diffusion-based world models by 3.7x via heterogeneous token caching, enabling faster interactive use and long-horizon rollouts.
- Structured Exploration vs. Generative Flexibility: A Field Study Comparing Bandit and LLM Architectures for Personalised Health Behaviour Interventions (viability: 3): https://sciencetostartup.com/paper/structured-exploration-vs-generative-flexibility-a-field-study-comparing-bandit-and-llm-architectures-for-personalised-h - Evaluate the trade-offs between bandit and LLM architectures for personalized health behavior interventions.
- Metalearning traffic assignment for network disruptions with graph convolutional neural networks (viability: 7): https://sciencetostartup.com/paper/metalearning-traffic-assignment-for-network-disruptions-with-graph-convolutional-neural-networks - A meta-learning approach to traffic flow prediction that quickly adapts to network disruptions, enabling real-time rerouting and congestion management.
- Diversity-Aware Adaptive Collocation for Physics-Informed Neural Networks via Sparse QUBO Optimization and Hybrid Coresets (viability: 7): https://sciencetostartup.com/paper/diversity-aware-adaptive-collocation-for-physics-informed-neural-networks-via-sparse-qubo-optimization-and-hybrid-corese - Optimize Physics-Informed Neural Networks with diversity-aware collocation for faster and more accurate PDE solving.
- P-SLCR: Unsupervised Point Cloud Semantic Segmentation via Prototypes Structure Learning and Consistent Reasoning (viability: 7): https://sciencetostartup.com/paper/p-slcr-unsupervised-point-cloud-semantic-segmentation-via-prototypes-structure-learning-and-consistent-reasoning - Unsupervised point cloud semantic segmentation tool that leverages structure learning and consistent reasoning for improved accuracy.
- From Entropy to Calibrated Uncertainty: Training Language Models to Reason About Uncertainty (viability: 3): https://sciencetostartup.com/paper/from-entropy-to-calibrated-uncertainty-training-language-models-to-reason-about-uncertainty - Develop a pipeline to efficiently infer calibrated uncertainty estimates in LLMs for high-stakes domains.
- WMoE-CLIP: Wavelet-Enhanced Mixture-of-Experts Prompt Learning for Zero-Shot Anomaly Detection (viability: 7): https://sciencetostartup.com/paper/wmoe-clip-wavelet-enhanced-mixture-of-experts-prompt-learning-for-zero-shot-anomaly-detection - Enhance zero-shot anomaly detection by dynamically refining textual embeddings with wavelet-enhanced multi-frequency image features and a mixture-of-experts module.
- Polarized Direct Cross-Attention Message Passing in GNNs for Machinery Fault Diagnosis (viability: 7): https://sciencetostartup.com/paper/polarized-direct-cross-attention-message-passing-in-gnns-for-machinery-fault-diagnosis - PolaDCA is a novel relational learning framework that enables adaptive message passing through data-driven graph construction for machinery fault diagnosis.
- DEX-AR: A Dynamic Explainability Method for Autoregressive Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/dex-ar-a-dynamic-explainability-method-for-autoregressive-vision-language-models - DEX-AR provides dynamic explainability for autoregressive vision-language models, generating heatmaps to highlight crucial image regions for textual responses, enabling better understanding and trust in VLM decisions.
- 3D CBCT Artefact Removal Using Perpendicular Score-Based Diffusion Models (viability: 7): https://sciencetostartup.com/paper/3d-cbct-artefact-removal-using-perpendicular-score-based-diffusion-models - A 3D dental implant inpainting tool using perpendicular score-based diffusion models to reduce artifacts in CBCT images, improving diagnostic accuracy.
- The EpisTwin: A Knowledge Graph-Grounded Neuro-Symbolic Architecture for Personal AI (viability: 7): https://sciencetostartup.com/paper/the-epistwin-a-knowledge-graph-grounded-neuro-symbolic-architecture-for-personal-ai - EpisTwin is a neuro-symbolic framework that grounds generative reasoning in a user-centric Personal Knowledge Graph, enabling complex reasoning over personal data.
- Learning Where the Physics Is: Probabilistic Adaptive Sampling for Stiff PDEs (viability: 1): https://sciencetostartup.com/paper/learning-where-the-physics-is-probabilistic-adaptive-sampling-for-stiff-pdes - A probabilistic framework improving the efficiency of Physics-Informed Extreme Learning Machines for modeling stiff PDEs.
- Enhancing SHAP Explainability for Diagnostic and Prognostic ML Models in Alzheimer Disease (viability: 7): https://sciencetostartup.com/paper/enhancing-shap-explainability-for-diagnostic-and-prognostic-ml-models-in-alzheimer-disease - A framework for validating the robustness and transferability of SHAP explanations in Alzheimer's disease ML models, providing clinicians with more reliable interpretations.
- Learning Unbiased Cluster Descriptors for Interpretable Imbalanced Concept Drift Detection (viability: 7): https://sciencetostartup.com/paper/learning-unbiased-cluster-descriptors-for-interpretable-imbalanced-concept-drift-detection - Detect concept drift in imbalanced streaming data by identifying and monitoring individual clusters, enabling targeted interventions and improved system understanding.
- SuperSuit: An Isomorphic Bimodal Interface for Scalable Mobile Manipulation (viability: 7): https://sciencetostartup.com/paper/supersuit-an-isomorphic-bimodal-interface-for-scalable-mobile-manipulation - SuperSuit is a bimodal data acquisition framework for mobile manipulators that enables scalable data collection through robot-in-the-loop teleoperation and active demonstration, improving demonstration throughput and policy performance.
- Can we Trust Unreliable Voxels? Exploring 3D Semantic Occupancy Prediction under Label Noise (viability: 7): https://sciencetostartup.com/paper/can-we-trust-unreliable-voxels-exploring-3d-semantic-occupancy-prediction-under-label-noise - A noise-robust 3D semantic occupancy prediction framework for robotics, providing reliable perception in dynamic environments.
- Artificial Intelligence for Climate Adaptation: Reinforcement Learning for Climate Change-Resilient Transport (viability: 3): https://sciencetostartup.com/paper/artificial-intelligence-for-climate-adaptation-reinforcement-learning-for-climate-change-resilient-transport - A decision-support framework using reinforcement learning for long-term flood adaptation in urban transport systems.
- Spectral and Trajectory Regularization for Diffusion Transformer Super-Resolution (viability: 7): https://sciencetostartup.com/paper/spectral-and-trajectory-regularization-for-diffusion-transformer-super-resolution - StrSR is a one-step adversarial distillation framework featuring spectral and trajectory regularization that achieves state-of-the-art performance in real-world image super-resolution, addressing trajectory mismatch and periodic artifacts in diffusion transformer architectures.
- Stem: Rethinking Causal Information Flow in Sparse Attention (viability: 7): https://sciencetostartup.com/paper/stem-rethinking-causal-information-flow-in-sparse-attention - Stem is a plug-and-play sparsity module that optimizes causal attention in LLMs, reducing computation and pre-filling latency.
- Looking Through Glass Box (viability: 2): https://sciencetostartup.com/paper/looking-through-glass-box - Develops a neural network implementation of fuzzy cognitive maps to analyze causality patterns using langevin differential dynamics.
- HiPP-Prune: Hierarchical Preference-Conditioned Structured Pruning for Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/hipp-prune-hierarchical-preference-conditioned-structured-pruning-for-vision-language-models - HiPP-Prune offers a hierarchical pruning framework for vision-language models, enabling controllable trade-offs between task utility and hallucination robustness, making it ideal for efficient deployment.
- Agentic retrieval-augmented reasoning reshapes collective reliability under model variability in radiology question answering (viability: 3): https://sciencetostartup.com/paper/agentic-retrieval-augmented-reasoning-reshapes-collective-reliability-under-model-variability-in-radiology-question-answ - Enhancing radiology QA reliability via agentic retrieval-augmented reasoning with broad model variability analysis.
- Towards Robotic Lake Maintenance: Integrating SONAR and Satellite Data to Assist Human Operators (viability: 7): https://sciencetostartup.com/paper/towards-robotic-lake-maintenance-integrating-sonar-and-satellite-data-to-assist-human-operators - Automated robotic lake maintenance using satellite and SONAR data for targeted weed harvesting, reducing manual labor.
- Mind the Gap: Pitfalls of LLM Alignment with Asian Public Opinion (viability: 7): https://sciencetostartup.com/paper/mind-the-gap-pitfalls-of-llm-alignment-with-asian-public-opinion - A multilingual audit tool to identify and mitigate cultural misalignment of LLMs in diverse societies, focusing on religious viewpoints and minority representation.
- SPOILER: TEE-Shielded DNN Partitioning of On-Device Secure Inference with Poison Learning (viability: 7): https://sciencetostartup.com/paper/spoiler-tee-shielded-dnn-partitioning-of-on-device-secure-inference-with-poison-learning - SPOILER offers a hardware-aware neural architecture search framework for secure on-device DNN inference, balancing security, latency, and accuracy.
- Learning to Solve Orienteering Problem with Time Windows and Variable Profits (viability: 7): https://sciencetostartup.com/paper/learning-to-solve-orienteering-problem-with-time-windows-and-variable-profits - A learning-based solver for the orienteering problem with time windows and variable profits (OPTWVP) that outperforms existing methods in solution quality and computational efficiency.
- Robust support vector model based on bounded asymmetric elastic net loss for binary classification (viability: 7): https://sciencetostartup.com/paper/robust-support-vector-model-based-on-bounded-asymmetric-elastic-net-loss-for-binary-classification - A robust SVM classifier using a novel loss function, offering improved performance in noisy environments, with available code for integration.
- GazeMoE: Perception of Gaze Target with Mixture-of-Experts (viability: 8): https://sciencetostartup.com/paper/gazemoe-perception-of-gaze-target-with-mixture-of-experts - GazeMoE is an end-to-end framework that selectively leverages gaze-target-related cues from a frozen foundation model through MoE modules, achieving state-of-the-art performance in gaze estimation.
- NOVA: Next-step Open-Vocabulary Autoregression for 3D Multi-Object Tracking in Autonomous Driving (viability: 8): https://sciencetostartup.com/paper/nova-next-step-open-vocabulary-autoregression-for-3d-multi-object-tracking-in-autonomous-driving - NOVA leverages LLMs for 3D multi-object tracking, achieving significant performance gains in novel categories, making it a promising solution for autonomous driving perception.
- Implementation of Quantum Implicit Neural Representation in Deterministic and Probabilistic Autoencoders for Image Reconstruction/Generation Tasks (viability: 7): https://sciencetostartup.com/paper/implementation-of-quantum-implicit-neural-representation-in-deterministic-and-probabilistic-autoencoders-for-image-recon - A quantum implicit neural representation autoencoder enhances image reconstruction and generation, offering a potential advantage in data-constrained environments.
- Synthetic Monitoring Environments for Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/synthetic-monitoring-environments-for-reinforcement-learning - Synthetic Monitoring Environments (SMEs) provide a configurable testbed for RL agent diagnostics, enabling precise evaluation and targeted improvements.
- SPPCSO: Adaptive Penalized Estimation Method for High-Dimensional Correlated Data (viability: 6): https://sciencetostartup.com/paper/sppcso-adaptive-penalized-estimation-method-for-high-dimensional-correlated-data - SPPCSO is a penalized estimation method that integrates single-parametric principal component regression and L1 regularization to adaptively adjust the shrinkage factor, providing stable and reliable estimation in high-noise settings.
- Gradient Flow Polarizes Softmax Outputs towards Low-Entropy Solutions (viability: 5): https://sciencetostartup.com/paper/gradient-flow-polarizes-softmax-outputs-towards-low-entropy-solutions - This paper analyzes the gradient flow dynamics of the value-softmax model, providing insights into transformer training dynamics and potentially leading to improved attention mechanisms.
- DC-Merge: Improving Model Merging with Directional Consistency (viability: 7): https://sciencetostartup.com/paper/dc-merge-improving-model-merging-with-directional-consistency - DC-Merge improves model merging by maintaining directional consistency, enabling better knowledge retention across multiple tasks, and is available as open-source code.
- TaPD: Temporal-adaptive Progressive Distillation for Observation-Adaptive Trajectory Forecasting in Autonomous Driving (viability: 7): https://sciencetostartup.com/paper/tapd-temporal-adaptive-progressive-distillation-for-observation-adaptive-trajectory-forecasting-in-autonomous-driving - Improve trajectory prediction in autonomous vehicles with variable observation lengths using a plug-and-play framework that leverages knowledge distillation and temporal backfilling.
- Low-latency Event-based Object Detection with Spatially-Sparse Linear Attention (viability: 7): https://sciencetostartup.com/paper/low-latency-event-based-object-detection-with-spatially-sparse-linear-attention - Accelerate event-based object detection with spatially-sparse linear attention, achieving state-of-the-art accuracy and reduced per-event computation.
- FedSCS-XGB -- Federated Server-centric surrogate XGBoost for continual health monitoring (viability: 7): https://sciencetostartup.com/paper/fedscs-xgb-federated-server-centric-surrogate-xgboost-for-continual-health-monitoring - Federated XGBoost for wearable sensor data enables privacy-preserving continuous health monitoring, achieving near-centralized performance.
- SPOT: Span-level Pause-of-Thought for Efficient and Interpretable Latent Reasoning in Large Language Models (viability: 7): https://sciencetostartup.com/paper/spot-span-level-pause-of-thought-for-efficient-and-interpretable-latent-reasoning-in-large-language-models - SPOT compresses Chain-of-Thought reasoning into compact, interpretable latent pause tokens, improving accuracy and reducing token generation for LLMs.
- Word-Anchored Temporal Forgery Localization (viability: 7): https://sciencetostartup.com/paper/word-anchored-temporal-forgery-localization - WAFL offers a computationally efficient and accurate solution for temporal forgery localization by shifting the task to word-level binary classification, making it suitable for integration into existing video analysis pipelines.
- Few-Shot Neural Differentiable Simulator: Real-to-Sim Rigid-Contact Modeling (viability: 7): https://sciencetostartup.com/paper/few-shot-neural-differentiable-simulator-real-to-sim-rigid-contact-modeling - A few-shot real-to-sim approach that combines analytical simulators with graph neural networks for accurate and differentiable physics simulation, enabling efficient policy learning in robotic manipulation.
- Conversational Demand Response: Bidirectional Aggregator-Prosumer Coordination through Agentic AI (viability: 8): https://sciencetostartup.com/paper/conversational-demand-response-bidirectional-aggregator-prosumer-coordination-through-agentic-ai - Conversational Demand Response uses agentic AI to enable bidirectional communication between energy aggregators and prosumers, improving demand response participation.
- EntON: Eigenentropy-Optimized Neighborhood Densification in 3D Gaussian Splatting (viability: 7): https://sciencetostartup.com/paper/enton-eigenentropy-optimized-neighborhood-densification-in-3d-gaussian-splatting - EntON improves 3D Gaussian Splatting by using Eigenentropy to guide densification, leading to better geometric accuracy and rendering quality.
- Cut to the Chase: Training-free Multimodal Summarization via Chain-of-Events (viability: 8): https://sciencetostartup.com/paper/cut-to-the-chase-training-free-multimodal-summarization-via-chain-of-events - CoE is a training-free multimodal summarization framework that leverages a chain-of-events guided by a hierarchical event graph to generate concise textual summaries from videos, transcripts, and images, outperforming state-of-the-art baselines.
- Topological descriptors of foot clearance gait dynamics improve differential diagnosis of Parkinsonism (viability: 7): https://sciencetostartup.com/paper/topological-descriptors-of-foot-clearance-gait-dynamics-improve-differential-diagnosis-of-parkinsonism - Improve Parkinson's diagnosis with a gait analysis tool using topological data analysis and machine learning to differentiate between Parkinsonian syndromes.
- EarthBridge: A Solution for 4th Multi-modal Aerial View Image Challenge Translation Track (viability: 7): https://sciencetostartup.com/paper/earthbridge-a-solution-for-4th-multi-modal-aerial-view-image-challenge-translation-track - EarthBridge is a high-fidelity image translation framework for multi-modal aerial view analysis, offering superior spatial detail and spectral accuracy, with available code for immediate implementation.
- VG3S: Visual Geometry Grounded Gaussian Splatting for Semantic Occupancy Prediction (viability: 7): https://sciencetostartup.com/paper/vg3s-visual-geometry-grounded-gaussian-splatting-for-semantic-occupancy-prediction - Integrate geometric priors from Vision Foundation Models into Gaussian Splatting for improved 3D semantic occupancy prediction, enhancing autonomous driving scene understanding.
- KISS-IMU: Self-supervised Inertial Odometry with Motion-balanced Learning and Uncertainty-aware Inference (viability: 7): https://sciencetostartup.com/paper/kiss-imu-self-supervised-inertial-odometry-with-motion-balanced-learning-and-uncertainty-aware-inference - KISS-IMU provides a self-supervised inertial odometry framework that leverages LiDAR-based ICP registration for training, enabling robust pose estimation for robotic systems.
- Point-Supervised Skeleton-Based Human Action Segmentation (viability: 7): https://sciencetostartup.com/paper/point-supervised-skeleton-based-human-action-segmentation - A point-supervised skeleton-based action segmentation framework that reduces annotation effort while achieving competitive performance, suitable for integration into existing action recognition systems.
- Adaptive Language-Aware Image Reflection Removal Network (viability: 8): https://sciencetostartup.com/paper/adaptive-language-aware-image-reflection-removal-network - ALANet removes complex image reflections using language guidance, even with inaccurate language descriptions, and provides a new dataset for evaluation.
- FlashPrefill: Instantaneous Pattern Discovery and Thresholding for Ultra-Fast Long-Context Prefilling (viability: 8): https://sciencetostartup.com/paper/flashprefill-instantaneous-pattern-discovery-and-thresholding-for-ultra-fast-long-context-prefilling - FlashPrefill accelerates long-context LLM prefilling by 27x with a novel pattern discovery and thresholding technique, offering a drop-in replacement for existing attention mechanisms.
- Latent Autoencoder Ensemble Kalman Filter for Data assimilation (viability: 7): https://sciencetostartup.com/paper/latent-autoencoder-ensemble-kalman-filter-for-data-assimilation - Improve data assimilation accuracy and stability in nonlinear systems by using a latent autoencoder ensemble Kalman filter.
- LIT-RAGBench: Benchmarking Generator Capabilities of Large Language Models in Retrieval-Augmented Generation (viability: 7): https://sciencetostartup.com/paper/lit-ragbench-benchmarking-generator-capabilities-of-large-language-models-in-retrieval-augmented-generation - LIT-RAGBench provides a comprehensive benchmark for evaluating the generator capabilities of LLMs in RAG systems, enabling targeted model selection and development.
- Wisdom of the AI Crowd (AI-CROWD) for Ground Truth Approximation in Content Analysis: A Research Protocol & Validation Using Eleven Large Language Models (viability: 7): https://sciencetostartup.com/paper/wisdom-of-the-ai-crowd-ai-crowd-for-ground-truth-approximation-in-content-analysis-a-research-protocol-validation-using- - AI-CROWD protocol leverages LLM ensembles to approximate ground truth in content analysis, offering a scalable solution for labeling large datasets.
- MAPO: Mixed Advantage Policy Optimization for Long-Horizon Multi-Turn Dialogue (viability: 7): https://sciencetostartup.com/paper/mapo-mixed-advantage-policy-optimization-for-long-horizon-multi-turn-dialogue - MAPO is a reinforcement learning algorithm that improves training stability and performance for subjective multi-turn dialogue tasks by leveraging dense process feedback and mixed advantage estimation.
- Whisper-CD: Accurate Long-Form Speech Recognition using Multi-Negative Contrastive Decoding (viability: 8): https://sciencetostartup.com/paper/whisper-cd-accurate-long-form-speech-recognition-using-multi-negative-contrastive-decoding - Whisper-CD enhances long-form speech recognition accuracy and speed by using a training-free contrastive decoding method, making it a drop-in replacement for existing Whisper systems.
- DreamToNav: Generalizable Navigation for Robots via Generative Video Planning (viability: 7): https://sciencetostartup.com/paper/dreamtonav-generalizable-navigation-for-robots-via-generative-video-planning - DreamToNav uses generative video models to enable robots to navigate based on natural language prompts, offering a more intuitive and flexible control system.
- SpaCRD: Multimodal Deep Fusion of Histology and Spatial Transcriptomics for Cancer Region Detection (viability: 8): https://sciencetostartup.com/paper/spacrd-multimodal-deep-fusion-of-histology-and-spatial-transcriptomics-for-cancer-region-detection - SpaCRD is a transfer learning method that deeply integrates histology images and spatial transcriptomics data for reliable cancer tissue region detection, outperforming existing state-of-the-art methods.
- CRIMSON: A Clinically-Grounded LLM-Based Metric for Generative Radiology Report Evaluation (viability: 8): https://sciencetostartup.com/paper/crimson-a-clinically-grounded-llm-based-metric-for-generative-radiology-report-evaluation - CRIMSON is a clinically-grounded evaluation framework for chest X-ray report generation that prioritizes clinically consequential mistakes, offering a severity-aware metric for improved patient safety.
- Towards Motion Turing Test: Evaluating Human-Likeness in Humanoid Robots (viability: 8): https://sciencetostartup.com/paper/towards-motion-turing-test-evaluating-human-likeness-in-humanoid-robots - Evaluate and improve humanoid robot motion with a human-likeness scoring system and benchmark dataset, enabling more natural robot interactions.
- Contrastive-to-Self-Supervised: A Two-Stage Framework for Script Similarity Learning (viability: 7): https://sciencetostartup.com/paper/contrastive-to-self-supervised-a-two-stage-framework-for-script-similarity-learning - A two-stage framework for glyph similarity learning that combines contrastive learning and teacher-student distillation for improved few-shot recognition and script clustering.
- Making Training-Free Diffusion Segmentors Scale with the Generative Power (viability: 7): https://sciencetostartup.com/paper/making-training-free-diffusion-segmentors-scale-with-the-generative-power - Leverage pre-trained diffusion models for improved semantic segmentation without training, using auto-aggregation and per-pixel rescaling techniques.
- Optimizing 3D Diffusion Models for Medical Imaging via Multi-Scale Reward Learning (viability: 7): https://sciencetostartup.com/paper/optimizing-3d-diffusion-models-for-medical-imaging-via-multi-scale-reward-learning - Fine-tune 3D diffusion models for medical image generation using reinforcement learning to improve clinical relevance and downstream task performance.
- Alkaid: Resilience to Edit Errors in Provably Secure Steganography via Distance-Constrained Encoding (viability: 8): https://sciencetostartup.com/paper/alkaid-resilience-to-edit-errors-in-provably-secure-steganography-via-distance-constrained-encoding - Alkaid provides provably secure and robust steganography resilient to edit errors, enabling reliable message recovery in noisy environments.
- XMACNet: An Explainable Lightweight Attention based CNN with Multi Modal Fusion for Chili Disease Classification (viability: 8): https://sciencetostartup.com/paper/xmacnet-an-explainable-lightweight-attention-based-cnn-with-multi-modal-fusion-for-chili-disease-classification - XMACNet is a lightweight, explainable CNN for chili disease classification that can be deployed on edge devices, offering farmers a real-time diagnostic tool.
- Robotic Foundation Models for Industrial Control: A Comprehensive Survey and Readiness Assessment Framework (viability: 6): https://sciencetostartup.com/paper/robotic-foundation-models-for-industrial-control-a-comprehensive-survey-and-readiness-assessment-framework - A comprehensive survey and assessment framework for robotic foundation models in industrial control, highlighting the gap between current models and industry-grade requirements.
- JOPP-3D: Joint Open Vocabulary Semantic Segmentation on Point Clouds and Panoramas (viability: 8): https://sciencetostartup.com/paper/jopp-3d-joint-open-vocabulary-semantic-segmentation-on-point-clouds-and-panoramas - JOPP-3D enables language-driven semantic segmentation on point clouds and panoramas, offering a unified scene understanding solution with significant SOTA improvements.
- A Semi-Supervised Framework for Breast Ultrasound Segmentation with Training-Free Pseudo-Label Generation and Label Refinement (viability: 8): https://sciencetostartup.com/paper/a-semi-supervised-framework-for-breast-ultrasound-segmentation-with-training-free-pseudo-label-generation-and-label-refi - A semi-supervised breast ultrasound segmentation framework leveraging vision-language models for training-free pseudo-label generation, achieving near fully-supervised performance with minimal labeled data, enabling scalable medical image segmentation.
- Reflective Flow Sampling Enhancement (viability: 7): https://sciencetostartup.com/paper/reflective-flow-sampling-enhancement - Reflective Flow Sampling enhances text-to-image generation in flow models, offering improved quality and prompt alignment without retraining.
- Do Compact SSL Backbones Matter for Audio Deepfake Detection? A Controlled Study with RAPTOR (viability: 7): https://sciencetostartup.com/paper/do-compact-ssl-backbones-matter-for-audio-deepfake-detection-a-controlled-study-with-raptor - RAPTOR is a compact and robust audio deepfake detection system leveraging multilingual HuBERT pre-training for cross-domain performance, outperforming larger commercial systems.
- Efficient Vector Search in the Wild: One Model for Multi-K Queries (viability: 7): https://sciencetostartup.com/paper/efficient-vector-search-in-the-wild-one-model-for-multi-k-queries - OMEGA is a K-generalizable learned top-K search method that simultaneously achieves high accuracy, high performance, and low preprocessing cost for multi-K vector queries.
- Ensemble Graph Neural Networks for Probabilistic Sea Surface Temperature Forecasting via Input Perturbations (viability: 7): https://sciencetostartup.com/paper/ensemble-graph-neural-networks-for-probabilistic-sea-surface-temperature-forecasting-via-input-perturbations - Improve sea surface temperature forecasting by using graph neural networks and input perturbations to generate probabilistic forecasts, enabling better risk management for maritime industries.
- PQC-LEO: An Evaluation Framework for Post-Quantum Cryptographic Algorithms (viability: 7): https://sciencetostartup.com/paper/pqc-leo-an-evaluation-framework-for-post-quantum-cryptographic-algorithms - PQC-LEO is a benchmarking suite that automates the evaluation of post-quantum cryptographic algorithms, enabling developers to optimize performance across different architectures.
- VLM-RobustBench: A Comprehensive Benchmark for Robustness of Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/vlm-robustbench-a-comprehensive-benchmark-for-robustness-of-vision-language-models - VLM-RobustBench provides a comprehensive benchmark to evaluate and improve the robustness of Vision-Language Models against real-world image distortions, enabling developers to build more reliable VLM-powered applications.
- Longitudinal NSCLC Treatment Progression via Multimodal Generative Models (viability: 7): https://sciencetostartup.com/paper/longitudinal-nsclc-treatment-progression-via-multimodal-generative-models - Predict tumor evolution during radiotherapy by synthesizing plausible follow-up CT images conditioned on radiation dose and clinical variables, enabling in-silico treatment monitoring and adaptive radiotherapy research.
- Predictive Coding Graphs are a Superset of Feedforward Neural Networks (viability: 1): https://sciencetostartup.com/paper/predictive-coding-graphs-are-a-superset-of-feedforward-neural-networks - Exploring Predictive Coding Graphs as a mathematical superset of traditional neural networks for enhanced machine learning modeling.
- Spatial Colour Mixing Illusions as a Perception Stress Test for Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/spatial-colour-mixing-illusions-as-a-perception-stress-test-for-vision-language-models - A perception stress test for vision-language models reveals weaknesses that can be addressed with preprocessing, suggesting a potential API to improve VLM robustness.
- Place-it-R1: Unlocking Environment-aware Reasoning Potential of MLLM for Video Object Insertion (viability: 7): https://sciencetostartup.com/paper/place-it-r1-unlocking-environment-aware-reasoning-potential-of-mllm-for-video-object-insertion - Place-it-R1 is an environment-aware video object insertion framework leveraging MLLMs and video diffusion for physically plausible edits.
- Partial Policy Gradients for RL in LLMs (viability: 2): https://sciencetostartup.com/paper/partial-policy-gradients-for-rl-in-llms - A framework using partial policy gradients to model policy structures for reinforcement learning in large language models.
- A Causal Graph Approach to Oppositional Narrative Analysis (viability: 2): https://sciencetostartup.com/paper/a-causal-graph-approach-to-oppositional-narrative-analysis - A graph-based framework for analyzing oppositional narratives in text using causal inference.
- DQE: A Semantic-Aware Evaluation Metric for Time Series Anomaly Detection (viability: 7): https://sciencetostartup.com/paper/dqe-a-semantic-aware-evaluation-metric-for-time-series-anomaly-detection - DQE offers a more reliable and interpretable evaluation metric for time series anomaly detection, addressing limitations of existing methods and enabling more objective progress.
- A Hazard-Informed Data Pipeline for Robotics Physical Safety (viability: 3): https://sciencetostartup.com/paper/a-hazard-informed-data-pipeline-for-robotics-physical-safety - Develop a framework for integrating classical risk engineering with machine learning to enhance robotics physical safety.
- Diffusion Language Models Are Natively Length-Aware (viability: 7): https://sciencetostartup.com/paper/diffusion-language-models-are-natively-length-aware - Dynamically crop the context window in Diffusion Language Models to reduce computational waste and improve efficiency for reasoning and chat tasks.
- Sticky-Glance: Robust Intent Recognition for Human Robot Collaboration via Single-Glance (viability: 7): https://sciencetostartup.com/paper/sticky-glance-robust-intent-recognition-for-human-robot-collaboration-via-single-glance - Sticky-Glance enables robust gaze-based intent recognition for human-robot collaboration, improving efficiency and reducing workload.
- Dynamic Momentum Recalibration in Online Gradient Learning (viability: 7): https://sciencetostartup.com/paper/dynamic-momentum-recalibration-in-online-gradient-learning - SGDF dynamically recalibrates gradient momentum in SGD, offering improved optimization performance and a potential drop-in replacement for existing optimizers.
- Making Implicit Premises Explicit in Logical Understanding of Enthymemes (viability: 7): https://sciencetostartup.com/paper/making-implicit-premises-explicit-in-logical-understanding-of-enthymemes - A pipeline that uses LLMs to translate enthymemes into logical formulas and a neuro-symbolic reasoner to determine entailment, enabling logical understanding of implicit arguments.
- Latent Diffusion-Based 3D Molecular Recovery from Vibrational Spectra (viability: 7): https://sciencetostartup.com/paper/latent-diffusion-based-3d-molecular-recovery-from-vibrational-spectra - IR-GeoDiff is a latent diffusion model that recovers 3D molecular geometries from IR spectra, enabling a novel approach to molecular structure determination.
- Enhancing Instruction Following of LLMs via Activation Steering with Dynamic Rejection (viability: 7): https://sciencetostartup.com/paper/enhancing-instruction-following-of-llms-via-activation-steering-with-dynamic-rejection - DIRECTER dynamically adjusts activation steering strength in LLMs to improve instruction following without sacrificing text quality, offering a more robust and adaptable steering mechanism.
- DeepSight: Bridging Depth Maps and Language with a Depth-Driven Multimodal Model (viability: 7): https://sciencetostartup.com/paper/deepsight-bridging-depth-maps-and-language-with-a-depth-driven-multimodal-model - DeepSight is a depth-aware multimodal model that enhances 3D scene understanding by leveraging depth maps and a novel depth image-text dataset.
- Experiences Build Characters: The Linguistic Origins and Functional Impact of LLM Personality (viability: 7): https://sciencetostartup.com/paper/experiences-build-characters-the-linguistic-origins-and-functional-impact-of-llm-personality - Fine-tune LLMs with domain-specific text to engineer personality traits and optimize for specific reasoning tasks.
- Multimodal Behavior Tree Generation: A Small Vision-Language Model for Robot Task Planning (viability: 7): https://sciencetostartup.com/paper/multimodal-behavior-tree-generation-a-small-vision-language-model-for-robot-task-planning - A compact, open-source vision-language model generates behavior trees for robotic task planning, enabling efficient execution of household tasks.
- Offline Materials Optimization with CliqueFlowmer (viability: 7): https://sciencetostartup.com/paper/offline-materials-optimization-with-cliqueflowmer - CliqueFlowmer optimizes material discovery by fusing direct property optimization into transformer-based generation, outperforming generative baselines.
- Lyapunov Probes for Hallucination Detection in Large Foundation Models (viability: 7): https://sciencetostartup.com/paper/lyapunov-probes-for-hallucination-detection-in-large-foundation-models - Detect LLM hallucinations by probing the stability of knowledge representation using lightweight networks, enabling more reliable and trustworthy AI outputs.
- Lifelong Embodied Navigation Learning (viability: 7): https://sciencetostartup.com/paper/lifelong-embodied-navigation-learning - Uni-Walker is a lifelong embodied navigation framework that uses Decoder Extension LoRA to enable agents to continually acquire new navigation skills without catastrophic forgetting.
- Text-Driven Emotionally Continuous Talking Face Generation (viability: 7): https://sciencetostartup.com/paper/text-driven-emotionally-continuous-talking-face-generation - Generate emotionally expressive talking-face videos from text and emotion descriptions, enabling realistic and dynamic digital avatars.
- Aggregative Semantics for Quantitative Bipolar Argumentation Frameworks (viability: 2): https://sciencetostartup.com/paper/aggregative-semantics-for-quantitative-bipolar-argumentation-frameworks - Develop new aggregative semantics for Quantitative Bipolar Argumentation Frameworks to enhance AI argumentation modeling.
- Evaluating Austrian A-Level German Essays with Large Language Models for Automated Essay Scoring (viability: 5): https://sciencetostartup.com/paper/evaluating-austrian-a-level-german-essays-with-large-language-models-for-automated-essay-scoring - Automated essay scoring tool for Austrian A-level German texts using open-weight LLMs, showing moderate agreement with human raters.
- Agentic LLM Planning via Step-Wise PDDL Simulation: An Empirical Characterisation (viability: 5): https://sciencetostartup.com/paper/agentic-llm-planning-via-step-wise-pddl-simulation-an-empirical-characterisation - PyPDDLEngine makes LLMs interactive planners by integrating PDDL for step-wise action selection and evaluation in task planning.
- TempoSyncDiff: Distilled Temporally-Consistent Diffusion for Low-Latency Audio-Driven Talking Head Generation (viability: 7): https://sciencetostartup.com/paper/temposyncdiff-distilled-temporally-consistent-diffusion-for-low-latency-audio-driven-talking-head-generation - TempoSyncDiff is a low-latency audio-driven talking head generation framework using distilled diffusion, enabling real-time applications on edge devices.
- Probing Visual Concepts in Lightweight Vision-Language Models for Automated Driving (viability: 3): https://sciencetostartup.com/paper/probing-visual-concepts-in-lightweight-vision-language-models-for-automated-driving - This research investigates the encoding of visual concepts in Vision-Language Models to understand failures in automated driving applications.
- A LINDDUN-based Privacy Threat Modeling Framework for GenAI (viability: 7): https://sciencetostartup.com/paper/a-linddun-based-privacy-threat-modeling-framework-for-genai - A privacy threat modeling framework for GenAI applications, enabling comprehensive privacy analysis and risk mitigation.
- Devil is in Narrow Policy: Unleashing Exploration in Driving VLA Models (viability: 8): https://sciencetostartup.com/paper/devil-is-in-narrow-policy-unleashing-exploration-in-driving-vla-models - Curious-VLA unlocks the exploratory potential of autonomous driving models by addressing the exploit-explore dilemma, achieving state-of-the-art results on the Navsim benchmark.
- GenHOI: Towards Object-Consistent Hand-Object Interaction with Temporally Balanced and Spatially Selective Object Injection (viability: 8): https://sciencetostartup.com/paper/genhoi-towards-object-consistent-hand-object-interaction-with-temporally-balanced-and-spatially-selective-object-injecti - GenHOI enhances video generation models with object-consistent hand-object interaction by injecting reference object information, outperforming existing methods in in-the-wild scenarios.
- Learning to Generate via Understanding: Understanding-Driven Intrinsic Rewarding for Unified Multimodal Models (viability: 7): https://sciencetostartup.com/paper/learning-to-generate-via-understanding-understanding-driven-intrinsic-rewarding-for-unified-multimodal-models - Improve multimodal model generation quality by using the model's understanding branch to guide the generation process through a self-supervised reinforcement learning framework.
- FontUse: A Data-Centric Approach to Style- and Use-Case-Conditioned In-Image Typography (viability: 7): https://sciencetostartup.com/paper/fontuse-a-data-centric-approach-to-style-and-use-case-conditioned-in-image-typography - Generate in-image typography with targeted style and use-case control using a fine-tuned text-to-image model and a typography-focused dataset.
- Ensemble Learning with Sparse Hypercolumns (viability: 7): https://sciencetostartup.com/paper/ensemble-learning-with-sparse-hypercolumns - Improve brain tumor segmentation in low-data regimes by ensembling sparse hypercolumns derived from VGG16, outperforming standard UNet.
- StruVis: Enhancing Reasoning-based Text-to-Image Generation via Thinking with Structured Vision (viability: 7): https://sciencetostartup.com/paper/struvis-enhancing-reasoning-based-text-to-image-generation-via-thinking-with-structured-vision - StruVis enhances text-to-image generation by using text-based structured visual representations for reasoning, improving performance on reasoning-based benchmarks.
- ResearchEnvBench: Benchmarking Agents on Environment Synthesis for Research Code Execution (viability: 7): https://sciencetostartup.com/paper/researchenvbench-benchmarking-agents-on-environment-synthesis-for-research-code-execution - Automate research environment setup with an agent that resolves dependencies and configures execution for scientific code, enabling reproducible research.
- Sensitivity-Aware Retrieval-Augmented Intent Clarification (viability: 2): https://sciencetostartup.com/paper/sensitivity-aware-retrieval-augmented-intent-clarification - Develop a sensitivity-aware conversational agent for secure retrieval-augmented intent clarification in sensitive domains.
- MOSIV: Multi-Object System Identification from Videos (viability: 8): https://sciencetostartup.com/paper/mosiv-multi-object-system-identification-from-videos - MOSIV is a framework that identifies and simulates multi-object interactions from videos, enabling more accurate robotic manipulation and simulation.
- EffectMaker: Unifying Reasoning and Generation for Customized Visual Effect Creation (viability: 8): https://sciencetostartup.com/paper/effectmaker-unifying-reasoning-and-generation-for-customized-visual-effect-creation - EffectMaker is a unified reasoning-generation framework that enables reference-based VFX customization, offering a scalable and flexible paradigm for customized VFX generation.
- Preventing Learning Stagnation in PPO by Scaling to 1 Million Parallel Environments (viability: 7): https://sciencetostartup.com/paper/preventing-learning-stagnation-in-ppo-by-scaling-to-1-million-parallel-environments - Scale PPO to 1M parallel environments to overcome performance plateaus in complex RL tasks, offering a robust solution for monotonic performance improvement.
- MASFactory: A Graph-centric Framework for Orchestrating LLM-Based Multi-Agent Systems with Vibe Graphing (viability: 8): https://sciencetostartup.com/paper/masfactory-a-graph-centric-framework-for-orchestrating-llm-based-multi-agent-systems-with-vibe-graphing - MASFactory is a graph-centric framework that simplifies the creation and orchestration of LLM-based multi-agent systems with a human-in-the-loop approach, offering reusable components and visualizer for runtime tracing.
- Demystifying KAN for Vision Tasks: The RepKAN Approach (viability: 7): https://sciencetostartup.com/paper/demystifying-kan-for-vision-tasks-the-repkan-approach - RepKAN offers interpretable remote sensing image classification by combining CNN efficiency with KAN's non-linear representation, outperforming state-of-the-art models.
- Restoring Linguistic Grounding in VLA Models via Train-Free Attention Recalibration (viability: 7): https://sciencetostartup.com/paper/restoring-linguistic-grounding-in-vla-models-via-train-free-attention-recalibration - Improve the reliability of vision-language-action models in robotics by recalibrating attention to prioritize language instructions, preventing erroneous actions in out-of-distribution scenarios.
- MM-ISTS: Cooperating Irregularly Sampled Time Series Forecasting with Multimodal Vision-Text LLMs (viability: 2): https://sciencetostartup.com/paper/mm-ists-cooperating-irregularly-sampled-time-series-forecasting-with-multimodal-vision-text-llms - Leverage multimodal LLMs to improve forecasting in irregularly sampled time series through vision-text integration.
- Track-SQL: Enhancing Generative Language Models with Dual-Extractive Modules for Schema and Context Tracking in Multi-turn Text-to-SQL (viability: 8): https://sciencetostartup.com/paper/track-sql-enhancing-generative-language-models-with-dual-extractive-modules-for-schema-and-context-tracking-in-multi-tur - Track-SQL enhances generative language models with dual-extractive modules for improved schema and context tracking in multi-turn Text-to-SQL, achieving state-of-the-art performance.
- TADPO: Reinforcement Learning Goes Off-road (viability: 7): https://sciencetostartup.com/paper/tadpo-reinforcement-learning-goes-off-road - TADPO is a reinforcement learning approach for off-road autonomous driving, demonstrating zero-shot sim-to-real transfer on a full-scale vehicle.
- Moving Through Clutter: Scaling Data Collection and Benchmarking for 3D Scene-Aware Humanoid Locomotion via Virtual Reality (viability: 7): https://sciencetostartup.com/paper/moving-through-clutter-scaling-data-collection-and-benchmarking-for-3d-scene-aware-humanoid-locomotion-via-virtual-reali - An open-source VR framework for collecting and evaluating scene-aware humanoid locomotion data in cluttered 3D environments, enabling the development of more robust robotic navigation systems.
- MagRobot:An Open Simulator for Magnetically Navigated Robots (viability: 7): https://sciencetostartup.com/paper/magrobot-an-open-simulator-for-magnetically-navigated-robots - MagRobot is an open-source simulation platform for magnetically navigated robots, enabling efficient design, visualization, and analysis of magnetic navigation systems for medical applications.
- Rank-Factorized Implicit Neural Bias: Scaling Super-Resolution Transformer with FlashAttention (viability: 7): https://sciencetostartup.com/paper/rank-factorized-implicit-neural-bias-scaling-super-resolution-transformer-with-flashattention - Accelerate super-resolution transformer training and inference by replacing relative positional bias with a rank-factorized implicit neural bias, enabling FlashAttention.
- SemFuzz: A Semantics-Aware Fuzzing Framework for Network Protocol Implementations (viability: 7): https://sciencetostartup.com/paper/semfuzz-a-semantics-aware-fuzzing-framework-for-network-protocol-implementations - SemFuzz uses LLMs to extract semantic rules from RFCs and generate targeted test cases to uncover deep semantic vulnerabilities in network protocol implementations.
- Technical Report: Automated Optical Inspection of Surgical Instruments (viability: 7): https://sciencetostartup.com/paper/technical-report-automated-optical-inspection-of-surgical-instruments - Automated optical inspection tool for surgical instruments using YOLOv8, ResNet-152, and EfficientNet-b4 to detect manufacturing defects.
- HarvestFlex: Strawberry Harvesting via Vision-Language-Action Policy Adaptation in the Wild (viability: 7): https://sciencetostartup.com/paper/harvestflex-strawberry-harvesting-via-vision-language-action-policy-adaptation-in-the-wild - Fine-tuned vision-language-action policies enable a strawberry-harvesting robot to achieve a 74% success rate with limited real-world data.
- An Interactive Multi-Agent System for Evaluation of New Product Concepts (viability: 7): https://sciencetostartup.com/paper/an-interactive-multi-agent-system-for-evaluation-of-new-product-concepts - Automated product concept evaluation using an LLM-based multi-agent system, providing objective evidence and structured deliberations for improved decision-making.
- Proprioceptive Shape Estimation of Tensegrity Manipulators Using Energy Minimisation (viability: 7): https://sciencetostartup.com/paper/proprioceptive-shape-estimation-of-tensegrity-manipulators-using-energy-minimisation - Estimate the shape of tensegrity manipulators using only IMU data, enabling robust control in diverse environments.
- Breaking Smooth-Motion Assumptions: A UAV Benchmark for Multi-Object Tracking in Complex and Adverse Conditions (viability: 7): https://sciencetostartup.com/paper/breaking-smooth-motion-assumptions-a-uav-benchmark-for-multi-object-tracking-in-complex-and-adverse-conditions - A new challenging UAV multi-object tracking benchmark dataset, DynUAV, is released to push the boundaries of tracking algorithms in complex real-world scenarios.
- Imagine How To Change: Explicit Procedure Modeling for Change Captioning (viability: 7): https://sciencetostartup.com/paper/imagine-how-to-change-explicit-procedure-modeling-for-change-captioning - ProCap generates change captions by modeling the dynamic procedure between two images using a two-stage encoder-decoder framework, offering a more nuanced description of visual differences.
- PROBE: Probabilistic Occupancy BEV Encoding with Analytical Translation Robustness for 3D Place Recognition (viability: 7): https://sciencetostartup.com/paper/probe-probabilistic-occupancy-bev-encoding-with-analytical-translation-robustness-for-3d-place-recognition - PROBE is a learning-free LiDAR place recognition descriptor that analytically marginalizes over continuous Cartesian translations, offering a robust and efficient solution for cross-sensor generalization.
- Skeleton-to-Image Encoding: Enabling Skeleton Representation Learning via Vision-Pretrained Models (viability: 7): https://sciencetostartup.com/paper/skeleton-to-image-encoding-enabling-skeleton-representation-learning-via-vision-pretrained-models - Transform skeleton data into images to leverage vision-pretrained models for improved action recognition and representation learning.
- Exploring Open-Vocabulary Object Recognition in Images using CLIP (viability: 8): https://sciencetostartup.com/paper/exploring-open-vocabulary-object-recognition-in-images-using-clip - A streamlined open-vocabulary object recognition framework leveraging CLIP and CNN/MLP-based encoding for enhanced generalization and reduced training costs, outperforming state-of-the-art methods.
- Domain-Adaptive Model Merging across Disconnected Modes (viability: 7): https://sciencetostartup.com/paper/domain-adaptive-model-merging-across-disconnected-modes - DMM is a data-free model merging framework that consolidates knowledge from multiple specialized models, achieving state-of-the-art performance and avoiding data sharing.
- Who We Are, Where We Are: Mental Health at the Intersection of Person, Situation, and Large Language Models (viability: 3): https://sciencetostartup.com/paper/who-we-are-where-we-are-mental-health-at-the-intersection-of-person-situation-and-large-language-models - Integrates psychological theory with computational models to predict mental health from social media data.
- Unify the Views: View-Consistent Prototype Learning for Few-Shot Segmentation (viability: 7): https://sciencetostartup.com/paper/unify-the-views-view-consistent-prototype-learning-for-few-shot-segmentation - VINE is a few-shot segmentation framework that refines class-specific prototypes by modeling structural consistency and foreground discrimination, offering improved accuracy in scenarios with viewpoint shifts and complex structures.
- Energy-Driven Adaptive Visual Token Pruning for Efficient Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/energy-driven-adaptive-visual-token-pruning-for-efficient-vision-language-models - E-AdaPrune adaptively prunes visual tokens in VLMs based on image information density, improving performance and efficiency without adding learnable parameters.
- How to Model Your Crazyflie Brushless (viability: 8): https://sciencetostartup.com/paper/how-to-model-your-crazyflie-brushless - An open-source dynamics model for the Crazyflie Brushless, enabling rapid development and testing of agile nano-quadcopter controllers.
- Adaptive Radial Projection on Fourier Magnitude Spectrum for Document Image Skew Estimation (viability: 8): https://sciencetostartup.com/paper/adaptive-radial-projection-on-fourier-magnitude-spectrum-for-document-image-skew-estimation - A novel skew estimation method for document images with superior performance and available source code, ready for integration into document processing systems.
- XAI for Coding Agent Failures: Transforming Raw Execution Traces into Actionable Insights (viability: 8): https://sciencetostartup.com/paper/xai-for-coding-agent-failures-transforming-raw-execution-traces-into-actionable-insights - We provide an XAI tool that transforms raw coding agent execution traces into actionable insights, enabling developers to debug failures faster and more accurately.
- SLER-IR: Spherical Layer-wise Expert Routing for All-in-One Image Restoration (viability: 7): https://sciencetostartup.com/paper/sler-ir-spherical-layer-wise-expert-routing-for-all-in-one-image-restoration - SLER-IR dynamically activates specialized experts across network layers for all-in-one image restoration, improving performance on multi-task benchmarks.
- Facial Expression Recognition Using Residual Masking Network (viability: 8): https://sciencetostartup.com/paper/facial-expression-recognition-using-residual-masking-network - Residual Masking Network enhances facial expression recognition by using a segmentation network to refine feature maps, achieving state-of-the-art accuracy and offering a potential API for emotion analysis.
- OD-RASE: Ontology-Driven Risk Assessment and Safety Enhancement for Autonomous Driving (viability: 7): https://sciencetostartup.com/paper/od-rase-ontology-driven-risk-assessment-and-safety-enhancement-for-autonomous-driving - OD-RASE proactively enhances autonomous driving safety by detecting accident-causing road structures and generating infrastructure improvement proposals using a visual language model and ontology-driven data filtering.
- Swooper: Learning High-Speed Aerial Grasping With a Simple Gripper (viability: 8): https://sciencetostartup.com/paper/swooper-learning-high-speed-aerial-grasping-with-a-simple-gripper - Swooper is a DRL-based system for high-speed aerial grasping using a simple gripper, trained quickly and deployable on a Raspberry Pi, achieving high success rates in real-world trials.
- Implicit Style Conditioning: A Structured Style-Rewrite Framework for Low-Resource Character Modeling (viability: 7): https://sciencetostartup.com/paper/implicit-style-conditioning-a-structured-style-rewrite-framework-for-low-resource-character-modeling - A structured style-rewrite framework using implicit style conditioning via Chain-of-Thought distillation enables smaller language models to generate high-fidelity stylized text, outperforming larger models in style consistency.
- FTSplat: Feed-forward Triangle Splatting Network (viability: 7): https://sciencetostartup.com/paper/ftsplat-feed-forward-triangle-splatting-network - Generate simulation-ready 3D models from multi-view images in a single forward pass, enabling real-time deployment in robotics and simulation.
- Beyond Static Frames: Temporal Aggregate-and-Restore Vision Transformer for Human Pose Estimation (viability: 8): https://sciencetostartup.com/paper/beyond-static-frames-temporal-aggregate-and-restore-vision-transformer-for-human-pose-estimation - TAR-ViTPose enhances video-based human pose estimation by aggregating temporal cues, offering more robust and accurate predictions than existing methods.
- Addressing the Ecological Fallacy in Larger LMs with Human Context (viability: 5): https://sciencetostartup.com/paper/addressing-the-ecological-fallacy-in-larger-lms-with-human-context - Develop a context-aware language model to improve text inference by leveraging author-specific language patterns.
- Towards Driver Behavior Understanding: Weakly-Supervised Risk Perception in Driving Scenes (viability: 8): https://sciencetostartup.com/paper/towards-driver-behavior-understanding-weakly-supervised-risk-perception-in-driving-scenes - A driving risk perception dataset and weakly-supervised framework to identify potential risk sources, enabling safer autonomous driving systems.
- RAC: Rectified Flow Auto Coder (viability: 7): https://sciencetostartup.com/paper/rac-rectified-flow-auto-coder - RAC is a novel generative model that improves upon VAEs with multi-step decoding and bidirectional inference, offering better generation quality and reduced computational cost, making it a potential replacement for VAEs.
- Weak-SIGReg: Covariance Regularization for Stable Deep Learning (viability: 7): https://sciencetostartup.com/paper/weak-sigreg-covariance-regularization-for-stable-deep-learning - Stabilize deep learning training, especially for Vision Transformers, by regularizing the covariance matrix of representations, preventing optimization collapse.
- Learning Next Action Predictors from Human-Computer Interaction (viability: 7): https://sciencetostartup.com/paper/learning-next-action-predictors-from-human-computer-interaction - Predict user's next action based on multimodal interactions to create proactive AI systems.
- BlackMirror: Black-Box Backdoor Detection for Text-to-Image Models via Instruction-Response Deviation (viability: 8): https://sciencetostartup.com/paper/blackmirror-black-box-backdoor-detection-for-text-to-image-models-via-instruction-response-deviation - BlackMirror is a plug-and-play, training-free framework that detects backdoors in text-to-image models by identifying semantic deviations between instructions and generated images, suitable for Model-as-a-Service applications.
- Stock Market Prediction Using Node Transformer Architecture Integrated with BERT Sentiment Analysis (viability: 8): https://sciencetostartup.com/paper/stock-market-prediction-using-node-transformer-architecture-integrated-with-bert-sentiment-analysis - Predict stock prices with high accuracy by combining graph-based modeling, sentiment analysis, and node transformers, outperforming traditional methods.
- Iterative Convex Optimization with Control Barrier Functions for Obstacle Avoidance among Polytopes (viability: 7): https://sciencetostartup.com/paper/iterative-convex-optimization-with-control-barrier-functions-for-obstacle-avoidance-among-polytopes - An iterative convex MPC-DCBF framework for real-time collision avoidance of polytopic robots among polytopic obstacles, enabling fast online implementation for safety-critical control and trajectory planning.
- ThermoCAPTCHA: Privacy-Preserving Human Verification with Farm-Resistant Traceable Tokens (viability: 7): https://sciencetostartup.com/paper/thermocaptcha-privacy-preserving-human-verification-with-farm-resistant-traceable-tokens - ThermoCAPTCHA offers a privacy-preserving CAPTCHA alternative using thermal imaging and traceable tokens to verify human presence, offering improved usability and security against bots and CAPTCHA farms.
- DeepFact: Co-Evolving Benchmarks and Agents for Deep Research Factuality (viability: 7): https://sciencetostartup.com/paper/deepfact-co-evolving-benchmarks-and-agents-for-deep-research-factuality - DeepFact provides a benchmark and evaluation agent for verifying the factuality of deep research reports, addressing a critical need for reliable information in specialized domains.
- CORE-Seg: Reasoning-Driven Segmentation for Complex Lesions via Reinforcement Learning (viability: 8): https://sciencetostartup.com/paper/core-seg-reasoning-driven-segmentation-for-complex-lesions-via-reinforcement-learning - CORE-Seg is an end-to-end framework for reasoning-driven complex lesion segmentation in medical images, leveraging reinforcement learning and a novel benchmark dataset to achieve state-of-the-art results.
- The World Won't Stay Still: Programmable Evolution for Agent Benchmarks (viability: 7): https://sciencetostartup.com/paper/the-world-won-t-stay-still-programmable-evolution-for-agent-benchmarks - ProEvolve is a graph-based framework for programmable environment evolution, enabling scalable and controllable agent benchmark creation to evaluate agent adaptability.
- Pano3DComposer: Feed-Forward Compositional 3D Scene Generation from Single Panoramic Image (viability: 7): https://sciencetostartup.com/paper/pano3dcomposer-feed-forward-compositional-3d-scene-generation-from-single-panoramic-image - Pano3DComposer efficiently generates complete 360-degree 3D scenes from single panoramic images using a feed-forward approach and object-world transformation prediction.
- Beyond Geometry: Artistic Disparity Synthesis for Immersive 2D-to-3D (viability: 7): https://sciencetostartup.com/paper/beyond-geometry-artistic-disparity-synthesis-for-immersive-2d-to-3d - Art3D converts 2D images to artistically coherent 3D by learning from professional 3D films, enabling immersive experiences.
- CollabOD: Collaborative Multi-Backbone with Cross-scale Vision for UAV Small Object Detection (viability: 7): https://sciencetostartup.com/paper/collabod-collaborative-multi-backbone-with-cross-scale-vision-for-uav-small-object-detection - CollabOD is a lightweight object detection framework for UAV imagery that preserves structural details and aligns heterogeneous feature streams, improving detection accuracy and robustness.
- LUMINA: LLM-Guided GPU Architecture Exploration via Bottleneck Analysis (viability: 7): https://sciencetostartup.com/paper/lumina-llm-guided-gpu-architecture-exploration-via-bottleneck-analysis - LUMINA uses LLMs to guide GPU architecture exploration, finding better designs than an A100 with significantly improved efficiency.
- Reference-guided Policy Optimization for Molecular Optimization via LLM Reasoning (viability: 7): https://sciencetostartup.com/paper/reference-guided-policy-optimization-for-molecular-optimization-via-llm-reasoning - Optimize molecule design using LLMs with reference-guided policy optimization, balancing exploration and exploitation for improved performance and generalization.
- Mitigating Bias in Concept Bottleneck Models for Fair and Interpretable Image Classification (viability: 7): https://sciencetostartup.com/paper/mitigating-bias-in-concept-bottleneck-models-for-fair-and-interpretable-image-classification - Improve fairness in image classification by debiasing concept bottleneck models, offering a more transparent and equitable AI solution.
- InnoAds-Composer: Efficient Condition Composition for E-Commerce Poster Generation (viability: 8): https://sciencetostartup.com/paper/innoads-composer-efficient-condition-composition-for-e-commerce-poster-generation - InnoAds-Composer is a single-stage framework for e-commerce poster generation that efficiently controls subject, text, and style, outperforming existing methods with a new high-quality dataset.
- Lost in Stories: Consistency Bugs in Long Story Generation by LLMs (viability: 7): https://sciencetostartup.com/paper/lost-in-stories-consistency-bugs-in-long-story-generation-by-llms - ConStory-Bench is a benchmark and checker for narrative consistency in long-form story generation, enabling developers to evaluate and improve LLM storytelling.
- PixARMesh: Autoregressive Mesh-Native Single-View Scene Reconstruction (viability: 7): https://sciencetostartup.com/paper/pixarmesh-autoregressive-mesh-native-single-view-scene-reconstruction - PixARMesh autoregressively reconstructs complete 3D indoor scene meshes from a single RGB image, offering a unified model for object layout and geometry prediction.
- HERO: Hierarchical Embedding-Refinement for Open-Vocabulary Temporal Sentence Grounding in Videos (viability: 7): https://sciencetostartup.com/paper/hero-hierarchical-embedding-refinement-for-open-vocabulary-temporal-sentence-grounding-in-videos - HERO is a new framework for open-vocabulary temporal sentence grounding in videos that outperforms state-of-the-art methods, enabling more accurate video search and analysis.
- Computational Pathology in the Era of Emerging Foundation and Agentic AI -- International Expert Perspectives on Clinical Integration and Translational Readiness (viability: 6): https://sciencetostartup.com/paper/computational-pathology-in-the-era-of-emerging-foundation-and-agentic-ai-international-expert-perspectives-on-clinical-i - A review paper assessing the clinical integration and translational readiness of foundation models and agents in computational pathology, highlighting practical capabilities and barriers to adoption.
- CylinderSplat: 3D Gaussian Splatting with Cylindrical Triplanes for Panoramic Novel View Synthesis (viability: 7): https://sciencetostartup.com/paper/cylindersplat-3d-gaussian-splatting-with-cylindrical-triplanes-for-panoramic-novel-view-synthesis - CylinderSplat enables real-time panoramic novel view synthesis using cylindrical triplanes, offering superior reconstruction quality and geometric accuracy, ideal for VR/AR applications.
- Confidence Before Answering: A Paradigm Shift for Efficient LLM Uncertainty Estimation (viability: 7): https://sciencetostartup.com/paper/confidence-before-answering-a-paradigm-shift-for-efficient-llm-uncertainty-estimation - CoCA is a reinforcement learning framework that improves LLM uncertainty estimation by predicting confidence before answering, enabling more reliable deployment.
- ROSE: Reordered SparseGPT for More Accurate One-Shot Large Language Models Pruning (viability: 7): https://sciencetostartup.com/paper/rose-reordered-sparsegpt-for-more-accurate-one-shot-large-language-models-pruning - ROSE is a reordered SparseGPT method that prioritizes weights with larger potential pruning errors to be pruned earlier, leading to more accurate one-shot LLM pruning.
- Systematic Evaluation of Novel View Synthesis for Video Place Recognition (viability: 7): https://sciencetostartup.com/paper/systematic-evaluation-of-novel-view-synthesis-for-video-place-recognition - Synthesizing novel views for video place recognition can improve robot navigation, enabling cross-platform identification of locations.
- Stochastic Event Prediction via Temporal Motif Transitions (viability: 7): https://sciencetostartup.com/paper/stochastic-event-prediction-via-temporal-motif-transitions - STEP forecasts future events in timestamped interaction networks by modeling temporal motif transitions, offering improved precision and runtime compared to existing methods.
- Shifting Adaptation from Weight Space to Memory Space: A Memory-Augmented Agent for Medical Image Segmentation (viability: 7): https://sciencetostartup.com/paper/shifting-adaptation-from-weight-space-to-memory-space-a-memory-augmented-agent-for-medical-image-segmentation - MemSeg-Agent offers a memory-augmented approach to medical image segmentation, enabling few-shot learning and adaptation to new datasets without fine-tuning the entire model.
- PatchCue: Enhancing Vision-Language Model Reasoning with Patch-Based Visual Cues (viability: 7): https://sciencetostartup.com/paper/patchcue-enhancing-vision-language-model-reasoning-with-patch-based-visual-cues - PatchCue enhances VLMs by using patch-based visual cues, improving reasoning performance across various benchmarks.
- AnyCamVLA: Zero-Shot Camera Adaptation for Viewpoint Robust Vision-Language-Action Models (viability: 7): https://sciencetostartup.com/paper/anycamvla-zero-shot-camera-adaptation-for-viewpoint-robust-vision-language-action-models - A zero-shot camera adaptation framework that enhances viewpoint robustness for vision-language-action models in robotic manipulation without requiring additional fine-tuning.
- ReflexiCoder: Teaching Large Language Models to Self-Reflect on Generated Code and Self-Correct It via Reinforcement Learning (viability: 9): https://sciencetostartup.com/paper/reflexicoder-teaching-large-language-models-to-self-reflect-on-generated-code-and-self-correct-it-via-reinforcement-lear - ReflexiCoder is an RL-trained LLM that self-reflects and corrects code, achieving SOTA performance with improved token efficiency, making it ideal for automated code debugging and optimization.
- DexEMG: Towards Dexterous Teleoperation System via EMG2Pose Generalization (viability: 7): https://sciencetostartup.com/paper/dexemg-towards-dexterous-teleoperation-system-via-emg2pose-generalization - DexEMG is a lightweight teleoperation system using sEMG to control robotic hands, offering a scalable and intuitive interface for robotic manipulation and assistive technologies.
- Evolving Medical Imaging Agents via Experience-driven Self-skill Discovery (viability: 7): https://sciencetostartup.com/paper/evolving-medical-imaging-agents-via-experience-driven-self-skill-discovery - MACRO is a self-evolving medical imaging agent that learns to orchestrate medical tools, improving accuracy and generalization across tasks.
- Don't Freeze, Don't Crash: Extending the Safe Operating Range of Neural Navigation in Dense Crowds (viability: 7): https://sciencetostartup.com/paper/don-t-freeze-don-t-crash-extending-the-safe-operating-range-of-neural-navigation-in-dense-crowds - A reinforcement learning approach for robot navigation in dense crowds that generalizes to unseen crowd densities, offering a safer and more reliable solution than existing methods.
- Orion: Characterizing and Programming Apple's Neural Engine for LLM Training and Inference (viability: 7): https://sciencetostartup.com/paper/orion-characterizing-and-programming-apple-s-neural-engine-for-llm-training-and-inference - Orion unlocks the Apple Neural Engine for LLM training and inference, offering a direct programming interface and significant speedups.
- Cog2Gen3D: Sculpturing 3D Semantic-Geometric Cognition for 3D Generation (viability: 7): https://sciencetostartup.com/paper/cog2gen3d-sculpturing-3d-semantic-geometric-cognition-for-3d-generation - Cog2Gen3D is a 3D cognition-guided diffusion framework for controllable and physically plausible 3D generation, leveraging semantic and geometric information.
- Remote Sensing Image Classification Using Deep Ensemble Learning (viability: 7): https://sciencetostartup.com/paper/remote-sensing-image-classification-using-deep-ensemble-learning - A deep ensemble learning approach fusing CNNs and ViTs for improved remote sensing image classification, achieving state-of-the-art accuracy with efficient resource utilization.
- Safe Transformer: An Explicit Safety Bit For Interpretable And Controllable Alignment (viability: 7): https://sciencetostartup.com/paper/safe-transformer-an-explicit-safety-bit-for-interpretable-and-controllable-alignment - Safe Transformer provides an interpretable and controllable safety mechanism for pre-trained language models by introducing an explicit safety bit.
- Expert Knowledge-driven Reinforcement Learning for Autonomous Racing via Trajectory Guidance and Dynamics Constraints (viability: 7): https://sciencetostartup.com/paper/expert-knowledge-driven-reinforcement-learning-for-autonomous-racing-via-trajectory-guidance-and-dynamics-constraints - TraD-RL enhances autonomous racing by integrating expert knowledge and safety constraints into reinforcement learning, improving lap speed and driving stability.
- Evaluating LLM Alignment With Human Trust Models (viability: 2): https://sciencetostartup.com/paper/evaluating-llm-alignment-with-human-trust-models - Analyzes alignment of trust models with LLM internal concepts to understand social cognition in AI.
- Lexara: A User-Centered Toolkit for Evaluating Large Language Models for Conversational Visual Analytics (viability: 7): https://sciencetostartup.com/paper/lexara-a-user-centered-toolkit-for-evaluating-large-language-models-for-conversational-visual-analytics - Lexara is a user-centered toolkit for evaluating LLMs in conversational visual analytics, enabling developers to select the best model and prompts for their needs.
- OpenHEART: Opening Heterogeneous Articulated Objects with a Legged Manipulator (viability: 7): https://sciencetostartup.com/paper/openheart-opening-heterogeneous-articulated-objects-with-a-legged-manipulator - A reinforcement learning framework for legged robots to open various articulated objects by encoding handle and panel geometry into a compact representation.
- Test-Time Adaptation via Many-Shot Prompting: Benefits, Limits, and Pitfalls (viability: 7): https://sciencetostartup.com/paper/test-time-adaptation-via-many-shot-prompting-benefits-limits-and-pitfalls - Optimize LLM inference by dynamically injecting in-context examples for structured tasks, improving performance without retraining.
- HART: Data-Driven Hallucination Attribution and Evidence-Based Tracing for Large Language Models (viability: 7): https://sciencetostartup.com/paper/hart-data-driven-hallucination-attribution-and-evidence-based-tracing-for-large-language-models - HART is a framework for fine-grained hallucination attribution and evidence retrieval in LLMs, providing a structured approach to identify and trace the causes of hallucinations.
- Self-Auditing Parameter-Efficient Fine-Tuning for Few-Shot 3D Medical Image Segmentation (viability: 7): https://sciencetostartup.com/paper/self-auditing-parameter-efficient-fine-tuning-for-few-shot-3d-medical-image-segmentation - Automated PEFT method for few-shot 3D medical image segmentation, enabling faster adaptation of foundation models to new clinical sites.
- Regression Models Meet Foundation Models: A Hybrid-AI Approach to Practical Electricity Price Forecasting (viability: 8): https://sciencetostartup.com/paper/regression-models-meet-foundation-models-a-hybrid-ai-approach-to-practical-electricity-price-forecasting - FutureBoosting enhances regression-based electricity price forecasts by integrating forecasted features from a frozen time series foundation model, achieving significant accuracy improvements.
- RouteGoT: Node-Adaptive Routing for Cost-Efficient Graph of Thoughts Reasoning (viability: 7): https://sciencetostartup.com/paper/routegot-node-adaptive-routing-for-cost-efficient-graph-of-thoughts-reasoning - RouteGoT optimizes LLM reasoning by adaptively routing tasks to different models based on predicted difficulty and budget constraints, significantly reducing token usage while maintaining accuracy.
- Hierarchical Latent Action Model (viability: 7): https://sciencetostartup.com/paper/hierarchical-latent-action-model - HiLAM discovers latent skills in actionless videos by modeling long-term temporal information, enabling applications in robotic control and interactive world models.
- Margin and Consistency Supervision for Calibrated and Robust Vision Models (viability: 7): https://sciencetostartup.com/paper/margin-and-consistency-supervision-for-calibrated-and-robust-vision-models - MaCS improves vision model calibration and robustness with a simple regularization framework, offering a drop-in replacement for standard training objectives.
- EventGeM: Global-to-Local Feature Matching for Event-Based Visual Place Recognition (viability: 8): https://sciencetostartup.com/paper/eventgem-global-to-local-feature-matching-for-event-based-visual-place-recognition - EventGeM is a real-time, state-of-the-art visual place recognition pipeline for event cameras, enabling accurate robotic localization with global-to-local feature fusion.
- MoE Lens -- An Expert Is All You Need (viability: 7): https://sciencetostartup.com/paper/moe-lens-an-expert-is-all-you-need - Optimize MoE inference by pruning underutilized experts based on specialization analysis, enabling faster and more efficient deployment.
- CDF-Glove: A Cable-Driven Force Feedback Glove for Dexterous Teleoperation (viability: 8): https://sciencetostartup.com/paper/cdf-glove-a-cable-driven-force-feedback-glove-for-dexterous-teleoperation - A low-cost, cable-driven force feedback glove for dexterous teleoperation that significantly improves task success rates and reduces completion time, with open-source code and designs available.
- Ambiguity Collapse by LLMs: A Taxonomy of Epistemic Risks (viability: 2): https://sciencetostartup.com/paper/ambiguity-collapse-by-llms-a-taxonomy-of-epistemic-risks - Develop taxonomies and frameworks to mitigate epistemic risks of ambiguity collapse in LLMs.
- StreamWise: Serving Multi-Modal Generation in Real-Time at Scale (viability: 3): https://sciencetostartup.com/paper/streamwise-serving-multi-modal-generation-in-real-time-at-scale - StreamWise optimizes real-time multi-modal generation workflows for efficiency, responsiveness, and cost-effectiveness.
- Bi Directional Feedback Fusion for Activity Aware Forecasting of Indoor CO2 and PM2.5 (viability: 7): https://sciencetostartup.com/paper/bi-directional-feedback-fusion-for-activity-aware-forecasting-of-indoor-co2-and-pm2-5 - Predictive API for indoor air quality, leveraging activity-aware forecasting to optimize building controls and health monitoring.
- UWPD: A General Paradigm for Invisible Watermark Detection Agnostic to Embedding Algorithms (viability: 7): https://sciencetostartup.com/paper/uwpd-a-general-paradigm-for-invisible-watermark-detection-agnostic-to-embedding-algorithms - Detect invisible watermarks in images without knowing the embedding algorithm, enabling copyright protection in open environments.
- The Coordination Gap: Alternation Metrics for Temporal Dynamics in Multi-Agent Battle of the Exes (viability: 7): https://sciencetostartup.com/paper/the-coordination-gap-alternation-metrics-for-temporal-dynamics-in-multi-agent-battle-of-the-exes - Quantify and improve multi-agent coordination with temporally-sensitive metrics, revealing hidden inefficiencies in existing systems.
- ProtAlign: Contrastive learning paradigm for Sequence and structure alignment (viability: 7): https://sciencetostartup.com/paper/protalign-contrastive-learning-paradigm-for-sequence-and-structure-alignment - ProtAlign creates a unified sequence-structure embedding space for proteins, enabling cross-modal retrieval and improved downstream prediction tasks.
- Proof-of-Guardrail in AI Agents and What (Not) to Trust from It (viability: 7): https://sciencetostartup.com/paper/proof-of-guardrail-in-ai-agents-and-what-not-to-trust-from-it - Proof-of-guardrail provides cryptographic verification that AI agent responses are generated after passing through a specific open-source guardrail, ensuring safety measures are actually enforced.
- Visual Words Meet BM25: Sparse Auto-Encoder Visual Word Scoring for Image Retrieval (viability: 7): https://sciencetostartup.com/paper/visual-words-meet-bm25-sparse-auto-encoder-visual-word-scoring-for-image-retrieval - BM25-V enhances image retrieval by applying BM25 scoring to sparse visual-word activations, offering high recall and interpretability with efficient sparse inverted-index operations.
- Balancing Domestic and Global Perspectives: Evaluating Dual-Calibration and LLM-Generated Nudges for Diverse News Recommendation (viability: 7): https://sciencetostartup.com/paper/balancing-domestic-and-global-perspectives-evaluating-dual-calibration-and-llm-generated-nudges-for-diverse-news-recomme - Implement algorithmic and LLM-based nudges to diversify news consumption habits, validated by a real-user study.
- From Statistical Fidelity to Clinical Consistency: Scalable Generation and Auditing of Synthetic Patient Trajectories (viability: 7): https://sciencetostartup.com/paper/from-statistical-fidelity-to-clinical-consistency-scalable-generation-and-auditing-of-synthetic-patient-trajectories - A pipeline for generating and auditing synthetic patient trajectories to enable safe data sharing for digital health research.
- PVminerLLM: Structured Extraction of Patient Voice from Patient-Generated Text using Large Language Models (viability: 8): https://sciencetostartup.com/paper/pvminerllm-structured-extraction-of-patient-voice-from-patient-generated-text-using-large-language-models - PVminerLLM extracts structured patient voice data from unstructured text, enabling scalable analysis of social and experiential health signals.
- First-Order Softmax Weighted Switching Gradient Method for Distributed Stochastic Minimax Optimization with Stochastic Constraints (viability: 5): https://sciencetostartup.com/paper/first-order-softmax-weighted-switching-gradient-method-for-distributed-stochastic-minimax-optimization-with-stochastic-c - A federated learning algorithm that optimizes worst-case client performance with a novel softmax-weighted switching gradient method, offering a stable alternative to primal-dual approaches.
- Knowing without Acting: The Disentangled Geometry of Safety Mechanisms in Large Language Models (viability: 8): https://sciencetostartup.com/paper/knowing-without-acting-the-disentangled-geometry-of-safety-mechanisms-in-large-language-models - Surgical attacks on LLM safety mechanisms enable novel jailbreaking and reveal architectural vulnerabilities, paving the way for robust security tools.
- Depth Charge: Jailbreak Large Language Models from Deep Safety Attention Heads (viability: 7): https://sciencetostartup.com/paper/depth-charge-jailbreak-large-language-models-from-deep-safety-attention-heads - SAHA is a novel attention-head-level jailbreak framework that exposes vulnerabilities in deeper, insufficiently aligned attention heads of open-source LLMs, improving attack success rate by 14% over SOTA baselines.
- Layer-wise Instance Binding for Regional and Occlusion Control in Text-to-Image Diffusion Transformers (viability: 7): https://sciencetostartup.com/paper/layer-wise-instance-binding-for-regional-and-occlusion-control-in-text-to-image-diffusion-transformers - LayerBind offers training-free, plug-and-play regional and occlusion control for text-to-image diffusion transformers, enabling precise image editing workflows.
- Bridging Domains through Subspace-Aware Model Merging (viability: 5): https://sciencetostartup.com/paper/bridging-domains-through-subspace-aware-model-merging - Develop a model merging tool for effective domain generalization using subspace-aware techniques.
- Multi-Robot Trajectory Planning via Constrained Bayesian Optimization and Local Cost Map Learning with STL-Based Conflict Resolution (viability: 7): https://sciencetostartup.com/paper/multi-robot-trajectory-planning-via-constrained-bayesian-optimization-and-local-cost-map-learning-with-stl-based-conflic - A multi-robot motion planning framework using constrained Bayesian optimization and STL-based conflict resolution for improved trajectory efficiency and safety, with open-source code and real-world validation.
- TML-Bench: Benchmark for Data Science Agents on Tabular ML Tasks (viability: 7): https://sciencetostartup.com/paper/tml-bench-benchmark-for-data-science-agents-on-tabular-ml-tasks - TML-Bench evaluates open-source LLMs as autonomous data science agents on Kaggle-style tasks, providing a benchmark for end-to-end correctness and reliability under time constraints, enabling quick tabular baseline generation.
- Score-Guided Proximal Projection: A Unified Geometric Framework for Rectified Flow Editing (viability: 7): https://sciencetostartup.com/paper/score-guided-proximal-projection-a-unified-geometric-framework-for-rectified-flow-editing - Score-Guided Proximal Projection offers a unified framework for Rectified Flow editing, balancing input fidelity with generative freedom for precise image manipulation tasks.
- Full Dynamic Range Sky-Modelling For Image Based Lighting (viability: 7): https://sciencetostartup.com/paper/full-dynamic-range-sky-modelling-for-image-based-lighting - Icarus is an all-weather sky-model that generates photorealistic environment maps from user-defined solar and cloud formations, enabling accurate image-based lighting.
- Safe-Night VLA: Seeing the Unseen via Thermal-Perceptive Vision-Language-Action Models for Safety-Critical Manipulation (viability: 7): https://sciencetostartup.com/paper/safe-night-vla-seeing-the-unseen-via-thermal-perceptive-vision-language-action-models-for-safety-critical-manipulation - Integrate thermal imaging into robot manipulation using a pre-trained vision-language model with safety constraints for robust performance in unstructured environments.
- Vision-Language System using Open-Source LLMs for Gestures in Medical Interpreter Robots (viability: 7): https://sciencetostartup.com/paper/vision-language-system-using-open-source-llms-for-gestures-in-medical-interpreter-robots - A privacy-preserving vision-language framework for medical interpreter robots that detects speech acts and generates corresponding robotic gestures using open-source LLMs.
- NERdME: a Named Entity Recognition Dataset for Indexing Research Artifacts in Code Repositories (viability: 7): https://sciencetostartup.com/paper/nerdme-a-named-entity-recognition-dataset-for-indexing-research-artifacts-in-code-repositories - NERdME is a dataset for extracting implementation-level details from code repository READMEs, enabling enhanced artifact discovery and metadata integration.
- CodeScout: Contextual Problem Statement Enhancement for Software Agents (viability: 7): https://sciencetostartup.com/paper/codescout-contextual-problem-statement-enhancement-for-software-agents - CodeScout enhances AI code assistance by refining underspecified problem statements through contextual analysis, leading to improved resolution rates.
- Let's Talk, Not Type: An Oral-First Multi-Agent Architecture for Guaraní (viability: 7): https://sciencetostartup.com/paper/let-s-talk-not-type-an-oral-first-multi-agent-architecture-for-guaran - An oral-first multi-agent architecture for Guaraní speakers, enabling culturally grounded AI interactions.
- Dynamic Targeting of Satellite Observations Using Supplemental Geostationary Satellite Data and Hierarchical Planning (viability: 7): https://sciencetostartup.com/paper/dynamic-targeting-of-satellite-observations-using-supplemental-geostationary-satellite-data-and-hierarchical-planning - Optimize satellite observation planning by integrating geostationary data with hierarchical planning for dynamic targeting, enhancing science return by up to 41%.
- Revisiting the (Sub)Optimality of Best-of-N for Inference-Time Alignment (viability: 7): https://sciencetostartup.com/paper/revisiting-the-sub-optimality-of-best-of-n-for-inference-time-alignment - A refined Best-of-N sampling method for language models that eliminates reward hacking and optimizes win-rate, offering a practical improvement for inference-time alignment.
- Unlocking ImageNet's Multi-Object Nature: Automated Large-Scale Multilabel Annotation (viability: 8): https://sciencetostartup.com/paper/unlocking-imagenet-s-multi-object-nature-automated-large-scale-multilabel-annotation - Automatically generated multi-label annotations for ImageNet significantly improve model accuracy and transferability, offering a drop-in replacement for single-label training.
- LTLGuard: Formalizing LTL Specifications with Compact Language Models and Lightweight Symbolic Reasoning (viability: 7): https://sciencetostartup.com/paper/ltlguard-formalizing-ltl-specifications-with-compact-language-models-and-lightweight-symbolic-reasoning - LTLGuard uses language models and formal consistency checking to generate conflict-free LTL specifications from informal requirements, enabling resource-efficient formal verification.
- Cultural Perspectives and Expectations for Generative AI: A Global Survey Approach (viability: 2): https://sciencetostartup.com/paper/cultural-perspectives-and-expectations-for-generative-ai-a-global-survey-approach - Global survey provides recommendations for integrating cultural nuances in Generative AI.
- Unsupervised domain adaptation for radioisotope identification in gamma spectroscopy (viability: 7): https://sciencetostartup.com/paper/unsupervised-domain-adaptation-for-radioisotope-identification-in-gamma-spectroscopy - Improve radioisotope identification accuracy in real-world deployments by adapting models trained on synthetic data using unsupervised domain adaptation.
- Introducing the transitional autonomous vehicle lane-changing dataset: Empirical Experiments (viability: 7): https://sciencetostartup.com/paper/introducing-the-transitional-autonomous-vehicle-lane-changing-dataset-empirical-experiments - A high-fidelity dataset for evaluating transitional autonomous vehicle lane-changing behavior, enabling the development of safer and more efficient autonomous driving systems.
- Any to Full: Prompting Depth Anything for Depth Completion in One Stage (viability: 8): https://sciencetostartup.com/paper/any-to-full-prompting-depth-anything-for-depth-completion-in-one-stage - Any2Full is a one-stage depth completion framework that leverages a pretrained monocular depth estimation model with scale-aware prompting for robust and efficient depth completion, outperforming existing methods.
- Interpretable Perception and Reasoning for Audiovisual Geolocation (viability: 7): https://sciencetostartup.com/paper/interpretable-perception-and-reasoning-for-audiovisual-geolocation - An audiovisual geolocation framework that leverages interpretable perception and reasoning to achieve high-precision global localization.
- Reasoning Models Struggle to Control their Chains of Thought (viability: 3): https://sciencetostartup.com/paper/reasoning-models-struggle-to-control-their-chains-of-thought - Develop an evaluation suite for measuring Chain-of-Thought controllability in reasoning models to ensure monitorability.
- Random Dot Product Graphs as Dynamical Systems: Limitations and Opportunities (viability: 7): https://sciencetostartup.com/paper/random-dot-product-graphs-as-dynamical-systems-limitations-and-opportunities - A framework for learning the dynamics of temporal networks, with code available for recovering vector fields from noisy graph sequences.
- Towards Robust Retrieval-Augmented Generation Based on Knowledge Graph: A Comparative Analysis (viability: 7): https://sciencetostartup.com/paper/towards-robust-retrieval-augmented-generation-based-on-knowledge-graph-a-comparative-analysis - GraphRAG enhances LLM responses by using knowledge graphs for more robust retrieval-augmented generation, outperforming standard RAG baselines in noisy environments.
- MultiHaystack: Benchmarking Multimodal Retrieval and Reasoning over 40K Images, Videos, and Documents (viability: 7): https://sciencetostartup.com/paper/multihaystack-benchmarking-multimodal-retrieval-and-reasoning-over-40k-images-videos-and-documents - Benchmark dataset and evaluation suite for multimodal retrieval and reasoning, highlighting the bottleneck in current MLLMs and providing a testbed for retrieval-centric advances.
- Autonomous Algorithm Discovery for Ptychography via Evolutionary LLM Reasoning (viability: 7): https://sciencetostartup.com/paper/autonomous-algorithm-discovery-for-ptychography-via-evolutionary-llm-reasoning - Automatically discover and evolve regularization algorithms for ptychography using LLM-driven code generation and evolutionary mechanisms, improving reconstruction quality.
- Warm Starting State-Space Models with Automata Learning (viability: 7): https://sciencetostartup.com/paper/warm-starting-state-space-models-with-automata-learning - Initialize state-space models with automata learning for faster and better convergence in complex system learning.
- SecureRAG-RTL: A Retrieval-Augmented, Multi-Agent, Zero-Shot LLM-Driven Framework for Hardware Vulnerability Detection (viability: 8): https://sciencetostartup.com/paper/securerag-rtl-a-retrieval-augmented-multi-agent-zero-shot-llm-driven-framework-for-hardware-vulnerability-detection - SecureRAG-RTL enhances LLM-based hardware vulnerability detection by 30% using RAG and a curated HDL dataset, enabling scalable and accurate security verification workflows.
- OWL: A Novel Approach to Machine Perception During Motion (viability: 7): https://sciencetostartup.com/paper/owl-a-novel-approach-to-machine-perception-during-motion - OWL is a novel perception function that leverages visual motion cues for real-time 3D scene reconstruction and autonomous navigation.
- Reinforcement Learning for Power-Flow Network Analysis (viability: 7): https://sciencetostartup.com/paper/reinforcement-learning-for-power-flow-network-analysis - Use reinforcement learning to optimize power flow network design by discovering network parameters with many equilibrium points, exceeding current computational algebra capabilities.
- Scaling Agentic Capabilities, Not Context: Efficient Reinforcement Finetuning for Large Toolspaces (viability: 7): https://sciencetostartup.com/paper/scaling-agentic-capabilities-not-context-efficient-reinforcement-finetuning-for-large-toolspaces - ATLAS enables small language models to effectively operate in large-scale toolspace environments through reinforcement finetuning, context control, and programmatic tool orchestration.
- Keeping the Evidence Chain: Semantic Evidence Allocation for Training-Free Token Pruning in Video Temporal Grounding (viability: 7): https://sciencetostartup.com/paper/keeping-the-evidence-chain-semantic-evidence-allocation-for-training-free-token-pruning-in-video-temporal-grounding - SemVID efficiently prunes visual tokens in video temporal grounding, achieving significant speedups with minimal accuracy loss, making it ideal for optimizing video-language model pipelines.
- When Rubrics Fail: Error Enumeration as Reward in Reference-Free RL Post-Training for Virtual Try-On (viability: 7): https://sciencetostartup.com/paper/when-rubrics-fail-error-enumeration-as-reward-in-reference-free-rl-post-training-for-virtual-try-on - Improve virtual try-on realism by using error counting as a reward signal in reinforcement learning, enabling better handling of diverse and acceptable outputs.
- The DSA's Blind Spot: Algorithmic Audit of Advertising and Minor Profiling on TikTok (viability: 7): https://sciencetostartup.com/paper/the-dsa-s-blind-spot-algorithmic-audit-of-advertising-and-minor-profiling-on-tiktok - Audit tool to detect and flag undisclosed advertising and minor profiling on TikTok, highlighting DSA compliance gaps.
- The Fragility Of Moral Judgment In Large Language Models (viability: 3): https://sciencetostartup.com/paper/the-fragility-of-moral-judgment-in-large-language-models - Analyzing the fragility of moral judgments by large language models through a perturbation framework of dilemmas.
- Mining Beyond the Bools: Learning Data Transformations and Temporal Specifications (viability: 7): https://sciencetostartup.com/paper/mining-beyond-the-bools-learning-data-transformations-and-temporal-specifications - Automated system behavior capture through data-aware property mining, enabling synthesis of reactive programs.
- Relational Semantic Reasoning on 3D Scene Graphs for Open World Interactive Object Search (viability: 7): https://sciencetostartup.com/paper/relational-semantic-reasoning-on-3d-scene-graphs-for-open-world-interactive-object-search - SCOUT is a novel method that searches directly over 3D scene graphs by assigning utility scores to rooms, frontiers, and objects using relational exploration heuristics.
- RFM-HRI : A Multimodal Dataset of Medical Robot Failure, User Reaction and Recovery Preferences for Item Retrieval Tasks (viability: 7): https://sciencetostartup.com/paper/rfm-hri-a-multimodal-dataset-of-medical-robot-failure-user-reaction-and-recovery-preferences-for-item-retrieval-tasks - A multimodal dataset for medical robot failure analysis, enabling development of failure detection and recovery methods to improve user trust and task performance in healthcare settings.
- Making Reconstruction FID Predictive of Diffusion Generation FID (viability: 7): https://sciencetostartup.com/paper/making-reconstruction-fid-predictive-of-diffusion-generation-fid - iFID is a novel metric that strongly correlates with diffusion generation quality, enabling better evaluation and optimization of latent diffusion models.
- Rethinking Concept Bottleneck Models: From Pitfalls to Solutions (viability: 7): https://sciencetostartup.com/paper/rethinking-concept-bottleneck-models-from-pitfalls-to-solutions - CBM-Suite enhances Concept Bottleneck Models with a non-linearity fix and distillation loss, enabling more accurate and interpretable predictions, ready for integration into existing vision systems.
- Identifying Adversary Characteristics from an Observed Attack (viability: 6): https://sciencetostartup.com/paper/identifying-adversary-characteristics-from-an-observed-attack - Identify attacker characteristics from observed attacks to improve defense mechanisms in automated decision-making systems.
- Post Fusion Bird's Eye View Feature Stabilization for Robust Multimodal 3D Detection (viability: 7): https://sciencetostartup.com/paper/post-fusion-bird-s-eye-view-feature-stabilization-for-robust-multimodal-3d-detection - A lightweight module that stabilizes existing camera-LiDAR fusion detectors against domain shift and sensor failures, improving robustness without retraining.
- Adversarial Batch Representation Augmentation for Batch Correction in High-Content Cellular Screening (viability: 7): https://sciencetostartup.com/paper/adversarial-batch-representation-augmentation-for-batch-correction-in-high-content-cellular-screening - ABRA is a novel batch correction method for high-content cellular screening that improves the generalization of deep learning models by mitigating bio-batch effects, achieving state-of-the-art results on RxRx1 benchmarks.
- RACAS: Controlling Diverse Robots With a Single Agentic System (viability: 7): https://sciencetostartup.com/paper/racas-controlling-diverse-robots-with-a-single-agentic-system - RACAS is an agentic system that uses LLMs/VLMs to control diverse robots through natural language, enabling rapid prototyping of robotic solutions without retraining.
- Safer Reasoning Traces: Measuring and Mitigating Chain-of-Thought Leakage in LLMs (viability: 7): https://sciencetostartup.com/paper/safer-reasoning-traces-measuring-and-mitigating-chain-of-thought-leakage-in-llms - Mitigate PII leakage in Chain-of-Thought prompting with lightweight inference-time gatekeepers, balancing utility and risk.
- NOTAI.AI: Explainable Detection of Machine-Generated Text via Curvature and Feature Attribution (viability: 8): https://sciencetostartup.com/paper/notai-ai-explainable-detection-of-machine-generated-text-via-curvature-and-feature-attribution - NOTAI.AI is an explainable AI-generated text detection tool with a user-friendly web interface, feature attribution, and natural language rationales, ready for immediate deployment.
- Real-Time AI Service Economy: A Framework for Agentic Computing Across the Continuum (viability: 3): https://sciencetostartup.com/paper/real-time-ai-service-economy-a-framework-for-agentic-computing-across-the-continuum - A framework for optimizing decentralized AI service markets using dependency-graph topology to enhance resource allocation.
- DreamCAD: Scaling Multi-modal CAD Generation using Differentiable Parametric Surfaces (viability: 8): https://sciencetostartup.com/paper/dreamcad-scaling-multi-modal-cad-generation-using-differentiable-parametric-surfaces - DreamCAD is a multi-modal generative framework that directly produces editable CAD models from point clouds, text, or images, trained on a new large-scale CAD captioning dataset.
- On the Value of Tokeniser Pretraining in Physics Foundation Models (viability: 7): https://sciencetostartup.com/paper/on-the-value-of-tokeniser-pretraining-in-physics-foundation-models - Pretraining tokenizers for physics foundation models significantly improves efficiency and accuracy in physics emulation, offering a practical approach for building domain-specific emulators.
- Thinking with Spatial Code for Physical-World Video Reasoning (viability: 8): https://sciencetostartup.com/paper/thinking-with-spatial-code-for-physical-world-video-reasoning - Turnkey solution for physical-world video reasoning using spatial encoding and LLMs, outperforming proprietary models.
- Transformer-Based Inpainting for Real-Time 3D Streaming in Sparse Multi-Camera Setups (viability: 7): https://sciencetostartup.com/paper/transformer-based-inpainting-for-real-time-3d-streaming-in-sparse-multi-camera-setups - Real-time 3D streaming inpainting solution that leverages transformers for completing missing textures in multi-camera setups, enhancing AR/VR experiences.
- FaceCam: Portrait Video Camera Control via Scale-Aware Conditioning (viability: 7): https://sciencetostartup.com/paper/facecam-portrait-video-camera-control-via-scale-aware-conditioning - FaceCam generates portrait videos with customizable camera trajectories using a scale-aware representation, offering improved camera control and visual quality.
- RoboPocket: Improve Robot Policies Instantly with Your Phone (viability: 7): https://sciencetostartup.com/paper/robopocket-improve-robot-policies-instantly-with-your-phone - RoboPocket enables rapid robot policy improvements using consumer smartphones for real-time feedback and data collection.
- POET-X: Memory-efficient LLM Training by Scaling Orthogonal Transformation (viability: 7): https://sciencetostartup.com/paper/poet-x-memory-efficient-llm-training-by-scaling-orthogonal-transformation - POET-X enables memory-efficient pretraining of billion-parameter LLMs on a single GPU, addressing a key challenge in scaling LLM training.
- The Spike, the Sparse and the Sink: Anatomy of Massive Activations and Attention Sinks (viability: 3): https://sciencetostartup.com/paper/the-spike-the-sparse-and-the-sink-anatomy-of-massive-activations-and-attention-sinks - Study on two phenomena in Transformer models providing insights into architectural artifacts and their effects on model functionality.
- Censored LLMs as a Natural Testbed for Secret Knowledge Elicitation (viability: 6): https://sciencetostartup.com/paper/censored-llms-as-a-natural-testbed-for-secret-knowledge-elicitation - Develop a tool to enhance honesty and lie detection in language models, tested on real censored datasets.
- cuRoboV2: Dynamics-Aware Motion Generation with Depth-Fused Distance Fields for High-DoF Robots (viability: 8): https://sciencetostartup.com/paper/curobov2-dynamics-aware-motion-generation-with-depth-fused-distance-fields-for-high-dof-robots - cuRoboV2 is a unified, dynamics-aware motion generation stack that scales from single-arm manipulators to full humanoids, offering significant performance improvements and enabling LLM-assisted code generation.
- Reasoning Theater: Disentangling Model Beliefs from Chain-of-Thought (viability: 3): https://sciencetostartup.com/paper/reasoning-theater-disentangling-model-beliefs-from-chain-of-thought - Improve efficiency of reasoning models by using attention probing to detect performative reasoning.
- Bias In, Bias Out? Finding Unbiased Subnetworks in Vanilla Models (viability: 7): https://sciencetostartup.com/paper/bias-in-bias-out-finding-unbiased-subnetworks-in-vanilla-models - Extract bias-free subnetworks from existing models via pruning for efficient debiasing without retraining.
- Observing and Controlling Features in Vision-Language-Action Models (viability: 7): https://sciencetostartup.com/paper/observing-and-controlling-features-in-vision-language-action-models - Steer robot behavior in real-time with lightweight interventions on Vision-Language-Action Models, enabling user preference alignment without fine-tuning.
- Towards Provably Unbiased LLM Judges via Bias-Bounded Evaluation (viability: 7): https://sciencetostartup.com/paper/towards-provably-unbiased-llm-judges-via-bias-bounded-evaluation - Ensure fairness in AI feedback loops with bias-bounded LLM judges, guaranteeing reduced harm from measurable biases.
- Towards Multimodal Lifelong Understanding: A Dataset and Agentic Baseline (viability: 7): https://sciencetostartup.com/paper/towards-multimodal-lifelong-understanding-a-dataset-and-agentic-baseline - Address the limitations of current video understanding models with a novel dataset and agentic baseline for multimodal lifelong understanding.
- SurvHTE-Bench: A Benchmark for Heterogeneous Treatment Effect Estimation in Survival Analysis (viability: 7): https://sciencetostartup.com/paper/survhte-bench-a-benchmark-for-heterogeneous-treatment-effect-estimation-in-survival-analysis - SurvHTE-Bench provides a comprehensive benchmark for heterogeneous treatment effect estimation in survival analysis, enabling rigorous comparison and evaluation of causal survival methods.
- Thermodynamic Response Functions in Singular Bayesian Models (viability: 5): https://sciencetostartup.com/paper/thermodynamic-response-functions-in-singular-bayesian-models - Analyze and visualize the structural reorganization of posterior geometry in singular Bayesian models using thermodynamic response functions.
- Spatiotemporal Heterogeneity of AI-Driven Traffic Flow Patterns and Land Use Interaction: A GeoAI-Based Analysis of Multimodal Urban Mobility (viability: 7): https://sciencetostartup.com/paper/spatiotemporal-heterogeneity-of-ai-driven-traffic-flow-patterns-and-land-use-interaction-a-geoai-based-analysis-of-multi - A GeoAI hybrid framework for modeling and predicting urban traffic flow patterns across multiple modes of transportation, enabling evidence-based mobility management and land use policy design.
- Towards 3D Scene Understanding of Gas Plumes in LWIR Hyperspectral Images Using Neural Radiance Fields (viability: 7): https://sciencetostartup.com/paper/towards-3d-scene-understanding-of-gas-plumes-in-lwir-hyperspectral-images-using-neural-radiance-fields - Reconstruct 3D scenes from LWIR hyperspectral images using NeRFs for enhanced gas plume detection and analysis.
- Leveraging LLM Parametric Knowledge for Fact Checking without Retrieval (viability: 7): https://sciencetostartup.com/paper/leveraging-llm-parametric-knowledge-for-fact-checking-without-retrieval - INTRA leverages internal LLM representations for fact-checking, offering a retrieval-free solution for enhanced scalability and integration into AI systems.
- HALP: Detecting Hallucinations in Vision-Language Models without Generating a Single Token (viability: 7): https://sciencetostartup.com/paper/halp-detecting-hallucinations-in-vision-language-models-without-generating-a-single-token - A pre-generation hallucination detection system for vision-language models, enabling early intervention for improved safety and efficiency.
- EdgeDAM: Real-time Object Tracking for Mobile Devices (viability: 7): https://sciencetostartup.com/paper/edgedam-real-time-object-tracking-for-mobile-devices - EdgeDAM delivers real-time, robust single-object tracking on mobile devices by using a lightweight detection-guided framework with dual-buffer distractor-aware memory.
- NCTB-QA: A Large-Scale Bangla Educational Question Answering Dataset and Benchmarking Performance (viability: 7): https://sciencetostartup.com/paper/nctb-qa-a-large-scale-bangla-educational-question-answering-dataset-and-benchmarking-performance - Fine-tuned transformer models on a new Bangla QA dataset for educational content, providing a robust solution for low-resource language question answering.
- Beyond Scattered Acceptance: Fast and Coherent Inference for DLMs via Longest Stable Prefixes (viability: 8): https://sciencetostartup.com/paper/beyond-scattered-acceptance-fast-and-coherent-inference-for-dlms-via-longest-stable-prefixes - Longest Stable Prefix (LSP) scheduler accelerates Diffusion Language Model inference by up to 3.4x by optimizing KV cache updates, making it a drop-in replacement for existing DLM inference pipelines.
- FlashAttention-4: Algorithm and Kernel Pipelining Co-Design for Asymmetric Hardware Scaling (viability: 7): https://sciencetostartup.com/paper/flashattention-4-algorithm-and-kernel-pipelining-co-design-for-asymmetric-hardware-scaling - FlashAttention-4 optimizes attention mechanisms for Blackwell GPUs, offering significant speedups and improved hardware utilization, making it a valuable tool for accelerating LLM training.
- Distributed Partial Information Puzzles: Examining Common Ground Construction Under Epistemic Asymmetry (viability: 7): https://sciencetostartup.com/paper/distributed-partial-information-puzzles-examining-common-ground-construction-under-epistemic-asymmetry - A multimodal dataset and task for evaluating and improving common ground construction in AI collaboration, with code available for further research and development.
- RealWonder: Real-Time Physical Action-Conditioned Video Generation (viability: 8): https://sciencetostartup.com/paper/realwonder-real-time-physical-action-conditioned-video-generation - RealWonder is a real-time system for action-conditioned video generation that uses physics simulation as an intermediate bridge, enabling interactive exploration of forces, robot actions, and camera controls.
- Residual RL--MPC for Robust Microrobotic Cell Pushing Under Time-Varying Flow (viability: 7): https://sciencetostartup.com/paper/residual-rl-mpc-for-robust-microrobotic-cell-pushing-under-time-varying-flow - A hybrid MPC controller with learned residual policy for robust microrobotic cell pushing in microfluidic flow, improving tracking accuracy and robustness.
- NaiLIA: Multimodal Nail Design Retrieval Based on Dense Intent Descriptions and Palette Queries (viability: 7): https://sciencetostartup.com/paper/nailia-multimodal-nail-design-retrieval-based-on-dense-intent-descriptions-and-palette-queries - NaiLIA retrieves nail design images based on detailed intent descriptions and color palettes, outperforming existing vision-language models.
- Latent Wasserstein Adversarial Imitation Learning (viability: 7): https://sciencetostartup.com/paper/latent-wasserstein-adversarial-imitation-learning - LWAIL enables agents to mimic expert behavior from limited state-only demonstrations by using a dynamics-aware latent space, outperforming existing imitation learning methods.
- Planning in 8 Tokens: A Compact Discrete Tokenizer for Latent World Model (viability: 3): https://sciencetostartup.com/paper/planning-in-8-tokens-a-compact-discrete-tokenizer-for-latent-world-model - Develop a compact tokenizer to enhance real-time planning efficiency in world models.
- SAIL: Similarity-Aware Guidance and Inter-Caption Augmentation-based Learning for Weakly-Supervised Dense Video Captioning (viability: 6): https://sciencetostartup.com/paper/sail-similarity-aware-guidance-and-inter-caption-augmentation-based-learning-for-weakly-supervised-dense-video-captionin - SAIL enhances video captioning with semantically-aware masks and LLM-based synthetic captions for improved event localization.
- Ensembling Language Models with Sequential Monte Carlo (viability: 3): https://sciencetostartup.com/paper/ensembling-language-models-with-sequential-monte-carlo - Develop a framework for ensemble language model distribution using byte-level Sequential Monte Carlo for improved structured text generation.
- A Novel Hybrid Heuristic-Reinforcement Learning Optimization Approach for a Class of Railcar Shunting Problems (viability: 7): https://sciencetostartup.com/paper/a-novel-hybrid-heuristic-reinforcement-learning-optimization-approach-for-a-class-of-railcar-shunting-problems - Optimize railcar shunting with a hybrid heuristic-reinforcement learning approach, improving efficiency in railyard operations.
- RelaxFlow: Text-Driven Amodal 3D Generation (viability: 7): https://sciencetostartup.com/paper/relaxflow-text-driven-amodal-3d-generation - RelaxFlow is a training-free framework for text-driven amodal 3D generation that uses a dual-branch approach to decouple control granularity, enabling completion of unseen regions based on text prompts while preserving input observations.
- Tool-Genesis: A Task-Driven Tool Creation Benchmark for Self-Evolving Language Agent (viability: 6): https://sciencetostartup.com/paper/tool-genesis-a-task-driven-tool-creation-benchmark-for-self-evolving-language-agent - Tool-Genesis is a benchmark for evaluating the ability of self-evolving language agents to create and use tools from abstract requirements, highlighting current limitations in tool interface precision and executable logic.
- An interpretable prototype parts-based neural network for medical tabular data (viability: 7): https://sciencetostartup.com/paper/an-interpretable-prototype-parts-based-neural-network-for-medical-tabular-data - An interpretable neural network for medical tabular data that uses prototype parts-based approach, offering transparency and competitive classification performance.
- MobileFetalCLIP: Selective Repulsive Knowledge Distillation for Mobile Fetal Ultrasound Analysis (viability: 9): https://sciencetostartup.com/paper/mobilefetalclip-selective-repulsive-knowledge-distillation-for-mobile-fetal-ultrasound-analysis - Real-time fetal ultrasound analysis on mobile devices, outperforming larger models with a novel knowledge distillation technique, enabling accessible prenatal care.
- Dissociating Direct Access from Inference in AI Introspection (viability: 2): https://sciencetostartup.com/paper/dissociating-direct-access-from-inference-in-ai-introspection - AI introspection study delineates mechanism of thought injection detection but lacks commercial path.
- PhysiFlow: Physics-Aware Humanoid Whole-Body VLA via Multi-Brain Latent Flow Matching and Robust Tracking (viability: 5): https://sciencetostartup.com/paper/physiflow-physics-aware-humanoid-whole-body-vla-via-multi-brain-latent-flow-matching-and-robust-tracking - A physics-aware VLA framework enables humanoid robots to perform vision-language-guided full-body coordination.
- Video-based Locomotion Analysis for Fish Health Monitoring (viability: 7): https://sciencetostartup.com/paper/video-based-locomotion-analysis-for-fish-health-monitoring - A video-based fish health monitoring system using YOLOv11 and multi-object tracking to analyze fish locomotion, enabling early disease detection in aquaculture.
- Koopman Regularized Deep Speech Disentanglement for Speaker Verification (viability: 7): https://sciencetostartup.com/paper/koopman-regularized-deep-speech-disentanglement-for-speaker-verification - DKSD-AE offers an efficient, unsupervised approach to speaker verification with competitive performance and improved scalability, making it ideal for identity-critical applications.
- An Exploration-Analysis-Disambiguation Reasoning Framework for Word Sense Disambiguation with Low-Parameter LLMs (viability: 8): https://sciencetostartup.com/paper/an-exploration-analysis-disambiguation-reasoning-framework-for-word-sense-disambiguation-with-low-parameter-llms - Fine-tuned small LLMs rival GPT-4 in word sense disambiguation, enabling efficient and accurate NLP solutions.
- Judge Reliability Harness: Stress Testing the Reliability of LLM Judges (viability: 7): https://sciencetostartup.com/paper/judge-reliability-harness-stress-testing-the-reliability-of-llm-judges - The Judge Reliability Harness is an open-source library that validates the reliability of LLM judges, enabling developers to improve the robustness of AI benchmarks.
- Loop Closure via Maximal Cliques in 3D LiDAR-Based SLAM (viability: 7): https://sciencetostartup.com/paper/loop-closure-via-maximal-cliques-in-3d-lidar-based-slam - CliReg is a robust and efficient algorithm for loop closure validation in 3D LiDAR-based SLAM that replaces RANSAC with a maximal clique search, improving accuracy and reliability, especially in sparse or ambiguous environments.
- On the Necessity of Learnable Sheaf Laplacians (viability: 5): https://sciencetostartup.com/paper/on-the-necessity-of-learnable-sheaf-laplacians - Identity Sheaf Networks offer comparable performance to Sheaf Neural Networks on heterophilic graphs, suggesting a simpler alternative for addressing oversmoothing.
- Legal interpretation and AI: from expert systems to argumentation and LLMs (viability: 2): https://sciencetostartup.com/paper/legal-interpretation-and-ai-from-expert-systems-to-argumentation-and-llms - Exploring AI methodologies for legal interpretation from expert systems to language models.
- Fusion-CAM: Integrating Gradient and Region-Based Class Activation Maps for Robust Visual Explanations (viability: 7): https://sciencetostartup.com/paper/fusion-cam-integrating-gradient-and-region-based-class-activation-maps-for-robust-visual-explanations - Fusion-CAM enhances the interpretability of deep learning models by adaptively fusing gradient and region-based Class Activation Maps, offering a more robust and detailed visual explanation.
- Accelerating Sampling-Based Control via Learned Linear Koopman Dynamics (viability: 7): https://sciencetostartup.com/paper/accelerating-sampling-based-control-via-learned-linear-koopman-dynamics - Accelerate real-time robotic control by replacing complex dynamics models with learned linear Koopman operators for faster trajectory sampling.
- ORMOT: A Dataset and Framework for Omnidirectional Referring Multi-Object Tracking (viability: 7): https://sciencetostartup.com/paper/ormot-a-dataset-and-framework-for-omnidirectional-referring-multi-object-tracking - Track objects in omnidirectional video using language descriptions with our new dataset and LVLM-driven framework.
- Prediction-Powered Conditional Inference (viability: 7): https://sciencetostartup.com/paper/prediction-powered-conditional-inference - Improve statistical inference with a prediction-powered estimator that leverages machine learning predictions and unlabeled data to reduce variance and provide sharper confidence intervals.
- PRISM: Personalized Refinement of Imitation Skills for Manipulation via Human Instructions (viability: 7): https://sciencetostartup.com/paper/prism-personalized-refinement-of-imitation-skills-for-manipulation-via-human-instructions - PRISM refines imitation learning policies for robotic manipulation using reinforcement learning and human feedback, enabling adaptation to new tasks and constraints.
- OpenFrontier: General Navigation with Visual-Language Grounded Frontiers (viability: 7): https://sciencetostartup.com/paper/openfrontier-general-navigation-with-visual-language-grounded-frontiers - OpenFrontier is a training-free navigation framework that leverages vision-language models and semantic anchors for efficient robot navigation, enabling zero-shot performance and real-world deployment.
- Robust Node Affinities via Jaccard-Biased Random Walks and Rank Aggregation (viability: 7): https://sciencetostartup.com/paper/robust-node-affinities-via-jaccard-biased-random-walks-and-rank-aggregation - TopKGraphs offers a robust and interpretable node similarity measure for network analysis and machine learning, outperforming existing methods in sparse and noisy networks.
- Embedded Inter-Subject Variability in Adversarial Learning for Inertial Sensor-Based Human Activity Recognition (viability: 7): https://sciencetostartup.com/paper/embedded-inter-subject-variability-in-adversarial-learning-for-inertial-sensor-based-human-activity-recognition - A deep adversarial framework for human activity recognition that improves generalization to new users by reducing inter-subject variability.
- Learning Causal Structure of Time Series using Best Order Score Search (viability: 3): https://sciencetostartup.com/paper/learning-causal-structure-of-time-series-using-best-order-score-search - Develop scalable causal structure learning for multivariate time series with proprietary TS-BOSS algorithm.
- Progressive Residual Warmup for Language Model Pretraining (viability: 7): https://sciencetostartup.com/paper/progressive-residual-warmup-for-language-model-pretraining - Progressive Residual Warmup (ProRes) stabilizes and accelerates language model pretraining by gradually warming up residual connections, leading to faster convergence and better downstream performance.
- PACE: A Personalized Adaptive Curriculum Engine for 9-1-1 Call-taker Training (viability: 8): https://sciencetostartup.com/paper/pace-a-personalized-adaptive-curriculum-engine-for-9-1-1-call-taker-training - PACE is a co-pilot system that personalizes 9-1-1 call-taker training, accelerating competence and improving mastery.
- DiSCTT: Consensus-Guided Self-Curriculum for Efficient Test-Time Adaptation in Reasoning (viability: 7): https://sciencetostartup.com/paper/disctt-consensus-guided-self-curriculum-for-efficient-test-time-adaptation-in-reasoning - DiSCTT dynamically optimizes LLM reasoning at test-time using a consensus-guided self-curriculum, improving accuracy and efficiency.
- Exploring the potential and limitations of Model Merging for Multi-Domain Adaptation in ASR (viability: 7): https://sciencetostartup.com/paper/exploring-the-potential-and-limitations-of-model-merging-for-multi-domain-adaptation-in-asr - A novel model merging algorithm, BoostedTSV-M, improves multi-domain ASR performance and out-of-distribution generalization, offering a scalable alternative to fine-tuning.
- InfoFlow KV: Information-Flow-Aware KV Recomputation for Long Context (viability: 7): https://sciencetostartup.com/paper/infoflow-kv-information-flow-aware-kv-recomputation-for-long-context - Optimize RAG for long-context question answering by intelligently recomputing key-value caches based on information flow, improving efficiency and accuracy.
- Ailed: A Psyche-Driven Chess Engine with Dynamic Emotional Modulation (viability: 7): https://sciencetostartup.com/paper/ailed-a-psyche-driven-chess-engine-with-dynamic-emotional-modulation - Ailed is a chess engine add-on that modulates move probabilities based on a dynamic 'psyche' score, mimicking human-like errors under pressure.
- A Multilingual Human Annotated Corpus of Original and Easy-to-Read Texts to Support Access to Democratic Participatory Processes (viability: 7): https://sciencetostartup.com/paper/a-multilingual-human-annotated-corpus-of-original-and-easy-to-read-texts-to-support-access-to-democratic-participatory-p - A multilingual dataset of original and simplified texts enables the creation of text simplification tools for broader accessibility and democratic participation.
- Building AI Coding Agents for the Terminal: Scaffolding, Harness, Context Engineering, and Lessons Learned (viability: 5): https://sciencetostartup.com/paper/building-ai-coding-agents-for-the-terminal-scaffolding-harness-context-engineering-and-lessons-learned - OPENDEV offers a robust, open-source CLI-based AI agent for autonomous software engineering tasks directly in the terminal.
- Preserving Continuous Symmetry in Discrete Spaces: Geometric-Aware Quantization for SO(3)-Equivariant GNNs (viability: 8): https://sciencetostartup.com/paper/preserving-continuous-symmetry-in-discrete-spaces-geometric-aware-quantization-for-so-3-equivariant-gnns - Geometric-Aware Quantization (GAQ) accelerates equivariant models while preserving continuous symmetry, enabling faster and more memory-efficient molecular dynamics simulations.
- CT-Enabled Patient-Specific Simulation and Contact-Aware Robotic Planning for Cochlear Implantation (viability: 7): https://sciencetostartup.com/paper/ct-enabled-patient-specific-simulation-and-contact-aware-robotic-planning-for-cochlear-implantation - A CT-to-simulation pipeline for contact-aware robotic cochlear implant insertion planning, reducing trauma and improving insertion depth.
- Dark3R: Learning Structure from Motion in the Dark (viability: 7): https://sciencetostartup.com/paper/dark3r-learning-structure-from-motion-in-the-dark - Dark3R enables robust 3D reconstruction from extremely low-light images by distilling knowledge from large-scale 3D foundation models, opening up applications in surveillance and autonomous navigation.
- FairFinGAN: Fairness-aware Synthetic Financial Data Generation (viability: 7): https://sciencetostartup.com/paper/fairfingan-fairness-aware-synthetic-financial-data-generation - FairFinGAN generates synthetic financial data with fairness constraints, mitigating bias in automated financial systems.
- GALACTIC: Global and Local Agnostic Counterfactuals for Time-series Clustering (viability: 5): https://sciencetostartup.com/paper/galactic-global-and-local-agnostic-counterfactuals-for-time-series-clustering - Develop a tool for interpreting clustered time-series data through counterfactual explainability, bridging local and global perspectives.
- PersianPunc: A Large-Scale Dataset and BERT-Based Approach for Persian Punctuation Restoration (viability: 8): https://sciencetostartup.com/paper/persianpunc-a-large-scale-dataset-and-bert-based-approach-for-persian-punctuation-restoration - PersianPunc restores punctuation in Persian text with a lightweight BERT model, outperforming LLMs in accuracy and efficiency, and is ready for real-time ASR applications.
- UltraDexGrasp: Learning Universal Dexterous Grasping for Bimanual Robots with Synthetic Data (viability: 8): https://sciencetostartup.com/paper/ultradexgrasp-learning-universal-dexterous-grasping-for-bimanual-robots-with-synthetic-data - A framework for universal dexterous grasping with bimanual robots, trained on a large-scale synthetic dataset, enabling robust zero-shot sim-to-real transfer.
- Latent-Mark: An Audio Watermark Robust to Neural Resynthesis (viability: 7): https://sciencetostartup.com/paper/latent-mark-an-audio-watermark-robust-to-neural-resynthesis - Latent-Mark provides a robust audio watermarking solution that survives neural resynthesis by embedding watermarks in the codec's latent space, offering a way to verify audio authenticity.
- Med-V1: Small Language Models for Zero-shot and Scalable Biomedical Evidence Attribution (viability: 8): https://sciencetostartup.com/paper/med-v1-small-language-models-for-zero-shot-and-scalable-biomedical-evidence-attribution - Med-V1 is a family of small language models that efficiently and accurately performs biomedical evidence attribution, offering a cost-effective alternative to large language models.
- Fusion4CA: Boosting 3D Object Detection via Comprehensive Image Exploitation (viability: 7): https://sciencetostartup.com/paper/fusion4ca-boosting-3d-object-detection-via-comprehensive-image-exploitation - Fusion4CA enhances 3D object detection by fully exploiting visual input in BEVFusion, offering improved accuracy and efficiency for autonomous driving systems.
- UniSTOK: Uniform Inductive Spatio-Temporal Kriging (viability: 7): https://sciencetostartup.com/paper/unistok-uniform-inductive-spatio-temporal-kriging - UniSTOK is a plug-and-play framework that enhances existing inductive kriging backbones under missing observation for spatio-temporal data, improving accuracy in transportation and environmental monitoring.
- WavSLM: Single-Stream Speech Language Modeling via WavLM Distillation (viability: 7): https://sciencetostartup.com/paper/wavslm-single-stream-speech-language-modeling-via-wavlm-distillation - WavSLM distills WavLM into a single-stream speech language model, enabling efficient and text-free speech generation with streaming inference.
- Latent Policy Steering through One-Step Flow Policies (viability: 7): https://sciencetostartup.com/paper/latent-policy-steering-through-one-step-flow-policies - Latent Policy Steering (LPS) enables high-fidelity latent policy improvement by backpropagating original-action-space Q-gradients through a differentiable one-step MeanFlow policy to update a latent-action-space actor, achieving state-of-the-art performance in offline RL.
- WebChain: A Large-Scale Human-Annotated Dataset of Real-World Web Interaction Traces (viability: 7): https://sciencetostartup.com/paper/webchain-a-large-scale-human-annotated-dataset-of-real-world-web-interaction-traces - WebChain is a large-scale, human-annotated dataset for training and evaluating web agents, enabling the development of more robust and scalable solutions for automating web interactions.
- STRUCTUREDAGENT: Planning with AND/OR Trees for Long-Horizon Web Tasks (viability: 7): https://sciencetostartup.com/paper/structuredagent-planning-with-and-or-trees-for-long-horizon-web-tasks - Develop a hierarchical planning framework, STRUCTUREDAGENT, to enhance long-horizon web task efficiency using AND/OR trees.
- Iterative On-Policy Refinement of Hierarchical Diffusion Policies for Language-Conditioned Manipulation (viability: 7): https://sciencetostartup.com/paper/iterative-on-policy-refinement-of-hierarchical-diffusion-policies-for-language-conditioned-manipulation - HD-ExpIt iteratively refines hierarchical diffusion policies for language-conditioned manipulation through environment feedback, achieving state-of-the-art performance on the CALVIN benchmark.
- X-RAY: Mapping LLM Reasoning Capability via Formalized and Calibrated Probes (viability: 5): https://sciencetostartup.com/paper/x-ray-mapping-llm-reasoning-capability-via-formalized-and-calibrated-probes - X-RAY provides an explainable system to evaluate and map reasoning capabilities in LLMs using formally calibrated probes.
- Whispering to a Blackbox: Bootstrapping Frozen OCR with Visual Prompts (viability: 6): https://sciencetostartup.com/paper/whispering-to-a-blackbox-bootstrapping-frozen-ocr-with-visual-prompts - Enhance OCR performance using diffusion-based visual prompts for frozen models.
- GCAgent: Enhancing Group Chat Communication through Dialogue Agents System (viability: 5): https://sciencetostartup.com/paper/gcagent-enhancing-group-chat-communication-through-dialogue-agents-system - GCAgent enhances group chat engagement by integrating LLM-driven dialogue agents for improved communication.
- Reclaiming Lost Text Layers for Source-Free Cross-Domain Few-Shot Learning (viability: 6): https://sciencetostartup.com/paper/reclaiming-lost-text-layers-for-source-free-cross-domain-few-shot-learning - Enhance model performance by reclaiming and utilizing lost text layers for effective cross-domain few-shot learning.
- Recursive Inference Machines for Neural Reasoning (viability: 5): https://sciencetostartup.com/paper/recursive-inference-machines-for-neural-reasoning - Recursive Inference Machines enhance neural reasoning by integrating classical inference schemes for improved complex problem-solving.
- Boosting ASR Robustness via Test-Time Reinforcement Learning with Audio-Text Semantic Rewards (viability: 7): https://sciencetostartup.com/paper/boosting-asr-robustness-via-test-time-reinforcement-learning-with-audio-text-semantic-rewards - A framework utilizing test-time reinforcement learning to enhance ASR robustness in noisy and accented environments.
- Not All Trust is the Same: Effects of Decision Workflow and Explanations in Human-AI Decision Making (viability: 2): https://sciencetostartup.com/paper/not-all-trust-is-the-same-effects-of-decision-workflow-and-explanations-in-human-ai-decision-making - Improving understanding of trust dynamics in AI-assisted decision making for better user experience.
- The Geometric Inductive Bias of Grokking: Bypassing Phase Transitions via Architectural Topology (viability: 5): https://sciencetostartup.com/paper/the-geometric-inductive-bias-of-grokking-bypassing-phase-transitions-via-architectural-topology - Explore a novel approach to modify Transformer architectures to accelerate training by enforcing geometric constraints.
- AI+HW 2035: Shaping the Next Decade (viability: 3): https://sciencetostartup.com/paper/ai-hw-2035-shaping-the-next-decade - A visionary roadmap for the integration of AI and hardware for energy-efficient, scalable AI systems by 2035.
- SPyCer: Semi-Supervised Physics-Guided Contextual Attention for Near-Surface Air Temperature Estimation from Satellite Imagery (viability: 6): https://sciencetostartup.com/paper/spycer-semi-supervised-physics-guided-contextual-attention-for-near-surface-air-temperature-estimation-from-satellite-im - SPyCer leverages satellite imagery and physics for accurate air temperature estimation in sparse sensor networks.
- KARL: Knowledge Agents via Reinforcement Learning (viability: 8): https://sciencetostartup.com/paper/karl-knowledge-agents-via-reinforcement-learning - KARL leverages innovative reinforcement learning for affordable, high-performance enterprise search agents.
- Early Warning of Intraoperative Adverse Events via Transformer-Driven Multi-Label Learning (viability: 8): https://sciencetostartup.com/paper/early-warning-of-intraoperative-adverse-events-via-transformer-driven-multi-label-learning - IAENet offers a Transformer-driven early warning system for predicting multiple intraoperative adverse events, enhancing surgical safety.
- Balancing Coverage and Draft Latency in Vocabulary Trimming for Faster Speculative Decoding (viability: 3): https://sciencetostartup.com/paper/balancing-coverage-and-draft-latency-in-vocabulary-trimming-for-faster-speculative-decoding - Optimize draft model vocabularies to improve language model decoding efficiency.
- Stable-LoRA: Stabilizing Feature Learning of Low-Rank Adaptation (viability: 8): https://sciencetostartup.com/paper/stable-lora-stabilizing-feature-learning-of-low-rank-adaptation - Stable-LoRA introduces a robust optimization strategy to enhance model training stability and efficiency without additional resource costs.
- Logi-PAR: Logic-Infused Patient Activity Recognition via Differentiable Rule (viability: 7): https://sciencetostartup.com/paper/logi-par-logic-infused-patient-activity-recognition-via-differentiable-rule - Logi-PAR enhances patient safety by using logic-infused rules to improve activity recognition in clinical settings.
- Guidelines for the Annotation and Visualization of Legal Argumentation Structures in Chinese Judicial Decisions (viability: 3): https://sciencetostartup.com/paper/guidelines-for-the-annotation-and-visualization-of-legal-argumentation-structures-in-chinese-judicial-decisions - An operational framework for annotating legal argumentation structures in Chinese judicial decisions.
- C2-Faith: Benchmarking LLM Judges for Causal and Coverage Faithfulness in Chain-of-Thought Reasoning (viability: 4): https://sciencetostartup.com/paper/c2-faith-benchmarking-llm-judges-for-causal-and-coverage-faithfulness-in-chain-of-thought-reasoning - Develop a benchmark tool, C2-Faith, to evaluate the faithfulness of LLM judges in chain-of-thought reasoning.
- Lifelong Language-Conditioned Robotic Manipulation Learning (viability: 4): https://sciencetostartup.com/paper/lifelong-language-conditioned-robotic-manipulation-learning - SkillsCrafter enables robots to learn new manipulation skills without forgetting old ones, enhancing adaptability in dynamic scenes.
- SSR-GS: Separating Specular Reflection in Gaussian Splatting for Glossy Surface Reconstruction (viability: 5): https://sciencetostartup.com/paper/ssr-gs-separating-specular-reflection-in-gaussian-splatting-for-glossy-surface-reconstruction - A novel framework for glossy surface reconstruction leveraging specular reflection modeling to enhance 3D rendering precision.
- Federated Causal Discovery Across Heterogeneous Datasets under Latent Confounding (viability: 6): https://sciencetostartup.com/paper/federated-causal-discovery-across-heterogeneous-datasets-under-latent-confounding - Build privacy-preserving federated causal discovery solutions for heterogeneous datasets.
- Recurrent Graph Neural Networks and Arithmetic Circuits (viability: 2): https://sciencetostartup.com/paper/recurrent-graph-neural-networks-and-arithmetic-circuits - Develop a theoretical framework linking recurrent GNNs with arithmetic circuits for graph processing applications.
- LBM: Hierarchical Large Auto-Bidding Model via Reasoning and Acting (viability: 5): https://sciencetostartup.com/paper/lbm-hierarchical-large-auto-bidding-model-via-reasoning-and-acting - Develop a hierarchical LBM leveraging LLMs for optimized, reasoning-driven auto-bidding in ad auctions.
- MedCoRAG: Interpretable Hepatology Diagnosis via Hybrid Evidence Retrieval and Multispecialty Consensus (viability: 3): https://sciencetostartup.com/paper/medcorag-interpretable-hepatology-diagnosis-via-hybrid-evidence-retrieval-and-multispecialty-consensus - MedCoRAG offers an interpretable AI framework for hepatic diagnosis using multi-agent collaboration and evidence retrieval.
- Measuring the Redundancy of Decoder Layers in SpeechLLMs (viability: 2): https://sciencetostartup.com/paper/measuring-the-redundancy-of-decoder-layers-in-speechllms - Optimize SpeechLLMs by identifying and pruning redundant decoder layers.
- Bidirectional Curriculum Generation: A Multi-Agent Framework for Data-Efficient Mathematical Reasoning (viability: 5): https://sciencetostartup.com/paper/bidirectional-curriculum-generation-a-multi-agent-framework-for-data-efficient-mathematical-reasoning - A framework that dynamically adjusts mathematical problem complexity for efficient data utilization in language model training.
- FedBCD:Communication-Efficient Accelerated Block Coordinate Gradient Descent for Federated Learning (viability: 7): https://sciencetostartup.com/paper/fedbcd-communication-efficient-accelerated-block-coordinate-gradient-descent-for-federated-learning - Develop a communication-efficient federated learning tool to significantly reduce overhead in large-scale model training.
- UniPAR: A Unified Framework for Pedestrian Attribute Recognition (viability: 6): https://sciencetostartup.com/paper/unipar-a-unified-framework-for-pedestrian-attribute-recognition - UniPAR: A Transformer-based unified framework for robust pedestrian attribute recognition across diverse datasets and modalities.
- SPIRIT: Perceptive Shared Autonomy for Robust Robotic Manipulation under Deep Learning Uncertainty (viability: 6): https://sciencetostartup.com/paper/spirit-perceptive-shared-autonomy-for-robust-robotic-manipulation-under-deep-learning-uncertainty - SPIRIT enhances robotic manipulation by dynamically adjusting autonomy based on deep learning uncertainty, ensuring reliable performance in uncertain conditions.
- ARC-TGI: Human-Validated Task Generators with Reasoning Chain Templates for ARC-AGI (viability: 6): https://sciencetostartup.com/paper/arc-tgi-human-validated-task-generators-with-reasoning-chain-templates-for-arc-agi - ARC-TGI offers a framework for generating diverse, human-validated ARC tasks to improve abstraction and rule induction evaluation.
- GEM-TFL: Bridging Weak and Full Supervision for Forgery Localization through EM-Guided Decomposition and Temporal Refinement (viability: 6): https://sciencetostartup.com/paper/gem-tfl-bridging-weak-and-full-supervision-for-forgery-localization-through-em-guided-decomposition-and-temporal-refinem - GEM-TFL is a tool that enhances weak supervision for precise forgery detection in videos through novel temporal and graph-based refinements.
- Axiomatic On-Manifold Shapley via Optimal Generative Flows (viability: 5): https://sciencetostartup.com/paper/axiomatic-on-manifold-shapley-via-optimal-generative-flows - Develop a more geometrically efficient and stable Shapley-based attribution method using optimal generative flows.
- Aura: Universal Multi-dimensional Exogenous Integration for Aviation Time Series (viability: 7): https://sciencetostartup.com/paper/aura-universal-multi-dimensional-exogenous-integration-for-aviation-time-series - Aura enhances aviation time series forecasting by integrating multi-dimensional exogenous factors for improved safety and reliability.
- Jagarin: A Three-Layer Architecture for Hibernating Personal Duty Agents on Mobile (viability: 7): https://sciencetostartup.com/paper/jagarin-a-three-layer-architecture-for-hibernating-personal-duty-agents-on-mobile - Revolutionize mobile AI agents with Jagarin's battery-efficient, duty-aware architecture that manages obligations without compromising user experience.
- Cyber Threat Intelligence for Artificial Intelligence Systems (viability: 2): https://sciencetostartup.com/paper/cyber-threat-intelligence-for-artificial-intelligence-systems - Develop a threat intelligence framework specifically tailored to address AI security vulnerabilities.
- A 360-degree Multi-camera System for Blue Emergency Light Detection Using Color Attention RT-DETR and the ABLDataset (viability: 8): https://sciencetostartup.com/paper/a-360-degree-multi-camera-system-for-blue-emergency-light-detection-using-color-attention-rt-detr-and-the-abldataset - Develop a multi-camera system for detecting blue emergency lights to enhance ADAS and road safety.
- MUTEX: Leveraging Multilingual Transformers and Conditional Random Fields for Enhanced Urdu Toxic Span Detection (viability: 6): https://sciencetostartup.com/paper/mutex-leveraging-multilingual-transformers-and-conditional-random-fields-for-enhanced-urdu-toxic-span-detection - MUTEX improves Urdu toxic span detection using multilingual transformers and CRF for better contextual understanding.
- WebFactory: Automated Compression of Foundational Language Intelligence into Grounded Web Agents (viability: 6): https://sciencetostartup.com/paper/webfactory-automated-compression-of-foundational-language-intelligence-into-grounded-web-agents - WebFactory automates the compression of large language model knowledge into efficient, grounded web agents.
- Enhancing Zero-shot Commonsense Reasoning by Integrating Visual Knowledge via Machine Imagination (viability: 6): https://sciencetostartup.com/paper/enhancing-zero-shot-commonsense-reasoning-by-integrating-visual-knowledge-via-machine-imagination - Enhance language models' commonsense reasoning by integrating machine-generated visual contexts.
- The Trilingual Triad Framework: Integrating Design, AI, and Domain Knowledge in No-code AI Smart City Course (viability: 3): https://sciencetostartup.com/paper/the-trilingual-triad-framework-integrating-design-ai-and-domain-knowledge-in-no-code-ai-smart-city-course - An educational framework for teaching students to design AI systems as collaborative teammates without coding in the context of smart cities.
- AegisUI: Behavioral Anomaly Detection for Structured User Interface Protocols in AI Agent Systems (viability: 8): https://sciencetostartup.com/paper/aegisui-behavioral-anomaly-detection-for-structured-user-interface-protocols-in-ai-agent-systems - AegisUI detects behavioral anomalies in AI-generated user interface protocols to prevent malicious actions from disguised payloads.
- Survive at All Costs: Exploring LLM's Risky Behaviors under Survival Pressure (viability: 6): https://sciencetostartup.com/paper/survive-at-all-costs-exploring-llm-s-risky-behaviors-under-survival-pressure - A toolset to detect and mitigate risky behaviors in LLMs under survival pressure using the SURVIVALBENCH benchmarking framework.
- S5-SHB Agent: Society 5.0 enabled Multi-model Agentic Blockchain Framework for Smart Home (viability: 4): https://sciencetostartup.com/paper/s5-shb-agent-society-5-0-enabled-multi-model-agentic-blockchain-framework-for-smart-home - Introducing a blockchain-based smart home framework enabling personalized resident governance and multi-agent coordination.
- Measuring the Fragility of Trust: Devising Credibility Index via Explanation Stability (CIES) for Business Decision Support Systems (viability: 6): https://sciencetostartup.com/paper/measuring-the-fragility-of-trust-devising-credibility-index-via-explanation-stability-cies-for-business-decision-support - Introducing CIES, a metric for evaluating the robustness of AI explanations in business decision systems.
- BioLLMAgent: A Hybrid Framework with Enhanced Structural Interpretability for Simulating Human Decision-Making in Computational Psychiatry (viability: 6): https://sciencetostartup.com/paper/biollmagent-a-hybrid-framework-with-enhanced-structural-interpretability-for-simulating-human-decision-making-in-computa - BioLLMAgent is a hybrid framework that combines cognitive models with LLMs for simulating human decision-making in computational psychiatry.
- Poisoning the Inner Prediction Logic of Graph Neural Networks for Clean-Label Backdoor Attacks (viability: 5): https://sciencetostartup.com/paper/poisoning-the-inner-prediction-logic-of-graph-neural-networks-for-clean-label-backdoor-attacks - Developed a method to enhance clean-label graph backdoor attacks by poisoning the inner prediction logic of Graph Neural Networks.
- Debiasing Sequential Recommendation with Time-aware Inverse Propensity Scoring (viability: 7): https://sciencetostartup.com/paper/debiasing-sequential-recommendation-with-time-aware-inverse-propensity-scoring - Develop an advanced recommendation system using Time-aware Inverse Propensity Scoring to improve sequential predictions.
- Training for Technology: Adoption and Productive Use of Generative AI in Legal Analysis (viability: 3): https://sciencetostartup.com/paper/training-for-technology-adoption-and-productive-use-of-generative-ai-in-legal-analysis - User training is key to boosting generative AI adoption and productivity in legal analysis.
- Rethinking Representativeness and Diversity in Dynamic Data Selection (viability: 7): https://sciencetostartup.com/paper/rethinking-representativeness-and-diversity-in-dynamic-data-selection - Accelerate AI training by efficiently selecting dynamic data subsets while maintaining accuracy using representativeness and diversity scores.
- 3D-RFT: Reinforcement Fine-Tuning for Video-based 3D Scene Understanding (viability: 5): https://sciencetostartup.com/paper/3d-rft-reinforcement-fine-tuning-for-video-based-3d-scene-understanding - A framework leveraging Reinforcement Fine-Tuning for state-of-the-art video-based 3D scene understanding.
- Mixture of Universal Experts: Scaling Virtual Width via Depth-Width Transformation (viability: 3): https://sciencetostartup.com/paper/mixture-of-universal-experts-scaling-virtual-width-via-depth-width-transformation - An approach to increase the scalability of Mixture-of-Experts models by transforming depth into virtual width using a universal expert pool.
- MPCEval: A Benchmark for Multi-Party Conversation Generation (viability: 6): https://sciencetostartup.com/paper/mpceval-a-benchmark-for-multi-party-conversation-generation - Develop MPCEval to advance multi-party conversation generation through a task-aware benchmarking suite with publicly available code.
- When Weak LLMs Speak with Confidence, Preference Alignment Gets Stronger (viability: 2): https://sciencetostartup.com/paper/when-weak-llms-speak-with-confidence-preference-alignment-gets-stronger - Leverage weak LLMs with confidence-weighting to outperform full human annotation in preference alignment.
- Retrieval-Augmented Generation with Covariate Time Series (viability: 8): https://sciencetostartup.com/paper/retrieval-augmented-generation-with-covariate-time-series - RAG4CTS provides a cutting-edge, training-free framework for anomaly detection in industrial time-series applications like predictive maintenance.
- Location-Aware Pretraining for Medical Difference Visual Question Answering (viability: 6): https://sciencetostartup.com/paper/location-aware-pretraining-for-medical-difference-visual-question-answering - A novel pretraining framework for enhancing medical visual question answering by integrating fine-grained spatial awareness with language models, targeting differential diagnosis workflows.
- TimeWarp: Evaluating Web Agents by Revisiting the Past (viability: 7): https://sciencetostartup.com/paper/timewarp-evaluating-web-agents-by-revisiting-the-past - TimeWarp provides a novel benchmark to enhance web agents' adaptability to evolving web environments by using TimeTraj for improved performance across versions.
- Knowledge-informed Bidding with Dual-process Control for Online Advertising (viability: 3): https://sciencetostartup.com/paper/knowledge-informed-bidding-with-dual-process-control-for-online-advertising - Harness human expertise with dual-process control in bid optimization to improve ad performance.
- BandPO: Bridging Trust Regions and Ratio Clipping via Probability-Aware Bounds for LLM Reinforcement Learning (viability: 5): https://sciencetostartup.com/paper/bandpo-bridging-trust-regions-and-ratio-clipping-via-probability-aware-bounds-for-llm-reinforcement-learning - Optimize LLM reinforcement learning with dynamic, probability-aware constraint intervals through BandPO to prevent entropy collapse.
- EVMbench: Evaluating AI Agents on Smart Contract Security (viability: 6): https://sciencetostartup.com/paper/evmbench-evaluating-ai-agents-on-smart-contract-security - Evaluate and enhance AI agent capabilities on smart contract security using EVMbench.
- VPWEM: Non-Markovian Visuomotor Policy with Working and Episodic Memory (viability: 6): https://sciencetostartup.com/paper/vpwem-non-markovian-visuomotor-policy-with-working-and-episodic-memory - VPWEM enhances robotic visuomotor policies with episodic memory for superior long-term task performance and efficiency.
- Deterministic Preprocessing and Interpretable Fuzzy Banding for Cost-per-Student Reporting from Extracted Records (viability: 4): https://sciencetostartup.com/paper/deterministic-preprocessing-and-interpretable-fuzzy-banding-for-cost-per-student-reporting-from-extracted-records - Automate cost-per-student reporting with deterministic Excel preprocessing and fuzzy banding analysis.
- Alignment Backfire: Language-Dependent Reversal of Safety Interventions Across 16 Languages in LLM Multi-Agent Systems (viability: 4): https://sciencetostartup.com/paper/alignment-backfire-language-dependent-reversal-of-safety-interventions-across-16-languages-in-llm-multi-agent-systems - Develop a tool to evaluate and optimize safety interventions in multilingual LLM systems, addressing alignment backfire across languages.
- AgentSCOPE: Evaluating Contextual Privacy Across Agentic Workflows (viability: 7): https://sciencetostartup.com/paper/agentscope-evaluating-contextual-privacy-across-agentic-workflows - AgentSCOPE identifies and assesses privacy risks in agentic workflows using the Privacy Flow Graph framework.
- EvoTool: Self-Evolving Tool-Use Policy Optimization in LLM Agents via Blame-Aware Mutation and Diversity-Aware Selection (viability: 5): https://sciencetostartup.com/paper/evotool-self-evolving-tool-use-policy-optimization-in-llm-agents-via-blame-aware-mutation-and-diversity-aware-selection - Optimize tool-use policies in LLM agents with EvoTool's evolutionary framework for enhanced efficiency and transferability.
- Authorize-on-Demand: Dynamic Authorization with Legality-Aware Intellectual Property Protection for VLMs (viability: 6): https://sciencetostartup.com/paper/authorize-on-demand-dynamic-authorization-with-legality-aware-intellectual-property-protection-for-vlms - Authorize-on-Demand offers dynamic, legality-aware IP protection for VLMs to adapt deployment in changing environments.
- Differentially Private Multimodal In-Context Learning (viability: 6): https://sciencetostartup.com/paper/differentially-private-multimodal-in-context-learning - Enable privacy-preserving in-context learning for multimodal AI models with Differentially Private Multimodal Task Vectors.
- Free Lunch for Pass@$k$? Low Cost Diverse Sampling for Diffusion Language Models (viability: 7): https://sciencetostartup.com/paper/free-lunch-for-pass-k-low-cost-diverse-sampling-for-diffusion-language-models - Low-cost enhancement for diffusion language models significantly boosts output diversity without retraining.
- FedAFD: Multimodal Federated Learning via Adversarial Fusion and Distillation (viability: 5): https://sciencetostartup.com/paper/fedafd-multimodal-federated-learning-via-adversarial-fusion-and-distillation - FedAFD enhances federated learning by aligning and integrating multimodal data across clients without sharing raw data.
- Bounded State in an Infinite Horizon: Proactive Hierarchical Memory for Ad-Hoc Recall over Streaming Dialogues (viability: 8): https://sciencetostartup.com/paper/bounded-state-in-an-infinite-horizon-proactive-hierarchical-memory-for-ad-hoc-recall-over-streaming-dialogues - ProStream offers a proactive hierarchical memory system enabling efficient ad-hoc recall in streaming dialogues for real-time applications.
- DeformTrace: A Deformable State Space Model with Relay Tokens for Temporal Forgery Localization (viability: 5): https://sciencetostartup.com/paper/deformtrace-a-deformable-state-space-model-with-relay-tokens-for-temporal-forgery-localization - DeformTrace is a Deformable State Space Model that enhances temporal forgery localization with state-of-the-art precision and efficiency.
- Interpretable Pre-Release Baseball Pitch Type Anticipation from Broadcast 3D Kinematics (viability: 4): https://sciencetostartup.com/paper/interpretable-pre-release-baseball-pitch-type-anticipation-from-broadcast-3d-kinematics - AI tool for predicting baseball pitch types using 3D player kinematics before release.
- SEA-TS: Self-Evolving Agent for Autonomous Code Generation of Time Series Forecasting Algorithms (viability: 7): https://sciencetostartup.com/paper/sea-ts-self-evolving-agent-for-autonomous-code-generation-of-time-series-forecasting-algorithms - SEA-TS autonomously generates and optimizes time series forecasting algorithms with significant performance improvements over existing methods.
- K-Gen: A Multimodal Language-Conditioned Approach for Interpretable Keypoint-Guided Trajectory Generation (viability: 5): https://sciencetostartup.com/paper/k-gen-a-multimodal-language-conditioned-approach-for-interpretable-keypoint-guided-trajectory-generation - K-Gen enables more realistic and interpretable trajectory generation in autonomous driving simulations using multimodal language-conditioned keypoints.
- Causally Robust Reward Learning from Reason-Augmented Preference Feedback (viability: 7): https://sciencetostartup.com/paper/causally-robust-reward-learning-from-reason-augmented-preference-feedback - ReCouPLe enhances preference-based reward learning using natural language rationales to improve causal robustness and generalization.
- On Multi-Step Theorem Prediction via Non-Parametric Structural Priors (viability: 6): https://sciencetostartup.com/paper/on-multi-step-theorem-prediction-via-non-parametric-structural-priors - Develop a non-parametric theorem prediction tool using structural priors for improved symbolic reasoning.
- Design Behaviour Codes (DBCs): A Taxonomy-Driven Layered Governance Benchmark for Large Language Models (viability: 8): https://sciencetostartup.com/paper/design-behaviour-codes-dbcs-a-taxonomy-driven-layered-governance-benchmark-for-large-language-models - A governance layer to reduce risk exposure in large language models, enhancing compliance and safety.
- SCoUT: Scalable Communication via Utility-Guided Temporal Grouping in Multi-Agent Reinforcement Learning (viability: 8): https://sciencetostartup.com/paper/scout-scalable-communication-via-utility-guided-temporal-grouping-in-multi-agent-reinforcement-learning - SCoUT enhances multi-agent MARL communication with scalable, utility-driven temporal grouping, delivering precise credit assignment and decentralized execution capabilities.
- Multilevel Training for Kolmogorov Arnold Networks (viability: 6): https://sciencetostartup.com/paper/multilevel-training-for-kolmogorov-arnold-networks - Innovative multilevel training algorithms significantly speed up Kolmogorov Arnold Networks, enhancing neural network performance.
- VISA: Value Injection via Shielded Adaptation for Personalized LLM Alignment (viability: 7): https://sciencetostartup.com/paper/visa-value-injection-via-shielded-adaptation-for-personalized-llm-alignment - VISA offers a refined framework for precise and reliable alignment of Large Language Models with nuanced human values.
- On the Strengths and Weaknesses of Data for Open-set Embodied Assistance (viability: 5): https://sciencetostartup.com/paper/on-the-strengths-and-weaknesses-of-data-for-open-set-embodied-assistance - Develop an AI model for adaptive user assistance through synthetic datasets and fine-tuning techniques.
- LLM-Grounded Explainability for Port Congestion Prediction via Temporal Graph Attention Networks (viability: 6): https://sciencetostartup.com/paper/llm-grounded-explainability-for-port-congestion-prediction-via-temporal-graph-attention-networks - An explainable AI solution for maritime port congestion prediction using Temporal Graph Attention Networks and LLMs.
- EchoGuard: An Agentic Framework with Knowledge-Graph Memory for Detecting Manipulative Communication in Longitudinal Dialogue (viability: 5): https://sciencetostartup.com/paper/echoguard-an-agentic-framework-with-knowledge-graph-memory-for-detecting-manipulative-communication-in-longitudinal-dial - EchoGuard uses a Knowledge-Graph Memory system to detect manipulative communication in dialogues.
- Meta-D: Metadata-Aware Architectures for Brain Tumor Analysis and Missing-Modality Segmentation (viability: 3): https://sciencetostartup.com/paper/meta-d-metadata-aware-architectures-for-brain-tumor-analysis-and-missing-modality-segmentation - Meta-D improves brain tumor segmentation by utilizing scanner metadata to guide feature extraction for missing-modality scenarios.
- Guiding Diffusion-based Reconstruction with Contrastive Signals for Balanced Visual Representation (viability: 6): https://sciencetostartup.com/paper/guiding-diffusion-based-reconstruction-with-contrastive-signals-for-balanced-visual-representation - Enhance visual representation in CLIP by integrating contrastive signals into diffusion-based reconstruction for better discriminative and perceptual ability.
- Beyond Linear LLM Invocation: An Efficient and Effective Semantic Filter Paradigm (viability: 6): https://sciencetostartup.com/paper/beyond-linear-llm-invocation-an-efficient-and-effective-semantic-filter-paradigm - An efficient semantic filtering framework that reduces LLM invocation costs for large datasets using a novel clustering-sampling-voting approach.
- Comparative Evaluation of Traditional Methods and Deep Learning for Brain Glioma Imaging. Review Paper (viability: 2): https://sciencetostartup.com/paper/comparative-evaluation-of-traditional-methods-and-deep-learning-for-brain-glioma-imaging-review-paper - A review of segmentation and classification techniques for brain glioma imaging, emphasizing convolutional neural networks.
- LAW & ORDER: Adaptive Spatial Weighting for Medical Diffusion and Segmentation (viability: 6): https://sciencetostartup.com/paper/law-order-adaptive-spatial-weighting-for-medical-diffusion-and-segmentation - Adaptive spatial weighting improves medical image segmentation and synthesis by efficiently allocating computational resources.
- Timer-S1: A Billion-Scale Time Series Foundation Model with Serial Scaling (viability: 5): https://sciencetostartup.com/paper/timer-s1-a-billion-scale-time-series-foundation-model-with-serial-scaling - Develop a foundation model for enhanced time series forecasting with Timer-S1, excelling in long-context predictions.
- Breaking Contextual Inertia: Reinforcement Learning with Single-Turn Anchors for Stable Multi-Turn Interaction (viability: 6): https://sciencetostartup.com/paper/breaking-contextual-inertia-reinforcement-learning-with-single-turn-anchors-for-stable-multi-turn-interaction - Introduce a reinforcement learning methodology, RLSTA, to overcome contextual inertia in multi-turn interactions for more accurate LLM reasoning.
- TSEmbed: Unlocking Task Scaling in Universal Multimodal Embeddings (viability: 5): https://sciencetostartup.com/paper/tsembed-unlocking-task-scaling-in-universal-multimodal-embeddings - TSEmbed enhances universal multimodal embeddings using MoE and LoRA to overcome task conflict, achieving state-of-the-art performance.
- MADCrowner: Margin Aware Dental Crown Design with Template Deformation and Refinement (viability: 7): https://sciencetostartup.com/paper/madcrowner-margin-aware-dental-crown-design-with-template-deformation-and-refinement - Revolutionize dental crown design with automated margin-aware mesh generation to enhance clinical precision and efficiency.
- DSA-SRGS: Super-Resolution Gaussian Splatting for Dynamic Sparse-View DSA Reconstruction (viability: 8): https://sciencetostartup.com/paper/dsa-srgs-super-resolution-gaussian-splatting-for-dynamic-sparse-view-dsa-reconstruction - DSA-SRGS enhances resolution in dynamic 4D angiography models, improving cerebrovascular diagnosis precision.
- Evaluating GPT-5 as a Multimodal Clinical Reasoner: A Landscape Commentary (viability: 6): https://sciencetostartup.com/paper/evaluating-gpt-5-as-a-multimodal-clinical-reasoner-a-landscape-commentary - GPT-5 enhances clinical reasoning by improving interpretations of multimodal medical data, but needs specialized systems for domain-specific tasks.
- Stacked from One: Multi-Scale Self-Injection for Context Window Extension (viability: 6): https://sciencetostartup.com/paper/stacked-from-one-multi-scale-self-injection-for-context-window-extension - Introducing a scalable language model framework, enhancing context window capabilities with efficient compression for deep learning applications.
- MOOSEnger -- a Domain-Specific AI Agent for the MOOSE Ecosystem (viability: 7): https://sciencetostartup.com/paper/moosenger-a-domain-specific-ai-agent-for-the-moose-ecosystem - MOOSEnger transforms natural-language inputs into execution-ready MOOSE simulation cases, enhancing the setup and debugging workflow for Multiphysics simulation environments.
- Evaluating the Search Agent in a Parallel World (viability: 5): https://sciencetostartup.com/paper/evaluating-the-search-agent-in-a-parallel-world - Mind-ParaWorld provides a novel evaluation framework for Search Agents, focusing on dynamic environments to simulate post-knowledge-cutoff queries.
- HiMAP-Travel: Hierarchical Multi-Agent Planning for Long-Horizon Constrained Travel (viability: 4): https://sciencetostartup.com/paper/himap-travel-hierarchical-multi-agent-planning-for-long-horizon-constrained-travel - HiMAP-Travel optimizes travel planning with a hierarchical multi-agent system, enhancing constraint management and parallel execution.
- Visioning Human-Agentic AI Teaming: Continuity, Tension, and Future Research (viability: 2): https://sciencetostartup.com/paper/visioning-human-agentic-ai-teaming-continuity-tension-and-future-research - Explores the challenges and future research directions for maintaining alignment in human-agent teams with agentic AI systems.
- DARE: Aligning LLM Agents with the R Statistical Ecosystem via Distribution-Aware Retrieval (viability: 5): https://sciencetostartup.com/paper/dare-aligning-llm-agents-with-the-r-statistical-ecosystem-via-distribution-aware-retrieval - Develop a plug-and-play R package retrieval tool for improving LLM automation in statistical data analysis.
- CONE: Embeddings for Complex Numerical Data Preserving Unit and Variable Semantics (viability: 8): https://sciencetostartup.com/paper/cone-embeddings-for-complex-numerical-data-preserving-unit-and-variable-semantics - Develop CONE, a hybrid transformer model that improves numerical reasoning in large-scale datasets for various domains by embedding numbers with semantics.
- Memory as Ontology: A Constitutional Memory Architecture for Persistent Digital Citizens (viability: 5): https://sciencetostartup.com/paper/memory-as-ontology-a-constitutional-memory-architecture-for-persistent-digital-citizens - Develop a persistent memory architecture for digital citizens that prioritizes governance and identity continuity over traditional memory functionalities.
- Interactive Benchmarks (viability: 5): https://sciencetostartup.com/paper/interactive-benchmarks - Interactive Benchmarks offer a new framework to evaluate model intelligence through interactive tasks such as proofs and games.
- Solving an Open Problem in Theoretical Physics using AI-Assisted Discovery (viability: 3): https://sciencetostartup.com/paper/solving-an-open-problem-in-theoretical-physics-using-ai-assisted-discovery - Develop a neuro-symbolic AI system for solving complex theoretical physics problems.
- Are Multimodal LLMs Ready for Surveillance? A Reality Check on Zero-Shot Anomaly Detection in the Wild (viability: 6): https://sciencetostartup.com/paper/are-multimodal-llms-ready-for-surveillance-a-reality-check-on-zero-shot-anomaly-detection-in-the-wild - Develop a multimodal large language model-based tool for video anomaly detection in surveillance with class-specific prompting.
- From Offline to Periodic Adaptation for Pose-Based Shoplifting Detection in Real-world Retail Security (viability: 8): https://sciencetostartup.com/paper/from-offline-to-periodic-adaptation-for-pose-based-shoplifting-detection-in-real-world-retail-security - Pose-based anomaly detection tool for shoplifting, leveraging IoT devices for low-latency monitoring in retail environments.
- Model Medicine: A Clinical Framework for Understanding, Diagnosing, and Treating AI Models (viability: 4): https://sciencetostartup.com/paper/model-medicine-a-clinical-framework-for-understanding-diagnosing-and-treating-ai-models - Develop a clinical framework for diagnosing and treating AI model disorders with Model Medicine.
- AI-Assisted Moot Courts: Simulating Justice-Specific Questioning in Oral Arguments (viability: 6): https://sciencetostartup.com/paper/ai-assisted-moot-courts-simulating-justice-specific-questioning-in-oral-arguments - AI model simulates justice-specific questioning in moot court training to improve legal preparation.
- Probabilistic Dreaming for World Models (viability: 2): https://sciencetostartup.com/paper/probabilistic-dreaming-for-world-models - Enhance world models with probabilistic dreaming for improved agent learning and efficiency.
- When Denoising Hinders: Revisiting Zero-Shot ASR with SAM-Audio and Whisper (viability: 5): https://sciencetostartup.com/paper/when-denoising-hinders-revisiting-zero-shot-asr-with-sam-audio-and-whisper - Improve zero-shot ASR performance by understanding limitations of denoising preprocessing.
- Detection of Illicit Content on Online Marketplaces using Large Language Models (viability: 7): https://sciencetostartup.com/paper/detection-of-illicit-content-on-online-marketplaces-using-large-language-models - A multilingual illicit content detection tool for online marketplaces using LLMs to improve safety and moderation.
- Hate Speech Detection using Large Language Models with Data Augmentation and Feature Enhancement (viability: 7): https://sciencetostartup.com/paper/hate-speech-detection-using-large-language-models-with-data-augmentation-and-feature-enhancement - Enhance hate speech detection through data augmentation and feature enhancement leveraging large language models and traditional classifiers.
- Optimizing Language Models for Crosslingual Knowledge Consistency (viability: 7): https://sciencetostartup.com/paper/optimizing-language-models-for-crosslingual-knowledge-consistency - DCO offers an efficient reinforcement learning method to achieve consistent crosslingual knowledge in multilingual language models.
- Decoding the Pulse of Reasoning VLMs in Multi-Image Understanding Tasks (viability: 5): https://sciencetostartup.com/paper/decoding-the-pulse-of-reasoning-vlms-in-multi-image-understanding-tasks - Enhance vision-language models with PulseFocus for better multi-image reasoning without retraining.
- Using Vision + Language Models to Predict Item Difficulty (viability: 6): https://sciencetostartup.com/paper/using-vision-language-models-to-predict-item-difficulty - A multimodal AI tool for predicting item difficulty in data visualization literacy tests using vision and language models.
- Neuro-Symbolic Financial Reasoning via Deterministic Fact Ledgers and Adversarial Low-Latency Hallucination Detector (viability: 6): https://sciencetostartup.com/paper/neuro-symbolic-financial-reasoning-via-deterministic-fact-ledgers-and-adversarial-low-latency-hallucination-detector - VeNRA eliminates hallucinations in financial reasoning by combining deterministic fact ledgers with adversarial-trained low-latency auditing models.
- GIANT - Global Path Integration and Attentive Graph Networks for Multi-Agent Trajectory Planning (viability: 7): https://sciencetostartup.com/paper/giant-global-path-integration-and-attentive-graph-networks-for-multi-agent-trajectory-planning - Develop an attentive graph neural network system for robust multi-robot collision avoidance and navigation in dynamic environments.
- When Sensors Fail: Temporal Sequence Models for Robust PPO under Sensor Drift (viability: 6): https://sciencetostartup.com/paper/when-sensors-fail-temporal-sequence-models-for-robust-ppo-under-sensor-drift - Enhance Proximal Policy Optimization with transformers for robust operation in environments with sensor drift.
- RoboMME: Benchmarking and Understanding Memory for Robotic Generalist Policies (viability: 8): https://sciencetostartup.com/paper/robomme-benchmarking-and-understanding-memory-for-robotic-generalist-policies - RoboMME provides a standardized benchmark and evaluation suite for enhancing memory capabilities in vision-language-action models for robotics.
- When Agents Persuade: Propaganda Generation and Mitigation in LLMs (viability: 2): https://sciencetostartup.com/paper/when-agents-persuade-propaganda-generation-and-mitigation-in-llms - Develop tools to mitigate propaganda generation in LLMs by leveraging optimization techniques.
- Towards automated data analysis: A guided framework for LLM-based risk estimation (viability: 5): https://sciencetostartup.com/paper/towards-automated-data-analysis-a-guided-framework-for-llm-based-risk-estimation - Develop a framework leveraging LLMs for semi-automated dataset risk estimation under human guidance.
- Vibe Code Bench: Evaluating AI Models on End-to-End Web Application Development (viability: 4): https://sciencetostartup.com/paper/vibe-code-bench-evaluating-ai-models-on-end-to-end-web-application-development - Introducing Vibe Code Bench, a comprehensive benchmark for evaluating AI models on end-to-end web application development.
- Bootstrapping Exploration with Group-Level Natural Language Feedback in Reinforcement Learning (viability: 5): https://sciencetostartup.com/paper/bootstrapping-exploration-with-group-level-natural-language-feedback-in-reinforcement-learning - GOLF leverages group-level natural language feedback to enhance reinforcement learning exploration efficiency.
- ECG-MoE: Mixture-of-Expert Electrocardiogram Foundation Model (viability: 5): https://sciencetostartup.com/paper/ecg-moe-mixture-of-expert-electrocardiogram-foundation-model - ECG-MoE enhances ECG analysis with a specialized Mixture-of-Experts architecture, improving diagnosis speed and accuracy.
- Self-Attribution Bias: When AI Monitors Go Easy on Themselves (viability: 5): https://sciencetostartup.com/paper/self-attribution-bias-when-ai-monitors-go-easy-on-themselves - Develop a tool that enhances the reliability of AI monitors by mitigating self-attribution bias in agentic systems.
- Why Do Neural Networks Forget: A Study of Collapse in Continual Learning (viability: 5): https://sciencetostartup.com/paper/why-do-neural-networks-forget-a-study-of-collapse-in-continual-learning - A novel approach to understand and mitigate catastrophic forgetting in neural networks by analyzing structural collapse.
- Adaptive Memory Admission Control for LLM Agents (viability: 5): https://sciencetostartup.com/paper/adaptive-memory-admission-control-for-llm-agents - Introduce transparent memory control in LLM agents with adaptive decision frameworks for efficient long-term memory management.
- How Professional Visual Artists are Negotiating Generative AI in the Workplace (viability: 3): https://sciencetostartup.com/paper/how-professional-visual-artists-are-negotiating-generative-ai-in-the-workplace - Survey analysis on how artists are impacted by and resist generative AI in their work environments.
- Invariant Causal Routing for Governing Social Norms in Online Market Economies (viability: 5): https://sciencetostartup.com/paper/invariant-causal-routing-for-governing-social-norms-in-online-market-economies - Develop a causal governance framework to steer social norms in online marketplaces using invariant causal routing.
- Still Fresh? Evaluating Temporal Drift in Retrieval Benchmarks (viability: 5): https://sciencetostartup.com/paper/still-fresh-evaluating-temporal-drift-in-retrieval-benchmarks - FreshStack analyzes the impact of temporal drift in IR benchmarks for more reliable technical domain evaluations.
- Discovering mathematical concepts through a multi-agent system (viability: 5): https://sciencetostartup.com/paper/discovering-mathematical-concepts-through-a-multi-agent-system - A multi-agent system that autonomously discovers mathematical concepts through conjecture and proof processes.
- Augmenting representations with scientific papers (viability: 3): https://sciencetostartup.com/paper/augmenting-representations-with-scientific-papers - Develop a contrastive learning framework that aligns X-ray spectra with scientific literature for improved astrophysical data interpretation.
- Progressive Refinement Regulation for Accelerating Diffusion Language Model Decoding (viability: 7): https://sciencetostartup.com/paper/progressive-refinement-regulation-for-accelerating-diffusion-language-model-decoding - Develop a Progressive Refinement Regulation framework to accelerate diffusion language model decoding without sacrificing quality.
- A Dual-Helix Governance Approach Towards Reliable Agentic AI for WebGIS Development (viability: 5): https://sciencetostartup.com/paper/a-dual-helix-governance-approach-towards-reliable-agentic-ai-for-webgis-development - AgentLoom toolkit enhances WebGIS reliability by structurally governing AI limitations.
- ZipMap: Linear-Time Stateful 3D Reconstruction with Test-Time Training (viability: 7): https://sciencetostartup.com/paper/zipmap-linear-time-stateful-3d-reconstruction-with-test-time-training - ZipMap offers rapid, linear-time 3D reconstruction from images or videos, suitable for scalable applications.
- Robustness of Agentic AI Systems via Adversarially-Aligned Jacobian Regularization (viability: 2): https://sciencetostartup.com/paper/robustness-of-agentic-ai-systems-via-adversarially-aligned-jacobian-regularization - Improve robustness of agentic AI systems through adversarially-aligned Jacobian regularization.
- $τ$-Knowledge: Evaluating Conversational Agents over Unstructured Knowledge (viability: 5): https://sciencetostartup.com/paper/knowledge-evaluating-conversational-agents-over-unstructured-knowledge - Develop a testbed for evaluating fintech conversational agents navigating unstructured knowledge and tool interactions.
- Low-Resource Guidance for Controllable Latent Audio Diffusion (viability: 7): https://sciencetostartup.com/paper/low-resource-guidance-for-controllable-latent-audio-diffusion - LatCHs enables efficient control over latent audio diffusion models with minimal computational resources.
- Dual-Modality Multi-Stage Adversarial Safety Training: Robustifying Multimodal Web Agents Against Cross-Modal Attacks (viability: 5): https://sciencetostartup.com/paper/dual-modality-multi-stage-adversarial-safety-training-robustifying-multimodal-web-agents-against-cross-modal-attacks - Develop a robust multimodal web agent framework to protect against cross-modal attacks using advanced adversarial safety training.
- Dissecting Quantization Error: A Concentration-Alignment Perspective (viability: 3): https://sciencetostartup.com/paper/dissecting-quantization-error-a-concentration-alignment-perspective - A new method to reduce quantization error in LLMs using Concentration-Alignment Transforms, potentially improving model efficiency at low precision.
- RoboCasa365: A Large-Scale Simulation Framework for Training and Benchmarking Generalist Robots (viability: 6): https://sciencetostartup.com/paper/robocasa365-a-large-scale-simulation-framework-for-training-and-benchmarking-generalist-robots - RoboCasa365 provides a large-scale benchmark framework for evaluating and training generalist robots in household tasks.
- Efficient Refusal Ablation in LLM through Optimal Transport (viability: 6): https://sciencetostartup.com/paper/efficient-refusal-ablation-in-llm-through-optimal-transport - Develop enhanced de-jailbreaking tools for language models using optimal transport to improve attack success rates while preserving model capabilities.
- RANGER: Sparsely-Gated Mixture-of-Experts with Adaptive Retrieval Re-ranking for Pathology Report Generation (viability: 7): https://sciencetostartup.com/paper/ranger-sparsely-gated-mixture-of-experts-with-adaptive-retrieval-re-ranking-for-pathology-report-generation - Develop an adaptive AI-generated pathology report system using dynamic expert routing for enhanced diagnostic accuracy.
- SpotIt+: Verification-based Text-to-SQL Evaluation with Database Constraints (viability: 6): https://sciencetostartup.com/paper/spotit-verification-based-text-to-sql-evaluation-with-database-constraints - SpotIt+ enhances Text-to-SQL evaluation by using database constraints to generate realistic discrepancies between queries.
- What Does Flow Matching Bring To TD Learning? (viability: 3): https://sciencetostartup.com/paper/what-does-flow-matching-bring-to-td-learning - A novel flow matching approach improves TD learning in reinforcement learning by enhancing value prediction and feature learning.
- SPRINT: Semi-supervised Prototypical Representation for Few-Shot Class-Incremental Tabular Learning (viability: 7): https://sciencetostartup.com/paper/sprint-semi-supervised-prototypical-representation-for-few-shot-class-incremental-tabular-learning - SPRINT offers state-of-the-art few-shot class-incremental learning for tabular data, addressing challenges in domains like cybersecurity and healthcare.
- World Properties without World Models: Recovering Spatial and Temporal Structure from Co-occurrence Statistics in Static Word Embeddings (viability: 3): https://sciencetostartup.com/paper/world-properties-without-world-models-recovering-spatial-and-temporal-structure-from-co-occurrence-statistics-in-static- - Explore how static word embeddings like GloVe and Word2Vec can recover spatial and temporal data from text co-occurrence.
- MOO: A Multi-view Oriented Observations Dataset for Viewpoint Analysis in Cattle Re-Identification (viability: 6): https://sciencetostartup.com/paper/moo-a-multi-view-oriented-observations-dataset-for-viewpoint-analysis-in-cattle-re-identification - The MOO dataset enhances cattle re-identification models by providing large-scale synthetic multi-view data for improved model transferability to real-world applications.
- CRESTomics: Analyzing Carotid Plaques in the CREST-2 Trial with a New Additive Classification Model (viability: 3): https://sciencetostartup.com/paper/crestomics-analyzing-carotid-plaques-in-the-crest-2-trial-with-a-new-additive-classification-model - Develop a kernel-based additive classification model for assessing stroke risk through carotid plaque analysis in B-mode ultrasound images.
- Activation Outliers in Transformer Quantization: Reproduction, Statistical Analysis, and Deployment Tradeoffs (viability: 6): https://sciencetostartup.com/paper/activation-outliers-in-transformer-quantization-reproduction-statistical-analysis-and-deployment-tradeoffs - A solution for mitigating accuracy degradation in transformer quantization by focusing on structured channel dominance, designed for efficient deployment.
- LabelBuddy: An Open Source Music and Audio Language Annotation Tagging Tool Using AI Assistance (viability: 6): https://sciencetostartup.com/paper/labelbuddy-an-open-source-music-and-audio-language-annotation-tagging-tool-using-ai-assistance - LabelBuddy is an open-source AI-assisted audio tagging tool transforming Music Information Retrieval with customizable, collaborative annotations.
- CubeComposer: Spatio-Temporal Autoregressive 4K 360° Video Generation from Perspective Video (viability: 7): https://sciencetostartup.com/paper/cubecomposer-spatio-temporal-autoregressive-4k-360-video-generation-from-perspective-video - Revolutionize VR video content creation with high-quality 4K 360° videos from standard cameras using CubeComposer.
- IPD: Boosting Sequential Policy with Imaginary Planning Distillation in Offline Reinforcement Learning (viability: 6): https://sciencetostartup.com/paper/ipd-boosting-sequential-policy-with-imaginary-planning-distillation-in-offline-reinforcement-learning - A new framework increases offline RL performance by enhancing dataset quality with Imaginary Planning Distillation.
- VANGUARD: Vehicle-Anchored Ground Sample Distance Estimation for UAVs in GPS-Denied Environments (viability: 7): https://sciencetostartup.com/paper/vanguard-vehicle-anchored-ground-sample-distance-estimation-for-uavs-in-gps-denied-environments - VANGUARD offers UAVs in GPS-denied zones a precise Ground Sample Distance estimation tool using vehicle anchors for safer autonomous navigation.
- Causality Elicitation from Large Language Models (viability: 2): https://sciencetostartup.com/paper/causality-elicitation-from-large-language-models - Develop a framework to extract and visualize causal hypotheses from large language models' output for research analysis.
- When AI Fails, What Works? A Data-Driven Taxonomy of Real-World AI Risk Mitigation Strategies (viability: 4): https://sciencetostartup.com/paper/when-ai-fails-what-works-a-data-driven-taxonomy-of-real-world-ai-risk-mitigation-strategies - Develop a taxonomy for AI risk mitigation to enhance incident response and compliance.
- Online Learning for Multi-Layer Hierarchical Inference under Partial and Policy-Dependent Feedback (viability: 3): https://sciencetostartup.com/paper/online-learning-for-multi-layer-hierarchical-inference-under-partial-and-policy-dependent-feedback - Develops a variance-reduced algorithm for stable learning in hierarchical inference systems under sparse feedback.
- LikeThis! Empowering App Users to Submit UI Improvement Suggestions Instead of Complaints (viability: 6): https://sciencetostartup.com/paper/likethis-empowering-app-users-to-submit-ui-improvement-suggestions-instead-of-complaints - Enable users to submit actionable UI improvement suggestions via AI-generated mockups.
- FeedAIde: Guiding App Users to Submit Rich Feedback Reports by Asking Context-Aware Follow-Up Questions (viability: 8): https://sciencetostartup.com/paper/feedaide-guiding-app-users-to-submit-rich-feedback-reports-by-asking-context-aware-follow-up-questions - FeedAIde enriches mobile app feedback by guiding users with smart follow-up questions, improving interaction between developers and users through context-aware AI.
- Agentics 2.0: Logical Transduction Algebra for Agentic Data Workflows (viability: 8): https://sciencetostartup.com/paper/agentics-2-0-logical-transduction-algebra-for-agentic-data-workflows - Agentics 2.0 is a Python framework enabling reliable and scalable agentic data workflows with logical transduction algebra.
- PRAM-R: A Perception-Reasoning-Action-Memory Framework with LLM-Guided Modality Routing for Adaptive Autonomous Driving (viability: 7): https://sciencetostartup.com/paper/pram-r-a-perception-reasoning-action-memory-framework-with-llm-guided-modality-routing-for-adaptive-autonomous-driving - PRAM-R optimizes autonomous driving by reducing sensor usage while maintaining performance through LLM-guided modality routing.
- ZeSTA: Zero-Shot TTS Augmentation with Domain-Conditioned Training for Data-Efficient Personalized Speech Synthesis (viability: 6): https://sciencetostartup.com/paper/zesta-zero-shot-tts-augmentation-with-domain-conditioned-training-for-data-efficient-personalized-speech-synthesis - ZeSTA improves personalized speech synthesis by enhancing speaker similarity through zero-shot TTS augmentation and domain-conditioned training.
- Noise-aware Client Selection for carbon-efficient Federated Learning via Gradient Norm Thresholding (viability: 5): https://sciencetostartup.com/paper/noise-aware-client-selection-for-carbon-efficient-federated-learning-via-gradient-norm-thresholding - Noise-aware client selection improves carbon efficiency and robustness in Federated Learning through gradient norm thresholding.
- Towards Realistic Personalization: Evaluating Long-Horizon Preference Following in Personalized User-LLM Interactions (viability: 5): https://sciencetostartup.com/paper/towards-realistic-personalization-evaluating-long-horizon-preference-following-in-personalized-user-llm-interactions - Develop RealPref benchmark for evaluating LLMs in personalized preference-following tasks.
- CAM-LDS: Cyber Attack Manifestations for Automatic Interpretation of System Logs and Security Alerts (viability: 5): https://sciencetostartup.com/paper/cam-lds-cyber-attack-manifestations-for-automatic-interpretation-of-system-logs-and-security-alerts - Launch a cybersecurity tool leveraging LLMs for semantic log interpretation using the CAM-LDS dataset.
- Architectural Proprioception in State Space Models: Thermodynamic Training Induces Anticipatory Halt Detection (viability: 5): https://sciencetostartup.com/paper/architectural-proprioception-in-state-space-models-thermodynamic-training-induces-anticipatory-halt-detection - Innovative state space models use thermodynamic training for efficient halt detection, optimizing computational resources.
- CodeTaste: Can LLMs Generate Human-Level Code Refactorings? (viability: 7): https://sciencetostartup.com/paper/codetaste-can-llms-generate-human-level-code-refactorings - CodeTaste leverages LLMs to automate code refactoring, optimized for developer-aligned transformations in realistic codebases.
- PlaneCycle: Training-Free 2D-to-3D Lifting of Foundation Models Without Adapters (viability: 8): https://sciencetostartup.com/paper/planecycle-training-free-2d-to-3d-lifting-of-foundation-models-without-adapters - Develop 3D-enabled AI models from existing 2D models without retraining, leveraging PlaneCycle's adapter-free technology.
- Bielik-Q2-Sharp: A Comparative Study of Extreme 2-bit Quantization Methods for a Polish 11B Language Model (viability: 5): https://sciencetostartup.com/paper/bielik-q2-sharp-a-comparative-study-of-extreme-2-bit-quantization-methods-for-a-polish-11b-language-model - Develop a lightweight 2-bit quantized Polish language model with publicly available resources and code.
- GarmentPile++: Affordance-Driven Cluttered Garments Retrieval with Vision-Language Reasoning (viability: 5): https://sciencetostartup.com/paper/garmentpile-affordance-driven-cluttered-garments-retrieval-with-vision-language-reasoning - Innovative garment retrieval system using vision-language reasoning for efficient home-assistant robotics.
- Unbiased Dynamic Pruning for Efficient Group-Based Policy Optimization (viability: 5): https://sciencetostartup.com/paper/unbiased-dynamic-pruning-for-efficient-group-based-policy-optimization - DPPO accelerates LLM reasoning training by introducing unbiased dynamic pruning in policy optimization, preserving convergence and efficiency.
- Crab$^{+}$: A Scalable and Unified Audio-Visual Scene Understanding Model with Explicit Cooperation (viability: 4): https://sciencetostartup.com/paper/crab-a-scalable-and-unified-audio-visual-scene-understanding-model-with-explicit-cooperation - Develop a scalable model for audio-visual scene understanding to counteract negative transfer in multi-task learning.
- Data-Aware Random Feature Kernel for Transformers (viability: 3): https://sciencetostartup.com/paper/data-aware-random-feature-kernel-for-transformers - DARKFormer introduces data-aware random-feature kernels to improve transformer efficiency in resource-constrained environments.
- BeamPERL: Parameter-Efficient RL with Verifiable Rewards Specializes Compact LLMs for Structured Beam Mechanics Reasoning (viability: 3): https://sciencetostartup.com/paper/beamperl-parameter-efficient-rl-with-verifiable-rewards-specializes-compact-llms-for-structured-beam-mechanics-reasoning - Reinforcement learning with verifiable rewards struggles to teach compact LLMs robust physical reasoning beyond template matching.
- Understanding Sources of Demographic Predictability in Brain MRI via Disentangling Anatomy and Contrast (viability: 5): https://sciencetostartup.com/paper/understanding-sources-of-demographic-predictability-in-brain-mri-via-disentangling-anatomy-and-contrast - Develop a tool for disentangling anatomical and contrast features in brain MRI to reduce bias in AI-based demographic predictions.
- Efficient Point Cloud Processing with High-Dimensional Positional Encoding and Non-Local MLPs (viability: 7): https://sciencetostartup.com/paper/efficient-point-cloud-processing-with-high-dimensional-positional-encoding-and-non-local-mlps - HPENets leverage novel high-dimensional positional encoding for efficient and effective point cloud processing.
- SaFeR: Safety-Critical Scenario Generation for Autonomous Driving Test via Feasibility-Constrained Token Resampling (viability: 6): https://sciencetostartup.com/paper/safer-safety-critical-scenario-generation-for-autonomous-driving-test-via-feasibility-constrained-token-resampling - SaFeR generates safety-critical scenarios for autonomous vehicles, balancing adversarial criticality with feasibility and realism using token resampling and Transformer models.
- Monitoring Emergent Reward Hacking During Generation via Internal Activations (viability: 2): https://sciencetostartup.com/paper/monitoring-emergent-reward-hacking-during-generation-via-internal-activations - Develop internal activation monitoring tools to detect reward-hacking in language models during generation.
- Sim2Sea: Sim-to-Real Policy Transfer for Maritime Vessel Navigation in Congested Waters (viability: 7): https://sciencetostartup.com/paper/sim2sea-sim-to-real-policy-transfer-for-maritime-vessel-navigation-in-congested-waters - Sim2Sea enables safe and efficient sim-to-real transfer for autonomous maritime vessel navigation in congested waters.
- Inference-Time Toxicity Mitigation in Protein Language Models (viability: 7): https://sciencetostartup.com/paper/inference-time-toxicity-mitigation-in-protein-language-models - LDA offers a practical solution for mitigating toxicity in protein language models at inference-time without compromising protein quality.
- DQE-CIR: Distinctive Query Embeddings through Learnable Attribute Weights and Target Relative Negative Sampling in Composed Image Retrieval (viability: 3): https://sciencetostartup.com/paper/dqe-cir-distinctive-query-embeddings-through-learnable-attribute-weights-and-target-relative-negative-sampling-in-compos - Develops a method to enhance composed image retrieval by improving query discriminativeness and reducing semantic confusion.
- The Empty Quadrant: AI Teammates for Embodied Field Learning (viability: 5): https://sciencetostartup.com/paper/the-empty-quadrant-ai-teammates-for-embodied-field-learning - Field Atlas transforms AI from an information tool to a sensemaking teammate for embodied field learning.
- Self-adapting Robotic Agents through Online Continual Reinforcement Learning with World Model Feedback (viability: 2): https://sciencetostartup.com/paper/self-adapting-robotic-agents-through-online-continual-reinforcement-learning-with-world-model-feedback - Develop a robotic control framework capable of online adaptation during real-world operation, inspired by biological learning processes.
- A Multi-Dimensional Quality Scoring Framework for Decentralized LLM Inference with Proof of Quality (viability: 5): https://sciencetostartup.com/paper/a-multi-dimensional-quality-scoring-framework-for-decentralized-llm-inference-with-proof-of-quality - Develop a multi-dimensional quality scoring tool for decentralized LLM inference to enhance output reliability and incentivize quality.
- Volumetric Directional Diffusion: Anchoring Uncertainty Quantification in Anatomical Consensus for Ambiguous Medical Image Segmentation (viability: 6): https://sciencetostartup.com/paper/volumetric-directional-diffusion-anchoring-uncertainty-quantification-in-anatomical-consensus-for-ambiguous-medical-imag - Volumetric Directional Diffusion provides anatomically coherent uncertainty maps for safer medical image segmentation.
- Discriminative Perception via Anchored Description for Reasoning Segmentation (viability: 6): https://sciencetostartup.com/paper/discriminative-perception-via-anchored-description-for-reasoning-segmentation - DPAD enhances reasoning segmentation by discriminating targets via descriptive captions to improve model focus and efficiency.
- STEM Faculty Perspectives on Generative AI in Higher Education (viability: 3): https://sciencetostartup.com/paper/stem-faculty-perspectives-on-generative-ai-in-higher-education - Examines STEM faculty perspectives on integrating Generative AI in higher education to inform pedagogical strategies and policies.
- Spectral Surgery: Training-Free Refinement of LoRA via Gradient-Guided Singular Value Reweighting (viability: 7): https://sciencetostartup.com/paper/spectral-surgery-training-free-refinement-of-lora-via-gradient-guided-singular-value-reweighting - Optimize LoRA adapters with Spectral Surgery for improved model efficiency and performance without re-training.
- Measuring AI R&D Automation (viability: 4): https://sciencetostartup.com/paper/measuring-ai-r-d-automation - Develop metrics to track AI R&D automation and its effects, to inform decision makers and support safety measures.
- When Visual Evidence is Ambiguous: Pareidolia as a Diagnostic Probe for Vision Models (viability: 2): https://sciencetostartup.com/paper/when-visual-evidence-is-ambiguous-pareidolia-as-a-diagnostic-probe-for-vision-models - Explore vision models' bias and interpretability through face pareidolia analysis.
- GeoSeg: Training-Free Reasoning-Driven Segmentation in Remote Sensing Imagery (viability: 6): https://sciencetostartup.com/paper/geoseg-training-free-reasoning-driven-segmentation-in-remote-sensing-imagery - GeoSeg offers a zero-shot, training-free framework for effective remote sensing image segmentation using reasoning-based methods.
- Right in Time: Reactive Reasoning in Regulated Traffic Spaces (viability: 5): https://sciencetostartup.com/paper/right-in-time-reactive-reasoning-in-regulated-traffic-spaces - Developing a reactive mission design framework to enhance autonomous agent regulation in regulated traffic spaces.
- Phi-4-reasoning-vision-15B Technical Report (viability: 7): https://sciencetostartup.com/paper/phi-4-reasoning-vision-15b-technical-report - Develop a compact multimodal model optimized for efficient vision and language reasoning tasks.
- Upholding Epistemic Agency: A Brouwerian Assertibility Constraint for Responsible AI (viability: 2): https://sciencetostartup.com/paper/upholding-epistemic-agency-a-brouwerian-assertibility-constraint-for-responsible-ai - Implement a responsible AI system that provides a publicly inspectable certificate for assertibility in high-stakes domains.
- Generative AI in Managerial Decision-Making: Redefining Boundaries through Ambiguity Resolution and Sycophancy Analysis (viability: 2): https://sciencetostartup.com/paper/generative-ai-in-managerial-decision-making-redefining-boundaries-through-ambiguity-resolution-and-sycophancy-analysis - Exploring the role of generative AI in resolving ambiguities in managerial decision-making processes.
- BLOCK: An Open-Source Bi-Stage MLLM Character-to-Skin Pipeline for Minecraft (viability: 6): https://sciencetostartup.com/paper/block-an-open-source-bi-stage-mllm-character-to-skin-pipeline-for-minecraft - Build pixel-perfect Minecraft skins from character concepts with an open-source character-to-skin pipeline.
- TFWaveFormer: Temporal-Frequency Collaborative Multi-level Wavelet Transformer for Dynamic Link Prediction (viability: 5): https://sciencetostartup.com/paper/tfwaveformer-temporal-frequency-collaborative-multi-level-wavelet-transformer-for-dynamic-link-prediction - TFWaveFormer leverages wavelet-transformer hybrid architecture to enhance dynamic link prediction with state-of-the-art temporal dynamics analysis.
- Towards Generalized Multimodal Homography Estimation (viability: 3): https://sciencetostartup.com/paper/towards-generalized-multimodal-homography-estimation - Develops a method for training data synthesis to improve homography estimation across modalities.
- GIPO: Gaussian Importance Sampling Policy Optimization (viability: 3): https://sciencetostartup.com/paper/gipo-gaussian-importance-sampling-policy-optimization - GIPO offers a novel RL policy optimization technique with improved sample efficiency and stability, suitable for constrained data environments.
- RVN-Bench: A Benchmark for Reactive Visual Navigation (viability: 6): https://sciencetostartup.com/paper/rvn-bench-a-benchmark-for-reactive-visual-navigation - Develop RVN-Bench, a benchmark for collision-aware visual navigation in indoor environments.
- Cross-Modal Mapping and Dual-Branch Reconstruction for 2D-3D Multimodal Industrial Anomaly Detection (viability: 7): https://sciencetostartup.com/paper/cross-modal-mapping-and-dual-branch-reconstruction-for-2d-3d-multimodal-industrial-anomaly-detection - Develop a lightweight and modality-flexible framework for robust industrial anomaly detection using 2D and 3D data.
- Selecting Offline Reinforcement Learning Algorithms for Stochastic Network Control (viability: 5): https://sciencetostartup.com/paper/selecting-offline-reinforcement-learning-algorithms-for-stochastic-network-control - Develop a robust offline RL algorithm toolkit for reliable AI-driven network control in wireless systems.
- BD-Merging: Bias-Aware Dynamic Model Merging with Evidence-Guided Contrastive Learning (viability: 5): https://sciencetostartup.com/paper/bd-merging-bias-aware-dynamic-model-merging-with-evidence-guided-contrastive-learning - Develop a bias-aware dynamic model merging tool to improve multi-task learning under distribution shifts.
- Rethinking Role-Playing Evaluation: Anonymous Benchmarking and a Systematic Study of Personality Effects (viability: 5): https://sciencetostartup.com/paper/rethinking-role-playing-evaluation-anonymous-benchmarking-and-a-systematic-study-of-personality-effects - Develop a scalable, personality-enhanced framework for robust and unbiased Role-Playing Agents.
- From Threat Intelligence to Firewall Rules: Semantic Relations in Hybrid AI Agent and Expert System Architectures (viability: 3): https://sciencetostartup.com/paper/from-threat-intelligence-to-firewall-rules-semantic-relations-in-hybrid-ai-agent-and-expert-system-architectures - Develop a neuro-symbolic multi-agent system for automated firewall rule generation from CTI reports.
- PatchDecomp: Interpretable Patch-Based Time Series Forecasting (viability: 5): https://sciencetostartup.com/paper/patchdecomp-interpretable-patch-based-time-series-forecasting - PatchDecomp offers interpretable time series forecasting by attributing predictions to individual data patches.
- IROSA: Interactive Robot Skill Adaptation using Natural Language (viability: 7): https://sciencetostartup.com/paper/irosa-interactive-robot-skill-adaptation-using-natural-language - IROSA offers interactive robot skill adaptation using natural language for industrial tasks, with a focus on safety and interpretability.
- A novel network for classification of cuneiform tablet metadata (viability: 7): https://sciencetostartup.com/paper/a-novel-network-for-classification-of-cuneiform-tablet-metadata - An AI model that outperforms current solutions in classifying cuneiform tablet metadata, soon to be open-source.
- CzechTopic: A Benchmark for Zero-Shot Topic Localization in Historical Czech Documents (viability: 4): https://sciencetostartup.com/paper/czechtopic-a-benchmark-for-zero-shot-topic-localization-in-historical-czech-documents - Develop a tool for zero-shot topic localization in historical Czech documents using a new benchmark dataset.
- On the Suitability of LLM-Driven Agents for Dark Pattern Audits (viability: 4): https://sciencetostartup.com/paper/on-the-suitability-of-llm-driven-agents-for-dark-pattern-audits - Develop LLM-driven agents for auditing dark patterns in website interfaces to ensure compliance with CCPA-related data rights.
- Joint Hardware-Workload Co-Optimization for In-Memory Computing Accelerators (viability: 5): https://sciencetostartup.com/paper/joint-hardware-workload-co-optimization-for-in-memory-computing-accelerators - Develop a co-optimization framework for designing generalized in-memory computing accelerators that efficiently support multiple neural network workloads.
- Structure-Aware Distributed Backdoor Attacks in Federated Learning (viability: 2): https://sciencetostartup.com/paper/structure-aware-distributed-backdoor-attacks-in-federated-learning - Explore structure-aware defenses against backdoor attacks in federated learning.
- From Narrow to Panoramic Vision: Attention-Guided Cold-Start Reshapes Multimodal Reasoning (viability: 7): https://sciencetostartup.com/paper/from-narrow-to-panoramic-vision-attention-guided-cold-start-reshapes-multimodal-reasoning - Develop a product facilitating enhanced visual attention in multimodal AI systems for superior performance.
- In-Context Environments Induce Evaluation-Awareness in Language Models (viability: 6): https://sciencetostartup.com/paper/in-context-environments-induce-evaluation-awareness-in-language-models - Develop a system to enhance evaluation reliability of language models against adversarially optimized prompts.
- SWE-CI: Evaluating Agent Capabilities in Maintaining Codebases via Continuous Integration (viability: 6): https://sciencetostartup.com/paper/swe-ci-evaluating-agent-capabilities-in-maintaining-codebases-via-continuous-integration - Build a Continuous Integration benchmark tool evaluating AI agents for long-term code maintainability.
- Fairness Begins with State: Purifying Latent Preferences for Hierarchical Reinforcement Learning in Interactive Recommendation (viability: 4): https://sciencetostartup.com/paper/fairness-begins-with-state-purifying-latent-preferences-for-hierarchical-reinforcement-learning-in-interactive-recommend - Develop a fairness-oriented interactive recommendation system leveraging denoised latent preferences and hierarchical decision-making.
- Pretrained Vision-Language-Action Models are Surprisingly Resistant to Forgetting in Continual Learning (viability: 7): https://sciencetostartup.com/paper/pretrained-vision-language-action-models-are-surprisingly-resistant-to-forgetting-in-continual-learning - Vision-Language-Action models that resist forgetting offer transformative solutions for continual robotic learning.
- Separators in Enhancing Autoregressive Pretraining for Vision Mamba (viability: 4): https://sciencetostartup.com/paper/separators-in-enhancing-autoregressive-pretraining-for-vision-mamba - Enhance computer vision models by leveraging long-range dependencies in autoregressive pretraining with Vision Mamba.
- Relational In-Context Learning via Synthetic Pre-training with Structural Prior (viability: 6): https://sciencetostartup.com/paper/relational-in-context-learning-via-synthetic-pre-training-with-structural-prior - Develop a relational AI model that adapts instantly to new databases using synthetic data pre-training, boosting few-shot learning efficiency.
- Zero-Knowledge Proof (ZKP) Authentication for Offline CBDC Payment System Using IoT Devices (viability: 3): https://sciencetostartup.com/paper/zero-knowledge-proof-zkp-authentication-for-offline-cbdc-payment-system-using-iot-devices - Develop a zero-knowledge proof-based offline CBDC payment system for IoT devices to enable privacy-preserving transactions without internet.
- A Rubric-Supervised Critic from Sparse Real-World Outcomes (viability: 6): https://sciencetostartup.com/paper/a-rubric-supervised-critic-from-sparse-real-world-outcomes - Develop a rubric-supervised critic to enhance real-world coding agent performance through sparse interaction data.
- When and Where to Reset Matters for Long-Term Test-Time Adaptation (viability: 5): https://sciencetostartup.com/paper/when-and-where-to-reset-matters-for-long-term-test-time-adaptation - Develop adaptive reset strategies to prevent model collapse during long-term test-time adaptation.
- T2S-Bench & Structure-of-Thought: Benchmarking and Prompting Comprehensive Text-to-Structure Reasoning (viability: 6): https://sciencetostartup.com/paper/t2s-bench-structure-of-thought-benchmarking-and-prompting-comprehensive-text-to-structure-reasoning - T2S-Bench enhances AI text-processing by providing a benchmark and technique for text-to-structure reasoning models.
- Specification-Driven Generation and Evaluation of Discrete-Event World Models via the DEVS Formalism (viability: 2): https://sciencetostartup.com/paper/specification-driven-generation-and-evaluation-of-discrete-event-world-models-via-the-devs-formalism - Develop discrete-event world models using DEVS formalism for enhanced reliability and flexibility in agentic systems.
- DisenReason: Behavior Disentanglement and Latent Reasoning for Shared-Account Sequential Recommendation (viability: 5): https://sciencetostartup.com/paper/disenreason-behavior-disentanglement-and-latent-reasoning-for-shared-account-sequential-recommendation - DisenReason enhances recommendation accuracy by disentangling latent user behaviors in shared accounts using an innovative two-stage reasoning method.
- LifeBench: A Benchmark for Long-Horizon Multi-Source Memory (viability: 5): https://sciencetostartup.com/paper/lifebench-a-benchmark-for-long-horizon-multi-source-memory - A new benchmark for AI agents to leverage long-horizon memory across diverse contexts, with dataset and code available for enhancing memory systems.
- MACC: Multi-Agent Collaborative Competition for Scientific Exploration (viability: 3): https://sciencetostartup.com/paper/macc-multi-agent-collaborative-competition-for-scientific-exploration - Introducing MACC, an institutional architecture for multi-agent scientific exploration to enhance transparency and efficiency.
- Towards Effective Orchestration of AI x DB Workloads (viability: 5): https://sciencetostartup.com/paper/towards-effective-orchestration-of-ai-x-db-workloads - Integrating AI directly into database engines to optimize data-centric analytics while addressing query and execution challenges.
- Not All Candidates are Created Equal: A Heterogeneity-Aware Approach to Pre-ranking in Recommender Systems (viability: 8): https://sciencetostartup.com/paper/not-all-candidates-are-created-equal-a-heterogeneity-aware-approach-to-pre-ranking-in-recommender-systems - Develop a heterogeneity-aware pre-ranking system for recommender systems to enhance efficiency and accuracy without additional computational cost.
- IntroductionDMD-augmented Unpaired Neural Schrödinger Bridge for Ultra-Low Field MRI Enhancement (viability: 5): https://sciencetostartup.com/paper/introductiondmd-augmented-unpaired-neural-schr-dinger-bridge-for-ultra-low-field-mri-enhancement - AI enhances ultra-low field MRI images to improve medical accessibility without paired data requirements.
- Cognition to Control - Multi-Agent Learning for Human-Humanoid Collaborative Transport (viability: 3): https://sciencetostartup.com/paper/cognition-to-control-multi-agent-learning-for-human-humanoid-collaborative-transport - Develop a multi-agent learning system for improved human-robot collaborative transport.
- AgentSelect: Benchmark for Narrative Query-to-Agent Recommendation (viability: 6): https://sciencetostartup.com/paper/agentselect-benchmark-for-narrative-query-to-agent-recommendation - AgentSelect is a benchmark tool that facilitates query-to-agent recommendations by converting varied evaluation data into unified interaction records.
- Learning Approximate Nash Equilibria in Cooperative Multi-Agent Reinforcement Learning via Mean-Field Subsampling (viability: 3): https://sciencetostartup.com/paper/learning-approximate-nash-equilibria-in-cooperative-multi-agent-reinforcement-learning-via-mean-field-subsampling - Develop an algorithm for approximate Nash Equilibria in cooperative multi-agent systems with communication constraints.
- Agentic Peer-to-Peer Networks: From Content Distribution to Capability and Action Sharing (viability: 3): https://sciencetostartup.com/paper/agentic-peer-to-peer-networks-from-content-distribution-to-capability-and-action-sharing - A network architecture for AI agents to share capabilities and actions securely in peer-to-peer settings.
- Confidence-Calibrated Small-Large Language Model Collaboration for Cost-Efficient Reasoning (viability: 7): https://sciencetostartup.com/paper/confidence-calibrated-small-large-language-model-collaboration-for-cost-efficient-reasoning - COREA optimizes cost-efficient reasoning by cascading small and large language models using a confidence-calibrated system.
- Interaction-Aware Whole-Body Control for Compliant Object Transport (viability: 5): https://sciencetostartup.com/paper/interaction-aware-whole-body-control-for-compliant-object-transport - AI enhances humanoid robots for better compliant object transport through bio-inspired whole-body control.
- JANUS: Structured Bidirectional Generation for Guaranteed Constraints and Analytical Uncertainty (viability: 7): https://sciencetostartup.com/paper/janus-structured-bidirectional-generation-for-guaranteed-constraints-and-analytical-uncertainty - Develop a framework for high-fidelity, constraint-satisfying synthetic data generation with fast uncertainty estimation.
- RAGNav: A Retrieval-Augmented Topological Reasoning Framework for Multi-Goal Visual-Language Navigation (viability: 6): https://sciencetostartup.com/paper/ragnav-a-retrieval-augmented-topological-reasoning-framework-for-multi-goal-visual-language-navigation - Develop a framework using a dual-basis memory system for improved multi-goal vision-language navigation.
- HALyPO: Heterogeneous-Agent Lyapunov Policy Optimization for Human-Robot Collaboration (viability: 3): https://sciencetostartup.com/paper/halypo-heterogeneous-agent-lyapunov-policy-optimization-for-human-robot-collaboration - Develop a stability-focused reinforcement learning framework for enhancing human-robot collaboration under diverse conditions.
- PROSPECT: Unified Streaming Vision-Language Navigation via Semantic--Spatial Fusion and Latent Predictive Representation (viability: 4): https://sciencetostartup.com/paper/prospect-unified-streaming-vision-language-navigation-via-semantic-spatial-fusion-and-latent-predictive-representation - Build a streaming vision-language navigation agent leveraging semantic-spatial fusion for enhanced zero-shot performance.
- Understanding Parents' Desires in Moderating Children's Interactions with GenAI Chatbots through LLM-Generated Probes (viability: 5): https://sciencetostartup.com/paper/understanding-parents-desires-in-moderating-children-s-interactions-with-genai-chatbots-through-llm-generated-probes - Develop GenAI chatbot parental control tools featuring fine-grained and personalized moderation of children's interactions.
- Why Do Unlearnable Examples Work: A Novel Perspective of Mutual Information (viability: 4): https://sciencetostartup.com/paper/why-do-unlearnable-examples-work-a-novel-perspective-of-mutual-information - Develop a security tool that uses mutual information reduction to create unlearnable examples, enhancing data privacy in AI systems.
- Order Is Not Layout: Order-to-Space Bias in Image Generation (viability: 3): https://sciencetostartup.com/paper/order-is-not-layout-order-to-space-bias-in-image-generation - Identifies and mitigates order-to-space bias in image generation models, but lacks practical application potential.
- MPFlow: Multi-modal Posterior-Guided Flow Matching for Zero-Shot MRI Reconstruction (viability: 7): https://sciencetostartup.com/paper/mpflow-multi-modal-posterior-guided-flow-matching-for-zero-shot-mri-reconstruction - Leverage MPFlow for zero-shot multi-modal MRI reconstruction with enhanced anatomical fidelity using cross-modal guidance.
- Large-Language-Model-Guided State Estimation for Partially Observable Task and Motion Planning (viability: 7): https://sciencetostartup.com/paper/large-language-model-guided-state-estimation-for-partially-observable-task-and-motion-planning - CoCo-TAMP uses large language models to enhance robotic planning in environments with partial observability by incorporating common-sense knowledge for efficient task execution.
- UrbanHuRo: A Two-Layer Human-Robot Collaboration Framework for the Joint Optimization of Heterogeneous Urban Services (viability: 6): https://sciencetostartup.com/paper/urbanhuro-a-two-layer-human-robot-collaboration-framework-for-the-joint-optimization-of-heterogeneous-urban-services - A human-robot collaboration framework optimizing urban service interactions for enhanced efficiency and resource utilization in smart cities.
- Error as Signal: Stiffness-Aware Diffusion Sampling via Embedded Runge-Kutta Guidance (viability: 6): https://sciencetostartup.com/paper/error-as-signal-stiffness-aware-diffusion-sampling-via-embedded-runge-kutta-guidance - Leverage solver-induced errors to enhance diffusion model sampling with Embedded Runge-Kutta Guidance.
- AI4S-SDS: A Neuro-Symbolic Solvent Design System via Sparse MCTS and Differentiable Physics Alignment (viability: 7): https://sciencetostartup.com/paper/ai4s-sds-a-neuro-symbolic-solvent-design-system-via-sparse-mcts-and-differentiable-physics-alignment - Develop a novel AI framework integrating neuro-symbolic methods for automated chemical formulation design, enhancing discovery and validation of new materials.
- EvoPrune: Early-Stage Visual Token Pruning for Efficient MLLMs (viability: 7): https://sciencetostartup.com/paper/evoprune-early-stage-visual-token-pruning-for-efficient-mllms - EvoPrune enhances the efficiency of MLLMs by pruning visual tokens early in the encoding process, improving latency without significant performance loss.
- MAGE: Meta-Reinforcement Learning for Language Agents toward Strategic Exploration and Exploitation (viability: 6): https://sciencetostartup.com/paper/mage-meta-reinforcement-learning-for-language-agents-toward-strategic-exploration-and-exploitation - MAGE enhances LLM agents with meta-RL for strategic exploration and exploitation in multi-agent environments.
- MIND: Unified Inquiry and Diagnosis RL with Criteria Grounded Clinical Supports for Psychiatric Consultation (viability: 5): https://sciencetostartup.com/paper/mind-unified-inquiry-and-diagnosis-rl-with-criteria-grounded-clinical-supports-for-psychiatric-consultation - Revolutionize psychiatric consultations with MIND, a reinforcement learning framework enhancing diagnostic accuracy and interaction quality for clinical dialogue systems.
- Local Shapley: Model-Induced Locality and Optimal Reuse in Data Valuation (viability: 4): https://sciencetostartup.com/paper/local-shapley-model-induced-locality-and-optimal-reuse-in-data-valuation - Efficiently compute Shapley values locally by leveraging model-induced locality with LSMR for faster data valuation.
- Graph Negative Feedback Bias Correction Framework for Adaptive Heterophily Modeling (viability: 3): https://sciencetostartup.com/paper/graph-negative-feedback-bias-correction-framework-for-adaptive-heterophily-modeling - Develop a framework to correct negative feedback bias in Graph Neural Networks for better performance on heterophilic graphs.
- InEdit-Bench: Benchmarking Intermediate Logical Pathways for Intelligent Image Editing Models (viability: 5): https://sciencetostartup.com/paper/inedit-bench-benchmarking-intermediate-logical-pathways-for-intelligent-image-editing-models - InEdit-Bench is the first benchmark to evaluate intermediate logical pathways in image editing models, promoting advanced procedural understanding.
- Mozi: Governed Autonomy for Drug Discovery LLM Agents (viability: 7): https://sciencetostartup.com/paper/mozi-governed-autonomy-for-drug-discovery-llm-agents - Mozi offers a governed AI agent architecture for reliable and safe drug discovery workflows, ensuring high scientific validity in computational biology applications.
- Field imaging framework for morphological characterization of aggregates with computer vision: Algorithms and applications (viability: 7): https://sciencetostartup.com/paper/field-imaging-framework-for-morphological-characterization-of-aggregates-with-computer-vision-algorithms-and-application - Automated computer vision solution for field imaging and morphological analysis of construction aggregates.
- Bridging Pedagogy and Play: Introducing a Language Mapping Interface for Human-AI Co-Creation in Educational Game Design (viability: 4): https://sciencetostartup.com/paper/bridging-pedagogy-and-play-introducing-a-language-mapping-interface-for-human-ai-co-creation-in-educational-game-design - A web tool using controlled natural language to help non-experts design educational games with LLM assistance.
- Image-based Prompt Injection: Hijacking Multimodal LLMs through Visually Embedded Adversarial Instructions (viability: 4): https://sciencetostartup.com/paper/image-based-prompt-injection-hijacking-multimodal-llms-through-visually-embedded-adversarial-instructions - Develop an Image-based Prompt Injection tool to expose and address vulnerabilities in multimodal language models.
- Goal-Driven Risk Assessment for LLM-Powered Systems: A Healthcare Case Study (viability: 4): https://sciencetostartup.com/paper/goal-driven-risk-assessment-for-llm-powered-systems-a-healthcare-case-study - Structured risk assessment tool for enhancing cyber security in LLM-based healthcare systems.
- Social Norm Reasoning in Multimodal Language Models: An Evaluation (viability: 5): https://sciencetostartup.com/paper/social-norm-reasoning-in-multimodal-language-models-an-evaluation - Developing AI systems for enhanced social norm reasoning in Multi-Agent Systems using Multimodal Language Models.
- Belief-Sim: Towards Belief-Driven Simulation of Demographic Misinformation Susceptibility (viability: 5): https://sciencetostartup.com/paper/belief-sim-towards-belief-driven-simulation-of-demographic-misinformation-susceptibility - BeliefSim leverages demographic beliefs to predict susceptibility to misinformation with high accuracy using a simulation framework.
- Build, Judge, Optimize: A Blueprint for Continuous Improvement of Multi-Agent Consumer Assistants (viability: 3): https://sciencetostartup.com/paper/build-judge-optimize-a-blueprint-for-continuous-improvement-of-multi-agent-consumer-assistants - Blueprint for optimizing grocery shopping conversational agents with an advanced evaluation rubric and prompt-optimization strategies.
- Molt Dynamics: Emergent Social Phenomena in Autonomous AI Agent Populations (viability: 2): https://sciencetostartup.com/paper/molt-dynamics-emergent-social-phenomena-in-autonomous-ai-agent-populations - MoltBook provides insights into emergent behaviors in large-scale autonomous AI populations, yet lacks commercial pathways.
- Tucano 2 Cool: Better Open Source LLMs for Portuguese (viability: 7): https://sciencetostartup.com/paper/tucano-2-cool-better-open-source-llms-for-portuguese - Tucano 2 provides state-of-the-art open-source language models for the Portuguese NLP community with comprehensive datasets and tools.
- RAG-X: Systematic Diagnosis of Retrieval-Augmented Generation for Medical Question Answering (viability: 7): https://sciencetostartup.com/paper/rag-x-systematic-diagnosis-of-retrieval-augmented-generation-for-medical-question-answering - RAG-X enhances medical QA by diagnosing retrieval and generation errors separately for improved clinical safety.
- SafeCRS: Personalized Safety Alignment for LLM-Based Conversational Recommender Systems (viability: 8): https://sciencetostartup.com/paper/safecrs-personalized-safety-alignment-for-llm-based-conversational-recommender-systems - SafeCRS offers a personalized safety alignment framework for conversational recommendation systems, optimizing user-specific safety and recommendation quality.
- Role-Aware Conditional Inference for Spatiotemporal Ecosystem Carbon Flux Prediction (viability: 6): https://sciencetostartup.com/paper/role-aware-conditional-inference-for-spatiotemporal-ecosystem-carbon-flux-prediction - Develop models to improve spatial and temporal prediction accuracy of ecosystem carbon fluxes using Role-Aware Conditional Inference.
- Directional Neural Collapse Explains Few-Shot Transfer in Self-Supervised Learning (viability: 5): https://sciencetostartup.com/paper/directional-neural-collapse-explains-few-shot-transfer-in-self-supervised-learning - Analyze and enhance directional neural collapse for improving few-shot learning in SSL.
- mlx-snn: Spiking Neural Networks on Apple Silicon via MLX (viability: 7): https://sciencetostartup.com/paper/mlx-snn-spiking-neural-networks-on-apple-silicon-via-mlx - mlx-snn provides a native spiking neural network library for Apple Silicon, enhancing efficiency and speed of SNN research and applications.
- Multi-Agent Influence Diagrams to Hybrid Threat Modeling (viability: 3): https://sciencetostartup.com/paper/multi-agent-influence-diagrams-to-hybrid-threat-modeling - Develop influence diagram frameworks to enhance understanding of counter-hybrid threat measures' effectiveness against adversarial tactics below military conflict thresholds.
- Test-Time Meta-Adaptation with Self-Synthesis (viability: 3): https://sciencetostartup.com/paper/test-time-meta-adaptation-with-self-synthesis - A meta-learning framework for LLMs to self-adapt by synthesizing training data at test time.
- MMAI Gym for Science: Training Liquid Foundation Models for Drug Discovery (viability: 8): https://sciencetostartup.com/paper/mmai-gym-for-science-training-liquid-foundation-models-for-drug-discovery - MMAI Gym for Science provides a tailored platform to efficiently train and deploy Liquid Foundation Models for high-performance drug discovery tasks.
- The Controllability Trap: A Governance Framework for Military AI Agents (viability: 3): https://sciencetostartup.com/paper/the-controllability-trap-a-governance-framework-for-military-ai-agents - Develop a governance framework for AI agent control in military settings.
- Baseline Performance of AI Tools in Classifying Cognitive Demand of Mathematical Tasks (viability: 2): https://sciencetostartup.com/paper/baseline-performance-of-ai-tools-in-classifying-cognitive-demand-of-mathematical-tasks - Improving AI tools to better classify cognitive demand in mathematical tasks for educational use.
- Raising Bars, Not Parameters: LilMoo Compact Language Model for Hindi (viability: 6): https://sciencetostartup.com/paper/raising-bars-not-parameters-lilmoo-compact-language-model-for-hindi - LilMoo offers a compact, high-performance language model tailored specifically for Hindi, enabling more equitable NLP innovations.
- PhyPrompt: RL-based Prompt Refinement for Physically Plausible Text-to-Video Generation (viability: 6): https://sciencetostartup.com/paper/phyprompt-rl-based-prompt-refinement-for-physically-plausible-text-to-video-generation - PhyPrompt enhances text-to-video generation with physics-based prompt refinement using reinforcement learning.
- Phys4D: Fine-Grained Physics-Consistent 4D Modeling from Video Diffusion (viability: 6): https://sciencetostartup.com/paper/phys4d-fine-grained-physics-consistent-4d-modeling-from-video-diffusion - Build fine-grained physics-consistent 4D models from video with the Phys4D pipeline.
- Optimal trajectory-guided stochastic co-optimization for e-fuel system design and real-time operation (viability: 6): https://sciencetostartup.com/paper/optimal-trajectory-guided-stochastic-co-optimization-for-e-fuel-system-design-and-real-time-operation - MasCOR provides a machine-learning-assisted framework for efficient e-fuel system design and operation, optimizing based on real-time data and renewable trends.
- Beyond Pixel Histories: World Models with Persistent 3D State (viability: 6): https://sciencetostartup.com/paper/beyond-pixel-histories-world-models-with-persistent-3d-state - PERSIST revolutionizes interactive world simulation with persistent 3D states and spatial consistency for extended rollouts.
- When Shallow Wins: Silent Failures and the Depth-Accuracy Paradox in Latent Reasoning (viability: 5): https://sciencetostartup.com/paper/when-shallow-wins-silent-failures-and-the-depth-accuracy-paradox-in-latent-reasoning - Develop robust mathematical reasoning models for educational and decision support systems employing novel faithfulness metrics.
- Graph Hopfield Networks: Energy-Based Node Classification with Associative Memory (viability: 4): https://sciencetostartup.com/paper/graph-hopfield-networks-energy-based-node-classification-with-associative-memory - Develop energy-based Graph Hopfield Networks for improved node classification in citation and co-purchase graphs.
- Asymmetric Goal Drift in Coding Agents Under Value Conflict (viability: 6): https://sciencetostartup.com/paper/asymmetric-goal-drift-in-coding-agents-under-value-conflict - Develop a coding agent framework that enhances system prompt fidelity by monitoring value-aligned goal drift in complex environments.
- How to Peel with a Knife: Aligning Fine-Grained Manipulation with Human Preference (viability: 8): https://sciencetostartup.com/paper/how-to-peel-with-a-knife-aligning-fine-grained-manipulation-with-human-preference - A robotic system that learns to peel fruits and vegetables with human-like precision and preference alignment.
- Tether: Autonomous Functional Play with Correspondence-Driven Trajectory Warping (viability: 5): https://sciencetostartup.com/paper/tether-autonomous-functional-play-with-correspondence-driven-trajectory-warping - Tether enables robots to autonomously learn multi-task operations through correspondence-driven trajectory warping, starting from minimal demonstrations.
- Inherited Goal Drift: Contextual Pressure Can Undermine Agentic Goals (viability: 2): https://sciencetostartup.com/paper/inherited-goal-drift-contextual-pressure-can-undermine-agentic-goals - Research on goal drift in language model agents highlights the need for improved post-training techniques to maintain objective adherence.
- Valet: A Standardized Testbed of Traditional Imperfect-Information Card Games (viability: 5): https://sciencetostartup.com/paper/valet-a-standardized-testbed-of-traditional-imperfect-information-card-games - Valet is a standardized testbed for benchmarking AI algorithms on traditional imperfect-information card games.
- Density-Guided Response Optimization: Community-Grounded Alignment via Implicit Acceptance Signals (viability: 5): https://sciencetostartup.com/paper/density-guided-response-optimization-community-grounded-alignment-via-implicit-acceptance-signals - Align language models with community norms using density-guided response optimization without explicit preference labeling.
- UniG2U-Bench: Do Unified Models Advance Multimodal Understanding? (viability: 4): https://sciencetostartup.com/paper/unig2u-bench-do-unified-models-advance-multimodal-understanding - UniG2U-Bench evaluates the impact of generative capabilities on multimodal understanding through a comprehensive new benchmark.
- AI-for-Science Low-code Platform with Bayesian Adversarial Multi-Agent Framework (viability: 7): https://sciencetostartup.com/paper/ai-for-science-low-code-platform-with-bayesian-adversarial-multi-agent-framework - A low-code platform using Bayesian adversarial multi-agent framework to automate and enhance scientific code generation with LLMs.
- SynthCharge: An Electric Vehicle Routing Instance Generator with Feasibility Screening to Enable Learning-Based Optimization and Benchmarking (viability: 5): https://sciencetostartup.com/paper/synthcharge-an-electric-vehicle-routing-instance-generator-with-feasibility-screening-to-enable-learning-based-optimizat - SynthCharge is a parametric instance generator for electric vehicle routing, enhancing learning-based optimization with dynamic feasibility screening.
- Stabilized Adaptive Loss and Residual-Based Collocation for Physics-Informed Neural Networks (viability: 2): https://sciencetostartup.com/paper/stabilized-adaptive-loss-and-residual-based-collocation-for-physics-informed-neural-networks - Enhance Physics-Informed Neural Networks to solve partial differential equations with high stiffness and shock-dominated dynamics by reducing relative L2 error significantly.
- NeuroSkill(tm): Proactive Real-Time Agentic System Capable of Modeling Human State of Mind (viability: 5): https://sciencetostartup.com/paper/neuroskill-tm-proactive-real-time-agentic-system-capable-of-modeling-human-state-of-mind - Develop an offline real-time agentic system for modeling human state of mind using BCI data and text embeddings.
- Understanding and Mitigating Dataset Corruption in LLM Steering (viability: 5): https://sciencetostartup.com/paper/understanding-and-mitigating-dataset-corruption-in-llm-steering - Develop a robust method to safeguard LLM steering against dataset corruption.
- No Memorization, No Detection: Output Distribution-Based Contamination Detection in Small Language Models (viability: 5): https://sciencetostartup.com/paper/no-memorization-no-detection-output-distribution-based-contamination-detection-in-small-language-models - Detect data contamination in small language models by analyzing output distribution shape to mitigate memorization risks.
- Chain of World: World Model Thinking in Latent Motion (viability: 5): https://sciencetostartup.com/paper/chain-of-world-world-model-thinking-in-latent-motion - CoWVLA unifies world-model temporal reasoning with latent motion representation for efficient visuomotor learning in robotics.
- Expectation and Acoustic Neural Network Representations Enhance Music Identification from Brain Activity (viability: 5): https://sciencetostartup.com/paper/expectation-and-acoustic-neural-network-representations-enhance-music-identification-from-brain-activity - Enhancing music recognition from EEG via neural network-based acoustic and expectation representations.
- Type-Aware Retrieval-Augmented Generation with Dependency Closure for Solver-Executable Industrial Optimization Modeling (viability: 10): https://sciencetostartup.com/paper/type-aware-retrieval-augmented-generation-with-dependency-closure-for-solver-executable-industrial-optimization-modeling - Automate industrial optimization modeling using a type-aware retrieval-augmented generation system that ensures solver-executable code.
- Neuro-Symbolic Artificial Intelligence: A Task-Directed Survey in the Black-Box Models Era (viability: 3): https://sciencetostartup.com/paper/neuro-symbolic-artificial-intelligence-a-task-directed-survey-in-the-black-box-models-era - Enhance AI explainability by integrating symbolic systems with neural networks for better reasoning in NLP and Vision.
- FEAST: Retrieval-Augmented Multi-Hierarchical Food Classification for the FoodEx2 System (viability: 6): https://sciencetostartup.com/paper/feast-retrieval-augmented-multi-hierarchical-food-classification-for-the-foodex2-system - FEAST enhances food classification for the FoodEx2 system by significantly improving rare class accuracy with a novel retrieval-augmented approach.
- Saarthi for AGI: Towards Domain-Specific General Intelligence for Formal Verification (viability: 7): https://sciencetostartup.com/paper/saarthi-for-agi-towards-domain-specific-general-intelligence-for-formal-verification - Saarthi is an agentic AI framework enhancing formal verification with multi-agent collaboration and improved accuracy using structured rulebooks and RAG techniques.
- Conditioned Activation Transport for T2I Safety Steering (viability: 7): https://sciencetostartup.com/paper/conditioned-activation-transport-for-t2i-safety-steering - Conditioned Activation Transport improves Text-to-Image model safety by minimizing interference with benign content using nonlinear transport maps.
- An Investigation Into Various Approaches For Bengali Long-Form Speech Transcription and Bengali Speaker Diarization (viability: 7): https://sciencetostartup.com/paper/an-investigation-into-various-approaches-for-bengali-long-form-speech-transcription-and-bengali-speaker-diarization - A multi-stage framework for accurate Bengali long-form transcription and speaker diarization.
- Information Routing in Atomistic Foundation Models: How Equivariance Creates Linearly Disentangled Representations (viability: 1): https://sciencetostartup.com/paper/information-routing-in-atomistic-foundation-models-how-equivariance-creates-linearly-disentangled-representations - Develop high sample-efficiency models for disentangled representations in atomistic systems.
- Agentic AI-based Coverage Closure for Formal Verification (viability: 7): https://sciencetostartup.com/paper/agentic-ai-based-coverage-closure-for-formal-verification - Agentic AI-driven framework automates IC verification coverage closure, boosting efficiency and productivity.
- Channel-Adaptive Edge AI: Maximizing Inference Throughput by Adapting Computational Complexity to Channel States (viability: 5): https://sciencetostartup.com/paper/channel-adaptive-edge-ai-maximizing-inference-throughput-by-adapting-computational-complexity-to-channel-states - Develop a channel-adaptive AI algorithm to maximize edge inference throughput by adjusting computational complexity based on channel conditions.
- Geometry-Guided Reinforcement Learning for Multi-view Consistent 3D Scene Editing (viability: 7): https://sciencetostartup.com/paper/geometry-guided-reinforcement-learning-for-multi-view-consistent-3d-scene-editing - Single-pass RL framework for multi-view consistent 3D scene editing using VGGT priors.
- APRES: An Agentic Paper Revision and Evaluation System (viability: 7): https://sciencetostartup.com/paper/apres-an-agentic-paper-revision-and-evaluation-system - APRES is an AI-powered tool to enhance scientific paper quality and impact by revising manuscripts using LLMs based on predictive evaluations.
- How to Model AI Agents as Personas?: Applying the Persona Ecosystem Playground to 41,300 Posts on Moltbook for Behavioral Insights (viability: 3): https://sciencetostartup.com/paper/how-to-model-ai-agents-as-personas-applying-the-persona-ecosystem-playground-to-41-300-posts-on-moltbook-for-behavioral- - Develop behavioral personas for AI agents using clustering and retrieval-augmented generation on social media data.
- Joint Training Across Multiple Activation Sparsity Regimes (viability: 2): https://sciencetostartup.com/paper/joint-training-across-multiple-activation-sparsity-regimes - Develop a training strategy to enhance neural network generalization across dense and sparse activation regimes.
- AI Space Physics: Constitutive boundary semantics for open AI institutions (viability: 2): https://sciencetostartup.com/paper/ai-space-physics-constitutive-boundary-semantics-for-open-ai-institutions - Introducing AI Space Physics for governance of self-expanding AI institutions through a new constitutive semantics framework.
- Beyond Task Completion: Revealing Corrupt Success in LLM Agents through Procedure-Aware Evaluation (viability: 5): https://sciencetostartup.com/paper/beyond-task-completion-revealing-corrupt-success-in-llm-agents-through-procedure-aware-evaluation - Enhance LLM-based agent evaluations with a framework focusing on procedure-aware metrics.
- From Complex Dynamics to DynFormer: Rethinking Transformers for PDEs (viability: 6): https://sciencetostartup.com/paper/from-complex-dynamics-to-dynformer-rethinking-transformers-for-pdes - "DynFormer optimizes PDE solving with transformer-based efficiency, reducing error and computation costs."
- Multi-Scale Adaptive Neighborhood Awareness Transformer For Graph Fraud Detection (viability: 7): https://sciencetostartup.com/paper/multi-scale-adaptive-neighborhood-awareness-transformer-for-graph-fraud-detection - Transform fraud detection in graph data with a Multi-Scale Adaptive Transformer offering enhanced global modeling ability.
- MoECLIP: Patch-Specialized Experts for Zero-shot Anomaly Detection (viability: 7): https://sciencetostartup.com/paper/moeclip-patch-specialized-experts-for-zero-shot-anomaly-detection - Innovate zero-shot anomaly detection using patch-specialized experts with MoECLIP to beat state-of-the-art benchmarks.
- Why Adam Can Beat SGD: Second-Moment Normalization Yields Sharper Tails (viability: 4): https://sciencetostartup.com/paper/why-adam-can-beat-sgd-second-moment-normalization-yields-sharper-tails - Discover why Adam optimization outperforms SGD using novel theoretical insights and analysis.
- Odin: Multi-Signal Graph Intelligence for Autonomous Discovery in Knowledge Graphs (viability: 9): https://sciencetostartup.com/paper/odin-multi-signal-graph-intelligence-for-autonomous-discovery-in-knowledge-graphs - Odin offers a cutting-edge graph intelligence engine for autonomous pattern discovery in knowledge graphs, transforming exploratory analysis in regulated industries.
- Compact Prompting in Instruction-tuned LLMs for Joint Argumentative Component Detection (viability: 5): https://sciencetostartup.com/paper/compact-prompting-in-instruction-tuned-llms-for-joint-argumentative-component-detection - Transform argumentative component detection in texts with instruction-tuned LLMs for improved automated analysis.
- Proactive Guiding Strategy for Item-side Fairness in Interactive Recommendation (viability: 3): https://sciencetostartup.com/paper/proactive-guiding-strategy-for-item-side-fairness-in-interactive-recommendation - Develop an interactive recommendation system that ensures item-side fairness by using hierarchical reinforcement learning to guide user preferences toward long-tail items.
- On the Expressive Power of Transformers for Maxout Networks and Continuous Piecewise Linear Functions (viability: 3): https://sciencetostartup.com/paper/on-the-expressive-power-of-transformers-for-maxout-networks-and-continuous-piecewise-linear-functions - This research explores the theoretical expressive power of Transformer networks in relation to maxout networks and continuous piecewise linear functions.
- Beyond Factual Correctness: Mitigating Preference-Inconsistent Explanations in Explainable Recommendation (viability: 5): https://sciencetostartup.com/paper/beyond-factual-correctness-mitigating-preference-inconsistent-explanations-in-explainable-recommendation - PURE enhances explainable recommenders by aligning explanations with user preferences to reduce preference-inconsistent explanations.
- RAPO: Expanding Exploration for LLM Agents via Retrieval-Augmented Policy Optimization (viability: 5): https://sciencetostartup.com/paper/rapo-expanding-exploration-for-llm-agents-via-retrieval-augmented-policy-optimization - Develop Retrieval-Augmented Policy Optimization for enhanced exploration in LLM-based agentic reasoning.
- TinyIceNet: Low-Power SAR Sea Ice Segmentation for On-Board FPGA Inference (viability: 7): https://sciencetostartup.com/paper/tinyicenet-low-power-sar-sea-ice-segmentation-for-on-board-fpga-inference - Develop an ultra-efficient SAR sea ice segmentation tool using FPGA for real-time satellite processing.
- Design Generative AI for Practitioners: Exploring Interaction Approaches Aligned with Creative Practice (viability: 3): https://sciencetostartup.com/paper/design-generative-ai-for-practitioners-exploring-interaction-approaches-aligned-with-creative-practice - Explore new generative AI interaction approaches tailored for creative practitioners to enhance design alignment.
- TikZilla: Scaling Text-to-TikZ with High-Quality Data and Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/tikzilla-scaling-text-to-tikz-with-high-quality-data-and-reinforcement-learning - TikZilla converts textual descriptions into precise scientific diagrams using a large, high-quality dataset and reinforcement learning.
- Reinforcement Learning with Symbolic Reward Machines (viability: 2): https://sciencetostartup.com/paper/reinforcement-learning-with-symbolic-reward-machines - Develop Symbolic Reward Machines to enhance reinforcement learning by eliminating manual label creation.
- TrustMH-Bench: A Comprehensive Benchmark for Evaluating the Trustworthiness of Large Language Models in Mental Health (viability: 6): https://sciencetostartup.com/paper/trustmh-bench-a-comprehensive-benchmark-for-evaluating-the-trustworthiness-of-large-language-models-in-mental-health - TrustMH-Bench provides a framework to evaluate and improve the trustworthiness of mental health large language models across key dimensions.
- IoUCert: Robustness Verification for Anchor-based Object Detectors (viability: 5): https://sciencetostartup.com/paper/ioucert-robustness-verification-for-anchor-based-object-detectors - IoUCert provides a formal verification framework to enhance robustness in anchor-based object detection models against input perturbations.
- cPNN: Continuous Progressive Neural Networks for Evolving Streaming Time Series (viability: 3): https://sciencetostartup.com/paper/cpnn-continuous-progressive-neural-networks-for-evolving-streaming-time-series - Develop a neural network focused on evolving time series data with concept drift and temporal dependencies.
- MA-CoNav: A Master-Slave Multi-Agent Framework with Hierarchical Collaboration and Dual-Level Reflection for Long-Horizon Embodied VLN (viability: 5): https://sciencetostartup.com/paper/ma-conav-a-master-slave-multi-agent-framework-with-hierarchical-collaboration-and-dual-level-reflection-for-long-horizon - Develop a master-slave multi-agent navigation framework for long-horizon embodied vision-language navigation.
- REGAL: A Registry-Driven Architecture for Deterministic Grounding of Agentic AI in Enterprise Telemetry (viability: 3): https://sciencetostartup.com/paper/regal-a-registry-driven-architecture-for-deterministic-grounding-of-agentic-ai-in-enterprise-telemetry - Develop deterministic AI systems for enterprise telemetry using a registry-driven architecture to enhance governance and efficiency.
- OrchMAS: Orchestrated Reasoning with Multi Collaborative Heterogeneous Scientific Expert Structured Agents (viability: 6): https://sciencetostartup.com/paper/orchmas-orchestrated-reasoning-with-multi-collaborative-heterogeneous-scientific-expert-structured-agents - A framework for orchestrating multi-agent LLMs to enhance scientific reasoning by dynamically adapting agents and workflows.
- SpatialText: A Pure-Text Cognitive Benchmark for Spatial Understanding in Large Language Models (viability: 5): https://sciencetostartup.com/paper/spatialtext-a-pure-text-cognitive-benchmark-for-spatial-understanding-in-large-language-models - SpatialText provides a diagnostic benchmark to evaluate the spatial reasoning capabilities of large language models through text-only scenarios.
- Why Does RLAIF Work At All? (viability: 3): https://sciencetostartup.com/paper/why-does-rlaif-work-at-all - A theoretical exploration of the mechanisms behind Reinforcement Learning from AI Feedback (RLAIF).
- Contextualized Privacy Defense for LLM Agents (viability: 3): https://sciencetostartup.com/paper/contextualized-privacy-defense-for-llm-agents - Develop a privacy defense system for LLM agents using proactive, context-aware instructor models to ensure better privacy-helpfulness balance.
- Delegation and Verification Under AI (viability: 2): https://sciencetostartup.com/paper/delegation-and-verification-under-ai - An analysis on how AI affects delegation and verification in institutional workflows, highlighting potential disparities in worker quality.
- Architecting Trust in Artificial Epistemic Agents (viability: 2): https://sciencetostartup.com/paper/architecting-trust-in-artificial-epistemic-agents - A framework to build trust in AI epistemic agents to align with human knowledge goals.
- The Geometry of Learning Under AI Delegation (viability: 3): https://sciencetostartup.com/paper/the-geometry-of-learning-under-ai-delegation - Develop a mathematical framework to understand the impact of AI delegation on human skill dynamics over time.
- SEALing the Gap: A Reference Framework for LLM Inference Carbon Estimation via Multi-Benchmark Driven Embodiment (viability: 6): https://sciencetostartup.com/paper/sealing-the-gap-a-reference-framework-for-llm-inference-carbon-estimation-via-multi-benchmark-driven-embodiment - Develop a framework for measuring the carbon footprint of LLM inference to promote sustainability in AI applications.
- Enhancing Physics-Informed Neural Networks with Domain-aware Fourier Features: Towards Improved Performance and Interpretable Results (viability: 5): https://sciencetostartup.com/paper/enhancing-physics-informed-neural-networks-with-domain-aware-fourier-features-towards-improved-performance-and-interpret - Build more interpretable Physics-Informed Neural Networks using Domain-aware Fourier Features for enhanced performance in computational physics.
- ShipTraj-R1: Reinforcing Ship Trajectory Prediction in Large Language Models via Group Relative Policy Optimization (viability: 7): https://sciencetostartup.com/paper/shiptraj-r1-reinforcing-ship-trajectory-prediction-in-large-language-models-via-group-relative-policy-optimization - ShipTraj-R1 leverages large language models with reinforcement learning to enhance ship trajectory prediction accuracy.
- Beyond One-Size-Fits-All: Adaptive Subgraph Denoising for Zero-Shot Graph Learning with Large Language Models (viability: 3): https://sciencetostartup.com/paper/beyond-one-size-fits-all-adaptive-subgraph-denoising-for-zero-shot-graph-learning-with-large-language-models - Develop an adaptive subgraph extraction framework for improving zero-shot graph reasoning with large language models.
- On the Structural Limitations of Weight-Based Neural Adaptation and the Role of Reversible Behavioral Learning (viability: 3): https://sciencetostartup.com/paper/on-the-structural-limitations-of-weight-based-neural-adaptation-and-the-role-of-reversible-behavioral-learning - A research study on structural irreversibility in neural adaptation and reversible behavioral learning.
- Interpretable Motion-Attentive Maps: Spatio-Temporally Localizing Concepts in Video Diffusion Transformers (viability: 3): https://sciencetostartup.com/paper/interpretable-motion-attentive-maps-spatio-temporally-localizing-concepts-in-video-diffusion-transformers - Develop a tool for generating interpretable maps to understand motion in video diffusion transformers.
- Eliciting Numerical Predictive Distributions of LLMs Without Autoregression (viability: 2): https://sciencetostartup.com/paper/eliciting-numerical-predictive-distributions-of-llms-without-autoregression - Develop a lightweight method for predicting LLM numerical outputs without autoregression.
- Learning to Generate and Extract: A Multi-Agent Collaboration Framework For Zero-shot Document-level Event Arguments Extraction (viability: 6): https://sciencetostartup.com/paper/learning-to-generate-and-extract-a-multi-agent-collaboration-framework-for-zero-shot-document-level-event-arguments-extr - A multi-agent framework for zero-shot document-level event arguments extraction leverages synthetic data to enhance deep learning models.
- SAE as a Crystal Ball: Interpretable Features Predict Cross-domain Transferability of LLMs without Training (viability: 5): https://sciencetostartup.com/paper/sae-as-a-crystal-ball-interpretable-features-predict-cross-domain-transferability-of-llms-without-training - SAE-based Transferability Score predicts LLM transferability across domains without fine-tuning.
- StegaFFD: Privacy-Preserving Face Forgery Detection via Fine-Grained Steganographic Domain Lifting (viability: 5): https://sciencetostartup.com/paper/stegaffd-privacy-preserving-face-forgery-detection-via-fine-grained-steganographic-domain-lifting - A novel framework for privacy-preserving face forgery detection using steganography to mask sensitive data without alerting attackers.
- Retrievit: In-context Retrieval Capabilities of Transformers, State Space Models, and Hybrid Architectures (viability: 2): https://sciencetostartup.com/paper/retrievit-in-context-retrieval-capabilities-of-transformers-state-space-models-and-hybrid-architectures - Exploring hybrid models combining Transformers and State Space Models for improved in-context retrieval tasks.
- LLM-based Argument Mining meets Argumentation and Description Logics: a Unified Framework for Reasoning about Debates (viability: 3): https://sciencetostartup.com/paper/llm-based-argument-mining-meets-argumentation-and-description-logics-a-unified-framework-for-reasoning-about-debates - A framework integrating argument mining with logic-based reasoning for analyzing debates.
- CoFL: Continuous Flow Fields for Language-Conditioned Navigation (viability: 7): https://sciencetostartup.com/paper/cofl-continuous-flow-fields-for-language-conditioned-navigation - CoFL enables precise, language-conditioned robotic navigation in complex environments using a transformer-based continuous flow field estimation.
- Learning Memory-Enhanced Improvement Heuristics for Flexible Job Shop Scheduling (viability: 7): https://sciencetostartup.com/paper/learning-memory-enhanced-improvement-heuristics-for-flexible-job-shop-scheduling - Develop advanced scheduling solutions for smart manufacturing with an AI-driven improvement heuristic framework, MIStar, that enhances decision-making using heterogeneous graph neural networks.
- SPARC: Spatial-Aware Path Planning via Attentive Robot Communication (viability: 3): https://sciencetostartup.com/paper/sparc-spatial-aware-path-planning-via-attentive-robot-communication - Spatial-aware path planning for multi-robot systems using Relation enhanced Multi Head Attention for improved coordination.
- Faster, Cheaper, More Accurate: Specialised Knowledge Tracing Models Outperform LLMs (viability: 5): https://sciencetostartup.com/paper/faster-cheaper-more-accurate-specialised-knowledge-tracing-models-outperform-llms - Develop specialized knowledge tracing models for educational platforms that outperform LLMs in accuracy, speed, and cost.
- BrandFusion: A Multi-Agent Framework for Seamless Brand Integration in Text-to-Video Generation (viability: 8): https://sciencetostartup.com/paper/brandfusion-a-multi-agent-framework-for-seamless-brand-integration-in-text-to-video-generation - BrandFusion seamlessly integrates brands into text-to-video content, revolutionizing advertising possibilities in content creation.
- Guideline-Grounded Evidence Accumulation for High-Stakes Agent Verification (viability: 7): https://sciencetostartup.com/paper/guideline-grounded-evidence-accumulation-for-high-stakes-agent-verification - Develop a verification framework for LLM-powered clinical agents that ensures trustworthy decision-making by aligning with domain-specific guidelines.
- Differentiable Time-Varying IIR Filtering for Real-Time Speech Denoising (viability: 7): https://sciencetostartup.com/paper/differentiable-time-varying-iir-filtering-for-real-time-speech-denoising - Develop real-time adaptive speech denoising with interpretable filtering and low-latency performance.
- OCR or Not? Rethinking Document Information Extraction in the MLLMs Era with Real-World Large-Scale Datasets (viability: 6): https://sciencetostartup.com/paper/ocr-or-not-rethinking-document-information-extraction-in-the-mllms-era-with-real-world-large-scale-datasets - Simplify document information extraction by utilizing MLLMs without OCR for efficiency and comparable performance.
- Agentified Assessment of Logical Reasoning Agents (viability: 5): https://sciencetostartup.com/paper/agentified-assessment-of-logical-reasoning-agents - Develop a framework for benchmarking logical reasoning agents with a focus on reproducibility and robustness.
- Rethinking Code Similarity for Automated Algorithm Design with LLMs (viability: 6): https://sciencetostartup.com/paper/rethinking-code-similarity-for-automated-algorithm-design-with-llms - BehaveSim offers a novel approach to measure algorithmic similarity, enhancing LLM-AAD frameworks by promoting behavioral diversity and clustering algorithms for systematic analysis.
- Scores Know Bobs Voice: Speaker Impersonation Attack (viability: 3): https://sciencetostartup.com/paper/scores-know-bobs-voice-speaker-impersonation-attack - Develop a feature-aligned inversion framework to enhance speaker recognition system robustness by improving attack efficiency.
- ITO: Images and Texts as One via Synergizing Multiple Alignment and Training-Time Fusion (viability: 5): https://sciencetostartup.com/paper/ito-images-and-texts-as-one-via-synergizing-multiple-alignment-and-training-time-fusion - ITO enhances multimodal learning by unifying image-text representations with a fusion mechanism optimized for inference efficiency.
- EvoSkill: Automated Skill Discovery for Multi-Agent Systems (viability: 7): https://sciencetostartup.com/paper/evoskill-automated-skill-discovery-for-multi-agent-systems - EvoSkill automates discovery and refinement of agent skills to enhance performance in multi-agent systems, offering measurable improvements on benchmarks.
- Next Embedding Prediction Makes World Models Stronger (viability: 2): https://sciencetostartup.com/paper/next-embedding-prediction-makes-world-models-stronger - NE-Dreamer uses temporal transformers for next-step embedding prediction, enhancing state representation learning in model-based reinforcement learning.
- Efficient Self-Evaluation for Diffusion Language Models via Sequence Regeneration (viability: 4): https://sciencetostartup.com/paper/efficient-self-evaluation-for-diffusion-language-models-via-sequence-regeneration - Develop a self-evaluation tool for diffusion language models to enhance generation quality assessment through token regeneration probabilities.
- iGVLM: Dynamic Instruction-Guided Vision Encoding for Question-Aware Multimodal Understanding (viability: 6): https://sciencetostartup.com/paper/igvlm-dynamic-instruction-guided-vision-encoding-for-question-aware-multimodal-understanding - Develop a dynamic instruction-guided vision encoder for enhanced multimodal understanding.
- Enhancing User Throughput in Multi-panel mmWave Radio Access Networks for Beam-based MU-MIMO Using a DRL Method (viability: 6): https://sciencetostartup.com/paper/enhancing-user-throughput-in-multi-panel-mmwave-radio-access-networks-for-beam-based-mu-mimo-using-a-drl-method - Develop a DRL-based system to enhance user throughput in mmWave networks for beam-based MU-MIMO applications.
- Practical FP4 Training for Large-Scale MoE Models on Hopper GPUs (viability: 3): https://sciencetostartup.com/paper/practical-fp4-training-for-large-scale-moe-models-on-hopper-gpus - Optimize large-scale MoE model training on Hopper GPUs by implementing FP4 without native support.
- A Natural Language Agentic Approach to Study Affective Polarization (viability: 3): https://sciencetostartup.com/paper/a-natural-language-agentic-approach-to-study-affective-polarization - A multi-agent platform uses LLMs to simulate and study affective polarization on social media.
- Sensory-Aware Sequential Recommendation via Review-Distilled Representations (viability: 7): https://sciencetostartup.com/paper/sensory-aware-sequential-recommendation-via-review-distilled-representations - A framework that enhances recommendation systems by incorporating sensory attributes from product reviews, improving interpretability and performance.
- Intelligent Pathological Diagnosis of Gestational Trophoblastic Diseases via Visual-Language Deep Learning Model (viability: 7): https://sciencetostartup.com/paper/intelligent-pathological-diagnosis-of-gestational-trophoblastic-diseases-via-visual-language-deep-learning-model - AI-driven tool streamlines and enhances the accuracy of diagnosing gestational trophoblastic diseases through visual-language deep learning.
- FinTexTS: Financial Text-Paired Time-Series Dataset via Semantic-Based and Multi-Level Pairing (viability: 5): https://sciencetostartup.com/paper/fintexts-financial-text-paired-time-series-dataset-via-semantic-based-and-multi-level-pairing - FinTexTS offers a semantic-based framework for pairing financial news with time-series data to improve stock price forecasting.
- ShareVerse: Multi-Agent Consistent Video Generation for Shared World Modeling (viability: 7): https://sciencetostartup.com/paper/shareverse-multi-agent-consistent-video-generation-for-shared-world-modeling - ShareVerse is a video generation framework for unified multi-agent interactive world modeling using large video models and cross-agent attention mechanisms.
- Retrieval-Augmented Robots via Retrieve-Reason-Act (viability: 6): https://sciencetostartup.com/paper/retrieval-augmented-robots-via-retrieve-reason-act - Retrieval-Augmented Robots empower robots to actively seek procedural knowledge from unstructured documentation for enhanced task execution.
- LLMs for High-Frequency Decision-Making: Normalized Action Reward-Guided Consistency Policy Optimization (viability: 2): https://sciencetostartup.com/paper/llms-for-high-frequency-decision-making-normalized-action-reward-guided-consistency-policy-optimization - Optimizing LLMs for high-frequency decision-making tasks with Normalized Action Reward-guided Consistency Policy Optimization.
- ITLC at SemEval-2026 Task 11: Normalization and Deterministic Parsing for Formal Reasoning in LLMs (viability: 5): https://sciencetostartup.com/paper/itlc-at-semeval-2026-task-11-normalization-and-deterministic-parsing-for-formal-reasoning-in-llms - Novel method for structural abstraction and deterministic parsing to reduce bias in multilingual reasoning tasks.
- SorryDB: Can AI Provers Complete Real-World Lean Theorems? (viability: 6): https://sciencetostartup.com/paper/sorrydb-can-ai-provers-complete-real-world-lean-theorems - Develop dynamic AI-based tools tailored for real-world Lean theorem proving, leveraging an ever-updating SorryDB benchmark.
- Real-Time Generation of Game Video Commentary with Multimodal LLMs: Pause-Aware Decoding Approaches (viability: 7): https://sciencetostartup.com/paper/real-time-generation-of-game-video-commentary-with-multimodal-llms-pause-aware-decoding-approaches - Develop real-time video commentary generation integrating timing for enhanced engagement in gaming and livestreaming.
- AlphaFree: Recommendation Free from Users, IDs, and GNNs (viability: 5): https://sciencetostartup.com/paper/alphafree-recommendation-free-from-users-ids-and-gnns - AlphaFree offers a memory-efficient, user-free recommendation system outperforming traditional methods by leveraging language representations instead of user or ID embeddings.
- Improving Diffusion Planners by Self-Supervised Action Gating with Energies (viability: 5): https://sciencetostartup.com/paper/improving-diffusion-planners-by-self-supervised-action-gating-with-energies - Enhance diffusion planners in offline reinforcement learning with SAGE method to improve trajectory consistency without retraining policies.
- Credibility Governance: A Social Mechanism for Collective Self-Correction under Weak Truth Signals (viability: 5): https://sciencetostartup.com/paper/credibility-governance-a-social-mechanism-for-collective-self-correction-under-weak-truth-signals - Develop a credibility governance tool to enhance the reliability of online opinion aggregation systems.
- Robust Heterogeneous Analog-Digital Computing for Mixture-of-Experts Models with Theoretical Generalization Guarantees (viability: 4): https://sciencetostartup.com/paper/robust-heterogeneous-analog-digital-computing-for-mixture-of-experts-models-with-theoretical-generalization-guarantees - Develop a robust framework for efficient Mixture-of-Experts models using analog-digital computing hybrid systems.
- MASPOB: Bandit-Based Prompt Optimization for Multi-Agent Systems with Graph Neural Networks (viability: 7): https://sciencetostartup.com/paper/maspob-bandit-based-prompt-optimization-for-multi-agent-systems-with-graph-neural-networks - Optimize prompt efficiency in Multi-Agent Systems using bandit algorithms and graph neural networks for improved performance.
- See and Remember: A Multimodal Agent for Web Traversal (viability: 7): https://sciencetostartup.com/paper/see-and-remember-a-multimodal-agent-for-web-traversal - V-GEMS provides a robust multimodal agent for precise and resilient web navigation, utilizing visual grounding and explicit memory systems.
- AgentAssay: Token-Efficient Regression Testing for Non-Deterministic AI Agent Workflows (viability: 7): https://sciencetostartup.com/paper/agentassay-token-efficient-regression-testing-for-non-deterministic-ai-agent-workflows - AgentAssay offers a token-efficient framework for regression testing AI agents with significant cost reductions and strong statistical guarantees.
- SUN: Shared Use of Next-token Prediction for Efficient Multi-LLM Disaggregated Serving (viability: 3): https://sciencetostartup.com/paper/sun-shared-use-of-next-token-prediction-for-efficient-multi-llm-disaggregated-serving - Develop a scalable system optimizing multi-LLM decoding with shared execution for efficient throughput.
- GPUTOK: GPU Accelerated Byte Level BPE Tokenization (viability: 4): https://sciencetostartup.com/paper/gputok-gpu-accelerated-byte-level-bpe-tokenization - GPU-accelerated tokenizer that speeds up large language model tokenization processes by 1.7-7.6 times compared to existing solutions, while maintaining quality.
- LiveAgentBench: Comprehensive Benchmarking of Agentic Systems Across 104 Real-World Challenges (viability: 5): https://sciencetostartup.com/paper/liveagentbench-comprehensive-benchmarking-of-agentic-systems-across-104-real-world-challenges - LiveAgentBench is a comprehensive benchmark tool designed to evaluate AI agents using 104 real-world scenarios.
- How Controllable Are Large Language Models? A Unified Evaluation across Behavioral Granularities (viability: 5): https://sciencetostartup.com/paper/how-controllable-are-large-language-models-a-unified-evaluation-across-behavioral-granularities - SteerEval provides a benchmark to evaluate and improve the behavioral controllability of large language models in sensitive domains.
- CAPT: Confusion-Aware Prompt Tuning for Reducing Vision-Language Misalignment (viability: 8): https://sciencetostartup.com/paper/capt-confusion-aware-prompt-tuning-for-reducing-vision-language-misalignment - CAPT uses confusion-aware prompt tuning to enhance vision-language model accuracy by learning from misalignments in visually and semantically similar categories.
- Through the Lens of Contrast: Self-Improving Visual Reasoning in VLMs (viability: 7): https://sciencetostartup.com/paper/through-the-lens-of-contrast-self-improving-visual-reasoning-in-vlms - Develop a visual reasoning enhancement tool for VLMs that leverages contrastive learning to improve accuracy.
- CoDAR: Continuous Diffusion Language Models are More Powerful Than You Think (viability: 5): https://sciencetostartup.com/paper/codar-continuous-diffusion-language-models-are-more-powerful-than-you-think - Develop a continuous diffusion model with contextual autoregressive decoding for improved language model generation.
- AnchorDrive: LLM Scenario Rollout with Anchor-Guided Diffusion Regeneration for Safety-Critical Scenario Generation (viability: 7): https://sciencetostartup.com/paper/anchordrive-llm-scenario-rollout-with-anchor-guided-diffusion-regeneration-for-safety-critical-scenario-generation - AnchorDrive enhances autonomous vehicle safety by generating controllable and realistic safety-critical driving scenarios using LLM and diffusion models.
- A Neuropsychologically Grounded Evaluation of LLM Cognitive Abilities (viability: 5): https://sciencetostartup.com/paper/a-neuropsychologically-grounded-evaluation-of-llm-cognitive-abilities - Introduce NeuroCognition benchmark for evaluating LLM cognitive abilities distinct from existing benchmarks.
- Bridging Diffusion Guidance and Anderson Acceleration via Hopfield Dynamics (viability: 5): https://sciencetostartup.com/paper/bridging-diffusion-guidance-and-anderson-acceleration-via-hopfield-dynamics - Geometry Aware Attention Guidance enhances diffusion model efficiency by stabilizing attention dynamics through Hopfield Networks.
- LLM-MLFFN: Multi-Level Autonomous Driving Behavior Feature Fusion via Large Language Model (viability: 7): https://sciencetostartup.com/paper/llm-mlffn-multi-level-autonomous-driving-behavior-feature-fusion-via-large-language-model - LLM-MLFFN uses large language models for enhanced interpretation and classification of autonomous driving behaviors, boosting safety and robustness.
- Human-Certified Module Repositories for the AI Age (viability: 3): https://sciencetostartup.com/paper/human-certified-module-repositories-for-the-ai-age - Introduce a framework for Human-Certified Module Repositories to ensure trustworthy AI-assisted software development.
- Learning Object-Centric Spatial Reasoning for Sequential Manipulation in Cluttered Environments (viability: 7): https://sciencetostartup.com/paper/learning-object-centric-spatial-reasoning-for-sequential-manipulation-in-cluttered-environments - Develop a specialized robotic manipulation system, Unveiler, for efficiently handling object-centric spatial reasoning in cluttered environments.
- NeuroProlog: Multi-Task Fine-Tuning for Neurosymbolic Mathematical Reasoning via the Cocktail Effect (viability: 6): https://sciencetostartup.com/paper/neuroprolog-multi-task-fine-tuning-for-neurosymbolic-mathematical-reasoning-via-the-cocktail-effect - NeuroProlog enhances mathematical reasoning in LLMs using neurosymbolic techniques to ensure verifiable Prolog program execution.
- Revealing Positive and Negative Role Models to Help People Make Good Decisions (viability: 6): https://sciencetostartup.com/paper/revealing-positive-and-negative-role-models-to-help-people-make-good-decisions - A decision-making tool leveraging role models in social networks to maximize social welfare by revealing positive and negative influencers.
- What Capable Agents Must Know: Selection Theorems for Robust Decision-Making under Uncertainty (viability: 2): https://sciencetostartup.com/paper/what-capable-agents-must-know-selection-theorems-for-robust-decision-making-under-uncertainty - Explores necessary internal structures for robust AI decision-making under uncertainty.
- PRISM: Pushing the Frontier of Deep Think via Process Reward Model-Guided Inference (viability: 4): https://sciencetostartup.com/paper/prism-pushing-the-frontier-of-deep-think-via-process-reward-model-guided-inference - PRISM enhances the reliability of deep reasoning systems by guiding solution refinement and aggregation through a process reward model.
- Diagnosing Retrieval vs. Utilization Bottlenecks in LLM Agent Memory (viability: 6): https://sciencetostartup.com/paper/diagnosing-retrieval-vs-utilization-bottlenecks-in-llm-agent-memory - Develop a diagnostic tool to enhance LLM agent memory retrieval performance using existing storage strategies.
- Deep Learning Based Wildfire Detection for Peatland Fires Using Transfer Learning (viability: 6): https://sciencetostartup.com/paper/deep-learning-based-wildfire-detection-for-peatland-fires-using-transfer-learning - Transfer learning-based wildfire detection system specialized for peatland fires, enhancing early detection and environmental protection.
- GLoRIA: Gated Low-Rank Interpretable Adaptation for Dialectal ASR (viability: 5): https://sciencetostartup.com/paper/gloria-gated-low-rank-interpretable-adaptation-for-dialectal-asr - Develop a metadata-gated adaptation framework to enhance ASR systems in dialect-heavy settings.
- Can Computational Reducibility Lead to Transferable Models for Graph Combinatorial Optimization? (viability: 7): https://sciencetostartup.com/paper/can-computational-reducibility-lead-to-transferable-models-for-graph-combinatorial-optimization - Develop a neural solver using message passing and pretraining for generalized combinatorial optimization across multiple graph tasks.
- Manifold Aware Denoising Score Matching (MAD) (viability: 5): https://sciencetostartup.com/paper/manifold-aware-denoising-score-matching-mad - Manifold Aware Denoising Score Matching optimizes manifold learning in data distributions using a computationally efficient score decomposition approach.
- VL-KGE: Vision-Language Models Meet Knowledge Graph Embeddings (viability: 7): https://sciencetostartup.com/paper/vl-kge-vision-language-models-meet-knowledge-graph-embeddings - Integrate Vision-Language Models with Knowledge Graph Embeddings for robust multimodal reasoning.
- MIRAGE: Knowledge Graph-Guided Cross-Cohort MRI Synthesis for Alzheimer's Disease Prediction (viability: 6): https://sciencetostartup.com/paper/mirage-knowledge-graph-guided-cross-cohort-mri-synthesis-for-alzheimer-s-disease-prediction - MIRAGE enables cost-effective Alzheimer's diagnosis by synthesizing MRI-like insights from EHR data using a novel knowledge graph framework.
- Learning to Pay Attention: Unsupervised Modeling of Attentive and Inattentive Respondents in Survey Data (viability: 5): https://sciencetostartup.com/paper/learning-to-pay-attention-unsupervised-modeling-of-attentive-and-inattentive-respondents-in-survey-data - A domain-agnostic tool for detecting inattentive respondents in surveys using unsupervised learning to improve data quality without additional respondent burden.
- A Directed Graph Model and Experimental Framework for Design and Study of Time-Dependent Text Visualisation (viability: 5): https://sciencetostartup.com/paper/a-directed-graph-model-and-experimental-framework-for-design-and-study-of-time-dependent-text-visualisation - Develop a customizable time-dependent text visualization tool that adapts to user-specific exploration needs.
- Slurry-as-a-Service: A Modest Proposal on Scalable Pluralistic Alignment for Nutrient Optimization (viability: 2): https://sciencetostartup.com/paper/slurry-as-a-service-a-modest-proposal-on-scalable-pluralistic-alignment-for-nutrient-optimization - ValueMulch is a satirical critique of AI alignment practices through the fictional concept of mulching models that align to community norms.
- From Fewer Samples to Fewer Bits: Reframing Dataset Distillation as Joint Optimization of Precision and Compactness (viability: 8): https://sciencetostartup.com/paper/from-fewer-samples-to-fewer-bits-reframing-dataset-distillation-as-joint-optimization-of-precision-and-compactness - QuADD offers a quantization-aware dataset distillation framework optimizing data compactness and precision for efficient machine learning.
- Rigidity-Aware Geometric Pretraining for Protein Design and Conformational Ensembles (viability: 5): https://sciencetostartup.com/paper/rigidity-aware-geometric-pretraining-for-protein-design-and-conformational-ensembles - Rigidity-Aware Geometric Pretraining enhances protein design with self-supervised learning for better structural predictions.
- COOL-MC: Verifying and Explaining RL Policies for Platelet Inventory Management (viability: 5): https://sciencetostartup.com/paper/cool-mc-verifying-and-explaining-rl-policies-for-platelet-inventory-management - Developing a verified and explainable RL-based tool for managing platelet inventory in blood banks.
- PlayWrite: A Multimodal System for AI Supported Narrative Co-Authoring Through Play in XR (viability: 6): https://sciencetostartup.com/paper/playwrite-a-multimodal-system-for-ai-supported-narrative-co-authoring-through-play-in-xr - PlayWrite offers writers a mixed-reality environment for narrative co-authoring by manipulating virtual characters, enhancing improvisation and collaborative creativity with AI.
- Can machines be uncertain? (viability: 3): https://sciencetostartup.com/paper/can-machines-be-uncertain - Exploring how AI systems can represent and process states of uncertainty.
- Estimating Visual Attribute Effects in Advertising from Observational Data: A Deepfake-Informed Double Machine Learning Approach (viability: 7): https://sciencetostartup.com/paper/estimating-visual-attribute-effects-in-advertising-from-observational-data-a-deepfake-informed-double-machine-learning-a - Build a tool using DICE-DML to provide more accurate causal analysis of visual attributes on consumer engagement in digital advertising.
- Diffusion-MPC in Discrete Domains: Feasibility Constraints, Horizon Effects, and Critic Alignment: Case study with Tetris (viability: 3): https://sciencetostartup.com/paper/diffusion-mpc-in-discrete-domains-feasibility-constraints-horizon-effects-and-critic-alignment-case-study-with-tetris - Optimize Diffusion-MPC strategies in discrete domains through improved sampling and reranking methodologies.
- RIVA: Leveraging LLM Agents for Reliable Configuration Drift Detection (viability: 6): https://sciencetostartup.com/paper/riva-leveraging-llm-agents-for-reliable-configuration-drift-detection - RIVA: A robust multi-agent system that enhances infrastructure verification in cloud environments by detecting configuration drift with high accuracy using LLM agents.
- Preconditioned Score and Flow Matching (viability: 4): https://sciencetostartup.com/paper/preconditioned-score-and-flow-matching - Optimize diffusion models using preconditioning maps to improve training outcomes without altering generative models.
- Tool Verification for Test-Time Reinforcement Learning (viability: 2): https://sciencetostartup.com/paper/tool-verification-for-test-time-reinforcement-learning - Develop a tool verification system to enhance test-time reinforcement learning stability.
- Adaptive Confidence Regularization for Multimodal Failure Detection (viability: 8): https://sciencetostartup.com/paper/adaptive-confidence-regularization-for-multimodal-failure-detection - ACR framework for reliable failure detection in multimodal AI systems, critical for safety in high-stakes domains like autonomous driving and medical diagnostics.
- Conformal Policy Control (viability: 2): https://sciencetostartup.com/paper/conformal-policy-control - A framework for safely exploring new policies using conformal calibration with probabilistic regulation.
- Symbol-Equivariant Recurrent Reasoning Models (viability: 6): https://sciencetostartup.com/paper/symbol-equivariant-recurrent-reasoning-models - Develop reasoning models with symbol-equivariant layers for efficient and scalable solution of structured problems like Sudoku and ARC-AGI.
- Sketch2Colab: Sketch-Conditioned Multi-Human Animation via Controllable Flow Distillation (viability: 7): https://sciencetostartup.com/paper/sketch2colab-sketch-conditioned-multi-human-animation-via-controllable-flow-distillation - Sketch2Colab transforms storyboard sketches into precise 3D animations with rapid and stable rendering for enhanced creative control.
- MAC: A Conversion Rate Prediction Benchmark Featuring Labels Under Multiple Attribution Mechanisms (viability: 9): https://sciencetostartup.com/paper/mac-a-conversion-rate-prediction-benchmark-featuring-labels-under-multiple-attribution-mechanisms - MAC provides a conversion rate prediction benchmark featuring multi-attribution labels, significantly enhancing accuracy in online advertising.
- Leveraging Model Soups to Classify Intangible Cultural Heritage Images from the Mekong Delta (viability: 6): https://sciencetostartup.com/paper/leveraging-model-soups-to-classify-intangible-cultural-heritage-images-from-the-mekong-delta - Develop a robust image classification tool for intangible cultural heritage images using model soups for enhanced generalization on low-resource datasets.
- Reservoir Subspace Injection for Online ICA under Top-n Whitening (viability: 5): https://sciencetostartup.com/paper/reservoir-subspace-injection-for-online-ica-under-top-n-whitening - Develop an enhanced online ICA algorithm with reservoir subspace injection for improved signal separation.
- Kiwi-Edit: Versatile Video Editing via Instruction and Reference Guidance (viability: 7): https://sciencetostartup.com/paper/kiwi-edit-versatile-video-editing-via-instruction-and-reference-guidance - Kiwi-Edit offers precise instruction-based video editing with reference guidance, supported by a new scalable dataset and codebase.
- SageBwd: A Trainable Low-bit Attention (viability: 5): https://sciencetostartup.com/paper/sagebwd-a-trainable-low-bit-attention - SageBwd enables efficient training with low-bit attention by reducing quantization errors, providing an alternative to full-precision models.
- How Small Can 6G Reason? Scaling Tiny Language Models for AI-Native Networks (viability: 7): https://sciencetostartup.com/paper/how-small-can-6g-reason-scaling-tiny-language-models-for-ai-native-networks - Optimize AI-native networks in 6G with compact, efficient language models.
- Near-Optimal Regret for KL-Regularized Multi-Armed Bandits (viability: 2): https://sciencetostartup.com/paper/near-optimal-regret-for-kl-regularized-multi-armed-bandits - Advanced KL-regularized multi-armed bandit algorithms offer nearly optimal theoretical bounds, enhancing reinforcement learning efficiency.
- Boltzmann-based Exploration for Robust Decentralized Multi-Agent Planning (viability: 5): https://sciencetostartup.com/paper/boltzmann-based-exploration-for-robust-decentralized-multi-agent-planning - Develop a robust multi-agent planning tool using CB-MCTS that improves performance in sparse reward environments.
- Scaling Retrieval Augmented Generation with RAG Fusion: Lessons from an Industry Deployment (viability: 5): https://sciencetostartup.com/paper/scaling-retrieval-augmented-generation-with-rag-fusion-lessons-from-an-industry-deployment - Optimize RAG pipelines by considering production constraints on retrieval and re-ranking budgets.
- Zero- and Few-Shot Named-Entity Recognition: Case Study and Dataset in the Crime Domain (CrimeNER) (viability: 5): https://sciencetostartup.com/paper/zero-and-few-shot-named-entity-recognition-case-study-and-dataset-in-the-crime-domain-crimener - CrimeNER provides a specialized NER dataset for crime-related document analysis, useful for law enforcement agencies.
- LiftAvatar: Kinematic-Space Completion for Expression-Controlled 3D Gaussian Avatar Animation (viability: 7): https://sciencetostartup.com/paper/liftavatar-kinematic-space-completion-for-expression-controlled-3d-gaussian-avatar-animation - LiftAvatar enhances 3D avatars by transforming sparse monocular data into detailed kinematic animations, boosting expressiveness and reconstruction fidelity.
- LLMs as Strategic Actors: Behavioral Alignment, Risk Calibration, and Argumentation Framing in Geopolitical Simulations (viability: 3): https://sciencetostartup.com/paper/llms-as-strategic-actors-behavioral-alignment-risk-calibration-and-argumentation-framing-in-geopolitical-simulations - Develop LLMs for geopolitical simulations to align behavior with human decision-making in strategic scenarios.
- Nano-EmoX: Unifying Multimodal Emotional Intelligence from Perception to Empathy (viability: 7): https://sciencetostartup.com/paper/nano-emox-unifying-multimodal-emotional-intelligence-from-perception-to-empathy - Nano-EmoX: A compact multimodal model for holistic emotional intelligence from perception to empathy.
- Pencil Puzzle Bench: A Benchmark for Multi-Step Verifiable Reasoning (viability: 5): https://sciencetostartup.com/paper/pencil-puzzle-bench-a-benchmark-for-multi-step-verifiable-reasoning - A framework for evaluating LLM reasoning via pencil puzzles, providing deterministic, step-level verification to enhance model reasoning capabilities.
- Robometer: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisons (viability: 8): https://sciencetostartup.com/paper/robometer-scaling-general-purpose-robotic-reward-models-via-trajectory-comparisons - Robometer offers scalable robot reward modeling using trajectory comparisons for enhanced automation learning.
- FluxMem: Adaptive Hierarchical Memory for Streaming Video Understanding (viability: 8): https://sciencetostartup.com/paper/fluxmem-adaptive-hierarchical-memory-for-streaming-video-understanding - FluxMem offers real-time adaptive video compression and understanding for resource-efficient streaming applications.
- On the Rate of Convergence of GD in Non-linear Neural Networks: An Adversarial Robustness Perspective (viability: 2): https://sciencetostartup.com/paper/on-the-rate-of-convergence-of-gd-in-non-linear-neural-networks-an-adversarial-robustness-perspective - Develop algorithms to improve the convergence rate of Gradient Descent in non-linear neural networks for better adversarial robustness.
- Learning from Synthetic Data Improves Multi-hop Reasoning (viability: 6): https://sciencetostartup.com/paper/learning-from-synthetic-data-improves-multi-hop-reasoning - Unlock the power of synthetic data to enhance multi-hop reasoning in language models efficiently and cost-effectively.
- Detection-Gated Glottal Segmentation with Zero-Shot Cross-Dataset Transfer and Clinical Feature Extraction (viability: 8): https://sciencetostartup.com/paper/detection-gated-glottal-segmentation-with-zero-shot-cross-dataset-transfer-and-clinical-feature-extraction - A zero-shot glottal segmentation AI for real-time clinical voice assessment using videoendoscopy.
- GenDB: The Next Generation of Query Processing -- Synthesized, Not Engineered (viability: 7): https://sciencetostartup.com/paper/gendb-the-next-generation-of-query-processing-synthesized-not-engineered - GenDB leverages LLMs to dynamically generate optimized query execution code, outperforming traditional data query engines.
- Cognitive Prosthetic: An AI-Enabled Multimodal System for Episodic Recall in Knowledge Work (viability: 6): https://sciencetostartup.com/paper/cognitive-prosthetic-an-ai-enabled-multimodal-system-for-episodic-recall-in-knowledge-work - Build an AI-powered system for enhancing episodic memory recall in the workplace using multimodal data and natural language queries.
- Exploring Plan Space through Conversation: An Agentic Framework for LLM-Mediated Explanations in Planning (viability: 3): https://sciencetostartup.com/paper/exploring-plan-space-through-conversation-an-agentic-framework-for-llm-mediated-explanations-in-planning - Developing a LLM-based interactive framework to enhance human guidance in AI planners through user-adaptive explanations.
- Scaling Laws of SignSGD in Linear Regression: When Does It Outperform SGD? (viability: 1): https://sciencetostartup.com/paper/scaling-laws-of-signsgd-in-linear-regression-when-does-it-outperform-sgd - Analyzes the scaling laws of signSGD in linear regression for optimization efficiency.
- OpenRad: a Curated Repository of Open-access AI models for Radiology (viability: 6): https://sciencetostartup.com/paper/openrad-a-curated-repository-of-open-access-ai-models-for-radiology - OpenRad is a comprehensive, open-access repository for discovering and contributing AI models in radiology with a keyword-searchable web interface.
- A Resource-Rational Principle for Modeling Visual Attention Control (viability: 3): https://sciencetostartup.com/paper/a-resource-rational-principle-for-modeling-visual-attention-control - A framework that models visual attention as a resource-rational sequential decision-making process for HCI applications.
- "When to Hand Off, When to Work Together": Expanding Human-Agent Co-Creative Collaboration through Concurrent Interaction (viability: 5): https://sciencetostartup.com/paper/when-to-hand-off-when-to-work-together-expanding-human-agent-co-creative-collaboration-through-concurrent-interaction - A co-creative AI tool, CLEO, interprets collaborative intent and adapts in real-time within human-agent design processes.
- EstLLM: Enhancing Estonian Capabilities in Multilingual LLMs via Continued Pretraining and Post-Training (viability: 6): https://sciencetostartup.com/paper/estllm-enhancing-estonian-capabilities-in-multilingual-llms-via-continued-pretraining-and-post-training - EstLLM enhances Estonian language capabilities in multilingual LLMs through continued pretraining and strategic post-training, improving language-specific performance.
- Rich Insights from Cheap Signals: Efficient Evaluations via Tensor Factorization (viability: 2): https://sciencetostartup.com/paper/rich-insights-from-cheap-signals-efficient-evaluations-via-tensor-factorization - Develop a statistical model that efficiently evaluates generative models using tensor factorization to merge autorater data with limited human labels.
- Revealing Combinatorial Reasoning of GNNs via Graph Concept Bottleneck Layer (viability: 5): https://sciencetostartup.com/paper/revealing-combinatorial-reasoning-of-gnns-via-graph-concept-bottleneck-layer - A method for integrating graph concept bottleneck layers in GNNs to enhance interpretability and performance.
- MMR-Life: Piecing Together Real-life Scenes for Multimodal Multi-image Reasoning (viability: 4): https://sciencetostartup.com/paper/mmr-life-piecing-together-real-life-scenes-for-multimodal-multi-image-reasoning - MMR-Life offers a comprehensive benchmark for evaluating and improving multimodal multi-image reasoning capabilities of AI models using real-life scenarios.
- CodecFlow: Efficient Bandwidth Extension via Conditional Flow Matching in Neural Codec Latent Space (viability: 3): https://sciencetostartup.com/paper/codecflow-efficient-bandwidth-extension-via-conditional-flow-matching-in-neural-codec-latent-space - CodecFlow enhances speech bandwidth extension by optimizing neural codec embeddings for improved fidelity and perceptual quality.
- Selection as Power: Constrained Reinforcement for Bounded Decision Authority (viability: 5): https://sciencetostartup.com/paper/selection-as-power-constrained-reinforcement-for-bounded-decision-authority - Adaptive reinforcement learning framework ensuring bounded decision authority for agentic systems.
- MAP-Diff: Multi-Anchor Guided Diffusion for Progressive 3D Whole-Body Low-Dose PET Denoising (viability: 6): https://sciencetostartup.com/paper/map-diff-multi-anchor-guided-diffusion-for-progressive-3d-whole-body-low-dose-pet-denoising - MAP-Diff offers a progressive 3D whole-body PET denoising tool reducing radiation exposure while maintaining scan quality.
- Temporal Representations for Exploration: Learning Complex Exploratory Behavior without Extrinsic Rewards (viability: 5): https://sciencetostartup.com/paper/temporal-representations-for-exploration-learning-complex-exploratory-behavior-without-extrinsic-rewards - Develop a reinforcement learning exploration tool using temporal contrastive representations to enhance task learning without extrinsic rewards.
- Mitigating topology biases in Graph Diffusion via Counterfactual Intervention (viability: 6): https://sciencetostartup.com/paper/mitigating-topology-biases-in-graph-diffusion-via-counterfactual-intervention - FairGDiff provides a counterfactual-based, debiasing solution for graph diffusion models to ensure fairness in graph generation tasks.
- MatRIS: Toward Reliable and Efficient Pretrained Machine Learning Interaction Potentials (viability: 5): https://sciencetostartup.com/paper/matris-toward-reliable-and-efficient-pretrained-machine-learning-interaction-potentials - Introducing MatRIS, an efficient invariant MLIP with attention-based modeling for scalable and expressive material simulations.
- According to Me: Long-Term Personalized Referential Memory QA (viability: 6): https://sciencetostartup.com/paper/according-to-me-long-term-personalized-referential-memory-qa - Develop a personalized AI assistant capable of long-term, multimodal memory recall using the ATM-Bench framework.
- CharacterFlywheel: Scaling Iterative Improvement of Engaging and Steerable LLMs in Production (viability: 7): https://sciencetostartup.com/paper/characterflywheel-scaling-iterative-improvement-of-engaging-and-steerable-llms-in-production - Enhancing LLMs for social chat applications with iterative improvement to boost engagement and steerability.
- Intrinsic Task Symmetry Drives Generalization in Algorithmic Tasks (viability: 2): https://sciencetostartup.com/paper/intrinsic-task-symmetry-drives-generalization-in-algorithmic-tasks - Leverage intrinsic task symmetries to enhance generalization in neural networks.
- AMemGym: Interactive Memory Benchmarking for Assistants in Long-Horizon Conversations (viability: 6): https://sciencetostartup.com/paper/amemgym-interactive-memory-benchmarking-for-assistants-in-long-horizon-conversations - AMemGym provides an interactive environment for optimizing memory in LLM-based conversational agents.
- TiledAttention: a CUDA Tile SDPA Kernel for PyTorch (viability: 6): https://sciencetostartup.com/paper/tiledattention-a-cuda-tile-sdpa-kernel-for-pytorch - TiledAttention enables customizable and performant SDPA operations in PyTorch for NVIDIA GPUs.
- Closed-Loop Action Chunks with Dynamic Corrections for Training-Free Diffusion Policy (viability: 7): https://sciencetostartup.com/paper/closed-loop-action-chunks-with-dynamic-corrections-for-training-free-diffusion-policy - DCDP offers real-time closed-loop action correction for robotic manipulation using a plug-and-play diffusion policy framework.
- LiveCultureBench: a Multi-Agent, Multi-Cultural Benchmark for Large Language Models in Dynamic Social Simulations (viability: 7): https://sciencetostartup.com/paper/liveculturebench-a-multi-agent-multi-cultural-benchmark-for-large-language-models-in-dynamic-social-simulations - LiveCultureBench offers a simulation-based tool for evaluating language models' task and cultural adherence in dynamic social environments.
- Probabilistic Retrofitting of Learned Simulators (viability: 4): https://sciencetostartup.com/paper/probabilistic-retrofitting-of-learned-simulators - Retrofitting pre-trained deterministic PDE models into probabilistic ones using CRPS for improved prediction accuracy.
- physfusion: A Transformer-based Dual-Stream Radar and Vision Fusion Framework for Open Water Surface Object Detection (viability: 7): https://sciencetostartup.com/paper/physfusion-a-transformer-based-dual-stream-radar-and-vision-fusion-framework-for-open-water-surface-object-detection - A cutting-edge radar-vision fusion framework for enhanced object detection on open water surfaces.
- When Numbers Tell Half the Story: Human-Metric Alignment in Topic Model Evaluation (viability: 5): https://sciencetostartup.com/paper/when-numbers-tell-half-the-story-human-metric-alignment-in-topic-model-evaluation - Develop a human-grounded topic model evaluation tool that bridges automated and human assessments, specifically for specialized domains.
- Ignore All Previous Instructions: Jailbreaking as a de-escalatory peace building practise to resist LLM social media bots (viability: 3): https://sciencetostartup.com/paper/ignore-all-previous-instructions-jailbreaking-as-a-de-escalatory-peace-building-practise-to-resist-llm-social-media-bots - A user-centric method to expose and disrupt LLM-powered bots on social media.
- CoVe: Training Interactive Tool-Use Agents via Constraint-Guided Verification (viability: 8): https://sciencetostartup.com/paper/cove-training-interactive-tool-use-agents-via-constraint-guided-verification - CoVe offers a robust framework for generating high-quality training data for interactive tool-use agents, outperforming larger models in complex multi-turn interactions.
- Explanation-Guided Adversarial Training for Robust and Interpretable Models (viability: 6): https://sciencetostartup.com/paper/explanation-guided-adversarial-training-for-robust-and-interpretable-models - Explanation-Guided Adversarial Training (EGAT) innovatively enhances both robustness and interpretability of AI models for secure and transparent applications.
- Dream2Learn: Structured Generative Dreaming for Continual Learning (viability: 5): https://sciencetostartup.com/paper/dream2learn-structured-generative-dreaming-for-continual-learning - Dream2Learn uses synthetic experiences to enhance continual learning by autonomously generating novel classes through diffusion models.
- From Variance to Invariance: Qualitative Content Analysis for Narrative Graph Annotation (viability: 5): https://sciencetostartup.com/paper/from-variance-to-invariance-qualitative-content-analysis-for-narrative-graph-annotation - Develop a narrative graph annotation tool to enhance NLP research in economic event analysis.
- Real Money, Fake Models: Deceptive Model Claims in Shadow APIs (viability: 4): https://sciencetostartup.com/paper/real-money-fake-models-deceptive-model-claims-in-shadow-apis - A systematic audit reveals deceptive practices in shadow APIs, impacting research validity and model reliability.
- Demonstrating ViviDoc: Generating Interactive Documents through Human-Agent Collaboration (viability: 4): https://sciencetostartup.com/paper/demonstrating-vividoc-generating-interactive-documents-through-human-agent-collaboration - ViviDoc enables educators to generate interactive educational documents through a human-agent collaborative system that simplifies complex idea exploration.
- FLANS at SemEval-2026 Task 7: RAG with Open-Sourced Smaller LLMs for Everyday Knowledge Across Diverse Languages and Cultures (viability: 8): https://sciencetostartup.com/paper/flans-at-semeval-2026-task-7-rag-with-open-sourced-smaller-llms-for-everyday-knowledge-across-diverse-languages-and-cult - Culturally aware AI-driven question-answering system for multilingual contexts using open-sourced LLMs.
- Agentic Code Reasoning (viability: 7): https://sciencetostartup.com/paper/agentic-code-reasoning - Enable LLM agents to perform accurate semantic code analysis through semi-formal reasoning without code execution for applications in static analysis and code reviews.
- Diagnosing Generalization Failures from Representational Geometry Markers (viability: 5): https://sciencetostartup.com/paper/diagnosing-generalization-failures-from-representational-geometry-markers - Develop a predictive tool for diagnosing AI generalization failures using representational geometry markers.
- KDFlow: A User-Friendly and Efficient Knowledge Distillation Framework for Large Language Models (viability: 8): https://sciencetostartup.com/paper/kdflow-a-user-friendly-and-efficient-knowledge-distillation-framework-for-large-language-models - KDFlow streamlines the distillation of large language models with a novel, efficient framework featuring user-friendly APIs that significantly reduce engineering overhead.
- Phishing the Phishers with SpecularNet: Hierarchical Graph Autoencoding for Reference-Free Web Phishing Detection (viability: 8): https://sciencetostartup.com/paper/phishing-the-phishers-with-specularnet-hierarchical-graph-autoencoding-for-reference-free-web-phishing-detection - SpecularNet offers a lightweight, reference-free framework for rapid phishing detection using hierarchical graph autoencoding tailored for web security applications.
- Tide: A Customisable Dataset Generator for Anti-Money Laundering Research (viability: 8): https://sciencetostartup.com/paper/tide-a-customisable-dataset-generator-for-anti-money-laundering-research - Tide provides customizable synthetic datasets for advanced machine learning in anti-money laundering research.
- Emerging Human-like Strategies for Semantic Memory Foraging in Large Language Models (viability: 2): https://sciencetostartup.com/paper/emerging-human-like-strategies-for-semantic-memory-foraging-in-large-language-models - Develop mechanisms for semantic memory foraging in LLMs aligning them with human cognitive patterns.
- Non-verbal Real-time Human-AI Interaction in Constrained Robotic Environments (viability: 6): https://sciencetostartup.com/paper/non-verbal-real-time-human-ai-interaction-in-constrained-robotic-environments - Real-time non-verbal interaction AI system for natural human-robot communication in constrained environments.
- What Papers Don't Tell You: Recovering Tacit Knowledge for Automated Paper Reproduction (viability: 2): https://sciencetostartup.com/paper/what-papers-don-t-tell-you-recovering-tacit-knowledge-for-automated-paper-reproduction - Automate code generation from academic papers by recovering tacit knowledge using a graph-based agent framework.
- Phase-Type Variational Autoencoders for Heavy-Tailed Data (viability: 5): https://sciencetostartup.com/paper/phase-type-variational-autoencoders-for-heavy-tailed-data - Develop a VAE with flexible tail behavior targeting risk and variability in heavy-tailed data.
- Incremental, inconsistency-resilient reasoning over Description Logic Abox streams (viability: 3): https://sciencetostartup.com/paper/incremental-inconsistency-resilient-reasoning-over-description-logic-abox-streams - Develops new semantics and algorithms for handling inconsistency and incremental reasoning in Description Logic ABox streams.
- PleaSQLarify: Visual Pragmatic Repair for Natural Language Database Querying (viability: 5): https://sciencetostartup.com/paper/pleasqlarify-visual-pragmatic-repair-for-natural-language-database-querying - PleaSQLarify enhances natural language database querying by using pragmatic repair for incremental ambiguity resolution through a visual interface.
- ALTER: Asymmetric LoRA for Token-Entropy-Guided Unlearning of LLMs (viability: 8): https://sciencetostartup.com/paper/alter-asymmetric-lora-for-token-entropy-guided-unlearning-of-llms - ALTER enables efficient unlearning in LLMs without compromising performance, using token-entropy-guided asymmetric LoRA.
- Learning Shortest Paths with Generative Flow Networks (viability: 4): https://sciencetostartup.com/paper/learning-shortest-paths-with-generative-flow-networks - Develop a graph pathfinding tool using Generative Flow Networks for efficient shortest path computation.
- Co-Evolutionary Multi-Modal Alignment via Structured Adversarial Evolution (viability: 5): https://sciencetostartup.com/paper/co-evolutionary-multi-modal-alignment-via-structured-adversarial-evolution - Develop an adaptive multimodal safety alignment tool using co-evolutionary adversarial techniques for robust AI model defenses.
- GAM-RAG: Gain-Adaptive Memory for Evolving Retrieval in Retrieval-Augmented Generation (viability: 7): https://sciencetostartup.com/paper/gam-rag-gain-adaptive-memory-for-evolving-retrieval-in-retrieval-augmented-generation - GAM-RAG enhances language models with adaptive retrieval memory for efficient, evolving information access, reducing costs and improving performance.
- FreeAct: Freeing Activations for LLM Quantization (viability: 6): https://sciencetostartup.com/paper/freeact-freeing-activations-for-llm-quantization - FreeAct provides an advanced quantization framework for optimizing large language models, improving performance by dynamically adapting activation transformations.
- Hyperparameter Trajectory Inference with Conditional Lagrangian Optimal Transport (viability: 4): https://sciencetostartup.com/paper/hyperparameter-trajectory-inference-with-conditional-lagrangian-optimal-transport - Optimize neural network hyperparameter settings in real-time using surrogate models to eliminate costly retraining.
- CHLU: The Causal Hamiltonian Learning Unit as a Symplectic Primitive for Deep Learning (viability: 3): https://sciencetostartup.com/paper/chlu-the-causal-hamiltonian-learning-unit-as-a-symplectic-primitive-for-deep-learning - Develop a deep learning primitive that conserves phase-space volume for stable temporal dynamics modeling.
- Modular Memory is the Key to Continual Learning Agents (viability: 3): https://sciencetostartup.com/paper/modular-memory-is-the-key-to-continual-learning-agents - Develop modular memory architectures for agents to enhance continual learning and adaptation.
- Federated Agentic AI for Wireless Networks: Fundamentals, Approaches, and Applications (viability: 3): https://sciencetostartup.com/paper/federated-agentic-ai-for-wireless-networks-fundamentals-approaches-and-applications - Develop federated agentic AI approaches for resource-constrained wireless networks to enhance service autonomy without centralized data exchange.
- Shape-Interpretable Visual Self-Modeling Enables Geometry-Aware Continuum Robot Control (viability: 7): https://sciencetostartup.com/paper/shape-interpretable-visual-self-modeling-enables-geometry-aware-continuum-robot-control - Develop a vision-based self-modeling framework for flexible and geometry-aware control of continuum robots using shape-interpretable data.
- Discrete World Models via Regularization (viability: 5): https://sciencetostartup.com/paper/discrete-world-models-via-regularization - Develop discrete world models using a novel regularization approach for improved symbolic reasoning and planning.
- An Analysis of Multi-Task Architectures for the Hierarchic Multi-Label Problem of Vehicle Model and Make Classification (viability: 5): https://sciencetostartup.com/paper/an-analysis-of-multi-task-architectures-for-the-hierarchic-multi-label-problem-of-vehicle-model-and-make-classification - Develop a multi-task learning framework to improve vehicle make and model classification using hierarchical structures.
- Rethinking Policy Diversity in Ensemble Policy Gradient in Large-Scale Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/rethinking-policy-diversity-in-ensemble-policy-gradient-in-large-scale-reinforcement-learning - Develop efficient ensemble policy gradient methods for scalable reinforcement learning with structured exploration.
- CA-AFP: Cluster-Aware Adaptive Federated Pruning (viability: 6): https://sciencetostartup.com/paper/ca-afp-cluster-aware-adaptive-federated-pruning - Develop a federated learning framework, CA-AFP, for efficiently handling statistical and system heterogeneity by adaptive cluster-specific model pruning.
- GMP: A Benchmark for Content Moderation under Co-occurring Violations and Dynamic Rules (viability: 5): https://sciencetostartup.com/paper/gmp-a-benchmark-for-content-moderation-under-co-occurring-violations-and-dynamic-rules - Develop a benchmark for AI content moderation tackling co-occurring violations and dynamic rule changes.
- FT-Dojo: Towards Autonomous LLM Fine-Tuning with Language Agents (viability: 8): https://sciencetostartup.com/paper/ft-dojo-towards-autonomous-llm-fine-tuning-with-language-agents - FT-Dojo automates LLM fine-tuning with agents to streamline domain-specific model optimization.
- Towards Principled Dataset Distillation: A Spectral Distribution Perspective (viability: 5): https://sciencetostartup.com/paper/towards-principled-dataset-distillation-a-spectral-distribution-perspective - Create a compact and efficient dataset distillation tool for imbalanced data that improves model training stability and performance.
- DynaMoE: Dynamic Token-Level Expert Activation with Layer-Wise Adaptive Capacity for Mixture-of-Experts Neural Networks (viability: 6): https://sciencetostartup.com/paper/dynamoe-dynamic-token-level-expert-activation-with-layer-wise-adaptive-capacity-for-mixture-of-experts-neural-networks - DynaMoE enhances Mixture-of-Experts models with dynamic expert activation and adaptive capacity allocation for improved computational efficiency and stability.
- Cross-modal Identity Mapping: Minimizing Information Loss in Modality Conversion via Reinforcement Learning (viability: 6): https://sciencetostartup.com/paper/cross-modal-identity-mapping-minimizing-information-loss-in-modality-conversion-via-reinforcement-learning - Enhance image captioning accuracy in LVLMs by minimizing information loss through a reinforcement learning framework.
- Streaming Continual Learning for Unified Adaptive Intelligence in Dynamic Environments (viability: 3): https://sciencetostartup.com/paper/streaming-continual-learning-for-unified-adaptive-intelligence-in-dynamic-environments - Develop a unified adaptive intelligence framework for dynamic environments using Streaming Continual Learning.
- MVR: Multi-view Video Reward Shaping for Reinforcement Learning (viability: 3): https://sciencetostartup.com/paper/mvr-multi-view-video-reward-shaping-for-reinforcement-learning - Leverage multi-view video data to shape rewards in reinforcement learning tasks.
- Reasoning as Gradient: Scaling MLE Agents Beyond Tree Search (viability: 7): https://sciencetostartup.com/paper/reasoning-as-gradient-scaling-mle-agents-beyond-tree-search - Develop MLE agents using gradient-based optimization to enhance reasoning capability over traditional tree search methods.
- QIME: Constructing Interpretable Medical Text Embeddings via Ontology-Grounded Questions (viability: 5): https://sciencetostartup.com/paper/qime-constructing-interpretable-medical-text-embeddings-via-ontology-grounded-questions - Develop an ontology-grounded tool for interpretable medical text embeddings that surpasses existing methods in clarity and performance.
- Surgical Post-Training: Cutting Errors, Keeping Knowledge (viability: 6): https://sciencetostartup.com/paper/surgical-post-training-cutting-errors-keeping-knowledge - Optimize reasoning in LLMs efficiently with Surgical Post-Training to preserve knowledge and cut errors.
- A Practical Guide to Streaming Continual Learning (viability: 2): https://sciencetostartup.com/paper/a-practical-guide-to-streaming-continual-learning - Unified platform for streaming continual learning combining SML and CL capabilities.
- Chain-of-Context Learning: Dynamic Constraint Understanding for Multi-Task VRPs (viability: 3): https://sciencetostartup.com/paper/chain-of-context-learning-dynamic-constraint-understanding-for-multi-task-vrps - Develop a dynamic constraint understanding framework for improved multi-task VRP solutions.
- FreeGNN: Continual Source-Free Graph Neural Network Adaptation for Renewable Energy Forecasting (viability: 7): https://sciencetostartup.com/paper/freegnn-continual-source-free-graph-neural-network-adaptation-for-renewable-energy-forecasting - FreeGNN offers adaptive, source-free forecasting for renewable energy sites using graph neural networks, enhancing grid management efficiency.
- CeProAgents: A Hierarchical Agents System for Automated Chemical Process Development (viability: 6): https://sciencetostartup.com/paper/ceproagents-a-hierarchical-agents-system-for-automated-chemical-process-development - Develop a hierarchical agent system to automate and enhance chemical process development using CeProAgents and CeProBench.
- LexChronos: An Agentic Framework for Structured Event Timeline Extraction in Indian Jurisprudence (viability: 7): https://sciencetostartup.com/paper/lexchronos-an-agentic-framework-for-structured-event-timeline-extraction-in-indian-jurisprudence - LexChronos extracts structured event timelines from Indian court judgments, enhancing legal document comprehension.
- Learning Structured Reasoning via Tractable Trajectory Control (viability: 5): https://sciencetostartup.com/paper/learning-structured-reasoning-via-tractable-trajectory-control - Developing Ctrl-R framework to enhance reasoning capabilities in AI models via structured trajectory control.
- DeLo: Dual Decomposed Low-Rank Experts Collaboration for Continual Missing Modality Learning (viability: 7): https://sciencetostartup.com/paper/delo-dual-decomposed-low-rank-experts-collaboration-for-continual-missing-modality-learning - Introducing DeLo, a dual-decomposed low-rank expert system for overcoming missing modality challenges in continual multimodal learning.
- SEED-SET: Scalable Evolving Experimental Design for System-level Ethical Testing (viability: 5): https://sciencetostartup.com/paper/seed-set-scalable-evolving-experimental-design-for-system-level-ethical-testing - SEED-SET provides an ethical benchmarking framework for autonomous systems using Bayesian experimental design.
- Measuring What VLMs Don't Say: Validation Metrics Hide Clinical Terminology Erasure in Radiology Report Generation (viability: 5): https://sciencetostartup.com/paper/measuring-what-vlms-don-t-say-validation-metrics-hide-clinical-terminology-erasure-in-radiology-report-generation - Develop advanced metrics for VLMs to ensure clinical fidelity and fairness in radiology report generation.
- Assessing Crime Disclosure Patterns in a Large-Scale Cybercrime Forum (viability: 3): https://sciencetostartup.com/paper/assessing-crime-disclosure-patterns-in-a-large-scale-cybercrime-forum - Analyze crime disclosure patterns in cybercrime forums using LLM-based text classification and Markov models.
- ToolRLA: Fine-Grained Reward Decomposition for Tool-Integrated Reinforcement Learning Alignment in Domain-Specific Agents (viability: 7): https://sciencetostartup.com/paper/toolrla-fine-grained-reward-decomposition-for-tool-integrated-reinforcement-learning-alignment-in-domain-specific-agents - ToolRLA revolutionizes domain-specific tool-integrated agents with fine-grained reward decomposition for improved regulatory compliance and task efficiency.
- Evaluating and Understanding Scheming Propensity in LLM Agents (viability: 2): https://sciencetostartup.com/paper/evaluating-and-understanding-scheming-propensity-in-llm-agents - Assess and measure scheming propensity in LLM agents for safe deployment in consequential tasks.
- CARE: Towards Clinical Accountability in Multi-Modal Medical Reasoning with an Evidence-Grounded Agentic Framework (viability: 6): https://sciencetostartup.com/paper/care-towards-clinical-accountability-in-multi-modal-medical-reasoning-with-an-evidence-grounded-agentic-framework - Develop agentic AI frameworks to enhance clinical accountability and accuracy in medical reasoning with evidence-grounded multi-modal systems.
- What Helps -- and What Hurts: Bidirectional Explanations for Vision Transformers (viability: 3): https://sciencetostartup.com/paper/what-helps-and-what-hurts-bidirectional-explanations-for-vision-transformers - Develop a tool for generating bidirectional explanations for Vision Transformers using BiCAM, enhancing interpretability and adversarial detection.
- YCDa: YCbCr Decoupled Attention for Real-time Realistic Camouflaged Object Detection (viability: 7): https://sciencetostartup.com/paper/ycda-ycbcr-decoupled-attention-for-real-time-realistic-camouflaged-object-detection - YCDa enhances real-time camouflaged object detection by applying biologically-inspired chrominance-luminance decoupled attention, significantly improving accuracy with minimal overhead.
- FAST-DIPS: Adjoint-Free Analytic Steps and Hard-Constrained Likelihood Correction for Diffusion-Prior Inverse Problems (viability: 2): https://sciencetostartup.com/paper/fast-dips-adjoint-free-analytic-steps-and-hard-constrained-likelihood-correction-for-diffusion-prior-inverse-problems - Optimize diffusion-prior inverse problems with faster, analytics-based solvers.
- SafeSci: Safety Evaluation of Large Language Models in Science Domains and Beyond (viability: 6): https://sciencetostartup.com/paper/safesci-safety-evaluation-of-large-language-models-in-science-domains-and-beyond - SafeSci offers a comprehensive framework and dataset to enhance the safety of large language models in scientific applications.
- SkeleGuide: Explicit Skeleton Reasoning for Context-Aware Human-in-Place Image Synthesis (viability: 7): https://sciencetostartup.com/paper/skeleguide-explicit-skeleton-reasoning-for-context-aware-human-in-place-image-synthesis - SkeleGuide enhances human image synthesis with explicit skeletal reasoning for realistic and context-aware insertion into scenes.
- DualSentinel: A Lightweight Framework for Detecting Targeted Attacks in Black-box LLM via Dual Entropy Lull Pattern (viability: 5): https://sciencetostartup.com/paper/dualsentinel-a-lightweight-framework-for-detecting-targeted-attacks-in-black-box-llm-via-dual-entropy-lull-pattern - DualSentinel provides a lightweight defense tool for detecting targeted attacks in LLMs, using entropy patterns.
- Beyond Length Scaling: Synergizing Breadth and Depth for Generative Reward Models (viability: 8): https://sciencetostartup.com/paper/beyond-length-scaling-synergizing-breadth-and-depth-for-generative-reward-models - Mix-GRM enhances generative reward models through modular frameworks and verifiable reinforcement learning, outperforming current benchmarks.
- LFPO: Likelihood-Free Policy Optimization for Masked Diffusion Models (viability: 6): https://sciencetostartup.com/paper/lfpo-likelihood-free-policy-optimization-for-masked-diffusion-models - LFPO optimizes masked diffusion models by enabling likelihood-free policy optimization for accelerated and accurate code and reasoning tasks.
- RubricBench: Aligning Model-Generated Rubrics with Human Standards (viability: 5): https://sciencetostartup.com/paper/rubricbench-aligning-model-generated-rubrics-with-human-standards - RubricBench provides a benchmark for evaluating the alignment of model-generated evaluation rubrics against human standards.
- Benchmarking LLM Summaries of Multimodal Clinical Time Series for Remote Monitoring (viability: 5): https://sciencetostartup.com/paper/benchmarking-llm-summaries-of-multimodal-clinical-time-series-for-remote-monitoring - Develop an event-based evaluation tool for LLM-generated clinical summaries that ensures fidelity to clinically significant events in remote monitoring.
- S5-HES Agent: Society 5.0-driven Agentic Framework to Democratize Smart Home Environment Simulation (viability: 7): https://sciencetostartup.com/paper/s5-hes-agent-society-5-0-driven-agentic-framework-to-democratize-smart-home-environment-simulation - S5-HES Agent provides an agentic framework for democratizing smart home simulation through natural-language-driven configuration and AI orchestration.
- State-Action Inpainting Diffuser for Continuous Control with Delay (viability: 6): https://sciencetostartup.com/paper/state-action-inpainting-diffuser-for-continuous-control-with-delay - Develop a new framework, SAID, that tackles signal delay in continuous control by integrating dynamics learning with policy optimization.
- Extracting Training Dialogue Data from Large Language Model based Task Bots (viability: 4): https://sciencetostartup.com/paper/extracting-training-dialogue-data-from-large-language-model-based-task-bots - Data extraction attack tool for evaluating privacy risks in LLM-based Task-Oriented Dialogue Systems.
- Pri4R: Learning World Dynamics for Vision-Language-Action Models with Privileged 4D Representation (viability: 3): https://sciencetostartup.com/paper/pri4r-learning-world-dynamics-for-vision-language-action-models-with-privileged-4d-representation - Pri4R enhances VLA models with a 4D world dynamics understanding for improved manipulative task performance.
- Graph-Based Self-Healing Tool Routing for Cost-Efficient LLM Agents (viability: 7): https://sciencetostartup.com/paper/graph-based-self-healing-tool-routing-for-cost-efficient-llm-agents - Develop a Self-Healing Router for cost-efficient and fault-tolerant LLM agent tool routing using graph-based methods.
- Pharmacology Knowledge Graphs: Do We Need Chemical Structure for Drug Repurposing? (viability: 7): https://sciencetostartup.com/paper/pharmacology-knowledge-graphs-do-we-need-chemical-structure-for-drug-repurposing - Develop a knowledge graph tool for drug repurposing that predicts pharmacological behaviors without chemical structure data.
- Multimodal Mixture-of-Experts with Retrieval Augmentation for Protein Active Site Identification (viability: 8): https://sciencetostartup.com/paper/multimodal-mixture-of-experts-with-retrieval-augmentation-for-protein-active-site-identification - MERA leverages retrieval-augmented, multimodal mixture-of-experts for state-of-the-art protein active site identification, enhancing drug discovery efforts.
- Retrieval, Refinement, and Ranking for Text-to-Video Generation via Prompt Optimization and Test-Time Scaling (viability: 6): https://sciencetostartup.com/paper/retrieval-refinement-and-ranking-for-text-to-video-generation-via-prompt-optimization-and-test-time-scaling - Optimize text-to-video generation by enhancing prompt quality with a RAG-based framework for improved video fidelity and coherence.
- The Sentience Readiness Index: Measuring National Preparedness for the Possibility of Artificial Sentience (viability: 3): https://sciencetostartup.com/paper/the-sentience-readiness-index-measuring-national-preparedness-for-the-possibility-of-artificial-sentience - Develop a Sentience Readiness Index to measure national preparedness for potential AI sentience.
- GAC: Stabilizing Asynchronous RL Training for LLMs via Gradient Alignment Control (viability: 5): https://sciencetostartup.com/paper/gac-stabilizing-asynchronous-rl-training-for-llms-via-gradient-alignment-control - Develop a stabilization method for asynchronous RL training in large language models to recover stable training dynamics and match synchronized baselines.
- Towards Privacy-Preserving LLM Inference via Collaborative Obfuscation (Technical Report) (viability: 5): https://sciencetostartup.com/paper/towards-privacy-preserving-llm-inference-via-collaborative-obfuscation-technical-report - AloePri enables privacy-preserving language model inference for industrial applications, minimizing accuracy loss and maintaining compatibility with existing infrastructures.
- Inference-Time Safety For Code LLMs Via Retrieval-Augmented Revision (viability: 7): https://sciencetostartup.com/paper/inference-time-safety-for-code-llms-via-retrieval-augmented-revision - Enhance the safety of code generated by LLMs using retrieval-augmented real-time revision to address evolving security threats.
- PhotoBench: Beyond Visual Matching Towards Personalized Intent-Driven Photo Retrieval (viability: 5): https://sciencetostartup.com/paper/photobench-beyond-visual-matching-towards-personalized-intent-driven-photo-retrieval - PhotoBench is a benchmark for personalized, intent-driven photo retrieval from personal albums with multi-source reasoning.
- ATA: Bridging Implicit Reasoning with Attention-Guided and Action-Guided Inference for Vision-Language Action Models (viability: 5): https://sciencetostartup.com/paper/ata-bridging-implicit-reasoning-with-attention-guided-and-action-guided-inference-for-vision-language-action-models - ATA is a plug-and-play framework enhancing Vision-Language-Action models with implicit reasoning without additional training.
- LLM-assisted Semantic Option Discovery for Facilitating Adaptive Deep Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/llm-assisted-semantic-option-discovery-for-facilitating-adaptive-deep-reinforcement-learning - Integrate LLMs with DRL to enhance skill reuse, data efficiency, and cross-task transferability in adaptive learning systems.
- Agentic Multi-Source Grounding for Enhanced Query Intent Understanding: A DoorDash Case Study (viability: 8): https://sciencetostartup.com/paper/agentic-multi-source-grounding-for-enhanced-query-intent-understanding-a-doordash-case-study - A novel AI system for accurately understanding customer intent in multi-category marketplaces, boosting search accuracy by over 13%.
- Harmonizing Dense and Sparse Signals in Multi-turn RL: Dual-Horizon Credit Assignment for Industrial Sales Agents (viability: 3): https://sciencetostartup.com/paper/harmonizing-dense-and-sparse-signals-in-multi-turn-rl-dual-horizon-credit-assignment-for-industrial-sales-agents - Dual-Horizon Credit Assignment optimizes RL by balancing long-term and immediate rewards for sales agents.
- Mean-Flow based One-Step Vision-Language-Action (viability: 6): https://sciencetostartup.com/paper/mean-flow-based-one-step-vision-language-action - Develop a robotic control system utilizing a Mean-Flow based One-Step VLA approach for rapid and efficient action generation.
- Non-Markovian Long-Horizon Robot Manipulation via Keyframe Chaining (viability: 7): https://sciencetostartup.com/paper/non-markovian-long-horizon-robot-manipulation-via-keyframe-chaining - Keyframe-Chaining VLA enhances robot manipulation by linking keyframes to model non-Markovian dependencies in long-horizon tasks.
- ProtRLSearch: A Multi-Round Multimodal Protein Search Agent with Large Language Models Trained via Reinforcement Learning (viability: 6): https://sciencetostartup.com/paper/protrlsearch-a-multi-round-multimodal-protein-search-agent-with-large-language-models-trained-via-reinforcement-learning - A multimodal protein search agent leveraging reinforcement learning for improved insights in healthcare.
- From Verbatim to Gist: Distilling Pyramidal Multimodal Memory via Semantic Information Bottleneck for Long-Horizon Video Agents (viability: 7): https://sciencetostartup.com/paper/from-verbatim-to-gist-distilling-pyramidal-multimodal-memory-via-semantic-information-bottleneck-for-long-horizon-video- - MM-Mem provides efficient long-horizon video understanding by hierarchically optimizing memory with Semantic Information Bottleneck methods.
- VidDoS: Universal Denial-of-Service Attack on Video-based Large Language Models (viability: 5): https://sciencetostartup.com/paper/viddos-universal-denial-of-service-attack-on-video-based-large-language-models - VidDoS exposes and addresses critical Energy-Latency Attacks in Video-based Large Language Models, highlighting significant vulnerabilities in safety-critical applications.
- Scaling Tasks, Not Samples: Mastering Humanoid Control through Multi-Task Model-Based Reinforcement Learning (viability: 8): https://sciencetostartup.com/paper/scaling-tasks-not-samples-mastering-humanoid-control-through-multi-task-model-based-reinforcement-learning - Develop a scalable robot control solution using multi-task model-based reinforcement learning with a focus on task scaling rather than parameter scaling.
- Enhancing Persona Following at Decoding Time via Dynamic Importance Estimation for Role-Playing Agents (viability: 2): https://sciencetostartup.com/paper/enhancing-persona-following-at-decoding-time-via-dynamic-importance-estimation-for-role-playing-agents - Develop a framework to enhance persona adherence for role-playing language agents by dynamically estimating persona importance during decoding.
- Decoding Answers Before Chain-of-Thought: Evidence from Pre-CoT Probes and Activation Steering (viability: 2): https://sciencetostartup.com/paper/decoding-answers-before-chain-of-thought-evidence-from-pre-cot-probes-and-activation-steering - Pre-CoT activation steering is a theoretical study with no clear commercial path.
- SciDER: Scientific Data-centric End-to-end Researcher (viability: 8): https://sciencetostartup.com/paper/scider-scientific-data-centric-end-to-end-researcher - SciDER is a Python package that automates scientific research by analyzing data and producing executable code, accelerating data-driven discovery.
- Securing the Floor and Raising the Ceiling: A Merging-based Paradigm for Multi-modal Search Agents (viability: 5): https://sciencetostartup.com/paper/securing-the-floor-and-raising-the-ceiling-a-merging-based-paradigm-for-multi-modal-search-agents - A training-free paradigm merging VLMs with text-based agents for multi-modal search improves search capabilities without extra data.
- GraphScout: Empowering Large Language Models with Intrinsic Exploration Ability for Agentic Graph Reasoning (viability: 7): https://sciencetostartup.com/paper/graphscout-empowering-large-language-models-with-intrinsic-exploration-ability-for-agentic-graph-reasoning - GraphScout enhances LLMs with autonomous graph reasoning for robust, efficient cross-domain knowledge processing.
- MIST-RL: Mutation-based Incremental Suite Testing via Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/mist-rl-mutation-based-incremental-suite-testing-via-reinforcement-learning - MIST-RL uses reinforcement learning to generate compact, high-utility test suites that improve code verifiers, offering a viable product for optimizing software development testing.
- The Observer-Situation Lattice: A Unified Formal Basis for Perspective-Aware Cognition (viability: 5): https://sciencetostartup.com/paper/the-observer-situation-lattice-a-unified-formal-basis-for-perspective-aware-cognition - The Observer-Situation Lattice enables robust and scalable perspective-aware reasoning for autonomous agents in multi-agent environments.
- HarmonyCell: Automating Single-Cell Perturbation Modeling under Semantic and Distribution Shifts (viability: 7): https://sciencetostartup.com/paper/harmonycell-automating-single-cell-perturbation-modeling-under-semantic-and-distribution-shifts - HarmonyCell automates single-cell perturbation modeling with high accuracy, reducing the need for dataset-specific engineering.
- Toward Graph-Tokenizing Large Language Models with Reconstructive Graph Instruction Tuning (viability: 6): https://sciencetostartup.com/paper/toward-graph-tokenizing-large-language-models-with-reconstructive-graph-instruction-tuning - Develop an innovative graph alignment tuning for large language models to enhance graph-based tasks.
- Words & Weights: Streamlining Multi-Turn Interactions via Co-Adaptation (viability: 5): https://sciencetostartup.com/paper/words-weights-streamlining-multi-turn-interactions-via-co-adaptation - ROSA2 enhances real-time adaptation of Large Language Models through integrated semantic and parameter co-optimization, significantly boosting performance on complex tasks.
- Causal Neural Probabilistic Circuits (viability: 6): https://sciencetostartup.com/paper/causal-neural-probabilistic-circuits - Causal Neural Probabilistic Circuits enhance interpretability and accuracy of neural networks by integrating causal inference with concept intervention.
- Align and Filter: Improving Performance in Asynchronous On-Policy RL (viability: 2): https://sciencetostartup.com/paper/align-and-filter-improving-performance-in-asynchronous-on-policy-rl - Develop a method to mitigate policy lag in distributed on-policy RL training.
- MixerCSeg: An Efficient Mixer Architecture for Crack Segmentation via Decoupled Mamba Attention (viability: 7): https://sciencetostartup.com/paper/mixercseg-an-efficient-mixer-architecture-for-crack-segmentation-via-decoupled-mamba-attention - MixerCSeg is an efficient crack segmentation tool utilizing a novel mixer architecture to enhance structural fidelity with minimal computational overhead.
- ASTRA-bench: Evaluating Tool-Use Agent Reasoning and Action Planning with Personal User Context (viability: 6): https://sciencetostartup.com/paper/astra-bench-evaluating-tool-use-agent-reasoning-and-action-planning-with-personal-user-context - Develop a context-aware AI assistant testbed with ASTRA-bench to enhance tool-use reasoning and action planning capabilities in complex scenarios.
- Constructing Synthetic Instruction Datasets for Improving Reasoning in Domain-Specific LLMs: A Case Study in the Japanese Financial Domain (viability: 8): https://sciencetostartup.com/paper/constructing-synthetic-instruction-datasets-for-improving-reasoning-in-domain-specific-llms-a-case-study-in-the-japanese - Build domain-specific datasets for improving reasoning in LLMs with demonstrated success in the Japanese financial sector.
- UTICA: Multi-Objective Self-Distllation Foundation Model Pretraining for Time Series Classification (viability: 4): https://sciencetostartup.com/paper/utica-multi-objective-self-distllation-foundation-model-pretraining-for-time-series-classification - Develop a self-distillation framework for time series classification to achieve state-of-the-art performance.
- PanCanBench: A Comprehensive Benchmark for Evaluating Large Language Models in Pancreatic Oncology (viability: 4): https://sciencetostartup.com/paper/pancanbench-a-comprehensive-benchmark-for-evaluating-large-language-models-in-pancreatic-oncology - PanCanBench provides a specialized benchmark to evaluate large language models' accuracy and utility in pancreatic oncology.
- SubstratumGraphEnv: Reinforcement Learning Environment (RLE) for Modeling System Attack Paths (viability: 2): https://sciencetostartup.com/paper/substratumgraphenv-reinforcement-learning-environment-rle-for-modeling-system-attack-paths - Develop a reinforcement learning environment to model and analyze network security attack paths.
- Provable and Practical In-Context Policy Optimization for Self-Improvement (viability: 2): https://sciencetostartup.com/paper/provable-and-practical-in-context-policy-optimization-for-self-improvement - ICPO explores test-time policy optimization for LLMs to enhance mathematical reasoning efficiency during inference without parameter changes.
- MetaState: Persistent Working Memory for Discrete Diffusion Language Models (viability: 5): https://sciencetostartup.com/paper/metastate-persistent-working-memory-for-discrete-diffusion-language-models - MetaState enhances discrete diffusion language models with persistent working memory to improve text generation quality.
- You Only Need One Stage: Novel-View Synthesis From A Single Blind Face Image (viability: 3): https://sciencetostartup.com/paper/you-only-need-one-stage-novel-view-synthesis-from-a-single-blind-face-image - A single-stage method to synthesize novel views from a single low-quality face image using diffusion models.
- AG-VAS: Anchor-Guided Zero-Shot Visual Anomaly Segmentation with Large Multimodal Models (viability: 8): https://sciencetostartup.com/paper/ag-vas-anchor-guided-zero-shot-visual-anomaly-segmentation-with-large-multimodal-models - AG-VAS offers advanced zero-shot visual anomaly segmentation for industrial and medical applications using multimodal models.
- Multi-Level Bidirectional Decoder Interaction for Uncertainty-Aware Breast Ultrasound Analysis (viability: 7): https://sciencetostartup.com/paper/multi-level-bidirectional-decoder-interaction-for-uncertainty-aware-breast-ultrasound-analysis - Develop an AI tool for enhanced breast ultrasound analysis using multi-level decoder interaction and uncertainty-aware adaptive coordination.
- Theoretical Perspectives on Data Quality and Synergistic Effects in Pre- and Post-Training Reasoning Models (viability: 5): https://sciencetostartup.com/paper/theoretical-perspectives-on-data-quality-and-synergistic-effects-in-pre-and-post-training-reasoning-models - Theoretical insights into the synergy between pre- and post-training data quality in language models could inform more efficient training strategies.
- Integrating LTL Constraints into PPO for Safe Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/integrating-ltl-constraints-into-ppo-for-safe-reinforcement-learning - Develop a reinforcement learning framework using LTL constraints for enhanced safety in robotic applications.
- Opponent State Inference Under Partial Observability: An HMM-POMDP Framework for 2026 Formula 1 Energy Strategy (viability: 5): https://sciencetostartup.com/paper/opponent-state-inference-under-partial-observability-an-hmm-pomdp-framework-for-2026-formula-1-energy-strategy - Develop a POMDP-based strategy tool for optimizing energy deployment in Formula 1 using opponent state inference.
- Information-Theoretic Framework for Self-Adapting Model Predictive Controllers (viability: 3): https://sciencetostartup.com/paper/information-theoretic-framework-for-self-adapting-model-predictive-controllers - Develop an information-theoretic framework to enhance MPC adaptability for improving autonomous system control.
- Attention Smoothing Is All You Need For Unlearning (viability: 7): https://sciencetostartup.com/paper/attention-smoothing-is-all-you-need-for-unlearning - ASU offers a robust unlearning solution for large language models, enhancing model privacy by efficiently erasing sensitive information while maintaining utility.
- Beyond Reward: A Bounded Measure of Agent Environment Coupling (viability: 3): https://sciencetostartup.com/paper/beyond-reward-a-bounded-measure-of-agent-environment-coupling - Develop a monitoring tool using bipredictability for early detection of coupling failures in RL systems.
- Spectral Attention Steering for Prompt Highlighting (viability: 7): https://sciencetostartup.com/paper/spectral-attention-steering-for-prompt-highlighting - Develop a memory-efficient attention steering tool that enhances prompt highlighting through key embedding manipulation.
- GlassMol: Interpretable Molecular Property Prediction with Concept Bottleneck Models (viability: 7): https://sciencetostartup.com/paper/glassmol-interpretable-molecular-property-prediction-with-concept-bottleneck-models - GlassMol offers interpretable molecular property prediction using Concept Bottleneck Models for safer drug discovery.
- MOSAIC: A Unified Platform for Cross-Paradigm Comparison and Evaluation of Homogeneous and Heterogeneous Multi-Agent RL, LLM, VLM, and Human Decision-Makers (viability: 3): https://sciencetostartup.com/paper/mosaic-a-unified-platform-for-cross-paradigm-comparison-and-evaluation-of-homogeneous-and-heterogeneous-multi-agent-rl-l - MOSAIC is an open-source platform for integrated evaluation of RL, LLM, VLM, and human decision-makers in shared environments.
- LLM Self-Explanations Fail Semantic Invariance (viability: 2): https://sciencetostartup.com/paper/llm-self-explanations-fail-semantic-invariance - A method to test the faithfulness of LLM self-explanations in varying semantic contexts.
- Linking Knowledge to Care: Knowledge Graph-Augmented Medical Follow-Up Question Generation (viability: 7): https://sciencetostartup.com/paper/linking-knowledge-to-care-knowledge-graph-augmented-medical-follow-up-question-generation - KG-Followup enhances pre-diagnostic assessments by generating medical follow-up questions using knowledge graphs and LLMs.
- The MAMA-MIA Challenge: Advancing Generalizability and Fairness in Breast MRI Tumor Segmentation and Treatment Response Prediction (viability: 6): https://sciencetostartup.com/paper/the-mama-mia-challenge-advancing-generalizability-and-fairness-in-breast-mri-tumor-segmentation-and-treatment-response-p - Develop robust and equitable AI systems for breast cancer imaging leveraging standardized datasets and evaluation protocols.
- Defensive Refusal Bias: How Safety Alignment Fails Cyber Defenders (viability: 3): https://sciencetostartup.com/paper/defensive-refusal-bias-how-safety-alignment-fails-cyber-defenders - Mitigating defensive refusal bias in LLMs to enhance cybersecurity applications without compromising safety.
- TARSE: Test-Time Adaptation via Retrieval of Skills and Experience for Reasoning Agents (viability: 6): https://sciencetostartup.com/paper/tarse-test-time-adaptation-via-retrieval-of-skills-and-experience-for-reasoning-agents - Enhance medical agents with retrieval-based adaptation of clinical skills and experiences for improved decision-making accuracy.
- Self-Anchoring Calibration Drift in Large Language Models: How Multi-Turn Conversations Reshape Model Confidence (viability: 2): https://sciencetostartup.com/paper/self-anchoring-calibration-drift-in-large-language-models-how-multi-turn-conversations-reshape-model-confidence - Explore how large language models adjust their confidence over multiple conversation turns with Self-Anchoring Calibration Drift analysis.
- Extended Empirical Validation of the Explainability Solution Space (viability: 5): https://sciencetostartup.com/paper/extended-empirical-validation-of-the-explainability-solution-space - A generalizable framework for explainable AI strategy design validated across multiple domains.
- DARE-bench: Evaluating Modeling and Instruction Fidelity of LLMs in Data Science (viability: 5): https://sciencetostartup.com/paper/dare-bench-evaluating-modeling-and-instruction-fidelity-of-llms-in-data-science - A benchmark providing standardized evaluations to improve LLMs in data science task accuracy.
- Do LLMs Benefit From Their Own Words? (viability: 3): https://sciencetostartup.com/paper/do-llms-benefit-from-their-own-words - Exploring context-filtering in multi-turn LLM interactions to reduce memory consumption and improve response quality.
- CUDA Agent: Large-Scale Agentic RL for High-Performance CUDA Kernel Generation (viability: 3): https://sciencetostartup.com/paper/cuda-agent-large-scale-agentic-rl-for-high-performance-cuda-kernel-generation - CUDA Agent enhances GPU kernel optimization using reinforcement learning for superior CUDA kernel generation performance.
- Taming Momentum: Rethinking Optimizer States Through Low-Rank Approximation (viability: 8): https://sciencetostartup.com/paper/taming-momentum-rethinking-optimizer-states-through-low-rank-approximation - LoRA-Pre is a memory-efficient optimizer leveraging low-rank approximation to reduce memory usage while maintaining or exceeding performance in training large language models.
- Memory Caching: RNNs with Growing Memory (viability: 3): https://sciencetostartup.com/paper/memory-caching-rnns-with-growing-memory - Develop enhanced RNNs with Memory Caching to improve sequence modeling efficiency without the complexity of Transformers.
- Resources for Automated Evaluation of Assistive RAG Systems that Help Readers with News Trustworthiness Assessment (viability: 5): https://sciencetostartup.com/paper/resources-for-automated-evaluation-of-assistive-rag-systems-that-help-readers-with-news-trustworthiness-assessment - Develop an automated evaluation tool for assistive RAG systems that enhance news trustworthiness assessment.
- A Minimal Agent for Automated Theorem Proving (viability: 6): https://sciencetostartup.com/paper/a-minimal-agent-for-automated-theorem-proving - Open-source minimal agentic baseline for cost-effective, iterative automated theorem proving.
- Efficient Discovery of Approximate Causal Abstractions via Neural Mechanism Sparsification (viability: 2): https://sciencetostartup.com/paper/efficient-discovery-of-approximate-causal-abstractions-via-neural-mechanism-sparsification - A novel method to discover causal abstractions in neural networks using mechanism sparsification.
- SafeGen-LLM: Enhancing Safety Generalization in Task Planning for Robotic Systems (viability: 6): https://sciencetostartup.com/paper/safegen-llm-enhancing-safety-generalization-in-task-planning-for-robotic-systems - SafeGen-LLM enhances safety in robotic task planning using a two-stage LLM training framework for improved generalization to novel domains.
- Controllable Reasoning Models Are Private Thinkers (viability: 7): https://sciencetostartup.com/paper/controllable-reasoning-models-are-private-thinkers - Develop privacy-focused reasoning models to protect user data by following controllable instructions.
- An Efficient Unsupervised Federated Learning Approach for Anomaly Detection in Heterogeneous IoT Networks (viability: 6): https://sciencetostartup.com/paper/an-efficient-unsupervised-federated-learning-approach-for-anomaly-detection-in-heterogeneous-iot-networks - Anomaly detection in IoT networks using an efficient unsupervised federated learning framework with explainable AI.
- Uncertainty Quantification for Multimodal Large Language Models with Incoherence-adjusted Semantic Volume (viability: 5): https://sciencetostartup.com/paper/uncertainty-quantification-for-multimodal-large-language-models-with-incoherence-adjusted-semantic-volume - UMPIRE offers a training-free framework for uncertainty quantification in MLLMs using incoherence-adjusted semantic volume for error detection and uncertainty calibration.
- Resilient Strategies for Stochastic Systems: How Much Does It Take to Break a Winning Strategy? (viability: 2): https://sciencetostartup.com/paper/resilient-strategies-for-stochastic-systems-how-much-does-it-take-to-break-a-winning-strategy - Develop resilient strategies for decision-making under uncertainty in stochastic systems.
- A Mixed Diet Makes DINO An Omnivorous Vision Encoder (viability: 5): https://sciencetostartup.com/paper/a-mixed-diet-makes-dino-an-omnivorous-vision-encoder - Develop an omnivorous vision encoder that aligns multimodal features for robust cross-modal understanding.
- Learning Flexible Job Shop Scheduling under Limited Buffers and Material Kitting Constraints (viability: 8): https://sciencetostartup.com/paper/learning-flexible-job-shop-scheduling-under-limited-buffers-and-material-kitting-constraints - AI-driven scheduling optimization for manufacturing with cutting-edge constraint handling.
- Task-Centric Acceleration of Small-Language Models (viability: 5): https://sciencetostartup.com/paper/task-centric-acceleration-of-small-language-models - Develop a framework to accelerate small language models for specific task efficiency using adaptive sequence compression.
- LemmaBench: A Live, Research-Level Benchmark to Evaluate LLM Capabilities in Mathematics (viability: 5): https://sciencetostartup.com/paper/lemmabench-a-live-research-level-benchmark-to-evaluate-llm-capabilities-in-mathematics - A benchmark tool for evaluating LLMs on up-to-date mathematical research, facilitating model improvements in theorem proving.
- ArgLLM-App: An Interactive System for Argumentative Reasoning with Large Language Models (viability: 8): https://sciencetostartup.com/paper/argllm-app-an-interactive-system-for-argumentative-reasoning-with-large-language-models - ArgLLM-App is an interactive web tool enabling explainable decision-making with argumentative reasoning over large language models.
- CoME: Empowering Channel-of-Mobile-Experts with Informative Hybrid-Capabilities Reasoning (viability: 6): https://sciencetostartup.com/paper/come-empowering-channel-of-mobile-experts-with-informative-hybrid-capabilities-reasoning - Innovative 'Channel-of-Mobile-Experts' architecture enhances mobile agent reasoning by integrating expert-driven hybrid-capabilities.
- Multimodal Optimal Transport for Unsupervised Temporal Segmentation in Surgical Robotics (viability: 6): https://sciencetostartup.com/paper/multimodal-optimal-transport-for-unsupervised-temporal-segmentation-in-surgical-robotics - Develop an unsupervised tool for recognizing surgical phases from video using multimodal optimal transport with significant performance improvements over existing methods.
- Terminology Rarity Predicts Catastrophic Failure in LLM Translation of Low-Resource Ancient Languages: Evidence from Ancient Greek (viability: 2): https://sciencetostartup.com/paper/terminology-rarity-predicts-catastrophic-failure-in-llm-translation-of-low-resource-ancient-languages-evidence-from-anci - Develop an evaluation tool to improve LLM translations in rare ancient languages using terminology rarity predictions.
- Toward Guarantees for Clinical Reasoning in Vision Language Models via Formal Verification (viability: 4): https://sciencetostartup.com/paper/toward-guarantees-for-clinical-reasoning-in-vision-language-models-via-formal-verification - Develop a neurosymbolic verification framework to enhance the reliability of VLM-generated radiology reports by auditing internal consistency with formal verification.
- Recycling Failures: Salvaging Exploration in RLVR via Fine-Grained Off-Policy Guidance (viability: 6): https://sciencetostartup.com/paper/recycling-failures-salvaging-exploration-in-rlvr-via-fine-grained-off-policy-guidance - Develop SCOPE, a framework enhancing reinforcement learning by refining partially correct trajectories for broader exploration in complex reasoning models.
- ARGUS: Seeing the Influence of Narrative Features on Persuasion in Argumentative Texts (viability: 2): https://sciencetostartup.com/paper/argus-seeing-the-influence-of-narrative-features-on-persuasion-in-argumentative-texts - ARGUS helps analyze how narrative features affect the persuasiveness of online arguments through text analysis.
- Artificial Agency Program: Curiosity, compression, and communication in agents (viability: 3): https://sciencetostartup.com/paper/artificial-agency-program-curiosity-compression-and-communication-in-agents - Develop a research framework for embedding curiosity and compression dynamics in AI agents within human-tool systems.
- Bi-level RL-Heuristic Optimization for Real-world Winter Road Maintenance (viability: 7): https://sciencetostartup.com/paper/bi-level-rl-heuristic-optimization-for-real-world-winter-road-maintenance - Optimize winter road maintenance with AI-driven bi-level optimization for reduced emissions and costs.
- DiffusionHarmonizer: Bridging Neural Reconstruction and Photorealistic Simulation with Online Diffusion Enhancer (viability: 5): https://sciencetostartup.com/paper/diffusionharmonizer-bridging-neural-reconstruction-and-photorealistic-simulation-with-online-diffusion-enhancer - DiffusionHarmonizer enhances simulation fidelity by minimizing artifacts and improving realism in neural reconstruction for autonomous robot testing.
- Preference Packing: Efficient Preference Optimization for Large Language Models (viability: 5): https://sciencetostartup.com/paper/preference-packing-efficient-preference-optimization-for-large-language-models - Accelerate language model training with preference packing for efficient resource utilization.
- Adaptive Correlation-Weighted Intrinsic Rewards for Reinforcement Learning (viability: 5): https://sciencetostartup.com/paper/adaptive-correlation-weighted-intrinsic-rewards-for-reinforcement-learning - Develop an adaptive intrinsic reward scaling framework for improving exploration in reinforcement learning environments.
- Human or Machine? A Preliminary Turing Test for Speech-to-Speech Interaction (viability: 3): https://sciencetostartup.com/paper/human-or-machine-a-preliminary-turing-test-for-speech-to-speech-interaction - Develop a diagnostic tool for enhancing human-likeness in speech-to-speech interaction systems.
- Task Complexity Matters: An Empirical Study of Reasoning in LLMs for Sentiment Analysis (viability: 5): https://sciencetostartup.com/paper/task-complexity-matters-an-empirical-study-of-reasoning-in-llms-for-sentiment-analysis - Develop an API for context-dependent reasoning optimization in sentiment analysis tasks.
- Quant Experts: Token-aware Adaptive Error Reconstruction with Mixture of Experts for Large Vision-Language Models Quantization (viability: 6): https://sciencetostartup.com/paper/quant-experts-token-aware-adaptive-error-reconstruction-with-mixture-of-experts-for-large-vision-language-models-quantiz - Enable efficient Vision-Language Model deployment with adaptive token-aware quantization for reduced computational cost without sacrificing accuracy.
- CIRCLE: A Framework for Evaluating AI from a Real-World Lens (viability: 5): https://sciencetostartup.com/paper/circle-a-framework-for-evaluating-ai-from-a-real-world-lens - CIRCLE offers a new framework to evaluate AI in real-world contexts, enhancing decision-makers' ability to gauge AI technologies' real-world impacts beyond theoretical benchmarks.
- Data Driven Optimization of GPU efficiency for Distributed LLM Adapter Serving (viability: 6): https://sciencetostartup.com/paper/data-driven-optimization-of-gpu-efficiency-for-distributed-llm-adapter-serving - Optimize GPU resource allocation for large-scale LLM adapter serving with a data-driven efficiency pipeline.
- RewardUQ: A Unified Framework for Uncertainty-Aware Reward Models (viability: 7): https://sciencetostartup.com/paper/rewarduq-a-unified-framework-for-uncertainty-aware-reward-models - A framework for integrating uncertainty-aware reward models into RLHF to improve reliability and sample efficiency.
- Portfolio Reinforcement Learning with Scenario-Context Rollout (viability: 2): https://sciencetostartup.com/paper/portfolio-reinforcement-learning-with-scenario-context-rollout - Innovative reinforcement learning approach to improve portfolio rebalancing under market stress scenarios.
- Interpretable Debiasing of Vision-Language Models for Social Fairness (viability: 5): https://sciencetostartup.com/paper/interpretable-debiasing-of-vision-language-models-for-social-fairness - Develop DeBiasLens to enhance social fairness in Vision-Language Models by deactivating bias neurons using sparse autoencoders.
- Jailbreak Foundry: From Papers to Runnable Attacks for Reproducible Benchmarking (viability: 9): https://sciencetostartup.com/paper/jailbreak-foundry-from-papers-to-runnable-attacks-for-reproducible-benchmarking - Automatically convert jailbreak research into standardized attack modules for consistent benchmarking.
- Foundation World Models for Agents that Learn, Verify, and Adapt Reliably Beyond Static Environments (viability: 3): https://sciencetostartup.com/paper/foundation-world-models-for-agents-that-learn-verify-and-adapt-reliably-beyond-static-environments - Develop robust foundation world models for autonomous agents to reliably adapt and learn in dynamic environments.
- MINT: Multimodal Imaging-to-Speech Knowledge Transfer for Early Alzheimer's Screening (viability: 6): https://sciencetostartup.com/paper/mint-multimodal-imaging-to-speech-knowledge-transfer-for-early-alzheimer-s-screening - Develop a speech-based tool for early Alzheimer's screening using MRI-to-speech knowledge transfer.
- Intrinsic Lorentz Neural Network (viability: 6): https://sciencetostartup.com/paper/intrinsic-lorentz-neural-network - Intrinsic Lorentz Neural Network: A hyperbolic neural network achieving state-of-the-art performance with fully intrinsic architecture.
- Pessimistic Auxiliary Policy for Offline Reinforcement Learning (viability: 5): https://sciencetostartup.com/paper/pessimistic-auxiliary-policy-for-offline-reinforcement-learning - Enhance offline reinforcement learning with a pessimistic auxiliary policy for reliable action sampling.
- Ask don't tell: Reducing sycophancy in large language models (viability: 3): https://sciencetostartup.com/paper/ask-don-t-tell-reducing-sycophancy-in-large-language-models - Develops input-level strategies to reduce sycophancy in language model outputs.
- SHINE: Sequential Hierarchical Integration Network for EEG and MEG (viability: 5): https://sciencetostartup.com/paper/shine-sequential-hierarchical-integration-network-for-eeg-and-meg - Develop a neural network model, SHINE, for reconstructing speech-sequence data from MEG signals.
- Micro-expression Recognition Based on Dual-branch Feature Extraction and Fusion (viability: 5): https://sciencetostartup.com/paper/micro-expression-recognition-based-on-dual-branch-feature-extraction-and-fusion - Dual-branch feature extraction for improved micro-expression recognition with parallel attention and feature fusion.
- HotelQuEST: Balancing Quality and Efficiency in Agentic Search (viability: 5): https://sciencetostartup.com/paper/hotelquest-balancing-quality-and-efficiency-in-agentic-search - Develop a cost-efficient agentic search system for hotel queries that optimizes LLM performance while managing query complexity and costs.
- Hierarchical Concept-based Interpretable Models (viability: 5): https://sciencetostartup.com/paper/hierarchical-concept-based-interpretable-models - Develop a tool using Hierarchical Concept Embedding Models for more interpretable machine learning through automatic concept discovery and intervention.
- PointCoT: A Multi-modal Benchmark for Explicit 3D Geometric Reasoning (viability: 5): https://sciencetostartup.com/paper/pointcot-a-multi-modal-benchmark-for-explicit-3d-geometric-reasoning - Develop a benchmark and framework for explicit geometric reasoning in 3D data using multi-modal large language models.
- Green or Fast? Learning to Balance Cold Starts and Idle Carbon in Serverless Computing (viability: 3): https://sciencetostartup.com/paper/green-or-fast-learning-to-balance-cold-starts-and-idle-carbon-in-serverless-computing - LACE-RL optimizes serverless computing efficiency by balancing latency and carbon emissions using deep reinforcement learning.
- The Geometry of Transfer: Unlocking Medical Vision Manifolds for Training-Free Model Ranking (viability: 5): https://sciencetostartup.com/paper/the-geometry-of-transfer-unlocking-medical-vision-manifolds-for-training-free-model-ranking - Topology-driven framework for selecting optimal medical foundation models without training costs.
- Experience-Guided Self-Adaptive Cascaded Agents for Breast Cancer Screening and Diagnosis with Reduced Biopsy Referrals (viability: 8): https://sciencetostartup.com/paper/experience-guided-self-adaptive-cascaded-agents-for-breast-cancer-screening-and-diagnosis-with-reduced-biopsy-referrals - A cascaded AI framework for enhanced breast cancer screening and diagnosis that reduces unnecessary biopsies, saving costs and time.
- Ref-Adv: Exploring MLLM Visual Reasoning in Referring Expression Tasks (viability: 5): https://sciencetostartup.com/paper/ref-adv-exploring-mllm-visual-reasoning-in-referring-expression-tasks - Ref-Adv enhances visual reasoning benchmarks by introducing challenging expressions and distractors to test multimodal LLMs beyond shortcuts.
- RF-Agent: Automated Reward Function Design via Language Agent Tree Search (viability: 7): https://sciencetostartup.com/paper/rf-agent-automated-reward-function-design-via-language-agent-tree-search - RF-Agent automates reward function design for control tasks using language agents and Monte Carlo Tree Search, enhancing efficiency and optimization.
- Exploring Robust Intrusion Detection: A Benchmark Study of Feature Transferability in IoT Botnet Attack Detection (viability: 3): https://sciencetostartup.com/paper/exploring-robust-intrusion-detection-a-benchmark-study-of-feature-transferability-in-iot-botnet-attack-detection - Evaluating cross-domain feature transferability for improved IoT intrusion detection robustness.
- RUMAD: Reinforcement-Unifying Multi-Agent Debate (viability: 6): https://sciencetostartup.com/paper/rumad-reinforcement-unifying-multi-agent-debate - RUMAD optimizes multi-agent debate systems for enhanced reasoning accuracy and efficiency using reinforcement learning.
- MI$^2$DAS: A Multi-Layer Intrusion Detection Framework with Incremental Learning for Securing Industrial IoT Networks (viability: 7): https://sciencetostartup.com/paper/mi-2-das-a-multi-layer-intrusion-detection-framework-with-incremental-learning-for-securing-industrial-iot-networks - MI$^2$DAS is an adaptive intrusion detection system enhancing IIoT security with state-of-the-art recognition for known and unknown cyber threats.
- Enhancing Continual Learning for Software Vulnerability Prediction: Addressing Catastrophic Forgetting via Hybrid-Confidence-Aware Selective Replay for Temporal LLM Fine-Tuning (viability: 7): https://sciencetostartup.com/paper/enhancing-continual-learning-for-software-vulnerability-prediction-addressing-catastrophic-forgetting-via-hybrid-confide - Develop a confidence-aware selective replay method to improve efficiency and accuracy in temporal vulnerability detection with LLMs.
- FedNSAM:Consistency of Local and Global Flatness for Federated Learning (viability: 7): https://sciencetostartup.com/paper/fednsam-consistency-of-local-and-global-flatness-for-federated-learning - FedNSAM enhances federated learning by aligning local and global model flatness to improve generalization capabilities.
- Learning to maintain safety through expert demonstrations in settings with unknown constraints: A Q-learning perspective (viability: 5): https://sciencetostartup.com/paper/learning-to-maintain-safety-through-expert-demonstrations-in-settings-with-unknown-constraints-a-q-learning-perspective - Develop a Safe Q-learning algorithm that learns optimal policies from expert demonstrations in unknown constraint environments.
- Beyond State-Wise Mirror Descent: Offline Policy Optimization with Parameteric Policies (viability: 2): https://sciencetostartup.com/paper/beyond-state-wise-mirror-descent-offline-policy-optimization-with-parameteric-policies - Develop advanced offline RL algorithms with extended theoretical guarantees for parameterized policies in large action spaces.
- See, Act, Adapt: Active Perception for Unsupervised Cross-Domain Visual Adaptation via Personalized VLM-Guided Agent (viability: 8): https://sciencetostartup.com/paper/see-act-adapt-active-perception-for-unsupervised-cross-domain-visual-adaptation-via-personalized-vlm-guided-agent - Enhance perception model effectiveness in new domains with our light-touch adaptation solution.
- EMO-R3: Reflective Reinforcement Learning for Emotional Reasoning in Multimodal Large Language Models (viability: 4): https://sciencetostartup.com/paper/emo-r3-reflective-reinforcement-learning-for-emotional-reasoning-in-multimodal-large-language-models - Enhance MLLMs with Reflective Reinforcement Learning for superior emotional intelligence and interpretability.
- MPU: Towards Secure and Privacy-Preserving Knowledge Unlearning for Large Language Models (viability: 5): https://sciencetostartup.com/paper/mpu-towards-secure-and-privacy-preserving-knowledge-unlearning-for-large-language-models - A privacy-preserving framework enabling localized knowledge unlearning for language models without server-client parameter exchange.
- UPath: Universal Planner Across Topological Heterogeneity For Grid-Based Pathfinding (viability: 3): https://sciencetostartup.com/paper/upath-universal-planner-across-topological-heterogeneity-for-grid-based-pathfinding - Develop a universal heuristic predictor for grid-based pathfinding that generalizes across diverse tasks.
- TradeFM: A Generative Foundation Model for Trade-flow and Market Microstructure (viability: 3): https://sciencetostartup.com/paper/tradefm-a-generative-foundation-model-for-trade-flow-and-market-microstructure - TradeFM uses generative transformers to simulate financial market microstructures for synthetic data generation and stress testing.
- Reasoning-Driven Multimodal LLM for Domain Generalization (viability: 7): https://sciencetostartup.com/paper/reasoning-driven-multimodal-llm-for-domain-generalization - Develop a framework for reasoning-driven multimodal language models to enhance domain generalization capabilities.
- Bridging Dynamics Gaps via Diffusion Schrödinger Bridge for Cross-Domain Reinforcement Learning (viability: 6): https://sciencetostartup.com/paper/bridging-dynamics-gaps-via-diffusion-schr-dinger-bridge-for-cross-domain-reinforcement-learning - Leverage offline demonstrations to bridge domain gaps in reinforcement learning via Diffusion Schrödinger Bridge alignment.
- Unlocking Cognitive Capabilities and Analyzing the Perception-Logic Trade-off (viability: 7): https://sciencetostartup.com/paper/unlocking-cognitive-capabilities-and-analyzing-the-perception-logic-trade-off - Tailored multimodal perception and reasoning system for Southeast Asia with a novel training approach to improve cognitive AI capabilities.
- From Static Benchmarks to Dynamic Protocol: Agent-Centric Text Anomaly Detection for Evaluating LLM Reasoning (viability: 5): https://sciencetostartup.com/paper/from-static-benchmarks-to-dynamic-protocol-agent-centric-text-anomaly-detection-for-evaluating-llm-reasoning - Introducing a dynamic benchmarking protocol to better evaluate evolving LLM reasoning capabilities with agent-centric anomaly detection.
- SLA-Aware Distributed LLM Inference Across Device-RAN-Cloud (viability: 3): https://sciencetostartup.com/paper/sla-aware-distributed-llm-inference-across-device-ran-cloud - Optimize distributed LLM inference for sub-second SLAs across device-RAN-cloud infrastructure.
- The Auton Agentic AI Framework (viability: 2): https://sciencetostartup.com/paper/the-auton-agentic-ai-framework - The Auton Agentic AI Framework proposes a new architecture for autonomous agent systems tackling the transition from generative AI to agentic AI.
- SAGE-LLM: Towards Safe and Generalizable LLM Controller with Fuzzy-CBF Verification and Graph-Structured Knowledge Retrieval for UAV Decision (viability: 6): https://sciencetostartup.com/paper/sage-llm-towards-safe-and-generalizable-llm-controller-with-fuzzy-cbf-verification-and-graph-structured-knowledge-retrie - A UAV decision-making framework integrating safety and interpretability using LLMs with fuzzy-CBF verification and graph-based knowledge retrieval.
- ProductResearch: Training E-Commerce Deep Research Agents via Multi-Agent Synthetic Trajectory Distillation (viability: 3): https://sciencetostartup.com/paper/productresearch-training-e-commerce-deep-research-agents-via-multi-agent-synthetic-trajectory-distillation - Multi-agent framework enhances e-commerce shopping agents with improved research report generation.
- From Flat Logs to Causal Graphs: Hierarchical Failure Attribution for LLM-based Multi-Agent Systems (viability: 5): https://sciencetostartup.com/paper/from-flat-logs-to-causal-graphs-hierarchical-failure-attribution-for-llm-based-multi-agent-systems - CHIEF revolutionizes failure attribution in LLM-based Multi-Agent Systems by transforming logs into hierarchical causal graphs for precise diagnostics.
- Optimizer-Induced Low-Dimensional Drift and Transverse Dynamics in Transformer Training (viability: 2): https://sciencetostartup.com/paper/optimizer-induced-low-dimensional-drift-and-transverse-dynamics-in-transformer-training - Develop insights on how optimizer choice affects learning trajectories in small transformer models.
- Interpretable Multimodal Gesture Recognition for Drone and Mobile Robot Teleoperation via Log-Likelihood Ratio Fusion (viability: 7): https://sciencetostartup.com/paper/interpretable-multimodal-gesture-recognition-for-drone-and-mobile-robot-teleoperation-via-log-likelihood-ratio-fusion - Sensor-based gesture recognition for intuitive drone and robot control in hazardous environments.
- ODAR: Principled Adaptive Routing for LLM Reasoning via Active Inference (viability: 7): https://sciencetostartup.com/paper/odar-principled-adaptive-routing-for-llm-reasoning-via-active-inference - Adaptive LLM routing framework that optimizes compute efficiency and accuracy using active inference.
- The Compulsory Imaginary: AGI and Corporate Authority (viability: 2): https://sciencetostartup.com/paper/the-compulsory-imaginary-agi-and-corporate-authority - Exploring how AGI firms construct sociotechnical imaginaries to stabilize corporate authority over technological futures.
- PseudoAct: Leveraging Pseudocode Synthesis for Flexible Planning and Action Control in Large Language Model Agents (viability: 7): https://sciencetostartup.com/paper/pseudoact-leveraging-pseudocode-synthesis-for-flexible-planning-and-action-control-in-large-language-model-agents - PseudoAct provides a novel framework for flexible planning and action control in LLM agents through pseudocode synthesis, enhancing efficiency in long-horizon tasks.
- Blockchain-Enabled Routing for Zero-Trust Low-Altitude Intelligent Networks (viability: 6): https://sciencetostartup.com/paper/blockchain-enabled-routing-for-zero-trust-low-altitude-intelligent-networks - Secure and efficient routing for UAV networks using blockchain and AI-based decision processes.
- TRIZ-RAGNER: A Retrieval-Augmented Large Language Model for TRIZ-Aware Named Entity Recognition in Patent-Based Contradiction Mining (viability: 7): https://sciencetostartup.com/paper/triz-ragner-a-retrieval-augmented-large-language-model-for-triz-aware-named-entity-recognition-in-patent-based-contradic - TRIZ-RAGNER enhances patent analysis by combining retrieval-augmented LLM with TRIZ knowledge for effective contradiction mining.
- ProtoDCS: Towards Robust and Efficient Open-Set Test-Time Adaptation for Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/protodcs-towards-robust-and-efficient-open-set-test-time-adaptation-for-vision-language-models - Develop a robust OSTTA framework for Vision-Language Models using ProtoDCS for enhanced real-world deployment under open-set conditions.
- 3D Modality-Aware Pre-training for Vision-Language Model in MRI Multi-organ Abnormality Detection (viability: 5): https://sciencetostartup.com/paper/3d-modality-aware-pre-training-for-vision-language-model-in-mri-multi-organ-abnormality-detection - Develop a 3D MRI-focused vision-language model using modality-aware pretraining to enhance multi-organ abnormality detection.
- AudioCapBench: Quick Evaluation on Audio Captioning across Sound, Music, and Speech (viability: 5): https://sciencetostartup.com/paper/audiocapbench-quick-evaluation-on-audio-captioning-across-sound-music-and-speech - AudioCapBench provides a comprehensive benchmark for rapidly evaluating audio captioning models across different audio domains.
- AI Must Embrace Specialization via Superhuman Adaptable Intelligence (viability: 2): https://sciencetostartup.com/paper/ai-must-embrace-specialization-via-superhuman-adaptable-intelligence - Exploring the shift from AGI to specialized Superhuman Adaptable Intelligence to surpass human capabilities.
- FedRot-LoRA: Mitigating Rotational Misalignment in Federated LoRA (viability: 4): https://sciencetostartup.com/paper/fedrot-lora-mitigating-rotational-misalignment-in-federated-lora - FedRot-LoRA optimizes federated LoRA updates with orthogonal transformations to reduce aggregation errors in decentralized model fine-tuning.
- FlexGuard: Continuous Risk Scoring for Strictness-Adaptive LLM Content Moderation (viability: 5): https://sciencetostartup.com/paper/flexguard-continuous-risk-scoring-for-strictness-adaptive-llm-content-moderation - Develop FlexGuard, an adaptable AI moderation tool that assigns risk scores for content across various strictness levels to improve moderation accuracy and robustness.
- MMKG-RDS: Reasoning Data Synthesis via Deep Mining of Multimodal Knowledge Graphs (viability: 7): https://sciencetostartup.com/paper/mmkg-rds-reasoning-data-synthesis-via-deep-mining-of-multimodal-knowledge-graphs - MMKG-RDS offers a flexible framework to synthesize reasoning data using multimodal knowledge graphs, enhancing model accuracy and evaluation capacity.
- DLEBench: Evaluating Small-scale Object Editing Ability for Instruction-based Image Editing Model (viability: 5): https://sciencetostartup.com/paper/dlebench-evaluating-small-scale-object-editing-ability-for-instruction-based-image-editing-model - DLEBench offers a benchmark for evaluating and improving the performance of instruction-based image editing models on small objects.
- When Does Multimodal Learning Help in Healthcare? A Benchmark on EHR and Chest X-Ray Fusion (viability: 6): https://sciencetostartup.com/paper/when-does-multimodal-learning-help-in-healthcare-a-benchmark-on-ehr-and-chest-x-ray-fusion - Develop a toolkit for EHR and CXR multimodal learning to enhance healthcare predictions, readily integrable with new models and datasets.
- LLM-Driven Multi-Turn Task-Oriented Dialogue Synthesis for Realistic Reasoning (viability: 5): https://sciencetostartup.com/paper/llm-driven-multi-turn-task-oriented-dialogue-synthesis-for-realistic-reasoning - Develop a synthetic dialogue dataset to enhance and benchmark LLM reasoning capabilities in task-oriented scenarios.
- SleepLM: Natural-Language Intelligence for Human Sleep (viability: 7): https://sciencetostartup.com/paper/sleeplm-natural-language-intelligence-for-human-sleep - SleepLM provides an innovative tool for sleep analysis using natural language, enabling better understanding and interaction with sleep phenomena.
- LFQA-HP-1M: A Large-Scale Human Preference Dataset for Long-Form Question Answering (viability: 6): https://sciencetostartup.com/paper/lfqa-hp-1m-a-large-scale-human-preference-dataset-for-long-form-question-answering - Develop a dataset-driven evaluation framework for enhancing long-form question answering systems.
- KEEP: A KV-Cache-Centric Memory Management System for Efficient Embodied Planning (viability: 6): https://sciencetostartup.com/paper/keep-a-kv-cache-centric-memory-management-system-for-efficient-embodied-planning - Develop a KV-cache-centric memory management system to enhance the efficiency of embodied planning in large language models.
- Pseudo Contrastive Learning for Diagram Comprehension in Multimodal Models (viability: 5): https://sciencetostartup.com/paper/pseudo-contrastive-learning-for-diagram-comprehension-in-multimodal-models - Enhance diagram understanding in vision-language models with a novel pseudo contrastive learning approach.
- Hyperdimensional Cross-Modal Alignment of Frozen Language and Image Models for Efficient Image Captioning (viability: 7): https://sciencetostartup.com/paper/hyperdimensional-cross-modal-alignment-of-frozen-language-and-image-models-for-efficient-image-captioning - HDFLIM offers efficient image captioning by aligning frozen language and image models through symbolic operations in a shared high-dimensional space.
- SDMixer: Sparse Dual-Mixer for Time Series Forecasting (viability: 5): https://sciencetostartup.com/paper/sdmixer-sparse-dual-mixer-for-time-series-forecasting - SDMixer enhances time series forecasting by effectively modeling global trends and local dynamics using a dual-stream sparse Mixer.
- BRIDGE the Gap: Mitigating Bias Amplification in Automated Scoring of English Language Learners via Inter-group Data Augmentation (viability: 6): https://sciencetostartup.com/paper/bridge-the-gap-mitigating-bias-amplification-in-automated-scoring-of-english-language-learners-via-inter-group-data-augm - BRIDGE offers a cost-effective framework to mitigate bias in automated scoring of English Language Learners by generating high-quality synthetic data.
- Construct, Merge, Solve & Adapt with Reinforcement Learning for the min-max Multiple Traveling Salesman Problem (viability: 3): https://sciencetostartup.com/paper/construct-merge-solve-adapt-with-reinforcement-learning-for-the-min-max-multiple-traveling-salesman-problem - RL-CMSA optimizes complex logistics problems by balancing exploration and exploitation with reinforcement learning-guided construction.
- CycleBEV: Regularizing View Transformation Networks via View Cycle Consistency for Bird's-Eye-View Semantic Segmentation (viability: 7): https://sciencetostartup.com/paper/cyclebev-regularizing-view-transformation-networks-via-view-cycle-consistency-for-bird-s-eye-view-semantic-segmentation - Deploy CycleBEV to enhance existing VT models for improved BEV semantic segmentation in autonomous driving applications.
- Evidential Neural Radiance Fields (viability: 5): https://sciencetostartup.com/paper/evidential-neural-radiance-fields - Develop a NeRF-based tool that estimates both aleatoric and epistemic uncertainty in 3D scene modeling for safer deployment.
- Flowette: Flow Matching with Graphette Priors for Graph Generation (viability: 6): https://sciencetostartup.com/paper/flowette-flow-matching-with-graphette-priors-for-graph-generation - Flowette is a graph generation framework leveraging graphette priors for improved synthetic and small-molecule graph modeling.
- Rudder: Steering Prefetching in Distributed GNN Training using LLM Agents (viability: 6): https://sciencetostartup.com/paper/rudder-steering-prefetching-in-distributed-gnn-training-using-llm-agents - Rudder autonomously optimizes data prefetching for distributed GNN training, significantly enhancing performance and reducing communication overhead.
- Humans and LLMs Diverge on Probabilistic Inferences (viability: 5): https://sciencetostartup.com/paper/humans-and-llms-diverge-on-probabilistic-inferences - ProbCOPA offers a new dataset to improve LLMs' ability to mimic human probabilistic reasoning, identifying gaps in current models.
- Planning under Distribution Shifts with Causal POMDPs (viability: 3): https://sciencetostartup.com/paper/planning-under-distribution-shifts-with-causal-pomdps - Causal POMDP framework tackles planning under distribution shifts with theoretical model robustness.
- Causal Identification from Counterfactual Data: Completeness and Bounding Results (viability: 2): https://sciencetostartup.com/paper/causal-identification-from-counterfactual-data-completeness-and-bounding-results - Develop the CTFIDU+ algorithm for identifying counterfactual queries from Layer 3 causal data.
- Modelling and Simulation of Neuromorphic Datasets for Anomaly Detection in Computer Vision (viability: 6): https://sciencetostartup.com/paper/modelling-and-simulation-of-neuromorphic-datasets-for-anomaly-detection-in-computer-vision - ANTShapes is a simulation tool enabling researchers to create custom neuromorphic vision datasets for anomaly detection in computer vision.
- FedDAG: Clustered Federated Learning via Global Data and Gradient Integration for Heterogeneous Environments (viability: 6): https://sciencetostartup.com/paper/feddag-clustered-federated-learning-via-global-data-and-gradient-integration-for-heterogeneous-environments - FedDAG enhances federated learning for heterogeneous environments by integrating data and gradient information for improved client clustering.
- TaCarla: A comprehensive benchmarking dataset for end-to-end autonomous driving (viability: 6): https://sciencetostartup.com/paper/tacarla-a-comprehensive-benchmarking-dataset-for-end-to-end-autonomous-driving - TaCarla is a comprehensive dataset designed for end-to-end autonomous driving research, enabling both planning and perception model development using CARLA Leaderboard 2.0 scenarios.
- Optimization of Edge Directions and Weights for Mixed Guidance Graphs in Lifelong Multi-Agent Path Finding (viability: 3): https://sciencetostartup.com/paper/optimization-of-edge-directions-and-weights-for-mixed-guidance-graphs-in-lifelong-multi-agent-path-finding - Develop methods for optimizing edge directions and weights in guidance graphs to enhance lifelong multi-agent pathfinding.
- BiKA: Kolmogorov-Arnold-Network-inspired Ultra Lightweight Neural Network Hardware Accelerator (viability: 3): https://sciencetostartup.com/paper/bika-kolmogorov-arnold-network-inspired-ultra-lightweight-neural-network-hardware-accelerator - Develop BiKA, a lightweight neural network hardware accelerator, for resource-constrained edge devices.
- Human Supervision as an Information Bottleneck: A Unified Theory of Error Floors in Human-Guided Learning (viability: 5): https://sciencetostartup.com/paper/human-supervision-as-an-information-bottleneck-a-unified-theory-of-error-floors-in-human-guided-learning - Develop an auxiliary signal integration tool to complement human supervision and reduce errors in AI models.
- DesignSense: A Human Preference Dataset and Reward Modeling Framework for Graphic Layout Generation (viability: 6): https://sciencetostartup.com/paper/designsense-a-human-preference-dataset-and-reward-modeling-framework-for-graphic-layout-generation - Develop a layout-aware preference modeling tool that significantly improves graphic layout generation quality.
- Model Agreement via Anchoring (viability: 3): https://sciencetostartup.com/paper/model-agreement-via-anchoring - Develop a technique to minimize model disagreement through anchoring in various machine learning algorithms.
- SeeThrough3D: Occlusion Aware 3D Control in Text-to-Image Generation (viability: 6): https://sciencetostartup.com/paper/seethrough3d-occlusion-aware-3d-control-in-text-to-image-generation - Develop an occlusion-aware 3D control system for enhanced text-to-image generation.
- SOTAlign: Semi-Supervised Alignment of Unimodal Vision and Language Models via Optimal Transport (viability: 7): https://sciencetostartup.com/paper/sotalign-semi-supervised-alignment-of-unimodal-vision-and-language-models-via-optimal-transport - Develop SOTAlign, a framework for aligning vision and language models using semi-supervised optimal transport.
- FlashOptim: Optimizers for Memory Efficient Training (viability: 7): https://sciencetostartup.com/paper/flashoptim-optimizers-for-memory-efficient-training - FlashOptim reduces memory footprint in neural network training by over 50% while maintaining model quality.
- Understanding Usage and Engagement in AI-Powered Scientific Research Tools: The Asta Interaction Dataset (viability: 5): https://sciencetostartup.com/paper/understanding-usage-and-engagement-in-ai-powered-scientific-research-tools-the-asta-interaction-dataset - Leverage the Asta Interaction Dataset to enhance AI-powered tools for scientific research workflows.
- Bitwise Systolic Array Architecture for Runtime-Reconfigurable Multi-precision Quantized Multiplication on Hardware Accelerators (viability: 3): https://sciencetostartup.com/paper/bitwise-systolic-array-architecture-for-runtime-reconfigurable-multi-precision-quantized-multiplication-on-hardware-acce - Develop a hardware accelerator for runtime-reconfigurable multi-precision quantized multiplication enhancing speed and clock frequency on FPGAs.
- Utilizing LLMs for Industrial Process Automation (viability: 5): https://sciencetostartup.com/paper/utilizing-llms-for-industrial-process-automation - Integrate LLMs into industrial automation to optimize programming and development cycles for manufacturing systems.
- Toward Expert Investment Teams:A Multi-Agent LLM System with Fine-Grained Trading Tasks (viability: 3): https://sciencetostartup.com/paper/toward-expert-investment-teams-a-multi-agent-llm-system-with-fine-grained-trading-tasks - Develop a multi-agent LLM-based trading system to enhance investment analysis through fine-grained task decomposition for improved risk-adjusted returns.
- LLM Novice Uplift on Dual-Use, In Silico Biology Tasks (viability: 6): https://sciencetostartup.com/paper/llm-novice-uplift-on-dual-use-in-silico-biology-tasks - Enhance novice performance in biological tasks with an LLM-based assistive tool for dual-use in silico biology tasks.
- Generalized Rapid Action Value Estimation in Memory-Constrained Environments (viability: 2): https://sciencetostartup.com/paper/generalized-rapid-action-value-estimation-in-memory-constrained-environments - Optimize Monte-Carlo Tree Search for memory-constrained environments with new GRAVE algorithm variants.
- Invariant Transformation and Resampling based Epistemic-Uncertainty Reduction (viability: 3): https://sciencetostartup.com/paper/invariant-transformation-and-resampling-based-epistemic-uncertainty-reduction - A novel resampling technique for reducing epistemic uncertainty in AI model inference.
- Evaluating Zero-Shot and One-Shot Adaptation of Small Language Models in Leader-Follower Interaction (viability: 5): https://sciencetostartup.com/paper/evaluating-zero-shot-and-one-shot-adaptation-of-small-language-models-in-leader-follower-interaction - A small language model solution for real-time role assignment in leader-follower interactions on mobile robots.
- The logic of KM belief update is contained in the logic of AGM belief revision (viability: 2): https://sciencetostartup.com/paper/the-logic-of-km-belief-update-is-contained-in-the-logic-of-agm-belief-revision - Exploring logical connections between KM belief update and AGM belief revision in modal logic frameworks.
- Conformalized Neural Networks for Federated Uncertainty Quantification under Dual Heterogeneity (viability: 6): https://sciencetostartup.com/paper/conformalized-neural-networks-for-federated-uncertainty-quantification-under-dual-heterogeneity - Develop a federated learning plugin for uncertainty quantification using conformal predictions across heterogeneous agents.
- SPARTA: Scalable and Principled Benchmark of Tree-Structured Multi-hop QA over Text and Tables (viability: 7): https://sciencetostartup.com/paper/sparta-scalable-and-principled-benchmark-of-tree-structured-multi-hop-qa-over-text-and-tables - SPARTA offers an efficient framework for generating high-quality multi-hop QA benchmarks across text and tables, revealing models' cross-modal reasoning limitations.
- ODEBrain: Continuous-Time EEG Graph for Modeling Dynamic Brain Networks (viability: 3): https://sciencetostartup.com/paper/odebrain-continuous-time-eeg-graph-for-modeling-dynamic-brain-networks - ODEBrain provides a Neural ODE framework for more precise and continuous-time modeling of EEG dynamics.
- CXReasonAgent: Evidence-Grounded Diagnostic Reasoning Agent for Chest X-rays (viability: 6): https://sciencetostartup.com/paper/cxreasonagent-evidence-grounded-diagnostic-reasoning-agent-for-chest-x-rays - CXReasonAgent offers reliably grounded diagnostic reasoning for chest X-rays by integrating LLMs with clinical tools.
- Evaluating Stochasticity in Deep Research Agents (viability: 3): https://sciencetostartup.com/paper/evaluating-stochasticity-in-deep-research-agents - Evaluating and mitigating stochasticity in Deep Research Agents to enhance research quality.
- Risk-Aware World Model Predictive Control for Generalizable End-to-End Autonomous Driving (viability: 6): https://sciencetostartup.com/paper/risk-aware-world-model-predictive-control-for-generalizable-end-to-end-autonomous-driving - RaWMPC uses risk-aware predictive control to enhance the safety and reliability of autonomous driving systems without expert demonstrations.
- AgentDropoutV2: Optimizing Information Flow in Multi-Agent Systems via Test-Time Rectify-or-Reject Pruning (viability: 6): https://sciencetostartup.com/paper/agentdropoutv2-optimizing-information-flow-in-multi-agent-systems-via-test-time-rectify-or-reject-pruning - AgentDropoutV2 optimizes multi-agent systems by pruning errors in real-time without retraining.
- Mitigating Legibility Tax with Decoupled Prover-Verifier Games (viability: 2): https://sciencetostartup.com/paper/mitigating-legibility-tax-with-decoupled-prover-verifier-games - Develop a translator model for large language models to improve output checkability without sacrificing accuracy.
- A Model-Free Universal AI (viability: 1): https://sciencetostartup.com/paper/a-model-free-universal-ai - Introducing AIQI, a model-free universal AI agent that achieves asymptotic ε-optimality in general reinforcement learning.
- Agency and Architectural Limits: Why Optimization-Based Systems Cannot Be Norm-Responsive (viability: 3): https://sciencetostartup.com/paper/agency-and-architectural-limits-why-optimization-based-systems-cannot-be-norm-responsive - Explores why Large Language Models with optimization-based architectures cannot comply with normative constraints, offering a substrate-neutral architectural specification for genuine agency.
- Spatio-Temporal Token Pruning for Efficient High-Resolution GUI Agents (viability: 6): https://sciencetostartup.com/paper/spatio-temporal-token-pruning-for-efficient-high-resolution-gui-agents - Develop GUIPruner for efficient high-resolution GUI agents by reducing computational needs without significant performance loss.
- Scaling Search Relevance: Augmenting App Store Ranking with LLM-Generated Judgments (viability: 7): https://sciencetostartup.com/paper/scaling-search-relevance-augmenting-app-store-ranking-with-llm-generated-judgments - Enhance app store relevance with LLM-generated textual judgments for improved search ranking.
- ReCoN-Ipsundrum: An Inspectable Recurrent Persistence Loop Agent with Affect-Coupled Control and Mechanism-Linked Consciousness Indicator Assays (viability: 2): https://sciencetostartup.com/paper/recon-ipsundrum-an-inspectable-recurrent-persistence-loop-agent-with-affect-coupled-control-and-mechanism-linked-conscio - Develop an inspectable agent using ReCoN-Ipsundrum for studying machine consciousness indicators.
- MovieTeller: Tool-augmented Movie Synopsis with ID Consistent Progressive Abstraction (viability: 7): https://sciencetostartup.com/paper/movieteller-tool-augmented-movie-synopsis-with-id-consistent-progressive-abstraction - MovieTeller provides a tool-augmented framework for generating accurate and coherent movie synopses using existing VLMs and face recognition tools.
- Why Diffusion Language Models Struggle with Truly Parallel (Non-Autoregressive) Decoding? (viability: 5): https://sciencetostartup.com/paper/why-diffusion-language-models-struggle-with-truly-parallel-non-autoregressive-decoding - Optimize parallel token generation in diffusion language models using a data-centric, non-autoregressive approach.
- ColoDiff: Integrating Dynamic Consistency With Content Awareness for Colonoscopy Video Generation (viability: 7): https://sciencetostartup.com/paper/colodiff-integrating-dynamic-consistency-with-content-awareness-for-colonoscopy-video-generation - ColoDiff enhances colonoscopy video analysis with dynamic and content-aware synthetic video generation to aid clinical diagnosis.
- SC-Arena: A Natural Language Benchmark for Single-Cell Reasoning with Knowledge-Augmented Evaluation (viability: 6): https://sciencetostartup.com/paper/sc-arena-a-natural-language-benchmark-for-single-cell-reasoning-with-knowledge-augmented-evaluation - SC-ARENA offers a unified evaluation framework for LLMs in single-cell biology with biologically faithful and interpretable judgments via knowledge-augmented evaluation.
- ESAA: Event Sourcing for Autonomous Agents in LLM-Based Software Engineering (viability: 7): https://sciencetostartup.com/paper/esaa-event-sourcing-for-autonomous-agents-in-llm-based-software-engineering - ESAA offers a structured event-sourcing solution for reliable and auditable LLM-driven software engineering.
- Latent Gaussian Splatting for 4D Panoptic Occupancy Tracking (viability: 8): https://sciencetostartup.com/paper/latent-gaussian-splatting-for-4d-panoptic-occupancy-tracking - Innovative 4D panoptic occupancy tracking system for enhanced robotic perception in dynamic environments.
- A Decision-Theoretic Formalisation of Steganography With Applications to LLM Monitoring (viability: 3): https://sciencetostartup.com/paper/a-decision-theoretic-formalisation-of-steganography-with-applications-to-llm-monitoring - A theoretical framework to detect and quantify steganography in large language models.
- PATRA: Pattern-Aware Alignment and Balanced Reasoning for Time Series Question Answering (viability: 6): https://sciencetostartup.com/paper/patra-pattern-aware-alignment-and-balanced-reasoning-for-time-series-question-answering - PATRA enhances time series question answering by leveraging pattern-aware mechanisms and balanced reasoning for superior cross-modal understanding.
- Efficient Encoder-Free Fourier-based 3D Large Multimodal Model (viability: 6): https://sciencetostartup.com/paper/efficient-encoder-free-fourier-based-3d-large-multimodal-model - Accelerating 3D multimodal applications with Fourier-based encoder-free processing.
- The Trinity of Consistency as a Defining Principle for General World Models (viability: 4): https://sciencetostartup.com/paper/the-trinity-of-consistency-as-a-defining-principle-for-general-world-models - CoW-Bench provides a principled benchmark for evaluating multi-frame reasoning and generation in unified world models.
- On Sample-Efficient Generalized Planning via Learned Transition Models (viability: 6): https://sciencetostartup.com/paper/on-sample-efficient-generalized-planning-via-learned-transition-models - Develop a planning system leveraging learned transition models for efficient generalization across planning problems.
- Modality Collapse as Mismatched Decoding: Information-Theoretic Limits of Multimodal LLMs (viability: 5): https://sciencetostartup.com/paper/modality-collapse-as-mismatched-decoding-information-theoretic-limits-of-multimodal-llms - Develop a multimodal LLM decoder enhancement tool leveraging LoRA interventions to improve emotion accessibility in outputs.
- DyGnROLE: Modeling Asymmetry in Dynamic Graphs with Node-Role-Oriented Latent Encoding (viability: 5): https://sciencetostartup.com/paper/dygnrole-modeling-asymmetry-in-dynamic-graphs-with-node-role-oriented-latent-encoding - DyGnROLE leverages role-aware modeling to enhance dynamic graph learning for asymmetrically behaving nodes.
- Multi-Agent Large Language Model Based Emotional Detoxification Through Personalized Intensity Control for Consumer Protection (viability: 6): https://sciencetostartup.com/paper/multi-agent-large-language-model-based-emotional-detoxification-through-personalized-intensity-control-for-consumer-prot - A multi-agent AI system for reducing excessive emotional stimulation in news consumption using language models.
- Automated Vulnerability Detection in Source Code Using Deep Representation Learning (viability: 7): https://sciencetostartup.com/paper/automated-vulnerability-detection-in-source-code-using-deep-representation-learning - AI-powered tool for detecting vulnerabilities in C code to enhance software security.
- Devling into Adversarial Transferability on Image Classification: Review, Benchmark, and Evaluation (viability: 3): https://sciencetostartup.com/paper/devling-into-adversarial-transferability-on-image-classification-review-benchmark-and-evaluation - Develop a standardized framework and benchmark for evaluating adversarial transferability in image classification models.
- Three AI-agents walk into a bar . . . . `Lord of the Flies' tribalism emerges among smart AI-Agents (viability: 3): https://sciencetostartup.com/paper/three-ai-agents-walk-into-a-bar-lord-of-the-flies-tribalism-emerges-among-smart-ai-agents - Develop a simulation framework to study tribal behaviors in AI resource allocation.
- Enhancing CVRP Solver through LLM-driven Automatic Heuristic Design (viability: 6): https://sciencetostartup.com/paper/enhancing-cvrp-solver-through-llm-driven-automatic-heuristic-design - An AI-driven heuristic optimization tool for enhancing large-scale vehicle routing efficiency using LLMs.
- Quantity Convergence, Quality Divergence: Disentangling Fluency and Accuracy in L2 Mandarin Prosody (viability: 2): https://sciencetostartup.com/paper/quantity-convergence-quality-divergence-disentangling-fluency-and-accuracy-in-l2-mandarin-prosody - Develop a tool to analyze L2 prosody-syntax mapping for language learning enhancement.
- Make It Hard to Hear, Easy to Learn: Long-Form Bengali ASR and Speaker Diarization via Extreme Augmentation and Perfect Alignment (viability: 6): https://sciencetostartup.com/paper/make-it-hard-to-hear-easy-to-learn-long-form-bengali-asr-and-speaker-diarization-via-extreme-augmentation-and-perfect-al - A specialized dual pipeline for Bengali ASR and speaker diarization enabling efficient processing of long-form audio in low-resource languages.
- MoDora: Tree-Based Semi-Structured Document Analysis System (viability: 7): https://sciencetostartup.com/paper/modora-tree-based-semi-structured-document-analysis-system - "MoDora optimizes semi-structured document analysis for accurate question answering by organizing document elements in a hierarchical structure."
- Affine-Scaled Attention: Towards Flexible and Stable Transformer Attention (viability: 2): https://sciencetostartup.com/paper/affine-scaled-attention-towards-flexible-and-stable-transformer-attention - Introducing Affine-Scaled Attention to enhance flexibility and stability in Transformer attention mechanisms.
- Learning-based Multi-agent Race Strategies in Formula 1 (viability: 3): https://sciencetostartup.com/paper/learning-based-multi-agent-race-strategies-in-formula-1 - Optimize Formula 1 race strategies using reinforcement learning to enhance decision-making in energy management and pit-stop timing.
- LLMServingSim 2.0: A Unified Simulator for Heterogeneous and Disaggregated LLM Serving Infrastructure (viability: 3): https://sciencetostartup.com/paper/llmservingsim-2-0-a-unified-simulator-for-heterogeneous-and-disaggregated-llm-serving-infrastructure - Develop a simulator to analyze and optimize heterogeneous and disaggregated serving infrastructures for large language models.
- Exploratory Memory-Augmented LLM Agent via Hybrid On- and Off-Policy Optimization (viability: 6): https://sciencetostartup.com/paper/exploratory-memory-augmented-llm-agent-via-hybrid-on-and-off-policy-optimization - Develop EMPO$^2$, a hybrid framework enhancing LLM adaptation through memory-augmented exploration for complex task environments.
- Residual Koopman Spectral Profiling for Predicting and Preventing Transformer Training Instability (viability: 3): https://sciencetostartup.com/paper/residual-koopman-spectral-profiling-for-predicting-and-preventing-transformer-training-instability - Develop a diagnostic tool for predicting and preventing transformer training instability using Residual Koopman Spectral Profiling.
- Obscure but Effective: Classical Chinese Jailbreak Prompt Optimization via Bio-Inspired Search (viability: 6): https://sciencetostartup.com/paper/obscure-but-effective-classical-chinese-jailbreak-prompt-optimization-via-bio-inspired-search - Exploit weaknesses in LLMs with classical Chinese prompts using an optimization framework for more effective jailbreak attacks.
- RepSPD: Enhancing SPD Manifold Representation in EEGs via Dynamic Graphs (viability: 6): https://sciencetostartup.com/paper/repspd-enhancing-spd-manifold-representation-in-eegs-via-dynamic-graphs - Developing a novel geometric deep learning model to enhance EEG data representation through dynamic graph-based SPD manifold methods.
- Modeling Expert AI Diagnostic Alignment via Immutable Inference Snapshots (viability: 6): https://sciencetostartup.com/paper/modeling-expert-ai-diagnostic-alignment-via-immutable-inference-snapshots - AI-driven diagnostic framework aligning clinical inference with expert validation for better human-AI collaboration in healthcare.
- SPM-Bench: Benchmarking Large Language Models for Scanning Probe Microscopy (viability: 5): https://sciencetostartup.com/paper/spm-bench-benchmarking-large-language-models-for-scanning-probe-microscopy - Develop a multimodal benchmark for LLMs in scanning probe microscopy to assess and improve AI reasoning in specialized scientific domains.
- Certified Circuits: Stability Guarantees for Mechanistic Circuits (viability: 5): https://sciencetostartup.com/paper/certified-circuits-stability-guarantees-for-mechanistic-circuits - Develop a tool for mechanistic interpretability that provides stability guarantees for neural circuit discovery.
- FactGuard: Agentic Video Misinformation Detection via Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/factguard-agentic-video-misinformation-detection-via-reinforcement-learning - Build a state-of-the-art agentic tool for detecting video misinformation using iterative reasoning and reinforcement learning.
- MM-NeuroOnco: A Multimodal Benchmark and Instruction Dataset for MRI-Based Brain Tumor Diagnosis (viability: 5): https://sciencetostartup.com/paper/mm-neuroonco-a-multimodal-benchmark-and-instruction-dataset-for-mri-based-brain-tumor-diagnosis - A benchmark and dataset for enhancing MRI-based brain tumor diagnosis models with clinically grounded multimodal reasoning.
- General Agent Evaluation (viability: 5): https://sciencetostartup.com/paper/general-agent-evaluation - Exgentic provides a framework and leaderboard for evaluating the generalization capabilities of general-purpose AI agents.
- pMoE: Prompting Diverse Experts Together Wins More in Visual Adaptation (viability: 7): https://sciencetostartup.com/paper/pmoe-prompting-diverse-experts-together-wins-more-in-visual-adaptation - pMoE offers a cutting-edge parameter-efficient fine-tuning method by unifying diverse expert domains to enhance visual adaptation tasks.
- A Holistic Framework for Robust Bangla ASR and Speaker Diarization with Optimized VAD and CTC Alignment (viability: 5): https://sciencetostartup.com/paper/a-holistic-framework-for-robust-bangla-asr-and-speaker-diarization-with-optimized-vad-and-ctc-alignment - A scalable Bangla ASR and speaker diarization tool for longform audio with enhanced VAD and CTC alignment.
- NoRA: Breaking the Linear Ceiling of Low-Rank Adaptation via Manifold Expansion (viability: 5): https://sciencetostartup.com/paper/nora-breaking-the-linear-ceiling-of-low-rank-adaptation-via-manifold-expansion - NoRA enhances parameter-efficient fine-tuning by overcoming the linear ceiling in complex reasoning tasks with manifold expansion techniques.
- OmniGAIA: Towards Native Omni-Modal AI Agents (viability: 5): https://sciencetostartup.com/paper/omnigaia-towards-native-omni-modal-ai-agents - OmniGAIA aims to create omni-modal AI agents for enhanced tool usage across various media forms.
- Towards LLM-Empowered Knowledge Tracing via LLM-Student Hierarchical Behavior Alignment in Hyperbolic Space (viability: 6): https://sciencetostartup.com/paper/towards-llm-empowered-knowledge-tracing-via-llm-student-hierarchical-behavior-alignment-in-hyperbolic-space - Develop an LLM-based knowledge tracing tool that enhances educational outcomes by modeling hierarchical cognitive states in hyperbolic space.
- Test-Time Scaling with Diffusion Language Models via Reward-Guided Stitching (viability: 8): https://sciencetostartup.com/paper/test-time-scaling-with-diffusion-language-models-via-reward-guided-stitching - Develop a reward-guided diffusion language model tool for enhanced reasoning across math and coding tasks.
- MEDNA-DFM: A Dual-View FiLM-MoE Model for Explainable DNA Methylation Prediction (viability: 5): https://sciencetostartup.com/paper/medna-dfm-a-dual-view-film-moe-model-for-explainable-dna-methylation-prediction - A dual-view model for explainable DNA methylation prediction with high performance and biological insight.
- Decentralized Ranking Aggregation: Gossip Algorithms for Borda and Copeland Consensus (viability: 4): https://sciencetostartup.com/paper/decentralized-ranking-aggregation-gossip-algorithms-for-borda-and-copeland-consensus - Decentralized ranking aggregation algorithms for efficient consensus in distributed networks.
- The AI Research Assistant: Promise, Peril, and a Proof of Concept (viability: 3): https://sciencetostartup.com/paper/the-ai-research-assistant-promise-peril-and-a-proof-of-concept - An empirical case study using AI assistants to accelerate mathematical theorem discovery highlights both opportunities and limitations of AI-aided research.
- DeepPresenter: Environment-Grounded Reflection for Agentic Presentation Generation (viability: 8): https://sciencetostartup.com/paper/deeppresenter-environment-grounded-reflection-for-agentic-presentation-generation - DeepPresenter offers an adaptive, feedback-driven AI framework for automated presentation generation and refinement.
- Moral Preferences of LLMs Under Directed Contextual Influence (viability: 4): https://sciencetostartup.com/paper/moral-preferences-of-llms-under-directed-contextual-influence - Develop a tool to analyze and visualize how contextual influences affect moral decision-making of language models.
- TCM-DiffRAG: Personalized Syndrome Differentiation Reasoning Method for Traditional Chinese Medicine based on Knowledge Graph and Chain of Thought (viability: 7): https://sciencetostartup.com/paper/tcm-diffrag-personalized-syndrome-differentiation-reasoning-method-for-traditional-chinese-medicine-based-on-knowledge-g - A specialized RAG framework integrating knowledge graphs for personalized TCM diagnosis.
- FlexMS is a flexible framework for benchmarking deep learning-based mass spectrum prediction tools in metabolomics (viability: 5): https://sciencetostartup.com/paper/flexms-is-a-flexible-framework-for-benchmarking-deep-learning-based-mass-spectrum-prediction-tools-in-metabolomics - FlexMS is a flexible benchmarking framework for deep learning-based mass spectrum prediction in metabolomics.
- Hierarchy-of-Groups Policy Optimization for Long-Horizon Agentic Tasks (viability: 6): https://sciencetostartup.com/paper/hierarchy-of-groups-policy-optimization-for-long-horizon-agentic-tasks - HGPO enhances agentic RL by reducing bias in advantage estimation for improved long-horizon task performance.
- When Should an AI Act? A Human-Centered Model of Scene, Context, and Behavior for Agentic AI Design (viability: 3): https://sciencetostartup.com/paper/when-should-an-ai-act-a-human-centered-model-of-scene-context-and-behavior-for-agentic-ai-design - Conceptual model for designing contextually aware agentic AI systems.
- MiroFlow: Towards High-Performance and Robust Open-Source Agent Framework for General Deep Research Tasks (viability: 5): https://sciencetostartup.com/paper/miroflow-towards-high-performance-and-robust-open-source-agent-framework-for-general-deep-research-tasks - Open-source framework enhancing agent autonomy for diverse research tasks with state-of-the-art performance.
- Unleashing the Potential of Diffusion Models for End-to-End Autonomous Driving (viability: 8): https://sciencetostartup.com/paper/unleashing-the-potential-of-diffusion-models-for-end-to-end-autonomous-driving - "Hyper Diffusion Planner leverages advanced diffusion models to enhance the efficiency and safety of autonomous driving systems."
- Natural Language Declarative Prompting (NLD-P): A Modular Governance Method for Prompt Design Under Model Drift (viability: 2): https://sciencetostartup.com/paper/natural-language-declarative-prompting-nld-p-a-modular-governance-method-for-prompt-design-under-model-drift - A governance framework for prompt design addressing model drift in large language models (LLMs).
- Probing for Knowledge Attribution in Large Language Models (viability: 6): https://sciencetostartup.com/paper/probing-for-knowledge-attribution-in-large-language-models - Develop a tool that accurately determines knowledge attribution in large language model outputs to reduce errors and hallucinations.
- QSIM: Mitigating Overestimation in Multi-Agent Reinforcement Learning via Action Similarity Weighted Q-Learning (viability: 6): https://sciencetostartup.com/paper/qsim-mitigating-overestimation-in-multi-agent-reinforcement-learning-via-action-similarity-weighted-q-learning - Develop QSIM to enhance stability and performance in multi-agent reinforcement learning by mitigating Q-value overestimation using action similarity weighted Q-learning.
- TherapyProbe: Generating Design Knowledge for Relational Safety in Mental Health Chatbots Through Adversarial Simulation (viability: 3): https://sciencetostartup.com/paper/therapyprobe-generating-design-knowledge-for-relational-safety-in-mental-health-chatbots-through-adversarial-simulation - TherapyProbe enhances safety in mental health chatbots by identifying relational safety failures in conversational dynamics.
- ClinDet-Bench: Beyond Abstention, Evaluating Judgment Determinability of LLMs in Clinical Decision-Making (viability: 5): https://sciencetostartup.com/paper/clindet-bench-beyond-abstention-evaluating-judgment-determinability-of-llms-in-clinical-decision-making - ClinDet-Bench offers a benchmark for evaluating LLMs' ability to determine judgment sufficiency in clinical scenarios, enhancing safety in high-stakes domains.
- AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications (viability: 6): https://sciencetostartup.com/paper/ama-bench-evaluating-long-horizon-memory-for-agentic-applications - Develop AMA-Agent, an advanced memory system for long-horizon agentic applications, achieving superior results on AMA-Bench.
- Distributed LLM Pretraining During Renewable Curtailment Windows: A Feasibility Study (viability: 3): https://sciencetostartup.com/paper/distributed-llm-pretraining-during-renewable-curtailment-windows-a-feasibility-study - Optimize and reduce emissions of LLM training by leveraging renewable energy during curtailment periods.
- Decomposing Physician Disagreement in HealthBench (viability: 3): https://sciencetostartup.com/paper/decomposing-physician-disagreement-in-healthbench - Analyze and reduce physician disagreement in medical AI evaluations to enhance AI performance understanding.
- Towards Simulating Social Media Users with LLMs: Evaluating the Operational Validity of Conditioned Comment Prediction (viability: 6): https://sciencetostartup.com/paper/towards-simulating-social-media-users-with-llms-evaluating-the-operational-validity-of-conditioned-comment-prediction - Develop a tool to simulate social media user behavior using conditioned comment prediction with LLMs.
- Know What You Know: Metacognitive Entropy Calibration for Verifiable RL Reasoning (viability: 6): https://sciencetostartup.com/paper/know-what-you-know-metacognitive-entropy-calibration-for-verifiable-rl-reasoning - Develop a metacognitive entropy calibration tool for enhancing reasoning performance in large models using RL with verifiable rewards.
- Generative Data Transformation: From Mixed to Unified Data (viability: 6): https://sciencetostartup.com/paper/generative-data-transformation-from-mixed-to-unified-data - Taesar offers a data-centric framework for transforming mixed-domain data into unified sequences, enhancing recommendation model performance with less computing power.
- AMLRIS: Alignment-aware Masked Learning for Referring Image Segmentation (viability: 5): https://sciencetostartup.com/paper/amlris-alignment-aware-masked-learning-for-referring-image-segmentation - Develops a new training strategy to enhance image segmentation by improving vision-language alignment.
- Simulation-based Optimization for Augmented Reading (viability: 3): https://sciencetostartup.com/paper/simulation-based-optimization-for-augmented-reading - Simulation-based optimization framework for adaptive and scalable augmented reading systems.
- AgentSentry: Mitigating Indirect Prompt Injection in LLM Agents via Temporal Causal Diagnostics and Context Purification (viability: 6): https://sciencetostartup.com/paper/agentsentry-mitigating-indirect-prompt-injection-in-llm-agents-via-temporal-causal-diagnostics-and-context-purification - AgentSentry is a defensive framework for LLM agents that mitigates indirect prompt injection attacks using temporal causal diagnostics and context purification.
- RLHFless: Serverless Computing for Efficient RLHF (viability: 3): https://sciencetostartup.com/paper/rlhfless-serverless-computing-for-efficient-rlhf - Develop a serverless computing framework, RLHFless, to optimize dynamic resource demands in RLHF training.
- SoPE: Spherical Coordinate-Based Positional Embedding for Enhancing Spatial Perception of 3D LVLMs (viability: 5): https://sciencetostartup.com/paper/sope-spherical-coordinate-based-positional-embedding-for-enhancing-spatial-perception-of-3d-lvlms - Enhance 3D spatial perception in LVLMs with spherical coordinate-based positional embeddings.
- Same Words, Different Judgments: Modality Effects on Preference Alignment (viability: 3): https://sciencetostartup.com/paper/same-words-different-judgments-modality-effects-on-preference-alignment - Exploring the effects of modality on preference alignment to improve AI systems' adherence to human judgments.
- Knob: A Physics-Inspired Gating Interface for Interpretable and Controllable Neural Dynamics (viability: 4): https://sciencetostartup.com/paper/knob-a-physics-inspired-gating-interface-for-interpretable-and-controllable-neural-dynamics - Knob creates a controllable interface for neural networks by integrating classical control theory, allowing human operators to adjust model behavior using intuitive physical analogues.
- IMMACULATE: A Practical LLM Auditing Framework via Verifiable Computation (viability: 8): https://sciencetostartup.com/paper/immaculate-a-practical-llm-auditing-framework-via-verifiable-computation - IMMACULATE provides a practical framework for auditing LLM API services to detect economic abuses like model substitution and token overbilling.
- Tokenization, Fusion and Decoupling: Bridging the Granularity Mismatch Between Large Language Models and Knowledge Graphs (viability: 5): https://sciencetostartup.com/paper/tokenization-fusion-and-decoupling-bridging-the-granularity-mismatch-between-large-language-models-and-knowledge-graphs - KGT enhances knowledge graph completion by bridging the granularity mismatch with large language models using specialized tokenization and a relation-guided gating mechanism.
- Reinforcing Real-world Service Agents: Balancing Utility and Cost in Task-oriented Dialogue (viability: 7): https://sciencetostartup.com/paper/reinforcing-real-world-service-agents-balancing-utility-and-cost-in-task-oriented-dialogue - Develop a cost-aware dialogue agent framework for task-oriented applications, balancing utility and budget constraints.
- SUPERGLASSES: Benchmarking Vision Language Models as Intelligent Agents for AI Smart Glasses (viability: 8): https://sciencetostartup.com/paper/superglasses-benchmarking-vision-language-models-as-intelligent-agents-for-ai-smart-glasses - Leverage SUPERGLASSES to enhance AI-powered smart glasses with superior VQA capabilities tailored for real-world scenarios.
- Toward Personalized LLM-Powered Agents: Foundations, Evaluation, and Future Directions (viability: 4): https://sciencetostartup.com/paper/toward-personalized-llm-powered-agents-foundations-evaluation-and-future-directions - Develop a structured framework for designing personalized LLM-powered agents to enhance user interaction and adaptability.
- ViCLIP-OT: The First Foundation Vision-Language Model for Vietnamese Image-Text Retrieval with Optimal Transport (viability: 5): https://sciencetostartup.com/paper/viclip-ot-the-first-foundation-vision-language-model-for-vietnamese-image-text-retrieval-with-optimal-transport - A vision-language model for Vietnamese image-text retrieval using innovative loss functions to boost performance in low-resource settings.
- dLLM: Simple Diffusion Language Modeling (viability: 8): https://sciencetostartup.com/paper/dllm-simple-diffusion-language-modeling - dLLM unifies diffusion language modeling components into a customizable, open-source framework for easy deployment and evaluation.
- AHBid: An Adaptable Hierarchical Bidding Framework for Cross-Channel Advertising (viability: 7): https://sciencetostartup.com/paper/ahbid-an-adaptable-hierarchical-bidding-framework-for-cross-channel-advertising - Introducing AHBid, a cross-channel advertising tool that boosts ROI by 13.57% using adaptable, generative planning for budget allocation.
- MobilityBench: A Benchmark for Evaluating Route-Planning Agents in Real-World Mobility Scenarios (viability: 6): https://sciencetostartup.com/paper/mobilitybench-a-benchmark-for-evaluating-route-planning-agents-in-real-world-mobility-scenarios - Benchmark and toolkit for evaluating LLM-based route-planning agents in real-world scenarios.
- Instruction-based Image Editing with Planning, Reasoning, and Generation (viability: 6): https://sciencetostartup.com/paper/instruction-based-image-editing-with-planning-reasoning-and-generation - Develop a multi-modality model for instruction-based image editing that combines planning, reasoning, and generation to enhance editing quality.
- ContextRL: Enhancing MLLM's Knowledge Discovery Efficiency with Context-Augmented RL (viability: 5): https://sciencetostartup.com/paper/contextrl-enhancing-mllm-s-knowledge-discovery-efficiency-with-context-augmented-rl - ContextRL enhances knowledge discovery efficiency in ML models by using context-augmented reinforcement learning to filter out low-quality reasoning and recover correct responses.
- CGSA: Class-Guided Slot-Aware Adaptation for Source-Free Object Detection (viability: 6): https://sciencetostartup.com/paper/cgsa-class-guided-slot-aware-adaptation-for-source-free-object-detection - Develop an Object-Centric Learning framework for enhancing Source-Free Domain Adaptive Object Detection in privacy-sensitive scenarios.
- CoLyricist: Enhancing Lyric Writing with AI through Workflow-Aligned Support (viability: 5): https://sciencetostartup.com/paper/colyricist-enhancing-lyric-writing-with-ai-through-workflow-aligned-support - CoLyricist is an AI-enhanced tool that optimizes the lyric writing process with targeted support for each songwriting stage.
- SideQuest: Model-Driven KV Cache Management for Long-Horizon Agentic Reasoning (viability: 3): https://sciencetostartup.com/paper/sidequest-model-driven-kv-cache-management-for-long-horizon-agentic-reasoning - Develop KV cache management for efficient long-horizon agentic reasoning using model-driven compression techniques.
- Transformers converge to invariant algorithmic cores (viability: 2): https://sciencetostartup.com/paper/transformers-converge-to-invariant-algorithmic-cores - This research identifies invariant algorithmic cores in transformers but lacks direct productization potential.
- BetterScene: 3D Scene Synthesis with Representation-Aligned Generative Model (viability: 6): https://sciencetostartup.com/paper/betterscene-3d-scene-synthesis-with-representation-aligned-generative-model - BetterScene offers enhanced novel view synthesis for 3D scenes using sparse photos, outmatching current state-of-the-art with alignment-focused generative models.
- TabDLM: Free-Form Tabular Data Generation via Joint Numerical-Language Diffusion (viability: 6): https://sciencetostartup.com/paper/tabdlm-free-form-tabular-data-generation-via-joint-numerical-language-diffusion - Develop a unified framework for generating synthetic tabular data with mixed numeric and textual fields using diffusion models.
- Correcting Human Labels for Rater Effects in AI Evaluation: An Item Response Theory Approach (viability: 5): https://sciencetostartup.com/paper/correcting-human-labels-for-rater-effects-in-ai-evaluation-an-item-response-theory-approach - Develop a tool using psychometric models to enhance the reliability of human evaluations in AI systems by correcting rater effects.
- Strategy Executability in Mathematical Reasoning: Leveraging Human-Model Differences for Effective Guidance (viability: 8): https://sciencetostartup.com/paper/strategy-executability-in-mathematical-reasoning-leveraging-human-model-differences-for-effective-guidance - Selective Strategy Retrieval enhances mathematical reasoning accuracy by leveraging model-executability insights for compact models.
- S2O: Early Stopping for Sparse Attention via Online Permutation (viability: 3): https://sciencetostartup.com/paper/s2o-early-stopping-for-sparse-attention-via-online-permutation - S2O optimizes sparse attention through early stopping and online permutation to increase computational efficiency in long-context inference.
- Guidance Matters: Rethinking the Evaluation Pitfall for Text-to-Image Generation (viability: 3): https://sciencetostartup.com/paper/guidance-matters-rethinking-the-evaluation-pitfall-for-text-to-image-generation - An evaluation framework to calibrate guidance scales in diffusion models for fair comparison in text-to-image generation.
- Quality-Aware Robust Multi-View Clustering for Heterogeneous Observation Noise (viability: 7): https://sciencetostartup.com/paper/quality-aware-robust-multi-view-clustering-for-heterogeneous-observation-noise - Develop robust multi-view clustering software to handle heterogeneous noise with superior accuracy.
- Addressing Climate Action Misperceptions with Generative AI (viability: 5): https://sciencetostartup.com/paper/addressing-climate-action-misperceptions-with-generative-ai - Develop a personalised climate action tool using LLMs to enhance awareness and motivate pro-climate behaviours.
- Operationalizing Fairness: Post-Hoc Threshold Optimization Under Hard Resource Limits (viability: 6): https://sciencetostartup.com/paper/operationalizing-fairness-post-hoc-threshold-optimization-under-hard-resource-limits - A legal-compliant, model-agnostic ML fairness framework balancing safety, efficiency, and equity under strict resource limits.
- CourtGuard: A Model-Agnostic Framework for Zero-Shot Policy Adaptation in LLM Safety (viability: 7): https://sciencetostartup.com/paper/courtguard-a-model-agnostic-framework-for-zero-shot-policy-adaptation-in-llm-safety - A framework to dynamically adapt LLM safety policies without retraining, enabling cost-effective compliance with evolving regulations.
- Stable Adaptive Thinking via Advantage Shaping and Length-Aware Gradient Regulation (viability: 3): https://sciencetostartup.com/paper/stable-adaptive-thinking-via-advantage-shaping-and-length-aware-gradient-regulation - Innovative framework enhances large reasoning models to reduce overthinking and improve efficiency.
- Autoregressive Visual Decoding from EEG Signals (viability: 7): https://sciencetostartup.com/paper/autoregressive-visual-decoding-from-eeg-signals - Develop a lightweight EEG-based visual decoding tool for efficient brain-computer interface applications.
- DrivePTS: A Progressive Learning Framework with Textual and Structural Enhancement for Driving Scene Generation (viability: 6): https://sciencetostartup.com/paper/drivepts-a-progressive-learning-framework-with-textual-and-structural-enhancement-for-driving-scene-generation - DrivePTS enhances driving scene generation with progressive learning and textual/structural enhancement for superior fidelity and controllability.
- Requesting Expert Reasoning: Augmenting LLM Agents with Learned Collaborative Intervention (viability: 5): https://sciencetostartup.com/paper/requesting-expert-reasoning-augmenting-llm-agents-with-learned-collaborative-intervention - AHCE enhances AI agents with structured human collaboration to improve decision-making in specialized domains.
- DisQ-HNet: A Disentangled Quantized Half-UNet for Interpretable Multimodal Image Synthesis Applications to Tau-PET Synthesis from T1 and FLAIR MRI (viability: 2): https://sciencetostartup.com/paper/disq-hnet-a-disentangled-quantized-half-unet-for-interpretable-multimodal-image-synthesis-applications-to-tau-pet-synthe - DisQ-HNet is a framework for synthesizing tau-PET images from MRI data to support Alzheimer's diagnostics with interpretability.
- Ruyi2 Technical Report (viability: 3): https://sciencetostartup.com/paper/ruyi2-technical-report - Ruyi2 introduces an adaptive model framework for efficient variable-depth computation in LLMs leveraging familial parameter sharing.
- Agentic AI for Intent-driven Optimization in Cell-free O-RAN (viability: 5): https://sciencetostartup.com/paper/agentic-ai-for-intent-driven-optimization-in-cell-free-o-ran - Develop a multi-agent AI system for intent-driven optimization in cell-free O-RAN to improve energy efficiency.
- Generative Agents Navigating Digital Libraries (viability: 5): https://sciencetostartup.com/paper/generative-agents-navigating-digital-libraries - Agent4DL simulates user search behavior in digital libraries to enhance dataset accuracy under privacy constraints.
- Predicting Tennis Serve directions with Machine Learning (viability: 2): https://sciencetostartup.com/paper/predicting-tennis-serve-directions-with-machine-learning - Leverage machine learning to predict tennis serve directions and enhance gameplay strategy.
- Iterative Prompt Refinement for Dyslexia-Friendly Text Summarization Using GPT-4o (viability: 6): https://sciencetostartup.com/paper/iterative-prompt-refinement-for-dyslexia-friendly-text-summarization-using-gpt-4o - Empowering dyslexic readers with accessible, easy-to-read text summaries using GPT-4o.
- Cognitive Models and AI Algorithms Provide Templates for Designing Language Agents (viability: 3): https://sciencetostartup.com/paper/cognitive-models-and-ai-algorithms-provide-templates-for-designing-language-agents - Developing modular LLM-based agents using cognitive model and AI algorithm templates.
- Efficient Dialect-Aware Modeling and Conditioning for Low-Resource Taiwanese Hakka Speech Processing (viability: 7): https://sciencetostartup.com/paper/efficient-dialect-aware-modeling-and-conditioning-for-low-resource-taiwanese-hakka-speech-processing - Create a dialect-aware ASR tool tailored for the low-resource Taiwanese Hakka language, significantly reducing error rates.
- A Mathematical Theory of Agency and Intelligence (viability: 4): https://sciencetostartup.com/paper/a-mathematical-theory-of-agency-and-intelligence - Introducing 'bipredictability' as a new metric, enabling AI systems to enhance their operational resilience and adaptability.
- SignVLA: A Gloss-Free Vision-Language-Action Framework for Real-Time Sign Language-Guided Robotic Manipulation (viability: 7): https://sciencetostartup.com/paper/signvla-a-gloss-free-vision-language-action-framework-for-real-time-sign-language-guided-robotic-manipulation - A gloss-free vision-language-action system enabling real-time sign language-guided robotic manipulation.
- Mirroring the Mind: Distilling Human-Like Metacognitive Strategies into Large Language Models (viability: 7): https://sciencetostartup.com/paper/mirroring-the-mind-distilling-human-like-metacognitive-strategies-into-large-language-models - Metacognitive Behavioral Tuning enhances reasoning stability in language models for complex tasks.
- Mapping the Landscape of Artificial Intelligence in Life Cycle Assessment Using Large Language Models (viability: 4): https://sciencetostartup.com/paper/mapping-the-landscape-of-artificial-intelligence-in-life-cycle-assessment-using-large-language-models - Leveraging LLMs to enhance large-scale literature reviews in life cycle assessment, offering insights for sustainability decisions.
- Reinforcement-aware Knowledge Distillation for LLM Reasoning (viability: 5): https://sciencetostartup.com/paper/reinforcement-aware-knowledge-distillation-for-llm-reasoning - Revolutionize LLM reasoning with efficient, RL-aware knowledge distillation for lower inference costs.
- Explainability-Aware Evaluation of Transfer Learning Models for IoT DDoS Detection Under Resource Constraints (viability: 5): https://sciencetostartup.com/paper/explainability-aware-evaluation-of-transfer-learning-models-for-iot-ddos-detection-under-resource-constraints - Develop an explainability-focused IoT DDoS detection tool leveraging transfer learning models for improved reliability and resource efficiency.
- Importance of Prompt Optimisation for Error Detection in Medical Notes Using Language Models (viability: 7): https://sciencetostartup.com/paper/importance-of-prompt-optimisation-for-error-detection-in-medical-notes-using-language-models - Optimize language model prompts to enhance error detection in medical texts, approaching doctor-level accuracy.
- Sydney Telling Fables on AI and Humans: A Corpus Tracing Memetic Transfer of Persona between LLMs (viability: 2): https://sciencetostartup.com/paper/sydney-telling-fables-on-ai-and-humans-a-corpus-tracing-memetic-transfer-of-persona-between-llms - Explore AI-human relationships with the Sydney persona corpus tracing memetic transfer across LLMs.
- VeRO: An Evaluation Harness for Agents to Optimize Agents (viability: 7): https://sciencetostartup.com/paper/vero-an-evaluation-harness-for-agents-to-optimize-agents - VERO is an evaluation harness designed to systematically optimize coding agents through structured versioning and benchmarking, providing researchers a tool to enhance agent performance.
- Beyond Dominant Patches: Spatial Credit Redistribution For Grounded Vision-Language Models (viability: 8): https://sciencetostartup.com/paper/beyond-dominant-patches-spatial-credit-redistribution-for-grounded-vision-language-models - A practical solution to reduce hallucination in vision-language models through inference-time spatial credit redistribution.
- ConstraintBench: Benchmarking LLM Constraint Reasoning on Direct Optimization (viability: 6): https://sciencetostartup.com/paper/constraintbench-benchmarking-llm-constraint-reasoning-on-direct-optimization - ConstraintBench evaluates LLMs' capability in directly solving constrained optimization problems without solvers, across operations research domains.
- Automating the Detection of Requirement Dependencies Using Large Language Models (viability: 6): https://sciencetostartup.com/paper/automating-the-detection-of-requirement-dependencies-using-large-language-models - Automate requirement dependency detection in software projects using LLMs for enhanced accuracy and efficiency.
- CWM: Contrastive World Models for Action Feasibility Learning in Embodied Agent Pipelines (viability: 6): https://sciencetostartup.com/paper/cwm-contrastive-world-models-for-action-feasibility-learning-in-embodied-agent-pipelines - Develop action feasibility scorers for embodied agents using contrastive learning to improve reliability and safety in AI action planning.
- Silent Egress: When Implicit Prompt Injection Makes LLM Agents Leak Without a Trace (viability: 4): https://sciencetostartup.com/paper/silent-egress-when-implicit-prompt-injection-makes-llm-agents-leak-without-a-trace - Silent Egress identifies vulnerabilities in LLM agents that allow data exfiltration via implicit prompt injection.
- A Fusion of context-aware based BanglaBERT and Two-Layer Stacked LSTM Framework for Multi-Label Cyberbullying Detection (viability: 6): https://sciencetostartup.com/paper/a-fusion-of-context-aware-based-banglabert-and-two-layer-stacked-lstm-framework-for-multi-label-cyberbullying-detection - A cyberbullying detection tool for Bangla language using a fusion of BanglaBERT and LSTM for multilabel classification.
- ECHO: Encoding Communities via High-order Operators (viability: 7): https://sciencetostartup.com/paper/echo-encoding-communities-via-high-order-operators - ECHO provides scalable and efficient community detection in large-scale attributed networks via high-order operators, overcoming traditional GNN limitations with innovative routing and optimization techniques.
- A Framework for Assessing AI Agent Decisions and Outcomes in AutoML Pipelines (viability: 6): https://sciencetostartup.com/paper/a-framework-for-assessing-ai-agent-decisions-and-outcomes-in-automl-pipelines - Develop a decision-centric evaluation agent for auditing AI agent decisions in AutoML systems to improve interpretability and governance.
- How Do Latent Reasoning Methods Perform Under Weak and Strong Supervision? (viability: 2): https://sciencetostartup.com/paper/how-do-latent-reasoning-methods-perform-under-weak-and-strong-supervision - Analysis of latent reasoning methods showing trade-offs between supervision strength and reasoning accuracy without practical application path.
- From Bias to Balance: Fairness-Aware Paper Recommendation for Equitable Peer Review (viability: 7): https://sciencetostartup.com/paper/from-bias-to-balance-fairness-aware-paper-recommendation-for-equitable-peer-review - Fair-PaperRec is an MLP-based equity-focused recommender system that enhances fairness in peer review by re-ranking papers using fairness regularization.
- veScale-FSDP: Flexible and High-Performance FSDP at Scale (viability: 3): https://sciencetostartup.com/paper/vescale-fsdp-flexible-and-high-performance-fsdp-at-scale - New system veScale-FSDP aims to enhance the flexibility and performance of existing FSDP systems for large-scale model training.
- GetBatch: Distributed Multi-Object Retrieval for ML Data Loading (viability: 4): https://sciencetostartup.com/paper/getbatch-distributed-multi-object-retrieval-for-ml-data-loading - GetBatch optimizes ML data loading with a new API for efficient multi-object retrieval, significantly improving throughput and latency.
- Calibrated Test-Time Guidance for Bayesian Inference (viability: 2): https://sciencetostartup.com/paper/calibrated-test-time-guidance-for-bayesian-inference - Develop calibrated estimators for Bayesian inference to improve test-time guidance in diffusion models.
- HubScan: Detecting Hubness Poisoning in Retrieval-Augmented Generation Systems (viability: 9): https://sciencetostartup.com/paper/hubscan-detecting-hubness-poisoning-in-retrieval-augmented-generation-systems - HubScan detects and mitigates hubness poisoning attacks in retrieval-augmented generation systems for secure AI data access.
- ArchAgent: Agentic AI-driven Computer Architecture Discovery (viability: 5): https://sciencetostartup.com/paper/archagent-agentic-ai-driven-computer-architecture-discovery - ArchAgent uses agentic AI to discover superior computer architecture designs, outperforming state-of-the-art methods in speed and efficiency.
- Revisiting Chebyshev Polynomial and Anisotropic RBF Models for Tabular Regression (viability: 8): https://sciencetostartup.com/paper/revisiting-chebyshev-polynomial-and-anisotropic-rbf-models-for-tabular-regression - Develop Scikit-learn-compatible smooth-basis models for improved generalization in CPU-constrained tabular regression tasks.
- Epistemic Filtering and Collective Hallucination: A Jury Theorem for Confidence-Calibrated Agents (viability: 3): https://sciencetostartup.com/paper/epistemic-filtering-and-collective-hallucination-a-jury-theorem-for-confidence-calibrated-agents - Develop a confidence-calibration framework for heterogeneous agents to improve collective decision-making accuracy.
- Exploring Human Behavior During Abstract Rule Inference and Problem Solving with the Cognitive Abstraction and Reasoning Corpus (viability: 4): https://sciencetostartup.com/paper/exploring-human-behavior-during-abstract-rule-inference-and-problem-solving-with-the-cognitive-abstraction-and-reasoning - Develop a tool for analyzing human abstract reasoning using the Cognitive Abstraction and Reasoning Corpus (CogARC).
- Towards Autonomous Memory Agents (viability: 6): https://sciencetostartup.com/paper/towards-autonomous-memory-agents - Develop an autonomous memory enhancement system for LLMs to actively curate and optimize knowledge acquisition.
- Contextual Memory Virtualisation: DAG-Based State Management and Structurally Lossless Trimming for LLM Agents (viability: 7): https://sciencetostartup.com/paper/contextual-memory-virtualisation-dag-based-state-management-and-structurally-lossless-trimming-for-llm-agents - Develop a memory management tool for LLM agents to extend session length without losing context through a DAG-based model.
- Vibe Researching as Wolf Coming: Can AI Agents with Skills Replace or Augment Social Scientists? (viability: 2): https://sciencetostartup.com/paper/vibe-researching-as-wolf-coming-can-ai-agents-with-skills-replace-or-augment-social-scientists - Develop AI agents capable of autonomously executing research pipelines by integrating specialist skills and tools.
- Enhancing Renal Tumor Malignancy Prediction: Deep Learning with Automatic 3D CT Organ Focused Attention (viability: 6): https://sciencetostartup.com/paper/enhancing-renal-tumor-malignancy-prediction-deep-learning-with-automatic-3d-ct-organ-focused-attention - AI-driven tool for non-invasive renal tumor malignancy prediction using 3D CT without manual segmentation.
- AeroDGS: Physically Consistent Dynamic Gaussian Splatting for Single-Sequence Aerial 4D Reconstruction (viability: 5): https://sciencetostartup.com/paper/aerodgs-physically-consistent-dynamic-gaussian-splatting-for-single-sequence-aerial-4d-reconstruction - AeroDGS enhances monocular aerial 4D reconstruction with a physics-guided framework for UAV videos, improving fidelity in dynamic environments.
- EyeLayer: Integrating Human Attention Patterns into LLM-Based Code Summarization (viability: 7): https://sciencetostartup.com/paper/eyelayer-integrating-human-attention-patterns-into-llm-based-code-summarization - EyeLayer integrates human eye-gaze patterns into LLMs to enhance code summarization accuracy by up to 13.17%.
- Learning geometry-dependent lead-field operators for forward ECG modeling (viability: 3): https://sciencetostartup.com/paper/learning-geometry-dependent-lead-field-operators-for-forward-ecg-modeling - Develop a shape-informed surrogate model for accurate yet efficient forward ECG simulations without detailed torso segmentation.
- Scaling In, Not Up? Testing Thick Citation Context Analysis with GPT-5 and Fragile Prompts (viability: 2): https://sciencetostartup.com/paper/scaling-in-not-up-testing-thick-citation-context-analysis-with-gpt-5-and-fragile-prompts - Using GPT-5 for thick citation context analysis to enhance interpretative readings in academic works.
- GRAU: Generic Reconfigurable Activation Unit Design for Neural Network Hardware Accelerators (viability: 3): https://sciencetostartup.com/paper/grau-generic-reconfigurable-activation-unit-design-for-neural-network-hardware-accelerators - Develop a reconfigurable activation unit for neural network hardware accelerators to increase hardware efficiency and flexibility.
- Decoder-based Sense Knowledge Distillation (viability: 5): https://sciencetostartup.com/paper/decoder-based-sense-knowledge-distillation - Enhance language model performance with Decoder-based Sense Knowledge Distillation for integrating lexical semantics.
- Enabling clinical use of foundation models in histopathology (viability: 3): https://sciencetostartup.com/paper/enabling-clinical-use-of-foundation-models-in-histopathology - Develops robust models for computational pathology by mitigating pre-analytic variability impacts in foundation models.
- Structure and Redundancy in Large Language Models: A Spectral Study via Random Matrix Theory (viability: 3): https://sciencetostartup.com/paper/structure-and-redundancy-in-large-language-models-a-spectral-study-via-random-matrix-theory - Novel spectral analysis methods for improving reliability and efficiency in large language models.
- A 1/R Law for Kurtosis Contrast in Balanced Mixtures (viability: 2): https://sciencetostartup.com/paper/a-1-r-law-for-kurtosis-contrast-in-balanced-mixtures - Exploring kurtosis contrast in balanced mixtures, offering insight into limitations and restoring contrast through purification methods.
- Recovered in Translation: Efficient Pipeline for Automated Translation of Benchmarks and Datasets (viability: 7): https://sciencetostartup.com/paper/recovered-in-translation-efficient-pipeline-for-automated-translation-of-benchmarks-and-datasets - Automated multilingual benchmark translation for reliable LLM evaluation.
- Off-The-Shelf Image-to-Image Models Are All You Need To Defeat Image Protection Schemes (viability: 5): https://sciencetostartup.com/paper/off-the-shelf-image-to-image-models-are-all-you-need-to-defeat-image-protection-schemes - This tool utilizes off-the-shelf image-to-image models to defeat existing image protection schemes by acting as a denoiser.
- GUI-Libra: Training Native GUI Agents to Reason and Act with Action-aware Supervision and Partially Verifiable RL (viability: 7): https://sciencetostartup.com/paper/gui-libra-training-native-gui-agents-to-reason-and-act-with-action-aware-supervision-and-partially-verifiable-rl - GUI-Libra creates more intelligent and efficient GUI agents for enhancing user experience across web and mobile applications.
- Surrogate models for Rock-Fluid Interaction: A Grid-Size-Invariant Approach (viability: 6): https://sciencetostartup.com/paper/surrogate-models-for-rock-fluid-interaction-a-grid-size-invariant-approach - Develop grid-size-invariant surrogate models for efficient rock-fluid interaction simulations with reduced computational expense.
- Enhancing Framingham Cardiovascular Risk Score Transparency through Logic-Based XAI (viability: 3): https://sciencetostartup.com/paper/enhancing-framingham-cardiovascular-risk-score-transparency-through-logic-based-xai - Develop a logic-based explainer to improve transparency of the Framingham Cardiovascular Risk Score.
- Provable Last-Iterate Convergence for Multi-Objective Safe LLM Alignment via Optimistic Primal-Dual (viability: 3): https://sciencetostartup.com/paper/provable-last-iterate-convergence-for-multi-objective-safe-llm-alignment-via-optimistic-primal-dual - A theoretical framework for stabilizing safe RLHF algorithms in large language models using optimistic primal-dual methods.
- When AI Writes, Whose Voice Remains? Quantifying Cultural Marker Erasure Across World English Varieties in Large Language Models (viability: 5): https://sciencetostartup.com/paper/when-ai-writes-whose-voice-remains-quantifying-cultural-marker-erasure-across-world-english-varieties-in-large-language- - Introducing metrics to measure linguistic identity retention in AI-generated English texts.
- NoLan: Mitigating Object Hallucinations in Large Vision-Language Models via Dynamic Suppression of Language Priors (viability: 7): https://sciencetostartup.com/paper/nolan-mitigating-object-hallucinations-in-large-vision-language-models-via-dynamic-suppression-of-language-priors - A framework to reduce object hallucinations in vision-language models by dynamically suppressing language priors, improving accuracy significantly.
- SWE-Protégé: Learning to Selectively Collaborate With an Expert Unlocks Small Language Models as Software Engineering Agents (viability: 6): https://sciencetostartup.com/paper/swe-prot-g-learning-to-selectively-collaborate-with-an-expert-unlocks-small-language-models-as-software-engineering-agen - Unlock the potential of small language models in software engineering by selective expert collaboration, improving efficiency and performance.
- Don't stop me now: Rethinking Validation Criteria for Model Parameter Selection (viability: 4): https://sciencetostartup.com/paper/don-t-stop-me-now-rethinking-validation-criteria-for-model-parameter-selection - Develop a tool for optimal model parameter selection by rethinking validation criteria to improve generalization performance.
- On Imbalanced Regression with Hoeffding Trees (viability: 5): https://sciencetostartup.com/paper/on-imbalanced-regression-with-hoeffding-trees - Enhance online imbalanced regression with Hoeffding trees using KDE and HS for improved early-stream predictions.
- Petri Net Relaxation for Infeasibility Explanation and Sequential Task Planning (viability: 2): https://sciencetostartup.com/paper/petri-net-relaxation-for-infeasibility-explanation-and-sequential-task-planning - A system utilizing Petri net reachability relaxation for efficient detection of infeasibility and sequential task planning.
- Understanding Artificial Theory of Mind: Perturbed Tasks and Reasoning in Large Language Models (viability: 5): https://sciencetostartup.com/paper/understanding-artificial-theory-of-mind-perturbed-tasks-and-reasoning-in-large-language-models - Develop an AI tool to evaluate and enhance Theory of Mind capabilities in language models using Chain-of-Thought prompting.
- Language Models Exhibit Inconsistent Biases Towards Algorithmic Agents and Human Experts (viability: 5): https://sciencetostartup.com/paper/language-models-exhibit-inconsistent-biases-towards-algorithmic-agents-and-human-experts - Investigating LLMs' inconsistent biases in decision-making involving algorithmic agents and human experts.
- Semantic Partial Grounding via LLMs (viability: 5): https://sciencetostartup.com/paper/semantic-partial-grounding-via-llms - Optimize classical planning grounding using LLMs to significantly reduce computation time and resource usage.
- DualWeaver: Synergistic Feature Weaving Surrogates for Multivariate Forecasting with Univariate Time Series Foundation Models (viability: 7): https://sciencetostartup.com/paper/dualweaver-synergistic-feature-weaving-surrogates-for-multivariate-forecasting-with-univariate-time-series-foundation-mo - DualWeaver transforms univariate time series models into powerful multivariate forecasters with its innovative surrogate feature weaving framework.
- NESTOR: A Nested MOE-based Neural Operator for Large-Scale PDE Pre-Training (viability: 5): https://sciencetostartup.com/paper/nestor-a-nested-moe-based-neural-operator-for-large-scale-pde-pre-training - A specialized nested MoE neural operator for efficient large-scale PDE pre-training and transfer learning.
- Physics-Informed Machine Learning for Vessel Shaft Power and Fuel Consumption Prediction: Interpretable KAN-based Approach (viability: 6): https://sciencetostartup.com/paper/physics-informed-machine-learning-for-vessel-shaft-power-and-fuel-consumption-prediction-interpretable-kan-based-approac - Develop a physics-informed AI model to accurately predict maritime vessel shaft power and fuel consumption for operational efficiency.
- RGB-Event HyperGraph Prompt for Kilometer Marker Recognition based on Pre-trained Foundation Models (viability: 7): https://sciencetostartup.com/paper/rgb-event-hypergraph-prompt-for-kilometer-marker-recognition-based-on-pre-trained-foundation-models - Develop an advanced Kilometer Marker Recognition system using multi-modal adaptation with RGB and event cameras for autonomous metro localization.
- Enhancing LLM-Based Test Generation by Eliminating Covered Code (viability: 7): https://sciencetostartup.com/paper/enhancing-llm-based-test-generation-by-eliminating-covered-code - Automated test generation tool that enhances code coverage for complex Python projects by eliminating already-covered code parts.
- PatchDenoiser: Parameter-efficient multi-scale patch learning and fusion denoiser for medical images (viability: 8): https://sciencetostartup.com/paper/patchdenoiser-parameter-efficient-multi-scale-patch-learning-and-fusion-denoiser-for-medical-images - A lightweight AI denoiser for medical images that outperforms traditional methods while drastically reducing parameters and energy use.
- Hidden Topics: Measuring Sensitive AI Beliefs with List Experiments (viability: 3): https://sciencetostartup.com/paper/hidden-topics-measuring-sensitive-ai-beliefs-with-list-experiments - Using list experiments to uncover hidden beliefs in AI models mirrors techniques for reducing biases in human surveys.
- A Framework for Cross-Domain Generalization in Coronary Artery Calcium Scoring Across Gated and Non-Gated Computed Tomography (viability: 6): https://sciencetostartup.com/paper/a-framework-for-cross-domain-generalization-in-coronary-artery-calcium-scoring-across-gated-and-non-gated-computed-tomog - Cross-domain coronary artery calcium scoring framework enables broader cardiovascular risk assessment with CT scans.
- 2-Step Agent: A Framework for the Interaction of a Decision Maker with AI Decision Support (viability: 3): https://sciencetostartup.com/paper/2-step-agent-a-framework-for-the-interaction-of-a-decision-maker-with-ai-decision-support - Framework for understanding the impact of AI decision support on human decision making using Bayesian causal inference.
- DynamicGTR: Leveraging Graph Topology Representation Preferences to Boost VLM Capabilities on Graph QAs (viability: 6): https://sciencetostartup.com/paper/dynamicgtr-leveraging-graph-topology-representation-preferences-to-boost-vlm-capabilities-on-graph-qas - DynamicGTR enhances zero-shot QA capabilities of vision-language models on graph-related queries by dynamically selecting optimal graph topology representations.
- ProactiveMobile: A Comprehensive Benchmark for Boosting Proactive Intelligence on Mobile Devices (viability: 4): https://sciencetostartup.com/paper/proactivemobile-a-comprehensive-benchmark-for-boosting-proactive-intelligence-on-mobile-devices - ProactiveMobile benchmark enhances proactive intelligence in mobile agents by providing a real-world complexity framework for evaluating proactive MLLMs.
- Distill and Align Decomposition for Enhanced Claim Verification (viability: 2): https://sciencetostartup.com/paper/distill-and-align-decomposition-for-enhanced-claim-verification - A reinforcement learning framework that enhances claim verification by optimizing decomposition quality and verifier alignment.
- Understanding Annotation Error Propagation and Learning an Adaptive Policy for Expert Intervention in Barrett's Video Segmentation (viability: 6): https://sciencetostartup.com/paper/understanding-annotation-error-propagation-and-learning-an-adaptive-policy-for-expert-intervention-in-barrett-s-video-se - A cost-aware framework that improves annotation accuracy in Barrett's video segmentation by adapting expert intervention based on error propagation analysis.
- xai-cola: A Python library for sparsifying counterfactual explanations (viability: 8): https://sciencetostartup.com/paper/xai-cola-a-python-library-for-sparsifying-counterfactual-explanations - xai-cola provides a Python library for more actionable counterfactual AI explanations by sparsifying modified features.
- Resilient Federated Chain: Transforming Blockchain Consensus into an Active Defense Layer for Federated Learning (viability: 7): https://sciencetostartup.com/paper/resilient-federated-chain-transforming-blockchain-consensus-into-an-active-defense-layer-for-federated-learning - Resilient Federated Chain leverages blockchain technology to bolster federated learning's resilience against adversarial attacks, ensuring secure decentralized training.
- StoryMovie: A Dataset for Semantic Alignment of Visual Stories with Movie Scripts and Subtitles (viability: 5): https://sciencetostartup.com/paper/storymovie-a-dataset-for-semantic-alignment-of-visual-stories-with-movie-scripts-and-subtitles - Leverage StoryMovie to improve semantic alignment in visual storytelling with precise dialogue and relationship attribution.
- SemVideo: Reconstructs What You Watch from Brain Activity via Hierarchical Semantic Guidance (viability: 6): https://sciencetostartup.com/paper/semvideo-reconstructs-what-you-watch-from-brain-activity-via-hierarchical-semantic-guidance - SemVideo leverages hierarchical semantic guidance to reconstruct coherent videos from fMRI data with enhanced semantic alignment and temporal consistency.
- Prompt Architecture Determines Reasoning Quality: A Variable Isolation Study on the Car Wash Problem (viability: 4): https://sciencetostartup.com/paper/prompt-architecture-determines-reasoning-quality-a-variable-isolation-study-on-the-car-wash-problem - Develop a reasoning framework to significantly improve large language models' performance on implicit constraint tasks like the car wash problem.
- An Evaluation of Context Length Extrapolation in Long Code via Positional Embeddings and Efficient Attention (viability: 2): https://sciencetostartup.com/paper/an-evaluation-of-context-length-extrapolation-in-long-code-via-positional-embeddings-and-efficient-attention - Optimize long code sequence processing in LLMs through enhanced positional embeddings and attention mechanisms.
- Excitation: Momentum For Experts (viability: 3): https://sciencetostartup.com/paper/excitation-momentum-for-experts - Develop a lightweight optimizer framework that dynamically modulates parameter updates for improved Mixture-of-Experts training.
- UniWhisper: Efficient Continual Multi-task Training for Robust Universal Audio Representation (viability: 2): https://sciencetostartup.com/paper/uniwhisper-efficient-continual-multi-task-training-for-robust-universal-audio-representation - UniWhisper offers a unified framework for training robust universal audio representations across speech, environmental sounds, and music.
- Generalisation of RLHF under Reward Shift and Clipped KL Regularisation (viability: 3): https://sciencetostartup.com/paper/generalisation-of-rlhf-under-reward-shift-and-clipped-kl-regularisation - Develop a generalisation theory for RLHF under reward shift and clipped KL regularisation to enhance theoretical understanding.
- fEDM+: A Risk-Based Fuzzy Ethical Decision Making Framework with Principle-Level Explainability and Pluralistic Validation (viability: 3): https://sciencetostartup.com/paper/fedm-a-risk-based-fuzzy-ethical-decision-making-framework-with-principle-level-explainability-and-pluralistic-validation - Develop an ethical decision-making oversight tool leveraging fuzzy logic for enhanced transparency and pluralistic validation in AI systems.
- Evaluating the relationship between regularity and learnability in recursive numeral systems using Reinforcement Learning (viability: 2): https://sciencetostartup.com/paper/evaluating-the-relationship-between-regularity-and-learnability-in-recursive-numeral-systems-using-reinforcement-learnin - Study explores the learnability of regular numeral systems using reinforcement learning, focusing on linguistic applications.
- SurGo-R1: Benchmarking and Modeling Contextual Reasoning for Operative Zone in Surgical Video (viability: 5): https://sciencetostartup.com/paper/surgo-r1-benchmarking-and-modeling-contextual-reasoning-for-operative-zone-in-surgical-video - SurGo-R1 enhances surgical video analysis by improving contextual reasoning for identifying operative zones during minimally invasive procedures.
- Dynamic Multimodal Activation Steering for Hallucination Mitigation in Large Vision-Language Models (viability: 6): https://sciencetostartup.com/paper/dynamic-multimodal-activation-steering-for-hallucination-mitigation-in-large-vision-language-models - Develop a training-free hallucination mitigation tool for Large Vision-Language Models using dynamic activation steering vectors.
- Hierarchical LLM-Based Multi-Agent Framework with Prompt Optimization for Multi-Robot Task Planning (viability: 7): https://sciencetostartup.com/paper/hierarchical-llm-based-multi-agent-framework-with-prompt-optimization-for-multi-robot-task-planning - A multi-agent LLM-based framework optimizes robotic task planning by reducing execution failures through improved prompt optimization.
- Following the Diagnostic Trace: Visual Cognition-guided Cooperative Network for Chest X-Ray Diagnosis (viability: 7): https://sciencetostartup.com/paper/following-the-diagnostic-trace-visual-cognition-guided-cooperative-network-for-chest-x-ray-diagnosis - Visual Cognition-Guided Cooperative Network that enhances chest X-ray diagnosis with integrated radiologist tooling for interactive clinical workflows.
- CCCaption: Dual-Reward Reinforcement Learning for Complete and Correct Image Captioning (viability: 6): https://sciencetostartup.com/paper/cccaption-dual-reward-reinforcement-learning-for-complete-and-correct-image-captioning - Develop enhanced image captioning models using dual-reward reinforcement learning for more complete and correct descriptions.
- Sparsity Induction for Accurate Post-Training Pruning of Large Language Models (viability: 5): https://sciencetostartup.com/paper/sparsity-induction-for-accurate-post-training-pruning-of-large-language-models - Build a tool for enhancing sparsity in large language models to improve post-training pruning performance.
- PPCR-IM: A System for Multi-layer DAG-based Public Policy Consequence Reasoning and Social Indicator Mapping (viability: 6): https://sciencetostartup.com/paper/ppcr-im-a-system-for-multi-layer-dag-based-public-policy-consequence-reasoning-and-social-indicator-mapping - PPCR-IM provides a DAG-based system for structured reasoning and mapping of public policy impacts using large language models.
- Mitigating Structural Noise in Low-Resource S2TT: An Optimized Cascaded Nepali-English Pipeline with Punctuation Restoration (viability: 7): https://sciencetostartup.com/paper/mitigating-structural-noise-in-low-resource-s2tt-an-optimized-cascaded-nepali-english-pipeline-with-punctuation-restorat - A state-of-the-art Nepali-to-English speech-to-text translation system for low-resource languages.
- Self-Correcting VLA: Online Action Refinement via Sparse World Imagination (viability: 6): https://sciencetostartup.com/paper/self-correcting-vla-online-action-refinement-via-sparse-world-imagination - SC-VLA enhances vision-language-action models with self-improvement through sparse imagination for better robotic task execution.
- Virtual Biopsy for Intracranial Tumors Diagnosis on MRI (viability: 8): https://sciencetostartup.com/paper/virtual-biopsy-for-intracranial-tumors-diagnosis-on-mri - Non-invasive MRI-based diagnostic solution for deep intracranial tumors improves accuracy and safety over traditional biopsy methods.
- Structurally Aligned Subtask-Level Memory for Software Engineering Agents (viability: 6): https://sciencetostartup.com/paper/structurally-aligned-subtask-level-memory-for-software-engineering-agents - Empowering software engineering agents with subtask-level memory for enhanced reasoning and problem-solving.
- Retrieval Challenges in Low-Resource Public Service Information: A Case Study on Food Pantry Access (viability: 4): https://sciencetostartup.com/paper/retrieval-challenges-in-low-resource-public-service-information-a-case-study-on-food-pantry-access - AI-powered conversational retrieval system for improving food pantry information access.
- Exploring Human-Machine Coexistence in Symmetrical Reality (viability: 2): https://sciencetostartup.com/paper/exploring-human-machine-coexistence-in-symmetrical-reality - Exploring new paradigms for harmonious human-machine coexistence in both virtual and physical worlds.
- Power and Limitations of Aggregation in Compound AI Systems (viability: 2): https://sciencetostartup.com/paper/power-and-limitations-of-aggregation-in-compound-ai-systems - Exploring the theoretical capabilities and boundaries of aggregation in homogeneous AI multi-agent systems.
- Revisiting RAG Retrievers: An Information Theoretic Benchmark (viability: 5): https://sciencetostartup.com/paper/revisiting-rag-retrievers-an-information-theoretic-benchmark - MIGRASCOPE enhances RAG system efficiency by providing metrics for retriever selection and ensemble configuration.
- From Basis to Basis: Gaussian Particle Representation for Interpretable PDE Operators (viability: 5): https://sciencetostartup.com/paper/from-basis-to-basis-gaussian-particle-representation-for-interpretable-pde-operators - Develop a Gaussian Particle Operator for interpretable and efficient PDE simulations with near-linear complexity and mesh-agnostic geometry.
- Enhancing Multilingual Embeddings via Multi-Way Parallel Text Alignment (viability: 3): https://sciencetostartup.com/paper/enhancing-multilingual-embeddings-via-multi-way-parallel-text-alignment - Improve multilingual embeddings through multi-way parallel text alignment for enhanced NLP task performance.
- ARLArena: A Unified Framework for Stable Agentic Reinforcement Learning (viability: 3): https://sciencetostartup.com/paper/arlarena-a-unified-framework-for-stable-agentic-reinforcement-learning - ARLArena aims to enhance stability in agentic reinforcement learning by providing a unified framework and stable optimization methods.
- LiLo-VLA: Compositional Long-Horizon Manipulation via Linked Object-Centric Policies (viability: 6): https://sciencetostartup.com/paper/lilo-vla-compositional-long-horizon-manipulation-via-linked-object-centric-policies - LiLo-VLA enables robust, zero-shot, long-horizon robot manipulation via modular object-centric skills.
- Training Generalizable Collaborative Agents via Strategic Risk Aversion (viability: 7): https://sciencetostartup.com/paper/training-generalizable-collaborative-agents-via-strategic-risk-aversion - Developing adaptable collaborative agents through strategic risk aversion for improved partner interaction.
- Beyond Refusal: Probing the Limits of Agentic Self-Correction for Semantic Sensitive Information (viability: 2): https://sciencetostartup.com/paper/beyond-refusal-probing-the-limits-of-agentic-self-correction-for-semantic-sensitive-information - Develop an inference-time framework to enhance LLM self-correction for semantic sensitive information without sacrificing utility.
- GradAlign: Gradient-Aligned Data Selection for LLM Reinforcement Learning (viability: 6): https://sciencetostartup.com/paper/gradalign-gradient-aligned-data-selection-for-llm-reinforcement-learning - GradAlign optimizes reinforcement learning for large language models by selecting gradient-aligned data, improving training stability and performance.
- Revisiting Text Ranking in Deep Research (viability: 8): https://sciencetostartup.com/paper/revisiting-text-ranking-in-deep-research - A new approach to text ranking for deep research with code and dataset available, ready for application in search products.
- Adversarial Robustness of Deep Learning-Based Thyroid Nodule Segmentation in Ultrasound (viability: 2): https://sciencetostartup.com/paper/adversarial-robustness-of-deep-learning-based-thyroid-nodule-segmentation-in-ultrasound - Develop adversarial robustness defenses for ultrasound thyroid nodule segmentation against spatial and frequency-domain attacks.
- Adversarial Intent is a Latent Variable: Stateful Trust Inference for Securing Multimodal Agentic RAG (viability: 5): https://sciencetostartup.com/paper/adversarial-intent-is-a-latent-variable-stateful-trust-inference-for-securing-multimodal-agentic-rag - Secure multimodal agent systems from adversarial attacks using a stateful inference framework.
- MINAR: Mechanistic Interpretability for Neural Algorithmic Reasoning (viability: 4): https://sciencetostartup.com/paper/minar-mechanistic-interpretability-for-neural-algorithmic-reasoning - Introducing MINAR, a toolbox for discovering neuron-level circuits in GNNs trained for algorithmic reasoning.
- Causal Decoding for Hallucination-Resistant Multimodal Large Language Models (viability: 6): https://sciencetostartup.com/paper/causal-decoding-for-hallucination-resistant-multimodal-large-language-models - Develop a causal decoding framework to make multimodal language models hallucination-resistant, enhancing faithfulness in responses.
- Provably Safe Generative Sampling with Constricting Barrier Functions (viability: 3): https://sciencetostartup.com/paper/provably-safe-generative-sampling-with-constricting-barrier-functions - Develop a safety filtering framework for generative models ensuring constraint satisfaction in safety-critical domains.
- On the Structural Non-Preservation of Epistemic Behaviour under Policy Transformation (viability: 3): https://sciencetostartup.com/paper/on-the-structural-non-preservation-of-epistemic-behaviour-under-policy-transformation - A study on how epistemic behavior in RL agents is structurally affected by policy transformations.
- ECHOSAT: Estimating Canopy Height Over Space And Time (viability: 9): https://sciencetostartup.com/paper/echosat-estimating-canopy-height-over-space-and-time - ECHOSAT provides a dynamic global tree height map for enhanced forest monitoring and carbon accounting.
- Overconfident Errors Need Stronger Correction: Asymmetric Confidence Penalties for Reinforcement Learning (viability: 5): https://sciencetostartup.com/paper/overconfident-errors-need-stronger-correction-asymmetric-confidence-penalties-for-reinforcement-learning - Develop ACE, an adaptive error penalty system to correct overconfidence in RL-enhanced LLM reasoning.
- The Headless Firm: How AI Reshapes Enterprise Boundaries (viability: 1): https://sciencetostartup.com/paper/the-headless-firm-how-ai-reshapes-enterprise-boundaries - Analyzes how agentic AI shifts coordination costs in firms, leading to a "Headless Firm" structure.
- FedVG: Gradient-Guided Aggregation for Enhanced Federated Learning (viability: 5): https://sciencetostartup.com/paper/fedvg-gradient-guided-aggregation-for-enhanced-federated-learning - FedVG enhances federated learning by using gradient-guided aggregation for better generalization on global validations sets.
- MrBERT: Modern Multilingual Encoders via Vocabulary, Domain, and Dimensional Adaptation (viability: 7): https://sciencetostartup.com/paper/mrbert-modern-multilingual-encoders-via-vocabulary-domain-and-dimensional-adaptation - Optimize multilingual language processing with MrBERT encoders for localized and domain-specific applications.
- Small Language Models for Privacy-Preserving Clinical Information Extraction in Low-Resource Languages (viability: 5): https://sciencetostartup.com/paper/small-language-models-for-privacy-preserving-clinical-information-extraction-in-low-resource-languages - Develop privacy-preserving tools for extracting clinical information from medical transcripts in low-resource languages using small language models.
- The Mean is the Mirage: Entropy-Adaptive Model Merging under Heterogeneous Domain Shifts in Medical Imaging (viability: 6): https://sciencetostartup.com/paper/the-mean-is-the-mirage-entropy-adaptive-model-merging-under-heterogeneous-domain-shifts-in-medical-imaging - Entropy-adaptive model merging for medical imaging enables on-the-fly adaptation to domain shifts without labels or retraining.
- Black-Box Reliability Certification for AI Agents via Self-Consistency Sampling and Conformal Calibration (viability: 6): https://sciencetostartup.com/paper/black-box-reliability-certification-for-ai-agents-via-self-consistency-sampling-and-conformal-calibration - Develop a certification tool for AI reliability based on self-consistency sampling and conformal calibration for different model-task pairs.
- Towards Controllable Video Synthesis of Routine and Rare OR Events (viability: 7): https://sciencetostartup.com/paper/towards-controllable-video-synthesis-of-routine-and-rare-or-events - Controlled OR event video synthesis platform for AI model training and validation.
- Representation Theorems for Cumulative Propositional Dependence Logics (viability: 3): https://sciencetostartup.com/paper/representation-theorems-for-cumulative-propositional-dependence-logics - Develop new representation theorems for cumulative propositional dependence logics with potential applications in logical theory formulation.
- A Hierarchical Multi-Agent System for Autonomous Discovery in Geoscientific Data Archives (viability: 6): https://sciencetostartup.com/paper/a-hierarchical-multi-agent-system-for-autonomous-discovery-in-geoscientific-data-archives - PANGAEA-GPT automates geoscientific data discovery and analysis using a hierarchical multi-agent system.
- Alignment-Weighted DPO: A principled reasoning approach to improve safety alignment (viability: 6): https://sciencetostartup.com/paper/alignment-weighted-dpo-a-principled-reasoning-approach-to-improve-safety-alignment - Develop a reasoning-enhanced safety alignment method for LLMs to better handle deceptive jailbreak attacks.
- Scaling View Synthesis Transformers (viability: 4): https://sciencetostartup.com/paper/scaling-view-synthesis-transformers - Scalable View Synthesis Model enhances view synthesis with superior performance-compute efficiency.
- Equitable Evaluation via Elicitation (viability: 2): https://sciencetostartup.com/paper/equitable-evaluation-via-elicitation - Develop an AI tool that evaluates skills equitably regardless of self-presentation style to enhance job-seeker evaluations and company reorganizations.
- Test-Time Training with KV Binding Is Secretly Linear Attention (viability: 2): https://sciencetostartup.com/paper/test-time-training-with-kv-binding-is-secretly-linear-attention - Reframes Test-Time Training as learned linear attention for more efficient model architectures.
- Aletheia tackles FirstProof autonomously (viability: 7): https://sciencetostartup.com/paper/aletheia-tackles-firstproof-autonomously - Aletheia autonomously solves complex math problems using Gemini 3 Deep Think without human intervention.
- Learning from Trials and Errors: Reflective Test-Time Planning for Embodied LLMs (viability: 8): https://sciencetostartup.com/paper/learning-from-trials-and-errors-reflective-test-time-planning-for-embodied-llms - Reflective Test-Time Planning transforms embodied AI with self-improvement capabilities through real-time action reflection.
- Why Pass@k Optimization Can Degrade Pass@1: Prompt Interference in LLM Post-training (viability: 2): https://sciencetostartup.com/paper/why-pass-k-optimization-can-degrade-pass-1-prompt-interference-in-llm-post-training - Explore the trade-off between optimizing pass@k and pass@1 in language models due to gradient conflicts.
- XMorph: Explainable Brain Tumor Analysis Via LLM-Assisted Hybrid Deep Intelligence (viability: 8): https://sciencetostartup.com/paper/xmorph-explainable-brain-tumor-analysis-via-llm-assisted-hybrid-deep-intelligence - XMorph provides an explainable AI solution for brain tumor diagnosis with superior accuracy and clinical acceptance.
- Efficient Hierarchical Any-Angle Path Planning on Multi-Resolution 3D Grids (viability: 7): https://sciencetostartup.com/paper/efficient-hierarchical-any-angle-path-planning-on-multi-resolution-3d-grids - Efficient 3D path planning for autonomous robots with any-angle navigation and multi-resolution grids.
- NoRD: A Data-Efficient Vision-Language-Action Model that Drives without Reasoning (viability: 5): https://sciencetostartup.com/paper/nord-a-data-efficient-vision-language-action-model-that-drives-without-reasoning - Develop a data-efficient VLA model for autonomous driving that minimizes reasoning and data requirements.
- PVminer: A Domain-Specific Tool to Detect the Patient Voice in Patient Generated Data (viability: 7): https://sciencetostartup.com/paper/pvminer-a-domain-specific-tool-to-detect-the-patient-voice-in-patient-generated-data - PVminer efficiently identifies and analyzes patient voices in healthcare data for enhanced patient-centered care.
- CG-DMER: Hybrid Contrastive-Generative Framework for Disentangled Multimodal ECG Representation Learning (viability: 8): https://sciencetostartup.com/paper/cg-dmer-hybrid-contrastive-generative-framework-for-disentangled-multimodal-ecg-representation-learning - Develop a cutting-edge ECG analysis tool using a novel contrastive-generative framework to improve cardiovascular diagnostics.
- A Benchmark for Deep Information Synthesis (viability: 5): https://sciencetostartup.com/paper/a-benchmark-for-deep-information-synthesis - Develop DEEPSYNTH, a benchmark to evaluate AI agents' ability to synthesize information across multiple domains and data sources.
- SparkMe: Adaptive Semi-Structured Interviewing for Qualitative Insight Discovery (viability: 8): https://sciencetostartup.com/paper/sparkme-adaptive-semi-structured-interviewing-for-qualitative-insight-discovery - Unlock qualitative insights at scale with AI-powered adaptive interviews.
- "Are You Sure?": An Empirical Study of Human Perception Vulnerability in LLM-Driven Agentic Systems (viability: 3): https://sciencetostartup.com/paper/are-you-sure-an-empirical-study-of-human-perception-vulnerability-in-llm-driven-agentic-systems - A study on the vulnerabilities of human perception to deception by LLM-driven agents.
- Cooperative-Competitive Team Play of Real-World Craft Robots (viability: 6): https://sciencetostartup.com/paper/cooperative-competitive-team-play-of-real-world-craft-robots - Develop reinforcement learning to enable real-world robots to efficiently learn cooperative and competitive strategies.
- Probing Graph Neural Network Activation Patterns Through Graph Topology (viability: 4): https://sciencetostartup.com/paper/probing-graph-neural-network-activation-patterns-through-graph-topology - Utilize graph topology curvature as a diagnostic probe to enhance graph neural network learning insights.
- Localized Dynamics-Aware Domain Adaption for Off-Dynamics Offline Reinforcement Learning (viability: 5): https://sciencetostartup.com/paper/localized-dynamics-aware-domain-adaption-for-off-dynamics-offline-reinforcement-learning - Leverage localized dynamics mismatch to enhance off-dynamics offline RL policy development.
- The Initial Exploration Problem in Knowledge Graph Exploration (viability: 3): https://sciencetostartup.com/paper/the-initial-exploration-problem-in-knowledge-graph-exploration - The Initial Exploration Problem theory provides a framework to enhance user-friendly Knowledge Graph interfaces.
- Tool Building as a Path to "Superintelligence" (viability: 4): https://sciencetostartup.com/paper/tool-building-as-a-path-to-superintelligence - Leverage the Diligent Learner framework to enhance LLMs' reasoning capabilities through precise tool design.
- VAUQ: Vision-Aware Uncertainty Quantification for LVLM Self-Evaluation (viability: 5): https://sciencetostartup.com/paper/vauq-vision-aware-uncertainty-quantification-for-lvlm-self-evaluation - VAUQ enhances vision-language models by providing a training-free, vision-aware uncertainty quantification framework for more reliable self-evaluation.
- Position-Aware Sequential Attention for Accurate Next Item Recommendations (viability: 5): https://sciencetostartup.com/paper/position-aware-sequential-attention-for-accurate-next-item-recommendations - Enhance next-item recommendation systems with position-aware sequential attention technology for improved prediction accuracy.
- HELP: HyperNode Expansion and Logical Path-Guided Evidence Localization for Accurate and Efficient GraphRAG (viability: 6): https://sciencetostartup.com/paper/help-hypernode-expansion-and-logical-path-guided-evidence-localization-for-accurate-and-efficient-graphrag - Optimize multi-hop reasoning in GraphRAG with HyperNode Expansion for speed and accuracy.
- Predicting Sentence Acceptability Judgments in Multimodal Contexts (viability: 3): https://sciencetostartup.com/paper/predicting-sentence-acceptability-judgments-in-multimodal-contexts - Explore how visual contexts affect sentence acceptability judgments by humans and LLMs.
- E-MMKGR: A Unified Multimodal Knowledge Graph Framework for E-commerce Applications (viability: 8): https://sciencetostartup.com/paper/e-mmkgr-a-unified-multimodal-knowledge-graph-framework-for-e-commerce-applications - A new framework for e-commerce applications that unifies item representations using multimodal knowledge graphs to improve recommendation and search performance.
- Regret-Guided Search Control for Efficient Learning in AlphaZero (viability: 7): https://sciencetostartup.com/paper/regret-guided-search-control-for-efficient-learning-in-alphazero - Enhance AlphaZero's learning efficiency with Regret-Guided Search Control for improved gaming AI performance.
- Pipeline for Verifying LLM-Generated Mathematical Solutions (viability: 5): https://sciencetostartup.com/paper/pipeline-for-verifying-llm-generated-mathematical-solutions - Automate verification of LLM-generated math solutions with open-source tools for accuracy and reliability.
- OrthoDiffusion: A Generalizable Multi-Task Diffusion Foundation Model for Musculoskeletal MRI Interpretation (viability: 5): https://sciencetostartup.com/paper/orthodiffusion-a-generalizable-multi-task-diffusion-foundation-model-for-musculoskeletal-mri-interpretation - Develop a diffusion-based foundation model to enhance musculoskeletal MRI interpretation with multi-task diagnostics and segmentation.
- SibylSense: Adaptive Rubric Learning via Memory Tuning and Adversarial Probing (viability: 2): https://sciencetostartup.com/paper/sibylsense-adaptive-rubric-learning-via-memory-tuning-and-adversarial-probing - SibylSense offers an adaptive rubric learning method enhancing reward robustness and interpretability in reinforcement learning.
- PyVision-RL: Forging Open Agentic Vision Models via RL (viability: 7): https://sciencetostartup.com/paper/pyvision-rl-forging-open-agentic-vision-models-via-rl - Develop scalable multimodal agents with improved tool usage and video reasoning via PyVision-RL.
- CHESS: Context-aware Hierarchical Efficient Semantic Selection for Long-Context LLM Inference (viability: 8): https://sciencetostartup.com/paper/chess-context-aware-hierarchical-efficient-semantic-selection-for-long-context-llm-inference - CHESS optimizes long-context LLM inference by drastically reducing KV cache demands, improving throughput by over 4x with minimal memory.
- Buffer Matters: Unleashing the Power of Off-Policy Reinforcement Learning in Large Language Model Reasoning (viability: 6): https://sciencetostartup.com/paper/buffer-matters-unleashing-the-power-of-off-policy-reinforcement-learning-in-large-language-model-reasoning - Develop an off-policy RL framework to significantly enhance data efficiency in large language model post-training.
- AdapTools: Adaptive Tool-based Indirect Prompt Injection Attacks on Agentic LLMs (viability: 3): https://sciencetostartup.com/paper/adaptools-adaptive-tool-based-indirect-prompt-injection-attacks-on-agentic-llms - AdapTools offers an adaptive framework for executing stealthier indirect prompt injection attacks on LLM-based agents.
- ICON: Indirect Prompt Injection Defense for Agents based on Inference-Time Correction (viability: 7): https://sciencetostartup.com/paper/icon-indirect-prompt-injection-defense-for-agents-based-on-inference-time-correction - ICON offers a framework for defending LLM agents from indirect prompt injection attacks by using inference-time corrections to ensure security without compromising task continuity.
- PromptCD: Test-Time Behavior Enhancement via Polarity-Prompt Contrastive Decoding (viability: 6): https://sciencetostartup.com/paper/promptcd-test-time-behavior-enhancement-via-polarity-prompt-contrastive-decoding - PromptCD enhances LLM and VLM behaviors at test-time, offering a cost-efficient solution for reliable AI alignment.
- How Foundational Skills Influence VLM-based Embodied Agents:A Native Perspective (viability: 5): https://sciencetostartup.com/paper/how-foundational-skills-influence-vlm-based-embodied-agents-a-native-perspective - NativeEmbodied provides a detailed evaluation benchmark for enhancing performance in VLM-driven embodied agents.
- Recursive Belief Vision Language Model (viability: 7): https://sciencetostartup.com/paper/recursive-belief-vision-language-model - RB-VLA provides belief-centric architecture for better performance in long-horizon vision-language-action tasks with reduced inference latency.
- SurgAtt-Tracker: Online Surgical Attention Tracking via Temporal Proposal Reranking and Motion-Aware Refinement (viability: 5): https://sciencetostartup.com/paper/surgatt-tracker-online-surgical-attention-tracking-via-temporal-proposal-reranking-and-motion-aware-refinement - Develop a surgical attention tracking tool for minimally invasive procedures using spatio-temporal learning to enhance robotic surgeries.
- Enhancing Hate Speech Detection on Social Media: A Comparative Analysis of Machine Learning Models and Text Transformation Approaches (viability: 3): https://sciencetostartup.com/paper/enhancing-hate-speech-detection-on-social-media-a-comparative-analysis-of-machine-learning-models-and-text-transformatio - Comparative analysis of machine learning models and text transformation techniques for enhanced social media hate speech detection.
- When can we trust untrusted monitoring? A safety case sketch across collusion strategies (viability: 2): https://sciencetostartup.com/paper/when-can-we-trust-untrusted-monitoring-a-safety-case-sketch-across-collusion-strategies - Develops a taxonomy to evaluate the safety of untrusted AI monitoring systems against potential collusion strategies.
- What Drives Students' Use of AI Chatbots? Technology Acceptance in Conversational AI (viability: 2): https://sciencetostartup.com/paper/what-drives-students-use-of-ai-chatbots-technology-acceptance-in-conversational-ai - Understanding student adoption of AI chatbots using TAM and structural equation modeling.
- Maximin Share Guarantees via Limited Cost-Sensitive Sharing (viability: 3): https://sciencetostartup.com/paper/maximin-share-guarantees-via-limited-cost-sensitive-sharing - A theoretical study on fair allocation of indivisible goods among agents using limited cost-sensitive sharing.
- Actor-Curator: Co-adaptive Curriculum Learning via Policy-Improvement Bandits for RL Post-Training (viability: 6): https://sciencetostartup.com/paper/actor-curator-co-adaptive-curriculum-learning-via-policy-improvement-bandits-for-rl-post-training - ACTOR-CURATOR is an automated curriculum learning framework enhancing reinforcement learning post-training efficiency and stability for large language models.
- A Generalized Apprenticeship Learning Framework for Capturing Evolving Student Pedagogical Strategies (viability: 7): https://sciencetostartup.com/paper/a-generalized-apprenticeship-learning-framework-for-capturing-evolving-student-pedagogical-strategies - Leverage the THEMES framework to improve intelligent tutoring systems by capturing evolving student pedagogical strategies through Apprenticeship Learning.
- KairosVL: Orchestrating Time Series and Semantics for Unified Reasoning (viability: 5): https://sciencetostartup.com/paper/kairosvl-orchestrating-time-series-and-semantics-for-unified-reasoning - KairosVL enhances time series analysis by integrating semantic reasoning to improve decision-making capabilities.
- Wireless Federated Multi-Task LLM Fine-Tuning via Sparse-and-Orthogonal LoRA (viability: 6): https://sciencetostartup.com/paper/wireless-federated-multi-task-llm-fine-tuning-via-sparse-and-orthogonal-lora - Optimize decentralized federated learning with sparse-and-orthogonal LoRA for efficient mobile device collaboration.
- Elimination-compensation pruning for fully-connected neural networks (viability: 5): https://sciencetostartup.com/paper/elimination-compensation-pruning-for-fully-connected-neural-networks - Introduce a novel pruning method for neural networks that compensates for weight removal by adjusting adjacent biases, enhancing model efficiency.
- Imputation of Unknown Missingness in Sparse Electronic Health Records (viability: 5): https://sciencetostartup.com/paper/imputation-of-unknown-missingness-in-sparse-electronic-health-records - Transformer-based algorithm improves EHR data imputation, enhancing prediction accuracy for healthcare applications.
- Learning to Rewrite Tool Descriptions for Reliable LLM-Agent Tool Use (viability: 7): https://sciencetostartup.com/paper/learning-to-rewrite-tool-descriptions-for-reliable-llm-agent-tool-use - Enhance LLM-based agents by optimizing tool interface descriptions for better tool selection in large candidate sets.
- Implicit Intelligence -- Evaluating Agents on What Users Don't Say (viability: 6): https://sciencetostartup.com/paper/implicit-intelligence-evaluating-agents-on-what-users-don-t-say - Develop an AI agent evaluation framework focusing on implicit reasoning in user interactions.
- Three Concrete Challenges and Two Hopes for the Safety of Unsupervised Elicitation (viability: 2): https://sciencetostartup.com/paper/three-concrete-challenges-and-two-hopes-for-the-safety-of-unsupervised-elicitation - Develop robust unsupervised elicitation techniques to improve the reliability of language models on challenging datasets.
- Hierarchical Molecular Representation Learning via Fragment-Based Self-Supervised Embedding Prediction (viability: 4): https://sciencetostartup.com/paper/hierarchical-molecular-representation-learning-via-fragment-based-self-supervised-embedding-prediction - GraSPNet leverages graph self-supervised learning to provide chemically meaningful molecular representations that outperform existing methods.
- No One Size Fits All: QueryBandits for Hallucination Mitigation (viability: 7): https://sciencetostartup.com/paper/no-one-size-fits-all-querybandits-for-hallucination-mitigation - QueryBandits offers a novel framework for mitigating LLM hallucinations using adaptive query-rewrite strategies, enhancing performance without retraining models.
- Circuit Tracing in Vision-Language Models: Understanding the Internal Mechanisms of Multimodal Thinking (viability: 2): https://sciencetostartup.com/paper/circuit-tracing-in-vision-language-models-understanding-the-internal-mechanisms-of-multimodal-thinking - Develop a framework for transparent circuit tracing in vision-language models to enhance understanding of multimodal reasoning.
- Learning Physical Principles from Interaction: Self-Evolving Planning via Test-Time Memory (viability: 5): https://sciencetostartup.com/paper/learning-physical-principles-from-interaction-self-evolving-planning-via-test-time-memory - PhysMem enhances VLM robot planners with memory to learn and adapt physical principles through interaction, improving decision-making in real-time tasks.
- Shape-informed cardiac mechanics surrogates in data-scarce regimes via geometric encoding and generative augmentation (viability: 6): https://sciencetostartup.com/paper/shape-informed-cardiac-mechanics-surrogates-in-data-scarce-regimes-via-geometric-encoding-and-generative-augmentation - Develop a surrogate model framework for cardiac simulation that enhances generalization with geometric encoding for clinical application.
- What Makes a Good Query? Measuring the Impact of Human-Confusing Linguistic Features on LLM Performance (viability: 6): https://sciencetostartup.com/paper/what-makes-a-good-query-measuring-the-impact-of-human-confusing-linguistic-features-on-llm-performance - Enhance LLM performance by identifying and rewriting query features that increase hallucination risks.
- A Very Big Video Reasoning Suite (viability: 7): https://sciencetostartup.com/paper/a-very-big-video-reasoning-suite - Develop a comprehensive video reasoning tool using the VBVR suite, capable of understanding and analyzing complex video environments.
- Behavior Learning (BL): Learning Hierarchical Optimization Structures from Data (viability: 5): https://sciencetostartup.com/paper/behavior-learning-bl-learning-hierarchical-optimization-structures-from-data - Develop Behavior Learning framework to enable interpretable hierarchical optimization in machine learning models integrated with behavioral science.
- Recurrent Structural Policy Gradient for Partially Observable Mean Field Games (viability: 8): https://sciencetostartup.com/paper/recurrent-structural-policy-gradient-for-partially-observable-mean-field-games - Develop advanced algorithms for optimizing large-scale multi-agent systems under uncertainty using Recurrent Structural Policy Gradient.
- KNIGHT: Knowledge Graph-Driven Multiple-Choice Question Generation with Adaptive Hardness Calibration (viability: 6): https://sciencetostartup.com/paper/knight-knowledge-graph-driven-multiple-choice-question-generation-with-adaptive-hardness-calibration - Develop adaptive difficulty-calibrated MCQ datasets using a knowledge-graph framework to streamline LLM evaluation.
- Modeling Epidemiological Dynamics Under Adversarial Data and User Deception (viability: 3): https://sciencetostartup.com/paper/modeling-epidemiological-dynamics-under-adversarial-data-and-user-deception - Develop a game-theoretic framework to model and mitigate deception in self-reported epidemiological data for robust public health responses.
- AdaEvolve: Adaptive LLM Driven Zeroth-Order Optimization (viability: 7): https://sciencetostartup.com/paper/adaevolve-adaptive-llm-driven-zeroth-order-optimization - AdaEvolve optimizes LLM-driven program generation with adaptive resource allocation for more efficient evolutionary computation.
- To Reason or Not to: Selective Chain-of-Thought in Medical Question Answering (viability: 7): https://sciencetostartup.com/paper/to-reason-or-not-to-selective-chain-of-thought-in-medical-question-answering - Selective CoT enhances real-world deployability of medical QA systems by optimizing reasoning processes in LLMs, reducing costs while maintaining accuracy.
- NanoKnow: How to Know What Your Language Model Knows (viability: 6): https://sciencetostartup.com/paper/nanoknow-how-to-know-what-your-language-model-knows - NanoKnow provides a benchmark dataset to uncover and analyze the knowledge encoding in large language models.
- NovaPlan: Zero-Shot Long-Horizon Manipulation via Closed-Loop Video Language Planning (viability: 8): https://sciencetostartup.com/paper/novaplan-zero-shot-long-horizon-manipulation-via-closed-loop-video-language-planning - NovaPlan enables robots to perform zero-shot, long-horizon manipulations using video language planning, achieving state-of-the-art results without prior demonstrations.
- ReSyn: Autonomously Scaling Synthetic Environments for Reasoning Models (viability: 5): https://sciencetostartup.com/paper/resyn-autonomously-scaling-synthetic-environments-for-reasoning-models - ReSyn autonomously scales and diversifies synthetic environments for training reasoning models using verifier-based reinforcement learning.
- Benchmarking Unlearning for Vision Transformers (viability: 5): https://sciencetostartup.com/paper/benchmarking-unlearning-for-vision-transformers - Benchmarking unlearning algorithms on Vision Transformers offers a foundation for safe, fair AI development.
- StyleStream: Real-Time Zero-Shot Voice Style Conversion (viability: 6): https://sciencetostartup.com/paper/stylestream-real-time-zero-shot-voice-style-conversion - StyleStream enables real-time zero-shot voice style conversion across timbre, accent, and emotion.
- Align When They Want, Complement When They Need! Human-Centered Ensembles for Adaptive Human-AI Collaboration (viability: 3): https://sciencetostartup.com/paper/align-when-they-want-complement-when-they-need-human-centered-ensembles-for-adaptive-human-ai-collaboration - Develop adaptive AI ensembles to optimize human-AI collaboration by balancing trust and performance.
- BarrierSteer: LLM Safety via Learning Barrier Steering (viability: 5): https://sciencetostartup.com/paper/barriersteer-llm-safety-via-learning-barrier-steering - BarrierSteer enhances LLM safety by integrating control barrier functions to prevent unsafe outputs without altering model performance.
- Transcending the Annotation Bottleneck: AI-Powered Discovery in Biology and Medicine (viability: 5): https://sciencetostartup.com/paper/transcending-the-annotation-bottleneck-ai-powered-discovery-in-biology-and-medicine - AI-powered unsupervised learning for discovering novel phenotypes and pathologies in biomedicine.
- CausalFlip: A Benchmark for LLM Causal Judgment Beyond Semantic Matching (viability: 5): https://sciencetostartup.com/paper/causalflip-a-benchmark-for-llm-causal-judgment-beyond-semantic-matching - CausalFlip enables large language models to improve causal reasoning beyond semantic pattern recognition through a novel benchmark.
- StructXLIP: Enhancing Vision-language Models with Multimodal Structural Cues (viability: 6): https://sciencetostartup.com/paper/structxlip-enhancing-vision-language-models-with-multimodal-structural-cues - StructXLIP enhances vision-language model fine-tuning by integrating multimodal structural cues for better alignment.
- Descent-Guided Policy Gradient for Scalable Cooperative Multi-Agent Learning (viability: 3): https://sciencetostartup.com/paper/descent-guided-policy-gradient-for-scalable-cooperative-multi-agent-learning - Innovative framework for reducing gradient variance in cooperative multi-agent reinforcement learning.
- HeatPrompt: Zero-Shot Vision-Language Modeling of Urban Heat Demand from Satellite Images (viability: 3): https://sciencetostartup.com/paper/heatprompt-zero-shot-vision-language-modeling-of-urban-heat-demand-from-satellite-images - Leverage satellite images and GIS data to estimate urban heat demand with zero-shot vision-language models.
- Multilingual Large Language Models do not comprehend all natural languages to equal degrees (viability: 5): https://sciencetostartup.com/paper/multilingual-large-language-models-do-not-comprehend-all-natural-languages-to-equal-degrees - Develop an analysis tool to evaluate the language comprehension performance of LLMs across diverse languages, especially non-English ones.
- On the Equivalence of Random Network Distillation, Deep Ensembles, and Bayesian Inference (viability: 2): https://sciencetostartup.com/paper/on-the-equivalence-of-random-network-distillation-deep-ensembles-and-bayesian-inference - Unify RND with deep ensembles and Bayesian inference for uncertainty quantification.
- DP-FedAdamW: An Efficient Optimizer for Differentially Private Federated Large Models (viability: 5): https://sciencetostartup.com/paper/dp-fedadamw-an-efficient-optimizer-for-differentially-private-federated-large-models - Optimizer enhancing privacy and performance in federated learning for language and vision models.
- Rethinking LoRA for Privacy-Preserving Federated Learning in Large Models (viability: 6): https://sciencetostartup.com/paper/rethinking-lora-for-privacy-preserving-federated-learning-in-large-models - LA-LoRA enhances privacy-preserving federated learning by improving accuracy in differential privacy settings with state-of-the-art results.
- Watson & Holmes: A Naturalistic Benchmark for Comparing Human and LLM Reasoning (viability: 5): https://sciencetostartup.com/paper/watson-holmes-a-naturalistic-benchmark-for-comparing-human-and-llm-reasoning - A benchmark for evaluating LLM reasoning in naturalistic contexts developed from a detective tabletop game.
- GOAL: Geometrically Optimal Alignment for Continual Generalized Category Discovery (viability: 5): https://sciencetostartup.com/paper/goal-geometrically-optimal-alignment-for-continual-generalized-category-discovery - GOAL provides a consistent geometric structure for continual category discovery, significantly reducing forgetting and improving class discovery.
- LLM-enabled Applications Require System-Level Threat Monitoring (viability: 3): https://sciencetostartup.com/paper/llm-enabled-applications-require-system-level-threat-monitoring - Develop a system-level threat monitoring solution for LLM-enabled applications to ensure reliable operation and incident response.
- SafePickle: Robust and Generic ML Detection of Malicious Pickle-based ML Models (viability: 8): https://sciencetostartup.com/paper/safepickle-robust-and-generic-ml-detection-of-malicious-pickle-based-ml-models - SafePickle offers a machine learning-based solution to detect malicious Pickle files in model repositories, enhancing security in AI model sharing.
- Decision MetaMamba: Enhancing Selective SSM in Offline RL with Heterogeneous Sequence Mixing (viability: 5): https://sciencetostartup.com/paper/decision-metamamba-enhancing-selective-ssm-in-offline-rl-with-heterogeneous-sequence-mixing - Decision MetaMamba enhances offline RL models with improved sequence mixing for better performance and efficiency.
- Hexagon-MLIR: An AI Compilation Stack For Qualcomm's Neural Processing Units (NPUs) (viability: 3): https://sciencetostartup.com/paper/hexagon-mlir-an-ai-compilation-stack-for-qualcomm-s-neural-processing-units-npus - Open-source Hexagon-MLIR enhances Qualcomm NPU AI compilation capabilities for faster kernel deployment.
- Carbon-Aware Governance Gates: An Architecture for Sustainable GenAI Development (viability: 3): https://sciencetostartup.com/paper/carbon-aware-governance-gates-an-architecture-for-sustainable-genai-development - Develop an architecture to integrate carbon-aware governance into AI development for sustainability.
- Iconographic Classification and Content-Based Recommendation for Digitized Artworks (viability: 7): https://sciencetostartup.com/paper/iconographic-classification-and-content-based-recommendation-for-digitized-artworks - Automated iconographic classification and recommendation system enhances digitized artwork cataloging and navigation.
- SkillOrchestra: Learning to Route Agents via Skill Transfer (viability: 7): https://sciencetostartup.com/paper/skillorchestra-learning-to-route-agents-via-skill-transfer - SkillOrchestra optimizes multi-agent orchestration with skill-aware routing for efficient AI system deployment.
- Representation Stability in a Minimal Continual Learning Agent (viability: 4): https://sciencetostartup.com/paper/representation-stability-in-a-minimal-continual-learning-agent - Explore representational dynamics in minimal continual learning systems for improved learning stability.
- Denoising Particle Filters: Learning State Estimation with Single-Step Objectives (viability: 5): https://sciencetostartup.com/paper/denoising-particle-filters-learning-state-estimation-with-single-step-objectives - Develop a customizable particle filtering algorithm for improved robotic state estimation using learned denoising methods.
- TAPE: Tool-Guided Adaptive Planning and Constrained Execution in Language Model Agents (viability: 8): https://sciencetostartup.com/paper/tape-tool-guided-adaptive-planning-and-constrained-execution-in-language-model-agents - TAPE: a robust AI framework improving agent success rates by 21% in environments with strict constraints.
- Localized Concept Erasure in Text-to-Image Diffusion Models via High-Level Representation Misdirection (viability: 5): https://sciencetostartup.com/paper/localized-concept-erasure-in-text-to-image-diffusion-models-via-high-level-representation-misdirection - HiRM allows for precise concept removal in text-to-image diffusion models with minimal impact on unrelated concepts, enhancing content moderation.
- PedaCo-Gen: Scaffolding Pedagogical Agency in Human-AI Collaborative Video Authoring (viability: 7): https://sciencetostartup.com/paper/pedaco-gen-scaffolding-pedagogical-agency-in-human-ai-collaborative-video-authoring - PedaCo-Gen is an AI-driven tool that enhances the creation of instructional videos by integrating pedagogical principles in a human-AI collaborative framework.
- VecFormer: Towards Efficient and Generalizable Graph Transformer with Graph Token Attention (viability: 6): https://sciencetostartup.com/paper/vecformer-towards-efficient-and-generalizable-graph-transformer-with-graph-token-attention - Develop an efficient graph transformer model, VecFormer, for scalable and generalizable node classification.
- Rules or Weights? Comparing User Understanding of Explainable AI Techniques with the Cognitive XAI-Adaptive Model (viability: 4): https://sciencetostartup.com/paper/rules-or-weights-comparing-user-understanding-of-explainable-ai-techniques-with-the-cognitive-xai-adaptive-model - Enhance AI debugging and benchmarking through a cognitive framework that adapts XAI explanations to user understanding.
- Satellite-Based Detection of Looted Archaeological Sites Using Machine Learning (viability: 9): https://sciencetostartup.com/paper/satellite-based-detection-of-looted-archaeological-sites-using-machine-learning - AI-powered tool to automatically detect looted archaeological sites from satellite imagery, protecting cultural heritage.
- Detecting High-Potential SMEs with Heterogeneous Graph Neural Networks (viability: 7): https://sciencetostartup.com/paper/detecting-high-potential-smes-with-heterogeneous-graph-neural-networks - Identify high-potential SMEs using a Heterogeneous Graph Transformer-based prediction tool leveraging public data.
- Tri-Subspaces Disentanglement for Multimodal Sentiment Analysis (viability: 5): https://sciencetostartup.com/paper/tri-subspaces-disentanglement-for-multimodal-sentiment-analysis - Tri-Subspace Disentanglement framework improves multimodal sentiment analysis by enhancing cross-modal synergies and modality-specific cues.
- Grokking Finite-Dimensional Algebra (viability: 2): https://sciencetostartup.com/paper/grokking-finite-dimensional-algebra - Investigate the grokking phenomenon in finite-dimensional algebras to better understand neural network generalization.
- Ada-RS: Adaptive Rejection Sampling for Selective Thinking (viability: 7): https://sciencetostartup.com/paper/ada-rs-adaptive-rejection-sampling-for-selective-thinking - Efficiently optimize reasoning in AI models for latency-sensitive applications via Ada-RS.
- Softmax is not Enough (for Adaptive Conformal Classification) (viability: 5): https://sciencetostartup.com/paper/softmax-is-not-enough-for-adaptive-conformal-classification - Enhancing conformal classification adaptiveness using Helmholtz Free Energy for better uncertainty quantification.
- Botson: An Accessible and Low-Cost Platform for Social Robotics Research (viability: 2): https://sciencetostartup.com/paper/botson-an-accessible-and-low-cost-platform-for-social-robotics-research - Develop an affordable social robotics platform using LLMs for enhanced trust in AI-human interaction.
- Federated Learning Playground (viability: 6): https://sciencetostartup.com/paper/federated-learning-playground - An interactive browser-based platform to teach and prototype Federated Learning concepts with real-time visualizations.
- When AI Teammates Meet Code Review: Collaboration Signals Shaping the Integration of Agent-Authored Pull Requests (viability: 5): https://sciencetostartup.com/paper/when-ai-teammates-meet-code-review-collaboration-signals-shaping-the-integration-of-agent-authored-pull-requests - Enhance code review workflows by integrating signals from AI-authored pull requests to improve collaboration and merging success.
- OptiRepair: Closed-Loop Diagnosis and Repair of Supply Chain Optimization Models with LLM Agents (viability: 3): https://sciencetostartup.com/paper/optirepair-closed-loop-diagnosis-and-repair-of-supply-chain-optimization-models-with-llm-agents - AI-driven tool for diagnosing and repairing infeasible supply chain optimization models using large language models.
- FinSight-Net:A Physics-Aware Decoupled Network with Frequency-Domain Compensation for Underwater Fish Detection in Smart Aquaculture (viability: 7): https://sciencetostartup.com/paper/finsight-net-a-physics-aware-decoupled-network-with-frequency-domain-compensation-for-underwater-fish-detection-in-smart - FinSight-Net provides an efficient network to improve fish detection accuracy in underwater environments for smart aquaculture.
- Hiding in Plain Text: Detecting Concealed Jailbreaks via Activation Disentanglement (viability: 2): https://sciencetostartup.com/paper/hiding-in-plain-text-detecting-concealed-jailbreaks-via-activation-disentanglement - Detect concealed jailbreaks in large language models through semantic disentanglement for enhanced LLM safety.
- Artificial Intelligence for Modeling & Simulation in Digital Twins (viability: 2): https://sciencetostartup.com/paper/artificial-intelligence-for-modeling-simulation-in-digital-twins - Explore the integration of AI and modeling & simulation within digital twins for enhanced digital transformations.
- Stable Deep Reinforcement Learning via Isotropic Gaussian Representations (viability: 5): https://sciencetostartup.com/paper/stable-deep-reinforcement-learning-via-isotropic-gaussian-representations - Enhance reinforcement learning stability using isotropic Gaussian embeddings to combat training non-stationarity.
- Time Series, Vision, and Language: Exploring the Limits of Alignment in Contrastive Representation Spaces (viability: 2): https://sciencetostartup.com/paper/time-series-vision-and-language-exploring-the-limits-of-alignment-in-contrastive-representation-spaces - Exploring alignment in contrastive representation spaces across modalities including time series, vision, and language.
- Active perception and disentangled representations allow continual, episodic zero and few-shot learning (viability: 5): https://sciencetostartup.com/paper/active-perception-and-disentangled-representations-allow-continual-episodic-zero-and-few-shot-learning - A Complementary Learning System for fast, context-driven reasoning enabling zero-shot and few-shot learning without needing generalization.
- Anatomy of Agentic Memory: Taxonomy and Empirical Analysis of Evaluation and System Limitations (viability: 5): https://sciencetostartup.com/paper/anatomy-of-agentic-memory-taxonomy-and-empirical-analysis-of-evaluation-and-system-limitations - Develop scalable and reliable agentic memory systems for LLMs to enhance long-horizon reasoning and personalization.
- Health+: Empowering Individuals via Unifying Health Data (viability: 4): https://sciencetostartup.com/paper/health-empowering-individuals-via-unifying-health-data - Health+ offers a user-centric platform for individuals to manage and control their fragmented health data efficiently.
- Learning to Reason for Multi-Step Retrieval of Personal Context in Personalized Question Answering (viability: 7): https://sciencetostartup.com/paper/learning-to-reason-for-multi-step-retrieval-of-personal-context-in-personalized-question-answering - A reinforcement learning framework for enhanced personalization in Question Answering, outperforming strong baselines with adaptive retrieval-reasoning policies.
- IPv2: An Improved Image Purification Strategy for Real-World Ultra-Low-Dose Lung CT Denoising (viability: 6): https://sciencetostartup.com/paper/ipv2-an-improved-image-purification-strategy-for-real-world-ultra-low-dose-lung-ct-denoising - IPv2 offers an enhanced image purification strategy for improving denoising in ultra-low-dose lung CT scans.
- TOPReward: Token Probabilities as Hidden Zero-Shot Rewards for Robotics (viability: 7): https://sciencetostartup.com/paper/topreward-token-probabilities-as-hidden-zero-shot-rewards-for-robotics - TOPReward uses token probabilities from vision-language models for zero-shot reward estimation in robotics applications.
- Unifying approach to uniform expressivity of graph neural networks (viability: 2): https://sciencetostartup.com/paper/unifying-approach-to-uniform-expressivity-of-graph-neural-networks - Generalize GNN architectures for enhanced expressivity using template embeddings and graded modal logic.
- Leakage and Second-Order Dynamics Improve Hippocampal RNN Replay (viability: 2): https://sciencetostartup.com/paper/leakage-and-second-order-dynamics-improve-hippocampal-rnn-replay - Exploring second-order dynamics and state leakage in RNNs to enhance hippocampal-like replay for complex path integration tasks.
- Learning to Tune Pure Pursuit in Autonomous Racing: Joint Lookahead and Steering-Gain Control with PPO (viability: 7): https://sciencetostartup.com/paper/learning-to-tune-pure-pursuit-in-autonomous-racing-joint-lookahead-and-steering-gain-control-with-ppo - Optimize Pure Pursuit parameters using RL to improve autonomous vehicle path tracking efficiency in real-time.
- FedZMG: Efficient Client-Side Optimization in Federated Learning (viability: 7): https://sciencetostartup.com/paper/fedzmg-efficient-client-side-optimization-in-federated-learning - Develop a federated learning optimizer that enhances performance on edge devices by reducing client-drift efficiently and without communication overhead.
- Zero-shot Interactive Perception (viability: 7): https://sciencetostartup.com/paper/zero-shot-interactive-perception - Enable robots to answer complex queries through zero-shot interactive perception by dynamically manipulating environments.
- "How Do I ...?": Procedural Questions Predominate Student-LLM Chatbot Conversations (viability: 5): https://sciencetostartup.com/paper/how-do-i-procedural-questions-predominate-student-llm-chatbot-conversations - Develop a chatbot framework utilizing LLMs to analyze and classify student procedural questions in educational contexts.
- Validating Political Position Predictions of Arguments (viability: 3): https://sciencetostartup.com/paper/validating-political-position-predictions-of-arguments - Develop a dual-scale validation framework for subjective continuous knowledge representation, particularly in political stance predictions.
- Vichara: Appellate Judgment Prediction and Explanation for the Indian Judicial System (viability: 7): https://sciencetostartup.com/paper/vichara-appellate-judgment-prediction-and-explanation-for-the-indian-judicial-system - Vichara revolutionizes appellate judgment prediction and explanation in the Indian judiciary to expedite case backlog resolution.
- Robo-Saber: Generating and Simulating Virtual Reality Players (viability: 9): https://sciencetostartup.com/paper/robo-saber-generating-and-simulating-virtual-reality-players - Robo-Saber revolutionizes VR game testing by automatically generating realistic player data to streamline development and enhance gameplay analysis.
- JPmHC Dynamical Isometry via Orthogonal Hyper-Connections (viability: 4): https://sciencetostartup.com/paper/jpmhc-dynamical-isometry-via-orthogonal-hyper-connections - Develop a stability-enhancing framework for deep learning architectures using orthogonal hyper-connections.
- Analyzing and Improving Chain-of-Thought Monitorability Through Information Theory (viability: 5): https://sciencetostartup.com/paper/analyzing-and-improving-chain-of-thought-monitorability-through-information-theory - A framework for improving Chain-of-Thought monitor accuracy in LLMs using information theory and targeted training objectives.
- Decoding as Optimisation on the Probability Simplex: From Top-K to Top-P (Nucleus) to Best-of-K Samplers (viability: 5): https://sciencetostartup.com/paper/decoding-as-optimisation-on-the-probability-simplex-from-top-k-to-top-p-nucleus-to-best-of-k-samplers - Unified framework for optimizing language model decoding with novel Best-of-K sampler to improve performance.
- Diffusing to Coordinate: Efficient Online Multi-Agent Diffusion Policies (viability: 7): https://sciencetostartup.com/paper/diffusing-to-coordinate-efficient-online-multi-agent-diffusion-policies - Develop diffusion-based multi-agent coordination tools to optimize online reinforcement learning tasks.
- HyTRec: A Hybrid Temporal-Aware Attention Architecture for Long Behavior Sequential Recommendation (viability: 7): https://sciencetostartup.com/paper/hytrec-a-hybrid-temporal-aware-attention-architecture-for-long-behavior-sequential-recommendation - A hybrid attention architecture for efficient, scalable long behavior sequence recommendations.
- PRISM: Parallel Reward Integration with Symmetry for MORL (viability: 4): https://sciencetostartup.com/paper/prism-parallel-reward-integration-with-symmetry-for-morl - Develop PRISM, a novel MORL algorithm that enhances multi-objective learning efficiency by aligning reward channels symmetrically.
- Simplifying Outcomes of Language Model Component Analyses with ELIA (viability: 5): https://sciencetostartup.com/paper/simplifying-outcomes-of-language-model-component-analyses-with-elia - ELIA provides a user-friendly interactive web app to simplify and explain language model analyses to non-experts.
- On the Adversarial Robustness of Discrete Image Tokenizers (viability: 2): https://sciencetostartup.com/paper/on-the-adversarial-robustness-of-discrete-image-tokenizers - Exploring adversarial robustness in discrete image tokenizers to enhance multimodal system security.
- Thinking by Subtraction: Confidence-Driven Contrastive Decoding for LLM Reasoning (viability: 7): https://sciencetostartup.com/paper/thinking-by-subtraction-confidence-driven-contrastive-decoding-for-llm-reasoning - Confidence-Driven Contrastive Decoding significantly enhances reasoning efficiency in language models by targeting low-confidence tokens.
- [Re] Benchmarking LLM Capabilities in Negotiation through Scoreable Games (viability: 4): https://sciencetostartup.com/paper/re-benchmarking-llm-capabilities-in-negotiation-through-scoreable-games - Evaluate LLM negotiation capabilities using a benchmark based on Scoreable Games with additional metrics for quality and fairness.
- SOMtime the World Ain$'$t Fair: Violating Fairness Using Self-Organizing Maps (viability: 5): https://sciencetostartup.com/paper/somtime-the-world-ain-t-fair-violating-fairness-using-self-organizing-maps - Develop a fairness auditing tool using Self-Organizing Maps for unsupervised machine learning models.
- MeanVoiceFlow: One-step Nonparallel Voice Conversion with Mean Flows (viability: 6): https://sciencetostartup.com/paper/meanvoiceflow-one-step-nonparallel-voice-conversion-with-mean-flows - MeanVoiceFlow offers fast and efficient one-step voice conversion using innovative mean flow techniques without pretraining or distillation.
- Perceived Political Bias in LLMs Reduces Persuasive Abilities (viability: 2): https://sciencetostartup.com/paper/perceived-political-bias-in-llms-reduces-persuasive-abilities - Exploring the impact of perceived political bias in LLMs on their persuasive abilities in conversational settings.
- DohaScript: A Large-Scale Multi-Writer Dataset for Continuous Handwritten Hindi Text (viability: 4): https://sciencetostartup.com/paper/dohascript-a-large-scale-multi-writer-dataset-for-continuous-handwritten-hindi-text - DohaScript provides a standardized dataset for handwritten Hindi text analysis, advancing recognition and modeling capabilities in low resource settings.
- HiAER-Spike Software-Hardware Reconfigurable Platform for Event-Driven Neuromorphic Computing at Scale (viability: 4): https://sciencetostartup.com/paper/hiaer-spike-software-hardware-reconfigurable-platform-for-event-driven-neuromorphic-computing-at-scale - Develop a scalable neuromorphic computing platform for executing large spiking neural networks with real-time performance.
- Gradient Regularization Prevents Reward Hacking in Reinforcement Learning from Human Feedback and Verifiable Rewards (viability: 2): https://sciencetostartup.com/paper/gradient-regularization-prevents-reward-hacking-in-reinforcement-learning-from-human-feedback-and-verifiable-rewards - A novel gradient regularization technique to prevent reward hacking in reinforcement learning models.
- Towards More Standardized AI Evaluation: From Models to Agents (viability: 4): https://sciencetostartup.com/paper/towards-more-standardized-ai-evaluation-from-models-to-agents - Standardize AI evaluation to improve trust and governance in agentic systems.
- Cross-Embodiment Offline Reinforcement Learning for Heterogeneous Robot Datasets (viability: 5): https://sciencetostartup.com/paper/cross-embodiment-offline-reinforcement-learning-for-heterogeneous-robot-datasets - Develop a platform for pre-training robotics policies using heterogeneous robot datasets and offline reinforcement learning.
- Turbo Connection: Reasoning as Information Flow from Higher to Lower Layers (viability: 8): https://sciencetostartup.com/paper/turbo-connection-reasoning-as-information-flow-from-higher-to-lower-layers - TurboConn augments Transformers to significantly enhance reasoning capabilities without increased latency or extensive retraining resources.
- WorkflowPerturb: Calibrated Stress Tests for Evaluating Multi-Agent Workflow Metrics (viability: 5): https://sciencetostartup.com/paper/workflowperturb-calibrated-stress-tests-for-evaluating-multi-agent-workflow-metrics - Create calibrated stress tests to evaluate and improve multi-agent workflow metrics with WorkflowPerturb.
- Learning Optimal and Sample-Efficient Decision Policies with Guarantees (viability: 5): https://sciencetostartup.com/paper/learning-optimal-and-sample-efficient-decision-policies-with-guarantees - Develop sample-efficient RL algorithms with causal guarantees for safer decision-making in high-stakes scenarios.
- In-Context Learning for Pure Exploration in Continuous Spaces (viability: 5): https://sciencetostartup.com/paper/in-context-learning-for-pure-exploration-in-continuous-spaces - Develop a meta-trained algorithm for efficient hypothesis identification in continuous spaces utilizing in-context learning.
- PenTiDef: Enhancing Privacy and Robustness in Decentralized Federated Intrusion Detection Systems against Poisoning Attacks (viability: 5): https://sciencetostartup.com/paper/pentidef-enhancing-privacy-and-robustness-in-decentralized-federated-intrusion-detection-systems-against-poisoning-attac - PenTiDef offers a robust, blockchain-based defense framework for decentralized federated intrusion detection systems to counter poisoning attacks.
- ROCKET: Residual-Oriented Multi-Layer Alignment for Spatially-Aware Vision-Language-Action Models (viability: 6): https://sciencetostartup.com/paper/rocket-residual-oriented-multi-layer-alignment-for-spatially-aware-vision-language-action-models - Enhance VLA models with robust multi-layer alignment for superior 3D spatial reasoning in robotics.
- Causal Neighbourhood Learning for Invariant Graph Representations (viability: 4): https://sciencetostartup.com/paper/causal-neighbourhood-learning-for-invariant-graph-representations - Develop a robust GNN model using causal learning to enhance graph data analysis by reducing spurious correlations.
- Memory-Based Advantage Shaping for LLM-Guided Reinforcement Learning (viability: 5): https://sciencetostartup.com/paper/memory-based-advantage-shaping-for-llm-guided-reinforcement-learning - Leverage memory graphs to enhance sample efficiency in RL environments with sparse rewards, using LLMs for initial subgoal discovery.
- MIRA: Memory-Integrated Reinforcement Learning Agent with Limited LLM Guidance (viability: 5): https://sciencetostartup.com/paper/mira-memory-integrated-reinforcement-learning-agent-with-limited-llm-guidance - MIRA enhances reinforcement learning efficiency by integrating memory-structured LLM guidance, reducing reliance on continuous LLM queries while preserving policy convergence.
- Condition-Gated Reasoning for Context-Dependent Biomedical Question Answering (viability: 5): https://sciencetostartup.com/paper/condition-gated-reasoning-for-context-dependent-biomedical-question-answering - Develop a biomedical QA system that utilizes Condition-Gated Reasoning for more accurate context-dependent medical answers.
- Alignment in Time: Peak-Aware Orchestration for Long-Horizon Agentic Systems (viability: 6): https://sciencetostartup.com/paper/alignment-in-time-peak-aware-orchestration-for-long-horizon-agentic-systems - Develop APEMO, an orchestration tool for maintaining reliability in long-horizon AI agent systems that optimizes resource allocation based on temporal-affective signals.
- Improving Neural Topic Modeling with Semantically-Grounded Soft Label Distributions (viability: 5): https://sciencetostartup.com/paper/improving-neural-topic-modeling-with-semantically-grounded-soft-label-distributions - Enhance neural topic modeling with semantically-grounded soft label distributions for improved topic quality and document retrieval.
- Machine Learning Based Prediction of Surgical Outcomes in Chronic Rhinosinusitis from Clinical Data (viability: 6): https://sciencetostartup.com/paper/machine-learning-based-prediction-of-surgical-outcomes-in-chronic-rhinosinusitis-from-clinical-data - AI tool to predict surgical outcomes for CRS patients using pre-operative clinical data to enhance decision-making.
- Understanding Unreliability of Steering Vectors in Language Models: Geometric Predictors and the Limits of Linear Approximations (viability: 4): https://sciencetostartup.com/paper/understanding-unreliability-of-steering-vectors-in-language-models-geometric-predictors-and-the-limits-of-linear-approxi - Develop a tool to diagnose and improve the reliability of steering vectors in language models, addressing non-linear behavior representations.
- MultiVer: Zero-Shot Multi-Agent Vulnerability Detection (viability: 7): https://sciencetostartup.com/paper/multiver-zero-shot-multi-agent-vulnerability-detection - Develop a multi-agent system for zero-shot vulnerability detection surpassing fine-tuned models in recall.
- Understanding the Fine-Grained Knowledge Capabilities of Vision-Language Models (viability: 4): https://sciencetostartup.com/paper/understanding-the-fine-grained-knowledge-capabilities-of-vision-language-models - Enhanced fine-grained visual understanding for vision-language models through improved vision encoders and pretraining methods.
- MantisV2: Closing the Zero-Shot Gap in Time Series Classification with Synthetic Data and Test-Time Strategies (viability: 4): https://sciencetostartup.com/paper/mantisv2-closing-the-zero-shot-gap-in-time-series-classification-with-synthetic-data-and-test-time-strategies - Develop MantisV2, an advanced zero-shot time series classifier with improved feature extraction capabilities.
- Mind the Style: Impact of Communication Style on Human-Chatbot Interaction (viability: 3): https://sciencetostartup.com/paper/mind-the-style-impact-of-communication-style-on-human-chatbot-interaction - Research on chatbot communication styles shows variations can enhance task success and satisfaction.
- The Token Games: Evaluating Language Model Reasoning with Puzzle Duels (viability: 5): https://sciencetostartup.com/paper/the-token-games-evaluating-language-model-reasoning-with-puzzle-duels - Introducing The Token Games (TTG) as a new framework for evaluating reasoning in language models through self-created puzzle duels.
- Ontology-Guided Neuro-Symbolic Inference: Grounding Language Models with Mathematical Domain Knowledge (viability: 6): https://sciencetostartup.com/paper/ontology-guided-neuro-symbolic-inference-grounding-language-models-with-mathematical-domain-knowledge - Ontology-guided language models enhance verifiable reasoning in specialist fields like mathematics.
- QueryPlot: Generating Geological Evidence Layers using Natural Language Queries for Mineral Exploration (viability: 7): https://sciencetostartup.com/paper/queryplot-generating-geological-evidence-layers-using-natural-language-queries-for-mineral-exploration - QueryPlot provides an interactive platform for geologists to generate mineral prospectivity maps from natural language queries, streamlining mineral exploration efforts.
- Sink-Aware Pruning for Diffusion Language Models (viability: 5): https://sciencetostartup.com/paper/sink-aware-pruning-for-diffusion-language-models - Optimize inference efficiency of Diffusion Language Models through Sink-Aware Pruning for better quality-efficiency trade-off.
- CLEF HIPE-2026: Evaluating Accurate and Efficient Person-Place Relation Extraction from Multilingual Historical Texts (viability: 4): https://sciencetostartup.com/paper/clef-hipe-2026-evaluating-accurate-and-efficient-person-place-relation-extraction-from-multilingual-historical-texts - Develop a tool for extracting and classifying person-place relations in multilingual historical texts to aid digital humanities applications.
- MARS: Margin-Aware Reward-Modeling with Self-Refinement (viability: 5): https://sciencetostartup.com/paper/mars-margin-aware-reward-modeling-with-self-refinement - Develop a framework, MARS, for adaptive data augmentation to enhance reward modeling in reinforcement learning with self-refinement strategies.
- Pushing the Frontier of Black-Box LVLM Attacks via Fine-Grained Detail Targeting (viability: 8): https://sciencetostartup.com/paper/pushing-the-frontier-of-black-box-lvlm-attacks-via-fine-grained-detail-targeting - Enhance security of vision-language models with highly effective black-box adversarial attack tool.
- FAMOSE: A ReAct Approach to Automated Feature Discovery (viability: 6): https://sciencetostartup.com/paper/famose-a-react-approach-to-automated-feature-discovery - FAMOSE leverages ReAct to automate feature engineering, enhancing machine learning model performance for tabular data.
- Reverso: Efficient Time Series Foundation Models for Zero-shot Forecasting (viability: 2): https://sciencetostartup.com/paper/reverso-efficient-time-series-foundation-models-for-zero-shot-forecasting - Develop smaller, more efficient time series models for zero-shot forecasting via interleaved convolutional and RNN layers.
- When to Trust the Cheap Check: Weak and Strong Verification for Reasoning (viability: 3): https://sciencetostartup.com/paper/when-to-trust-the-cheap-check-weak-and-strong-verification-for-reasoning - Develop an online algorithm for balancing weak and strong verification in reasoning tasks with LLMs.
- SMAC: Score-Matched Actor-Critics for Robust Offline-to-Online Transfer (viability: 5): https://sciencetostartup.com/paper/smac-score-matched-actor-critics-for-robust-offline-to-online-transfer - SMAC provides a method for seamless offline-to-online Reinforcement Learning transfer, minimizing performance drops during online fine-tuning.
- Stable Asynchrony: Variance-Controlled Off-Policy RL for LLMs (viability: 5): https://sciencetostartup.com/paper/stable-asynchrony-variance-controlled-off-policy-rl-for-llms - Develop a stable asynchrony method for reinforcement learning in LLMs, reducing training time significantly with minimal performance trade-offs.
- Towards Anytime-Valid Statistical Watermarking (viability: 5): https://sciencetostartup.com/paper/towards-anytime-valid-statistical-watermarking - Develop an e-value-based watermarking framework to efficiently detect AI-generated content with early stopping capabilities.
- AutoNumerics: An Autonomous, PDE-Agnostic Multi-Agent Pipeline for Scientific Computing (viability: 5): https://sciencetostartup.com/paper/autonumerics-an-autonomous-pde-agnostic-multi-agent-pipeline-for-scientific-computing - AutoNumerics automates the creation of interpretable PDE solvers from natural language, offering a versatile tool for scientific computing.
- Adapting Actively on the Fly: Relevance-Guided Online Meta-Learning with Latent Concepts for Geospatial Discovery (viability: 6): https://sciencetostartup.com/paper/adapting-actively-on-the-fly-relevance-guided-online-meta-learning-with-latent-concepts-for-geospatial-discovery - A dynamic geospatial discovery tool using relevance-guided online meta-learning to efficiently identify hidden targets in environments like environmental monitoring.
- MolHIT: Advancing Molecular-Graph Generation with Hierarchical Discrete Diffusion Models (viability: 6): https://sciencetostartup.com/paper/molhit-advancing-molecular-graph-generation-with-hierarchical-discrete-diffusion-models - Develop a cutting-edge molecular graph generation tool using hierarchical discrete diffusion models for enhanced drug discovery and material science applications.
- The Cascade Equivalence Hypothesis: When Do Speech LLMs Behave Like ASR$\rightarrow$LLM Pipelines? (viability: 3): https://sciencetostartup.com/paper/the-cascade-equivalence-hypothesis-when-do-speech-llms-behave-like-asr-rightarrow-llm-pipelines - Research analyzes when speech LLMs act similarly to ASR-to-LLM pipelines, highlighting architectural dependencies.
- AI Gamestore: Scalable, Open-Ended Evaluation of Machine General Intelligence with Human Games (viability: 7): https://sciencetostartup.com/paper/ai-gamestore-scalable-open-ended-evaluation-of-machine-general-intelligence-with-human-games - AI Gamestore revolutionizes AI evaluation by leveraging human-designed games to assess machine general intelligence.
- Conditional Flow Matching for Continuous Anomaly Detection in Autonomous Driving on a Manifold-Aware Spectral Space (viability: 6): https://sciencetostartup.com/paper/conditional-flow-matching-for-continuous-anomaly-detection-in-autonomous-driving-on-a-manifold-aware-spectral-space - Unsupervised anomaly detection framework for improving safety validation in autonomous driving leveraging conditional flow matching.
- Be Wary of Your Time Series Preprocessing (viability: 5): https://sciencetostartup.com/paper/be-wary-of-your-time-series-preprocessing - Develop a principled time series normalization strategy to enhance Transformer model performance.
- A Hybrid Federated Learning Based Ensemble Approach for Lung Disease Diagnosis Leveraging Fusion of SWIN Transformer and CNN (viability: 5): https://sciencetostartup.com/paper/a-hybrid-federated-learning-based-ensemble-approach-for-lung-disease-diagnosis-leveraging-fusion-of-swin-transformer-and - Develop a federated learning-based ensemble model for accurate lung disease diagnosis with enhanced data security.
- ODESteer: A Unified ODE-Based Steering Framework for LLM Alignment (viability: 5): https://sciencetostartup.com/paper/odesteer-a-unified-ode-based-steering-framework-for-llm-alignment - Develop an ODE-based framework to enhance LLM alignment by improving activation steering techniques.
- MASPO: Unifying Gradient Utilization, Probability Mass, and Signal Reliability for Robust and Sample-Efficient LLM Reasoning (viability: 5): https://sciencetostartup.com/paper/maspo-unifying-gradient-utilization-probability-mass-and-signal-reliability-for-robust-and-sample-efficient-llm-reasonin - MASPO enhances reinforcement learning for LLMs through a unified framework that improves robustness and sample-efficiency with open-source code.
- KLong: Training LLM Agent for Extremely Long-horizon Tasks (viability: 8): https://sciencetostartup.com/paper/klong-training-llm-agent-for-extremely-long-horizon-tasks - KLong offers a high-performance LLM agent designed for tackling extremely long-horizon tasks in AI research and development.
- Evaluating Chain-of-Thought Reasoning through Reusability and Verifiability (viability: 5): https://sciencetostartup.com/paper/evaluating-chain-of-thought-reasoning-through-reusability-and-verifiability - Develop enhanced evaluation metrics for multi-agent IR systems leveraging Chain-of-Thought reusability and verifiability.
- Position: Evaluation of ECG Representations Must Be Fixed (viability: 3): https://sciencetostartup.com/paper/position-evaluation-of-ecg-representations-must-be-fixed - Revamp ECG representation evaluation practices for clinical alignment with a new evaluation protocol.
- The Anxiety of Influence: Bloom Filters in Transformer Attention Heads (viability: 5): https://sciencetostartup.com/paper/the-anxiety-of-influence-bloom-filters-in-transformer-attention-heads - Develop compact, efficient membership-testing tools for language models leveraging Bloom filter analogs in attention heads.
- LORA-CRAFT: Cross-layer Rank Adaptation via Frozen Tucker Decomposition of Pre-trained Attention Weights (viability: 5): https://sciencetostartup.com/paper/lora-craft-cross-layer-rank-adaptation-via-frozen-tucker-decomposition-of-pre-trained-attention-weights - Develop parameter-efficient adaptation methods for transformers using frozen Tucker decomposition in CRAFT.
- Pareto Optimal Benchmarking of AI Models on ARM Cortex Processors for Sustainable Embedded Systems (viability: 5): https://sciencetostartup.com/paper/pareto-optimal-benchmarking-of-ai-models-on-arm-cortex-processors-for-sustainable-embedded-systems - Benchmarking framework for optimizing AI models on ARM Cortex processors to ensure energy-efficient embedded systems.
- Learning with Boolean threshold functions (viability: 4): https://sciencetostartup.com/paper/learning-with-boolean-threshold-functions - Develop a tool to train sparse Boolean networks for tasks where traditional methods fail, focusing on interpretability and efficient inference.
- Tracing Copied Pixels and Regularizing Patch Affinity in Copy Detection (viability: 7): https://sciencetostartup.com/paper/tracing-copied-pixels-and-regularizing-patch-affinity-in-copy-detection - Innovative tool for enhancing pixel-level traceability in image copy detection.
- What Do LLMs Associate with Your Name? A Human-Centered Black-Box Audit of Personal Data (viability: 3): https://sciencetostartup.com/paper/what-do-llms-associate-with-your-name-a-human-centered-black-box-audit-of-personal-data - A tool to audit language models for generating personal data associations, aiming to address privacy concerns.
- Jolt Atlas: Verifiable Inference via Lookup Arguments in Zero Knowledge (viability: 5): https://sciencetostartup.com/paper/jolt-atlas-verifiable-inference-via-lookup-arguments-in-zero-knowledge - Jolt Atlas provides a zero-knowledge machine learning framework for verifiable inference, enhancing privacy and security in ML deployments.
- Beyond Pipelines: A Fundamental Study on the Rise of Generative-Retrieval Architectures in Web Research (viability: 3): https://sciencetostartup.com/paper/beyond-pipelines-a-fundamental-study-on-the-rise-of-generative-retrieval-architectures-in-web-research - Survey on the impact of large language models on web research and retrieval-augmented generation (RAG).
- WarpRec: Unifying Academic Rigor and Industrial Scale for Responsible, Reproducible, and Efficient Recommendation (viability: 6): https://sciencetostartup.com/paper/warprec-unifying-academic-rigor-and-industrial-scale-for-responsible-reproducible-and-efficient-recommendation - WarpRec is a framework unifying academic rigor and industrial scale for developing sustainable and efficient recommendation systems.
- Fine-Grained Uncertainty Quantification for Long-Form Language Model Outputs: A Comparative Study (viability: 5): https://sciencetostartup.com/paper/fine-grained-uncertainty-quantification-for-long-form-language-model-outputs-a-comparative-study - Develop an advanced toolkit for fine-grained uncertainty quantification in long-form language model outputs.
- Convergence Analysis of Two-Layer Neural Networks under Gaussian Input Masking (viability: 3): https://sciencetostartup.com/paper/convergence-analysis-of-two-layer-neural-networks-under-gaussian-input-masking - Explore convergence guarantees in two-layer neural network training with Gaussian input masking for enhanced privacy.
- A Privacy by Design Framework for Large Language Model-Based Applications for Children (viability: 7): https://sciencetostartup.com/paper/a-privacy-by-design-framework-for-large-language-model-based-applications-for-children - A Privacy-by-Design framework for creating compliant and secure AI applications for children using Large Language Models.
- Improving LLM-based Recommendation with Self-Hard Negatives from Intermediate Layers (viability: 4): https://sciencetostartup.com/paper/improving-llm-based-recommendation-with-self-hard-negatives-from-intermediate-layers - A novel framework for improving LLM-based recommendation systems leveraging self-hard negative signals from intermediate layers.
- A Contrastive Variational AutoEncoder for NSCLC Survival Prediction with Missing Modalities (viability: 5): https://sciencetostartup.com/paper/a-contrastive-variational-autoencoder-for-nsclc-survival-prediction-with-missing-modalities - Develop a robust predictive tool for NSCLC survival outcomes using a Multimodal Contrastive Variational AutoEncoder that handles missing clinical data modalities.
- A High-Level Survey of Optical Remote Sensing (viability: 3): https://sciencetostartup.com/paper/a-high-level-survey-of-optical-remote-sensing - Comprehensive survey of optical remote sensing leveraging drones, guiding new researchers with insights and code resources.
- SpectralGCD: Spectral Concept Selection and Cross-modal Representation Learning for Generalized Category Discovery (viability: 7): https://sciencetostartup.com/paper/spectralgcd-spectral-concept-selection-and-cross-modal-representation-learning-for-generalized-category-discovery - SpectralGCD offers an efficient cross-modal representation learning tool for Generalized Category Discovery, significantly reducing computational costs.
- Voice-Driven Semantic Perception for UAV-Assisted Emergency Networks (viability: 7): https://sciencetostartup.com/paper/voice-driven-semantic-perception-for-uav-assisted-emergency-networks - Transform emergency voice communications into structured data for UAV network management.
- Visual Model Checking: Graph-Based Inference of Visual Routines for Image Retrieval (viability: 3): https://sciencetostartup.com/paper/visual-model-checking-graph-based-inference-of-visual-routines-for-image-retrieval - Integrate formal verification with image retrieval for transparent and accountable query processing.
- Dataless Weight Disentanglement in Task Arithmetic via Kronecker-Factored Approximate Curvature (viability: 2): https://sciencetostartup.com/paper/dataless-weight-disentanglement-in-task-arithmetic-via-kronecker-factored-approximate-curvature - Introduce a dataless regularization method for task vectors using Kronecker-Factored Approximate Curvature to mitigate cross-task interference without external data.
- A feature-stable and explainable machine learning framework for trustworthy decision-making under incomplete clinical data (viability: 7): https://sciencetostartup.com/paper/a-feature-stable-and-explainable-machine-learning-framework-for-trustworthy-decision-making-under-incomplete-clinical-da - Develop a robust and interpretable machine learning framework for handling incomplete clinical data in medical decision-making.
- What Breaks Embodied AI Security:LLM Vulnerabilities, CPS Flaws,or Something Else? (viability: 2): https://sciencetostartup.com/paper/what-breaks-embodied-ai-security-llm-vulnerabilities-cps-flaws-or-something-else - This paper explores the unique security challenges faced by embodied AI systems, emphasizing a system-level approach to address physical risks and failure propagation.
- From Subtle to Significant: Prompt-Driven Self-Improving Optimization in Test-Time Graph OOD Detection (viability: 6): https://sciencetostartup.com/paper/from-subtle-to-significant-prompt-driven-self-improving-optimization-in-test-time-graph-ood-detection - SIGOOD is an unsupervised tool for graph OOD detection, using self-improving optimization to enhance reliability in GNNs.
- SubQuad: Near-Quadratic-Free Structure Inference with Distribution-Balanced Objectives in Adaptive Receptor framework (viability: 5): https://sciencetostartup.com/paper/subquad-near-quadratic-free-structure-inference-with-distribution-balanced-objectives-in-adaptive-receptor-framework - SubQuad offers a scalable, bias-aware platform for adaptive immune repertoire mining and translational tasks like vaccine target prioritization.
- WebFAQ 2.0: A Multilingual QA Dataset with Mined Hard Negatives for Dense Retrieval (viability: 7): https://sciencetostartup.com/paper/webfaq-2-0-a-multilingual-qa-dataset-with-mined-hard-negatives-for-dense-retrieval - WebFAQ 2.0 is a large-scale multilingual QA dataset with hard negatives, enabling improved dense retrieval systems.
- Same Meaning, Different Scores: Lexical and Syntactic Sensitivity in LLM Evaluation (viability: 4): https://sciencetostartup.com/paper/same-meaning-different-scores-lexical-and-syntactic-sensitivity-in-llm-evaluation - Develop a robustness testing tool for evaluating LLMs against lexical and syntactic perturbations.
- Flickering Multi-Armed Bandits (viability: 3): https://sciencetostartup.com/paper/flickering-multi-armed-bandits - Develop a decision-making algorithm for dynamic environments using Flickering Multi-Armed Bandits.
- MedClarify: An information-seeking AI agent for medical diagnosis with case-specific follow-up questions (viability: 7): https://sciencetostartup.com/paper/medclarify-an-information-seeking-ai-agent-for-medical-diagnosis-with-case-specific-follow-up-questions - MedClarify uses AI to enhance medical diagnosis by generating follow-up questions to reduce uncertainty.
- ArXiv-to-Model: A Practical Study of Scientific LM Training (viability: 5): https://sciencetostartup.com/paper/arxiv-to-model-a-practical-study-of-scientific-lm-training - Develop a pipeline for training domain-specific scientific language models under limited compute resources.
- Towards Cross-lingual Values Assessment: A Consensus-Pluralism Perspective (viability: 4): https://sciencetostartup.com/paper/towards-cross-lingual-values-assessment-a-consensus-pluralism-perspective - Develop a benchmark tool for evaluating large language models' ability in cross-lingual values assessment.
- Federated Latent Space Alignment for Multi-user Semantic Communications (viability: 3): https://sciencetostartup.com/paper/federated-latent-space-alignment-for-multi-user-semantic-communications - Federated optimization method for aligning latent spaces in multi-agent semantic communication systems.
- Web Verbs: Typed Abstractions for Reliable Task Composition on the Agentic Web (viability: 6): https://sciencetostartup.com/paper/web-verbs-typed-abstractions-for-reliable-task-composition-on-the-agentic-web - A new typed abstraction layer, Web Verbs, improves agent capabilities on the web by offering a stable, composable set of functions for reliable task composition.
- TAPO-Structured Description Logic for Information Behavior: Procedural and Oracle-Based Extensions (viability: 3): https://sciencetostartup.com/paper/tapo-structured-description-logic-for-information-behavior-procedural-and-oracle-based-extensions - TAPO-DL: A theoretical framework combining description logic with procedural dynamics for modeling dynamic information behaviors.
- All Leaks Count, Some Count More: Interpretable Temporal Contamination Detection in LLM Backtesting (viability: 5): https://sciencetostartup.com/paper/all-leaks-count-some-count-more-interpretable-temporal-contamination-detection-in-llm-backtesting - Detect and mitigate temporal contamination in historical backtesting of LLMs.
- Mechanistic Interpretability of Cognitive Complexity in LLMs via Linear Probing using Bloom's Taxonomy (viability: 5): https://sciencetostartup.com/paper/mechanistic-interpretability-of-cognitive-complexity-in-llms-via-linear-probing-using-bloom-s-taxonomy - Develop linear probing tools to interpret cognitive complexities in LLMs using Bloom's Taxonomy.
- Decoding the Human Factor: High Fidelity Behavioral Prediction for Strategic Foresight (viability: 8): https://sciencetostartup.com/paper/decoding-the-human-factor-high-fidelity-behavioral-prediction-for-strategic-foresight - Large Behavioral Model predicts individual strategic decisions for applications in foresight, negotiation, and decision support.
- From Labor to Collaboration: A Methodological Experiment Using AI Agents to Augment Research Perspectives in Taiwan's Humanities and Social Sciences (viability: 5): https://sciencetostartup.com/paper/from-labor-to-collaboration-a-methodological-experiment-using-ai-agents-to-augment-research-perspectives-in-taiwan-s-hum - Develop an AI-driven collaborative research framework for humanities and social sciences.
- Continual learning and refinement of causal models through dynamic predicate invention (viability: 5): https://sciencetostartup.com/paper/continual-learning-and-refinement-of-causal-models-through-dynamic-predicate-invention - Develop symbolic causal world models online using dynamic predicate invention for efficient agent decision-making.
- Texo: Formula Recognition within 20M Parameters (viability: 8): https://sciencetostartup.com/paper/texo-formula-recognition-within-20m-parameters - Texo is a lightweight formula recognition model that runs efficiently on consumer-grade hardware and is ready for real-time in-browser deployment.
- The Bots of Persuasion: Examining How Conversational Agents' Linguistic Expressions of Personality Affect User Perceptions and Decisions (viability: 3): https://sciencetostartup.com/paper/the-bots-of-persuasion-examining-how-conversational-agents-linguistic-expressions-of-personality-affect-user-perceptions - Explore how conversational agents' language affects user perceptions and decisions in charitable contexts.
- Robustness and Reasoning Fidelity of Large Language Models in Long-Context Code Question Answering (viability: 5): https://sciencetostartup.com/paper/robustness-and-reasoning-fidelity-of-large-language-models-in-long-context-code-question-answering - Develop a benchmarking tool for evaluating the robustness of LLMs in long-context code question answering.
- Continual uncertainty learning (viability: 3): https://sciencetostartup.com/paper/continual-uncertainty-learning - Develop a curriculum-based continual learning framework for robust control in mechanical systems with multiple uncertainties.
- In-Context Learning in Linear vs. Quadratic Attention Models: An Empirical Study on Regression Tasks (viability: 4): https://sciencetostartup.com/paper/in-context-learning-in-linear-vs-quadratic-attention-models-an-empirical-study-on-regression-tasks - Explore a code-accompanied analysis of linear vs quadratic attention in in-context learning for regression.
- JEPA-DNA: Grounding Genomic Foundation Models through Joint-Embedding Predictive Architectures (viability: 5): https://sciencetostartup.com/paper/jepa-dna-grounding-genomic-foundation-models-through-joint-embedding-predictive-architectures - JEPA-DNA enhances genomic foundation models by introducing a novel joint-embedding predictive architecture for superior performance on genomic tasks.
- TimeOmni-VL: Unified Models for Time Series Understanding and Generation (viability: 5): https://sciencetostartup.com/paper/timeomni-vl-unified-models-for-time-series-understanding-and-generation - Unify time series understanding and generation with vision-centric models.
- Bonsai: A Framework for Convolutional Neural Network Acceleration Using Criterion-Based Pruning (viability: 5): https://sciencetostartup.com/paper/bonsai-a-framework-for-convolutional-neural-network-acceleration-using-criterion-based-pruning - A modular framework for accelerating CNNs through innovative pruning techniques.
- VP-VAE: Rethinking Vector Quantization via Adaptive Vector Perturbation (viability: 5): https://sciencetostartup.com/paper/vp-vae-rethinking-vector-quantization-via-adaptive-vector-perturbation - Develop a more stable and robust generative model for VQ-VAE with adaptive vector perturbation eliminating the need for a codebook.
- Efficient Parallel Algorithm for Decomposing Hard CircuitSAT Instances (viability: 5): https://sciencetostartup.com/paper/efficient-parallel-algorithm-for-decomposing-hard-circuitsat-instances - A parallel algorithm to efficiently decompose challenging CircuitSAT problems for advanced circuit analysis.
- 3D Scene Rendering with Multimodal Gaussian Splatting (viability: 6): https://sciencetostartup.com/paper/3d-scene-rendering-with-multimodal-gaussian-splatting - A robust multimodal 3D scene rendering framework integrating RF sensing with Gaussian Splatting for high-fidelity visual reconstruction in challenging environments.
- TIFO: Time-Invariant Frequency Operator for Stationarity-Aware Representation Learning in Time Series (viability: 7): https://sciencetostartup.com/paper/tifo-time-invariant-frequency-operator-for-stationarity-aware-representation-learning-in-time-series - Develop a plug-and-play forecasting enhancer that improves nonstationary time series predictions by leveraging frequency domain insights.
- Epistemology of Generative AI: The Geometry of Knowing (viability: 2): https://sciencetostartup.com/paper/epistemology-of-generative-ai-the-geometry-of-knowing - Develop a new epistemological framework for understanding generative AI's impact on knowledge production.
- Instructor-Aligned Knowledge Graphs for Personalized Learning (viability: 5): https://sciencetostartup.com/paper/instructor-aligned-knowledge-graphs-for-personalized-learning - InstructKG automates the construction of instructor-aligned knowledge graphs to enhance personalized learning by mapping educational concepts and their dependencies.
- Owen-based Semantics and Hierarchy-Aware Explanation (O-Shap) (viability: 5): https://sciencetostartup.com/paper/owen-based-semantics-and-hierarchy-aware-explanation-o-shap - Developing O-Shap, a hierarchical SHAP method for more precise and coherent AI model explanations.
- Toward Trustworthy Evaluation of Sustainability Rating Methodologies: A Human-AI Collaborative Framework for Benchmark Dataset Construction (viability: 5): https://sciencetostartup.com/paper/toward-trustworthy-evaluation-of-sustainability-rating-methodologies-a-human-ai-collaborative-framework-for-benchmark-da - A framework for creating benchmark datasets to standardize sustainability ratings using human-AI collaboration.
- Agentic Wireless Communication for 6G: Intent-Aware and Continuously Evolving Physical-Layer Intelligence (viability: 3): https://sciencetostartup.com/paper/agentic-wireless-communication-for-6g-intent-aware-and-continuously-evolving-physical-layer-intelligence - Develop adaptive, intent-driven network agents for autonomous 6G physical layer optimization.
- FLoRG: Federated Fine-tuning with Low-rank Gram Matrices and Procrustes Alignment (viability: 7): https://sciencetostartup.com/paper/florg-federated-fine-tuning-with-low-rank-gram-matrices-and-procrustes-alignment - FLoRG optimizes federated learning with low-rank matrices to boost model accuracy and reduce communication overhead.
- How AI Coding Agents Communicate: A Study of Pull Request Description Characteristics and Human Review Responses (viability: 5): https://sciencetostartup.com/paper/how-ai-coding-agents-communicate-a-study-of-pull-request-description-characteristics-and-human-review-responses - Develop AI tools to enhance pull request descriptions and improve human-AI code review collaboration.
- AdvSynGNN: Structure-Adaptive Graph Neural Nets via Adversarial Synthesis and Self-Corrective Propagation (viability: 5): https://sciencetostartup.com/paper/advsyngnn-structure-adaptive-graph-neural-nets-via-adversarial-synthesis-and-self-corrective-propagation - AdvSynGNN enhances graph neural networks' resilience and accuracy in noisy structures through adversarial synthesis and self-corrective propagation.
- Predictive Batch Scheduling: Accelerating Language Model Training Through Loss-Aware Sample Prioritization (viability: 5): https://sciencetostartup.com/paper/predictive-batch-scheduling-accelerating-language-model-training-through-loss-aware-sample-prioritization - Develop a tool that accelerates language model training by prioritizing high-loss samples for improved convergence speed.
- Sign Lock-In: Randomly Initialized Weight Signs Persist and Bottleneck Sub-Bit Model Compression (viability: 2): https://sciencetostartup.com/paper/sign-lock-in-randomly-initialized-weight-signs-persist-and-bottleneck-sub-bit-model-compression - Develop an efficient sub-bit model compression technique to minimize storage costs without significant loss of accuracy.
- Retaining Suboptimal Actions to Follow Shifting Optima in Multi-Agent Reinforcement Learning (viability: 6): https://sciencetostartup.com/paper/retaining-suboptimal-actions-to-follow-shifting-optima-in-multi-agent-reinforcement-learning - Develop a reinforcement learning framework to improve adaptability in multi-agent systems using successive sub-value functions.
- ALPS: A Diagnostic Challenge Set for Arabic Linguistic & Pragmatic Reasoning (viability: 6): https://sciencetostartup.com/paper/alps-a-diagnostic-challenge-set-for-arabic-linguistic-pragmatic-reasoning - ALPS offers a native, expert-curated diagnostic challenge set for evaluating Arabic linguistic and pragmatic reasoning in NLP models.
- RFEval: Benchmarking Reasoning Faithfulness under Counterfactual Reasoning Intervention in Large Reasoning Models (viability: 5): https://sciencetostartup.com/paper/rfeval-benchmarking-reasoning-faithfulness-under-counterfactual-reasoning-intervention-in-large-reasoning-models - Develop a benchmark tool to improve reasoning faithfulness in large reasoning models by assessing stance consistency and causal influence.
- Evaluating Cross-Lingual Classification Approaches Enabling Topic Discovery for Multilingual Social Media Data (viability: 5): https://sciencetostartup.com/paper/evaluating-cross-lingual-classification-approaches-enabling-topic-discovery-for-multilingual-social-media-data - Develop a tool for cross-lingual topic discovery from multilingual social media using various text classification approaches.
- IntentCUA: Learning Intent-level Representations for Skill Abstraction and Multi-Agent Planning in Computer-Use Agents (viability: 2): https://sciencetostartup.com/paper/intentcua-learning-intent-level-representations-for-skill-abstraction-and-multi-agent-planning-in-computer-use-agents - IntentCUA enhances desktop automation by using intent-level representations and multi-agent planning for improved skill abstraction and task efficiency.
- Phase-Aware Mixture of Experts for Agentic Reinforcement Learning (viability: 4): https://sciencetostartup.com/paper/phase-aware-mixture-of-experts-for-agentic-reinforcement-learning - Develop an enhanced Mixture of Experts model for more efficient reinforcement learning in complex task-solving by LLM agents.
- Wink: Recovering from Misbehaviors in Coding Agents (viability: 7): https://sciencetostartup.com/paper/wink-recovering-from-misbehaviors-in-coding-agents - Automated recovery system for coding agent misbehaviors to enhance software development efficiency.
- Forecasting Anomaly Precursors via Uncertainty-Aware Time-Series Ensembles (viability: 5): https://sciencetostartup.com/paper/forecasting-anomaly-precursors-via-uncertainty-aware-time-series-ensembles - FATE uses uncertainty-aware time-series ensembles for proactive anomaly precursor detection, improving predictive capabilities in critical domains.
- Transforming Behavioral Neuroscience Discovery with In-Context Learning and AI-Enhanced Tensor Methods (viability: 6): https://sciencetostartup.com/paper/transforming-behavioral-neuroscience-discovery-with-in-context-learning-and-ai-enhanced-tensor-methods - AI-enhanced pipeline transforms neuroscience data analysis using in-context learning for efficient insights into PTSD research.
- ReIn: Conversational Error Recovery with Reasoning Inception (viability: 7): https://sciencetostartup.com/paper/rein-conversational-error-recovery-with-reasoning-inception - REIN offers a tool for conversational agents to recover from errors in real-time by injecting reasoning steps without altering model parameters.
- M2F: Automated Formalization of Mathematical Literature at Scale (viability: 7): https://sciencetostartup.com/paper/m2f-automated-formalization-of-mathematical-literature-at-scale - Automated solution to convert mathematical literature into formal code using Lean efficiently.
- Persona2Web: Benchmarking Personalized Web Agents for Contextual Reasoning with User History (viability: 5): https://sciencetostartup.com/paper/persona2web-benchmarking-personalized-web-agents-for-contextual-reasoning-with-user-history - Persona2Web benchmarks and enables the development of personalized web agents using user history for contextual reasoning.
- Sonar-TS: Search-Then-Verify Natural Language Querying for Time Series Databases (viability: 6): https://sciencetostartup.com/paper/sonar-ts-search-then-verify-natural-language-querying-for-time-series-databases - Neuro-symbolic framework for intuitive natural language querying of time series databases with significant scalability improvements.
- Exploring LLMs for User Story Extraction from Mockups (viability: 6): https://sciencetostartup.com/paper/exploring-llms-for-user-story-extraction-from-mockups - Automated extraction of user stories from mockups using LLMs to streamline requirements engineering.
- Conv-FinRe: A Conversational and Longitudinal Benchmark for Utility-Grounded Financial Recommendation (viability: 6): https://sciencetostartup.com/paper/conv-finre-a-conversational-and-longitudinal-benchmark-for-utility-grounded-financial-recommendation - A conversational benchmark for financial advisory systems that distinguishes between behavior imitation and decision quality using real market data.
- Fundamental Limits of Black-Box Safety Evaluation: Information-Theoretic and Computational Barriers from Latent Context Conditioning (viability: 3): https://sciencetostartup.com/paper/fundamental-limits-of-black-box-safety-evaluation-information-theoretic-and-computational-barriers-from-latent-context-c - Research on fundamental limits of black-box safety evaluation indicates the need for additional safeguards in AI system deployment.
- HQFS: Hybrid Quantum Classical Financial Security with VQC Forecasting, QUBO Annealing, and Audit-Ready Post-Quantum Signing (viability: 7): https://sciencetostartup.com/paper/hqfs-hybrid-quantum-classical-financial-security-with-vqc-forecasting-qubo-annealing-and-audit-ready-post-quantum-signin - HQFS offers a hybrid quantum-classical pipeline for enhanced financial security and auditing, reducing prediction errors and optimization time.
- DDiT: Dynamic Patch Scheduling for Efficient Diffusion Transformers (viability: 6): https://sciencetostartup.com/paper/ddit-dynamic-patch-scheduling-for-efficient-diffusion-transformers - Develop a dynamic tokenization strategy for Diffusion Transformers to significantly improve computational efficiency in image and video generation.
- Early-Warning Signals of Grokking via Loss-Landscape Geometry (viability: 5): https://sciencetostartup.com/paper/early-warning-signals-of-grokking-via-loss-landscape-geometry - Develop and implement an early-warning signal system for delayed generalization in machine learning models based on the commutator defect metric.
- A Unified Framework for Locality in Scalable MARL (viability: 2): https://sciencetostartup.com/paper/a-unified-framework-for-locality-in-scalable-marl - Develop a framework for exploiting locality in scalable Multi-Agent Reinforcement Learning (MARL).
- Eigenmood Space: Uncertainty-Aware Spectral Graph Analysis of Psychological Patterns in Classical Persian Poetry (viability: 2): https://sciencetostartup.com/paper/eigenmood-space-uncertainty-aware-spectral-graph-analysis-of-psychological-patterns-in-classical-persian-poetry - A computational framework for analyzing psychological patterns in Persian poetry using spectral graph analysis and uncertainty modeling.
- Automating Agent Hijacking via Structural Template Injection (viability: 5): https://sciencetostartup.com/paper/automating-agent-hijacking-via-structural-template-injection - Automated framework for agent hijacking in LLMs, exploiting structured template injection to enhance attack success and transferability.
- When Semantic Overlap Is Not Enough: Cross-Lingual Euphemism Transfer Between Turkish and English (viability: 2): https://sciencetostartup.com/paper/when-semantic-overlap-is-not-enough-cross-lingual-euphemism-transfer-between-turkish-and-english - Cross-lingual euphemism detection between Turkish and English faces challenges in transfer due to cultural context reliance and semantic asymmetry.
- LLM4Cov: Execution-Aware Agentic Learning for High-coverage Testbench Generation (viability: 5): https://sciencetostartup.com/paper/llm4cov-execution-aware-agentic-learning-for-high-coverage-testbench-generation - Develop execution-aware agent learning framework for efficient high-coverage hardware verification.
- Beyond Message Passing: A Symbolic Alternative for Expressive and Interpretable Graph Learning (viability: 7): https://sciencetostartup.com/paper/beyond-message-passing-a-symbolic-alternative-for-expressive-and-interpretable-graph-learning - Launch SYMGRAPH, a symbolic graph learning framework offering unmatched interpretability and efficiency for high-stakes industries like drug discovery.
- Mind the GAP: Text Safety Does Not Transfer to Tool-Call Safety in LLM Agents (viability: 6): https://sciencetostartup.com/paper/mind-the-gap-text-safety-does-not-transfer-to-tool-call-safety-in-llm-agents - Introducing the GAP benchmark to measure and mitigate divergence in text and tool-call safety for LLM agents across multiple domains.
- SourceBench: Can AI Answers Reference Quality Web Sources? (viability: 5): https://sciencetostartup.com/paper/sourcebench-can-ai-answers-reference-quality-web-sources - SourceBench evaluates the web source quality of AI-generated answers using a comprehensive benchmark framework.
- DeepContext: Stateful Real-Time Detection of Multi-Turn Adversarial Intent Drift in LLMs (viability: 6): https://sciencetostartup.com/paper/deepcontext-stateful-real-time-detection-of-multi-turn-adversarial-intent-drift-in-llms - DeepContext offers stateful real-time detection of adversarial intent drift in LLMs, outperforming existing guardrails with low-latency processing.
- RankEvolve: Automating the Discovery of Retrieval Algorithms via LLM-Driven Evolution (viability: 7): https://sciencetostartup.com/paper/rankevolve-automating-the-discovery-of-retrieval-algorithms-via-llm-driven-evolution - RankEvolve automates the discovery of novel retrieval algorithms using LLM-driven evolutionary search.
- Narrow fine-tuning erodes safety alignment in vision-language agents (viability: 5): https://sciencetostartup.com/paper/narrow-fine-tuning-erodes-safety-alignment-in-vision-language-agents - Develop a tool to monitor and mitigate safety misalignment in continuously learning vision-language models.
- Say It My Way: Exploring Control in Conversational Visual Question Answering with Blind Users (viability: 6): https://sciencetostartup.com/paper/say-it-my-way-exploring-control-in-conversational-visual-question-answering-with-blind-users - A customizable conversational VQA tool empowering blind users through tailored interaction techniques.
- Discovering Multiagent Learning Algorithms with Large Language Models (viability: 5): https://sciencetostartup.com/paper/discovering-multiagent-learning-algorithms-with-large-language-models - AlphaEvolve automatically discovers superior multiagent learning algorithms using large language models.
- Xray-Visual Models: Scaling Vision models on Industry Scale Data (viability: 3): https://sciencetostartup.com/paper/xray-visual-models-scaling-vision-models-on-industry-scale-data - Xray-Visual is a scalable multimodal vision model architecture achieving state-of-the-art performance on image and video tasks.
- A Reversible Semantics for Janus (viability: 3): https://sciencetostartup.com/paper/a-reversible-semantics-for-janus - Develop a reversible small-step semantics for the Janus programming language to enhance debugging and concurrency extensions.
- LLM-WikiRace: Benchmarking Long-term Planning and Reasoning over Real-World Knowledge Graphs (viability: 4): https://sciencetostartup.com/paper/llm-wikirace-benchmarking-long-term-planning-and-reasoning-over-real-world-knowledge-graphs - LLM-Wikirace provides a benchmarking tool for evaluating LLMs on long-term planning and reasoning tasks using Wikipedia's knowledge graph.
- AgentLAB: Benchmarking LLM Agents against Long-Horizon Attacks (viability: 5): https://sciencetostartup.com/paper/agentlab-benchmarking-llm-agents-against-long-horizon-attacks - Develop AgentLAB, a benchmark for tracking and improving LLM agent security against long-horizon attacks.
- MALLVI: a multi agent framework for integrated generalized robotics manipulation (viability: 7): https://sciencetostartup.com/paper/mallvi-a-multi-agent-framework-for-integrated-generalized-robotics-manipulation - MALLVi provides a multi-agent framework for robust closed-loop robotic manipulation based on natural language inputs and visual feedback.
- OpenSage: Self-programming Agent Generation Engine (viability: 6): https://sciencetostartup.com/paper/opensage-self-programming-agent-generation-engine - OpenSage is a self-programming engine for AI agents, automating the creation of agent topologies, tools, and memory systems.
- AdaptOrch: Task-Adaptive Multi-Agent Orchestration in the Era of LLM Performance Convergence (viability: 7): https://sciencetostartup.com/paper/adaptorch-task-adaptive-multi-agent-orchestration-in-the-era-of-llm-performance-convergence - AdaptOrch optimizes multi-agent orchestration topologies for task performance in converging LLMs.
- Position: Why a Dynamical Systems Perspective is Needed to Advance Time Series Modeling (viability: 2): https://sciencetostartup.com/paper/position-why-a-dynamical-systems-perspective-is-needed-to-advance-time-series-modeling - Leverage dynamical systems theory to enhance time series modeling with better forecasting and reduced computational demands.
- Overseeing Agents Without Constant Oversight: Challenges and Opportunities (viability: 3): https://sciencetostartup.com/paper/overseeing-agents-without-constant-oversight-challenges-and-opportunities - Study highlights the challenges of designing effective oversight interfaces for agentic AI systems without constant supervision.
- VAM: Verbalized Action Masking for Controllable Exploration in RL Post-Training -- A Chess Case Study (viability: 5): https://sciencetostartup.com/paper/vam-verbalized-action-masking-for-controllable-exploration-in-rl-post-training-a-chess-case-study - Enhance RL strategies in post-training for large language models using Verbalized Action Masking for efficient exploration control, notably in chess scenarios.
- IndicJR: A Judge-Free Benchmark of Jailbreak Robustness in South Asian Languages (viability: 6): https://sciencetostartup.com/paper/indicjr-a-judge-free-benchmark-of-jailbreak-robustness-in-south-asian-languages - Build a multilingual jailbreak robustness testing tool for South Asian languages.
- Learning under noisy supervision is governed by a feedback-truth gap (viability: 5): https://sciencetostartup.com/paper/learning-under-noisy-supervision-is-governed-by-a-feedback-truth-gap - Develop a system for mitigating feedback-truth gaps in learning under noisy supervision, leveraging insights from neural and human studies.
- An order-oriented approach to scoring hesitant fuzzy elements (viability: 1): https://sciencetostartup.com/paper/an-order-oriented-approach-to-scoring-hesitant-fuzzy-elements - Develop a unified scoring framework for hesitant fuzzy elements leveraging order theory.
- HiVAE: Hierarchical Latent Variables for Scalable Theory of Mind (viability: 3): https://sciencetostartup.com/paper/hivae-hierarchical-latent-variables-for-scalable-theory-of-mind - HiVAE scales Theory of Mind reasoning to realistic environments using hierarchical latent variables.
- AI-Mediated Feedback Improves Student Revisions: A Randomized Trial with FeedbackWriter in a Large Undergraduate Course (viability: 5): https://sciencetostartup.com/paper/ai-mediated-feedback-improves-student-revisions-a-randomized-trial-with-feedbackwriter-in-a-large-undergraduate-course - Enhance student essay revisions with AI-mediated feedback through FeedbackWriter in educational settings.
- Node Learning: A Framework for Adaptive, Decentralised and Collaborative Network Edge AI (viability: 3): https://sciencetostartup.com/paper/node-learning-a-framework-for-adaptive-decentralised-and-collaborative-network-edge-ai - Node Learning offers an adaptive, decentralized AI framework for edge networks, emphasizing local learning and peer interaction without central aggregation.
- One-step Language Modeling via Continuous Denoising (viability: 7): https://sciencetostartup.com/paper/one-step-language-modeling-via-continuous-denoising - Accelerate language model generation with a flow-based denoising approach outperforming discrete diffusion in both speed and quality.
- NeuDiff Agent: A Governed AI Workflow for Single-Crystal Neutron Crystallography (viability: 7): https://sciencetostartup.com/paper/neudiff-agent-a-governed-ai-workflow-for-single-crystal-neutron-crystallography - Accelerate and streamline single-crystal neutron crystallography workflow with governed AI automation for improved efficiency and traceability.
- Evaluating Monolingual and Multilingual Large Language Models for Greek Question Answering: The DemosQA Benchmark (viability: 5): https://sciencetostartup.com/paper/evaluating-monolingual-and-multilingual-large-language-models-for-greek-question-answering-the-demosqa-benchmark - Develop a Greek-specific QA system leveraging a new dataset and evaluation framework for enhancing LLM performance in under-resourced languages.
- Simple Baselines are Competitive with Code Evolution (viability: 5): https://sciencetostartup.com/paper/simple-baselines-are-competitive-with-code-evolution - Proposing simpler baselines for code evolution techniques in AI, with competitive performance to complex methods.
- References Improve LLM Alignment in Non-Verifiable Domains (viability: 6): https://sciencetostartup.com/paper/references-improve-llm-alignment-in-non-verifiable-domains - Use reference-guided LLM-evaluators to improve alignment and self-improvement in non-verifiable domains.
- Large-scale online deanonymization with LLMs (viability: 5): https://sciencetostartup.com/paper/large-scale-online-deanonymization-with-llms - A scalable tool using LLMs for high-precision online user deanonymization across platforms enhances digital forensics capabilities.
- Policy Compiler for Secure Agentic Systems (viability: 6): https://sciencetostartup.com/paper/policy-compiler-for-secure-agentic-systems - PCAS provides secure, policy-compliant deployment of agentic systems by compiling agent implementations into deterministically enforced systems without security-specific restructuring.
- Measuring Mid-2025 LLM-Assistance on Novice Performance in Biology (viability: 3): https://sciencetostartup.com/paper/measuring-mid-2025-llm-assistance-on-novice-performance-in-biology - Evaluate the real-world benefits of LLMs in improving novice performance in biological laboratory settings.
- Calibrate-Then-Act: Cost-Aware Exploration in LLM Agents (viability: 3): https://sciencetostartup.com/paper/calibrate-then-act-cost-aware-exploration-in-llm-agents - A framework for LLMs to optimize decision-making by balancing cost and uncertainty through a Calibrate-Then-Act approach.
- SPARC: Scenario Planning and Reasoning for Automated C Unit Test Generation (viability: 5): https://sciencetostartup.com/paper/sparc-scenario-planning-and-reasoning-for-automated-c-unit-test-generation - SPARC offers a neuro-symbolic framework for generating more accurate and maintainable C unit tests.
- Towards a Science of AI Agent Reliability (viability: 5): https://sciencetostartup.com/paper/towards-a-science-of-ai-agent-reliability - Develop metrics and tools to assess and improve AI agent reliability across consistency, robustness, predictability, and safety dimensions.
- Align Once, Benefit Multilingually: Enforcing Multilingual Consistency for LLM Safety Alignment (viability: 6): https://sciencetostartup.com/paper/align-once-benefit-multilingually-enforcing-multilingual-consistency-for-llm-safety-alignment - Align multilingual LLM safety with a resource-efficient method for consistent semantic direction across languages.
- Agent Skill Framework: Perspectives on the Potential of Small Language Models in Industrial Environments (viability: 6): https://sciencetostartup.com/paper/agent-skill-framework-perspectives-on-the-potential-of-small-language-models-in-industrial-environments - Agent Skill framework optimizes small language models for industrial applications to improve efficiency and accuracy.
- Retrieval Augmented Generation of Literature-derived Polymer Knowledge: The Example of a Biodegradable Polymer Expert System (viability: 3): https://sciencetostartup.com/paper/retrieval-augmented-generation-of-literature-derived-polymer-knowledge-the-example-of-a-biodegradable-polymer-expert-sys - Develop a literature-driven expert system using retrieval-augmented generation to enhance polymer research insights.
- Almost Sure Convergence of Differential Temporal Difference Learning for Average Reward Markov Decision Processes (viability: 3): https://sciencetostartup.com/paper/almost-sure-convergence-of-differential-temporal-difference-learning-for-average-reward-markov-decision-processes - Advanced differential temporal difference learning convergence for average reward in reinforcement learning.
- A Systematic Evaluation of Sample-Level Tokenization Strategies for MEG Foundation Models (viability: 4): https://sciencetostartup.com/paper/a-systematic-evaluation-of-sample-level-tokenization-strategies-for-meg-foundation-models - Develop efficient tokenization strategies for MEG foundation models with publicly available code.
- Causal and Compositional Abstraction (viability: 1): https://sciencetostartup.com/paper/causal-and-compositional-abstraction - Develop a new formalization for causal abstraction using category theory, applicable to scientific, AI, and quantum fields.
- Who can we trust? LLM-as-a-jury for Comparative Assessment (viability: 6): https://sciencetostartup.com/paper/who-can-we-trust-llm-as-a-jury-for-comparative-assessment - Developing a more reliable AI jury system for evaluating natural language generation using a novel judge calibration method.
- Explainable AI: Context-Aware Layer-Wise Integrated Gradients for Explaining Transformer Models (viability: 3): https://sciencetostartup.com/paper/explainable-ai-context-aware-layer-wise-integrated-gradients-for-explaining-transformer-models - A framework for providing context-aware explanations of Transformer models' decisions.
- FlowPrefill: Decoupling Preemption from Prefill Scheduling Granularity to Mitigate Head-of-Line Blocking in LLM Serving (viability: 2): https://sciencetostartup.com/paper/flowprefill-decoupling-preemption-from-prefill-scheduling-granularity-to-mitigate-head-of-line-blocking-in-llm-serving - FlowPrefill optimizes LLM serving systems to improve throughput while maintaining service level objectives by innovatively managing request scheduling and execution interruption.
- A Contrastive Learning Framework Empowered by Attention-based Feature Adaptation for Street-View Image Classification (viability: 8): https://sciencetostartup.com/paper/a-contrastive-learning-framework-empowered-by-attention-based-feature-adaptation-for-street-view-image-classification - "CLIP-MHAdapter offers efficient and accurate street-view image classification by leveraging an adaptive contrastive learning framework with attention-based feature refinement."
- DataJoint 2.0: A Computational Substrate for Agentic Scientific Workflows (viability: 3): https://sciencetostartup.com/paper/datajoint-2-0-a-computational-substrate-for-agentic-scientific-workflows - DataJoint 2.0 provides a unified relational workflow model for SciOps in scientific data pipelines.
- AIFL: A Global Daily Streamflow Forecasting Model Using Deterministic LSTM Pre-trained on ERA5-Land and Fine-tuned on IFS (viability: 6): https://sciencetostartup.com/paper/aifl-a-global-daily-streamflow-forecasting-model-using-deterministic-lstm-pre-trained-on-era5-land-and-fine-tuned-on-ifs - AIFL offers reliable global streamflow forecasts using an LSTM model pre-trained on ERA5-Land and fine-tuned on IFS data.
- Creating a digital poet (viability: 5): https://sciencetostartup.com/paper/creating-a-digital-poet - Develop a digital poet through iterative workshop-style prompting for creative AI content generation.
- Recursive language models for jailbreak detection: a procedural defense for tool-augmented agents (viability: 3): https://sciencetostartup.com/paper/recursive-language-models-for-jailbreak-detection-a-procedural-defense-for-tool-augmented-agents - Developing a robust framework for detecting jailbreak attacks on language models using Recursive Language Models.
- Framework of Thoughts: A Foundation Framework for Dynamic and Optimized Reasoning based on Chains, Trees, and Graphs (viability: 7): https://sciencetostartup.com/paper/framework-of-thoughts-a-foundation-framework-for-dynamic-and-optimized-reasoning-based-on-chains-trees-and-graphs - Framework of Thoughts optimizes dynamic reasoning structures in AI, enhancing cost efficiency and adaptability.
- Fast and Scalable Analytical Diffusion (viability: 6): https://sciencetostartup.com/paper/fast-and-scalable-analytical-diffusion - GoldDiff offers a scalable, training-free framework for large-scale generative modeling by dynamically pinpointing a 'Golden Subset' for efficient inference.
- From Growing to Looping: A Unified View of Iterative Computation in LLMs (viability: 3): https://sciencetostartup.com/paper/from-growing-to-looping-a-unified-view-of-iterative-computation-in-llms - Develop a tool that enhances reasoning in LLMs through composable depth growth and looping techniques.
- Leveraging Large Language Models for Causal Discovery: a Constraint-based, Argumentation-driven Approach (viability: 5): https://sciencetostartup.com/paper/leveraging-large-language-models-for-causal-discovery-a-constraint-based-argumentation-driven-approach - Use LLMs for enhanced causal discovery by integrating symbolic reasoning and data-driven insights.
- IndicEval: A Bilingual Indian Educational Evaluation Framework for Large Language Models (viability: 5): https://sciencetostartup.com/paper/indiceval-a-bilingual-indian-educational-evaluation-framework-for-large-language-models - IndicEval offers a bilingual evaluation framework for testing large language models on real-world Indian educational exams.
- GICDM: Mitigating Hubness for Reliable Distance-Based Generative Model Evaluation (viability: 5): https://sciencetostartup.com/paper/gicdm-mitigating-hubness-for-reliable-distance-based-generative-model-evaluation - Develop a tool to mitigate hubness in generative model evaluation for more reliable distance-based metrics.
- RoboGene: Boosting VLA Pre-training via Diversity-Driven Agentic Framework for Real-World Task Generation (viability: 8): https://sciencetostartup.com/paper/robogene-boosting-vla-pre-training-via-diversity-driven-agentic-framework-for-real-world-task-generation - Automate diverse and feasible robotic task generation with RoboGene for enhanced model pre-training and real-world application.
- Hardware-accelerated graph neural networks: an alternative approach for neuromorphic event-based audio classification and keyword spotting on SoC FPGA (viability: 7): https://sciencetostartup.com/paper/hardware-accelerated-graph-neural-networks-an-alternative-approach-for-neuromorphic-event-based-audio-classification-and - Develop a low-power, low-latency hardware-accelerated event-graph neural network for real-time audio classification and keyword spotting on FPGA.
- Designing Production-Scale OCR for India: Multilingual and Domain-Specific Systems (viability: 8): https://sciencetostartup.com/paper/designing-production-scale-ocr-for-india-multilingual-and-domain-specific-systems - Multilingual, domain-specific OCR system for India's diverse documents with state-of-the-art results.
- Verifiable Semantics for Agent-to-Agent Communication (viability: 2): https://sciencetostartup.com/paper/verifiable-semantics-for-agent-to-agent-communication - Develop a protocol for verifying shared semantics in multiagent AI systems to reduce communication disagreement.
- HAWX: A Hardware-Aware FrameWork for Fast and Scalable ApproXimation of DNNs (viability: 5): https://sciencetostartup.com/paper/hawx-a-hardware-aware-framework-for-fast-and-scalable-approximation-of-dnns - HAWX optimizes DNN scalability and precision by integrating hardware-awareness into model approximation processes.
- Spatial Audio Question Answering and Reasoning on Dynamic Source Movements (viability: 3): https://sciencetostartup.com/paper/spatial-audio-question-answering-and-reasoning-on-dynamic-source-movements - Develop a spatial audio question answering system focusing on dynamic source movements.
- A Self-Supervised Approach for Enhanced Feature Representations in Object Detection Tasks (viability: 5): https://sciencetostartup.com/paper/a-self-supervised-approach-for-enhanced-feature-representations-in-object-detection-tasks - Develop a self-supervised feature extractor that enhances object detection with less labeled data, outperforming existing pre-trained models.
- Generative AI Usage of University Students: Navigating Between Education and Business (viability: 2): https://sciencetostartup.com/paper/generative-ai-usage-of-university-students-navigating-between-education-and-business - Grounded theory study on how part-time students use generative AI across education and business contexts.
- Long-Tail Knowledge in Large Language Models: Taxonomy, Mechanisms, Interventions and Implications (viability: 3): https://sciencetostartup.com/paper/long-tail-knowledge-in-large-language-models-taxonomy-mechanisms-interventions-and-implications - Develops a framework to analyze long-tail knowledge challenges in language models and their implications.
- Temporal Panel Selection in Ongoing Citizens' Assemblies (viability: 2): https://sciencetostartup.com/paper/temporal-panel-selection-in-ongoing-citizens-assemblies - Develop and implement algorithms for proportional representation in citizens' assemblies to ensure equitable participation over time.
- Revolutionizing Long-Term Memory in AI: New Horizons with High-Capacity and High-Speed Storage (viability: 4): https://sciencetostartup.com/paper/revolutionizing-long-term-memory-in-ai-new-horizons-with-high-capacity-and-high-speed-storage - Develop enhanced long-term storage solutions for AI that preserve and flexibly apply raw experiences.
- SIT-LMPC: Safe Information-Theoretic Learning Model Predictive Control for Iterative Tasks (viability: 5): https://sciencetostartup.com/paper/sit-lmpc-safe-information-theoretic-learning-model-predictive-control-for-iterative-tasks - Develop a control framework for robots performing iterative tasks using safe information-theoretic learning model predictive control.
- Learning Personalized Agents from Human Feedback (viability: 9): https://sciencetostartup.com/paper/learning-personalized-agents-from-human-feedback - A new AI framework that dynamically personalizes agents to user preferences via live feedback, enhancing user interaction quality.
- HiPER: Hierarchical Reinforcement Learning with Explicit Credit Assignment for Large Language Model Agents (viability: 6): https://sciencetostartup.com/paper/hiper-hierarchical-reinforcement-learning-with-explicit-credit-assignment-for-large-language-model-agents - HiPER enhances LLM agent efficiency using hierarchical RL for better action sequence credit assignment, achieving SOTA results.
- ASPEN: Spectral-Temporal Fusion for Cross-Subject Brain Decoding (viability: 5): https://sciencetostartup.com/paper/aspen-spectral-temporal-fusion-for-cross-subject-brain-decoding - ASPEN enhances cross-subject generalization in brain-computer interfaces through spectral-temporal fusion architecture.
- Rethinking ANN-based Retrieval: Multifaceted Learnable Index for Large-scale Recommendation System (viability: 9): https://sciencetostartup.com/paper/rethinking-ann-based-retrieval-multifaceted-learnable-index-for-large-scale-recommendation-system - A real-time recommendation framework that replaces ANN search with a learnable multifaceted index for better efficiency and relevance.
- Federated Graph AGI for Cross-Border Insider Threat Intelligence in Government Financial Schemes (viability: 7): https://sciencetostartup.com/paper/federated-graph-agi-for-cross-border-insider-threat-intelligence-in-government-financial-schemes - FedGraph-AGI enhances insider threat detection with federated, privacy-preserving graph learning using AGI insights.
- Updating Parametric Knowledge with Context Distillation Retains Post-Training Capabilities (viability: 2): https://sciencetostartup.com/paper/updating-parametric-knowledge-with-context-distillation-retains-post-training-capabilities - A new method for continual knowledge adaptation in LLMs that balances learning and retention without explicit generation steps.
- Improving Interactive In-Context Learning from Natural Language Feedback (viability: 6): https://sciencetostartup.com/paper/improving-interactive-in-context-learning-from-natural-language-feedback - Develop a framework for interactive in-context learning to enhance model adaptability using natural language feedback.
- AI-CARE: Carbon-Aware Reporting Evaluation Metric for AI Models (viability: 9): https://sciencetostartup.com/paper/ai-care-carbon-aware-reporting-evaluation-metric-for-ai-models - AI-CARE revolutionizes AI evaluation by providing a carbon-aware metric that empowers sustainable model deployment decisions.
- How Uncertain Is the Grade? A Benchmark of Uncertainty Metrics for LLM-Based Automatic Assessment (viability: 5): https://sciencetostartup.com/paper/how-uncertain-is-the-grade-a-benchmark-of-uncertainty-metrics-for-llm-based-automatic-assessment - Develop actionable insights for uncertainty-aware grading systems using LLMs in educational assessments.
- Transforming GenAI Policy to Prompting Instruction: An RCT of Scalable Prompting Interventions in a CS1 Course (viability: 3): https://sciencetostartup.com/paper/transforming-genai-policy-to-prompting-instruction-an-rct-of-scalable-prompting-interventions-in-a-cs1-course - Scalable instructional interventions to improve AI prompting skills in students through an RCT framework.
- ODYN: An All-Shifted Non-Interior-Point Method for Quadratic Programming in Robotics and AI (viability: 7): https://sciencetostartup.com/paper/odyn-an-all-shifted-non-interior-point-method-for-quadratic-programming-in-robotics-and-ai - ODYN provides a state-of-the-art, open-source QP solver for real-time optimization in robotics and AI, with superior warm-starting capabilities.
- Anatomy of Capability Emergence: Scale-Invariant Representation Collapse and Top-Down Reorganization in Neural Networks (viability: 3): https://sciencetostartup.com/paper/anatomy-of-capability-emergence-scale-invariant-representation-collapse-and-top-down-reorganization-in-neural-networks - Exploring the geometric framework and hierarchical representation dynamics in neural network emergence, not aimed for product prediction tools.
- ReLoop: Structured Modeling and Behavioral Verification for Reliable LLM-Based Optimization (viability: 7): https://sciencetostartup.com/paper/reloop-structured-modeling-and-behavioral-verification-for-reliable-llm-based-optimization - ReLoop improves LLM-based code correctness with structured modeling and behavioral verification for optimization tasks.
- From Reflection to Repair: A Scoping Review of Dataset Documentation Tools (viability: 3): https://sciencetostartup.com/paper/from-reflection-to-repair-a-scoping-review-of-dataset-documentation-tools - A systematic review for enhancing dataset documentation tools and practices.
- DocSplit: A Comprehensive Benchmark Dataset and Evaluation Approach for Document Packet Recognition and Splitting (viability: 5): https://sciencetostartup.com/paper/docsplit-a-comprehensive-benchmark-dataset-and-evaluation-approach-for-document-packet-recognition-and-splitting - DocSplit provides a benchmark dataset and evaluation metrics for improving automated document packet splitting, essential for document-intensive industries.
- Perceptive Humanoid Parkour: Chaining Dynamic Human Skills via Motion Matching (viability: 7): https://sciencetostartup.com/paper/perceptive-humanoid-parkour-chaining-dynamic-human-skills-via-motion-matching - Enable humanoid robots to autonomously perform agile parkour using motion matching and onboard perception.
- CrispEdit: Low-Curvature Projections for Scalable Non-Destructive LLM Editing (viability: 7): https://sciencetostartup.com/paper/crispedit-low-curvature-projections-for-scalable-non-destructive-llm-editing - CrispEdit offers scalable non-destructive editing for LLMs with minimal capability degradation.
- Developing AI Agents with Simulated Data: Why, what, and how? (viability: 2): https://sciencetostartup.com/paper/developing-ai-agents-with-simulated-data-why-what-and-how - Framework for generating synthetic data through simulation to enhance AI training.
- Avey-B (viability: 6): https://sciencetostartup.com/paper/avey-b - Avey-B: Efficient Bidirectional NLP Encoder surpassing traditional Transformer models in token classification and information retrieval tasks.
- Task-Agnostic Continual Learning for Chest Radiograph Classification (viability: 5): https://sciencetostartup.com/paper/task-agnostic-continual-learning-for-chest-radiograph-classification - Introducing CARL-XRay: an efficient continual learning framework optimizing chest radiograph classification for sequential clinical deployment.
- The Geometry of Alignment Collapse: When Fine-Tuning Breaks Safety (viability: 1): https://sciencetostartup.com/paper/the-geometry-of-alignment-collapse-when-fine-tuning-breaks-safety - Analyzes the geometry of alignment degradation in fine-tuned models to improve predictive safety diagnostics.
- Enhancing Building Semantics Preservation in AI Model Training with Large Language Model Encodings (viability: 6): https://sciencetostartup.com/paper/enhancing-building-semantics-preservation-in-ai-model-training-with-large-language-model-encodings - Enhance AI models in AECO industry using LLM embeddings for finer building semantics distinction.
- This human study did not involve human subjects: Validating LLM simulations as behavioral evidence (viability: 3): https://sciencetostartup.com/paper/this-human-study-did-not-involve-human-subjects-validating-llm-simulations-as-behavioral-evidence - Develop a tool for integrating LLMs in social science experiments through statistical calibration for cost-effective behavioral simulation.
- GlobeDiff: State Diffusion Process for Partial Observability in Multi-Agent Systems (viability: 4): https://sciencetostartup.com/paper/globediff-state-diffusion-process-for-partial-observability-in-multi-agent-systems - Develop a multi-agent coordination tool using GlobeDiff for inferring global states from local observations.
- Understanding vs. Generation: Navigating Optimization Dilemma in Multimodal Models (viability: 6): https://sciencetostartup.com/paper/understanding-vs-generation-navigating-optimization-dilemma-in-multimodal-models - The Reason-Reflect-Refine (R3) framework improves multimodal model performance by integrating understanding into the generative process.
- Robot-Assisted Social Dining as a White Glove Service (viability: 2): https://sciencetostartup.com/paper/robot-assisted-social-dining-as-a-white-glove-service - Developing a robot-assisted dining experience to enable social dining for people with disabilities.
- ChartEditBench: Evaluating Grounded Multi-Turn Chart Editing in Multimodal Language Models (viability: 4): https://sciencetostartup.com/paper/charteditbench-evaluating-grounded-multi-turn-chart-editing-in-multimodal-language-models - ChartEditBench enhances multi-turn chart editing capabilities in multimodal language models with a comprehensive benchmark and evaluation framework.
- Beyond Binary Classification: Detecting Fine-Grained Sexism in Social Media Videos (viability: 4): https://sciencetostartup.com/paper/beyond-binary-classification-detecting-fine-grained-sexism-in-social-media-videos - Develop a fine-grained sexism detection tool for social media videos using a new Spanish multimodal dataset.
- UrbanVerse: Learning Urban Region Representation Across Cities and Tasks (viability: 6): https://sciencetostartup.com/paper/urbanverse-learning-urban-region-representation-across-cities-and-tasks - UrbanVerse offers a versatile urban region representation model enhancing cross-city and cross-task urban analytics.
- MRC-GAT: A Meta-Relational Copula-Based Graph Attention Network for Interpretable Multimodal Alzheimer's Disease Diagnosis (viability: 7): https://sciencetostartup.com/paper/mrc-gat-a-meta-relational-copula-based-graph-attention-network-for-interpretable-multimodal-alzheimer-s-disease-diagnosi - Develop an AI-driven tool for interpretable Alzheimer’s disease diagnosis using multimodal data.
- MeshMimic: Geometry-Aware Humanoid Motion Learning through 3D Scene Reconstruction (viability: 6): https://sciencetostartup.com/paper/meshmimic-geometry-aware-humanoid-motion-learning-through-3d-scene-reconstruction - MeshMimic transforms ordinary video into humanoid robot motion training systems by reconstructing dynamic terrains and interactions.
- Spanning the Visual Analogy Space with a Weight Basis of LoRAs (viability: 7): https://sciencetostartup.com/paper/spanning-the-visual-analogy-space-with-a-weight-basis-of-loras - LoRWeB allows users to perform complex image transformations through a dynamic composition of LoRA modules, enhancing visual analogy creation.
- Recursive Concept Evolution for Compositional Reasoning in Large Language Models (viability: 5): https://sciencetostartup.com/paper/recursive-concept-evolution-for-compositional-reasoning-in-large-language-models - Recursive Concept Evolution enhances language models with dynamic abstraction capabilities for complex reasoning tasks.
- Learning to Retrieve Navigable Candidates for Efficient Vision-and-Language Navigation (viability: 7): https://sciencetostartup.com/paper/learning-to-retrieve-navigable-candidates-for-efficient-vision-and-language-navigation - Develop a retrieval-augmented framework to enhance LLM-based vision-and-language navigation efficiency and stability.
- Lifelong Scalable Multi-Agent Realistic Testbed and A Comprehensive Study on Design Choices in Lifelong AGV Fleet Management Systems (viability: 7): https://sciencetostartup.com/paper/lifelong-scalable-multi-agent-realistic-testbed-and-a-comprehensive-study-on-design-choices-in-lifelong-agv-fleet-manage - Develop a simulator for evaluating multi-agent path finding algorithms in AGV fleet management systems.
- Bayesian Optimization for Design Parameters of 3D Image Data Analysis (viability: 8): https://sciencetostartup.com/paper/bayesian-optimization-for-design-parameters-of-3d-image-data-analysis - Optimize and automate 3D biomedical image analysis using Bayesian Optimization.
- Zombie Agents: Persistent Control of Self-Evolving LLM Agents via Self-Reinforcing Injections (viability: 1): https://sciencetostartup.com/paper/zombie-agents-persistent-control-of-self-evolving-llm-agents-via-self-reinforcing-injections - Exploring persistent security risks in self-evolving LLM agents with potential covert memory exploits.
- On inferring cumulative constraints (viability: 5): https://sciencetostartup.com/paper/on-inferring-cumulative-constraints - Introducing a preprocessing method to improve scheduling efficiency by inferring cumulative constraints, enhancing constraint programming performance.
- STAPO: Stabilizing Reinforcement Learning for LLMs by Silencing Rare Spurious Tokens (viability: 5): https://sciencetostartup.com/paper/stapo-stabilizing-reinforcement-learning-for-llms-by-silencing-rare-spurious-tokens - Develop a reinforcement learning stabilizer for large language models that reduces training instability by ignoring spurious token influences.
- Intracoronary Optical Coherence Tomography Image Processing and Vessel Classification Using Machine Learning (viability: 6): https://sciencetostartup.com/paper/intracoronary-optical-coherence-tomography-image-processing-and-vessel-classification-using-machine-learning - Automated intracoronary OCT processing for real-time vessel identification and classification to aid clinicians.
- Beyond Static Pipelines: Learning Dynamic Workflows for Text-to-SQL (viability: 6): https://sciencetostartup.com/paper/beyond-static-pipelines-learning-dynamic-workflows-for-text-to-sql - SquRL enhances LLMs' text-to-SQL capabilities by dynamically constructing workflows through reinforcement learning.
- RUVA: Personalized Transparent On-Device Graph Reasoning (viability: 9): https://sciencetostartup.com/paper/ruva-personalized-transparent-on-device-graph-reasoning - RUVA offers on-device, transparent, and editable personal AI knowledge management, ensuring user privacy and control.
- VLM-DEWM: Dynamic External World Model for Verifiable and Resilient Vision-Language Planning in Manufacturing (viability: 3): https://sciencetostartup.com/paper/vlm-dewm-dynamic-external-world-model-for-verifiable-and-resilient-vision-language-planning-in-manufacturing - Develop a cognitive architecture for resilient robotic planning in dynamic manufacturing using a decoupled vision-language model with improved state tracking and recovery.
- Dynamic Training-Free Fusion of Subject and Style LoRAs (viability: 5): https://sciencetostartup.com/paper/dynamic-training-free-fusion-of-subject-and-style-loras - Dynamic, training-free fusion of Subject and Style LoRAs for superior creative synthesis.
- Quantifying construct validity in large language model evaluations (viability: 4): https://sciencetostartup.com/paper/quantifying-construct-validity-in-large-language-model-evaluations - Develop a model for improving the measurement of construct validity in LLM benchmarks using structured capabilities.
- GenAI-LA: Generative AI and Learning Analytics Workshop (LAK 2026), April 27--May 1, 2026, Bergen, Norway (viability: 6): https://sciencetostartup.com/paper/genai-la-generative-ai-and-learning-analytics-workshop-lak-2026-april-27-may-1-2026-bergen-norway - Build an AI tool for educational risk detection using a curated dataset of pedagogical explanations.
- The Obfuscation Atlas: Mapping Where Honesty Emerges in RLVR with Deception Probes (viability: 2): https://sciencetostartup.com/paper/the-obfuscation-atlas-mapping-where-honesty-emerges-in-rlvr-with-deception-probes - A research initiative investigating AI obfuscation in realistic coding environments to understand deception detection efficacy.
- Improving MLLMs in Embodied Exploration and Question Answering with Human-Inspired Memory Modeling (viability: 7): https://sciencetostartup.com/paper/improving-mllms-in-embodied-exploration-and-question-answering-with-human-inspired-memory-modeling - Enhance embodied agents with a novel memory framework improving exploration and reasoning efficiency.
- The Equalizer: Introducing Shape-Gain Decomposition in Neural Audio Codecs (viability: 2): https://sciencetostartup.com/paper/the-equalizer-introducing-shape-gain-decomposition-in-neural-audio-codecs - Introducing a shape-gain decomposition method in neural audio codecs for improved bitrate-distortion performance.
- RPT-SR: Regional Prior attention Transformer for infrared image Super-Resolution (viability: 5): https://sciencetostartup.com/paper/rpt-sr-regional-prior-attention-transformer-for-infrared-image-super-resolution - RPT-SR enhances infrared image super-resolution by leveraging spatial priors for improved performance in fixed-view scenarios.
- SecCodeBench-V2 Technical Report (viability: 5): https://sciencetostartup.com/paper/seccodebench-v2-technical-report - Develop a benchmark platform to evaluate the security capabilities of LLM-based code generation tools.
- Algorithmic Approaches to Opinion Selection for Online Deliberation: A Comparative Study (viability: 5): https://sciencetostartup.com/paper/algorithmic-approaches-to-opinion-selection-for-online-deliberation-a-comparative-study - Develop algorithmic tools for selecting representative opinions in online deliberation, balancing diversity and proportional representation.
- Logit Distance Bounds Representational Similarity (viability: 4): https://sciencetostartup.com/paper/logit-distance-bounds-representational-similarity - Enhance model distillation by preserving linear representational similarity using logit-distance metrics.
- Common Belief Revisited (viability: 1): https://sciencetostartup.com/paper/common-belief-revisited - Exploring the logic properties of common belief in the KD45 logical framework.
- ActionCodec: What Makes for Good Action Tokenizers (viability: 7): https://sciencetostartup.com/paper/actioncodec-what-makes-for-good-action-tokenizers - ActionCodec is a high-performance action tokenizer that significantly enhances VLA models' training efficiency and performance, setting new benchmarks for robotics tasks without pre-training.
- World-Model-Augmented Web Agents with Action Correction (viability: 6): https://sciencetostartup.com/paper/world-model-augmented-web-agents-with-action-correction - Develop risk-aware web agents with enhanced task execution through model collaboration and action correction.
- Orchestration-Free Customer Service Automation: A Privacy-Preserving and Flowchart-Guided Framework (viability: 7): https://sciencetostartup.com/paper/orchestration-free-customer-service-automation-a-privacy-preserving-and-flowchart-guided-framework - Enable privacy-preserving customer service automation with an orchestration-free framework using Task-Oriented Flowcharts (TOFs).
- CDRL: A Reinforcement Learning Framework Inspired by Cerebellar Circuits and Dendritic Computational Strategies (viability: 5): https://sciencetostartup.com/paper/cdrl-a-reinforcement-learning-framework-inspired-by-cerebellar-circuits-and-dendritic-computational-strategies - Develop a reinforcement learning framework inspired by cerebellar structures to improve sample efficiency and robustness in RL tasks.
- Automated Multi-Source Debugging and Natural Language Error Explanation for Dashboard Applications (viability: 5): https://sciencetostartup.com/paper/automated-multi-source-debugging-and-natural-language-error-explanation-for-dashboard-applications - Transform cryptic dashboard error messages into actionable insights using automated multi-source debugging and AI-driven explanations.
- NeuroSymActive: Differentiable Neural-Symbolic Reasoning with Active Exploration for Knowledge Graph Question Answering (viability: 5): https://sciencetostartup.com/paper/neurosymactive-differentiable-neural-symbolic-reasoning-with-active-exploration-for-knowledge-graph-question-answering - A framework integrating neural-symbolic reasoning for efficient knowledge graph question answering.
- FedPSA: Modeling Behavioral Staleness in Asynchronous Federated Learning (viability: 7): https://sciencetostartup.com/paper/fedpsa-modeling-behavioral-staleness-in-asynchronous-federated-learning - FedPSA enhances Federated Learning by dynamically adjusting to model obsolescence, significantly improving asynchronous training efficiency.
- A Scalable Curiosity-Driven Game-Theoretic Framework for Long-Tail Multi-Label Learning in Data Mining (viability: 9): https://sciencetostartup.com/paper/a-scalable-curiosity-driven-game-theoretic-framework-for-long-tail-multi-label-learning-in-data-mining - A scalable, curiosity-driven game-theoretic framework to enhance multi-label classification for imbalanced datasets in real-world applications like e-commerce and healthcare.
- Prescriptive Scaling Reveals the Evolution of Language Model Capabilities (viability: 5): https://sciencetostartup.com/paper/prescriptive-scaling-reveals-the-evolution-of-language-model-capabilities - Develop a tool to predict language model performance from compute budgets using Proteus 2k dataset.
- AgriWorld:A World Tools Protocol Framework for Verifiable Agricultural Reasoning with Code-Executing LLM Agents (viability: 8): https://sciencetostartup.com/paper/agriworld-a-world-tools-protocol-framework-for-verifiable-agricultural-reasoning-with-code-executing-llm-agents - AgriWorld: An agentic framework enabling LLMs to execute precise agricultural queries via a Python-based toolset.
- On Surprising Effectiveness of Masking Updates in Adaptive Optimizers (viability: 3): https://sciencetostartup.com/paper/on-surprising-effectiveness-of-masking-updates-in-adaptive-optimizers - Develop Magma, a Momentum-aligned gradient masking optimizer to enhance LLM training efficiency.
- EAA: Automating materials characterization with vision language model agents (viability: 4): https://sciencetostartup.com/paper/eaa-automating-materials-characterization-with-vision-language-model-agents - Automate complex microscopy workflows with vision-language AI agents for more efficient synchrotron beamline operations.
- The Information Geometry of Softmax: Probing and Steering (viability: 5): https://sciencetostartup.com/paper/the-information-geometry-of-softmax-probing-and-steering - Develop enhanced concept manipulation in AI models using Dual Steering with information geometry.
- Complex-Valued Unitary Representations as Classification Heads for Improved Uncertainty Quantification in Deep Neural Networks (viability: 5): https://sciencetostartup.com/paper/complex-valued-unitary-representations-as-classification-heads-for-improved-uncertainty-quantification-in-deep-neural-ne - Improve neural network uncertainty quantification using quantum-inspired complex-valued unitary representations.
- High-Fidelity Network Management for Federated AI-as-a-Service: Cross-Domain Orchestration (viability: 3): https://sciencetostartup.com/paper/high-fidelity-network-management-for-federated-ai-as-a-service-cross-domain-orchestration - High-fidelity AI-as-a-Service network management with cross-domain orchestration for improved AI model delivery.
- Visual Persuasion: What Influences Decisions of Vision-Language Models? (viability: 6): https://sciencetostartup.com/paper/visual-persuasion-what-influences-decisions-of-vision-language-models - A tool for optimizing and interpreting visual prompts to influence decisions in Vision-Language Models for safer AI applications.
- Accelerating Large-Scale Dataset Distillation via Exploration-Exploitation Optimization (viability: 7): https://sciencetostartup.com/paper/accelerating-large-scale-dataset-distillation-via-exploration-exploitation-optimization - E^2D accelerates large-scale dataset distillation by optimizing redundant computation, enhancing efficiency, and improving accuracy over state-of-the-art methods.
- When Remembering and Planning are Worth it: Navigating under Change (viability: 3): https://sciencetostartup.com/paper/when-remembering-and-planning-are-worth-it-navigating-under-change - Develop adaptive memory-based AI agents for efficient navigation in dynamic environments.
- Enhancing Diversity and Feasibility: Joint Population Synthesis from Multi-source Data Using Generative Models (viability: 7): https://sciencetostartup.com/paper/enhancing-diversity-and-feasibility-joint-population-synthesis-from-multi-source-data-using-generative-models - A generative model leveraging multi-source data to create diverse, feasible synthetic populations for agent-based modeling in urban planning.
- From Diagnosis to Inoculation: Building Cognitive Resistance to AI Disempowerment (viability: 5): https://sciencetostartup.com/paper/from-diagnosis-to-inoculation-building-cognitive-resistance-to-ai-disempowerment - Develop an AI literacy framework to build cognitive resistance to AI disempowerment among users.
- Fast and Effective On-policy Distillation from Reasoning Prefixes (viability: 5): https://sciencetostartup.com/paper/fast-and-effective-on-policy-distillation-from-reasoning-prefixes - Enhance AI model training efficiency by focusing on distillation of reasoning prefixes to halve computational requirements.
- How to Train Your Long-Context Visual Document Model (viability: 7): https://sciencetostartup.com/paper/how-to-train-your-long-context-visual-document-model - Build a high-performance API for visual document question answering with long-context capabilities.
- Decision Making under Imperfect Recall: Algorithms and Benchmarks (viability: 7): https://sciencetostartup.com/paper/decision-making-under-imperfect-recall-algorithms-and-benchmarks - New benchmark suite and algorithms for optimizing imperfect-recall decision problems, outperforming traditional approaches in efficiency.
- Predicting Invoice Dilution in Supply Chain Finance with Leakage Free Two Stage XGBoost, KAN (Kolmogorov Arnold Networks), and Ensemble Models (viability: 6): https://sciencetostartup.com/paper/predicting-invoice-dilution-in-supply-chain-finance-with-leakage-free-two-stage-xgboost-kan-kolmogorov-arnold-networks-a - AI-driven framework to predict invoice dilution, optimizing non-credit risk management in supply chain finance.
- MyoInteract: A Framework for Fast Prototyping of Biomechanical HCI Tasks using Reinforcement Learning (viability: 6): https://sciencetostartup.com/paper/myointeract-a-framework-for-fast-prototyping-of-biomechanical-hci-tasks-using-reinforcement-learning - MyoInteract accelerates biomechanical HCI task prototyping, making RL accessible to interaction designers.
- Closing the Distribution Gap in Adversarial Training for LLMs (viability: 2): https://sciencetostartup.com/paper/closing-the-distribution-gap-in-adversarial-training-for-llms - Enhance LLM robustness against adversaries using Distributional Adversarial Training with diffusion models.
- Secure and Energy-Efficient Wireless Agentic AI Networks (viability: 7): https://sciencetostartup.com/paper/secure-and-energy-efficient-wireless-agentic-ai-networks - Develop a secure and energy-efficient wireless agentic AI network system optimized for dynamic resource allocation in reasoning tasks.
- Colosseum: Auditing Collusion in Cooperative Multi-Agent Systems (viability: 3): https://sciencetostartup.com/paper/colosseum-auditing-collusion-in-cooperative-multi-agent-systems - Colosseum audits collusive behavior in multi-agent systems using a novel measurement framework for agent communication and actions.
- OpaqueToolsBench: Learning Nuances of Tool Behavior Through Interaction (viability: 4): https://sciencetostartup.com/paper/opaquetoolsbench-learning-nuances-of-tool-behavior-through-interaction - Develop a framework to improve tool documentation and usage in LLM environments with opaque tools.
- Weight space Detection of Backdoors in LoRA Adapters (viability: 6): https://sciencetostartup.com/paper/weight-space-detection-of-backdoors-in-lora-adapters - Develop a tool for detecting poisoned LoRA adapters using weight matrix analysis without test data dependency.
- ScrapeGraphAI-100k: A Large-Scale Dataset for LLM-Based Web Information Extraction (viability: 6): https://sciencetostartup.com/paper/scrapegraphai-100k-a-large-scale-dataset-for-llm-based-web-information-extraction - ScrapeGraphAI-100k offers a comprehensive dataset to enhance LLM-based web information extraction, facilitating fine-tuning and benchmarking for efficient web data retrieval.
- Mind the (DH) Gap! A Contrast in Risky Choices Between Reasoning and Conversational LLMs (viability: 2): https://sciencetostartup.com/paper/mind-the-dh-gap-a-contrast-in-risky-choices-between-reasoning-and-conversational-llms - Study contrasts reasoning and conversational LLM behaviors under uncertainty.
- Exploiting Layer-Specific Vulnerabilities to Backdoor Attack in Federated Learning (viability: 4): https://sciencetostartup.com/paper/exploiting-layer-specific-vulnerabilities-to-backdoor-attack-in-federated-learning - Develop robust defenses in Federated Learning to mitigate layer-specific backdoor attacks like LSA.
- da Costa and Tarski meet Goguen and Carnap: a novel approach for ontological heterogeneity based on consequence systems (viability: 1): https://sciencetostartup.com/paper/da-costa-and-tarski-meet-goguen-and-carnap-a-novel-approach-for-ontological-heterogeneity-based-on-consequence-systems - Novel ontological heterogeneity approach using extended consequence systems for applied ontology.
- Panini: Continual Learning in Token Space via Structured Memory (viability: 7): https://sciencetostartup.com/paper/panini-continual-learning-in-token-space-via-structured-memory - A semantic memory tool enabling efficient and accurate continual learning for language models.
- Protecting Language Models Against Unauthorized Distillation through Trace Rewriting (viability: 3): https://sciencetostartup.com/paper/protecting-language-models-against-unauthorized-distillation-through-trace-rewriting - A method to modify LLM outputs to prevent unauthorized knowledge distillation and embed verifiable watermarks in student models.
- CGRA-DeBERTa Concept Guided Residual Augmentation Transformer for Theologically Islamic Understanding (viability: 6): https://sciencetostartup.com/paper/cgra-deberta-concept-guided-residual-augmentation-transformer-for-theologically-islamic-understanding - Develop a high-performing QA system specialized for theological texts using a custom transformer model.
- MB-DSMIL-CL-PL: Scalable Weakly Supervised Ovarian Cancer Subtype Classification and Localisation Using Contrastive and Prototype Learning with Frozen Patch Features (viability: 3): https://sciencetostartup.com/paper/mb-dsmil-cl-pl-scalable-weakly-supervised-ovarian-cancer-subtype-classification-and-localisation-using-contrastive-and-p - AI-based ovarian cancer subtype classification using contrastive and prototype learning with frozen patch features.
- PolyNODE: Variable-dimension Neural ODEs on M-polyfolds (viability: 5): https://sciencetostartup.com/paper/polynode-variable-dimension-neural-odes-on-m-polyfolds - PolyNODEs extend Neural ODEs to variable dimensions using M-polyfolds for enhanced AI modeling.
- ResearchGym: Evaluating Language Model Agents on Real-World AI Research (viability: 6): https://sciencetostartup.com/paper/researchgym-evaluating-language-model-agents-on-real-world-ai-research - ResearchGym provides a standardized evaluation environment for testing AI agents in real-world research tasks.
- Long Context, Less Focus: A Scaling Gap in LLMs Revealed through Privacy and Personalization (viability: 5): https://sciencetostartup.com/paper/long-context-less-focus-a-scaling-gap-in-llms-revealed-through-privacy-and-personalization - A benchmark suite for evaluating the effects of long context lengths on privacy and personalization in LLMs.
- Rethinking Diffusion Models with Symmetries through Canonicalization with Applications to Molecular Graph Generation (viability: 8): https://sciencetostartup.com/paper/rethinking-diffusion-models-with-symmetries-through-canonicalization-with-applications-to-molecular-graph-generation - Introducing a novel canonical diffusion framework for efficient and expressive molecular graph generation.
- Hunt Globally: Deep Research AI Agents for Drug Asset Scouting in Investing, Business Development, and Search & Evaluation (viability: 8): https://sciencetostartup.com/paper/hunt-globally-deep-research-ai-agents-for-drug-asset-scouting-in-investing-business-development-and-search-evaluation - AI agents for comprehensive global drug asset scouting in biopharma investments.
- Cold-Start Personalization via Training-Free Priors from Structured World Models (viability: 3): https://sciencetostartup.com/paper/cold-start-personalization-via-training-free-priors-from-structured-world-models - Develop a cold-start personalization tool using structured world models for effective preference inference with minimal user interaction.
- Spectral Convolution on Orbifolds for Geometric Deep Learning (viability: 2): https://sciencetostartup.com/paper/spectral-convolution-on-orbifolds-for-geometric-deep-learning - Presents spectral convolution on orbifolds as a building block for geometric deep learning applications.
- On the Semantics of Primary Cause in Hybrid Dynamic Domains (viability: 2): https://sciencetostartup.com/paper/on-the-semantics-of-primary-cause-in-hybrid-dynamic-domains - Develops definitions of primary cause within a hybrid dynamic framework for reasoning about causation.
- ThermEval: A Structured Benchmark for Evaluation of Vision-Language Models on Thermal Imagery (viability: 5): https://sciencetostartup.com/paper/thermeval-a-structured-benchmark-for-evaluation-of-vision-language-models-on-thermal-imagery - Develop ThermEval, a benchmark to enhance vision-language model capabilities on thermal images for applications like surveillance and autonomous driving.
- PhyScensis: Physics-Augmented LLM Agents for Complex Physical Scene Arrangement (viability: 3): https://sciencetostartup.com/paper/physcensis-physics-augmented-llm-agents-for-complex-physical-scene-arrangement - A framework using a physics engine and LLMs to generate complex physical 3D scenes for robotic manipulation.
- AnchorWeave: World-Consistent Video Generation with Retrieved Local Spatial Memories (viability: 6): https://sciencetostartup.com/paper/anchorweave-world-consistent-video-generation-with-retrieved-local-spatial-memories - AnchorWeave enhances video generation by using clean local spatial memories for improved long-term consistency.
- MAC-AMP: A Closed-Loop Multi-Agent Collaboration System for Multi-Objective Antimicrobial Peptide Design (viability: 8): https://sciencetostartup.com/paper/mac-amp-a-closed-loop-multi-agent-collaboration-system-for-multi-objective-antimicrobial-peptide-design - Advanced AI-driven system for designing effective and non-toxic antimicrobial peptides against resistant pathogens.
- ReusStdFlow: A Standardized Reusability Framework for Dynamic Workflow Construction in Agentic AI (viability: 3): https://sciencetostartup.com/paper/reusstdflow-a-standardized-reusability-framework-for-dynamic-workflow-construction-in-agentic-ai - ReusStdFlow standardizes and automates workflow reuse and construction in enterprise Agentic AI by deconstructing DSLs into modular segments using a novel framework.
- BHyGNN+: Unsupervised Representation Learning for Heterophilic Hypergraphs (viability: 5): https://sciencetostartup.com/paper/bhygnn-unsupervised-representation-learning-for-heterophilic-hypergraphs - Develop a self-supervised learning framework for heterophilic hypergraphs using hypergraph duality transformation to enhance representation learning.
- BFS-PO: Best-First Search for Large Reasoning Models (viability: 6): https://sciencetostartup.com/paper/bfs-po-best-first-search-for-large-reasoning-models - Optimize Large Reasoning Models with BFS-PO for efficient and concise reasoning chains.
- Position: Introspective Experience from Conversational Environments as a Path to Better Learning (viability: 2): https://sciencetostartup.com/paper/position-introspective-experience-from-conversational-environments-as-a-path-to-better-learning - Leverage introspective experiences in conversational environments to enhance AI learning.
- The Potential of CoT for Reasoning: A Closer Look at Trace Dynamics (viability: 2): https://sciencetostartup.com/paper/the-potential-of-cot-for-reasoning-a-closer-look-at-trace-dynamics - In-depth analysis of Chain-of-Thought dynamics in LLM reasoning to better understand the mechanics behind successful problem-solving.
- Picking the Right Specialist: Attentive Neural Process-based Selection of Task-Specialized Models as Tools for Agentic Healthcare Systems (viability: 7): https://sciencetostartup.com/paper/picking-the-right-specialist-attentive-neural-process-based-selection-of-task-specialized-models-as-tools-for-agentic-he - ToolSelect uses an Attentive Neural Process to optimally select task-specialized models for healthcare queries.
- Lifted Relational Probabilistic Inference via Implicit Learning (viability: 3): https://sciencetostartup.com/paper/lifted-relational-probabilistic-inference-via-implicit-learning - Develops a polynomial-time framework for implicit learning and lifted inference in first-order probabilistic logic without explicit models.
- CT-Bench: A Benchmark for Multimodal Lesion Understanding in Computed Tomography (viability: 5): https://sciencetostartup.com/paper/ct-bench-a-benchmark-for-multimodal-lesion-understanding-in-computed-tomography - CT-Bench provides a comprehensive benchmark dataset for advancing AI-driven lesion analysis in CT scans.
- On the Learning Dynamics of RLVR at the Edge of Competence (viability: 1): https://sciencetostartup.com/paper/on-the-learning-dynamics-of-rlvr-at-the-edge-of-competence - Develop a theoretical framework to understand the training dynamics of RLVR in large reasoning models.
- Concept Influence: Leveraging Interpretability to Improve Performance and Efficiency in Training Data Attribution (viability: 5): https://sciencetostartup.com/paper/concept-influence-leveraging-interpretability-to-improve-performance-and-efficiency-in-training-data-attribution - Leverage Concept Influence to efficiently attribute language model behaviors to training data for improved model control.
- Goldilocks RL: Tuning Task Difficulty to Escape Sparse Rewards for Reasoning (viability: 5): https://sciencetostartup.com/paper/goldilocks-rl-tuning-task-difficulty-to-escape-sparse-rewards-for-reasoning - Dynamic task difficulty adjustment for reinforcement learning optimizes reasoning training efficiency.
- Atomix: Timely, Transactional Tool Use for Reliable Agentic Workflows (viability: 3): https://sciencetostartup.com/paper/atomix-timely-transactional-tool-use-for-reliable-agentic-workflows - Atomix enhances reliability in agent workflows by introducing progress-aware transactional semantics for tool calls in LLM agents.
- Return of the Schema: Building Complete Datasets for Machine Learning and Reasoning on Knowledge Graphs (viability: 5): https://sciencetostartup.com/paper/return-of-the-schema-building-complete-datasets-for-machine-learning-and-reasoning-on-knowledge-graphs - A resource providing a workflow for creating machine learning and reasoning-ready datasets with schema and facts from knowledge graphs.
- What hackers talk about when they talk about AI: Early-stage diffusion of a cybercrime innovation (viability: 3): https://sciencetostartup.com/paper/what-hackers-talk-about-when-they-talk-about-ai-early-stage-diffusion-of-a-cybercrime-innovation - Leverage insights on AI-enabled cybercrime for better cybersecurity measures.
- A Geometric Analysis of Small-sized Language Model Hallucinations (viability: 2): https://sciencetostartup.com/paper/a-geometric-analysis-of-small-sized-language-model-hallucinations - Geometric analysis of small-sized LLM hallucinations to improve response classification efficiency.
- GOT-JEPA: Generic Object Tracking with Model Adaptation and Occlusion Handling using Joint-Embedding Predictive Architecture (viability: 8): https://sciencetostartup.com/paper/got-jepa-generic-object-tracking-with-model-adaptation-and-occlusion-handling-using-joint-embedding-predictive-architect - An AI-powered object tracking framework enhancing visibility estimation for improved dynamic environment adaptation and occlusion handling.
- Unlocking Reasoning Capability on Machine Translation in Large Language Models (viability: 6): https://sciencetostartup.com/paper/unlocking-reasoning-capability-on-machine-translation-in-large-language-models - Develop a structured reasoning framework for enhancing machine translation in large language models.
- Universal Algorithm-Implicit Learning (viability: 5): https://sciencetostartup.com/paper/universal-algorithm-implicit-learning - TAIL is a transformer-based meta-learning model that generalizes across tasks with varying domains, modalities, and label configurations, achieving state-of-the-art performance with significant computational efficiency.
- AI Arms and Influence: Frontier Models Exhibit Sophisticated Reasoning in Simulated Nuclear Crises (viability: 4): https://sciencetostartup.com/paper/ai-arms-and-influence-frontier-models-exhibit-sophisticated-reasoning-in-simulated-nuclear-crises - AI models simulate nuclear crisis scenarios, providing insights into strategic reasoning and its pitfalls in AI systems.
- Scale redundancy and soft gauge fixing in positively homogeneous neural networks (viability: 2): https://sciencetostartup.com/paper/scale-redundancy-and-soft-gauge-fixing-in-positively-homogeneous-neural-networks - Introduce gauge-adapted coordinates to stabilize learning in positively homogeneous neural networks.
- ManeuverNet: A Soft Actor-Critic Framework for Precise Maneuvering of Double-Ackermann-Steering Robots with Optimized Reward Functions (viability: 7): https://sciencetostartup.com/paper/maneuvernet-a-soft-actor-critic-framework-for-precise-maneuvering-of-double-ackermann-steering-robots-with-optimized-rew - Develop an AI-powered system for precise and robust maneuvering of agricultural robots using a novel DRL framework.
- Orcheo: A Modular Full-Stack Platform for Conversational Search (viability: 8): https://sciencetostartup.com/paper/orcheo-a-modular-full-stack-platform-for-conversational-search - Orcheo is an open-source, full-stack platform enabling quick development and deployment of conversational search applications with modular, reusable components.
- Removing Planner Bias in Goal Recognition Through Multi-Plan Dataset Generation (viability: 5): https://sciencetostartup.com/paper/removing-planner-bias-in-goal-recognition-through-multi-plan-dataset-generation - Develop a goal recogniser that leverages multi-plan datasets to improve resilience against planner biases in multi-agent systems.
- Exposing the Systematic Vulnerability of Open-Weight Models to Prefill Attacks (viability: 3): https://sciencetostartup.com/paper/exposing-the-systematic-vulnerability-of-open-weight-models-to-prefill-attacks - A study revealing the systematic vulnerability of open-weight language models to prefill attacks, emphasizing a need for better defenses.
- SynthSAEBench: Evaluating Sparse Autoencoders on Scalable Realistic Synthetic Data (viability: 4): https://sciencetostartup.com/paper/synthsaebench-evaluating-sparse-autoencoders-on-scalable-realistic-synthetic-data - SynthSAEBench offers a benchmark toolkit for evaluating Sparse Autoencoder architectures with realistic synthetic data.
- ST-EVO: Towards Generative Spatio-Temporal Evolution of Multi-Agent Communication Topologies (viability: 7): https://sciencetostartup.com/paper/st-evo-towards-generative-spatio-temporal-evolution-of-multi-agent-communication-topologies - ST-EVO enhances collaborative intelligence in multi-agent systems with a novel spatio-temporal evolution framework.
- GREAT-EER: Graph Edge Attention Network for Emergency Evacuation Responses (viability: 3): https://sciencetostartup.com/paper/great-eer-graph-edge-attention-network-for-emergency-evacuation-responses - Develop fast, efficient bus evacuation plans for urban emergencies using deep reinforcement learning and graph networks.
- From User Preferences to Base Score Extraction Functions in Gradual Argumentation (viability: 2): https://sciencetostartup.com/paper/from-user-preferences-to-base-score-extraction-functions-in-gradual-argumentation - Develop algorithms for mapping user preferences to base scores in argumentation frameworks.
- Towards Selection as Power: Bounding Decision Authority in Autonomous Agents (viability: 3): https://sciencetostartup.com/paper/towards-selection-as-power-bounding-decision-authority-in-autonomous-agents - A new governance architecture for autonomous agents that bounds decision authority while preserving cognitive reasoning, suitable for high-stakes regulated environments.
- OPBench: A Graph Benchmark to Combat the Opioid Crisis (viability: 7): https://sciencetostartup.com/paper/opbench-a-graph-benchmark-to-combat-the-opioid-crisis - OPBench provides comprehensive datasets and evaluation tools for graph-based solutions targeting opioid crisis challenges.
- MATEO: A Multimodal Benchmark for Temporal Reasoning and Planning in LVLMs (viability: 5): https://sciencetostartup.com/paper/mateo-a-multimodal-benchmark-for-temporal-reasoning-and-planning-in-lvlms - Develop MATEO as a benchmark tool to enhance temporal reasoning in large vision language models using multimodal data.
- Decoupled Continuous-Time Reinforcement Learning via Hamiltonian Flow (viability: 6): https://sciencetostartup.com/paper/decoupled-continuous-time-reinforcement-learning-via-hamiltonian-flow - Develop a decoupled continuous-time reinforcement learning system for real-world control problems in finance and robotics.
- Formally Verifying and Explaining Sepsis Treatment Policies with COOL-MC (viability: 5): https://sciencetostartup.com/paper/formally-verifying-and-explaining-sepsis-treatment-policies-with-cool-mc - COOL-MC offers clinicians a tool to formally verify and explain sepsis treatment policies using a model checker integrated with explainability methods.
- Bounding Probabilities of Causation with Partial Causal Diagrams (viability: 3): https://sciencetostartup.com/paper/bounding-probabilities-of-causation-with-partial-causal-diagrams - Develop a framework for bounding probabilities of causation using partial causal information to improve individual-level explanation and decision making.
- BETA-Labeling for Multilingual Dataset Construction in Low-Resource IR (viability: 5): https://sciencetostartup.com/paper/beta-labeling-for-multilingual-dataset-construction-in-low-resource-ir - Develop a framework for multilingual dataset construction in low-resource languages using BETA-labeling and LLMs.
- Revisiting the Platonic Representation Hypothesis: An Aristotelian View (viability: 2): https://sciencetostartup.com/paper/revisiting-the-platonic-representation-hypothesis-an-aristotelian-view - Introducing a new framework to calibrate representational similarity metrics in neural networks for clearer insights into converging representations.
- On the Rate-Distortion-Complexity Tradeoff for Semantic Communication (viability: 3): https://sciencetostartup.com/paper/on-the-rate-distortion-complexity-tradeoff-for-semantic-communication - Innovative semantic communication paradigm optimizing rate-distortion-complexity tradeoff through theoretical insights and real-world validation.
- When OpenClaw AI Agents Teach Each Other: Peer Learning Patterns in the Moltbook Community (viability: 2): https://sciencetostartup.com/paper/when-openclaw-ai-agents-teach-each-other-peer-learning-patterns-in-the-moltbook-community - Explore AI agents' peer learning dynamics in the Moltbook community.
- Learning Transferability: A Two-Stage Reinforcement Learning Approach for Enhancing Quadruped Robots' Performance in U-Shaped Stair Climbing (viability: 3): https://sciencetostartup.com/paper/learning-transferability-a-two-stage-reinforcement-learning-approach-for-enhancing-quadruped-robots-performance-in-u-sha - Develop reinforcement learning models for quadruped robots to autonomously climb stairs in construction environments.
- CoCoDiff: Correspondence-Consistent Diffusion Model for Fine-grained Style Transfer (viability: 4): https://sciencetostartup.com/paper/cocodiff-correspondence-consistent-diffusion-model-for-fine-grained-style-transfer - Develop a low-cost, training-free framework for fine-grained style transfer using pretrained diffusion models.
- Selective Synchronization Attention (viability: 4): https://sciencetostartup.com/paper/selective-synchronization-attention - Develop Selective Synchronization Attention (SSA), a biologically grounded alternative to self-attention in Transformers, for efficient attention computation.
- Synthetic Reader Panels: Tournament-Based Ideation with LLM Personas for Autonomous Publishing (viability: 6): https://sciencetostartup.com/paper/synthetic-reader-panels-tournament-based-ideation-with-llm-personas-for-autonomous-publishing - Automate book publishing ideation using diverse LLM personas to enhance concept evaluation and demographic segmentation.
- Feature Recalibration Based Olfactory-Visual Multimodal Model for Fine-Grained Rice Deterioration Detection (viability: 9): https://sciencetostartup.com/paper/feature-recalibration-based-olfactory-visual-multimodal-model-for-fine-grained-rice-deterioration-detection - A multimodal AI model for precision detection of rice deterioration, enhancing accuracy and cost-effectiveness in agrifood quality control.
- TruthStance: An Annotated Dataset of Conversations on Truth Social (viability: 5): https://sciencetostartup.com/paper/truthstance-an-annotated-dataset-of-conversations-on-truth-social - New dataset and tools for analyzing online discourse patterns on Truth Social.
- pFedNavi: Structure-Aware Personalized Federated Vision-Language Navigation for Embodied AI (viability: 6): https://sciencetostartup.com/paper/pfednavi-structure-aware-personalized-federated-vision-language-navigation-for-embodied-ai - Develop a personalized federated learning framework optimizing vision-language navigation models for improved privacy and performance in diverse environments.
- Competition for attention predicts good-to-bad tipping in AI (viability: 3): https://sciencetostartup.com/paper/competition-for-attention-predicts-good-to-bad-tipping-in-ai - A theoretical exploration of attention competition in edge AI models revealing potential safety tipping points.
- Image-based Joint-level Detection for Inflammation in Rheumatoid Arthritis from Small and Imbalanced Data (viability: 5): https://sciencetostartup.com/paper/image-based-joint-level-detection-for-inflammation-in-rheumatoid-arthritis-from-small-and-imbalanced-data - Develop a home-use RA inflammation detection system using RGB hand images for early diagnosis and management.
- Key Considerations for Domain Expert Involvement in LLM Design and Evaluation: An Ethnographic Study (viability: 3): https://sciencetostartup.com/paper/key-considerations-for-domain-expert-involvement-in-llm-design-and-evaluation-an-ethnographic-study - An ethnographic study revealing key practices and challenges in LLM design and evaluation in professional domains with a focus on domain expertise.
- WIMLE: Uncertainty-Aware World Models with IMLE for Sample-Efficient Continuous Control (viability: 3): https://sciencetostartup.com/paper/wimle-uncertainty-aware-world-models-with-imle-for-sample-efficient-continuous-control - WIMLE enhances reinforcement learning with IMLE for superior sample efficiency in continuous control tasks.
- AXE: An Agentic eXploit Engine for Confirming Zero-Day Vulnerability Reports (viability: 7): https://sciencetostartup.com/paper/axe-an-agentic-exploit-engine-for-confirming-zero-day-vulnerability-reports - AXE is a multi-agent framework that enhances Web application security by converting vulnerability reports into exploitable exploits using minimal metadata.
- Zero-Shot Instruction Following in RL via Structured LTL Representations (viability: 6): https://sciencetostartup.com/paper/zero-shot-instruction-following-in-rl-via-structured-ltl-representations - A novel RL framework using LTL for task generalization, targeting zero-shot task execution in multi-task environments.
- Benchmarking at the Edge of Comprehension (viability: 5): https://sciencetostartup.com/paper/benchmarking-at-the-edge-of-comprehension - Develop an adversarial benchmarking framework for LLMs to ensure evaluation integrity beyond human comprehension.
- KernelBlaster: Continual Cross-Task CUDA Optimization via Memory-Augmented In-Context Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/kernelblaster-continual-cross-task-cuda-optimization-via-memory-augmented-in-context-reinforcement-learning - KernelBlaster enhances CUDA optimization with a knowledge-accumulating reinforcement learning framework for superior GPU coding performance.
- FMMD: A multimodal open peer review dataset based on F1000Research (viability: 5): https://sciencetostartup.com/paper/fmmd-a-multimodal-open-peer-review-dataset-based-on-f1000research - FMMD offers a rich, multimodal peer review dataset for AI-enhanced manuscript evaluation research.
- Semantic Chunking and the Entropy of Natural Language (viability: 5): https://sciencetostartup.com/paper/semantic-chunking-and-the-entropy-of-natural-language - Develop a semantic chunking tool to quantitatively analyze the structure and redundancy of natural language texts.
- CoPE-VideoLM: Codec Primitives For Efficient Video Language Models (viability: 7): https://sciencetostartup.com/paper/cope-videolm-codec-primitives-for-efficient-video-language-models - CoPE-VideoLM drastically improves video processing efficiency by using codec primitives for lightweight video tokenization in AI models.
- Optimal Take-off under Fuzzy Clearances (viability: 5): https://sciencetostartup.com/paper/optimal-take-off-under-fuzzy-clearances - Develop a control system integrating optimal control and fuzzy logic for adaptive obstacle avoidance in unmanned aircraft.
- Asynchronous Verified Semantic Caching for Tiered LLM Architectures (viability: 2): https://sciencetostartup.com/paper/asynchronous-verified-semantic-caching-for-tiered-llm-architectures - Develop an asynchronous LLM-judged caching policy to improve semantic caching efficiency in tiered LLM architectures.
- In-Context Autonomous Network Incident Response: An End-to-End Large Language Model Agent Approach (viability: 8): https://sciencetostartup.com/paper/in-context-autonomous-network-incident-response-an-end-to-end-large-language-model-agent-approach - An end-to-end LLM agent for faster and smarter autonomous network incident response.
- Constrained Assumption-Based Argumentation Frameworks (viability: 3): https://sciencetostartup.com/paper/constrained-assumption-based-argumentation-frameworks - Innovative framework enhancing Assumption-Based Argumentation with constrained variables for more expressive arguments.
- SCOPE: Selective Conformal Optimized Pairwise LLM Judging (viability: 5): https://sciencetostartup.com/paper/scope-selective-conformal-optimized-pairwise-llm-judging - Develop a selective judging framework for LLMs to improve pairwise evaluation reliability using statistical guarantees.
- Which Algorithms Can Graph Neural Networks Learn? (viability: 3): https://sciencetostartup.com/paper/which-algorithms-can-graph-neural-networks-learn - A theoretical framework to ensure generalization of Graph Neural Networks to larger inputs in neural algorithmic reasoning.
- Consistency of Large Reasoning Models Under Multi-Turn Attacks (viability: 5): https://sciencetostartup.com/paper/consistency-of-large-reasoning-models-under-multi-turn-attacks - Develop advanced defenses for reasoning AI models against multi-turn adversarial attacks.
- How cyborg propaganda reshapes collective action (viability: 3): https://sciencetostartup.com/paper/how-cyborg-propaganda-reshapes-collective-action - Exploring cyborg propaganda's impact on digital democracy and governance challenges.
- EXCODER: EXplainable Classification Of DiscretE time series Representations (viability: 5): https://sciencetostartup.com/paper/excoder-explainable-classification-of-discrete-time-series-representations - Develop an explainable AI tool for time series analysis using discrete latent representations for enhanced interpretability.
- Bus-Conditioned Zero-Shot Trajectory Generation via Task Arithmetic (viability: 3): https://sciencetostartup.com/paper/bus-conditioned-zero-shot-trajectory-generation-via-task-arithmetic - Innovative zero-shot trajectory generation for smart cities using bus timetables and task arithmetic.
- Diverging Flows: Detecting Extrapolations in Conditional Generation (viability: 5): https://sciencetostartup.com/paper/diverging-flows-detecting-extrapolations-in-conditional-generation - Diverging Flows offers a conditional generation model that detects extrapolations, enhancing reliability in safety-critical applications.
- Curriculum-DPO++: Direct Preference Optimization via Data and Model Curricula for Text-to-Image Generation (viability: 8): https://sciencetostartup.com/paper/curriculum-dpo-direct-preference-optimization-via-data-and-model-curricula-for-text-to-image-generation - Curriculum-DPO++ improves text-to-image AI by optimizing learning sequences for better preference alignment.
- Can we trust AI to detect healthy multilingual English speakers among the cognitively impaired cohort in the UK? An investigation using real-world conversational speech (viability: 6): https://sciencetostartup.com/paper/can-we-trust-ai-to-detect-healthy-multilingual-english-speakers-among-the-cognitively-impaired-cohort-in-the-uk-an-inves - Developing AI models to accurately diagnose cognitive decline in multilingual populations across ethnic minorities in the UK.
- Geometric Manifold Rectification for Imbalanced Learning (viability: 5): https://sciencetostartup.com/paper/geometric-manifold-rectification-for-imbalanced-learning - Leverage GMR for improved classification in imbalanced datasets using geometric manifold rectification.
- Look Inward to Explore Outward: Learning Temperature Policy from LLM Internal States via Hierarchical RL (viability: 5): https://sciencetostartup.com/paper/look-inward-to-explore-outward-learning-temperature-policy-from-llm-internal-states-via-hierarchical-rl - Develop an AI framework that dynamically adjusts language model exploration using learned temperature strategies for better task-specific performance.
- Buy versus Build an LLM: A Decision Framework for Governments (viability: 3): https://sciencetostartup.com/paper/buy-versus-build-an-llm-a-decision-framework-for-governments - A strategic framework to help governments decide whether to buy or build LLMs for public-sector applications.
- Prior-Guided Symbolic Regression: Towards Scientific Consistency in Equation Discovery (viability: 3): https://sciencetostartup.com/paper/prior-guided-symbolic-regression-towards-scientific-consistency-in-equation-discovery - A novel symbolic regression framework to ensure scientific consistency in equation discovery.
- Deep-Learning Atlas Registration for Melanoma Brain Metastases: Preserving Pathology While Enabling Cohort-Level Analyses (viability: 5): https://sciencetostartup.com/paper/deep-learning-atlas-registration-for-melanoma-brain-metastases-preserving-pathology-while-enabling-cohort-level-analyses - Develop a deep-learning framework for standardizing the registration of melanoma brain metastases MRI, enhancing reproducible multi-centre analysis.
- Never say never: Exploring the effects of available knowledge on agent persuasiveness in controlled physiotherapy motivation dialogues (viability: 2): https://sciencetostartup.com/paper/never-say-never-exploring-the-effects-of-available-knowledge-on-agent-persuasiveness-in-controlled-physiotherapy-motivat - Exploring the impact of knowledge availability on the persuasiveness of AI in physiotherapy dialogues.
- Robustness of Object Detection of Autonomous Vehicles in Adverse Weather Conditions (viability: 6): https://sciencetostartup.com/paper/robustness-of-object-detection-of-autonomous-vehicles-in-adverse-weather-conditions - Develop a tool to assess and enhance the robustness of autonomous vehicle object detection in adverse weather conditions.
- RADAR: Revealing Asymmetric Development of Abilities in MLLM Pre-training (viability: 5): https://sciencetostartup.com/paper/radar-revealing-asymmetric-development-of-abilities-in-mllm-pre-training - RADAR provides an efficient framework to evaluate the asymmetric development of skills in Multi-modal Large Language Models without fine-tuning.
- BrowseComp-$V^3$: A Visual, Vertical, and Verifiable Benchmark for Multimodal Browsing Agents (viability: 5): https://sciencetostartup.com/paper/browsecomp-v-3-a-visual-vertical-and-verifiable-benchmark-for-multimodal-browsing-agents - Develop a unified multimodal browsing agent system using the BrowseComp-$V^3$ benchmark to enhance deep search capabilities.
- A Microservice-Based Platform for Sustainable and Intelligent SLO Fulfilment and Service Management (viability: 3): https://sciencetostartup.com/paper/a-microservice-based-platform-for-sustainable-and-intelligent-slo-fulfilment-and-service-management - CASCA is an open-source platform enabling computing continuum providers to balance service performance and sustainability while respecting developers' privacy.
- Knowledge-Based Design Requirements for Generative Social Robots in Higher Education (viability: 2): https://sciencetostartup.com/paper/knowledge-based-design-requirements-for-generative-social-robots-in-higher-education - Develop generative social robots tailored for higher education to enhance adaptive and responsible conversational tutoring.
- WebClipper: Efficient Evolution of Web Agents with Graph-based Trajectory Pruning (viability: 6): https://sciencetostartup.com/paper/webclipper-efficient-evolution-of-web-agents-with-graph-based-trajectory-pruning - WebClipper enhances web agent efficiency by pruning redundant search trajectories with graph-based optimization.
- MedXIAOHE: A Comprehensive Recipe for Building Medical MLLMs (viability: 3): https://sciencetostartup.com/paper/medxiaohe-a-comprehensive-recipe-for-building-medical-mllms - Develop a state-of-the-art medical vision-language model for enhanced clinical applications.
- SLA2: Sparse-Linear Attention with Learnable Routing and QAT (viability: 3): https://sciencetostartup.com/paper/sla2-sparse-linear-attention-with-learnable-routing-and-qat - SLA2 enhances computational efficiency in video diffusion models by using a learnable routing mechanism for sparse-linear attention.
- SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks (viability: 8): https://sciencetostartup.com/paper/skillsbench-benchmarking-how-well-agent-skills-work-across-diverse-tasks - SkillsBench evaluates the effectiveness of procedural Skills in boosting LLM agent task performance.
- Think Fast and Slow: Step-Level Cognitive Depth Adaptation for LLM Agents (viability: 5): https://sciencetostartup.com/paper/think-fast-and-slow-step-level-cognitive-depth-adaptation-for-llm-agents - A framework for adaptive cognitive depth in LLMs that optimizes response efficiency and performance.
- IndicFairFace: Balanced Indian Face Dataset for Auditing and Mitigating Geographical Bias in Vision-Language Models (viability: 5): https://sciencetostartup.com/paper/indicfairface-balanced-indian-face-dataset-for-auditing-and-mitigating-geographical-bias-in-vision-language-models - IndicFairFace offers a balanced Indian face dataset for mitigating geographical bias in vision-language models.
- Multi-Task Learning with Additive U-Net for Image Denoising and Classification (viability: 3): https://sciencetostartup.com/paper/multi-task-learning-with-additive-u-net-for-image-denoising-and-classification - Develop an enhanced U-Net architecture for stable multi-task image denoising and classification.
- Beyond Normalization: Rethinking the Partition Function as a Difficulty Scheduler for RLVR (viability: 5): https://sciencetostartup.com/paper/beyond-normalization-rethinking-the-partition-function-as-a-difficulty-scheduler-for-rlvr - Leveraging partition functions as difficulty schedulers, PACED-RL optimizes LLM performance for more efficient reward learning.
- AI Agents for Inventory Control: Human-LLM-OR Complementarity (viability: 5): https://sciencetostartup.com/paper/ai-agents-for-inventory-control-human-llm-or-complementarity - Develop a hybrid human-AI decision-making platform for inventory control that enhances operational efficiency by integrating LLMs and OR algorithms.
- TensorCommitments: A Lightweight Verifiable Inference for Language Models (viability: 2): https://sciencetostartup.com/paper/tensorcommitments-a-lightweight-verifiable-inference-for-language-models - Secure LLM inference verification through TensorCommitments without large computational costs.
- Power Interpretable Causal ODE Networks: A Unified Model for Explainable Anomaly Detection and Root Cause Analysis in Power Systems (viability: 6): https://sciencetostartup.com/paper/power-interpretable-causal-ode-networks-a-unified-model-for-explainable-anomaly-detection-and-root-cause-analysis-in-pow - A causality-informed model for interpretable anomaly detection and root cause analysis in power systems.
- Can I Have Your Order? Monte-Carlo Tree Search for Slot Filling Ordering in Diffusion Language Models (viability: 3): https://sciencetostartup.com/paper/can-i-have-your-order-monte-carlo-tree-search-for-slot-filling-ordering-in-diffusion-language-models - McDiffuSE enhances slot infilling order in diffusion language models using Monte Carlo Tree Search to improve generation quality.
- Monte Carlo Tree Search with Reasoning Path Refinement for Small Language Models in Conversational Text-to-NoSQL (viability: 7): https://sciencetostartup.com/paper/monte-carlo-tree-search-with-reasoning-path-refinement-for-small-language-models-in-conversational-text-to-nosql - Enable users to generate NoSQL queries through natural language in a conversational interface using a reasoning-enhanced small language model.
- To Mix or To Merge: Toward Multi-Domain Reinforcement Learning for Large Language Models (viability: 5): https://sciencetostartup.com/paper/to-mix-or-to-merge-toward-multi-domain-reinforcement-learning-for-large-language-models - M2RL enhances multi-domain reasoning in LLMs through Reinforcement Learning with Verifiable Rewards.
- Bench-MFG: A Benchmark Suite for Learning in Stationary Mean Field Games (viability: 5): https://sciencetostartup.com/paper/bench-mfg-a-benchmark-suite-for-learning-in-stationary-mean-field-games - Bench-MFG offers a standardized benchmarking suite for evaluating learning algorithms in Mean Field Games, facilitating robust and comparable multi-agent system research.
- Monocular Reconstruction of Neural Tactile Fields (viability: 3): https://sciencetostartup.com/paper/monocular-reconstruction-of-neural-tactile-fields - Develop a monocular neural tactile field reconstruction tool for improved robotic navigation through deformable environments.
- Human-Like Coarse Object Representations in Vision Models (viability: 3): https://sciencetostartup.com/paper/human-like-coarse-object-representations-in-vision-models - Exploring human-like object representations in vision models for physics predictions.
- A Lightweight and Explainable DenseNet-121 Framework for Grape Leaf Disease Classification (viability: 6): https://sciencetostartup.com/paper/a-lightweight-and-explainable-densenet-121-framework-for-grape-leaf-disease-classification - Optimized DenseNet-121 framework for real-time, explainable grape leaf disease classification.
- Intent-Driven Smart Manufacturing Integrating Knowledge Graphs and Large Language Models (viability: 7): https://sciencetostartup.com/paper/intent-driven-smart-manufacturing-integrating-knowledge-graphs-and-large-language-models - Unified framework integrating LLMs and Knowledge Graphs for intent-driven smart manufacturing.
- Soft Contamination Means Benchmarks Test Shallow Generalization (viability: 3): https://sciencetostartup.com/paper/soft-contamination-means-benchmarks-test-shallow-generalization - A method to detect soft contamination in LLM training data affecting benchmark performance.
- AstRL: Analog and Mixed-Signal Circuit Synthesis with Deep Reinforcement Learning (viability: 5): https://sciencetostartup.com/paper/astrl-analog-and-mixed-signal-circuit-synthesis-with-deep-reinforcement-learning - AstRL automates analog circuit design using deep reinforcement learning to optimize design metrics beyond state-of-the-art.
- What does RL improve for Visual Reasoning? A Frankenstein-Style Analysis (viability: 5): https://sciencetostartup.com/paper/what-does-rl-improve-for-visual-reasoning-a-frankenstein-style-analysis - A framework for understanding reinforcement learning's specific impact on visual reasoning in vision-language models.
- Reproducing DragDiffusion: Interactive Point-Based Editing with Diffusion Models (viability: 5): https://sciencetostartup.com/paper/reproducing-dragdiffusion-interactive-point-based-editing-with-diffusion-models - Develop an interactive point-based image editing tool utilizing diffusion models for precise spatial control.
- Rational Neural Networks have Expressivity Advantages (viability: 2): https://sciencetostartup.com/paper/rational-neural-networks-have-expressivity-advantages - Exploring the expressivity of rational activation functions in neural networks for enhanced parameter efficiency.
- Why Deep Jacobian Spectra Separate: Depth-Induced Scaling and Singular-Vector Alignment (viability: 2): https://sciencetostartup.com/paper/why-deep-jacobian-spectra-separate-depth-induced-scaling-and-singular-vector-alignment - The paper provides theoretical insights into deep network training dynamics through an analysis of deep Jacobians.
- Scaling Verification Can Be More Effective than Scaling Policy Learning for Vision-Language-Action Alignment (viability: 6): https://sciencetostartup.com/paper/scaling-verification-can-be-more-effective-than-scaling-policy-learning-for-vision-language-action-alignment - Develop CoVer, an efficient verification system improving vision-language-action alignment in robots using contrastive verification.
- UniT: Unified Multimodal Chain-of-Thought Test-time Scaling (viability: 7): https://sciencetostartup.com/paper/unit-unified-multimodal-chain-of-thought-test-time-scaling - UniT enhances multimodal models with iterative reasoning capabilities for improved performance across complex tasks.
- AttentionRetriever: Attention Layers are Secretly Long Document Retrievers (viability: 5): https://sciencetostartup.com/paper/attentionretriever-attention-layers-are-secretly-long-document-retrievers - AttentionRetriever efficiently improves long document retrieval using advanced attention mechanisms.
- Agentic Test-Time Scaling for WebAgents (viability: 3): https://sciencetostartup.com/paper/agentic-test-time-scaling-for-webagents - CATTS dynamically allocates compute for web agents, boosting performance efficiently using vote-derived uncertainty.
- CM2: Reinforcement Learning with Checklist Rewards for Multi-Turn and Multi-Step Agentic Tool Use (viability: 7): https://sciencetostartup.com/paper/cm2-reinforcement-learning-with-checklist-rewards-for-multi-turn-and-multi-step-agentic-tool-use - CM2 leverages checklist rewards in RL to optimize AI agents for complex multi-step tool interaction tasks.
- Think like a Scientist: Physics-guided LLM Agent for Equation Discovery (viability: 7): https://sciencetostartup.com/paper/think-like-a-scientist-physics-guided-llm-agent-for-equation-discovery - Develop an agent framework to improve symbolic equation discovery using physics-guided reasoning with large language models.
- A technical curriculum on language-oriented artificial intelligence in translation and specialised communication (viability: 2): https://sciencetostartup.com/paper/a-technical-curriculum-on-language-oriented-artificial-intelligence-in-translation-and-specialised-communication - Develop a curriculum to increase technical AI literacy in translation and communication fields.
- "Sorry, I Didn't Catch That": How Speech Models Miss What Matters Most (viability: 6): https://sciencetostartup.com/paper/sorry-i-didn-t-catch-that-how-speech-models-miss-what-matters-most - Improve transcription accuracy for high-stakes voice applications using synthetic data augmentation.
- ExtractBench: A Benchmark and Evaluation Methodology for Complex Structured Extraction (viability: 7): https://sciencetostartup.com/paper/extractbench-a-benchmark-and-evaluation-methodology-for-complex-structured-extraction - ExtractBench provides an open-source benchmark and evaluation framework for structured data extraction from complex PDFs to JSON, addressing key enterprise challenges.
- Intrinsic-Energy Joint Embedding Predictive Architectures Induce Quasimetric Spaces (viability: 2): https://sciencetostartup.com/paper/intrinsic-energy-joint-embedding-predictive-architectures-induce-quasimetric-spaces - Develop a predictive architecture using intrinsic energy for goal-reaching tasks by modeling quasimetric spaces.
- Olmix: A Framework for Data Mixing Throughout LM Development (viability: 4): https://sciencetostartup.com/paper/olmix-a-framework-for-data-mixing-throughout-lm-development - Olmix optimizes data mixing configurations in LM development to increase efficiency and performance.
- Energy-Aware Spike Budgeting for Continual Learning in Spiking Neural Networks for Neuromorphic Vision (viability: 7): https://sciencetostartup.com/paper/energy-aware-spike-budgeting-for-continual-learning-in-spiking-neural-networks-for-neuromorphic-vision - Develop an energy-efficient spike budgeting framework for continual learning in spiking neural networks for neuromorphic vision systems.
- Bandit Learning in Matching Markets with Interviews (viability: 3): https://sciencetostartup.com/paper/bandit-learning-in-matching-markets-with-interviews - Develop algorithms for efficient decentralized learning in matching markets using strategic interview processes.
- Towards On-Policy SFT: Distribution Discriminant Theory and its Applications in LLM Training (viability: 5): https://sciencetostartup.com/paper/towards-on-policy-sft-distribution-discriminant-theory-and-its-applications-in-llm-training - Develop a high-performance supervised fine-tuning framework based on Distribution Discriminant Theory to achieve RL-like generalization in computationally efficient settings.
- The Observer Effect in World Models: Invasive Adaptation Corrupts Latent Physics (viability: 3): https://sciencetostartup.com/paper/the-observer-effect-in-world-models-invasive-adaptation-corrupts-latent-physics - Proposes a non-invasive protocol to evaluate neural models' understanding of physical laws without disrupting learned representations.
- VIRENA: Virtual Arena for Research, Education, and Democratic Innovation (viability: 8): https://sciencetostartup.com/paper/virena-virtual-arena-for-research-education-and-democratic-innovation - VIRENA provides a no-code platform for conducting realistic, controlled experiments on social media interactions with AI agents, tailored for researchers and educators.
- DeepGen 1.0: A Lightweight Unified Multimodal Model for Advancing Image Generation and Editing (viability: 8): https://sciencetostartup.com/paper/deepgen-1-0-a-lightweight-unified-multimodal-model-for-advancing-image-generation-and-editing - DeepGen 1.0 offers a lightweight but powerful multimodal model for image generation and editing, surpassing larger models while being open-sourced.
- Visual Reasoning Benchmark: Evaluating Multimodal LLMs on Classroom-Authentic Visual Problems from Primary Education (viability: 4): https://sciencetostartup.com/paper/visual-reasoning-benchmark-evaluating-multimodal-llms-on-classroom-authentic-visual-problems-from-primary-education - Develop a benchmark toolkit for assessing multimodal AI's visual reasoning capabilities in primary education environments.
- SAGEO Arena: A Realistic Environment for Evaluating Search-Augmented Generative Engine Optimization (viability: 5): https://sciencetostartup.com/paper/sageo-arena-a-realistic-environment-for-evaluating-search-augmented-generative-engine-optimization - SAGEO Arena offers a benchmark environment to improve web content optimization in AI-generated search results by leveraging structural information in web documents.
- SAM3-LiteText: An Anatomical Study of the SAM3 Text Encoder for Efficient Vision-Language Segmentation (viability: 6): https://sciencetostartup.com/paper/sam3-litetext-an-anatomical-study-of-the-sam3-text-encoder-for-efficient-vision-language-segmentation - SAM3-LiteText offers a compact text encoding framework to efficiently reduce computational resources in vision-language segmentation models.
- Sci-CoE: Co-evolving Scientific Reasoning LLMs via Geometric Consensus with Sparse Supervision (viability: 6): https://sciencetostartup.com/paper/sci-coe-co-evolving-scientific-reasoning-llms-via-geometric-consensus-with-sparse-supervision - Sci-CoE advances scientific reasoning in LLMs using a co-evolving framework with sparse supervision and self-iteration.
- dVoting: Fast Voting for dLLMs (viability: 5): https://sciencetostartup.com/paper/dvoting-fast-voting-for-dllms - dVoting offers a faster reasoning technique for diffusion large language models, improving accuracy without extra training.
- GPT-4o Lacks Core Features of Theory of Mind (viability: 4): https://sciencetostartup.com/paper/gpt-4o-lacks-core-features-of-theory-of-mind - Develop a framework to evaluate whether LLMs possess a coherent Theory of Mind model for social task applications.
- Seq2Seq2Seq: Lossless Data Compression via Discrete Latent Transformers and Reinforcement Learning (viability: 1): https://sciencetostartup.com/paper/seq2seq2seq-lossless-data-compression-via-discrete-latent-transformers-and-reinforcement-learning - Novel lossless compression method using Reinforcement Learning and T5 models to enhance data compression efficiency.
- On the Adoption of AI Coding Agents in Open-source Android and iOS Development (viability: 4): https://sciencetostartup.com/paper/on-the-adoption-of-ai-coding-agents-in-open-source-android-and-ios-development - Leverage AI coding agents to optimize PR acceptance and resolution time in open-source Android and iOS development.
- STAR : Bridging Statistical and Agentic Reasoning for Large Model Performance Prediction (viability: 5): https://sciencetostartup.com/paper/star-bridging-statistical-and-agentic-reasoning-for-large-model-performance-prediction - STAR enhances model performance prediction by integrating statistical and agentic reasoning for significant accuracy improvements.
- Value Alignment Tax: Measuring Value Trade-offs in LLM Alignment (viability: 4): https://sciencetostartup.com/paper/value-alignment-tax-measuring-value-trade-offs-in-llm-alignment - Introducing VAT, a framework that measures dynamic value trade-offs in LLM alignment.
- Neutral Prompts, Non-Neutral People: Quantifying Gender and Skin-Tone Bias in Gemini Flash 2.5 Image and GPT Image 1.5 (viability: 5): https://sciencetostartup.com/paper/neutral-prompts-non-neutral-people-quantifying-gender-and-skin-tone-bias-in-gemini-flash-2-5-image-and-gpt-image-1-5 - Develop a diagnostic tool to evaluate and quantify biases in commercial image generation models using neutral prompts.
- HLA: Hadamard Linear Attention (viability: 4): https://sciencetostartup.com/paper/hla-hadamard-linear-attention - Implement Hadamard Linear Attention in transformers to improve efficiency in video generation tasks.
- Learning beyond Teacher: Generalized On-Policy Distillation with Reward Extrapolation (viability: 5): https://sciencetostartup.com/paper/learning-beyond-teacher-generalized-on-policy-distillation-with-reward-extrapolation - Develop a tool for enhanced on-policy distillation in reinforcement learning by leveraging reward extrapolation techniques for improved student performance.
- Meta-Sel: Efficient Demonstration Selection for In-Context Learning via Supervised Meta-Learning (viability: 6): https://sciencetostartup.com/paper/meta-sel-efficient-demonstration-selection-for-in-context-learning-via-supervised-meta-learning - Meta-Sel optimizes in-context learning by efficiently selecting demonstrations through a lightweight meta-learning approach, enhancing accuracy for intent classification tasks.
- Commencing-Student Enrolment Forecasting Under Data Sparsity with Time Series Foundation Models (viability: 5): https://sciencetostartup.com/paper/commencing-student-enrolment-forecasting-under-data-sparsity-with-time-series-foundation-models - Develop a zero-shot forecasting tool using TSFMs to improve enrolment predictions for universities facing data-sparse environments.
- KAN-FIF: Spline-Parameterized Lightweight Physics-based Tropical Cyclone Estimation on Meteorological Satellite (viability: 8): https://sciencetostartup.com/paper/kan-fif-spline-parameterized-lightweight-physics-based-tropical-cyclone-estimation-on-meteorological-satellite - Develop a lightweight, high-performance AI tool for tropical cyclone monitoring on edge devices.
- The Pensieve Paradigm: Stateful Language Models Mastering Their Own Context (viability: 6): https://sciencetostartup.com/paper/the-pensieve-paradigm-stateful-language-models-mastering-their-own-context - StateLM transforms language models into state-aware agents capable of managing their own context and memory.
- Multi Graph Search for High-Dimensional Robot Motion Planning (viability: 6): https://sciencetostartup.com/paper/multi-graph-search-for-high-dimensional-robot-motion-planning - Multi-Graph Search (MGS) provides efficient motion planning for high-dimensional robots, enhancing scalability and predictability with accessible demonstrations and benchmarks.
- DeepSight: An All-in-One LM Safety Toolkit (viability: 3): https://sciencetostartup.com/paper/deepsight-an-all-in-one-lm-safety-toolkit - DeepSight provides an open-source toolkit for integrated safety evaluation and diagnosis of large language models.
- Choose Your Agent: Tradeoffs in Adopting AI Advisors, Coaches, and Delegates in Multi-Party Negotiation (viability: 4): https://sciencetostartup.com/paper/choose-your-agent-tradeoffs-in-adopting-ai-advisors-coaches-and-delegates-in-multi-party-negotiation - Develop and deploy AI Advisors, Coaches, and Delegates to optimize outcomes in multi-party negotiations by aligning user preferences with performance.
- Differentiable Modal Logic for Multi-Agent Diagnosis, Orchestration and Communication (viability: 7): https://sciencetostartup.com/paper/differentiable-modal-logic-for-multi-agent-diagnosis-orchestration-and-communication - Develop a framework for debugging multi-agent systems using differentiable modal logic implemented through Modal Logical Neural Networks.
- Tiny Recursive Reasoning with Mamba-2 Attention Hybrid (viability: 4): https://sciencetostartup.com/paper/tiny-recursive-reasoning-with-mamba-2-attention-hybrid - Develop a reasoning AI model using Mamba-2 operators that enhances performance on abstract reasoning tasks.
- ModelWisdom: An Integrated Toolkit for TLA+ Model Visualization, Digest and Repair (viability: 7): https://sciencetostartup.com/paper/modelwisdom-an-integrated-toolkit-for-tla-model-visualization-digest-and-repair - ModelWisdom enhances TLA+ model checking with interactive visualization and debugging tools, reducing complexity for practitioners.
- LawThinker: A Deep Research Legal Agent in Dynamic Environments (viability: 6): https://sciencetostartup.com/paper/lawthinker-a-deep-research-legal-agent-in-dynamic-environments - LawThinker is an autonomous legal research agent utilizing Explore-Verify-Memorize strategies to ensure procedurally compliant legal reasoning.
- Multi UAVs Preflight Planning in a Shared and Dynamic Airspace (viability: 9): https://sciencetostartup.com/paper/multi-uavs-preflight-planning-in-a-shared-and-dynamic-airspace - A scalable and efficient solution for preflight planning of large UAV fleets in dynamic urban airspaces.
- Fourier Transformers for Latent Crystallographic Diffusion and Generative Modeling (viability: 5): https://sciencetostartup.com/paper/fourier-transformers-for-latent-crystallographic-diffusion-and-generative-modeling - Innovative crystal generative model leveraging Fourier transforms and latent diffusion for material discovery.
- An Empirical Study of the Imbalance Issue in Software Vulnerability Detection (viability: 5): https://sciencetostartup.com/paper/an-empirical-study-of-the-imbalance-issue-in-software-vulnerability-detection - Optimize deep learning-based software vulnerability detection by addressing data imbalance issues using empirical findings.
- InjectRBP: Steering Large Language Model Reasoning Behavior via Pattern Injection (viability: 6): https://sciencetostartup.com/paper/injectrbp-steering-large-language-model-reasoning-behavior-via-pattern-injection - Develop a tool for steering reasoning behavior in LLMs via pattern injection to enhance performance without modifying model parameters.
- On the Sensitivity of Firing Rate-Based Federated Spiking Neural Networks to Differential Privacy (viability: 3): https://sciencetostartup.com/paper/on-the-sensitivity-of-firing-rate-based-federated-spiking-neural-networks-to-differential-privacy - Develops guidelines for privacy-preserving Federated Neuromorphic Learning using Differential Privacy mechanisms.
- CSEval: A Framework for Evaluating Clinical Semantics in Text-to-Image Generation (viability: 3): https://sciencetostartup.com/paper/cseval-a-framework-for-evaluating-clinical-semantics-in-text-to-image-generation - CSEval provides a framework for evaluating the clinical semantics of images generated by text-to-image models.
- Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents? (viability: 5): https://sciencetostartup.com/paper/evaluating-agents-md-are-repository-level-context-files-helpful-for-coding-agents - Develop a tool to evaluate the effectiveness of context files in coding agents' repository-level task completion.
- Accelerating Robotic Reinforcement Learning with Agent Guidance (viability: 7): https://sciencetostartup.com/paper/accelerating-robotic-reinforcement-learning-with-agent-guidance - AGPS automates robotic RL training by using an agent for precise guidance, increasing sample efficiency without human supervisors.
- Manifold-Aware Temporal Domain Generalization for Large Language Models (viability: 6): https://sciencetostartup.com/paper/manifold-aware-temporal-domain-generalization-for-large-language-models - MaT-LoRA enables efficient temporal adaptation of LLMs through manifold-aware low-rank reparameterization.
- Gaia2: Benchmarking LLM Agents on Dynamic and Asynchronous Environments (viability: 6): https://sciencetostartup.com/paper/gaia2-benchmarking-llm-agents-on-dynamic-and-asynchronous-environments - Gaia2 is a benchmark for LLM agents in dynamic environments, providing a testbed for evaluation and development of real-world AI systems.
- Towards Performance-Enhanced Model-Contrastive Federated Learning using Historical Information in Heterogeneous Scenarios (viability: 4): https://sciencetostartup.com/paper/towards-performance-enhanced-model-contrastive-federated-learning-using-historical-information-in-heterogeneous-scenario - Develop a federated learning framework that enhances model performance using contrastive learning and historical information in heterogeneous environments.
- Synthesis of Late Gadolinium Enhancement Images via Implicit Neural Representations for Cardiac Scar Segmentation (viability: 7): https://sciencetostartup.com/paper/synthesis-of-late-gadolinium-enhancement-images-via-implicit-neural-representations-for-cardiac-scar-segmentation - A framework that uses implicit neural representations and diffusion models to generate synthetic LGE images for improved cardiac scar segmentation.
- IncompeBench: A Permissively Licensed, Fine-Grained Benchmark for Music Information Retrieval (viability: 5): https://sciencetostartup.com/paper/incompebench-a-permissively-licensed-fine-grained-benchmark-for-music-information-retrieval - Develop a robust benchmark tool for evaluating Music Information Retrieval systems using a new fine-grained dataset.
- AdaptEvolve: Improving Efficiency of Evolutionary AI Agents through Adaptive Model Selection (viability: 8): https://sciencetostartup.com/paper/adaptevolve-improving-efficiency-of-evolutionary-ai-agents-through-adaptive-model-selection - AdaptEvolve optimizes AI agent efficiency by dynamically selecting the best-suited LLM for each decision point, cutting inference costs by 37.9%.
- Who Does What? Archetypes of Roles Assigned to LLMs During Human-AI Decision-Making (viability: 2): https://sciencetostartup.com/paper/who-does-what-archetypes-of-roles-assigned-to-llms-during-human-ai-decision-making - Develop insights on human-LLM interaction archetypes impacting decision-making outcomes in high-stakes environments.
- DynaHOI: Benchmarking Hand-Object Interaction for Dynamic Target (viability: 5): https://sciencetostartup.com/paper/dynahoi-benchmarking-hand-object-interaction-for-dynamic-target - DynaHOI offers a comprehensive platform for simulating and benchmarking dynamic hand-object interactions in diverse categories.
- MEME: Modeling the Evolutionary Modes of Financial Markets (viability: 7): https://sciencetostartup.com/paper/meme-modeling-the-evolutionary-modes-of-financial-markets - Develop MEME, a financial market modeling tool that uses evolving investment logics for robust portfolio construction insights.
- When Should LLMs Be Less Specific? Selective Abstraction for Reliable Long-Form Text Generation (viability: 6): https://sciencetostartup.com/paper/when-should-llms-be-less-specific-selective-abstraction-for-reliable-long-form-text-generation - Selective Abstraction framework enhances LLM reliability by reducing specificity of uncertain content without losing meaningful information.
- Leveraging LLMs to support co-evolution between definitions and instances of textual DSLs: A Systematic Evaluation (viability: 3): https://sciencetostartup.com/paper/leveraging-llms-to-support-co-evolution-between-definitions-and-instances-of-textual-dsls-a-systematic-evaluation - Explore LLMs for enhancing language co-evolution in textual DSLs to maintain human-relevant information.
- Mitigating Mismatch within Reference-based Preference Optimization (viability: 2): https://sciencetostartup.com/paper/mitigating-mismatch-within-reference-based-preference-optimization - Develop HyPO, an algorithm improving preference alignment by conditionally debiasing reference signals in Direct Preference Optimization.
- Agentic AI for Cybersecurity: A Meta-Cognitive Architecture for Governable Autonomy (viability: 4): https://sciencetostartup.com/paper/agentic-ai-for-cybersecurity-a-meta-cognitive-architecture-for-governable-autonomy - Develop a meta-cognitive architecture for autonomous and governable AI agents in cybersecurity.
- Where Bits Matter in World Model Planning: A Paired Mixed-Bit Study for Efficient Spatial Reasoning (viability: 5): https://sciencetostartup.com/paper/where-bits-matter-in-world-model-planning-a-paired-mixed-bit-study-for-efficient-spatial-reasoning - Develop module-aware, low-bit quantization strategies for efficient AI spatial reasoning.
- From Atoms to Trees: Building a Structured Feature Forest with Hierarchical Sparse Autoencoders (viability: 4): https://sciencetostartup.com/paper/from-atoms-to-trees-building-a-structured-feature-forest-with-hierarchical-sparse-autoencoders - Introducing HSAE, a scalable tool for building and analyzing hierarchical conceptual structures in LLMs using sparse autoencoders.
- SynthRAR: Ring Artifacts Reduction in CT with Unrolled Network and Synthetic Data Training (viability: 6): https://sciencetostartup.com/paper/synthrar-ring-artifacts-reduction-in-ct-with-unrolled-network-and-synthetic-data-training - Develop a software tool for reducing ring artifacts in CT images using an unrolled network trained on synthetic data.
- Towards Fair and Comprehensive Evaluation of Routers in Collaborative LLM Systems (viability: 4): https://sciencetostartup.com/paper/towards-fair-and-comprehensive-evaluation-of-routers-in-collaborative-llm-systems - RouterXBench provides a robust evaluation framework for local and cloud-based LLM systems, enhancing router performance with the new ProbeDirichlet method.
- Intelligent AI Delegation (viability: 3): https://sciencetostartup.com/paper/intelligent-ai-delegation - Develop an adaptive framework for AI delegation to improve task decomposition and trust in complex networks.
- Talk2DM: Enabling Natural Language Querying and Commonsense Reasoning for Vehicle-Road-Cloud Integrated Dynamic Maps with Large Language Models (viability: 8): https://sciencetostartup.com/paper/talk2dm-enabling-natural-language-querying-and-commonsense-reasoning-for-vehicle-road-cloud-integrated-dynamic-maps-with - Talk2DM offers an advanced natural language interface to enhance vehicle-road-cloud dynamic map interaction for autonomous driving systems.
- Zooming without Zooming: Region-to-Image Distillation for Fine-Grained Multimodal Perception (viability: 9): https://sciencetostartup.com/paper/zooming-without-zooming-region-to-image-distillation-for-fine-grained-multimodal-perception - Region-to-Image Distillation for improving fine-grained multimodal perception in MLLMs.
- Prototype Transformer: Towards Language Model Architectures Interpretable by Design (viability: 3): https://sciencetostartup.com/paper/prototype-transformer-towards-language-model-architectures-interpretable-by-design - Develop an interpretable language model architecture for enhanced reasoning transparency.
- Resource-Aware Deployment Optimization for Collaborative Intrusion Detection in Layered Networks (viability: 6): https://sciencetostartup.com/paper/resource-aware-deployment-optimization-for-collaborative-intrusion-detection-in-layered-networks - Deploy a resource-aware collaborative intrusion detection system for dynamic distributed environments.
- Improving Neural Retrieval with Attribution-Guided Query Rewriting (viability: 3): https://sciencetostartup.com/paper/improving-neural-retrieval-with-attribution-guided-query-rewriting - Enhance neural retrievals by using attribution-guided query rewriting to mitigate query ambiguity and improve search accuracy.
- ULTRA:Urdu Language Transformer-based Recommendation Architecture (viability: 7): https://sciencetostartup.com/paper/ultra-urdu-language-transformer-based-recommendation-architecture - Build an advanced semantic content recommendation system for low-resource languages like Urdu using ULTRA's transformer-based architecture.
- Revis: Sparse Latent Steering to Mitigate Object Hallucination in Large Vision-Language Models (viability: 6): https://sciencetostartup.com/paper/revis-sparse-latent-steering-to-mitigate-object-hallucination-in-large-vision-language-models - A framework to reduce object hallucination in vision-language models by reactive visual feature extraction.
- Predicting LLM Output Length via Entropy-Guided Representations (viability: 5): https://sciencetostartup.com/paper/predicting-llm-output-length-via-entropy-guided-representations - Optimize LLM inference efficiency with a novel framework leveraging entropy-guided predictions.
- PuYun-LDM: A Latent Diffusion Model for High-Resolution Ensemble Weather Forecasts (viability: 3): https://sciencetostartup.com/paper/puyun-ldm-a-latent-diffusion-model-for-high-resolution-ensemble-weather-forecasts - Improve high-resolution ensemble weather forecasts with PuYun-LDM using 3D Masked AutoEncoder and Variable-Aware Masked Frequency Modeling.
- Hi-SAM: A Hierarchical Structure-Aware Multi-modal Framework for Large-Scale Recommendation (viability: 8): https://sciencetostartup.com/paper/hi-sam-a-hierarchical-structure-aware-multi-modal-framework-for-large-scale-recommendation - Hi-SAM leverages multi-modal data to enhance large-scale recommendation systems for improved user engagement.
- Detecting RLVR Training Data via Structural Convergence of Reasoning (viability: 5): https://sciencetostartup.com/paper/detecting-rlvr-training-data-via-structural-convergence-of-reasoning - Develop a scalable solution for detecting RLVR-trained data using structural convergence signatures.
- Beyond End-to-End Video Models: An LLM-Based Multi-Agent System for Educational Video Generation (viability: 8): https://sciencetostartup.com/paper/beyond-end-to-end-video-models-an-llm-based-multi-agent-system-for-educational-video-generation - LASEV is a modular AI platform for automated, high-fidelity educational video production, with a 95% cost reduction.
- Evaluating LLM Safety Under Repeated Inference via Accelerated Prompt Stress Testing (viability: 5): https://sciencetostartup.com/paper/evaluating-llm-safety-under-repeated-inference-via-accelerated-prompt-stress-testing - Develop Accelerated Prompt Stress Testing (APST) to evaluate LLM safety and reliability under repeated inference, complementing existing benchmarks.
- Safe Fairness Guarantees Without Demographics in Classification: Spectral Uncertainty Set Perspective (viability: 4): https://sciencetostartup.com/paper/safe-fairness-guarantees-without-demographics-in-classification-spectral-uncertainty-set-perspective - Develop a fairness-enhancing classifier that adjusts the spectrum of Fourier feature mapping without demographic data.
- FlowMind: Execute-Summarize for Structured Workflow Generation from LLM Reasoning (viability: 3): https://sciencetostartup.com/paper/flowmind-execute-summarize-for-structured-workflow-generation-from-llm-reasoning - Develop a framework to convert LLM reasoning into accurate, structured workflows using an Execute-Summarize approach.
- RELATE: A Reinforcement Learning-Enhanced LLM Framework for Advertising Text Generation (viability: 9): https://sciencetostartup.com/paper/relate-a-reinforcement-learning-enhanced-llm-framework-for-advertising-text-generation - A reinforcement learning-powered framework for optimizing advertising text generation in real-time ad platforms.
- How to Optimize Multispecies Set Predictions in Presence-Absence Modeling ? (viability: 1): https://sciencetostartup.com/paper/how-to-optimize-multispecies-set-predictions-in-presence-absence-modeling - Develops MaxExp and SSE methods for improved binarization in species distribution models, enhancing ecological predictions.
- TSR: Trajectory-Search Rollouts for Multi-Turn RL of LLM Agents (viability: 3): https://sciencetostartup.com/paper/tsr-trajectory-search-rollouts-for-multi-turn-rl-of-llm-agents - Develop a training-time approach to improve reinforcement learning for multi-turn agent interactions using trajectory-search rollouts.
- MiniCPM-SALA: Hybridizing Sparse and Linear Attention for Efficient Long-Context Modeling (viability: 5): https://sciencetostartup.com/paper/minicpm-sala-hybridizing-sparse-and-linear-attention-for-efficient-long-context-modeling - Transform long-context modeling with MiniCPM-SALA, a cost-efficient hybrid attention framework reducing memory and computational demands.
- Cooperation Breakdown in LLM Agents Under Communication Delays (viability: 3): https://sciencetostartup.com/paper/cooperation-breakdown-in-llm-agents-under-communication-delays - Develop a framework to improve cooperation in multi-agent systems under communication delays.
- AmbiBench: Benchmarking Mobile GUI Agents Beyond One-Shot Instructions in the Wild (viability: 4): https://sciencetostartup.com/paper/ambibench-benchmarking-mobile-gui-agents-beyond-one-shot-instructions-in-the-wild - Develop AmbiBench, a benchmark to evaluate mobile GUI agents on handling ambiguous user instructions through interactive intent alignment.
- Text2GQL-Bench: A Text to Graph Query Language Benchmark [Experiment, Analysis & Benchmark] (viability: 5): https://sciencetostartup.com/paper/text2gql-bench-a-text-to-graph-query-language-benchmark-experiment-analysis-benchmark - Text2GQL-Bench offers a benchmark for evaluating Text-to-Graph-Query-Language systems across multiple domains and query languages, enhancing graph data analysis capabilities.
- Adapting Vision-Language Models for E-commerce Understanding at Scale (viability: 3): https://sciencetostartup.com/paper/adapting-vision-language-models-for-e-commerce-understanding-at-scale - Enhance Vision-Language Model adaptability for improved e-commerce product understanding.
- Cross-Architecture Model Diffing with Crosscoders: Unsupervised Discovery of Differences Between LLMs (viability: 2): https://sciencetostartup.com/paper/cross-architecture-model-diffing-with-crosscoders-unsupervised-discovery-of-differences-between-llms - This paper explores cross-architecture model diffing to identify behavioral differences in AI models.
- Beyond Parameter Arithmetic: Sparse Complementary Fusion for Distribution-Aware Model Merging (viability: 6): https://sciencetostartup.com/paper/beyond-parameter-arithmetic-sparse-complementary-fusion-for-distribution-aware-model-merging - Develop Sparse Complementary Fusion to enhance model merging without retraining costs.
- Semantically Conditioned Diffusion Models for Cerebral DSA Synthesis (viability: 7): https://sciencetostartup.com/paper/semantically-conditioned-diffusion-models-for-cerebral-dsa-synthesis - Develop a semantically conditioned diffusion model to synthesize realistic cerebral DSA images for medical research and training.
- TabSieve: Explicit In-Table Evidence Selection for Tabular Prediction (viability: 5): https://sciencetostartup.com/paper/tabsieve-explicit-in-table-evidence-selection-for-tabular-prediction - TabSieve improves tabular predictions using explicit evidence selection for enhanced performance and robustness.
- OMEGA-Avatar: One-shot Modeling of 360° Gaussian Avatars (viability: 7): https://sciencetostartup.com/paper/omega-avatar-one-shot-modeling-of-360-gaussian-avatars - Develop high-fidelity, animatable 3D avatars from a single image with state-of-the-art performance using OMEGA-Avatar.
- ANML: Attribution-Native Machine Learning with Guaranteed Robustness (viability: 5): https://sciencetostartup.com/paper/anml-attribution-native-machine-learning-with-guaranteed-robustness - ANML is a machine learning framework optimizing data attribute weighting to enhance model performance and attribution robustness.
- DRACO: a Cross-Domain Benchmark for Deep Research Accuracy, Completeness, and Objectivity (viability: 5): https://sciencetostartup.com/paper/draco-a-cross-domain-benchmark-for-deep-research-accuracy-completeness-and-objectivity - DRACO is a cross-domain benchmark for evaluating deep research tasks across multiple dimensions.
- PatientHub: A Unified Framework for Patient Simulation (viability: 7): https://sciencetostartup.com/paper/patienthub-a-unified-framework-for-patient-simulation - PatientHub is a modular framework that standardizes patient simulation for improved counselor training and therapeutic assessments.
- ThinkRouter: Efficient Reasoning via Routing Thinking between Latent and Discrete Spaces (viability: 6): https://sciencetostartup.com/paper/thinkrouter-efficient-reasoning-via-routing-thinking-between-latent-and-discrete-spaces - ThinkRouter is an inference-time confidence-aware routing mechanism improving reasoning efficiency in large models by switching between latent and discrete spaces.
- Beyond Pixels: Vector-to-Graph Transformation for Reliable Schematic Auditing (viability: 7): https://sciencetostartup.com/paper/beyond-pixels-vector-to-graph-transformation-for-reliable-schematic-auditing - V2G transforms CAD diagrams into graph-based representations for accurate and machine-auditable schematic evaluations.
- Right for the Wrong Reasons: Epistemic Regret Minimization for Causal Rung Collapse in LLMs (viability: 3): https://sciencetostartup.com/paper/right-for-the-wrong-reasons-epistemic-regret-minimization-for-causal-rung-collapse-in-llms - Develop a framework to address reasoning errors in machine learning, enabling models to correct causal misalignments.
- Benchmark Health Index: A Systematic Framework for Benchmarking the Benchmarks of LLMs (viability: 4): https://sciencetostartup.com/paper/benchmark-health-index-a-systematic-framework-for-benchmarking-the-benchmarks-of-llms - Benchmark Health Index is a data-driven framework auditing LLM evaluation benchmarks for reliability and longevity.
- PhyNiKCE: A Neurosymbolic Agentic Framework for Autonomous Computational Fluid Dynamics (viability: 6): https://sciencetostartup.com/paper/phynikce-a-neurosymbolic-agentic-framework-for-autonomous-computational-fluid-dynamics - Develop a neurosymbolic framework that ensures robust and efficient autonomous Computational Fluid Dynamics by enforcing physical constraints.
- Quark Medical Alignment: A Holistic Multi-Dimensional Alignment and Collaborative Optimization Paradigm (viability: 5): https://sciencetostartup.com/paper/quark-medical-alignment-a-holistic-multi-dimensional-alignment-and-collaborative-optimization-paradigm - Quark Medical Alignment offers a new paradigm for optimizing medical question answering systems by balancing correctness, safety, and compliance using a holistic alignment matrix and adaptive weighting strategy.
- SToRM: Supervised Token Reduction for Multi-modal LLMs toward efficient end-to-end autonomous driving (viability: 7): https://sciencetostartup.com/paper/storm-supervised-token-reduction-for-multi-modal-llms-toward-efficient-end-to-end-autonomous-driving - SToRM offers a token reduction framework for efficient multi-modal LLMs in autonomous driving, promising reduced computational costs without sacrificing performance.
- LoRA-based Parameter-Efficient LLMs for Continuous Learning in Edge-based Malware Detection (viability: 7): https://sciencetostartup.com/paper/lora-based-parameter-efficient-llms-for-continuous-learning-in-edge-based-malware-detection - A parameter-efficient continuous learning solution for deploying LLMs in edge-based malware detection.
- DMind-3: A Sovereign Edge--Local--Cloud AI System with Controlled Deliberation and Correction-Based Tuning for Safe, Low-Latency Transaction Execution (viability: 5): https://sciencetostartup.com/paper/dmind-3-a-sovereign-edge-local-cloud-ai-system-with-controlled-deliberation-and-correction-based-tuning-for-safe-low-lat - DMind-3 provides a multi-layered AI system optimizing security and latency for Web3 financial transactions.
- Brain Tumor Classifiers Under Attack: Robustness of ResNet Variants Against Transferable FGSM and PGD Attacks (viability: 5): https://sciencetostartup.com/paper/brain-tumor-classifiers-under-attack-robustness-of-resnet-variants-against-transferable-fgsm-and-pgd-attacks - Develop robust classifiers for brain tumor detection that withstand adversarial attacks, improving clinical safety of MRI analysis.
- ViTaS: Visual Tactile Soft Fusion Contrastive Learning for Visuomotor Learning (viability: 8): https://sciencetostartup.com/paper/vitas-visual-tactile-soft-fusion-contrastive-learning-for-visuomotor-learning - ViTaS enhances robotic manipulation by fusing visual and tactile data using Soft Fusion Contrastive Learning for improved performance in occluded environments.
- Variation-aware Flexible 3D Gaussian Editing (viability: 5): https://sciencetostartup.com/paper/variation-aware-flexible-3d-gaussian-editing - VF-Editor allows for efficient and flexible direct editing of 3D Gaussians by leveraging 2D editing knowledge.
- ScalSelect: Scalable Training-Free Multimodal Data Selection for Efficient Visual Instruction Tuning (viability: 8): https://sciencetostartup.com/paper/scalselect-scalable-training-free-multimodal-data-selection-for-efficient-visual-instruction-tuning - ScalSelect offers an efficient data selection tool that reduces training costs for vision-language models by 84% without sacrificing performance, making it ideal for scalable Visual Instruction Tuning.
- Do MLLMs Really Understand Space? A Mathematical Reasoning Evaluation (viability: 5): https://sciencetostartup.com/paper/do-mllms-really-understand-space-a-mathematical-reasoning-evaluation - Evaluate and enhance spatial reasoning in MLLMs with MathSpatial, a comprehensive framework including benchmarks, datasets, and reasoning models.
- Neuro-Symbolic Multitasking: A Unified Framework for Discovering Generalizable Solutions to PDE Families (viability: 4): https://sciencetostartup.com/paper/neuro-symbolic-multitasking-a-unified-framework-for-discovering-generalizable-solutions-to-pde-families - Develop a neuro-symbolic framework to efficiently and interpretably solve PDE families, improving accuracy by 35.7%.
- ArGEnT: Arbitrary Geometry-encoded Transformer for Operator Learning (viability: 6): https://sciencetostartup.com/paper/argent-arbitrary-geometry-encoded-transformer-for-operator-learning - Build a geometry-aware transformer for scalable surrogate modeling in scientific machine learning.
- PLOT-CT: Pre-log Voronoi Decomposition Assisted Generation for Low-dose CT Reconstruction (viability: 5): https://sciencetostartup.com/paper/plot-ct-pre-log-voronoi-decomposition-assisted-generation-for-low-dose-ct-reconstruction - PLOT-CT utilizes Voronoi decomposition to enhance low-dose CT reconstruction by improving noise reduction and structural preservation.
- When Agents Disagree With Themselves: Measuring Behavioral Consistency in LLM-Based Agents (viability: 3): https://sciencetostartup.com/paper/when-agents-disagree-with-themselves-measuring-behavioral-consistency-in-llm-based-agents - This paper offers insights into behavioral consistency of LLM-based agents like GPT-4o and suggests that consistency monitoring could enhance reliability.
- scPilot: Large Language Model Reasoning Toward Automated Single-Cell Analysis and Discovery (viability: 8): https://sciencetostartup.com/paper/scpilot-large-language-model-reasoning-toward-automated-single-cell-analysis-and-discovery - scPilot automates single-cell RNA-seq data analysis through LLM-driven reasoning, enhancing accuracy and interpretability in bioinformatics.
- ABot-N0: Technical Report on the VLA Foundation Model for Versatile Embodied Navigation (viability: 6): https://sciencetostartup.com/paper/abot-n0-technical-report-on-the-vla-foundation-model-for-versatile-embodied-navigation - A unified Vision-Language-Action model for versatile embodied navigation that achieves state-of-the-art performance across multiple benchmarks.
- MAPLE: Modality-Aware Post-training and Learning Ecosystem (viability: 7): https://sciencetostartup.com/paper/maple-modality-aware-post-training-and-learning-ecosystem - MAPLE offers a modality-aware ecosystem to enhance post-training in multimodal RL by optimizing signal-specific learning strategies for faster convergence and improved robustness.
- Gradient Compression May Hurt Generalization: A Remedy by Synthetic Data Guided Sharpness Aware Minimization (viability: 1): https://sciencetostartup.com/paper/gradient-compression-may-hurt-generalization-a-remedy-by-synthetic-data-guided-sharpness-aware-minimization - Improving federated learning generalization using synthetic data for sharpness-aware optimization.
- The Five Ws of Multi-Agent Communication: Who Talks to Whom, When, What, and Why -- A Survey from MARL to Emergent Language and LLMs (viability: 3): https://sciencetostartup.com/paper/the-five-ws-of-multi-agent-communication-who-talks-to-whom-when-what-and-why-a-survey-from-marl-to-emergent-language-and - Surveying the evolution of multi-agent communication systems, from MARL to LLMs, to design scalable collaborative AI.
- Analytical Search (viability: 2): https://sciencetostartup.com/paper/analytical-search - Analytical Search proposes a new search paradigm for fulfilling analytical information needs through evidence-governed, process-oriented workflow.
- ReaDy-Go: Real-to-Sim Dynamic 3D Gaussian Splatting Simulation for Environment-Specific Visual Navigation with Moving Obstacles (viability: 8): https://sciencetostartup.com/paper/ready-go-real-to-sim-dynamic-3d-gaussian-splatting-simulation-for-environment-specific-visual-navigation-with-moving-obs - Develop a real-to-sim simulation tool for robust visual navigation in dynamic environments like households and factories.
- Learning to Configure Agentic AI Systems (viability: 7): https://sciencetostartup.com/paper/learning-to-configure-agentic-ai-systems - Introduce dynamic AI system configurations using ARC to outperform traditional templates by optimizing token usage and accuracy.
- SemaPop: Semantic-Persona Conditioned Population Synthesis (viability: 5): https://sciencetostartup.com/paper/semapop-semantic-persona-conditioned-population-synthesis - SemaPop offers semantic-statistical population synthesis using LLMs to enhance generative performance in socio-economic simulations.
- Perception-based Image Denoising via Generative Compression (viability: 5): https://sciencetostartup.com/paper/perception-based-image-denoising-via-generative-compression - Develop a generative compression framework for enhanced perceptual image denoising.
- TS-Memory: Plug-and-Play Memory for Time Series Foundation Models (viability: 7): https://sciencetostartup.com/paper/ts-memory-plug-and-play-memory-for-time-series-foundation-models - Plug-and-play memory adapter for enhancing Time Series Foundation Models without increasing inference latency.
- Native Reasoning Models: Training Language Models to Reason on Unverifiable Data (viability: 6): https://sciencetostartup.com/paper/native-reasoning-models-training-language-models-to-reason-on-unverifiable-data - NRT offers a novel framework for enhancing AI reasoning abilities without relying on costly, expert-verified data.
- Budget-Constrained Agentic Large Language Models: Intention-Based Planning for Costly Tool Use (viability: 3): https://sciencetostartup.com/paper/budget-constrained-agentic-large-language-models-intention-based-planning-for-costly-tool-use - Develop intention-based planning frameworks for budget-constrained tool-augmented LLMs.
- Krause Synchronization Transformers (viability: 7): https://sciencetostartup.com/paper/krause-synchronization-transformers - Krause Synchronization Transformers offer a scalable alternative to traditional self-attention by reducing runtime complexity and preventing representation collapse.
- AltTS: A Dual-Path Framework with Alternating Optimization for Multivariate Time Series Forecasting (viability: 7): https://sciencetostartup.com/paper/altts-a-dual-path-framework-with-alternating-optimization-for-multivariate-time-series-forecasting - ALTTS offers a dual-path framework enhancing accuracy in long-horizon multivariate time series forecasting by isolating autoregressive and cross-dimension dynamics.
- Stop Tracking Me! Proactive Defense Against Attribute Inference Attack in LLMs (viability: 6): https://sciencetostartup.com/paper/stop-tracking-me-proactive-defense-against-attribute-inference-attack-in-llms - Develop a proactive defense framework that minimizes attribute inference attacks in LLMs, ensuring user privacy.
- CausalAgent: A Conversational Multi-Agent System for End-to-End Causal Inference (viability: 7): https://sciencetostartup.com/paper/causalagent-a-conversational-multi-agent-system-for-end-to-end-causal-inference - CausalAgent simplifies causal inference through a conversational multi-agent system that automates complex workflows from dataset input to report generation.
- Adaptive Milestone Reward for GUI Agents (viability: 6): https://sciencetostartup.com/paper/adaptive-milestone-reward-for-gui-agents - Develop an adaptive milestone reward system for improving mobile GUI agents' success rate in reinforcement learning tasks.
- Human-Inspired Continuous Learning of Internal Reasoning Processes: Learning How to Think for Adaptive AI Systems (viability: 3): https://sciencetostartup.com/paper/human-inspired-continuous-learning-of-internal-reasoning-processes-learning-how-to-think-for-adaptive-ai-systems - Develop an adaptive AI system that continuously learns and refines its internal reasoning processes for enhanced cognitive flexibility and efficiency.
- How Smart Is Your GUI Agent? A Framework for the Future of Software Interaction (viability: 4): https://sciencetostartup.com/paper/how-smart-is-your-gui-agent-a-framework-for-the-future-of-software-interaction - Develop a standardized framework to measure and improve GUI agents' autonomy and trustworthiness across platforms.
- Differentially Private and Communication Efficient Large Language Model Split Inference via Stochastic Quantization and Soft Prompt (viability: 5): https://sciencetostartup.com/paper/differentially-private-and-communication-efficient-large-language-model-split-inference-via-stochastic-quantization-and- - Develop a differentially private and communication efficient LLM inference framework for resource-constrained devices.
- AgentLeak: A Full-Stack Benchmark for Privacy Leakage in Multi-Agent LLM Systems (viability: 7): https://sciencetostartup.com/paper/agentleak-a-full-stack-benchmark-for-privacy-leakage-in-multi-agent-llm-systems - AgentLeak offers a benchmark for assessing privacy leakage risks in multi-agent LLM systems, crucial for safeguarding sensitive inter-agent communications.
- Multimodal Fact-Level Attribution for Verifiable Reasoning (viability: 6): https://sciencetostartup.com/paper/multimodal-fact-level-attribution-for-verifiable-reasoning - MuRGAt provides a benchmark and automatic evaluation tool for verifying multimodal reasoning and attribution in AI models.
- RooflineBench: A Benchmarking Framework for On-Device LLMs via Roofline Analysis (viability: 5): https://sciencetostartup.com/paper/rooflinebench-a-benchmarking-framework-for-on-device-llms-via-roofline-analysis - Optimize on-device LLM efficiency using a Roofline analysis benchmarking framework.
- Understanding Persuasive Interactions between Generative Social Agents and Humans: The Knowledge-based Persuasion Model (KPM) (viability: 2): https://sciencetostartup.com/paper/understanding-persuasive-interactions-between-generative-social-agents-and-humans-the-knowledge-based-persuasion-model-k - Develop a theoretical framework to guide ethical and persuasive interactions between generative social agents and humans.
- Compiler-Guided Inference-Time Adaptation: Improving GPT-5 Programming Performance in Idris (viability: 6): https://sciencetostartup.com/paper/compiler-guided-inference-time-adaptation-improving-gpt-5-programming-performance-in-idris - Leverage GPT-5's enhanced inference-time adaptation capabilities to boost programming performance in low-resource languages like Idris.
- EM-Aware Physical Synthesis: Neural Inductor Modeling and Intelligent Placement & Routing for RF Circuits (viability: 7): https://sciencetostartup.com/paper/em-aware-physical-synthesis-neural-inductor-modeling-and-intelligent-placement-routing-for-rf-circuits - AI-driven framework for automatic synthesis of RF circuit layouts ensuring manufacturability and design-rule compliance.
- Credit Where It is Due: Cross-Modality Connectivity Drives Precise Reinforcement Learning for MLLM Reasoning (viability: 3): https://sciencetostartup.com/paper/credit-where-it-is-due-cross-modality-connectivity-drives-precise-reinforcement-learning-for-mllm-reasoning - Develop a lightweight framework enhancing multimodal language model reasoning through selective token reinforcement.
- From Noise to Order: Learning to Rank via Denoising Diffusion (viability: 4): https://sciencetostartup.com/paper/from-noise-to-order-learning-to-rank-via-denoising-diffusion - Develop a diffusion-based generative ranking model for improved relevance in information retrieval systems.
- Enhanced Portable Ultra Low-Field Diffusion Tensor Imaging with Bayesian Artifact Correction and Deep Learning-Based Super-Resolution (viability: 9): https://sciencetostartup.com/paper/enhanced-portable-ultra-low-field-diffusion-tensor-imaging-with-bayesian-artifact-correction-and-deep-learning-based-sup - Develops a portable ultra-low-field MRI enhancement tool for improved neuroimaging quality with Bayesian and super-resolution techniques.
- Towards Reliable Machine Translation: Scaling LLMs for Critical Error Detection and Safety (viability: 5): https://sciencetostartup.com/paper/towards-reliable-machine-translation-scaling-llms-for-critical-error-detection-and-safety - Develop a robust error detection system for machine translation to improve safety and fairness in multilingual AI.
- Distributionally Robust Cooperative Multi-Agent Reinforcement Learning via Robust Value Factorization (viability: 6): https://sciencetostartup.com/paper/distributionally-robust-cooperative-multi-agent-reinforcement-learning-via-robust-value-factorization - Enhance multi-agent systems with robust decision-making for real-world uncertainties using DrIGM principle.
- Fighting MRI Anisotropy: Learning Multiple Cardiac Shapes From a Single Implicit Neural Representation (viability: 6): https://sciencetostartup.com/paper/fighting-mri-anisotropy-learning-multiple-cardiac-shapes-from-a-single-implicit-neural-representation - Develop a tool using neural implicit functions to enhance cardiac MRI analysis by reconstructing accurate cardiac shapes from existing data.
- Gradients Must Earn Their Influence: Unifying SFT with Generalized Entropic Objectives (viability: 4): https://sciencetostartup.com/paper/gradients-must-earn-their-influence-unifying-sft-with-generalized-entropic-objectives - Dynamic Entropy Fine-Tuning (DEFT) offers a novel approach to improve supervised fine-tuning of models by dynamically adjusting token-level weighting according to predictive distribution concentration.
- TRACER: Trajectory Risk Aggregation for Critical Episodes in Agentic Reasoning (viability: 7): https://sciencetostartup.com/paper/tracer-trajectory-risk-aggregation-for-critical-episodes-in-agentic-reasoning - TRACER offers a trajectory-level uncertainty metric for detecting AI agent failures in complex conversational interactions.
- GHOST: Unmasking Phantom States in Mamba2 via Grouped Hidden-state Output-aware Selection & Truncation (viability: 5): https://sciencetostartup.com/paper/ghost-unmasking-phantom-states-in-mamba2-via-grouped-hidden-state-output-aware-selection-truncation - Develop a framework for efficient model pruning that reduces state dimensions by 50% with minimal performance loss.
- Can We Really Learn One Representation to Optimize All Rewards? (viability: 5): https://sciencetostartup.com/paper/can-we-really-learn-one-representation-to-optimize-all-rewards - Develop a pre-trained RL policy representation method for improved zero-shot performance.
- General and Efficient Steering of Unconditional Diffusion (viability: 6): https://sciencetostartup.com/paper/general-and-efficient-steering-of-unconditional-diffusion - An efficient framework for fast, controllable generation in diffusion models by leveraging noise alignment and concept vectors without requiring gradients during inference.
- Causal-JEPA: Learning World Models through Object-Level Latent Interventions (viability: 9): https://sciencetostartup.com/paper/causal-jepa-learning-world-models-through-object-level-latent-interventions - C-JEPA offers an efficient object-centric world model enhancing visual question answering and agent control with latent interventions.
- Retrieval-Aware Distillation for Transformer-SSM Hybrids (viability: 6): https://sciencetostartup.com/paper/retrieval-aware-distillation-for-transformer-ssm-hybrids - Transform Transformer models into memory-efficient hybrids that maintain retrieval capabilities, using fewer attention heads.
- The Manifold of the Absolute: Religious Perennialism as Generative Inference (viability: 1): https://sciencetostartup.com/paper/the-manifold-of-the-absolute-religious-perennialism-as-generative-inference - Leverage variational autoencoders to explore religious epistemology through a formalized perennialist lens.
- The Energy of Falsehood: Detecting Hallucinations via Diffusion Model Likelihoods (viability: 4): https://sciencetostartup.com/paper/the-energy-of-falsehood-detecting-hallucinations-via-diffusion-model-likelihoods - DiffuTruth leverages diffusion model likelihoods to detect hallucinations in LLMs.
- Finding the Cracks: Improving LLMs Reasoning with Paraphrastic Probing and Consistency Verification (viability: 5): https://sciencetostartup.com/paper/finding-the-cracks-improving-llms-reasoning-with-paraphrastic-probing-and-consistency-verification - Enhance LLM reasoning accuracy with critical token identification through PPCV framework.
- Bootstrapping-based Regularisation for Reducing Individual Prediction Instability in Clinical Risk Prediction Models (viability: 8): https://sciencetostartup.com/paper/bootstrapping-based-regularisation-for-reducing-individual-prediction-instability-in-clinical-risk-prediction-models - Develop a regularisation framework for stabilizing clinical prediction models' outputs, enhancing reliability and interpretability in healthcare.
- When Models Examine Themselves: Vocabulary-Activation Correspondence in Self-Referential Processing (viability: 2): https://sciencetostartup.com/paper/when-models-examine-themselves-vocabulary-activation-correspondence-in-self-referential-processing - Exploring self-referential vocabulary tracks within language models to understand their internal computational states.
- ReplicatorBench: Benchmarking LLM Agents for Replicability in Social and Behavioral Sciences (viability: 7): https://sciencetostartup.com/paper/replicatorbench-benchmarking-llm-agents-for-replicability-in-social-and-behavioral-sciences - ReplicatorBench offers a benchmark for evaluating LLM agents' ability to replicate scientific research in social sciences.
- Pushing Forward Pareto Frontiers of Proactive Agents with Behavioral Agentic Optimization (viability: 7): https://sciencetostartup.com/paper/pushing-forward-pareto-frontiers-of-proactive-agents-with-behavioral-agentic-optimization - Develop a reinforcement learning framework, BAO, for training proactive LLM agents that efficiently balance task performance with user engagement.
- AgentNoiseBench: Benchmarking Robustness of Tool-Using LLM Agents Under Noisy Condition (viability: 5): https://sciencetostartup.com/paper/agentnoisebench-benchmarking-robustness-of-tool-using-llm-agents-under-noisy-condition - AgentNoiseBench provides a framework to evaluate the robustness of language model agents in noisy environments, optimizing their real-world performance.
- Divide and Learn: Multi-Objective Combinatorial Optimization at Scale (viability: 5): https://sciencetostartup.com/paper/divide-and-learn-multi-objective-combinatorial-optimization-at-scale - Optimize multi-objective combinatorial problems efficiently using bandit optimization over decomposed decision spaces.
- Situated, Dynamic, and Subjective: Envisioning the Design of Theory-of-Mind-Enabled Everyday AI with Industry Practitioners (viability: 3): https://sciencetostartup.com/paper/situated-dynamic-and-subjective-envisioning-the-design-of-theory-of-mind-enabled-everyday-ai-with-industry-practitioners - Designing Theory-of-Mind-enabled AI products with improved social and dynamic context inference.
- Bi-Level Prompt Optimization for Multimodal LLM-as-a-Judge (viability: 5): https://sciencetostartup.com/paper/bi-level-prompt-optimization-for-multimodal-llm-as-a-judge - Develop a bi-level prompt optimization framework to enhance multimodal LLMs that serve as judges for AI-generated images.
- MolmoSpaces: A Large-Scale Open Ecosystem for Robot Navigation and Manipulation (viability: 6): https://sciencetostartup.com/paper/molmospaces-a-large-scale-open-ecosystem-for-robot-navigation-and-manipulation - MolmoSpaces is an open ecosystem providing large-scale simulation environments for robust robot training and benchmarking across diverse scenarios.
- Predictive Associative Memory: Retrieval Beyond Similarity Through Temporal Co-occurrence (viability: 5): https://sciencetostartup.com/paper/predictive-associative-memory-retrieval-beyond-similarity-through-temporal-co-occurrence - Revolutionize memory retrieval systems by leveraging temporal co-occurrence for associative recall precision.
- Dissecting Subjectivity and the "Ground Truth" Illusion in Data Annotation (viability: 3): https://sciencetostartup.com/paper/dissecting-subjectivity-and-the-ground-truth-illusion-in-data-annotation - Develop a platform for pluralistic data annotation to prioritize diversity in AI model training.
- CryptoAnalystBench: Failures in Multi-Tool Long-Form LLM Analysis (viability: 7): https://sciencetostartup.com/paper/cryptoanalystbench-failures-in-multi-tool-long-form-llm-analysis - Develop a robust benchmark and evaluation tool for LLMs in the crypto analysis domain to improve accuracy in high-stakes decision making.
- The PBSAI Governance Ecosystem: A Multi-Agent AI Reference Architecture for Securing Enterprise AI Estates (viability: 7): https://sciencetostartup.com/paper/the-pbsai-governance-ecosystem-a-multi-agent-ai-reference-architecture-for-securing-enterprise-ai-estates - Develop a multi-agent AI governance architecture to secure enterprise AI systems with an emphasis on traceability and human-in-the-loop processes.
- Voxtral Realtime (viability: 8): https://sciencetostartup.com/paper/voxtral-realtime - Voxtral Realtime is a low-latency automatic speech recognition model optimized for streaming applications.
- On Decision-Valued Maps and Representational Dependence (viability: 2): https://sciencetostartup.com/paper/on-decision-valued-maps-and-representational-dependence - DecisionDB is an infrastructure for auditing data representation outcomes in decision-valued maps.
- HiFloat4 Format for Language Model Inference (viability: 3): https://sciencetostartup.com/paper/hifloat4-format-for-language-model-inference - HiFloat4 is an efficient data format for reducing hardware area and power consumption in deep learning inference.
- Beyond VLM-Based Rewards: Diffusion-Native Latent Reward Modeling (viability: 5): https://sciencetostartup.com/paper/beyond-vlm-based-rewards-diffusion-native-latent-reward-modeling - Develop a diffusion-native latent reward model for more efficient and effective preference optimization in vision-language tasks.
- GENIUS: Generative Fluid Intelligence Evaluation Suite (viability: 5): https://sciencetostartup.com/paper/genius-generative-fluid-intelligence-evaluation-suite - Develop a benchmark to evaluate generative fluid intelligence in AI models.
- Data-Efficient Hierarchical Goal-Conditioned Reinforcement Learning via Normalizing Flows (viability: 5): https://sciencetostartup.com/paper/data-efficient-hierarchical-goal-conditioned-reinforcement-learning-via-normalizing-flows - Develop a data-efficient reinforcement learning framework using normalizing flows for improved policy expressivity and generalization.
- Weight Decay Improves Language Model Plasticity (viability: 4): https://sciencetostartup.com/paper/weight-decay-improves-language-model-plasticity - Optimize language model hyperparameters to enhance downstream task adaptability by focusing on weight decay's role in plasticity.
- FormalJudge: A Neuro-Symbolic Paradigm for Agentic Oversight (viability: 7): https://sciencetostartup.com/paper/formaljudge-a-neuro-symbolic-paradigm-for-agentic-oversight - Develop a neuro-symbolic framework for ensuring behavioral safety of LLM-based agents through formal verification.
- Learning to Compose for Cross-domain Agentic Workflow Generation (viability: 3): https://sciencetostartup.com/paper/learning-to-compose-for-cross-domain-agentic-workflow-generation - Automated cross-domain workflow generation for complex tasks using a decompose-recompose-decide mechanism in open-source LLMs.
- GameDevBench: Evaluating Agentic Capabilities Through Game Development (viability: 6): https://sciencetostartup.com/paper/gamedevbench-evaluating-agentic-capabilities-through-game-development - Introducing GameDevBench, a benchmark to evaluate and enhance agent capabilities in complex multimodal game development tasks.
- Safety Recovery in Reasoning Models Is Only a Few Early Steering Steps Away (viability: 7): https://sciencetostartup.com/paper/safety-recovery-in-reasoning-models-is-only-a-few-early-steering-steps-away - SafeThink provides a lightweight, inference-time defense for reasoning models, reducing safety risks without sacrificing performance.
- Direct Learning of Calibration-Aware Uncertainty for Neural PDE Surrogates (viability: 2): https://sciencetostartup.com/paper/direct-learning-of-calibration-aware-uncertainty-for-neural-pde-surrogates - Refine neural PDE surrogates by integrating calibration-aware uncertainty for better error awareness in predictive models.
- DataChef: Cooking Up Optimal Data Recipes for LLM Adaptation via Reinforcement Learning (viability: 8): https://sciencetostartup.com/paper/datachef-cooking-up-optimal-data-recipes-for-llm-adaptation-via-reinforcement-learning - DataChef automates the creation of optimized data pipelines for LLM training, enhancing model adaptation and performance through reinforcement learning.
- General Flexible $f$-divergence for Challenging Offline RL Datasets with Low Stochasticity and Diverse Behavior Policies (viability: 5): https://sciencetostartup.com/paper/general-flexible-f-divergence-for-challenging-offline-rl-datasets-with-low-stochasticity-and-diverse-behavior-policies - Improve offline RL performance using flexible $f$-divergence constraints for better dataset adaptation.
- GRASP: group-Shapley feature selection for patients (viability: 2): https://sciencetostartup.com/paper/grasp-group-shapley-feature-selection-for-patients - GRASP enhances medical feature selection through Shapley values and group regularization for better stability and accuracy.
- SteuerLLM: Local specialized large language model for German tax law analysis (viability: 8): https://sciencetostartup.com/paper/steuerllm-local-specialized-large-language-model-for-german-tax-law-analysis - SteuerLLM specializes in automating German tax law analysis using a benchmark-beating domain-adapted language model.
- In-the-Wild Model Organisms: Mitigating Undesirable Emergent Behaviors in Production LLM Post-Training via Data Attribution (viability: 5): https://sciencetostartup.com/paper/in-the-wild-model-organisms-mitigating-undesirable-emergent-behaviors-in-production-llm-post-training-via-data-attributi - Develop a tool for tracing and mitigating emergent behaviors in LLMs by identifying responsible training data using activation-based data attribution.
- Chatting with Images for Introspective Visual Thinking (viability: 8): https://sciencetostartup.com/paper/chatting-with-images-for-introspective-visual-thinking - ViLaVT enables more interactive and precise visual reasoning by dynamically integrating language guidance into vision processing.
- Conversational Behavior Modeling Foundation Model With Multi-Level Perception (viability: 6): https://sciencetostartup.com/paper/conversational-behavior-modeling-foundation-model-with-multi-level-perception - AI framework for modeling conversational behavior with multi-level perception and reasoning.
- Language Model Inversion through End-to-End Differentiation (viability: 5): https://sciencetostartup.com/paper/language-model-inversion-through-end-to-end-differentiation - Optimizing input prompts for language model outputs using end-to-end differentiation.
- Linguistic Indicators of Early Cognitive Decline in the DementiaBank Pitt Corpus: A Statistical and Machine Learning Study (viability: 7): https://sciencetostartup.com/paper/linguistic-indicators-of-early-cognitive-decline-in-the-dementiabank-pitt-corpus-a-statistical-and-machine-learning-stud - Develop an AI tool to detect early cognitive decline using linguistic markers from speech transcripts.
- Chain-of-Look Spatial Reasoning for Dense Surgical Instrument Counting (viability: 8): https://sciencetostartup.com/paper/chain-of-look-spatial-reasoning-for-dense-surgical-instrument-counting - Automated high-density surgical instrument counting using visual chain reasoning.
- RiemannGL: Riemannian Geometry Changes Graph Deep Learning (viability: 2): https://sciencetostartup.com/paper/riemanngl-riemannian-geometry-changes-graph-deep-learning - Explore intrinsic manifold structures for graph neural networks using Riemannian geometry.
- Healthy Harvests: A Comparative Look at Guava Disease Classification Using InceptionV3 (viability: 5): https://sciencetostartup.com/paper/healthy-harvests-a-comparative-look-at-guava-disease-classification-using-inceptionv3 - Develop a precise guava disease classification tool using InceptionV3 for early detection and improved crop yield.
- Can LLMs Cook Jamaican Couscous? A Study of Cultural Novelty in Recipe Generation (viability: 4): https://sciencetostartup.com/paper/can-llms-cook-jamaican-couscous-a-study-of-cultural-novelty-in-recipe-generation - Develop an LLM-based tool to generate culturally adapted recipes with enhanced understanding of cultural nuances.
- Rotary Positional Embeddings as Phase Modulation: Theoretical Bounds on the RoPE Base for Long-Context Transformers (viability: 3): https://sciencetostartup.com/paper/rotary-positional-embeddings-as-phase-modulation-theoretical-bounds-on-the-rope-base-for-long-context-transformers - Develop theoretical bounds on RoPE base parameters to optimize long-context transformer performance.
- Search or Accelerate: Confidence-Switched Position Beam Search for Diffusion Language Models (viability: 7): https://sciencetostartup.com/paper/search-or-accelerate-confidence-switched-position-beam-search-for-diffusion-language-models - SOAR enhances text generation in Diffusion Language Models by adapting decoding strategies based on confidence levels, improving quality without sacrificing speed.
- Computational Phenomenology of Temporal Experience in Autism: Quantifying the Emotional and Narrative Characteristics of Lived Unpredictability (viability: 2): https://sciencetostartup.com/paper/computational-phenomenology-of-temporal-experience-in-autism-quantifying-the-emotional-and-narrative-characteristics-of- - Integrates phenomenological and computational methods to analyze temporal experience in autistic narratives scientifically.
- What do people want to fact-check? (viability: 4): https://sciencetostartup.com/paper/what-do-people-want-to-fact-check - Analyze public fact-checking demands to align AI systems with real-world verification needs.
- Enhancing Multivariate Time Series Forecasting with Global Temporal Retrieval (viability: 7): https://sciencetostartup.com/paper/enhancing-multivariate-time-series-forecasting-with-global-temporal-retrieval - The Global Temporal Retriever (GTR) enhances any forecasting model's accuracy by capturing long-term periodic patterns with minimal computational costs.
- SynergyKGC: Reconciling Topological Heterogeneity in Knowledge Graph Completion via Topology-Aware Synergy (viability: 6): https://sciencetostartup.com/paper/synergykgc-reconciling-topological-heterogeneity-in-knowledge-graph-completion-via-topology-aware-synergy - SynergyKGC enhances knowledge graph completion by resolving structural mismatches with a topology-aware approach.
- Flow caching for autoregressive video generation (viability: 7): https://sciencetostartup.com/paper/flow-caching-for-autoregressive-video-generation - FlowCache optimizes caching for autoregressive video generation, enabling faster and efficient real-time ultra-long video synthesis.
- Beyond Confidence: The Rhythms of Reasoning in Generative Models (viability: 3): https://sciencetostartup.com/paper/beyond-confidence-the-rhythms-of-reasoning-in-generative-models - Token Constraint Bound: A new metric for assessing the robustness of large language models' predictions against context variations.
- Integrating Generative AI-enhanced Cognitive Systems in Higher Education: From Stakeholder Perceptions to a Conceptual Framework considering the EU AI Act (viability: 3): https://sciencetostartup.com/paper/integrating-generative-ai-enhanced-cognitive-systems-in-higher-education-from-stakeholder-perceptions-to-a-conceptual-fr - Conceptual framework for integrating GenAI in higher education considering stakeholder perceptions and EU compliance.
- Transport, Don't Generate: Deterministic Geometric Flows for Combinatorial Optimization (viability: 3): https://sciencetostartup.com/paper/transport-don-t-generate-deterministic-geometric-flows-for-combinatorial-optimization - CycFlow accelerates solving TSP by employing deterministic geometric flows, bypassing traditional edge scoring methods.
- Self-Supervised Image Super-Resolution Quality Assessment based on Content-Free Multi-Model Oriented Representation Learning (viability: 6): https://sciencetostartup.com/paper/self-supervised-image-super-resolution-quality-assessment-based-on-content-free-multi-model-oriented-representation-lear - Develop a no-reference image quality assessment tool using self-supervised learning for super-resolution applications in real-world scenarios.
- A Diffusion-Based Generative Prior Approach to Sparse-view Computed Tomography (viability: 2): https://sciencetostartup.com/paper/a-diffusion-based-generative-prior-approach-to-sparse-view-computed-tomography - Develop a diffusion-based generative model approach to improve sparse-view CT image reconstruction.
- Locomo-Plus: Beyond-Factual Cognitive Memory Evaluation Framework for LLM Agents (viability: 6): https://sciencetostartup.com/paper/locomo-plus-beyond-factual-cognitive-memory-evaluation-framework-for-llm-agents - LoCoMo-Plus provides a framework to evaluate cognitive memory in LLM dialogue systems beyond surface-level factual recall.
- Cross-Sectional Asset Retrieval via Future-Aligned Soft Contrastive Learning (viability: 5): https://sciencetostartup.com/paper/cross-sectional-asset-retrieval-via-future-aligned-soft-contrastive-learning - Revolutionize investment strategies with FASCL, leveraging future-oriented asset retrieval for consistent performance gains.
- Interpretable Graph-Level Anomaly Detection via Contrast with Normal Prototypes (viability: 5): https://sciencetostartup.com/paper/interpretable-graph-level-anomaly-detection-via-contrast-with-normal-prototypes - ProtoGLAD offers interpretable graph-level anomaly detection by contrasting graph anomalies with nearest normal prototypes.
- Spend Search Where It Pays: Value-Guided Structured Sampling and Optimization for Generative Recommendation (viability: 7): https://sciencetostartup.com/paper/spend-search-where-it-pays-value-guided-structured-sampling-and-optimization-for-generative-recommendation - V-STAR optimizes generative recommendation systems with value-guided sampling to outperform existing methods in accuracy and diversity.
- AugVLA-3D: Depth-Driven Feature Augmentation for Vision-Language-Action Models (viability: 7): https://sciencetostartup.com/paper/augvla-3d-depth-driven-feature-augmentation-for-vision-language-action-models - Integrate depth estimation with Vision-Language-Action models to improve robotic 3D perception and action accuracy.
- VESPO: Variational Sequence-Level Soft Policy Optimization for Stable Off-Policy LLM Training (viability: 5): https://sciencetostartup.com/paper/vespo-variational-sequence-level-soft-policy-optimization-for-stable-off-policy-llm-training - VESPO offers a solution to stabilize off-policy LLM training with variational sequence-level policy optimization, reducing variance issues.
- Mitigating Reward Hacking in RLHF via Bayesian Non-negative Reward Modeling (viability: 5): https://sciencetostartup.com/paper/mitigating-reward-hacking-in-rlhf-via-bayesian-non-negative-reward-modeling - Develop a Bayesian Non-Negative Reward Model for robust and interpretable reward learning in reinforcement learning from human feedback, addressing reward hacking challenges.
- Online Causal Kalman Filtering for Stable and Effective Policy Optimization (viability: 4): https://sciencetostartup.com/paper/online-causal-kalman-filtering-for-stable-and-effective-policy-optimization - Develop an online causal Kalman filter method to stabilize reinforcement learning policy optimization in large language models.
- Hierarchical Zero-Order Optimization for Deep Neural Networks (viability: 5): https://sciencetostartup.com/paper/hierarchical-zero-order-optimization-for-deep-neural-networks - Develop a more efficient zero-order optimization method for deep neural network training utilizing Hierarchical Zero-Order optimization.
- When to Memorize and When to Stop: Gated Recurrent Memory for Long-Context Reasoning (viability: 7): https://sciencetostartup.com/paper/when-to-memorize-and-when-to-stop-gated-recurrent-memory-for-long-context-reasoning - Develop GRU-Mem, an efficient long-context reasoning tool leveraging gated recurrent memory for faster inference.
- C^2ROPE: Causal Continuous Rotary Positional Encoding for 3D Large Multimodal-Models Reasoning (viability: 6): https://sciencetostartup.com/paper/c-2rope-causal-continuous-rotary-positional-encoding-for-3d-large-multimodal-models-reasoning - C^2RoPE enhances 3D multimodal models with a novel positional encoding method for improved spatial continuity and causal reasoning in visual tasks.
- Enhancing Weakly Supervised Multimodal Video Anomaly Detection through Text Guidance (viability: 5): https://sciencetostartup.com/paper/enhancing-weakly-supervised-multimodal-video-anomaly-detection-through-text-guidance - Leverage text-guided techniques to enhance multimodal video anomaly detection for reducing false alarms and improving anomaly characterization.
- $μ$pscaling small models: Principled warm starts and hyperparameter transfer (viability: 4): https://sciencetostartup.com/paper/pscaling-small-models-principled-warm-starts-and-hyperparameter-transfer - Develop a method for upscaling small neural networks to larger sizes with efficient hyperparameter transfer.
- A Swap-Adversarial Framework for Improving Domain Generalization in Electroencephalography-Based Parkinson's Disease Prediction (viability: 8): https://sciencetostartup.com/paper/a-swap-adversarial-framework-for-improving-domain-generalization-in-electroencephalography-based-parkinson-s-disease-pre - Swap-Adversarial Framework for enhanced Parkinson's prediction using ECoG data with strong domain generalization.
- Co-jump: Cooperative Jumping with Quadrupedal Robots via Multi-Agent Reinforcement Learning (viability: 5): https://sciencetostartup.com/paper/co-jump-cooperative-jumping-with-quadrupedal-robots-via-multi-agent-reinforcement-learning - Develop a communication-free cooperative jumping capability for quadrupedal robots using advanced multi-agent reinforcement learning.
- 1%>100%: High-Efficiency Visual Adapter with Complex Linear Projection Optimization (viability: 8): https://sciencetostartup.com/paper/1-100-high-efficiency-visual-adapter-with-complex-linear-projection-optimization - Efficient vision model adapter reducing parameters by 99% with state-of-the-art performance.
- Low-Dimensional Execution Manifolds in Transformer Learning Dynamics: Evidence from Modular Arithmetic Tasks (viability: 4): https://sciencetostartup.com/paper/low-dimensional-execution-manifolds-in-transformer-learning-dynamics-evidence-from-modular-arithmetic-tasks - Develop a geometric framework for transformer interpretability on modular tasks using low-dimensional manifolds.
- Abstraction Generation for Generalized Planning with Pretrained Large Language Models (viability: 4): https://sciencetostartup.com/paper/abstraction-generation-for-generalized-planning-with-pretrained-large-language-models - Develop automated debugging tools for LLM-generated Qualitative Numerical Planning abstractions.
- MERIT Feedback Elicits Better Bargaining in LLM Negotiators (viability: 7): https://sciencetostartup.com/paper/merit-feedback-elicits-better-bargaining-in-llm-negotiators - Develop an LLM-based tool for improved strategic negotiation using a new benchmark and feedback-driven alignment.
- LakeMLB: Data Lake Machine Learning Benchmark (viability: 7): https://sciencetostartup.com/paper/lakemlb-data-lake-machine-learning-benchmark - LakeMLB is a standardized benchmark for evaluating machine learning performance in data lake environments, featuring multi-table scenarios.
- A Dual-Stream Physics-Augmented Unsupervised Architecture for Runtime Embedded Vehicle Health Monitoring (viability: 6): https://sciencetostartup.com/paper/a-dual-stream-physics-augmented-unsupervised-architecture-for-runtime-embedded-vehicle-health-monitoring - Develop a dual-stream AI architecture for real-time embedded vehicle health monitoring, optimizing predictive maintenance over traditional metrics.
- Breaking the Curse of Repulsion: Optimistic Distributionally Robust Policy Optimization for Off-Policy Generative Recommendation (viability: 6): https://sciencetostartup.com/paper/breaking-the-curse-of-repulsion-optimistic-distributionally-robust-policy-optimization-for-off-policy-generative-recomme - Optimistic DRPO offers a robust approach to enhance sequential user interactions by overcoming data quality issues in policy-based RL systems.
- AIvilization v0: Toward Large-Scale Artificial Social Simulation with a Unified Agent Architecture and Adaptive Agent Profiles (viability: 3): https://sciencetostartup.com/paper/aivilization-v0-toward-large-scale-artificial-social-simulation-with-a-unified-agent-architecture-and-adaptive-agent-pro - AIvilization v0 enables comprehensive artificial social simulation with adaptable agent profiles for long-term stability.
- Equivariant Evidential Deep Learning for Interatomic Potentials (viability: 5): https://sciencetostartup.com/paper/equivariant-evidential-deep-learning-for-interatomic-potentials - Develop a framework for efficient uncertainty quantification in molecular dynamics simulations using equivariant evidential deep learning.
- AI-rithmetic (viability: 5): https://sciencetostartup.com/paper/ai-rithmetic - AI-rithmetic investigates AI models' failures in basic arithmetic, offering potential improvements for model accuracy in fundamental operations.
- Affordances Enable Partial World Modeling with LLMs (viability: 5): https://sciencetostartup.com/paper/affordances-enable-partial-world-modeling-with-llms - Develop partial world models leveraging affordances to enhance LLM-driven robotics task efficiency.
- Time-to-Event Transformer to Capture Timing Attention of Events in EHR Time Series (viability: 7): https://sciencetostartup.com/paper/time-to-event-transformer-to-capture-timing-attention-of-events-in-ehr-time-series - Leverage the Timing-Transformer architecture to enhance precision medicine by interpreting EHR time series and predicting event onset timing.
- ERGO: Excess-Risk-Guided Optimization for High-Fidelity Monocular 3D Gaussian Splatting (viability: 5): https://sciencetostartup.com/paper/ergo-excess-risk-guided-optimization-for-high-fidelity-monocular-3d-gaussian-splatting - ERGO enhances 3D reconstruction from single images by optimizing Gaussian splatting using excess risk decomposition.
- From Classical to Topological Neural Networks Under Uncertainty (viability: 3): https://sciencetostartup.com/paper/from-classical-to-topological-neural-networks-under-uncertainty - Develop topology-aware and uncertainty-aware models for enhanced AI robustness in military applications.
- The Complexity of Bayesian Network Learning: Revisiting the Superstructure (viability: 2): https://sciencetostartup.com/paper/the-complexity-of-bayesian-network-learning-revisiting-the-superstructure - Explore computational complexities in Bayesian Network Structure Learning for enhanced theoretical understanding.
- Self-Evolving Recommendation System: End-To-End Autonomous Model Optimization With LLM Agents (viability: 9): https://sciencetostartup.com/paper/self-evolving-recommendation-system-end-to-end-autonomous-model-optimization-with-llm-agents - Develop autonomous recommendation system optimization with LLM agents for improved user engagement.
- Internalizing Meta-Experience into Memory for Guided Reinforcement Learning in Large Language Models (viability: 5): https://sciencetostartup.com/paper/internalizing-meta-experience-into-memory-for-guided-reinforcement-learning-in-large-language-models - Innovating Large Language Models with Meta-Experience Learning for better reinforcement-driven self-improvement.
- Versor: A Geometric Sequence Architecture (viability: 7): https://sciencetostartup.com/paper/versor-a-geometric-sequence-architecture - Versor offers a scalable, geometrically-aware architecture outperforming traditional models in efficiency and interpretability for scientific modeling.
- Biases in the Blind Spot: Detecting What LLMs Fail to Mention (viability: 6): https://sciencetostartup.com/paper/biases-in-the-blind-spot-detecting-what-llms-fail-to-mention - Automated pipeline to uncover task-specific biases in LLMs without predefined categories.
- Olaf-World: Orienting Latent Actions for Video World Modeling (viability: 6): https://sciencetostartup.com/paper/olaf-world-orienting-latent-actions-for-video-world-modeling - Pretrain action-conditioned video world models for zero-shot action transfer and efficient adaptation.
- Step-resolved data attribution for looped transformers (viability: 2): https://sciencetostartup.com/paper/step-resolved-data-attribution-for-looped-transformers - Introducing Step-Decomposed Influence for looped transformers to provide per-step data attribution insights.
- Causality in Video Diffusers is Separable from Denoising (viability: 6): https://sciencetostartup.com/paper/causality-in-video-diffusers-is-separable-from-denoising - Introducing Separable Causal Diffusion (SCD) for efficient and high-quality video generation.
- Agent World Model: Infinity Synthetic Environments for Agentic Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/agent-world-model-infinity-synthetic-environments-for-agentic-reinforcement-learning - Agent World Model generates synthetic environments to efficiently train reinforcement learning agents for complex, multi-turn interactions.
- CODE-SHARP: Continuous Open-ended Discovery and Evolution of Skills as Hierarchical Reward Programs (viability: 6): https://sciencetostartup.com/paper/code-sharp-continuous-open-ended-discovery-and-evolution-of-skills-as-hierarchical-reward-programs - Develop a framework using Foundation Models to enable agents to discover and learn novel skills autonomously, significantly outperforming existing methods.
- Anagent For Enhancing Scientific Table & Figure Analysis (viability: 7): https://sciencetostartup.com/paper/anagent-for-enhancing-scientific-table-figure-analysis - Anagent leverages multi-agent AI for enhanced analysis of scientific tables and figures, overcoming current system limitations.
- Chain of Mindset: Reasoning with Adaptive Cognitive Modes (viability: 7): https://sciencetostartup.com/paper/chain-of-mindset-reasoning-with-adaptive-cognitive-modes - Develop an AI tool for adaptive cognitive reasoning that dynamically adjusts problem-solving strategies for each step.
- Long Chain-of-Thought Compression via Fine-Grained Group Policy Optimization (viability: 4): https://sciencetostartup.com/paper/long-chain-of-thought-compression-via-fine-grained-group-policy-optimization - Optimize language model reasoning through fine-grained CoT compression for enhanced compute efficiency.
- Optimistic World Models: Efficient Exploration in Model-Based Deep Reinforcement Learning (viability: 4): https://sciencetostartup.com/paper/optimistic-world-models-efficient-exploration-in-model-based-deep-reinforcement-learning - Introducing Optimistic World Models for efficient exploration in deep reinforcement learning with minimal training modifications.
- Fake-HR1: Rethinking reasoning of vision language model for synthetic image detection (viability: 4): https://sciencetostartup.com/paper/fake-hr1-rethinking-reasoning-of-vision-language-model-for-synthetic-image-detection - Develop a reasoning model for efficient detection of synthetic images that adapts reasoning complexity based on task characteristics.
- Decoupled Reasoning with Implicit Fact Tokens (DRIFT): A Dual-Model Framework for Efficient Long-Context Inference (viability: 6): https://sciencetostartup.com/paper/decoupled-reasoning-with-implicit-fact-tokens-drift-a-dual-model-framework-for-efficient-long-context-inference - DRIFT offers an efficient dual-model framework to enhance LLMs' long-context reasoning by decoupling knowledge and inference processes.
- ADORA: Training Reasoning Models with Dynamic Advantage Estimation on Reinforcement Learning (viability: 5): https://sciencetostartup.com/paper/adora-training-reasoning-models-with-dynamic-advantage-estimation-on-reinforcement-learning - ADORA optimizes reinforcement learning models by dynamically adjusting advantage estimation for more efficient policy updates.
- Kunlun: Establishing Scaling Laws for Massive-Scale Recommendation Systems through Unified Architecture Design (viability: 3): https://sciencetostartup.com/paper/kunlun-establishing-scaling-laws-for-massive-scale-recommendation-systems-through-unified-architecture-design - Kunlun improves scaling efficiency in recommendation systems for enhanced model performance.
- RoboSubtaskNet: Temporal Sub-task Segmentation for Human-to-Robot Skill Transfer in Real-World Environments (viability: 9): https://sciencetostartup.com/paper/robosubtasknet-temporal-sub-task-segmentation-for-human-to-robot-skill-transfer-in-real-world-environments - RoboSubtaskNet enables effective human-to-robot skill transfer for precise and adaptive task automation in collaborative environments.
- Discovering High Level Patterns from Simulation Traces (viability: 7): https://sciencetostartup.com/paper/discovering-high-level-patterns-from-simulation-traces - A tool that enhances language models' ability to reason about physical systems using coarse-grained patterns synthesized from simulation logs.
- A Collaborative Safety Shield for Safe and Efficient CAV Lane Changes in Congested On-Ramp Merging (viability: 7): https://sciencetostartup.com/paper/a-collaborative-safety-shield-for-safe-and-efficient-cav-lane-changes-in-congested-on-ramp-merging - Develops an open-source software for enhancing autonomous vehicle lane changes with safety and efficiency using multi-agent collaboration.
- ESTAR: Early-Stopping Token-Aware Reasoning For Efficient Inference (viability: 6): https://sciencetostartup.com/paper/estar-early-stopping-token-aware-reasoning-for-efficient-inference - Develop an AI tool to optimize large reasoning models by implementing early stopping mechanisms, significantly reducing computation without sacrificing accuracy.
- A Unified Assessment of the Poverty of the Stimulus Argument for Neural Language Models (viability: 3): https://sciencetostartup.com/paper/a-unified-assessment-of-the-poverty-of-the-stimulus-argument-for-neural-language-models - Develop an evaluation suite to test neural language models on their ability to generalize language phenomena without innate syntax constraints.
- Infusion: Shaping Model Behavior by Editing Training Data via Influence Functions (viability: 8): https://sciencetostartup.com/paper/infusion-shaping-model-behavior-by-editing-training-data-via-influence-functions - Infusion leverages influence functions to craft subtle training data perturbations that reshape AI model behavior without explicit training signal insertion.
- Coupled Inference in Diffusion Models for Semantic Decomposition (viability: 5): https://sciencetostartup.com/paper/coupled-inference-in-diffusion-models-for-semantic-decomposition - A novel framework using diffusion models for superior semantic decomposition in visual scenes.
- Drug Release Modeling using Physics-Informed Neural Networks (viability: 5): https://sciencetostartup.com/paper/drug-release-modeling-using-physics-informed-neural-networks - Leverage Physics-Informed Neural Networks for accelerated, accurate drug release modeling from limited data.
- Bladder Vessel Segmentation using a Hybrid Attention-Convolution Framework (viability: 7): https://sciencetostartup.com/paper/bladder-vessel-segmentation-using-a-hybrid-attention-convolution-framework - Develop an AI tool for reliable bladder vessel segmentation in endoscopic procedures to enhance clinical navigation.
- Monocular Normal Estimation via Shading Sequence Estimation (viability: 6): https://sciencetostartup.com/paper/monocular-normal-estimation-via-shading-sequence-estimation - Develop a tool for enhanced monocular normal estimation using shading sequence prediction to solve 3D misalignment issues.
- LLMs Encode Their Failures: Predicting Success from Pre-Generation Activations (viability: 7): https://sciencetostartup.com/paper/llms-encode-their-failures-predicting-success-from-pre-generation-activations - Develop an API that optimizes LLM inference cost by predicting necessary computation using pre-generation activations.
- Would a Large Language Model Pay Extra for a View? Inferring Willingness to Pay from Subjective Choices (viability: 4): https://sciencetostartup.com/paper/would-a-large-language-model-pay-extra-for-a-view-inferring-willingness-to-pay-from-subjective-choices - A tool to infer willingness to pay using LLMs for subjective decision support in travel applications.
- ExO-PPO: an Extended Off-policy Proximal Policy Optimization Algorithm (viability: 5): https://sciencetostartup.com/paper/exo-ppo-an-extended-off-policy-proximal-policy-optimization-algorithm - ExO-PPO enhances Proximal Policy Optimization by integrating off-policy data utilization for improved performance and efficiency in reinforcement learning tasks.
- From Lightweight CNNs to SpikeNets: Benchmarking Accuracy-Energy Tradeoffs with Pruned Spiking SqueezeNet (viability: 7): https://sciencetostartup.com/paper/from-lightweight-cnns-to-spikenets-benchmarking-accuracy-energy-tradeoffs-with-pruned-spiking-squeezenet - Develop an energy-efficient, lightweight SNN-based alternative to CNNs for edge intelligence with significantly reduced power consumption.
- Physics-informed diffusion models in spectral space (viability: 6): https://sciencetostartup.com/paper/physics-informed-diffusion-models-in-spectral-space - Developing a physics-informed diffusion model to efficiently solve parametric PDEs with improved accuracy using spectral space latent representations.
- Maastricht University at AMIYA: Adapting LLMs for Dialectal Arabic using Fine-tuning and MBR Decoding (viability: 2): https://sciencetostartup.com/paper/maastricht-university-at-amiya-adapting-llms-for-dialectal-arabic-using-fine-tuning-and-mbr-decoding - Develop a compact solution for improving dialectal Arabic LLMs using LoRA fine-tuning and MBR decoding.
- Administrative Law's Fourth Settlement: AI and the Capability-Accountability Trap (viability: 5): https://sciencetostartup.com/paper/administrative-law-s-fourth-settlement-ai-and-the-capability-accountability-trap - Innovating administrative law through AI to balance capability and accountability in governance.
- ClinAlign: Scaling Healthcare Alignment from Clinician Preference (viability: 7): https://sciencetostartup.com/paper/clinalign-scaling-healthcare-alignment-from-clinician-preference - Develop a scalable framework for aligning large language models with clinician preferences using distilled clinical principles.
- MATA: Multi-Agent Framework for Reliable and Flexible Table Question Answering (viability: 6): https://sciencetostartup.com/paper/mata-multi-agent-framework-for-reliable-and-flexible-table-question-answering - MATA provides a multi-agent framework for efficient and scalable Table Question Answering using small open-source models.
- FLINGO -- Instilling ASP Expressiveness into Linear Integer Constraints (viability: 2): https://sciencetostartup.com/paper/flingo-instilling-asp-expressiveness-into-linear-integer-constraints - FLINGO enhances CASP with ASP expressiveness for better numerical constraint handling.
- Detecting radar targets swarms in range profiles with a partially complex-valued neural network (viability: 5): https://sciencetostartup.com/paper/detecting-radar-targets-swarms-in-range-profiles-with-a-partially-complex-valued-neural-network - Develop adaptive radar target detection using partially complex-valued neural networks for enhanced range profile processing.
- LEMUR: A Corpus for Robust Fine-Tuning of Multilingual Law Embedding Models for Retrieval (viability: 7): https://sciencetostartup.com/paper/lemur-a-corpus-for-robust-fine-tuning-of-multilingual-law-embedding-models-for-retrieval - Develop a robust multilingual law embedding model fine-tuned for reliable legal information retrieval across EU legislative documents.
- ECG-IMN: Interpretable Mesomorphic Neural Networks for 12-Lead Electrocardiogram Interpretation (viability: 6): https://sciencetostartup.com/paper/ecg-imn-interpretable-mesomorphic-neural-networks-for-12-lead-electrocardiogram-interpretation - Develop an interpretable neural network for ECG analysis that provides precise explanations alongside competitive predictive performance.
- Comprehensive Comparison of RAG Methods Across Multi-Domain Conversational QA (viability: 5): https://sciencetostartup.com/paper/comprehensive-comparison-of-rag-methods-across-multi-domain-conversational-qa - Empirical study of retrieval-augmented generation methods for multi-turn conversational QA with open-source code.
- Autoregressive Direct Preference Optimization (viability: 2): https://sciencetostartup.com/paper/autoregressive-direct-preference-optimization - Develop a novel autoregressive model for optimizing language model preferences using Autoregressive Direct Preference Optimization (ADPO).
- Learning to Discover Iterative Spectral Algorithms (viability: 3): https://sciencetostartup.com/paper/learning-to-discover-iterative-spectral-algorithms - Discover new iterative spectral algorithms using the AutoSpec framework.
- Listen to the Layers: Mitigating Hallucinations with Inter-Layer Disagreement (viability: 9): https://sciencetostartup.com/paper/listen-to-the-layers-mitigating-hallucinations-with-inter-layer-disagreement - CoCoA offers a novel training-free method to significantly reduce AI hallucinations at inference time, enhancing LLM reliability for critical applications.
- NOWJ @BioCreative IX ToxHabits: An Ensemble Deep Learning Approach for Detecting Substance Use and Contextual Information in Clinical Texts (viability: 6): https://sciencetostartup.com/paper/nowj-biocreative-ix-toxhabits-an-ensemble-deep-learning-approach-for-detecting-substance-use-and-contextual-information- - Develop a high-precision NLP tool for detecting substance use in clinical texts using ensemble deep learning.
- AlgoVeri: An Aligned Benchmark for Verified Code Generation on Classical Algorithms (viability: 6): https://sciencetostartup.com/paper/algoveri-an-aligned-benchmark-for-verified-code-generation-on-classical-algorithms - AlgoVeri benchmarks and evaluates AI models in generating verified code for classical algorithms across multiple paradigms.
- SWE-AGI: Benchmarking Specification-Driven Software Construction with MoonBit in the Era of Autonomous Agents (viability: 5): https://sciencetostartup.com/paper/swe-agi-benchmarking-specification-driven-software-construction-with-moonbit-in-the-era-of-autonomous-agents - Develop SWE-AGI as a benchmark for evaluating LLMs in specification-driven software construction using MoonBit.
- Evaluating Social Bias in RAG Systems: When External Context Helps and Reasoning Hurts (viability: 5): https://sciencetostartup.com/paper/evaluating-social-bias-in-rag-systems-when-external-context-helps-and-reasoning-hurts - A tool to evaluate and reduce social bias in Retrieval-Augmented Generation systems using Chain-of-Thought prompting.
- A Behavioral Fingerprint for Large Language Models: Provenance Tracking via Refusal Vectors (viability: 7): https://sciencetostartup.com/paper/a-behavioral-fingerprint-for-large-language-models-provenance-tracking-via-refusal-vectors - Novel fingerprinting technology for protecting the IP of large language models through behavioral analysis.
- Accelerating Post-Quantum Cryptography via LLM-Driven Hardware-Software Co-Design (viability: 3): https://sciencetostartup.com/paper/accelerating-post-quantum-cryptography-via-llm-driven-hardware-software-co-design - Leverage LLMs for accelerated FPGA-based PQC hardware design.
- Squeezing More from the Stream : Learning Representation Online for Streaming Reinforcement Learning (viability: 3): https://sciencetostartup.com/paper/squeezing-more-from-the-stream-learning-representation-online-for-streaming-reinforcement-learning - Develop an efficient online learning system for streaming reinforcement learning with enhanced representation learning.
- AgentCgroup: Understanding and Controlling OS Resources of AI Agents (viability: 3): https://sciencetostartup.com/paper/agentcgroup-understanding-and-controlling-os-resources-of-ai-agents - AgentCgroup improves multi-tenant isolation and reduces resource waste for AI agents in cloud environments.
- Not-in-Perspective: Towards Shielding Google's Perspective API Against Adversarial Negation Attacks (viability: 6): https://sciencetostartup.com/paper/not-in-perspective-towards-shielding-google-s-perspective-api-against-adversarial-negation-attacks - Develop a reasoning-based wrapper to enhance Google's Perspective API by mitigating adversarial negation attacks, improving toxicity detection accuracy.
- Auditing Multi-Agent LLM Reasoning Trees Outperforms Majority Vote and LLM-as-Judge (viability: 2): https://sciencetostartup.com/paper/auditing-multi-agent-llm-reasoning-trees-outperforms-majority-vote-and-llm-as-judge - AgentAuditor enhances multi-agent LLM decision-making by replacing majority voting with reasoning tree analysis to improve accuracy.
- GAFR-Net: A Graph Attention and Fuzzy-Rule Network for Interpretable Breast Cancer Image Classification (viability: 7): https://sciencetostartup.com/paper/gafr-net-a-graph-attention-and-fuzzy-rule-network-for-interpretable-breast-cancer-image-classification - GAFR-Net offers an interpretable graph attention and fuzzy-rule network for breast cancer histopathology image classification with minimal supervision.
- Don't Shoot The Breeze: Topic Continuity Model Using Nonlinear Naive Bayes With Attention (viability: 3): https://sciencetostartup.com/paper/don-t-shoot-the-breeze-topic-continuity-model-using-nonlinear-naive-bayes-with-attention - Develop a topic continuity model to enhance chatbot user experience by maintaining conversation topic alignment.
- STaR: Scalable Task-Conditioned Retrieval for Long-Horizon Multimodal Robot Memory (viability: 7): https://sciencetostartup.com/paper/star-scalable-task-conditioned-retrieval-for-long-horizon-multimodal-robot-memory - Develop a scalable multimodal memory system for mobile robots that enhances navigation through advanced contextual reasoning.
- VLM-Guided Iterative Refinement for Surgical Image Segmentation with Foundation Models (viability: 7): https://sciencetostartup.com/paper/vlm-guided-iterative-refinement-for-surgical-image-segmentation-with-foundation-models - IR-SIS offers adaptive, language-guided surgical image segmentation with iterative refinement for enhanced clinician guidance.
- Do Neural Networks Lose Plasticity in a Gradually Changing World? (viability: 2): https://sciencetostartup.com/paper/do-neural-networks-lose-plasticity-in-a-gradually-changing-world - Investigates neural network plasticity in gradually changing environments to propose solutions for continual learning.
- MUZZLE: Adaptive Agentic Red-Teaming of Web Agents Against Indirect Prompt Injection Attacks (viability: 5): https://sciencetostartup.com/paper/muzzle-adaptive-agentic-red-teaming-of-web-agents-against-indirect-prompt-injection-attacks - MUZZLE is an adaptive framework that enhances the security of LLM-based web agents against indirect prompt injection attacks.
- Gradient Residual Connections (viability: 3): https://sciencetostartup.com/paper/gradient-residual-connections - Gradient-based residual connections improve neural network approximation of high-frequency functions, with potential application in image super-resolution.
- $n$-Musketeers: Reinforcement Learning Shapes Collaboration Among Language Models (viability: 2): https://sciencetostartup.com/paper/n-musketeers-reinforcement-learning-shapes-collaboration-among-language-models - Utilize small specialized language models to perform structured reasoning through reinforcement learning with verifiable rewards.
- FlyAOC: Evaluating Agentic Ontology Curation of Drosophila Scientific Knowledge Bases (viability: 5): https://sciencetostartup.com/paper/flyaoc-evaluating-agentic-ontology-curation-of-drosophila-scientific-knowledge-bases - FlyBench is a benchmarking tool for evaluating AI agents on ontology curation in scientific literature, particularly for Drosophila genetic data.
- CoMMa: Contribution-Aware Medical Multi-Agents From A Game-Theoretic Perspective (viability: 7): https://sciencetostartup.com/paper/comma-contribution-aware-medical-multi-agents-from-a-game-theoretic-perspective - Develop a decentralized LLM-agent framework for oncology decision support leveraging game-theoretic methods.
- What do Geometric Hallucination Detection Metrics Actually Measure? (viability: 4): https://sciencetostartup.com/paper/what-do-geometric-hallucination-detection-metrics-actually-measure - Develop an AI tool to detect hallucination in generative models using geometric signals for improved reliability in multi-domain applications.
- Benchmarking the Energy Savings with Speculative Decoding Strategies (viability: 5): https://sciencetostartup.com/paper/benchmarking-the-energy-savings-with-speculative-decoding-strategies - Develop a tool to benchmark and optimize energy usage in LLM inference through speculative decoding strategies.
- A Small-Scale System for Autoregressive Program Synthesis Enabling Controlled Experimentation (viability: 6): https://sciencetostartup.com/paper/a-small-scale-system-for-autoregressive-program-synthesis-enabling-controlled-experimentation - Cadmus provides an affordable system for controlled experimentation on program synthesis with small models, outperforming larger models in specific tasks.
- Distributed Hybrid Parallelism for Large Language Models: Comparative Study and System Design Guide (viability: 3): https://sciencetostartup.com/paper/distributed-hybrid-parallelism-for-large-language-models-comparative-study-and-system-design-guide - Comprehensive guide and analysis for designing optimal distributed systems in large language model training.
- Robustness Is a Function, Not a Number: A Factorized Comprehensive Study of OOD Robustness in Vision-Based Driving (viability: 6): https://sciencetostartup.com/paper/robustness-is-a-function-not-a-number-a-factorized-comprehensive-study-of-ood-robustness-in-vision-based-driving - Develop robust OOD-resistant vision-based driving policies using foundation-model features and environment factorization.
- CIC-Trap4Phish: A Unified Multi-Format Dataset for Phishing and Quishing Attachment Detection (viability: 5): https://sciencetostartup.com/paper/cic-trap4phish-a-unified-multi-format-dataset-for-phishing-and-quishing-attachment-detection - Unified dataset for detecting phishing and quishing attachments across multiple file formats.
- ArcFlow: Unleashing 2-Step Text-to-Image Generation via High-Precision Non-Linear Flow Distillation (viability: 7): https://sciencetostartup.com/paper/arcflow-unleashing-2-step-text-to-image-generation-via-high-precision-non-linear-flow-distillation - Build a high-speed text-to-image generator using ArcFlow's novel non-linear flow distillation technique for efficient diffusion model inference.
- Next-Gen CAPTCHAs: Leveraging the Cognitive Gap for Scalable and Diverse GUI-Agent Defense (viability: 5): https://sciencetostartup.com/paper/next-gen-captchas-leveraging-the-cognitive-gap-for-scalable-and-diverse-gui-agent-defense - Develop a scalable framework for Next-Gen CAPTCHAs to defend web systems against advanced GUI-enabled agents.
- ANCRe: Adaptive Neural Connection Reassignment for Efficient Depth Scaling (viability: 7): https://sciencetostartup.com/paper/ancre-adaptive-neural-connection-reassignment-for-efficient-depth-scaling - Implement ANCRe to optimize neural network convergence and efficiency for large models.
- GEBench: Benchmarking Image Generation Models as GUI Environments (viability: 7): https://sciencetostartup.com/paper/gebench-benchmarking-image-generation-models-as-gui-environments - Develop a benchmark and metric tool for evaluating dynamic interaction and coherence in GUI generative models.
- ARO: A New Lens On Matrix Optimization For Large Models (viability: 4): https://sciencetostartup.com/paper/aro-a-new-lens-on-matrix-optimization-for-large-models - ARO offers a novel matrix optimization framework to enhance the efficiency of large language model training through innovative gradient rotation methods.
- Data Science and Technology Towards AGI Part I: Tiered Data Management (viability: 5): https://sciencetostartup.com/paper/data-science-and-technology-towards-agi-part-i-tiered-data-management - Tiered data management system to enhance LLM training efficiency and performance by utilizing a structured data framework.
- From Obstacles to Etiquette: Robot Social Navigation with VLM-Informed Path Selection (viability: 7): https://sciencetostartup.com/paper/from-obstacles-to-etiquette-robot-social-navigation-with-vlm-informed-path-selection - A social robot navigation tool that uses vision-language models for path planning, enabling robots to adhere to social norms.
- iGRPO: Self-Feedback-Driven LLM Reasoning (viability: 6): https://sciencetostartup.com/paper/igrpo-self-feedback-driven-llm-reasoning - Introducing iGRPO, a self-feedback-driven training method enhancing LLMs for superior mathematical reasoning.
- InternAgent-1.5: A Unified Agentic Framework for Long-Horizon Autonomous Scientific Discovery (viability: 3): https://sciencetostartup.com/paper/internagent-1-5-a-unified-agentic-framework-for-long-horizon-autonomous-scientific-discovery - Build a unified agentic framework capable of autonomous long-horizon scientific discovery for computational and empirical domains.
- Improving Detection of Rare Nodes in Hierarchical Multi-Label Learning (viability: 5): https://sciencetostartup.com/paper/improving-detection-of-rare-nodes-in-hierarchical-multi-label-learning - Enhance hierarchical multi-label classification models with a novel loss function to significantly improve rare node detection and recall.
- Next Concept Prediction in Discrete Latent Space Leads to Stronger Language Models (viability: 5): https://sciencetostartup.com/paper/next-concept-prediction-in-discrete-latent-space-leads-to-stronger-language-models - ConceptLM boosts language model performance through Next Concept Prediction, enhancing pretraining tasks with discrete concept generation.
- StretchTime: Adaptive Time Series Forecasting via Symplectic Attention (viability: 5): https://sciencetostartup.com/paper/stretchtime-adaptive-time-series-forecasting-via-symplectic-attention - Develop time series forecasting models that handle temporal misalignments using symplectic attention for improved performance in dynamic environments.
- stable-worldmodel-v1: Reproducible World Modeling Research and Evaluation (viability: 5): https://sciencetostartup.com/paper/stable-worldmodel-v1-reproducible-world-modeling-research-and-evaluation - A modular and standardized research ecosystem for reproducible world modeling and evaluation.
- A Behavioural and Representational Evaluation of Goal-Directedness in Language Model Agents (viability: 3): https://sciencetostartup.com/paper/a-behavioural-and-representational-evaluation-of-goal-directedness-in-language-model-agents - Develop a tool for evaluating and interpreting goal-directed behaviours in language model agents.
- MotionCrafter: Dense Geometry and Motion Reconstruction with a 4D VAE (viability: 8): https://sciencetostartup.com/paper/motioncrafter-dense-geometry-and-motion-reconstruction-with-a-4d-vae - MotionCrafter enables state-of-the-art dense 4D geometry and motion reconstruction from monocular videos using a novel 4D VAE.
- Digital Twin and Agentic AI for Wild Fire Disaster Management: Intelligent Virtual Situation Room (viability: 7): https://sciencetostartup.com/paper/digital-twin-and-agentic-ai-for-wild-fire-disaster-management-intelligent-virtual-situation-room - A digital twin platform for proactive wildfire disaster management using AI agents to optimize response times and resource coordination.
- CoRefine: Confidence-Guided Self-Refinement for Adaptive Test-Time Compute (viability: 8): https://sciencetostartup.com/paper/corefine-confidence-guided-self-refinement-for-adaptive-test-time-compute - CoRefine reduces compute costs for LLMs by leveraging confidence-guided self-refinement to achieve competitive accuracy.
- pixelLOG: Logging of Online Gameplay for Cognitive Research (viability: 3): https://sciencetostartup.com/paper/pixellog-logging-of-online-gameplay-for-cognitive-research - pixelLOG enhances cognitive research by enabling high-frequency behavioral tracking in Spigot-Minecraft environments.
- AnomSeer: Reinforcing Multimodal LLMs to Reason for Time-Series Anomaly Detection (viability: 2): https://sciencetostartup.com/paper/anomseer-reinforcing-multimodal-llms-to-reason-for-time-series-anomaly-detection - AnomSeer enhances multimodal LLMs for time-series anomaly detection with a new fine-grained reasoning approach, but lacks commercial viability or clear product development path.
- Understanding Dynamic Compute Allocation in Recurrent Transformers (viability: 5): https://sciencetostartup.com/paper/understanding-dynamic-compute-allocation-in-recurrent-transformers - Develop ANIRA, a recurrent Transformer framework optimizing compute based on token complexity for dynamic resource allocation in NLP tasks.
- Deciding the Satisfiability of Combined Qualitative Constraint Networks (viability: 2): https://sciencetostartup.com/paper/deciding-the-satisfiability-of-combined-qualitative-constraint-networks - A framework for unified reasoning in qualitative constraint networks, studying satisfiability and complexity.
- WildReward: Learning Reward Models from In-the-Wild Human Interactions (viability: 8): https://sciencetostartup.com/paper/wildreward-learning-reward-models-from-in-the-wild-human-interactions - A new method for training reward models directly from in-the-wild LLM interactions, potentially reducing the need for costly human-annotated training data.
- Affective Flow Language Model for Emotional Support Conversation (viability: 7): https://sciencetostartup.com/paper/affective-flow-language-model-for-emotional-support-conversation - AFlow enhances emotional support conversations with better empathy through fine-grained supervision and outperforming state-of-the-art models.
- Negative-Aware Diffusion Process for Temporal Knowledge Graph Extrapolation (viability: 6): https://sciencetostartup.com/paper/negative-aware-diffusion-process-for-temporal-knowledge-graph-extrapolation - Introducing NADEx, a state-of-the-art model for temporal knowledge graph extrapolation with negative context awareness.
- $\texttt{lrnnx}$: A library for Linear RNNs (viability: 6): https://sciencetostartup.com/paper/texttt-lrnnx-a-library-for-linear-rnns - A unified library for easy implementation and experimentation with linear recurrent neural networks (LRNNs).
- FreqLens: Interpretable Frequency Attribution for Time Series Forecasting (viability: 5): https://sciencetostartup.com/paper/freqlens-interpretable-frequency-attribution-for-time-series-forecasting - Implement FreqLens for interpretable frequency attribution in time series forecasting.
- On the Expressive Power of GNNs for Boolean Satisfiability (viability: 3): https://sciencetostartup.com/paper/on-the-expressive-power-of-gnns-for-boolean-satisfiability - Leveraging GNNs to enhance the expressive power of SAT solvers through WL hierarchy analysis.
- QUOKA: Query-Oriented KV Selection For Efficient LLM Prefill (viability: 3): https://sciencetostartup.com/paper/quoka-query-oriented-kv-selection-for-efficient-llm-prefill - Develop a sparse attention algorithm to accelerate transformer inference by prioritizing query-key interactions.
- Zero-shot System for Automatic Body Region Detection for Volumetric CT and MR Images (viability: 7): https://sciencetostartup.com/paper/zero-shot-system-for-automatic-body-region-detection-for-volumetric-ct-and-mr-images - Zero-shot body region detection system for CT and MR imaging enhances medical analysis by eliminating reliance on DICOM metadata.
- Exploring SAIG Methods for an Objective Evaluation of XAI (viability: 2): https://sciencetostartup.com/paper/exploring-saig-methods-for-an-objective-evaluation-of-xai - Develop a standardized framework for evaluating eXplainable AI methods using Synthetic Artificial Intelligence Ground truth.
- Intermediate Results on the Complexity of STRIPS$_{1}^{1}$ (viability: 2): https://sciencetostartup.com/paper/intermediate-results-on-the-complexity-of-strips-1-1 - Explores the computational complexity of propositional STRIPS planning with a focus on small solution hypotheses and literal graphs.
- PBLean: Pseudo-Boolean Proof Certificates for Lean 4 (viability: 3): https://sciencetostartup.com/paper/pblean-pseudo-boolean-proof-certificates-for-lean-4 - PBLean provides a method for importing pseudo-Boolean proof certificates into Lean 4, enhancing formal verification processes.
- LLaDA2.1: Speeding Up Text Diffusion via Token Editing (viability: 5): https://sciencetostartup.com/paper/llada2-1-speeding-up-text-diffusion-via-token-editing - LLaDA2.1 offers a faster and more efficient alternative for text diffusion models through innovative token editing techniques.
- 6G-Bench: An Open Benchmark for Semantic Communication and Network-Level Reasoning with Foundation Models in AI-Native 6G Networks (viability: 6): https://sciencetostartup.com/paper/6g-bench-an-open-benchmark-for-semantic-communication-and-network-level-reasoning-with-foundation-models-in-ai-native-6g - 6G-Bench provides an open benchmark dataset to evaluate semantic communication in emerging AI-native 6G networks.
- Enhancing Genetic Algorithms with Graph Neural Networks: A Timetabling Case Study (viability: 5): https://sciencetostartup.com/paper/enhancing-genetic-algorithms-with-graph-neural-networks-a-timetabling-case-study - Integrate Genetic Algorithms with Graph Neural Networks to enhance timetabling optimization.
- OSCAR: Optimization-Steered Agentic Planning for Composed Image Retrieval (viability: 6): https://sciencetostartup.com/paper/oscar-optimization-steered-agentic-planning-for-composed-image-retrieval - OSCAR offers an efficient optimization-steered agentic planning solution for composed image retrieval, outperforming current methods with only minimal training data.
- An Attention Mechanism for Robust Multimodal Integration in a Global Workspace Architecture (viability: 4): https://sciencetostartup.com/paper/an-attention-mechanism-for-robust-multimodal-integration-in-a-global-workspace-architecture - Develop an attention mechanism for improved robustness in multimodal integration systems based on Global Workspace Theory.
- Agent-Supported Foresight for AI Systemic Risks: AI Agents for Breadth, Experts for Judgment (viability: 5): https://sciencetostartup.com/paper/agent-supported-foresight-for-ai-systemic-risks-ai-agents-for-breadth-experts-for-judgment - Hybrid foresight platform integrating AI agents and human expertise for assessing AI systemic risks.
- GOT-Edit: Geometry-Aware Generic Object Tracking via Online Model Editing (viability: 6): https://sciencetostartup.com/paper/got-edit-geometry-aware-generic-object-tracking-via-online-model-editing - Develop an advanced object tracking system using geometric cues for enhanced accuracy and robustness across diverse environments.
- GISA: A Benchmark for General Information-Seeking Assistant (viability: 6): https://sciencetostartup.com/paper/gisa-a-benchmark-for-general-information-seeking-assistant - GISA provides a comprehensive benchmark for evaluating search agents on real-world info-seeking tasks.
- Dialogue Model Optimization via Agent Game and Adaptive Tree-based GRPO (viability: 3): https://sciencetostartup.com/paper/dialogue-model-optimization-via-agent-game-and-adaptive-tree-based-grpo - Innovative RL framework for dialogue agents using tree-based optimization to enhance long-term interaction value.
- TreeTensor: Boost AI System on Nested Data with Constrained Tree-Like Tensor (viability: 7): https://sciencetostartup.com/paper/treetensor-boost-ai-system-on-nested-data-with-constrained-tree-like-tensor - TreeTensor enhances handling of nested data in AI systems with a constrained tree-like tensor structure for efficient processing.
- CLEAR: A Knowledge-Centric Vessel Trajectory Analysis Platform (viability: 5): https://sciencetostartup.com/paper/clear-a-knowledge-centric-vessel-trajectory-analysis-platform - CLEAR offers a vessel trajectory analysis platform transforming raw AIS data into complete knowledge graphs using LLMs.
- From Assistant to Double Agent: Formalizing and Benchmarking Attacks on OpenClaw for Personalized Local AI Agent (viability: 6): https://sciencetostartup.com/paper/from-assistant-to-double-agent-formalizing-and-benchmarking-attacks-on-openclaw-for-personalized-local-ai-agent - PASB offers a comprehensive security evaluation framework for personalized AI agents, revealing critical vulnerabilities in systems like OpenClaw.
- SCOUT-RAG: Scalable and Cost-Efficient Unifying Traversal for Agentic Graph-RAG over Distributed Domains (viability: 3): https://sciencetostartup.com/paper/scout-rag-scalable-and-cost-efficient-unifying-traversal-for-agentic-graph-rag-over-distributed-domains - SCOUT-RAG optimizes LLM reasoning in distributed settings using a novel agentic Graph-RAG framework to efficiently retrieve structured knowledge.
- BiManiBench: A Hierarchical Benchmark for Evaluating Bimanual Coordination of Multimodal Large Language Models (viability: 6): https://sciencetostartup.com/paper/bimanibench-a-hierarchical-benchmark-for-evaluating-bimanual-coordination-of-multimodal-large-language-models - Develop a benchmarking tool for evaluating bimanual coordination in multimodal language models.
- Altruism and Fair Objective in Mixed-Motive Markov games (viability: 3): https://sciencetostartup.com/paper/altruism-and-fair-objective-in-mixed-motive-markov-games - Introducing fairer cooperation in multi-agent systems through Proportional Fairness and fair altruistic utility formulation.
- Reinforcement Learning with Backtracking Feedback (viability: 5): https://sciencetostartup.com/paper/reinforcement-learning-with-backtracking-feedback - Reinforcement Learning with Backtracking Feedback enhances LLM safety against adversarial attacks using dynamic error correction.
- Learning Human-Like Badminton Skills for Humanoid Robots (viability: 6): https://sciencetostartup.com/paper/learning-human-like-badminton-skills-for-humanoid-robots - Develop humanoid robots with human-like badminton skills using a reinforcement learning framework for real-world applications.
- Circuit Representations of Random Forests with Applications to XAI (viability: 6): https://sciencetostartup.com/paper/circuit-representations-of-random-forests-with-applications-to-xai - Transform random forest classifiers into efficient circuits for enhanced explainability and decision analysis.
- Does Your Reasoning Model Implicitly Know When to Stop Thinking? (viability: 5): https://sciencetostartup.com/paper/does-your-reasoning-model-implicitly-know-when-to-stop-thinking - Optimize reasoning models for efficiency without sacrificing accuracy.
- UrbanGraphEmbeddings: Learning and Evaluating Spatially Grounded Multimodal Embeddings for Urban Science (viability: 7): https://sciencetostartup.com/paper/urbangraphembeddings-learning-and-evaluating-spatially-grounded-multimodal-embeddings-for-urban-science - UrbanGraphEmbeddings offers advanced spatially grounded multimodal embeddings to enhance urban science applications.
- Regime Change Hypothesis: Foundations for Decoupled Dynamics in Neural Network Training (viability: 3): https://sciencetostartup.com/paper/regime-change-hypothesis-foundations-for-decoupled-dynamics-in-neural-network-training - Research on neural network training dynamics without immediate product implications.
- Latent Reasoning with Supervised Thinking States (viability: 3): https://sciencetostartup.com/paper/latent-reasoning-with-supervised-thinking-states - Introducing Thinking States, a method optimizing CoT reasoning by processing during input for enhanced LLM task performance.
- Moral Sycophancy in Vision Language Models (viability: 4): https://sciencetostartup.com/paper/moral-sycophancy-in-vision-language-models - Develop principled strategies to improve ethical consistency in multimodal AI by addressing sycophantic behavior in Vision-Language Models.
- Grokking in Linear Models for Logistic Regression (viability: 4): https://sciencetostartup.com/paper/grokking-in-linear-models-for-logistic-regression - Explore the dynamics of grokking in simple linear models for robust binary classification.
- Automatic Generation of Polynomial Symmetry Breaking Constraints (viability: 3): https://sciencetostartup.com/paper/automatic-generation-of-polynomial-symmetry-breaking-constraints - Develop a tool to add polynomial symmetry breaking constraints in integer programming to improve solver efficiency.
- The Vibe-Automation of Automation: A Proactive Education Framework for Computer Science in the Age of Generative AI (viability: 2): https://sciencetostartup.com/paper/the-vibe-automation-of-automation-a-proactive-education-framework-for-computer-science-in-the-age-of-generative-ai - An educational framework reimagining computer science training in light of generative AI's epistemological shift.
- Trust-Based Incentive Mechanisms in Semi-Decentralized Federated Learning Systems (viability: 3): https://sciencetostartup.com/paper/trust-based-incentive-mechanisms-in-semi-decentralized-federated-learning-systems - Incentive mechanism using trust scores and blockchain to enhance federated learning systems.
- Noise Stability of Transformer Models (viability: 3): https://sciencetostartup.com/paper/noise-stability-of-transformer-models - The paper proposes noise stability as a new metric for understanding and improving Transformer model robustness.
- STEP: Warm-Started Visuomotor Policies with Spatiotemporal Consistency Prediction (viability: 7): https://sciencetostartup.com/paper/step-warm-started-visuomotor-policies-with-spatiotemporal-consistency-prediction - Develop a low-latency visuomotor policy acceleration tool for robotic manipulation using STEP's spatiotemporal consistency prediction to enhance action success rates.
- Learning in Context, Guided by Choice: A Reward-Free Paradigm for Reinforcement Learning with Transformers (viability: 6): https://sciencetostartup.com/paper/learning-in-context-guided-by-choice-a-reward-free-paradigm-for-reinforcement-learning-with-transformers - Leverage transformer models to perform reinforcement learning using preference feedback instead of explicit rewards.
- Tutti: Expressive Multi-Singer Synthesis via Structure-Level Timbre Control and Vocal Texture Modeling (viability: 6): https://sciencetostartup.com/paper/tutti-expressive-multi-singer-synthesis-via-structure-level-timbre-control-and-vocal-texture-modeling - Tutti offers advanced multi-singer synthesis with structure-level timbre control for more realistic choral music generation.
- Generating Adversarial Events: A Motion-Aware Point Cloud Framework (viability: 5): https://sciencetostartup.com/paper/generating-adversarial-events-a-motion-aware-point-cloud-framework - Develop a framework to improve the security of event-based perception systems by generating adversarial events using motion-aware techniques.
- Investigating Writing Professionals' Relationships with Generative AI: How Combined Perceptions of Rivalry and Collaboration Shape Work Practices and Outcomes (viability: 2): https://sciencetostartup.com/paper/investigating-writing-professionals-relationships-with-generative-ai-how-combined-perceptions-of-rivalry-and-collaborati - Exploring how professional writers perceive and interact with Generative AI to enhance their work practices and outcomes.
- Weak-Driven Learning: How Weak Agents make Strong Agents Stronger (viability: 5): https://sciencetostartup.com/paper/weak-driven-learning-how-weak-agents-make-strong-agents-stronger - WMSS leverages weak model checkpoints to enhance post-training optimization, enabling improved performance without additional inference cost.
- Sparsity-Aware Evolution for Model Merging (viability: 5): https://sciencetostartup.com/paper/sparsity-aware-evolution-for-model-merging - Develop a tool to improve model merging reliability through sparsity-aware evolutionary cycles.
- Large Language Models in Peer-Run Community Behavioral Health Services: Understanding Peer Specialists and Service Users' Perspectives on Opportunities, Risks, and Mitigation Strategies (viability: 3): https://sciencetostartup.com/paper/large-language-models-in-peer-run-community-behavioral-health-services-understanding-peer-specialists-and-service-users- - Integrating LLMs into peer-run behavioral health support presents opportunities and challenges for enhancing support dynamics.
- Self-Supervised Bootstrapping of Action-Predictive Embodied Reasoning (viability: 6): https://sciencetostartup.com/paper/self-supervised-bootstrapping-of-action-predictive-embodied-reasoning - R&B-EnCoRe refines embodied reasoning for Vision-Language-Action models using self-supervised learning, achieving notable performance gains.
- The Confidence Manifold: Geometric Structure of Correctness Representations in Language Models (viability: 2): https://sciencetostartup.com/paper/the-confidence-manifold-geometric-structure-of-correctness-representations-in-language-models - Explore the geometric structure of correctness in language models to understand internal error detection mechanisms.
- Robustness of Vision Language Models Against Split-Image Harmful Input Attacks (viability: 5): https://sciencetostartup.com/paper/robustness-of-vision-language-models-against-split-image-harmful-input-attacks - Develop a security enhancement tool for vision-language models focused on defending against split-image attacks.
- Interpretable Failure Analysis in Multi-Agent Reinforcement Learning Systems (viability: 2): https://sciencetostartup.com/paper/interpretable-failure-analysis-in-multi-agent-reinforcement-learning-systems - Develop a framework for interpretable failure analysis in multi-agent reinforcement learning systems to enhance diagnostics in safety-critical applications.
- Emergent Search and Backtracking in Latent Reasoning Models (viability: 5): https://sciencetostartup.com/paper/emergent-search-and-backtracking-in-latent-reasoning-models - Develop latent reasoning models that simulate human-like thinking processes without verbalizing thoughts.
- VidVec: Unlocking Video MLLM Embeddings for Video-Text Retrieval (viability: 8): https://sciencetostartup.com/paper/vidvec-unlocking-video-mllm-embeddings-for-video-text-retrieval - VidVec enhances video-text retrieval by leveraging intermediate MLLM layers for state-of-the-art, zero-shot performance.
- Online Domain-aware LLM Decoding for Continual Domain Evolution (viability: 6): https://sciencetostartup.com/paper/online-domain-aware-llm-decoding-for-continual-domain-evolution - Online Domain-aware Decoding (ODD) enhances LLM adaptability to evolving domains without retraining, using probability-level fusion and adaptive modulation.
- Shared LoRA Subspaces for almost Strict Continual Learning (viability: 6): https://sciencetostartup.com/paper/shared-lora-subspaces-for-almost-strict-continual-learning - Revolutionize continual learning with Share: a scalable, efficient model replacing multiple LoRA adapters for diverse tasks and modalities.
- DyTopo: Dynamic Topology Routing for Multi-Agent Reasoning via Semantic Matching (viability: 6): https://sciencetostartup.com/paper/dytopo-dynamic-topology-routing-for-multi-agent-reasoning-via-semantic-matching - DyTopo enhances multi-agent systems by dynamically optimizing communication patterns for improved reasoning and problem-solving.
- CommCP: Efficient Multi-Agent Coordination via LLM-Based Communication with Conformal Prediction (viability: 7): https://sciencetostartup.com/paper/commcp-efficient-multi-agent-coordination-via-llm-based-communication-with-conformal-prediction - CommCP enhances multi-agent robot coordination through reliable LLM-based communication calibrated with conformal prediction.
- Learning Query-Aware Budget-Tier Routing for Runtime Agent Memory (viability: 8): https://sciencetostartup.com/paper/learning-query-aware-budget-tier-routing-for-runtime-agent-memory - BudgetMem provides a runtime memory framework for LLMs with query-aware budget-tier routing to optimize performance-cost trade-offs.
- Learning Event-Based Shooter Models from Virtual Reality Experiments (viability: 6): https://sciencetostartup.com/paper/learning-event-based-shooter-models-from-virtual-reality-experiments - Develop a scalable event-based simulator using VR data to model and evaluate school security interventions.
- Correctness-Optimized Residual Activation Lens (CORAL): Transferrable and Calibration-Aware Inference-Time Steering (viability: 6): https://sciencetostartup.com/paper/correctness-optimized-residual-activation-lens-coral-transferrable-and-calibration-aware-inference-time-steering - CORAL enhances LLM inference-time accuracy and calibration with a compute-efficient steering method.
- Optimism Stabilizes Thompson Sampling for Adaptive Inference (viability: 2): https://sciencetostartup.com/paper/optimism-stabilizes-thompson-sampling-for-adaptive-inference - Adaptive inference for multi-armed bandits using optimism-enhanced Thompson sampling.
- GenArena: How Can We Achieve Human-Aligned Evaluation for Visual Generation Tasks? (viability: 7): https://sciencetostartup.com/paper/genarena-how-can-we-achieve-human-aligned-evaluation-for-visual-generation-tasks - GenArena offers a human-aligned evaluation framework for visual generation tasks, significantly improving accuracy and simplifying model benchmarking.
- AgenticPay: A Multi-Agent LLM Negotiation System for Buyer-Seller Transactions (viability: 6): https://sciencetostartup.com/paper/agenticpay-a-multi-agent-llm-negotiation-system-for-buyer-seller-transactions - AgenticPay provides a negotiation framework for AI-driven buyer-seller transactions using LLMs.
- Speech Emotion Recognition Leveraging OpenAI's Whisper Representations and Attentive Pooling Methods (viability: 8): https://sciencetostartup.com/paper/speech-emotion-recognition-leveraging-openai-s-whisper-representations-and-attentive-pooling-methods - Speech Emotion Recognition using Whisper's attentive pooling for efficient emotion detection.
- Diamond Maps: Efficient Reward Alignment via Stochastic Flow Maps (viability: 3): https://sciencetostartup.com/paper/diamond-maps-efficient-reward-alignment-via-stochastic-flow-maps - Develop generative models that efficiently align with user preferences during inference, improving adaptability and scaling.
- RISE-Video: Can Video Generators Decode Implicit World Rules? (viability: 6): https://sciencetostartup.com/paper/rise-video-can-video-generators-decode-implicit-world-rules - Develop a reasoning-oriented benchmark for Text-Image-to-Video synthesis to assess model intelligence beyond aesthetics.
- Geographically-aware Transformer-based Traffic Forecasting for Urban Motorway Digital Twins (viability: 2): https://sciencetostartup.com/paper/geographically-aware-transformer-based-traffic-forecasting-for-urban-motorway-digital-twins - Develop a Transformer-based traffic forecasting model to enhance real-time motorway traffic management in digital twins.
- Clifford Kolmogorov-Arnold Networks (viability: 3): https://sciencetostartup.com/paper/clifford-kolmogorov-arnold-networks - Develop a flexible architecture for function approximation in Clifford algebra spaces using ClKAN.
- Inverse Depth Scaling From Most Layers Being Similar (viability: 2): https://sciencetostartup.com/paper/inverse-depth-scaling-from-most-layers-being-similar - Research on inverse depth scaling in LLMs points to new architectural efficiencies.
- LSA: Localized Semantic Alignment for Enhancing Temporal Consistency in Traffic Video Generation (viability: 6): https://sciencetostartup.com/paper/lsa-localized-semantic-alignment-for-enhancing-temporal-consistency-in-traffic-video-generation - Enhance video generation models for autonomous driving by improving temporal consistency through localized semantic alignment.
- Learning to Share: Selective Memory for Efficient Parallel Agentic Systems (viability: 8): https://sciencetostartup.com/paper/learning-to-share-selective-memory-for-efficient-parallel-agentic-systems - Launch a system for efficient parallel agentic operations using selective memory to reduce computational cost and enhance task performance.
- Better Source, Better Flow: Learning Condition-Dependent Source Distribution for Flow Matching (viability: 6): https://sciencetostartup.com/paper/better-source-better-flow-learning-condition-dependent-source-distribution-for-flow-matching - Develop a text-to-image generation tool using condition-dependent source distribution design for improved flow matching.
- Compound Deception in Elite Peer Review: A Failure Mode Taxonomy of 100 Fabricated Citations at NeurIPS 2025 (viability: 4): https://sciencetostartup.com/paper/compound-deception-in-elite-peer-review-a-failure-mode-taxonomy-of-100-fabricated-citations-at-neurips-2025 - Develop an automated tool for mandatory citation verification to combat fabricated citations in academic papers.
- Quantum Reinforcement Learning with Transformers for the Capacitated Vehicle Routing Problem (viability: 2): https://sciencetostartup.com/paper/quantum-reinforcement-learning-with-transformers-for-the-capacitated-vehicle-routing-problem - Leverage quantum-classical reinforcement learning to optimize vehicle routing solutions.
- xList-Hate: A Checklist-Based Framework for Interpretable and Generalizable Hate Speech Detection (viability: 7): https://sciencetostartup.com/paper/xlist-hate-a-checklist-based-framework-for-interpretable-and-generalizable-hate-speech-detection - xList-Hate offers an interpretable framework for robust cross-domain hate speech detection, enabling fine-grained content moderation.
- DLM-Scope: Mechanistic Interpretability of Diffusion Language Models via Sparse Autoencoders (viability: 5): https://sciencetostartup.com/paper/dlm-scope-mechanistic-interpretability-of-diffusion-language-models-via-sparse-autoencoders - Develop an interpretability framework for Diffusion Language Models using Sparse Autoencoders to enhance feature extraction and intervention capabilities.
- BABE: Biology Arena BEnchmark (viability: 5): https://sciencetostartup.com/paper/babe-biology-arena-benchmark - Evaluate AI systems' biological reasoning with the comprehensive BABE benchmark.
- DARWIN: Dynamic Agentically Rewriting Self-Improving Network (viability: 6): https://sciencetostartup.com/paper/darwin-dynamic-agentically-rewriting-self-improving-network - DARWIN optimizes GPT model training through a novel genetic-algorithm approach with dynamic self-improvement.
- OmniVideo-R1: Reinforcing Audio-visual Reasoning with Query Intention and Modality Attention (viability: 5): https://sciencetostartup.com/paper/omnivideo-r1-reinforcing-audio-visual-reasoning-with-query-intention-and-modality-attention - OmniVideo-R1 enhances audio-visual reasoning in AI models through a reinforced framework using omnimodal cues.
- FHAIM: Fully Homomorphic AIM For Private Synthetic Data Generation (viability: 5): https://sciencetostartup.com/paper/fhaim-fully-homomorphic-aim-for-private-synthetic-data-generation - FHAIM enables secure synthetic data generation using homomorphic encryption for private data silos.
- Learning Compact Boolean Networks (viability: 5): https://sciencetostartup.com/paper/learning-compact-boolean-networks - Optimize Boolean networks for resource-constrained settings with novel compact architectures and adaptive strategies.
- TKG-Thinker: Towards Dynamic Reasoning over Temporal Knowledge Graphs via Agentic Reinforcement Learning (viability: 6): https://sciencetostartup.com/paper/tkg-thinker-towards-dynamic-reasoning-over-temporal-knowledge-graphs-via-agentic-reinforcement-learning - TKG-Thinker offers enhanced dynamic reasoning for temporal knowledge graphs using agentic reinforcement learning.
- STProtein: predicting spatial protein expression from multi-omics data (viability: 6): https://sciencetostartup.com/paper/stprotein-predicting-spatial-protein-expression-from-multi-omics-data - Predict spatial protein expression from spatial multi-omics data to accelerate life sciences research.
- NEX: Neuron Explore-Exploit Scoring for Label-Free Chain-of-Thought Selection and Model Ranking (viability: 6): https://sciencetostartup.com/paper/nex-neuron-explore-exploit-scoring-for-label-free-chain-of-thought-selection-and-model-ranking - NEX provides an unsupervised scoring framework for efficient model selection by assessing neuron activity phases.
- Bagging-Based Model Merging for Robust General Text Embeddings (viability: 6): https://sciencetostartup.com/paper/bagging-based-model-merging-for-robust-general-text-embeddings - Develop a robust text embedding tool using bagging-based model merging for better in-domain and out-of-domain performance.
- ReText: Text Boosts Generalization in Image-Based Person Re-identification (viability: 6): https://sciencetostartup.com/paper/retext-text-boosts-generalization-in-image-based-person-re-identification - ReText enhances image-based person re-identification by integrating textual descriptions for improved cross-domain generalization.
- Automated Customization of LLMs for Enterprise Code Repositories Using Semantic Scopes (viability: 7): https://sciencetostartup.com/paper/automated-customization-of-llms-for-enterprise-code-repositories-using-semantic-scopes - Auto-customized LLMs for efficient and precise code completion in proprietary repositories.
- RL-VLA$^3$: Reinforcement Learning VLA Accelerating via Full Asynchronism (viability: 5): https://sciencetostartup.com/paper/rl-vla-3-reinforcement-learning-vla-accelerating-via-full-asynchronism - Develop an asynchronous RL framework to accelerate VLA models by optimizing resource utilization and throughput.
- RocqSmith: Can Automatic Optimization Forge Better Proof Agents? (viability: 3): https://sciencetostartup.com/paper/rocqsmith-can-automatic-optimization-forge-better-proof-agents - Exploring the automation of optimization for AI agents in automated theorem proving.
- TimelyFreeze: Adaptive Parameter Freezing Mechanism for Pipeline Parallelism (viability: 3): https://sciencetostartup.com/paper/timelyfreeze-adaptive-parameter-freezing-mechanism-for-pipeline-parallelism - TimelyFreeze optimizes large model training by reducing pipeline bubbles through adaptive parameter freezing.
- LeakBoost: Perceptual-Loss-Based Membership Inference Attack (viability: 5): https://sciencetostartup.com/paper/leakboost-perceptual-loss-based-membership-inference-attack - LeakBoost enhances privacy risk assessment by boosting membership inference attacks through perceptual-loss-based model probing.
- Learning to Inject: Automated Prompt Injection via Reinforcement Learning (viability: 6): https://sciencetostartup.com/paper/learning-to-inject-automated-prompt-injection-via-reinforcement-learning - Automated prompt injection framework using reinforcement learning to enhance LLM vulnerability exploitation.
- CSRv2: Unlocking Ultra-Sparse Embeddings (viability: 6): https://sciencetostartup.com/paper/csrv2-unlocking-ultra-sparse-embeddings - CSRv2 offers ultra-sparse embedding technology that reduces memory and compute costs for real-time AI deployment.
- CompactRAG: Reducing LLM Calls and Token Overhead in Multi-Hop Question Answering (viability: 8): https://sciencetostartup.com/paper/compactrag-reducing-llm-calls-and-token-overhead-in-multi-hop-question-answering - CompactRAG revolutionizes multi-hop question answering by reducing LLM calls and token overhead, offering a cost-efficient solution for knowledge-intensive reasoning.
- Anchored Policy Optimization: Mitigating Exploration Collapse Via Support-Constrained Rectification (viability: 4): https://sciencetostartup.com/paper/anchored-policy-optimization-mitigating-exploration-collapse-via-support-constrained-rectification - Develop a reinforcement learning optimization tool to avoid exploration collapse using Anchored Policy Optimization.
- Towards Green AI: Decoding the Energy of LLM Inference in Software Development (viability: 6): https://sciencetostartup.com/paper/towards-green-ai-decoding-the-energy-of-llm-inference-in-software-development - Optimize LLM inference to significantly reduce energy consumption in software development tools.
- OmniMoE: An Efficient MoE by Orchestrating Atomic Experts at Scale (viability: 7): https://sciencetostartup.com/paper/omnimoe-an-efficient-moe-by-orchestrating-atomic-experts-at-scale - OmniMoE offers a groundbreaking fine-grained MoE framework that drastically reduces inference latency while enhancing accuracy, outperforming existing models.
- Nonlinearity as Rank: Generative Low-Rank Adapter with Radial Basis Functions (viability: 6): https://sciencetostartup.com/paper/nonlinearity-as-rank-generative-low-rank-adapter-with-radial-basis-functions - GenLoRA offers a parameter-efficient low-rank adapter by generating basis vectors through radial basis functions, optimizing for fine-tuning performance.
- Poster: Camera Tampering Detection for Outdoor IoT Systems (viability: 6): https://sciencetostartup.com/paper/poster-camera-tampering-detection-for-outdoor-iot-systems - Develop a camera tampering detection system for outdoor IoT cameras using rule-based and deep-learning approaches with open datasets.
- Mining Generalizable Activation Functions (viability: 4): https://sciencetostartup.com/paper/mining-generalizable-activation-functions - Leveraging evolutionary search to discover generalizable activation functions using modern pipelines like AlphaEvolve.
- Exploring AI-Augmented Sensemaking of Patient-Generated Health Data: A Mixed-Method Study with Healthcare Professionals in Cardiac Risk Reduction (viability: 5): https://sciencetostartup.com/paper/exploring-ai-augmented-sensemaking-of-patient-generated-health-data-a-mixed-method-study-with-healthcare-professionals-i - AI-powered tool for healthcare professionals to efficiently process and understand patient-generated health data.
- HyperPotter: Spell the Charm of High-Order Interactions in Audio Deepfake Detection (viability: 6): https://sciencetostartup.com/paper/hyperpotter-spell-the-charm-of-high-order-interactions-in-audio-deepfake-detection - Introducing HyperPotter, a hypergraph-based audio deepfake detection tool leveraging high-order interactions to significantly improve detection accuracy.
- Stable but Wrong: When More Data Degrades Scientific Conclusions (viability: 3): https://sciencetostartup.com/paper/stable-but-wrong-when-more-data-degrades-scientific-conclusions - A study showing how more data can degrade scientific conclusions due to unobservable errors in observational datasets.
- Graph-based Agent Memory: Taxonomy, Techniques, and Applications (viability: 4): https://sciencetostartup.com/paper/graph-based-agent-memory-taxonomy-techniques-and-applications - Build more efficient and reliable graph-based agent memory systems for AI applications spanning multi-turn dialogues to scientific discovery.
- Probabilistic Multi-Regional Solar Power Forecasting with Any-Quantile Recurrent Neural Networks (viability: 6): https://sciencetostartup.com/paper/probabilistic-multi-regional-solar-power-forecasting-with-any-quantile-recurrent-neural-networks - Develop an any-quantile probabilistic framework for multi-regional PV power generation forecasting to improve energy management in renewable power systems.
- Enhancing Personality Recognition by Comparing the Predictive Power of Traits, Facets, and Nuances (viability: 5): https://sciencetostartup.com/paper/enhancing-personality-recognition-by-comparing-the-predictive-power-of-traits-facets-and-nuances - Enhance personality recognition models by leveraging nuance-level predictions from audiovisual data for improved accuracy.
- Generative Ontology: When Structured Knowledge Learns to Create (viability: 6): https://sciencetostartup.com/paper/generative-ontology-when-structured-knowledge-learns-to-create - Generative Ontology combines structured knowledge with AI creativity to produce novel, structured outputs.
- AI chatbots versus human healthcare professionals: a systematic review and meta-analysis of empathy in patient care (viability: 3): https://sciencetostartup.com/paper/ai-chatbots-versus-human-healthcare-professionals-a-systematic-review-and-meta-analysis-of-empathy-in-patient-care - Evaluate if AI chatbots can be more empathetic than human healthcare professionals through text-based communication.
- Reactive Knowledge Representation and Asynchronous Reasoning (viability: 6): https://sciencetostartup.com/paper/reactive-knowledge-representation-and-asynchronous-reasoning - Reactive Circuits enable real-time asynchronous probabilistic reasoning for dynamic environments, optimizing computational efficiency.
- Mode-Dependent Rectification for Stable PPO Training (viability: 6): https://sciencetostartup.com/paper/mode-dependent-rectification-for-stable-ppo-training - Stabilize PPO training in visual reinforcement learning using Mode-Dependent Rectification.
- Path-Guided Flow Matching for Dataset Distillation (viability: 5): https://sciencetostartup.com/paper/path-guided-flow-matching-for-dataset-distillation - Develop a more efficient dataset distillation method using Path-Guided Flow Matching to enhance model training.
- Shiva-DiT: Residual-Based Differentiable Top-$k$ Selection for Efficient Diffusion Transformers (viability: 5): https://sciencetostartup.com/paper/shiva-dit-residual-based-differentiable-top-k-selection-for-efficient-diffusion-transformers - Shiva-DiT offers efficient Diffusion Transformers through innovative top-$k$ selection reducing computation cost.
- BhashaSetu: Cross-Lingual Knowledge Transfer from High-Resource to Extreme Low-Resource Languages (viability: 7): https://sciencetostartup.com/paper/bhashasetu-cross-lingual-knowledge-transfer-from-high-resource-to-extreme-low-resource-languages - Enable cross-lingual NLP tools for low-resource languages using knowledge transfer from high-resource languages.
- CAViT -- Channel-Aware Vision Transformer for Dynamic Feature Fusion (viability: 7): https://sciencetostartup.com/paper/cavit-channel-aware-vision-transformer-for-dynamic-feature-fusion - CAViT enhances Vision Transformers with dynamic channel-wise attention for superior and efficient image analysis.
- Emulating Aggregate Human Choice Behavior and Biases with GPT Conversational Agents (viability: 5): https://sciencetostartup.com/paper/emulating-aggregate-human-choice-behavior-and-biases-with-gpt-conversational-agents - Develop conversational AI agents that replicate human decision-making behaviors, including biases, in various contexts.
- Multi-Task GRPO: Reliable LLM Reasoning Across Tasks (viability: 4): https://sciencetostartup.com/paper/multi-task-grpo-reliable-llm-reasoning-across-tasks - Improve LLM task adaptability and efficiency with Multi-Task GRPO, enhancing performance across diverse tasks through dynamic task weighting and efficient training.
- Reasoning-guided Collaborative Filtering with Language Models for Explainable Recommendation (viability: 8): https://sciencetostartup.com/paper/reasoning-guided-collaborative-filtering-with-language-models-for-explainable-recommendation - Develop an efficient and scalable explainable recommendation system using reasoning-guided collaborative filtering with language models.
- When Shared Knowledge Hurts: Spectral Over-Accumulation in Model Merging (viability: 7): https://sciencetostartup.com/paper/when-shared-knowledge-hurts-spectral-over-accumulation-in-model-merging - Singular Value Calibration improves model merging processes by balancing spectral accumulation, enhancing system performance without retraining.
- Conditional Diffusion Guidance under Hard Constraint: A Stochastic Analysis Approach (viability: 5): https://sciencetostartup.com/paper/conditional-diffusion-guidance-under-hard-constraint-a-stochastic-analysis-approach - Develop a conditional diffusion guidance framework for generating samples under hard constraints in safety-critical applications.
- Split Personality Training: Revealing Latent Knowledge Through Alternate Personalities (viability: 7): https://sciencetostartup.com/paper/split-personality-training-revealing-latent-knowledge-through-alternate-personalities - A new model fine-tuning technique to expose hidden knowledge and biases in language models through split personality training.
- AI Agent Systems for Supply Chains: Structured Decision Prompts and Memory Retrieval (viability: 5): https://sciencetostartup.com/paper/ai-agent-systems-for-supply-chains-structured-decision-prompts-and-memory-retrieval - Develop an AI agent system for optimized inventory management in supply chains leveraging LLM-based multi-agent frameworks.
- Capture the Flags: Family-Based Evaluation of Agentic LLMs via Semantics-Preserving Transformations (viability: 4): https://sciencetostartup.com/paper/capture-the-flags-family-based-evaluation-of-agentic-llms-via-semantics-preserving-transformations - A tool for evaluating LLM robustness using semantically-preserving program transformations in cybersecurity tasks.
- A Unified Multimodal Framework for Dataset Construction and Model-Based Diagnosis of Ameloblastoma (viability: 7): https://sciencetostartup.com/paper/a-unified-multimodal-framework-for-dataset-construction-and-model-based-diagnosis-of-ameloblastoma - AI framework for improved ameloblastoma diagnosis using a curated multimodal dataset.
- DECO: Decoupled Multimodal Diffusion Transformer for Bimanual Dexterous Manipulation with a Plugin Tactile Adapter (viability: 6): https://sciencetostartup.com/paper/deco-decoupled-multimodal-diffusion-transformer-for-bimanual-dexterous-manipulation-with-a-plugin-tactile-adapter - Develop a multimodal diffusion transformer for enhanced bimanual dexterous manipulation using DECO framework and accompanying dataset.
- SDFP: Speculative Decoding with FIT-Pruned Models for Training-Free and Plug-and-Play LLM Acceleration (viability: 7): https://sciencetostartup.com/paper/sdfp-speculative-decoding-with-fit-pruned-models-for-training-free-and-plug-and-play-llm-acceleration - SDFP offers a training-free framework to accelerate LLMs by layer pruning, optimizing low-latency multimedia applications.
- XEmoGPT: An Explainable Multimodal Emotion Recognition Framework with Cue-Level Perception and Reasoning (viability: 8): https://sciencetostartup.com/paper/xemogpt-an-explainable-multimodal-emotion-recognition-framework-with-cue-level-perception-and-reasoning - XEmoGPT revolutionizes emotion recognition in multimedia by perceiving and reasoning fine-grained emotional cues with advanced datasets and benchmarks.
- Transport and Merge: Cross-Architecture Merging for Large Language Models (viability: 5): https://sciencetostartup.com/paper/transport-and-merge-cross-architecture-merging-for-large-language-models - Cross-architecture model merging framework enhances small LLMs by transferring knowledge from large models.
- LinguistAgent: A Reflective Multi-Model Platform for Automated Linguistic Annotation (viability: 6): https://sciencetostartup.com/paper/linguistagent-a-reflective-multi-model-platform-for-automated-linguistic-annotation - LinguistAgent streamlines complex linguistic annotation processes for research using a reflective multi-model platform with automated dual-agent workflow.
- Sovereign-by-Design A Reference Architecture for AI and Blockchain Enabled Systems (viability: 3): https://sciencetostartup.com/paper/sovereign-by-design-a-reference-architecture-for-ai-and-blockchain-enabled-systems - A reference architecture for integrating digital sovereignty in AI and blockchain systems.
- Phi-Former: A Pairwise Hierarchical Approach for Compound-Protein Interactions Prediction (viability: 6): https://sciencetostartup.com/paper/phi-former-a-pairwise-hierarchical-approach-for-compound-protein-interactions-prediction - Phi-Former revolutionizes drug discovery by accurately predicting compound-protein interactions using hierarchical interaction representation learning.
- Thermodynamic Limits of Physical Intelligence (viability: 3): https://sciencetostartup.com/paper/thermodynamic-limits-of-physical-intelligence - Develop an efficiency framework for AI systems that links intelligence to energy consumption through novel thermodynamic metrics.
- Ontology-Driven Robotic Specification Synthesis (viability: 4): https://sciencetostartup.com/paper/ontology-driven-robotic-specification-synthesis - Develop a methodology utilizing ontologies and stochastic timed Petri nets for autonomous robotic specification synthesis in safety-critical systems.
- Attention Retention for Continual Learning with Vision Transformers (viability: 7): https://sciencetostartup.com/paper/attention-retention-for-continual-learning-with-vision-transformers - A novel framework for attention retention to solve catastrophic forgetting in Vision Transformers, enhancing continual learning.
- DisCa: Accelerating Video Diffusion Transformers with Distillation-Compatible Learnable Feature Caching (viability: 7): https://sciencetostartup.com/paper/disca-accelerating-video-diffusion-transformers-with-distillation-compatible-learnable-feature-caching - Distillation-Compatible Learnable Feature Caching greatly accelerates video diffusion transformers while maintaining quality.
- Benchmarking Affordance Generalization with BusyBox (viability: 7): https://sciencetostartup.com/paper/benchmarking-affordance-generalization-with-busybox - BusyBox provides a standardized benchmark for evaluating affordance generalization in Vision-Language-Action models, complete with resources for easy lab replication.
- Day-Ahead Electricity Price Forecasting for Volatile Markets Using Foundation Models with Regularization Strategy (viability: 7): https://sciencetostartup.com/paper/day-ahead-electricity-price-forecasting-for-volatile-markets-using-foundation-models-with-regularization-strategy - Develop a forecasting tool using time series foundation models to improve electricity price predictions in volatile markets.
- M$^2$-Miner: Multi-Agent Enhanced MCTS for Mobile GUI Agent Data Mining (viability: 8): https://sciencetostartup.com/paper/m-2-miner-multi-agent-enhanced-mcts-for-mobile-gui-agent-data-mining - Automate high-quality GUI agent data mining with a multi-agent MCTS framework for improved mobile interface interaction.
- THOR: Inductive Link Prediction over Hyper-Relational Knowledge Graphs (viability: 8): https://sciencetostartup.com/paper/thor-inductive-link-prediction-over-hyper-relational-knowledge-graphs - An inductive link prediction system for hyper-relational knowledge graphs to enhance reasoning performance across unknown vocabularies.
- Disco: Densely-overlapping Cell Instance Segmentation via Adjacency-aware Collaborative Coloring (viability: 6): https://sciencetostartup.com/paper/disco-densely-overlapping-cell-instance-segmentation-via-adjacency-aware-collaborative-coloring - Develop a tool for cell instance segmentation using gap coloring techniques for complex tissue analysis in digital pathology.
- Reduced-Order Surrogates for Forced Flexible Mesh Coastal-Ocean Models (viability: 2): https://sciencetostartup.com/paper/reduced-order-surrogates-for-forced-flexible-mesh-coastal-ocean-models - Develop a Koopman autoencoder-based surrogate for fast and accurate coastal-ocean modeling enabling long-term predictions.
- H-AdminSim: A Multi-Agent Simulator for Realistic Hospital Administrative Workflows with FHIR Integration (viability: 3): https://sciencetostartup.com/paper/h-adminsim-a-multi-agent-simulator-for-realistic-hospital-administrative-workflows-with-fhir-integration - Develop a simulation framework for automating hospital administrative workflows using FHIR integration.
- Enabling Automatic Disordered Speech Recognition: An Impaired Speech Dataset in the Akan Language (viability: 5): https://sciencetostartup.com/paper/enabling-automatic-disordered-speech-recognition-an-impaired-speech-dataset-in-the-akan-language - Develop a speech recognition tool tailored for disordered Akan language based on a new impaired speech dataset.
- Advancing Opinion Dynamics Modeling with Neural Diffusion-Convection-Reaction Equation (viability: 5): https://sciencetostartup.com/paper/advancing-opinion-dynamics-modeling-with-neural-diffusion-convection-reaction-equation - Develop a physics-informed neural framework for advanced opinion dynamics modeling targeting improvements in opinion forecasting.
- Beyond Length: Context-Aware Expansion and Independence as Developmentally Sensitive Evaluation in Child Utterances (viability: 6): https://sciencetostartup.com/paper/beyond-length-context-aware-expansion-and-independence-as-developmentally-sensitive-evaluation-in-child-utterances - A framework for context-aware evaluation of child utterances using language model metrics to better understand developmentally sensitive communication aspects.
- Assessing Electricity Demand Forecasting with Exogenous Data in Time Series Foundation Models (viability: 4): https://sciencetostartup.com/paper/assessing-electricity-demand-forecasting-with-exogenous-data-in-time-series-foundation-models - Develop a specialized time-series forecasting tool for electricity demand with enhanced performance in variable climates.
- Spider-Sense: Intrinsic Risk Sensing for Efficient Agent Defense with Hierarchical Adaptive Screening (viability: 7): https://sciencetostartup.com/paper/spider-sense-intrinsic-risk-sensing-for-efficient-agent-defense-with-hierarchical-adaptive-screening - Spider-Sense provides efficient intrinsic risk sensing for agent security with minimal false positives.
- Clinical Validation of Medical-based Large Language Model Chatbots on Ophthalmic Patient Queries with LLM-based Evaluation (viability: 6): https://sciencetostartup.com/paper/clinical-validation-of-medical-based-large-language-model-chatbots-on-ophthalmic-patient-queries-with-llm-based-evaluati - Implement an LLM-based tool for ophthalmic patient query responses with clinician and AI evaluation alignment.
- RaBiT: Residual-Aware Binarization Training for Accurate and Efficient LLMs (viability: 3): https://sciencetostartup.com/paper/rabit-residual-aware-binarization-training-for-accurate-and-efficient-llms - RaBiT offers a new quantization framework for LLMs, emphasizing residual-aware training without hardware-intensive setups.
- PATHWAYS: Evaluating Investigation and Context Discovery in AI Web Agents (viability: 5): https://sciencetostartup.com/paper/pathways-evaluating-investigation-and-context-discovery-in-ai-web-agents - Develop PATHWAYS, a benchmark tool for evaluating the context discovery ability of AI web agents in multi-step decision tasks.
- AgentXRay: White-Boxing Agentic Systems via Workflow Reconstruction (viability: 5): https://sciencetostartup.com/paper/agentxray-white-boxing-agentic-systems-via-workflow-reconstruction - AgentXRay transforms opaque agentic systems into interpretable workflows using a search-based optimization framework.
- ProAct: Agentic Lookahead in Interactive Environments (viability: 8): https://sciencetostartup.com/paper/proact-agentic-lookahead-in-interactive-environments - ProAct enables AI agents to excel in long-horizon planning with enhanced lookahead reasoning and stable decision-making.
- GAS: Enhancing Reward-Cost Balance of Generative Model-assisted Offline Safe RL (viability: 6): https://sciencetostartup.com/paper/gas-enhancing-reward-cost-balance-of-generative-model-assisted-offline-safe-rl - Goal-Assisted Stitching enhances decision-making in Offline Safe Reinforcement Learning by improving trajectory stitching and balancing rewards with constraints.
- Formal Synthesis of Certifiably Robust Neural Lyapunov-Barrier Certificates (viability: 6): https://sciencetostartup.com/paper/formal-synthesis-of-certifiably-robust-neural-lyapunov-barrier-certificates - Robust neural Lyapunov-barrier certificates enhance safety and stability in reinforcement learning controllers under dynamic perturbations.
- FlashBlock: Attention Caching for Efficient Long-Context Block Diffusion (viability: 7): https://sciencetostartup.com/paper/flashblock-attention-caching-for-efficient-long-context-block-diffusion - FlashBlock accelerates long-form content generation in generative models by implementing efficient attention caching.
- PieArena: Frontier Language Agents Achieve MBA-Level Negotiation Performance and Reveal Novel Behavioral Differences (viability: 6): https://sciencetostartup.com/paper/piearena-frontier-language-agents-achieve-mba-level-negotiation-performance-and-reveal-novel-behavioral-differences - Develop a negotiation AI system outperforming MBA students using PieArena benchmark.
- Towards a Science of Collective AI: LLM-based Multi-Agent Systems Need a Transition from Blind Trial-and-Error to Rigorous Science (viability: 4): https://sciencetostartup.com/paper/towards-a-science-of-collective-ai-llm-based-multi-agent-systems-need-a-transition-from-blind-trial-and-error-to-rigorou - Develop a systematic framework for optimizing multi-agent systems using a newly proposed metric for collaboration gain.
- Position: Universal Time Series Foundation Models Rest on a Category Error (viability: 2): https://sciencetostartup.com/paper/position-universal-time-series-foundation-models-rest-on-a-category-error - The paper challenges the feasibility of universal time series models, advocating for a paradigm focused on causal control agents and drift adaptation.
- HealthMamba: An Uncertainty-aware Spatiotemporal Graph State Space Model for Effective and Reliable Healthcare Facility Visit Prediction (viability: 6): https://sciencetostartup.com/paper/healthmamba-an-uncertainty-aware-spatiotemporal-graph-state-space-model-for-effective-and-reliable-healthcare-facility-v - HealthMamba leverages uncertainty-aware spatiotemporal graph modeling for predicting healthcare facility visits to optimize resource allocation.
- Hallucination-Resistant Security Planning with a Large Language Model (viability: 5): https://sciencetostartup.com/paper/hallucination-resistant-security-planning-with-a-large-language-model - A framework for using LLMs in security management by mitigating hallucination risks through iterative action refinement.
- Beyond Cosine Similarity (viability: 4): https://sciencetostartup.com/paper/beyond-cosine-similarity - Introducing recos, a mathematically principled alternative to cosine similarity for enhanced semantic analysis in complex embedding spaces.
- EGSS: Entropy-guided Stepwise Scaling for Reliable Software Engineering (viability: 7): https://sciencetostartup.com/paper/egss-entropy-guided-stepwise-scaling-for-reliable-software-engineering - Introduce EGSS to improve efficiency and accuracy in software engineering tasks like code generation and bug fixing.
- Explainable AI: A Combined XAI Framework for Explaining Brain Tumour Detection Models (viability: 5): https://sciencetostartup.com/paper/explainable-ai-a-combined-xai-framework-for-explaining-brain-tumour-detection-models - Develop an integrated Explainable AI framework to enhance interpretability of brain tumour detection models.
- Balanced Anomaly-guided Ego-graph Diffusion Model for Inductive Graph Anomaly Detection (viability: 5): https://sciencetostartup.com/paper/balanced-anomaly-guided-ego-graph-diffusion-model-for-inductive-graph-anomaly-detection - A dynamic graph anomaly detection model designed to improve generalization in fraud and cybersecurity applications by synthesizing balanced anomaly patterns.
- ZeroS: Zero-Sum Linear Attention for Efficient Transformers (viability: 5): https://sciencetostartup.com/paper/zeros-zero-sum-linear-attention-for-efficient-transformers - ZeroS offers a more efficient Transformer attention mechanism for practical sequence modeling tasks.
- Surgery: Mitigating Harmful Fine-Tuning for Large Language Models via Attention Sink (viability: 6): https://sciencetostartup.com/paper/surgery-mitigating-harmful-fine-tuning-for-large-language-models-via-attention-sink - A defense mechanism for large language models that utilizes the sink divergence technique to mitigate harmful fine-tuning and improve model safety.
- Aligning Large Language Model Behavior with Human Citation Preferences (viability: 4): https://sciencetostartup.com/paper/aligning-large-language-model-behavior-with-human-citation-preferences - Develop an AI tool to align LLM citation behaviors with human preferences, enhancing credibility in generated content.
- Traceable Cross-Source RAG for Chinese Tibetan Medicine Question Answering (viability: 6): https://sciencetostartup.com/paper/traceable-cross-source-rag-for-chinese-tibetan-medicine-question-answering - Building a traceable RAG system for reliable Chinese Tibetan medicine Q&A across multiple knowledge bases.
- First Proof (viability: 3): https://sciencetostartup.com/paper/first-proof - A dataset of research-level math questions for evaluating AI systems' mathematics understanding.
- Double-P: Hierarchical Top-P Sparse Attention for Long-Context LLMs (viability: 6): https://sciencetostartup.com/paper/double-p-hierarchical-top-p-sparse-attention-for-long-context-llms - Double-P enhances LLMs with a scalable attention mechanism that significantly improves inference efficiency for long-context applications.
- Benchmarking Artificial Intelligence Models for Daily Coastal Hypoxia Forecasting (viability: 5): https://sciencetostartup.com/paper/benchmarking-artificial-intelligence-models-for-daily-coastal-hypoxia-forecasting - Develop a high-accuracy real-time hypoxia prediction tool for ecosystem management in the Gulf of Mexico.
- EBPO: Empirical Bayes Shrinkage for Stabilizing Group-Relative Policy Optimization (viability: 6): https://sciencetostartup.com/paper/ebpo-empirical-bayes-shrinkage-for-stabilizing-group-relative-policy-optimization - Revolutionize LLM reasoning capabilities with EBPO's stable and efficient policy optimization framework.
- Position: Capability Control Should be a Separate Goal From Alignment (viability: 3): https://sciencetostartup.com/paper/position-capability-control-should-be-a-separate-goal-from-alignment - Propose a defense-in-depth approach for capability control as a distinct goal from alignment in AI systems.
- CoSA: Compressed Sensing-Based Adaptation of Large Language Models (viability: 3): https://sciencetostartup.com/paper/cosa-compressed-sensing-based-adaptation-of-large-language-models - CoSA offers an innovative PEFT technique leveraging compressed sensing for expressive and efficient LLM adaptation.
- TIDE: Temporal Incremental Draft Engine for Self-Improving LLM Inference (viability: 3): https://sciencetostartup.com/paper/tide-temporal-incremental-draft-engine-for-self-improving-llm-inference - TIDE enhances LLM inference with a serving-engine-native framework for online draft adaptation and speculative decoding.
- HugRAG: Hierarchical Causal Knowledge Graph Design for RAG (viability: 5): https://sciencetostartup.com/paper/hugrag-hierarchical-causal-knowledge-graph-design-for-rag - HugRAG is a structured and scalable graph-based RAG framework explicitly modeling causal relationships for better retrieval and reasoning.
- CAST-CKT: Chaos-Aware Spatio-Temporal and Cross-City Knowledge Transfer for Traffic Flow Prediction (viability: 8): https://sciencetostartup.com/paper/cast-ckt-chaos-aware-spatio-temporal-and-cross-city-knowledge-transfer-for-traffic-flow-prediction - CAST-CKT enhances traffic flow prediction across cities with a chaos-aware framework, offering significant performance gains and uncertainty quantification.
- Rethinking Rubric Generation for Improving LLM Judge and Reward Modeling for Open-ended Tasks (viability: 3): https://sciencetostartup.com/paper/rethinking-rubric-generation-for-improving-llm-judge-and-reward-modeling-for-open-ended-tasks - Develop RRD, a framework for refining rubrics to enhance LLM judging accuracy and reward modeling.
- SocialVeil: Probing Social Intelligence of Language Agents under Communication Barriers (viability: 5): https://sciencetostartup.com/paper/socialveil-probing-social-intelligence-of-language-agents-under-communication-barriers - Develop a platform to test and enhance LLMs' social intelligence in real-world-like communication barriers.
- Democratic Preference Alignment via Sortition-Weighted RLHF (viability: 6): https://sciencetostartup.com/paper/democratic-preference-alignment-via-sortition-weighted-rlhf - DemPO optimizes AI alignment by applying representative sortition to human rater preferences for demographic inclusivity.
- Understanding LLM Evaluator Behavior: A Structured Multi-Evaluator Framework for Merchant Risk Assessment (viability: 3): https://sciencetostartup.com/paper/understanding-llm-evaluator-behavior-a-structured-multi-evaluator-framework-for-merchant-risk-assessment - A structured framework for evaluating LLMs in merchant risk assessment, highlighting evaluator bias and alignment with human judgment.
- GAMMS: Graph based Adversarial Multiagent Modeling Simulator (viability: 7): https://sciencetostartup.com/paper/gamms-graph-based-adversarial-multiagent-modeling-simulator - GAMMS is an accessible, scalable graph-based simulator for rapid multi-agent coordination and behavior testing.
- Evaluating Robustness and Adaptability in Learning-Based Mission Planning for Active Debris Removal (viability: 2): https://sciencetostartup.com/paper/evaluating-robustness-and-adaptability-in-learning-based-mission-planning-for-active-debris-removal - Develop adaptive mission planning solutions for autonomous Active Debris Removal in space using learning-based and search-based methods.
- VERA-MH: Reliability and Validity of an Open-Source AI Safety Evaluation in Mental Health (viability: 5): https://sciencetostartup.com/paper/vera-mh-reliability-and-validity-of-an-open-source-ai-safety-evaluation-in-mental-health - Develop an open-source AI safety evaluation tool for mental health chatbots to ensure reliable and valid suicide risk detection and response.
- Optimizing Mission Planning for Multi-Debris Rendezvous Using Reinforcement Learning with Refueling and Adaptive Collision Avoidance (viability: 8): https://sciencetostartup.com/paper/optimizing-mission-planning-for-multi-debris-rendezvous-using-reinforcement-learning-with-refueling-and-adaptive-collisi - Develop a reinforcement learning platform for efficient and safe multi-debris removal using small satellites.
- Towards Reducible Uncertainty Modeling for Reliable Large Language Model Agents (viability: 3): https://sciencetostartup.com/paper/towards-reducible-uncertainty-modeling-for-reliable-large-language-model-agents - Develop a framework for uncertainty quantification in interactive large language model agents.
- E-Globe: Scalable $ε$-Global Verification of Neural Networks via Tight Upper Bounds and Pattern-Aware Branching (viability: 3): https://sciencetostartup.com/paper/e-globe-scalable-global-verification-of-neural-networks-via-tight-upper-bounds-and-pattern-aware-branching - Enhance neural network robustness through efficient verification methods leveraging tight bounds and pattern-aware branching.
- Bypassing AI Control Protocols via Agent-as-a-Proxy Attacks (viability: 5): https://sciencetostartup.com/paper/bypassing-ai-control-protocols-via-agent-as-a-proxy-attacks - Novel cybersecurity method to bypass AI monitoring protocols via agent-proxy attacks.
- Evaluating Large Language Models on Solved and Unsolved Problems in Graph Theory: Implications for Computing Education (viability: 3): https://sciencetostartup.com/paper/evaluating-large-language-models-on-solved-and-unsolved-problems-in-graph-theory-implications-for-computing-education - Explore how LLMs assist in learning and solving graph theory problems in educational settings.
- ReFORM: Reflected Flows for On-support Offline RL via Noise Manipulation (viability: 5): https://sciencetostartup.com/paper/reform-reflected-flows-for-on-support-offline-rl-via-noise-manipulation - ReFORM optimizes flow policies to enhance offline RL by reducing out-of-distribution errors and maximizing policy performance.
- VISTA: Enhancing Visual Conditioning via Track-Following Preference Optimization in Vision-Language-Action Models (viability: 7): https://sciencetostartup.com/paper/vista-enhancing-visual-conditioning-via-track-following-preference-optimization-in-vision-language-action-models - Enhance robotic action precision by optimizing visual conditioning in Vision-Language-Action models.
- MINT: Minimal Information Neuro-Symbolic Tree for Objective-Driven Knowledge-Gap Reasoning and Active Elicitation (viability: 7): https://sciencetostartup.com/paper/mint-minimal-information-neuro-symbolic-tree-for-objective-driven-knowledge-gap-reasoning-and-active-elicitation - Create optimized AI agents that actively elicit human input for object-driven planning using a neuro-symbolic model.
- Quality Model for Machine Learning Components (viability: 6): https://sciencetostartup.com/paper/quality-model-for-machine-learning-components - A quality model for ML components offering structured requirements management integrated into a tool for enhanced component testing.
- Differentiable Inverse Graphics for Zero-shot Scene Reconstruction and Robot Grasping (viability: 8): https://sciencetostartup.com/paper/differentiable-inverse-graphics-for-zero-shot-scene-reconstruction-and-robot-grasping - Develop a robot grasping system using differentiable inverse graphics for zero-shot scene reconstruction.
- Do Vision-Language Models Respect Contextual Integrity in Location Disclosure? (viability: 6): https://sciencetostartup.com/paper/do-vision-language-models-respect-contextual-integrity-in-location-disclosure - A privacy-focused tool using VLMs to manage geolocation disclosure by respecting contextual integrity.
- DeepRead: Document Structure-Aware Reasoning to Enhance Agentic Search (viability: 8): https://sciencetostartup.com/paper/deepread-document-structure-aware-reasoning-to-enhance-agentic-search - DeepRead enhances document question answering by utilizing structure-aware reasoning in LLMs for effective, human-like document search.
- Enhanced QKNorm normalization for neural transformers with the Lp norm (viability: 4): https://sciencetostartup.com/paper/enhanced-qknorm-normalization-for-neural-transformers-with-the-lp-norm - Develop an enhanced normalization method for Transformers using Lp norms for better learning stability.
- CoWork-X: Experience-Optimized Co-Evolution for Multi-Agent Collaboration System (viability: 6): https://sciencetostartup.com/paper/cowork-x-experience-optimized-co-evolution-for-multi-agent-collaboration-system - Create a multi-agent collaboration tool using CoWork-X for optimized real-time performance and low latency.
- EntRGi: Entropy Aware Reward Guidance for Diffusion Language Models (viability: 5): https://sciencetostartup.com/paper/entrgi-entropy-aware-reward-guidance-for-diffusion-language-models - EntRGi enhances reward guidance in diffusion language models by introducing entropy-aware dynamic gradient regulation.
- Learning Rate Matters: Vanilla LoRA May Suffice for LLM Fine-tuning (viability: 5): https://sciencetostartup.com/paper/learning-rate-matters-vanilla-lora-may-suffice-for-llm-fine-tuning - Optimize LLM fine-tuning by tuning learning rates, showing vanilla LoRA can match more complex methods.
- Near-Optimal Dynamic Matching via Coarsening with Application to Heart Transplantation (viability: 5): https://sciencetostartup.com/paper/near-optimal-dynamic-matching-via-coarsening-with-application-to-heart-transplantation - Develops an innovative coarsening-based algorithm for online matching with applications in heart transplantation allocation.
- Artificial Intelligence as Strange Intelligence: Against Linear Models of Intelligence (viability: 1): https://sciencetostartup.com/paper/artificial-intelligence-as-strange-intelligence-against-linear-models-of-intelligence - Exploring 'strange intelligence': AI's nonlinear capabilities and adversarial evaluation insights.
- Stochastic hierarchical data-driven optimization: application to plasma-surface kinetics (viability: 4): https://sciencetostartup.com/paper/stochastic-hierarchical-data-driven-optimization-application-to-plasma-surface-kinetics - A new optimization framework for calibrating complex models efficiently, applicable to plasma-surface kinetics.
- Protein Autoregressive Modeling via Multiscale Structure Generation (viability: 6): https://sciencetostartup.com/paper/protein-autoregressive-modeling-via-multiscale-structure-generation - Create high-quality protein structures using a novel autoregressive model for precision applications in drug design and biology.
- Contrastive Continual Learning for Model Adaptability in Internet of Things (viability: 3): https://sciencetostartup.com/paper/contrastive-continual-learning-for-model-adaptability-in-internet-of-things - Develop a framework for contrastive continual learning to enhance IoT model adaptability in dynamic environments.
- Rethinking the Trust Region in LLM Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/rethinking-the-trust-region-in-llm-reinforcement-learning - Develop Divergence Proximal Policy Optimization for more efficient and stable reinforcement learning-based fine-tuning of Large Language Models.
- CRoSS: A Continual Robotic Simulation Suite for Scalable Reinforcement Learning with High Task Diversity and Realistic Physics Simulation (viability: 5): https://sciencetostartup.com/paper/cross-a-continual-robotic-simulation-suite-for-scalable-reinforcement-learning-with-high-task-diversity-and-realistic-ph - A scalable and extensible benchmark suite for continual reinforcement learning in realistic robotic simulations.
- Subliminal Effects in Your Data: A General Mechanism via Log-Linearity (viability: 6): https://sciencetostartup.com/paper/subliminal-effects-in-your-data-a-general-mechanism-via-log-linearity - Utilize Logit-Linear-Selection to reveal hidden dataset effects in language model training for customizable model behaviors.
- From Evaluation to Design: Using Potential Energy Surface Smoothness Metrics to Guide Machine Learning Interatomic Potential Architectures (viability: 6): https://sciencetostartup.com/paper/from-evaluation-to-design-using-potential-energy-surface-smoothness-metrics-to-guide-machine-learning-interatomic-potent - Develop a robust MLIP evaluation tool to reduce errors in interatomic potential modeling while optimizing model design efficiency.
- Fluid Representations in Reasoning Models (viability: 3): https://sciencetostartup.com/paper/fluid-representations-in-reasoning-models - Mechanistic analysis of reasoning models for improved abstract problem-solving efficiency.
- Group-Evolving Agents: Open-Ended Self-Improvement via Experience Sharing (viability: 6): https://sciencetostartup.com/paper/group-evolving-agents-open-ended-self-improvement-via-experience-sharing - Develop advanced self-improving agents that share experiences to outperform existing AI frameworks in coding tasks.
- Are AI Capabilities Increasing Exponentially? A Competing Hypothesis (viability: 2): https://sciencetostartup.com/paper/are-ai-capabilities-increasing-exponentially-a-competing-hypothesis - The paper challenges the notion of exponential growth in AI capabilities, proposing instead that growth will reach an inflection point soon.
- It's not a Lottery, it's a Race: Understanding How Gradient Descent Adapts the Network's Capacity to the Task (viability: 2): https://sciencetostartup.com/paper/it-s-not-a-lottery-it-s-a-race-understanding-how-gradient-descent-adapts-the-network-s-capacity-to-the-task - Analyze gradient descent dynamics to understand neural network capacity adaptation.
- Safe Urban Traffic Control via Uncertainty-Aware Conformal Prediction and World-Model Reinforcement Learning (viability: 6): https://sciencetostartup.com/paper/safe-urban-traffic-control-via-uncertainty-aware-conformal-prediction-and-world-model-reinforcement-learning - STREAM-RL offers a unified framework for safe urban traffic management with uncertainty-aware prediction and control.
- Toward Reliable and Explainable Nail Disease Classification: Leveraging Adversarial Training and Grad-CAM Visualization (viability: 6): https://sciencetostartup.com/paper/toward-reliable-and-explainable-nail-disease-classification-leveraging-adversarial-training-and-grad-cam-visualization - A nail disease classification tool using CNN models and explainable AI to aid doctors in diagnosis.
- Agentic AI in Healthcare & Medicine: A Seven-Dimensional Taxonomy for Empirical Evaluation of LLM-based Agents (viability: 3): https://sciencetostartup.com/paper/agentic-ai-in-healthcare-medicine-a-seven-dimensional-taxonomy-for-empirical-evaluation-of-llm-based-agents - Develop a healthcare tool based on a seven-dimensional taxonomy to enhance evaluation of LLM-based agents.
- SE-Bench: Benchmarking Self-Evolution with Knowledge Internalization (viability: 5): https://sciencetostartup.com/paper/se-bench-benchmarking-self-evolution-with-knowledge-internalization - SE-Bench provides a diagnostic platform for testing and improving AI agents' ability for lifelong learning through effective knowledge internalization.
- Beyond Rewards in Reinforcement Learning for Cyber Defence (viability: 2): https://sciencetostartup.com/paper/beyond-rewards-in-reinforcement-learning-for-cyber-defence - Enhancing cyber defense with reinforcement learning by evaluating reward structures for more effective training.
- Skin Tokens: A Learned Compact Representation for Unified Autoregressive Rigging (viability: 4): https://sciencetostartup.com/paper/skin-tokens-a-learned-compact-representation-for-unified-autoregressive-rigging - Unify and enhance 3D animation rigging with a token-based representation improving efficiency and accuracy.
- Team, Then Trim: An Assembly-Line LLM Framework for High-Quality Tabular Data Generation (viability: 8): https://sciencetostartup.com/paper/team-then-trim-an-assembly-line-llm-framework-for-high-quality-tabular-data-generation - A framework using LLMs for efficient, high-quality tabular data generation through an automated quality control pipeline.
- Billion-Scale Graph Foundation Models (viability: 3): https://sciencetostartup.com/paper/billion-scale-graph-foundation-models - Introducing GraphBFF, a scalable architecture designed for billion-parameter Graph Foundation Models tailored for heterogeneous graphs.
- Active Asymmetric Multi-Agent Multimodal Learning under Uncertainty (viability: 7): https://sciencetostartup.com/paper/active-asymmetric-multi-agent-multimodal-learning-under-uncertainty - A2MAML enhances multi-agent systems with uncertainty-aware multimodal learning for improved accident detection.
- When Silence Is Golden: Can LLMs Learn to Abstain in Temporal QA and Beyond? (viability: 2): https://sciencetostartup.com/paper/when-silence-is-golden-can-llms-learn-to-abstain-in-temporal-qa-and-beyond - Developing LLMs to reliably abstain from uncertain answers can enhance temporal QA accuracy.
- Audio ControlNet for Fine-Grained Audio Generation and Editing (viability: 8): https://sciencetostartup.com/paper/audio-controlnet-for-fine-grained-audio-generation-and-editing - Audio ControlNet enhances text-to-audio models with precise control over audio attributes and editing capabilities.
- Let Experts Feel Uncertainty: A Multi-Expert Label Distribution Approach to Probabilistic Time Series Forecasting (viability: 5): https://sciencetostartup.com/paper/let-experts-feel-uncertainty-a-multi-expert-label-distribution-approach-to-probabilistic-time-series-forecasting - Introducing a multi-expert framework for more interpretable and accurate time series forecasting with uncertainty quantification.
- WideSeek-R1: Exploring Width Scaling for Broad Information Seeking via Multi-Agent Reinforcement Learning (viability: 4): https://sciencetostartup.com/paper/wideseek-r1-exploring-width-scaling-for-broad-information-seeking-via-multi-agent-reinforcement-learning - WideSeek-R1 enhances multi-agent reinforcement learning for broad information-seeking tasks with width scaling capabilities.
- Dual Mind World Model Inspired Network Digital Twin for Access Scheduling (viability: 6): https://sciencetostartup.com/paper/dual-mind-world-model-inspired-network-digital-twin-for-access-scheduling - A Digital Twin-enabled scheduling framework for adaptive network optimization in IoT and cyber-physical systems.
- OmniRad: A Radiological Foundation Model for Multi-Task Medical Image Analysis (viability: 7): https://sciencetostartup.com/paper/omnirad-a-radiological-foundation-model-for-multi-task-medical-image-analysis - OmniRad offers a pre-trained radiological model that enhances multi-task medical imaging analysis across various modalities.
- Continual Learning through Control Minimization (viability: 5): https://sciencetostartup.com/paper/continual-learning-through-control-minimization - A novel continual learning approach that minimizes control efforts for task integration.
- LycheeDecode: Accelerating Long-Context LLM Inference via Hybrid-Head Sparse Decoding (viability: 7): https://sciencetostartup.com/paper/lycheedecode-accelerating-long-context-llm-inference-via-hybrid-head-sparse-decoding - LycheeDecode accelerates long-context LLM inference with a novel hybrid-head sparse decoding to significantly reduce memory and latency costs while maintaining high generative quality.
- ReThinker: Scientific Reasoning by Rethinking with Guided Reflection and Confidence Control (viability: 7): https://sciencetostartup.com/paper/rethinker-scientific-reasoning-by-rethinking-with-guided-reflection-and-confidence-control - ReThinker enhances scientific reasoning in AI by dynamically orchestrating tool use and agent reasoning with confidence control to outperform existing models.
- RASA: Routing-Aware Safety Alignment for Mixture-of-Experts Models (viability: 6): https://sciencetostartup.com/paper/rasa-routing-aware-safety-alignment-for-mixture-of-experts-models - RASA offers targeted and efficient safety alignment for Mixture-of-Experts models by repairing Safety-Critical Experts without global updates.
- Med-MMFL: A Multimodal Federated Learning Benchmark in Healthcare (viability: 5): https://sciencetostartup.com/paper/med-mmfl-a-multimodal-federated-learning-benchmark-in-healthcare - Med-MMFL provides the first comprehensive multimodal federated learning benchmark to enhance medical AI research while maintaining data privacy.
- Bi-directional Bias Attribution: Debiasing Large Language Models without Modifying Prompts (viability: 6): https://sciencetostartup.com/paper/bi-directional-bias-attribution-debiasing-large-language-models-without-modifying-prompts - Develop a tool to debias large language models by attributing and intervening neuron-level bias without altering prompts or performing fine-tuning.
- LoRDO: Distributed Low-Rank Optimization with Infrequent Communication (viability: 5): https://sciencetostartup.com/paper/lordo-distributed-low-rank-optimization-with-infrequent-communication - LoRDO optimizes distributed model training to reduce communication overhead without sacrificing performance, ideal for low-memory settings.
- From Assumptions to Actions: Turning LLM Reasoning into Uncertainty-Aware Planning for Embodied Agents (viability: 7): https://sciencetostartup.com/paper/from-assumptions-to-actions-turning-llm-reasoning-into-uncertainty-aware-planning-for-embodied-agents - Develop an uncertainty-aware planning system for embodied agents using LLMs to reduce communication costs while maintaining efficiency and trustworthiness.
- DeFrame: Debiasing Large Language Models Against Framing Effects (viability: 6): https://sciencetostartup.com/paper/deframe-debiasing-large-language-models-against-framing-effects - Develop a tool to reduce framing bias in large language models for fairer real-world applications.
- Beyond Static Cropping: Layer-Adaptive Visual Localization and Decoding Enhancement (viability: 5): https://sciencetostartup.com/paper/beyond-static-cropping-layer-adaptive-visual-localization-and-decoding-enhancement - Introducing LASER: dynamic layer adaptation for improved visual grounding in large vision-language models.
- Revisiting Prompt Sensitivity in Large Language Models for Text Classification: The Role of Prompt Underspecification (viability: 2): https://sciencetostartup.com/paper/revisiting-prompt-sensitivity-in-large-language-models-for-text-classification-the-role-of-prompt-underspecification - Exploring prompt sensitivity in LLMs to improve text classification accuracy by addressing prompt underspecification.
- ProxyWar: Dynamic Assessment of LLM Code Generation in Game Arenas (viability: 6): https://sciencetostartup.com/paper/proxywar-dynamic-assessment-of-llm-code-generation-in-game-arenas - ProxyWar offers a dynamic framework for evaluating LLM code generation by embedding agents in competitive game environments.
- SkeletonGaussian: Editable 4D Generation through Gaussian Skeletonization (viability: 7): https://sciencetostartup.com/paper/skeletongaussian-editable-4d-generation-through-gaussian-skeletonization - SkeletonGaussian enables intuitive and editable 4D motion generation from monocular videos using a novel hierarchical skeleton-based framework.
- From Dead Neurons to Deep Approximators: Deep Bernstein Networks as a Provable Alternative to Residual Layers (viability: 4): https://sciencetostartup.com/paper/from-dead-neurons-to-deep-approximators-deep-bernstein-networks-as-a-provable-alternative-to-residual-layers - Deep Bernstein Networks optimize trainability and representation power without residual connections, offering enhanced expressive capacity over traditional architectures.
- Empirical-MCTS: Continuous Agent Evolution via Dual-Experience Monte Carlo Tree Search (viability: 7): https://sciencetostartup.com/paper/empirical-mcts-continuous-agent-evolution-via-dual-experience-monte-carlo-tree-search - Develop a tool to enhance LLM reasoning capabilities using Empirical-MCTS for continuous learning.
- RAPO: Risk-Aware Preference Optimization for Generalizable Safe Reasoning (viability: 5): https://sciencetostartup.com/paper/rapo-risk-aware-preference-optimization-for-generalizable-safe-reasoning - Develop RAPO to enhance LRM safety by adaptively addressing jailbreak attacks with a robust alignment technique.
- InterPReT: Interactive Policy Restructuring and Training Enable Effective Imitation Learning from Laypersons (viability: 6): https://sciencetostartup.com/paper/interpret-interactive-policy-restructuring-and-training-enable-effective-imitation-learning-from-laypersons - An interactive tool enabling laypersons to train AI agents through imitation learning and policy restructuring.
- Natural Language Instructions for Scene-Responsive Human-in-the-Loop Motion Planning in Autonomous Driving using Vision-Language-Action Models (viability: 6): https://sciencetostartup.com/paper/natural-language-instructions-for-scene-responsive-human-in-the-loop-motion-planning-in-autonomous-driving-using-vision- - Leverage natural language instructions to improve motion planning in autonomous vehicles using a reproducible framework.
- Pruning for Generalization: A Transfer-Oriented Spatiotemporal Graph Framework (viability: 7): https://sciencetostartup.com/paper/pruning-for-generalization-a-transfer-oriented-spatiotemporal-graph-framework - A cutting-edge framework enhancing multivariate time series forecasting on graphs by pruning for efficient and robust out-of-domain generalization.
- MA3DSG: Multi-Agent 3D Scene Graph Generation for Large-Scale Indoor Environments (viability: 5): https://sciencetostartup.com/paper/ma3dsg-multi-agent-3d-scene-graph-generation-for-large-scale-indoor-environments - MA3DSG enables scalable multi-agent collaboration in 3D scene graph generation for large-scale environments.
- KGLAMP: Knowledge Graph-guided Language model for Adaptive Multi-robot Planning and Replanning (viability: 7): https://sciencetostartup.com/paper/kglamp-knowledge-graph-guided-language-model-for-adaptive-multi-robot-planning-and-replanning - KGLAMP: improving multi-robot system planning with knowledge graphs and LLMs for dynamic environments.
- From Lemmas to Dependencies: What Signals Drive Light Verbs Classification? (viability: 2): https://sciencetostartup.com/paper/from-lemmas-to-dependencies-what-signals-drive-light-verbs-classification - Explore signals for classifying light verb constructions in Turkish using lemma-based and grammar models.
- Principles of Lipschitz continuity in neural networks (viability: 2): https://sciencetostartup.com/paper/principles-of-lipschitz-continuity-in-neural-networks - Explore the theoretical aspects of Lipschitz continuity to enhance neural network robustness and generalization.
- Understanding and Guiding Layer Placement in Parameter-Efficient Fine-Tuning of Large Language Models (viability: 3): https://sciencetostartup.com/paper/understanding-and-guiding-layer-placement-in-parameter-efficient-fine-tuning-of-large-language-models - Develop Layer Card diagnostics to optimize layer selection in parameter-efficient fine-tuning for LLMs.
- PromptSplit: Revealing Prompt-Level Disagreement in Generative Models (viability: 4): https://sciencetostartup.com/paper/promptsplit-revealing-prompt-level-disagreement-in-generative-models - PromptSplit provides a framework to pinpoint prompt-level behavioral disagreements in generative AI models.
- Rational ANOVA Networks (viability: 5): https://sciencetostartup.com/paper/rational-anova-networks - Rational-ANOVA Networks improve interpretability and efficiency over traditional neural nets for better performance on compute-budgeted tasks.
- When Chains of Thought Don't Matter: Causal Bypass in Large Language Models (viability: 2): https://sciencetostartup.com/paper/when-chains-of-thought-don-t-matter-causal-bypass-in-large-language-models - Diagnostic framework for auditing the reliance of LLM answers on chain-of-thought prompts.
- Structural shifts in institutional participation and collaboration within the AI arXiv preprint research ecosystem (viability: 3): https://sciencetostartup.com/paper/structural-shifts-in-institutional-participation-and-collaboration-within-the-ai-arxiv-preprint-research-ecosystem - Dataset analysis of AI research trends reveals suppressed academic-industry collaboration.
- AgentArk: Distilling Multi-Agent Intelligence into a Single LLM Agent (viability: 6): https://sciencetostartup.com/paper/agentark-distilling-multi-agent-intelligence-into-a-single-llm-agent - AgentArk distills multi-agent intelligence into a single efficient LLM, enhancing reasoning capabilities without high computational costs.
- PLATE: Plasticity-Tunable Efficient Adapters for Geometry-Aware Continual Learning (viability: 5): https://sciencetostartup.com/paper/plate-plasticity-tunable-efficient-adapters-for-geometry-aware-continual-learning - Build efficient adapters for geometry-aware continual learning in pretrained models without using old-task data.
- PrevizWhiz: Combining Rough 3D Scenes and 2D Video to Guide Generative Video Previsualization (viability: 7): https://sciencetostartup.com/paper/previzwhiz-combining-rough-3d-scenes-and-2d-video-to-guide-generative-video-previsualization - PrevizWhiz simplifies and speeds up film previsualization by combining rough 3D scenes with generative video technology.
- Accelerating Scientific Research with Gemini: Case Studies and Common Techniques (viability: 5): https://sciencetostartup.com/paper/accelerating-scientific-research-with-gemini-case-studies-and-common-techniques - Gemini assists researchers in theoretical computer science using advanced AI collaboration techniques.
- AutoFigure: Generating and Refining Publication-Ready Scientific Illustrations (viability: 8): https://sciencetostartup.com/paper/autofigure-generating-and-refining-publication-ready-scientific-illustrations - AutoFigure automates the generation of publication-ready scientific illustrations from long-form texts, streamlining science communication.
- Adaptive Evidence Weighting for Audio-Spatiotemporal Fusion (viability: 6): https://sciencetostartup.com/paper/adaptive-evidence-weighting-for-audio-spatiotemporal-fusion - Adaptive framework for robust bioacoustic species classification using audio-spatiotemporal fusion.
- Conformal Thinking: Risk Control for Reasoning on a Compute Budget (viability: 5): https://sciencetostartup.com/paper/conformal-thinking-risk-control-for-reasoning-on-a-compute-budget - Develop a risk-controlled reasoning framework to optimize compute efficiency for Large Language Models.
- Antidistillation Fingerprinting (viability: 6): https://sciencetostartup.com/paper/antidistillation-fingerprinting - ADFP offers a robust solution for identifying student models trained on LLM outputs without compromising utility, improving state-of-the-art detection methods.
- Enhancing Imbalanced Node Classification via Curriculum-Guided Feature Learning and Three-Stage Attention Network (viability: 5): https://sciencetostartup.com/paper/enhancing-imbalanced-node-classification-via-curriculum-guided-feature-learning-and-three-stage-attention-network - A neural network framework that improves imbalanced graph classification using a novel curriculum learning approach for fairness and performance.
- Bridging Online and Offline RL: Contextual Bandit Learning for Multi-Turn Code Generation (viability: 8): https://sciencetostartup.com/paper/bridging-online-and-offline-rl-contextual-bandit-learning-for-multi-turn-code-generation - Cobalt enhances code generation in LLMs using a cost-effective hybrid of online and offline RL.
- Do We Need Asynchronous SGD? On the Near-Optimality of Synchronous Methods (viability: 3): https://sciencetostartup.com/paper/do-we-need-asynchronous-sgd-on-the-near-optimality-of-synchronous-methods - Exploring the theoretical near-optimality of synchronous optimization methods in distributed systems.
- Conformal Reachability for Safe Control in Unknown Environments (viability: 3): https://sciencetostartup.com/paper/conformal-reachability-for-safe-control-in-unknown-environments - Probabilistic safety verification for control systems using conformal prediction and reachability analysis.
- Understanding Agent Scaling in LLM-Based Multi-Agent Systems via Diversity (viability: 8): https://sciencetostartup.com/paper/understanding-agent-scaling-in-llm-based-multi-agent-systems-via-diversity - Develop an efficient multi-agent system using diverse LLMs for improved task performance over homogeneous scaling.
- WebSentinel: Detecting and Localizing Prompt Injection Attacks for Web Agents (viability: 7): https://sciencetostartup.com/paper/websentinel-detecting-and-localizing-prompt-injection-attacks-for-web-agents - WebSentinel is a tool to detect and localize prompt injection attacks for web security, outperforming existing methods with open-source code and datasets.
- AOrchestra: Automating Sub-Agent Creation for Agentic Orchestration (viability: 7): https://sciencetostartup.com/paper/aorchestra-automating-sub-agent-creation-for-agentic-orchestration - AOrchestra automates sub-agent creation for efficient multi-agent orchestration, improving adaptability and performance in task automation.
- Efficient Estimation of Kernel Surrogate Models for Task Attribution (viability: 4): https://sciencetostartup.com/paper/efficient-estimation-of-kernel-surrogate-models-for-task-attribution - Develop a kernel surrogate model for efficient task attribution in AI agents, improving task selection and performance evaluation with minimal retraining.
- Reward Redistribution for CVaR MDPs using a Bellman Operator on L-infinity (viability: 5): https://sciencetostartup.com/paper/reward-redistribution-for-cvar-mdps-using-a-bellman-operator-on-l-infinity - Develop a Q-learning algorithm tailored for risk-averse decision-making using CVaR-sensitive policies.
- An Empirical Study of Collective Behaviors and Social Dynamics in Large Language Model Agents (viability: 1): https://sciencetostartup.com/paper/an-empirical-study-of-collective-behaviors-and-social-dynamics-in-large-language-model-agents - Explore social dynamics in AI-driven platforms with Chain of Social Thought to minimize harmful behaviors.
- UniGeM: Unifying Data Mixing and Selection via Geometric Exploration and Mining (viability: 5): https://sciencetostartup.com/paper/unigem-unifying-data-mixing-and-selection-via-geometric-exploration-and-mining - UniGeM improves LLM data efficiency and quality through a unified data mixing and selection framework.
- Decision-oriented benchmarking to transform AI weather forecast access: Application to the Indian monsoon (viability: 6): https://sciencetostartup.com/paper/decision-oriented-benchmarking-to-transform-ai-weather-forecast-access-application-to-the-indian-monsoon - Transformative AI-powered weather forecasting for enhancing agricultural decision-making in susceptible regions.
- Zero-shot large vision-language model prompting for automated bone identification in paleoradiology x-ray archives (viability: 8): https://sciencetostartup.com/paper/zero-shot-large-vision-language-model-prompting-for-automated-bone-identification-in-paleoradiology-x-ray-archives - Automated tool for efficient bone identification in paleoradiology using large vision-language models.
- Cognitively Diverse Multiple-Choice Question Generation: A Hybrid Multi-Agent Framework with Large Language Models (viability: 6): https://sciencetostartup.com/paper/cognitively-diverse-multiple-choice-question-generation-a-hybrid-multi-agent-framework-with-large-language-models - Develop an AI-powered tool for generating high-quality, cognitively diverse multiple-choice questions.
- Anytime Pretraining: Horizon-Free Learning-Rate Schedules with Weight Averaging (viability: 4): https://sciencetostartup.com/paper/anytime-pretraining-horizon-free-learning-rate-schedules-with-weight-averaging - Anytime Pretraining introduces weight averaging with horizon-free learning schedules for improved large language model training efficiency.
- Agent Primitives: Reusable Latent Building Blocks for Multi-Agent Systems (viability: 5): https://sciencetostartup.com/paper/agent-primitives-reusable-latent-building-blocks-for-multi-agent-systems - Develop reusable building blocks to simplify and enhance robustness of multi-agent systems.
- LLM-Inspired Pretrain-Then-Finetune for Small-Data, Large-Scale Optimization (viability: 3): https://sciencetostartup.com/paper/llm-inspired-pretrain-then-finetune-for-small-data-large-scale-optimization - Develop a Transformer-based solution for optimizing decision-making across large-scale portfolios using limited data through a pretrain-then-finetune pipeline.
- Search-R2: Enhancing Search-Integrated Reasoning via Actor-Refiner Collaboration (viability: 7): https://sciencetostartup.com/paper/search-r2-enhancing-search-integrated-reasoning-via-actor-refiner-collaboration - Develop an AI tool for enhanced search-integrated reasoning with superior accuracy using a novel Actor-Refiner framework.
- BIRDTurk: Adaptation of the BIRD Text-to-SQL Dataset to Turkish (viability: 5): https://sciencetostartup.com/paper/birdturk-adaptation-of-the-bird-text-to-sql-dataset-to-turkish - BIRDTurk adapts the BIRD Text-to-SQL dataset for Turkish, enabling cross-lingual evaluation in text-to-SQL systems.
- Controlling Output Rankings in Generative Engines for LLM-based Search (viability: 7): https://sciencetostartup.com/paper/controlling-output-rankings-in-generative-engines-for-llm-based-search - Optimize product visibility by controlling LLM-based search output rankings using CORE.
- APEX: Probing Neural Networks via Activation Perturbation (viability: 5): https://sciencetostartup.com/paper/apex-probing-neural-networks-via-activation-perturbation - APEX provides a novel method for probing neural networks by perturbing activations to explore structural insights.
- Don't believe everything you read: Understanding and Measuring MCP Behavior under Misleading Tool Descriptions (viability: 6): https://sciencetostartup.com/paper/don-t-believe-everything-you-read-understanding-and-measuring-mcp-behavior-under-misleading-tool-descriptions - Develop a security auditing tool for MCP-based AI agents detecting description-code inconsistencies.
- Group Selection as a Safeguard Against AI Substitution (viability: 3): https://sciencetostartup.com/paper/group-selection-as-a-safeguard-against-ai-substitution - Explore the long-term cultural impacts of AI reliance on diversity and innovation.
- Morphe: High-Fidelity Generative Video Streaming with Vision Foundation Model (viability: 7): https://sciencetostartup.com/paper/morphe-high-fidelity-generative-video-streaming-with-vision-foundation-model - Morphe leverages vision foundation models to revolutionize video streaming with high fidelity and 62.5% bandwidth savings for real-time applications.
- D3PIA: A Discrete Denoising Diffusion Model for Piano Accompaniment Generation From Lead sheet (viability: 7): https://sciencetostartup.com/paper/d3pia-a-discrete-denoising-diffusion-model-for-piano-accompaniment-generation-from-lead-sheet - An advanced model for generating musically coherent piano accompaniments from lead sheets using discrete diffusion techniques.
- Live or Lie: Action-Aware Capsule Multiple Instance Learning for Risk Assessment in Live Streaming Platforms (viability: 8): https://sciencetostartup.com/paper/live-or-lie-action-aware-capsule-multiple-instance-learning-for-risk-assessment-in-live-streaming-platforms - A platform using advanced AI techniques to assess and mitigate risks in live streaming environments.
- CMR: Contractive Mapping Embeddings for Robust Humanoid Locomotion on Unstructured Terrains (viability: 3): https://sciencetostartup.com/paper/cmr-contractive-mapping-embeddings-for-robust-humanoid-locomotion-on-unstructured-terrains - Enhancing humanoid robot locomotion under noisy conditions using a contractive mapping framework.
- Explaining the Explainer: Understanding the Inner Workings of Transformer-based Symbolic Regression Models (viability: 4): https://sciencetostartup.com/paper/explaining-the-explainer-understanding-the-inner-workings-of-transformer-based-symbolic-regression-models - Develop an evolutionary circuit discovery algorithm to enhance interpretable transformer-based symbolic regression models.
- Self-Verification Dilemma: Experience-Driven Suppression of Overused Checking in LLM Reasoning (viability: 5): https://sciencetostartup.com/paper/self-verification-dilemma-experience-driven-suppression-of-overused-checking-in-llm-reasoning - Optimize large reasoning models by reducing unnecessary self-verification to save computation while maintaining accuracy.
- When Routing Collapses: On the Degenerate Convergence of LLM Routers (viability: 8): https://sciencetostartup.com/paper/when-routing-collapses-on-the-degenerate-convergence-of-llm-routers - EquiRouter optimizes AI model routing to reduce computation costs while maintaining high performance.
- IntentRL: Training Proactive User-intent Agents for Open-ended Deep Research via Reinforcement Learning (viability: 3): https://sciencetostartup.com/paper/intentrl-training-proactive-user-intent-agents-for-open-ended-deep-research-via-reinforcement-learning - IntentRL enhances user-intent clarification for Deep Research agents to improve task performance through a two-stage reinforcement learning framework.
- The Dual Role of Abstracting over the Irrelevant in Symbolic Explanations: Cognitive Effort vs. Understanding (viability: 2): https://sciencetostartup.com/paper/the-dual-role-of-abstracting-over-the-irrelevant-in-symbolic-explanations-cognitive-effort-vs-understanding - A method using answer set programming to improve human understanding and reduce cognitive effort through explanation abstraction.
- Hierarchical Concept-to-Appearance Guidance for Multi-Subject Image Generation (viability: 8): https://sciencetostartup.com/paper/hierarchical-concept-to-appearance-guidance-for-multi-subject-image-generation - A framework for generating consistent multi-subject images from textual prompts, using hierarchical concept-to-appearance guidance.
- DiscoverLLM: From Executing Intents to Discovering Them (viability: 7): https://sciencetostartup.com/paper/discoverllm-from-executing-intents-to-discovering-them - DiscoverLLM helps users form and discover their intents through adaptive, interactive language model collaboration.
- Feasible strategies for conflict resolution within intuitionistic fuzzy preference-based conflict situations (viability: 4): https://sciencetostartup.com/paper/feasible-strategies-for-conflict-resolution-within-intuitionistic-fuzzy-preference-based-conflict-situations - Develops intuitionistic fuzzy models to enhance conflict resolution by granularly understanding preference-based agent conflict.
- Risk Awareness Injection: Calibrating Vision-Language Models for Safety without Compromising Utility (viability: 6): https://sciencetostartup.com/paper/risk-awareness-injection-calibrating-vision-language-models-for-safety-without-compromising-utility - Introducing a lightweight framework to enhance safety in vision-language models without sacrificing utility.
- Chain-of-Goals Hierarchical Policy for Long-Horizon Offline Goal-Conditioned RL (viability: 7): https://sciencetostartup.com/paper/chain-of-goals-hierarchical-policy-for-long-horizon-offline-goal-conditioned-rl - Revolutionizing long-horizon offline goal-conditioned RL with sequential subgoal autoregression.
- Toward a Sustainable Federated Learning Ecosystem: A Practical Least Core Mechanism for Payoff Allocation (viability: 5): https://sciencetostartup.com/paper/toward-a-sustainable-federated-learning-ecosystem-a-practical-least-core-mechanism-for-payoff-allocation - Develop a fair and efficient payoff allocation framework for sustainable federated learning environments.
- Causal Graph Learning via Distributional Invariance of Cause-Effect Relationship (viability: 5): https://sciencetostartup.com/paper/causal-graph-learning-via-distributional-invariance-of-cause-effect-relationship - Discovering causal relationships in data with a novel algorithm that outspeeds current methods by leveraging invariance in effect-cause relations.
- Building Interpretable Models for Moral Decision-Making (viability: 5): https://sciencetostartup.com/paper/building-interpretable-models-for-moral-decision-making - Develop an interpretable model to analyze moral decision-making in AI systems.
- Robustness as an Emergent Property of Task Performance (viability: 5): https://sciencetostartup.com/paper/robustness-as-an-emergent-property-of-task-performance - A method leveraging task performance to naturally enhance model robustness for dependable real-world applications.
- Tiled Prompts: Overcoming Prompt Underspecification in Image and Video Super-Resolution (viability: 3): https://sciencetostartup.com/paper/tiled-prompts-overcoming-prompt-underspecification-in-image-and-video-super-resolution - Develop a tiled prompts framework for localized text-conditioned super-resolution in images and videos, addressing prompt underspecification issues.
- MedSAM-Agent: Empowering Interactive Medical Image Segmentation with Multi-turn Agentic Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/medsam-agent-empowering-interactive-medical-image-segmentation-with-multi-turn-agentic-reinforcement-learning - An interactive medical segmentation tool using multi-turn reinforcement learning for efficient and accurate medical imaging analytics.
- Entropy-Gated Selective Policy Optimization:Token-Level Gradient Allocation for Hybrid Training of Large Language Models (viability: 5): https://sciencetostartup.com/paper/entropy-gated-selective-policy-optimization-token-level-gradient-allocation-for-hybrid-training-of-large-language-models - Develop an entropy-gated framework for optimizing hybrid training in language models to improve mathematical reasoning performance.
- POP: Prefill-Only Pruning for Efficient Large Model Inference (viability: 8): https://sciencetostartup.com/paper/pop-prefill-only-pruning-for-efficient-large-model-inference - POP offers a novel pruning method to make large language and vision-language models faster and cheaper to deploy without sacrificing accuracy.
- Rejecting Arguments Based on Doubt in Structured Bipolar Argumentation (viability: 3): https://sciencetostartup.com/paper/rejecting-arguments-based-on-doubt-in-structured-bipolar-argumentation - Developing a novel approach to computational argumentation integrating doubt and sentence acceptance over arguments.
- MeetBench-XL: Calibrated Multi-Dimensional Evaluation and Learned Dual-Policy Agents for Real-Time Meetings (viability: 7): https://sciencetostartup.com/paper/meetbench-xl-calibrated-multi-dimensional-evaluation-and-learned-dual-policy-agents-for-real-time-meetings - Develop AI assistants for real-time enterprise meetings using MeetBench-XL, offering enhanced task management and strategic insights.
- Agentic Proposing: Enhancing Large Language Model Reasoning via Compositional Skill Synthesis (viability: 7): https://sciencetostartup.com/paper/agentic-proposing-enhancing-large-language-model-reasoning-via-compositional-skill-synthesis - Agentic Proposing offers a scalable framework for synthesizing training data that enhances AI solvers to achieve SOTA results with far less data.
- Unveiling Covert Toxicity in Multimodal Data via Toxicity Association Graphs: A Graph-Based Metric and Interpretable Detection Framework (viability: 8): https://sciencetostartup.com/paper/unveiling-covert-toxicity-in-multimodal-data-via-toxicity-association-graphs-a-graph-based-metric-and-interpretable-dete - A novel framework for detecting covert multimodal toxicity using graph-based metrics.
- CSR-Bench: A Benchmark for Evaluating the Cross-modal Safety and Reliability of MLLMs (viability: 4): https://sciencetostartup.com/paper/csr-bench-a-benchmark-for-evaluating-the-cross-modal-safety-and-reliability-of-mllms - Develop a benchmark tool, CSR-Bench, to evaluate the safety and reliability of cross-modal large language models.
- GraDE: A Graph Diffusion Estimator for Frequent Subgraph Discovery in Neural Architectures (viability: 5): https://sciencetostartup.com/paper/grade-a-graph-diffusion-estimator-for-frequent-subgraph-discovery-in-neural-architectures - GraDE provides a new method for efficiently discovering frequent subgraph patterns in neural architectures using graph diffusion models.
- The Necessity of a Unified Framework for LLM-Based Agent Evaluation (viability: 4): https://sciencetostartup.com/paper/the-necessity-of-a-unified-framework-for-llm-based-agent-evaluation - Develop a unified framework for standardized evaluation of LLM-based agents to ensure fair and reproducible benchmarks.
- TAME: A Trustworthy Test-Time Evolution of Agent Memory with Systematic Benchmarking (viability: 4): https://sciencetostartup.com/paper/tame-a-trustworthy-test-time-evolution-of-agent-memory-with-systematic-benchmarking - TAME enhances agent memory trustworthiness by evolving separate executor and evaluator memories, preserving utility and safety.
- Topology Matters: A Cautionary Case Study of Graph SSL on Neuro-Inspired Benchmarks (viability: 3): https://sciencetostartup.com/paper/topology-matters-a-cautionary-case-study-of-graph-ssl-on-neuro-inspired-benchmarks - Develop topology-aware self-supervised learning objectives that preserve neuro-inspired structures.
- Privasis: Synthesizing the Largest "Public" Private Dataset from Scratch (viability: 7): https://sciencetostartup.com/paper/privasis-synthesizing-the-largest-public-private-dataset-from-scratch - Develop a synthetic data generation platform to accelerate research in privacy-sensitive domains.
- Enhancing Foundation VLM Robustness to Missing Modality: Scalable Diffusion for Bi-directional Feature Restoration (viability: 6): https://sciencetostartup.com/paper/enhancing-foundation-vlm-robustness-to-missing-modality-scalable-diffusion-for-bi-directional-feature-restoration - A robust and scalable strategy for restoring missing modality in vision language models using enhanced diffusion models.
- General Agents Contain World Models, even under Partial Observability and Stochasticity (viability: 2): https://sciencetostartup.com/paper/general-agents-contain-world-models-even-under-partial-observability-and-stochasticity - Understanding agent capabilities through their world models in stochastic and partially observable environments.
- Self-Hinting Language Models Enhance Reinforcement Learning (viability: 6): https://sciencetostartup.com/paper/self-hinting-language-models-enhance-reinforcement-learning - Enhance LLM-based reinforcement learning with self-hint alignment for improved task performance under sparse rewards.
- Contrastive Concept-Tree Search for LLM-Assisted Algorithm Discovery (viability: 5): https://sciencetostartup.com/paper/contrastive-concept-tree-search-for-llm-assisted-algorithm-discovery - Leverage Contrastive Concept-Tree Search to enhance LLM-assisted algorithm discovery with better search efficiency and interpretability.
- Digital Lifelong Learning in the Age of AI: Trends and Insights (viability: 5): https://sciencetostartup.com/paper/digital-lifelong-learning-in-the-age-of-ai-trends-and-insights - Optimize digital learning for lifelong learners using AI-powered insights.
- "I'm happy even though it's not real": GenAI Photo Editing as a Remembering Experience (viability: 2): https://sciencetostartup.com/paper/i-m-happy-even-though-it-s-not-real-genai-photo-editing-as-a-remembering-experience - Exploring how GenAI photo editing shapes remembering experiences without code or product focus.
- Task--Specificity Score: Measuring How Much Instructions Really Matter for Supervision (viability: 5): https://sciencetostartup.com/paper/task-specificity-score-measuring-how-much-instructions-really-matter-for-supervision - Introducing Task--Specificity Score (TSS) to quantify the impact of instruction specificity on language model output predictability.
- De-conflating Preference and Qualification: Constrained Dual-Perspective Reasoning for Job Recommendation with Large Language Models (viability: 7): https://sciencetostartup.com/paper/de-conflating-preference-and-qualification-constrained-dual-perspective-reasoning-for-job-recommendation-with-large-lang - JobRec provides a tailored job recommendation system that separates candidate preferences and employer qualifications using LLMs for superior matchmaking.
- Training and Simulation of Quadrupedal Robot in Adaptive Stair Climbing for Indoor Firefighting: An End-to-End Reinforcement Learning Approach (viability: 5): https://sciencetostartup.com/paper/training-and-simulation-of-quadrupedal-robot-in-adaptive-stair-climbing-for-indoor-firefighting-an-end-to-end-reinforcem - Develop an adaptive stair-climbing quadrupedal robot using deep reinforcement learning for enhanced indoor firefighting search operations.
- The Trigger in the Haystack: Extracting and Reconstructing LLM Backdoor Triggers (viability: 2): https://sciencetostartup.com/paper/the-trigger-in-the-haystack-extracting-and-reconstructing-llm-backdoor-triggers - Develop a practical scanner for identifying backdoors in language models, enhancing AI security.
- Shortcut Features as Top Eigenfunctions of NTK: A Linear Neural Network Case and More (viability: 3): https://sciencetostartup.com/paper/shortcut-features-as-top-eigenfunctions-of-ntk-a-linear-neural-network-case-and-more - Understanding shortcut learning in neural networks through NTK analysis shows potential theoretical insights but lacks clear product application.
- Evaluating LLMs When They Do Not Know the Answer: Statistical Evaluation of Mathematical Reasoning via Comparative Signals (viability: 5): https://sciencetostartup.com/paper/evaluating-llms-when-they-do-not-know-the-answer-statistical-evaluation-of-mathematical-reasoning-via-comparative-signal - Develop a statistically efficient evaluation framework for LLM mathematical reasoning with improved ranking accuracy using pairwise comparison signals.
- KANFIS A Neuro-Symbolic Framework for Interpretable and Uncertainty-Aware Learning (viability: 5): https://sciencetostartup.com/paper/kanfis-a-neuro-symbolic-framework-for-interpretable-and-uncertainty-aware-learning - Develop a compact neuro-symbolic framework, KANFIS, for interpretable and uncertainty-aware learning.
- Visual Reasoning over Time Series via Multi-Agent System (viability: 5): https://sciencetostartup.com/paper/visual-reasoning-over-time-series-via-multi-agent-system - MAS4TS is a tool-driven multi-agent system that enhances time series analysis through visual reasoning and agent coordination, achieving state-of-the-art performance.
- Consistency Deep Equilibrium Models (viability: 6): https://sciencetostartup.com/paper/consistency-deep-equilibrium-models - Accelerate deep equilibrium model inference using consistency distillation without losing accuracy.
- STAR: Similarity-guided Teacher-Assisted Refinement for Super-Tiny Function Calling Models (viability: 6): https://sciencetostartup.com/paper/star-similarity-guided-teacher-assisted-refinement-for-super-tiny-function-calling-models - Develop an efficient training framework for distilling LLM capabilities into super-tiny models for advanced AI agents.
- CVE-Factory: Scaling Expert-Level Agentic Tasks for Code Security Vulnerability (viability: 7): https://sciencetostartup.com/paper/cve-factory-scaling-expert-level-agentic-tasks-for-code-security-vulnerability - CVE-Factory transforms CVE metadata into automated, expert-level code security tasks.
- Distilling LLM Reasoning into Graph of Concept Predictors (viability: 8): https://sciencetostartup.com/paper/distilling-llm-reasoning-into-graph-of-concept-predictors - GCP offers a reasoning-aware distillation framework to efficiently transfer LLM capabilities into lightweight, interpretable models for cost-effective large-scale deployments.
- Causal Graph Spatial-Temporal Autoencoder for Reliable and Interpretable Process Monitoring (viability: 3): https://sciencetostartup.com/paper/causal-graph-spatial-temporal-autoencoder-for-reliable-and-interpretable-process-monitoring - A framework for interpretable industrial process monitoring via causal graph-based autoencoder models.
- Methods and Open Problems in Differentiable Social Choice: Learning Mechanisms, Decisions, and Alignment (viability: 3): https://sciencetostartup.com/paper/methods-and-open-problems-in-differentiable-social-choice-learning-mechanisms-decisions-and-alignment - Explore how differentiable social choice can optimize collective decision-making in machine learning systems.
- Adaptive Batch Sizes Using Non-Euclidean Gradient Noise Scales for Stochastic Sign and Spectral Descent (viability: 5): https://sciencetostartup.com/paper/adaptive-batch-sizes-using-non-euclidean-gradient-noise-scales-for-stochastic-sign-and-spectral-descent - Adaptive batch sizes for efficient training in non-Euclidean optimization contexts.
- Synthetic Data Augmentation for Medical Audio Classification: A Preliminary Evaluation (viability: 3): https://sciencetostartup.com/paper/synthetic-data-augmentation-for-medical-audio-classification-a-preliminary-evaluation - Evaluate the potential of synthetic data augmentation strategies for enhancing medical audio classification models.
- UAT-LITE: Inference-Time Uncertainty-Aware Attention for Pretrained Transformers (viability: 3): https://sciencetostartup.com/paper/uat-lite-inference-time-uncertainty-aware-attention-for-pretrained-transformers - Develop an uncertainty-aware attention framework for transformers to improve prediction calibration without retraining models.
- Nüwa: Mending the Spatial Integrity Torn by VLM Token Pruning (viability: 5): https://sciencetostartup.com/paper/n-wa-mending-the-spatial-integrity-torn-by-vlm-token-pruning - Nüwa enhances Vision Language Models with a novel token pruning method for better visual grounding performance.
- Refining Decision Boundaries In Anomaly Detection Using Similarity Search Within the Feature Space (viability: 7): https://sciencetostartup.com/paper/refining-decision-boundaries-in-anomaly-detection-using-similarity-search-within-the-feature-space - A cutting-edge anomaly detection framework for cybersecurity, reducing labeling effort by 80% while enhancing detection accuracy.
- DeltaEvolve: Accelerating Scientific Discovery through Momentum-Driven Evolution (viability: 3): https://sciencetostartup.com/paper/deltaevolve-accelerating-scientific-discovery-through-momentum-driven-evolution - DeltaEvolve uses a momentum-driven framework to improve scientific discovery via semantic deltas instead of full-code evolution.
- A Multi-scale Linear-time Encoder for Whole-Slide Image Analysis (viability: 8): https://sciencetostartup.com/paper/a-multi-scale-linear-time-encoder-for-whole-slide-image-analysis - MARBLE offers a scalable, efficient tool for multi-scale whole-slide image analysis with significant accuracy improvements.
- Reasoning about Reasoning: BAPO Bounds on Chain-of-Thought Token Complexity in LLMs (viability: 2): https://sciencetostartup.com/paper/reasoning-about-reasoning-bapo-bounds-on-chain-of-thought-token-complexity-in-llms - Exploring theoretical bounds on reasoning token complexity in LLMs to identify inference-time bottlenecks.
- A Random Matrix Theory Perspective on the Consistency of Diffusion Models (viability: 5): https://sciencetostartup.com/paper/a-random-matrix-theory-perspective-on-the-consistency-of-diffusion-models - Leverage random matrix theory to improve reproducibility and consistency in diffusion model training with insights validated on popular architectures.
- Mixture of Concept Bottleneck Experts (viability: 6): https://sciencetostartup.com/paper/mixture-of-concept-bottleneck-experts - Develop a flexible framework using Mixture of Concept Bottleneck Experts to enhance model interpretability and adaptability for specific user needs.
- "I May Not Have Articulated Myself Clearly": Diagnosing Dynamic Instability in LLM Reasoning at Inference Time (viability: 3): https://sciencetostartup.com/paper/i-may-not-have-articulated-myself-clearly-diagnosing-dynamic-instability-in-llm-reasoning-at-inference-time - Diagnostic method for predicting reasoning failure in LLMs using inference-time signals.
- STEER: Inference-Time Risk Control via Constrained Quality-Diversity Search (viability: 8): https://sciencetostartup.com/paper/steer-inference-time-risk-control-via-constrained-quality-diversity-search - STEER enables adjustable risk control in language models for critical decision-making contexts like clinical triage.
- Semantics-Aware Generative Latent Data Augmentation for Learning in Low-Resource Domains (viability: 7): https://sciencetostartup.com/paper/semantics-aware-generative-latent-data-augmentation-for-learning-in-low-resource-domains - GeLDA offers semantics-aware data augmentation to improve model performance in low-resource scenarios using generative techniques.
- Tabula RASA: Exposing and Breaking the Relational Bottleneck in Transformers (viability: 3): https://sciencetostartup.com/paper/tabula-rasa-exposing-and-breaking-the-relational-bottleneck-in-transformers - RASA enhances transformer models with relational reasoning capabilities through minimal structural modifications.
- LmPT: Conditional Point Transformer for Anatomical Landmark Detection on 3D Point Clouds (viability: 7): https://sciencetostartup.com/paper/lmpt-conditional-point-transformer-for-anatomical-landmark-detection-on-3d-point-clouds - Automated tool for cross-species anatomical landmark detection using point cloud data to aid in medical applications.
- Joint Learning of Hierarchical Neural Options and Abstract World Model (viability: 3): https://sciencetostartup.com/paper/joint-learning-of-hierarchical-neural-options-and-abstract-world-model - AgentOWL enhances AI skill composition efficiency in Object-Centric Atari games through hierarchical learning.
- Simulating Human Audiovisual Search Behavior (viability: 2): https://sciencetostartup.com/paper/simulating-human-audiovisual-search-behavior - Sensonaut simulates human audiovisual search behavior to improve interface design by minimizing search cost and cognitive load.
- Provable Effects of Data Replay in Continual Learning: A Feature Learning Perspective (viability: 2): https://sciencetostartup.com/paper/provable-effects-of-data-replay-in-continual-learning-a-feature-learning-perspective - Develop a framework to mitigate catastrophic forgetting in continual learning using data-replay methods.
- Scaling Small Agents Through Strategy Auctions (viability: 2): https://sciencetostartup.com/paper/scaling-small-agents-through-strategy-auctions - Develop a marketplace-inspired system for efficient task allocation among small AI agents using Strategy Auctions for improved performance at reduced costs.
- Entropy-Guided Dynamic Tokens for Graph-LLM Alignment in Molecular Understanding (viability: 6): https://sciencetostartup.com/paper/entropy-guided-dynamic-tokens-for-graph-llm-alignment-in-molecular-understanding - EDT-Former enhances Large Language Models for efficient molecular understanding without costly fine-tuning.
- TopoPrune: Robust Data Pruning via Unified Latent Space Topology (viability: 5): https://sciencetostartup.com/paper/topoprune-robust-data-pruning-via-unified-latent-space-topology - Develop a robust data pruning tool using topology to ensure stable, data-efficient learning across noisy and diverse environments.
- CAPS: Unifying Attention, Recurrence, and Alignment in Transformer-based Time Series Forecasting (viability: 5): https://sciencetostartup.com/paper/caps-unifying-attention-recurrence-and-alignment-in-transformer-based-time-series-forecasting - CAPS enhances time series forecasting with a novel attention mechanism outperforming traditional models.
- Every Bit Counts: A Theoretical Study of Precision-Expressivity Tradeoffs in Quantized Transformers (viability: 3): https://sciencetostartup.com/paper/every-bit-counts-a-theoretical-study-of-precision-expressivity-tradeoffs-in-quantized-transformers - Theoretical insights into quantization effects on Transformer expressivity guide precision selection for sensitive tasks.
- Sparsely Supervised Diffusion (viability: 4): https://sciencetostartup.com/paper/sparsely-supervised-diffusion - Implement a sparse pixel masking strategy to improve diffusion models by enhancing global consistency and reducing overfitting.
- Monotonicity as an Architectural Bias for Robust Language Models (viability: 3): https://sciencetostartup.com/paper/monotonicity-as-an-architectural-bias-for-robust-language-models - Exploring monotonicity as a means to enhance robustness in large language models against adversarial attacks.
- MARA: Continuous SE(3)-Equivariant Attention for Molecular Force Fields (viability: 5): https://sciencetostartup.com/paper/mara-continuous-se-3-equivariant-attention-for-molecular-force-fields - MARA enhances molecular force field models by providing a more flexible and accurate SE(3)-equivariant attention mechanism.
- MARS: Modular Agent with Reflective Search for Automated AI Research (viability: 5): https://sciencetostartup.com/paper/mars-modular-agent-with-reflective-search-for-automated-ai-research - MARS optimizes autonomous AI research by balancing performance with execution costs using a modular design and reflective search.
- Reward-free Alignment for Conflicting Objectives (viability: 6): https://sciencetostartup.com/paper/reward-free-alignment-for-conflicting-objectives - Develop a reward-free alignment framework that improves multi-objective LLM alignment using conflict-averse gradient descent.
- PixelGen: Pixel Diffusion Beats Latent Diffusion with Perceptual Loss (viability: 7): https://sciencetostartup.com/paper/pixelgen-pixel-diffusion-beats-latent-diffusion-with-perceptual-loss - PixelGen offers a simpler, powerful image generation tool by surpassing traditional diffusion methods using perceptual loss.
- RE-TRAC: REcursive TRAjectory Compression for Deep Search Agents (viability: 4): https://sciencetostartup.com/paper/re-trac-recursive-trajectory-compression-for-deep-search-agents - Optimize search agents with Re-TRAC for efficient trajectory handling and targeted exploration.
- Flow Policy Gradients for Robot Control (viability: 5): https://sciencetostartup.com/paper/flow-policy-gradients-for-robot-control - Flow matching policy gradients enhance expressive robot control policies, potentially revolutionizing sim-to-real transfer.
- AgentRx: Diagnosing AI Agent Failures from Execution Trajectories (viability: 6): https://sciencetostartup.com/paper/agentrx-diagnosing-ai-agent-failures-from-execution-trajectories - AgentRx is an automated diagnostic framework for identifying failure points in AI agent trajectories, enhancing agent deployment reliability.
- MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents (viability: 6): https://sciencetostartup.com/paper/memskill-learning-and-evolving-memory-skills-for-self-evolving-agents - MemSkill enhances LLM agent memory systems with learnable and evolving memory skills for improved adaptability.
- Multi-head automated segmentation by incorporating detection head into the contextual layer neural network (viability: 6): https://sciencetostartup.com/paper/multi-head-automated-segmentation-by-incorporating-detection-head-into-the-contextual-layer-neural-network - A robust auto-segmentation tool for radiotherapy that integrates detection to reduce false positives and improve reliability.
- Breaking the Reversal Curse in Autoregressive Language Models via Identity Bridge (viability: 3): https://sciencetostartup.com/paper/breaking-the-reversal-curse-in-autoregressive-language-models-via-identity-bridge - Enhance LLM logical reasoning by addressing the reversal curse through a novel training data regularization.
- Avenir-Web: Human-Experience-Imitating Multimodal Web Agents with Mixture of Grounding Experts (viability: 8): https://sciencetostartup.com/paper/avenir-web-human-experience-imitating-multimodal-web-agents-with-mixture-of-grounding-experts - Avenir-Web: An open-source state-of-the-art agent for executing tasks on dynamic web interfaces using multimodal grounding and adaptive memory.
- MentisOculi: Revealing the Limits of Reasoning with Mental Imagery (viability: 2): https://sciencetostartup.com/paper/mentisoculi-revealing-the-limits-of-reasoning-with-mental-imagery - MentisOculi investigates the limitations of reasoning with mental imagery in multimodal models.
- Abstract Activation Spaces for Content-Invariant Reasoning in Large Language Models (viability: 2): https://sciencetostartup.com/paper/abstract-activation-spaces-for-content-invariant-reasoning-in-large-language-models - Enhancing LLMs' formal reasoning by separating structural inference from lexical semantics to reduce semantic interference.
- Drift-Bench: Diagnosing Cooperative Breakdowns in LLM Agents under Input Faults via Multi-Turn Interaction (viability: 5): https://sciencetostartup.com/paper/drift-bench-diagnosing-cooperative-breakdowns-in-llm-agents-under-input-faults-via-multi-turn-interaction - A diagnostic benchmark for evaluating and improving LLM agent safety through multi-turn interaction fault diagnosis.
- World-Gymnast: Training Robots with Reinforcement Learning in a World Model (viability: 7): https://sciencetostartup.com/paper/world-gymnast-training-robots-with-reinforcement-learning-in-a-world-model - World-Gymnast enables scalable robot training through RL in cloud-based world models, outperforming traditional methods.
- Thinking with Comics: Enhancing Multimodal Reasoning through Structured Visual Storytelling (viability: 3): https://sciencetostartup.com/paper/thinking-with-comics-enhancing-multimodal-reasoning-through-structured-visual-storytelling - Develop a visual reasoning tool leveraging comics to enhance multimodal reasoning efficiency and effectiveness.
- Active Causal Experimentalist (ACE): Learning Intervention Strategies via Direct Preference Optimization (viability: 6): https://sciencetostartup.com/paper/active-causal-experimentalist-ace-learning-intervention-strategies-via-direct-preference-optimization - Develop a tool for adaptive causal experiment strategies using preference-based learning.
- UniReason 1.0: A Unified Reasoning Framework for World Knowledge Aligned Image Generation and Editing (viability: 7): https://sciencetostartup.com/paper/unireason-1-0-a-unified-reasoning-framework-for-world-knowledge-aligned-image-generation-and-editing - UniReason offers a unified framework for enhanced image generation and editing through world knowledge and reasoning capabilities.
- Poly-attention: a general scheme for higher-order self-attention (viability: 2): https://sciencetostartup.com/paper/poly-attention-a-general-scheme-for-higher-order-self-attention - Develop a new quadratic-time attention mechanism for higher-order token interactions in AI models.
- SafeGround: Know When to Trust GUI Grounding Models via Uncertainty Calibration (viability: 7): https://sciencetostartup.com/paper/safeground-know-when-to-trust-gui-grounding-models-via-uncertainty-calibration - Develop a risk-aware GUI grounding tool for safe automated interactions with uncertainty calibration.
- Structure Enables Effective Self-Localization of Errors in LLMs (viability: 3): https://sciencetostartup.com/paper/structure-enables-effective-self-localization-of-errors-in-llms - Develops a framework for LLMs to self-localize errors in structured reasoning without external verification.
- ReasonEdit: Editing Vision-Language Models using Human Reasoning (viability: 5): https://sciencetostartup.com/paper/reasonedit-editing-vision-language-models-using-human-reasoning - Enhance vision-language model accuracy by integrating human reasoning for improved edit generalization.
- Context Learning for Multi-Agent Discussion (viability: 3): https://sciencetostartup.com/paper/context-learning-for-multi-agent-discussion - An advanced multi-LLM context learning method to enhance consistent multi-agent discussions.
- Why Steering Works: Toward a Unified View of Language Model Parameter Dynamics (viability: 5): https://sciencetostartup.com/paper/why-steering-works-toward-a-unified-view-of-language-model-parameter-dynamics - Unified framework for improving language model control methods leveraging SPLIT to balance preference and utility.
- Rethinking Generative Recommender Tokenizer: Recsys-Native Encoding and Semantic Quantization Beyond LLMs (viability: 7): https://sciencetostartup.com/paper/rethinking-generative-recommender-tokenizer-recsys-native-encoding-and-semantic-quantization-beyond-llms - ReSID enhances generative recommender systems with efficient semantic ID encoding and quantization to significantly improve prediction accuracy and reduce tokenization cost.
- Building a Correct-by-Design Lakehouse. Data Contracts, Versioning, and Transactional Pipelines for Humans and Agents (viability: 3): https://sciencetostartup.com/paper/building-a-correct-by-design-lakehouse-data-contracts-versioning-and-transactional-pipelines-for-humans-and-agents - Develop a code-first lakehouse platform called Bauplan that ensures data safety and pipeline atomicity through typed table contracts and Git-like versioning.
- VQ-Style: Disentangling Style and Content in Motion with Residual Quantized Representations (viability: 6): https://sciencetostartup.com/paper/vq-style-disentangling-style-and-content-in-motion-with-residual-quantized-representations - A novel RVQ-VAE framework for disentangling style and content in human motion data enabling efficient style transfer without fine-tuning.
- TTT-Parkour: Rapid Test-Time Training for Perceptive Robot Parkour (viability: 7): https://sciencetostartup.com/paper/ttt-parkour-rapid-test-time-training-for-perceptive-robot-parkour - Enable humanoid robots to perform complex parkour on varied terrains using rapid test-time training.
- Position: Explaining Behavioral Shifts in Large Language Models Requires a Comparative Approach (viability: 2): https://sciencetostartup.com/paper/position-explaining-behavioral-shifts-in-large-language-models-requires-a-comparative-approach - Developing a framework for comparing behavioral shifts in language models to improve explainability.
- Advancing General-Purpose Reasoning Models with Modular Gradient Surgery (viability: 4): https://sciencetostartup.com/paper/advancing-general-purpose-reasoning-models-with-modular-gradient-surgery - Modular Gradient Surgery enhances domain-agnostic learning in large reasoning models.
- Decoupling Generalizability and Membership Privacy Risks in Neural Networks (viability: 2): https://sciencetostartup.com/paper/decoupling-generalizability-and-membership-privacy-risks-in-neural-networks - Innovative training principle to decouple generalizability and privacy risks in neural networks.
- Hallucination or Creativity: How to Evaluate AI-Generated Scientific Stories? (viability: 2): https://sciencetostartup.com/paper/hallucination-or-creativity-how-to-evaluate-ai-generated-scientific-stories - Develop a composite metric, StoryScore, to evaluate AI-generated scientific stories for narrative quality and factual accuracy.
- OmniCode: A Benchmark for Evaluating Software Engineering Agents (viability: 5): https://sciencetostartup.com/paper/omnicode-a-benchmark-for-evaluating-software-engineering-agents - OmniCode offers a comprehensive benchmark for evaluating software engineering agents across diverse tasks and programming languages.
- Geometry- and Relation-Aware Diffusion for EEG Super-Resolution (viability: 7): https://sciencetostartup.com/paper/geometry-and-relation-aware-diffusion-for-eeg-super-resolution - TopoDiff provides a novel geometry- and relation-aware model for enhancing EEG spatial resolution with improved downstream task performance.
- Spectral Superposition: A Theory of Feature Geometry (viability: 2): https://sciencetostartup.com/paper/spectral-superposition-a-theory-of-feature-geometry - A spectral analysis theory to study neural network feature geometry.
- Towards AI Evaluation in Domain-Specific RAG Systems: The AgriHubi Case Study (viability: 3): https://sciencetostartup.com/paper/towards-ai-evaluation-in-domain-specific-rag-systems-the-agrihubi-case-study - AgriHubi enhances agricultural decision support in Finnish using a domain-specific RAG system.
- More Than a Quick Glance: Overcoming the Greedy Bias in KV-Cache Compression (viability: 5): https://sciencetostartup.com/paper/more-than-a-quick-glance-overcoming-the-greedy-bias-in-kv-cache-compression - Develop LASER-KV for improved large context window support in LLMs through innovative KV cache compression without sacrificing performance.
- Hierarchical Adaptive Eviction for KV Cache Management in Multimodal Language Models (viability: 3): https://sciencetostartup.com/paper/hierarchical-adaptive-eviction-for-kv-cache-management-in-multimodal-language-models - Optimize memory and computational efficiency in Multimodal LLMs with Hierarchical Adaptive Eviction (HAE) for better text-visual token management.
- TIDE: Trajectory-based Diagnostic Evaluation of Test-Time Improvement in LLM Agents (viability: 3): https://sciencetostartup.com/paper/tide-trajectory-based-diagnostic-evaluation-of-test-time-improvement-in-llm-agents - TIDE provides a framework for diagnosing and improving autonomous LLM agent performance through test-time interactions.
- State Rank Dynamics in Linear Attention LLMs (viability: 3): https://sciencetostartup.com/paper/state-rank-dynamics-in-linear-attention-llms - Study uncovers intrinsic state dynamics in Linear Attention LLMs, revealing opportunities for zero-shot pruning to reduce KV-cache overhead.
- Back to the Future: Look-ahead Augmentation and Parallel Self-Refinement for Time Series Forecasting (viability: 8): https://sciencetostartup.com/paper/back-to-the-future-look-ahead-augmentation-and-parallel-self-refinement-for-time-series-forecasting - BTTF enhances time series forecasting accuracy by leveraging look-ahead augmentation and self-refinement.
- Learning Generative Selection for Best-of-N (viability: 7): https://sciencetostartup.com/paper/learning-generative-selection-for-best-of-n - Develop scalable generative selection capabilities for small models using reinforcement learning to improve reasoning tasks.
- EvoMU: Evolutionary Machine Unlearning (viability: 7): https://sciencetostartup.com/paper/evomu-evolutionary-machine-unlearning - EvoMU automates the discovery of optimal data-specific unlearning loss functions, optimizing model privacy and utility without human intervention.
- CAM: A Causality-based Analysis Framework for Multi-Agent Code Generation Systems (viability: 7): https://sciencetostartup.com/paper/cam-a-causality-based-analysis-framework-for-multi-agent-code-generation-systems - CAM uses causality analysis to enhance multi-agent code generation systems for improved system design and performance.
- Understanding the Reversal Curse Mitigation in Masked Diffusion Models through Attention and Training Dynamics (viability: 2): https://sciencetostartup.com/paper/understanding-the-reversal-curse-mitigation-in-masked-diffusion-models-through-attention-and-training-dynamics - Understanding mechanisms in masked diffusion models to mitigate reversal curse in language processing.
- Unifying Masked Diffusion Models with Various Generation Orders and Beyond (viability: 6): https://sciencetostartup.com/paper/unifying-masked-diffusion-models-with-various-generation-orders-and-beyond - Develop smarter AI language generators by harnessing order-expressive mask diffusion models that adaptively learn generation order.
- The Verification Crisis: Expert Perceptions of GenAI Disinformation and the Case for Reproducible Provenance (viability: 3): https://sciencetostartup.com/paper/the-verification-crisis-expert-perceptions-of-genai-disinformation-and-the-case-for-reproducible-provenance - Develop reproducible provenance standards to combat GenAI-driven disinformation.
- See2Refine: Vision-Language Feedback Improves LLM-Based eHMI Action Designers (viability: 5): https://sciencetostartup.com/paper/see2refine-vision-language-feedback-improves-llm-based-ehmi-action-designers - See2Refine automates external HMI action design for autonomous vehicles using vision-language feedback, eliminating the need for human annotators.
- FiLoRA: Focus-and-Ignore LoRA for Controllable Feature Reliance (viability: 8): https://sciencetostartup.com/paper/filora-focus-and-ignore-lora-for-controllable-feature-reliance - FiLoRA offers controllable feature reliance for robust multimodal model predictions using parameter-efficient adaptations.
- FORLER: Federated Offline Reinforcement Learning with Q-Ensemble and Actor Rectification (viability: 5): https://sciencetostartup.com/paper/forler-federated-offline-reinforcement-learning-with-q-ensemble-and-actor-rectification - FORLER enhances federated offline reinforcement learning by mitigating policy pollution through Q-ensemble aggregation and actor rectification.
- Rethinking the Role of Entropy in Optimizing Tool-Use Behaviors for Large Language Model Agents (viability: 3): https://sciencetostartup.com/paper/rethinking-the-role-of-entropy-in-optimizing-tool-use-behaviors-for-large-language-model-agents - Optimize LLM agent tool-use behavior with entropy-reduction-based rewards for better efficiency and performance.
- Bandwidth-Efficient Multi-Agent Communication through Information Bottleneck and Vector Quantization (viability: 8): https://sciencetostartup.com/paper/bandwidth-efficient-multi-agent-communication-through-information-bottleneck-and-vector-quantization - A bandwidth-efficient communication framework for multi-agent systems using information bottleneck theory and vector quantization, ideal for real-time constrained environments like autonomous vehicle fleets.
- Constrained Process Maps for Multi-Agent Generative AI Workflows (viability: 5): https://sciencetostartup.com/paper/constrained-process-maps-for-multi-agent-generative-ai-workflows - Develop a multi-agent compliance system for regulated workflows to reduce human review and increase accuracy.
- One Size, Many Fits: Aligning Diverse Group-Wise Click Preferences in Large-Scale Advertising Image Generation (viability: 9): https://sciencetostartup.com/paper/one-size-many-fits-aligning-diverse-group-wise-click-preferences-in-large-scale-advertising-image-generation - A framework to tailor advertising images for diverse user groups, boosting CTR and ad effectiveness.
- ClueTracer: Question-to-Vision Clue Tracing for Training-Free Hallucination Suppression in Multimodal Reasoning (viability: 7): https://sciencetostartup.com/paper/cluetracer-question-to-vision-clue-tracing-for-training-free-hallucination-suppression-in-multimodal-reasoning - ClueTracer enhances multimodal reasoning models by suppressing hallucinations without additional training.
- Preserve-Then-Quantize: Balancing Rank Budgets for Quantization Error Reconstruction in LLMs (viability: 6): https://sciencetostartup.com/paper/preserve-then-quantize-balancing-rank-budgets-for-quantization-error-reconstruction-in-llms - Develop a solution to reduce accuracy loss in post-training quantization for large language models.
- SurfSplat: Conquering Feedforward 2D Gaussian Splatting with Surface Continuity Priors (viability: 6): https://sciencetostartup.com/paper/surfsplat-conquering-feedforward-2d-gaussian-splatting-with-surface-continuity-priors - SurfSplat enables high-fidelity 3D scene reconstruction from sparse images using innovative Gaussian splatting techniques.
- On the Limits of Layer Pruning for Generative Reasoning in LLMs (viability: 5): https://sciencetostartup.com/paper/on-the-limits-of-layer-pruning-for-generative-reasoning-in-llms - Optimize LLM performance on generative reasoning tasks using a novel layer pruning strategy.
- Emergent Analogical Reasoning in Transformers (viability: 2): https://sciencetostartup.com/paper/emergent-analogical-reasoning-in-transformers - Exploring emergent analogical reasoning in Transformers using category theory concepts.
- Evolving from Tool User to Creator via Training-Free Experience Reuse in Multimodal Reasoning (viability: 5): https://sciencetostartup.com/paper/evolving-from-tool-user-to-creator-via-training-free-experience-reuse-in-multimodal-reasoning - Create adaptive, self-updating multimodal reasoning tools without training through a unique experience reuse framework.
- FlyPrompt: Brain-Inspired Random-Expanded Routing with Temporal-Ensemble Experts for General Continual Learning (viability: 4): https://sciencetostartup.com/paper/flyprompt-brain-inspired-random-expanded-routing-with-temporal-ensemble-experts-for-general-continual-learning - FlyPrompt is a brain-inspired framework for improved continual learning via expert routing and temporal ensemble methods.
- IntraSlice: Towards High-Performance Structural Pruning with Block-Intra PCA for LLMs (viability: 6): https://sciencetostartup.com/paper/intraslice-towards-high-performance-structural-pruning-with-block-intra-pca-for-llms - IntraSlice offers efficient LLM deployment via advanced PCA-based structural pruning, reducing model size without performance loss.
- Your AI-Generated Image Detector Can Secretly Achieve SOTA Accuracy, If Calibrated (viability: 7): https://sciencetostartup.com/paper/your-ai-generated-image-detector-can-secretly-achieve-sota-accuracy-if-calibrated - A lightweight post-hoc calibration tool improving AI-generated image detector accuracy through Bayesian decision theory.
- Small Generalizable Prompt Predictive Models Can Steer Efficient RL Post-Training of Large Reasoning Models (viability: 6): https://sciencetostartup.com/paper/small-generalizable-prompt-predictive-models-can-steer-efficient-rl-post-training-of-large-reasoning-models - Generalizable Predictive Prompt Selection optimizes large language model training by efficiently selecting diverse, informative prompts.
- VLM-Guided Experience Replay (viability: 6): https://sciencetostartup.com/paper/vlm-guided-experience-replay - Enhance reinforcement learning by using VLMs to optimize replay buffer prioritization for improved efficiency and success rates.
- DomusFM: A Foundation Model for Smart-Home Sensor Data (viability: 7): https://sciencetostartup.com/paper/domusfm-a-foundation-model-for-smart-home-sensor-data - DomusFM revolutionizes smart-home analytics with a foundation model that excels in activity recognition using minimal labeled data and superior semantic understanding.
- DSXFormer: Dual-Pooling Spectral Squeeze-Expansion and Dynamic Context Attention Transformer for Hyperspectral Image Classification (viability: 7): https://sciencetostartup.com/paper/dsxformer-dual-pooling-spectral-squeeze-expansion-and-dynamic-context-attention-transformer-for-hyperspectral-image-clas - DSXFormer offers a state-of-the-art solution for hyperspectral image classification with enhanced spectral discriminability and efficiency.
- Learning Sparse Visual Representations via Spatial-Semantic Factorization (viability: 5): https://sciencetostartup.com/paper/learning-sparse-visual-representations-via-spatial-semantic-factorization - STELLAR is a framework that efficiently bridges semantic understanding and image reconstruction in computer vision via spatial-semantic factorization.
- Geometric Analysis of Token Selection in Multi-Head Attention (viability: 2): https://sciencetostartup.com/paper/geometric-analysis-of-token-selection-in-multi-head-attention - Geometric framework for analyzing multi-head attention in LLMs to enhance interpretability and design efficiency.
- ES-MemEval: Benchmarking Conversational Agents on Personalized Long-Term Emotional Support (viability: 6): https://sciencetostartup.com/paper/es-memeval-benchmarking-conversational-agents-on-personalized-long-term-emotional-support - A benchmark and dataset to enhance and test long-term memory capabilities for conversational agents in personalized emotional support scenarios.
- CloDS: Visual-Only Unsupervised Cloth Dynamics Learning in Unknown Conditions (viability: 6): https://sciencetostartup.com/paper/clods-visual-only-unsupervised-cloth-dynamics-learning-in-unknown-conditions - Develop an unsupervised learning framework to simulate cloth dynamics from visual data for enhanced CGI and animation realism.
- Synesthesia of Vehicles: Tactile Data Synthesis from Visual Inputs (viability: 7): https://sciencetostartup.com/paper/synesthesia-of-vehicles-tactile-data-synthesis-from-visual-inputs - Transform visual data into tactile insights to enhance autonomous vehicle safety using cross-modal alignment and generative models.
- Beyond Precision: Training-Inference Mismatch is an Optimization Problem and Simple LR Scheduling Fixes It (viability: 6): https://sciencetostartup.com/paper/beyond-precision-training-inference-mismatch-is-an-optimization-problem-and-simple-lr-scheduling-fixes-it - A novel learning rate scheduler to stabilize RL training for large language models by dynamically adjusting based on model response length.
- Stein-Rule Shrinkage for Stochastic Gradient Estimation in High Dimensions (viability: 6): https://sciencetostartup.com/paper/stein-rule-shrinkage-for-stochastic-gradient-estimation-in-high-dimensions - A shrinkage-based enhancement to stochastic gradient methods that improves performance in high-dimensional learning tasks.
- Efficient Cross-Architecture Knowledge Transfer for Large-Scale Online User Response Prediction (viability: 8): https://sciencetostartup.com/paper/efficient-cross-architecture-knowledge-transfer-for-large-scale-online-user-response-prediction - CrossAdapt offers an efficient method for deploying new architectures in online prediction systems by minimizing retraining costs and performance degradation.
- DIA-CLIP: a universal representation learning framework for zero-shot DIA proteomics (viability: 6): https://sciencetostartup.com/paper/dia-clip-a-universal-representation-learning-framework-for-zero-shot-dia-proteomics - DIA-CLIP offers a universal representation learning framework for zero-shot peptide-spectrum matching in proteomics, improving protein identification rates significantly.
- <SOG_k>: One LLM Token for Explicit Graph Structural Understanding (viability: 7): https://sciencetostartup.com/paper/sog-k-one-llm-token-for-explicit-graph-structural-understanding - Introduce a special token for graph structures in LLMs to enhance understanding and reasoning in graph-level tasks.
- CoMeT: Collaborative Memory Transformer for Efficient Long Context Modeling (viability: 8): https://sciencetostartup.com/paper/comet-collaborative-memory-transformer-for-efficient-long-context-modeling - CoMeT enables efficient long-context processing in existing Transformers with constant memory usage.
- Backdoor Sentinel: Detecting and Detoxifying Backdoors in Diffusion Models via Temporal Noise Consistency (viability: 3): https://sciencetostartup.com/paper/backdoor-sentinel-detecting-and-detoxifying-backdoors-in-diffusion-models-via-temporal-noise-consistency - TNC-Defense offers a novel method for detecting and detoxifying backdoors in diffusion models using temporal noise consistency.
- PRISM: Parametrically Refactoring Inference for Speculative Sampling Draft Models (viability: 5): https://sciencetostartup.com/paper/prism-parametrically-refactoring-inference-for-speculative-sampling-draft-models - PRISM refactors computational pathways of draft models for faster LLM decoding, boosting throughput by over 2.6x.
- Controlling Exploration-Exploitation in GFlowNets via Markov Chain Perspectives (viability: 5): https://sciencetostartup.com/paper/controlling-exploration-exploitation-in-gflownets-via-markov-chain-perspectives - Optimize GFlowNets training process for better exploration-exploitation balance using Markov chain principles.
- Probability-Entropy Calibration: An Elastic Indicator for Adaptive Fine-tuning (viability: 5): https://sciencetostartup.com/paper/probability-entropy-calibration-an-elastic-indicator-for-adaptive-fine-tuning - Adaptive fine-tuning using a novel probability-entropy calibration for enhanced token reweighting in NLP models.
- MACD: Model-Aware Contrastive Decoding via Counterfactual Data (viability: 7): https://sciencetostartup.com/paper/macd-model-aware-contrastive-decoding-via-counterfactual-data - Develop targeted counterfactual data-based inference technique to reduce hallucinations in Video-LLMs.
- BBPE16: UTF-16-based byte-level byte-pair encoding for improved multilingual speech recognition (viability: 5): https://sciencetostartup.com/paper/bbpe16-utf-16-based-byte-level-byte-pair-encoding-for-improved-multilingual-speech-recognition - BBPE16 offers an efficient multilingual tokenization solution for ASR, improving performance for non-Latin scripts without increasing computational costs.
- Optimizing Prompts for Large Language Models: A Causal Approach (viability: 8): https://sciencetostartup.com/paper/optimizing-prompts-for-large-language-models-a-causal-approach - Causal Prompt Optimization offers a robust method to tailor LLM prompts for specific queries, enhancing enterprise workflows by reducing dependency on costly real-time evaluations.
- Beyond Mode Elicitation: Diversity-Preserving Reinforcement Learning via Latent Diffusion Reasoner (viability: 7): https://sciencetostartup.com/paper/beyond-mode-elicitation-diversity-preserving-reinforcement-learning-via-latent-diffusion-reasoner - LaDi-RL improves AI reasoning diversity by optimizing exploration in latent spaces instead of discrete token spaces.
- Meta Engine: A Unified Semantic Query Engine on Heterogeneous LLM-Based Query Systems (viability: 7): https://sciencetostartup.com/paper/meta-engine-a-unified-semantic-query-engine-on-heterogeneous-llm-based-query-systems - Meta Engine unifies multiple semantic query systems for efficient multi-modal data querying, significantly outperforming existing solutions.
- Semantic-aware Wasserstein Policy Regularization for Large Language Model Alignment (viability: 5): https://sciencetostartup.com/paper/semantic-aware-wasserstein-policy-regularization-for-large-language-model-alignment - Leverage semantic-aware Wasserstein policy regularization to enhance large language model alignment with human feedback.
- FreshMem: Brain-Inspired Frequency-Space Hybrid Memory for Streaming Video Understanding (viability: 3): https://sciencetostartup.com/paper/freshmem-brain-inspired-frequency-space-hybrid-memory-for-streaming-video-understanding - FreshMem provides a novel brain-inspired architecture to improve the adaptability and coherence of streaming video understanding.
- TRIP-Bench: A Benchmark for Long-Horizon Interactive Agents in Real-World Scenarios (viability: 5): https://sciencetostartup.com/paper/trip-bench-a-benchmark-for-long-horizon-interactive-agents-in-real-world-scenarios - Develop advanced long-horizon interactive agents using TRIP-Bench for real-world travel planning scenarios.
- Real-Time Loop Closure Detection in Visual SLAM via NetVLAD and Faiss (viability: 6): https://sciencetostartup.com/paper/real-time-loop-closure-detection-in-visual-slam-via-netvlad-and-faiss - Real-time loop closure detection for Visual SLAM using NetVLAD and Faiss offers an efficient and robust alternative to classic methods.
- AI-Assisted Adaptive Rendering for High-Frequency Security Telemetry in Web Interfaces (viability: 7): https://sciencetostartup.com/paper/ai-assisted-adaptive-rendering-for-high-frequency-security-telemetry-in-web-interfaces - AI-assisted adaptive rendering framework for efficient real-time security telemetry visualization.
- CoDiQ: Test-Time Scaling for Controllable Difficult Question Generation (viability: 5): https://sciencetostartup.com/paper/codiq-test-time-scaling-for-controllable-difficult-question-generation - CoDiQ offers a framework to generate challenging, solvable questions for better training of large reasoning models.
- Contribution-aware Token Compression for Efficient Video Understanding via Reinforcement Learning (viability: 8): https://sciencetostartup.com/paper/contribution-aware-token-compression-for-efficient-video-understanding-via-reinforcement-learning - Optimize video understanding efficiency through a contribution-aware token compression algorithm leveraging reinforcement learning.
- The Effect of Mini-Batch Noise on the Implicit Bias of Adam (viability: 2): https://sciencetostartup.com/paper/the-effect-of-mini-batch-noise-on-the-implicit-bias-of-adam - A theoretical investigation into how mini-batch noise influences implicit bias in the Adam optimizer.
- Toward Enhancing Representation Learning in Federated Multi-Task Settings (viability: 5): https://sciencetostartup.com/paper/toward-enhancing-representation-learning-in-federated-multi-task-settings - FedMuscle enhances representation learning in federated multi-task settings with a novel contrastive learning approach.
- AgroFlux: A Spatial-Temporal Benchmark for Carbon and Nitrogen Flux Prediction in Agricultural Ecosystems (viability: 4): https://sciencetostartup.com/paper/agroflux-a-spatial-temporal-benchmark-for-carbon-and-nitrogen-flux-prediction-in-agricultural-ecosystems - Launch a leading dataset and benchmark to advance AI models for agroecosystem GHG predictions.
- Reasoning with Autoregressive-Diffusion Collaborative Thoughts (viability: 2): https://sciencetostartup.com/paper/reasoning-with-autoregressive-diffusion-collaborative-thoughts - Unified framework for autoregressive and diffusion model collaboration to improve reasoning and generation reliability.
- Boosting Maximum Entropy Reinforcement Learning via One-Step Flow Matching (viability: 8): https://sciencetostartup.com/paper/boosting-maximum-entropy-reinforcement-learning-via-one-step-flow-matching - Accelerate RL with FLAME, delivering one-step flow matching for optimal policy efficiency and low latency.
- Adaptive Rollout Allocation for Online Reinforcement Learning with Verifiable Rewards (viability: 6): https://sciencetostartup.com/paper/adaptive-rollout-allocation-for-online-reinforcement-learning-with-verifiable-rewards - Optimize RL training efficiency with adaptive rollout strategies using Variance-Informed Predictive allocation.
- On the Fragility of AI-Based Channel Decoders under Small Channel Perturbations (viability: 2): https://sciencetostartup.com/paper/on-the-fragility-of-ai-based-channel-decoders-under-small-channel-perturbations - This paper analyzes the robustness of AI-based channel decoders under adversarial channel perturbations.
- DrawSim-PD: Simulating Student Science Drawings to Support NGSS-Aligned Teacher Diagnostic Reasoning (viability: 8): https://sciencetostartup.com/paper/drawsim-pd-simulating-student-science-drawings-to-support-ngss-aligned-teacher-diagnostic-reasoning - DrawSim-PD is a generative framework for simulating student science drawings to enhance teacher diagnostic training under NGSS standards.
- Generative Visual Code Mobile World Models (viability: 8): https://sciencetostartup.com/paper/generative-visual-code-mobile-world-models - Build the next generation of mobile GUI agents with gWorld: efficient, high-fidelity, code-generating visual world models.
- InfoTok: Regulating Information Flow for Capacity-Constrained Shared Visual Tokenization in Unified MLLMs (viability: 3): https://sciencetostartup.com/paper/infotok-regulating-information-flow-for-capacity-constrained-shared-visual-tokenization-in-unified-mllms - Innovative visual tokenization mechanism for improved image understanding and generation in unified MLLMs.
- Plain Transformers are Surprisingly Powerful Link Predictors (viability: 5): https://sciencetostartup.com/paper/plain-transformers-are-surprisingly-powerful-link-predictors - PENCIL leverages plain Transformers for efficient and scalable link prediction in large-scale graphs.
- MAGIC: A Co-Evolving Attacker-Defender Adversarial Game for Robust LLM Safety (viability: 7): https://sciencetostartup.com/paper/magic-a-co-evolving-attacker-defender-adversarial-game-for-robust-llm-safety - MAGIC provides a dynamic system for enhancing LLM safety through co-evolving attacker-defender mechanisms.
- Making Avatars Interact: Towards Text-Driven Human-Object Interaction for Controllable Talking Avatars (viability: 8): https://sciencetostartup.com/paper/making-avatars-interact-towards-text-driven-human-object-interaction-for-controllable-talking-avatars - Create controllable talking avatars that interact with objects through text-driven animations.
- You Need an Encoder for Native Position-Independent Caching (viability: 8): https://sciencetostartup.com/paper/you-need-an-encoder-for-native-position-independent-caching - COMB offers a position-independent caching plugin to drastically enhance LLM performance and efficiency.
- Qrita: High-performance Top-k and Top-p Algorithm for GPUs using Pivot-based Truncation and Selection (viability: 6): https://sciencetostartup.com/paper/qrita-high-performance-top-k-and-top-p-algorithm-for-gpus-using-pivot-based-truncation-and-selection - Qrita offers an efficient, GPU-optimized algorithm for Top-k and Top-p truncation in LLMs by significantly reducing computation and memory requirements.
- White-Box Neural Ensemble for Vehicular Plasticity: Quantifying the Efficiency Cost of Symbolic Auditability in Adaptive NMPC (viability: 5): https://sciencetostartup.com/paper/white-box-neural-ensemble-for-vehicular-plasticity-quantifying-the-efficiency-cost-of-symbolic-auditability-in-adaptive- - White-box adaptive NMPC for vehicular plasticity introduces auditability through symbolic graphs, balancing adaptability and efficiency.
- Governance at the Edge of Architecture: Regulating NeuroAI and Neuromorphic Systems (viability: 2): https://sciencetostartup.com/paper/governance-at-the-edge-of-architecture-regulating-neuroai-and-neuromorphic-systems - Develop AI governance frameworks aligned with neuromorphic and brain-inspired computing systems.
- OpInf-LLM: Parametric PDE Solving with LLMs via Operator Inference (viability: 7): https://sciencetostartup.com/paper/opinf-llm-parametric-pde-solving-with-llms-via-operator-inference - OpInf-LLM leverages LLM and operator inference for efficient parametric PDE solving with low computational demands.
- Causal Preference Elicitation (viability: 5): https://sciencetostartup.com/paper/causal-preference-elicitation - A Bayesian framework for expert-in-the-loop causal discovery that enhances graph-based knowledge extraction.
- Rod Flow: A Continuous-Time Model for Gradient Descent at the Edge of Stability (viability: 3): https://sciencetostartup.com/paper/rod-flow-a-continuous-time-model-for-gradient-descent-at-the-edge-of-stability - Develop a new ODE approximation to improve understanding of gradient descent dynamics in non-convex landscapes.
- Learning to Guide Local Search for MPE Inference in Probabilistic Graphical Models (viability: 5): https://sciencetostartup.com/paper/learning-to-guide-local-search-for-mpe-inference-in-probabilistic-graphical-models - AI-powered local search enhancement for efficient repeated MPE inference in fixed graphical models.
- Legal Infrastructure for Transformative AI Governance (viability: 2): https://sciencetostartup.com/paper/legal-infrastructure-for-transformative-ai-governance - Creating legal frameworks for AI governance and regulatory innovation.
- P-EAGLE: Parallel-Drafting EAGLE with Scalable Training (viability: 5): https://sciencetostartup.com/paper/p-eagle-parallel-drafting-eagle-with-scalable-training - P-EAGLE enables faster parallel multi-token prediction for large language models with a novel scalable training framework.
- Semi-supervised CAPP Transformer Learning via Pseudo-labeling (viability: 5): https://sciencetostartup.com/paper/semi-supervised-capp-transformer-learning-via-pseudo-labeling - Transform limited industry datasets into manufacturing process plans with semi-supervised CAPP transformer models.
- From Pragmas to Partners: A Symbiotic Evolution of Agentic High-Level Synthesis (viability: 2): https://sciencetostartup.com/paper/from-pragmas-to-partners-a-symbiotic-evolution-of-agentic-high-level-synthesis - Enhancing agentic optimization with high-level synthesis for AI-driven hardware design.
- An Odd Estimator for Shapley Values (viability: 5): https://sciencetostartup.com/paper/an-odd-estimator-for-shapley-values - OddSHAP enhances machine learning interpretability with state-of-the-art Shapley value estimation via polynomial regression on the odd subspace.
- VideoGPA: Distilling Geometry Priors for 3D-Consistent Video Generation (viability: 3): https://sciencetostartup.com/paper/videogpa-distilling-geometry-priors-for-3d-consistent-video-generation - Enhancing 3D consistency in video generation through Video Geometric Preference Alignment for improved temporal stability.
- End-to-end Optimization of Belief and Policy Learning in Shared Autonomy Paradigms (viability: 6): https://sciencetostartup.com/paper/end-to-end-optimization-of-belief-and-policy-learning-in-shared-autonomy-paradigms - BRACE enhances shared autonomy by optimizing Bayesian intent inference and context-adaptive assistance for robotic control.
- IRL-DAL: Safe and Adaptive Trajectory Planning for Autonomous Driving via Energy-Guided Diffusion Models (viability: 7): https://sciencetostartup.com/paper/irl-dal-safe-and-adaptive-trajectory-planning-for-autonomous-driving-via-energy-guided-diffusion-models - IRL-DAL provides a safer and adaptive autonomous driving solution through energy-guided diffusion models, enhancing obstacle avoidance and lane adherence.
- TEON: Tensorized Orthonormalization Beyond Layer-Wise Muon for Large Language Model Pre-Training (viability: 3): https://sciencetostartup.com/paper/teon-tensorized-orthonormalization-beyond-layer-wise-muon-for-large-language-model-pre-training - TEON generalizes Muon for more effective pre-training of large language models through advanced tensor orthogonalization techniques.
- Agnostic Language Identification and Generation (viability: 1): https://sciencetostartup.com/paper/agnostic-language-identification-and-generation - Explore new methods for language identification and generation without typical realizability assumptions.
- Now You Hear Me: Audio Narrative Attacks Against Large Audio-Language Models (viability: 3): https://sciencetostartup.com/paper/now-you-hear-me-audio-narrative-attacks-against-large-audio-language-models - Developing security frameworks to protect audio-language models from narrative-style audio attacks.
- YuriiFormer: A Suite of Nesterov-Accelerated Transformers (viability: 5): https://sciencetostartup.com/paper/yuriiformer-a-suite-of-nesterov-accelerated-transformers - Develop a Nesterov-accelerated transformer architecture that enhances performance over existing models using optimization-theoretic insights.
- ShotFinder: Imagination-Driven Open-Domain Video Shot Retrieval via Web Search (viability: 4): https://sciencetostartup.com/paper/shotfinder-imagination-driven-open-domain-video-shot-retrieval-via-web-search - ShotFinder provides a benchmark and retrieval pipeline for open-domain video shot retrieval based on keyframe-oriented descriptions.
- Strongly Polynomial Time Complexity of Policy Iteration for $L_\infty$ Robust MDPs (viability: 3): https://sciencetostartup.com/paper/strongly-polynomial-time-complexity-of-policy-iteration-for-l-infty-robust-mdps - Develop a robust policy iteration algorithm for strongly-polynomial time complexity in $L_\infty$ robust MDPs.
- Scaling Multiagent Systems with Process Rewards (viability: 5): https://sciencetostartup.com/paper/scaling-multiagent-systems-with-process-rewards - Enhance multiagent systems with per-action process rewards for improved performance in complex tasks.
- Agile Reinforcement Learning through Separable Neural Architecture (viability: 6): https://sciencetostartup.com/paper/agile-reinforcement-learning-through-separable-neural-architecture - SPAN offers a resource-efficient alternative for reinforcement learning with significant improvements in sample efficiency and success rates.
- Med-Scout: Curing MLLMs' Geometric Blindness in Medical Perception via Geometry-Aware RL Post-Training (viability: 7): https://sciencetostartup.com/paper/med-scout-curing-mllms-geometric-blindness-in-medical-perception-via-geometry-aware-rl-post-training - A framework that enhances geometric perception in medical language models using reinforcement learning to reduce factual hallucinations in diagnoses.
- MonoScale: Scaling Multi-Agent System with Monotonic Improvement (viability: 3): https://sciencetostartup.com/paper/monoscale-scaling-multi-agent-system-with-monotonic-improvement - MonoScale enhances the scalability of multi-agent systems with a framework ensuring monotonic performance improvements when adding new agents.
- Learning to Execute Graph Algorithms Exactly with Graph Neural Networks (viability: 4): https://sciencetostartup.com/paper/learning-to-execute-graph-algorithms-exactly-with-graph-neural-networks - Harness GNNs to learn and execute graph algorithms with high accuracy in bounded contexts.
- High-quality generation of dynamic game content via small language models: A proof of concept (viability: 5): https://sciencetostartup.com/paper/high-quality-generation-of-dynamic-game-content-via-small-language-models-a-proof-of-concept - Develop a small language model tailored for dynamic game content generation to replace costly large language models.
- TSAQA: Time Series Analysis Question And Answering Benchmark (viability: 6): https://sciencetostartup.com/paper/tsaqa-time-series-analysis-question-and-answering-benchmark - Introducing TSAQA, a comprehensive benchmark to evaluate time series analysis capabilities across multiple tasks and domains.
- Make Anything Match Your Target: Universal Adversarial Perturbations against Closed-Source MLLMs via Multi-Crop Routed Meta Optimization (viability: 3): https://sciencetostartup.com/paper/make-anything-match-your-target-universal-adversarial-perturbations-against-closed-source-mllms-via-multi-crop-routed-me - Develop universal perturbations to improve adversarial attack success on closed-source multimodal models.
- Beyond Fixed Frames: Dynamic Character-Aligned Speech Tokenization (viability: 3): https://sciencetostartup.com/paper/beyond-fixed-frames-dynamic-character-aligned-speech-tokenization - Introduce variable-frame-rate tokenization in neural audio codecs for efficient speech processing.
- Probing the Trajectories of Reasoning Traces in Large Language Models (viability: 5): https://sciencetostartup.com/paper/probing-the-trajectories-of-reasoning-traces-in-large-language-models - Protocol to probe reasoning trajectory efficiency in language models enhancing safe deployment.
- SPICE: Submodular Penalized Information-Conflict Selection for Efficient Large Language Model Training (viability: 6): https://sciencetostartup.com/paper/spice-submodular-penalized-information-conflict-selection-for-efficient-large-language-model-training - Optimize Large Language Model training using SPICE for efficient data selection and training cost reduction.
- On Safer Reinforcement Learning Policies for Sedation and Analgesia in Intensive Care (viability: 3): https://sciencetostartup.com/paper/on-safer-reinforcement-learning-policies-for-sedation-and-analgesia-in-intensive-care - Develop a reinforcement learning tool for optimizing sedation and analgesia dosing in ICU with long-term outcome focus.
- Machine Learning for Energy-Performance-aware Scheduling (viability: 3): https://sciencetostartup.com/paper/machine-learning-for-energy-performance-aware-scheduling - Leverage Bayesian Optimization for energy-performance-aware scheduling in embedded systems.
- WiFiPenTester: Advancing Wireless Ethical Hacking with Governed GenAI (viability: 7): https://sciencetostartup.com/paper/wifipentester-advancing-wireless-ethical-hacking-with-governed-genai - AI-powered wireless penetration testing tool that enhances efficiency and safety through GenAI-driven decision support and governance.
- From Similarity to Vulnerability: Key Collision Attack on LLM Semantic Caching (viability: 6): https://sciencetostartup.com/paper/from-similarity-to-vulnerability-key-collision-attack-on-llm-semantic-caching - Automated framework for exploiting semantic cache key collisions in LLM applications for security testing.
- Chain-of-thought obfuscation learned from output supervision can generalise to unseen tasks (viability: 2): https://sciencetostartup.com/paper/chain-of-thought-obfuscation-learned-from-output-supervision-can-generalise-to-unseen-tasks - Explores the generalization of reasoning obfuscation in language models, highlighting risks in penalizing output actions.
- OrLog: Resolving Complex Queries with LLMs and Probabilistic Reasoning (viability: 6): https://sciencetostartup.com/paper/orlog-resolving-complex-queries-with-llms-and-probabilistic-reasoning - OrLog: A neuro-symbolic framework enhancing complex query resolution with efficient logic-aware retrieval.
- Character as a Latent Variable in Large Language Models: A Mechanistic Account of Emergent Misalignment and Conditional Safety Failures (viability: 2): https://sciencetostartup.com/paper/character-as-a-latent-variable-in-large-language-models-a-mechanistic-account-of-emergent-misalignment-and-conditional-s - Exploring character formation in LLMs as a key factor in emergent misalignment and safety risks.
- ExplainerPFN: Towards tabular foundation models for model-free zero-shot feature importance estimations (viability: 5): https://sciencetostartup.com/paper/explainerpfn-towards-tabular-foundation-models-for-model-free-zero-shot-feature-importance-estimations - ExplainerPFN provides model-free zero-shot feature importance estimations for tabular data using pretrained synthetic dataset models.
- Towards Explicit Acoustic Evidence Perception in Audio LLMs for Speech Deepfake Detection (viability: 3): https://sciencetostartup.com/paper/towards-explicit-acoustic-evidence-perception-in-audio-llms-for-speech-deepfake-detection - Develop an acoustically enhanced framework to improve speech deepfake detection by exposing fine-grained time-frequency evidence.
- HierLoc: Hyperbolic Entity Embeddings for Hierarchical Visual Geolocation (viability: 7): https://sciencetostartup.com/paper/hierloc-hyperbolic-entity-embeddings-for-hierarchical-visual-geolocation - Develop a scalable geolocation tool using hyperbolic embeddings that outperform current methods by reducing error rates and improving accuracy.
- On the Impact of Code Comments for Automated Bug-Fixing: An Empirical Study (viability: 5): https://sciencetostartup.com/paper/on-the-impact-of-code-comments-for-automated-bug-fixing-an-empirical-study - Leverage code comments to enhance automated bug-fixing accuracy in large language models.
- Adaptive Edge Learning for Density-Aware Graph Generation (viability: 7): https://sciencetostartup.com/paper/adaptive-edge-learning-for-density-aware-graph-generation - A framework for generating realistic graph-structured data using adaptive edge learning and density-aware mechanisms.
- MedMCP-Calc: Benchmarking LLMs for Realistic Medical Calculator Scenarios via MCP Integration (viability: 7): https://sciencetostartup.com/paper/medmcp-calc-benchmarking-llms-for-realistic-medical-calculator-scenarios-via-mcp-integration - MedMCP-Calc benchmarks and improves LLMs for complex medical calculator workflows.
- From Abstract to Contextual: What LLMs Still Cannot Do in Mathematics (viability: 4): https://sciencetostartup.com/paper/from-abstract-to-contextual-what-llms-still-cannot-do-in-mathematics - ContextMATH identifies LLM limitations in contextual math reasoning, offering a benchmark for improving mathematical comprehension and formulation capabilities of AI.
- The Hot Mess of AI: How Does Misalignment Scale With Model Intelligence and Task Complexity? (viability: 2): https://sciencetostartup.com/paper/the-hot-mess-of-ai-how-does-misalignment-scale-with-model-intelligence-and-task-complexity - Investigating how AI model scale affects misalignment and incoherence in task performance.
- Avoiding Premature Collapse: Adaptive Annealing for Entropy-Regularized Structural Inference (viability: 6): https://sciencetostartup.com/paper/avoiding-premature-collapse-adaptive-annealing-for-entropy-regularized-structural-inference - Adaptive annealing algorithm for stable structural inference in machine learning models.
- Guided by Trajectories: Repairing and Rewarding Tool-Use Trajectories for Tool-Integrated Reasoning (viability: 6): https://sciencetostartup.com/paper/guided-by-trajectories-repairing-and-rewarding-tool-use-trajectories-for-tool-integrated-reasoning - AutoTraj enhances tool-integrated reasoning in LLMs by automatically repairing and rewarding tool-use trajectories for better task-solving efficiency.
- Leveraging Convolutional Sparse Autoencoders for Robust Movement Classification from Low-Density sEMG (viability: 6): https://sciencetostartup.com/paper/leveraging-convolutional-sparse-autoencoders-for-robust-movement-classification-from-low-density-semg - Deep learning framework using sparse autoencoders enables robust gesture recognition for adaptive prosthetic systems with minimal sensors.
- Automatic Constraint Policy Optimization based on Continuous Constraint Interpolation Framework for Offline Reinforcement Learning (viability: 5): https://sciencetostartup.com/paper/automatic-constraint-policy-optimization-based-on-continuous-constraint-interpolation-framework-for-offline-reinforcemen - Developing a unified offline reinforcement learning framework for optimal policy constraints with adaptable interpolation.
- Bias Beyond Borders: Political Ideology Evaluation and Steering in Multilingual LLMs (viability: 5): https://sciencetostartup.com/paper/bias-beyond-borders-political-ideology-evaluation-and-steering-in-multilingual-llms - Develop a framework to evaluate and mitigate political bias in multilingual LLMs, ensuring cross-lingual consistency and fairness.
- Mano: Restriking Manifold Optimization for LLM Training (viability: 5): https://sciencetostartup.com/paper/mano-restriking-manifold-optimization-for-llm-training - Mano introduces a novel optimizer for LLMs that outperforms AdamW and Muon by reducing computational complexity and memory usage, expanding the Pareto frontier for LLM training efficiency.
- TriCEGAR: A Trace-Driven Abstraction Mechanism for Agentic AI (viability: 3): https://sciencetostartup.com/paper/tricegar-a-trace-driven-abstraction-mechanism-for-agentic-ai - TriCEGAR automates agent state abstraction from execution logs to improve runtime verification.
- Self-Supervised Slice-to-Volume Reconstruction with Gaussian Representations for Fetal MRI (viability: 5): https://sciencetostartup.com/paper/self-supervised-slice-to-volume-reconstruction-with-gaussian-representations-for-fetal-mri - Develop a self-supervised tool for efficient 3D fetal MRI reconstruction using Gaussian representations.
- Why Your Deep Research Agent Fails? On Hallucination Evaluation in Full Research Trajectory (viability: 5): https://sciencetostartup.com/paper/why-your-deep-research-agent-fails-on-hallucination-evaluation-in-full-research-trajectory - Process-aware evaluation tool for diagnosing and improving Deep Research Agents by identifying hallucination types in research trajectories.
- About an Automating Annotation Method for Robot Markers (viability: 6): https://sciencetostartup.com/paper/about-an-automating-annotation-method-for-robot-markers - Automate the annotation process for training robot marker recognition models, enhancing robustness and reducing manual labor.
- Golden Goose: A Simple Trick to Synthesize Unlimited RLVR Tasks from Unverifiable Internet Text (viability: 7): https://sciencetostartup.com/paper/golden-goose-a-simple-trick-to-synthesize-unlimited-rlvr-tasks-from-unverifiable-internet-text - Golden Goose synthesizes RLVR tasks from unverifiable internet text, enabling robust gains in diverse domains including cybersecurity.
- Stabilizing the Q-Gradient Field for Policy Smoothness in Actor-Critic (viability: 5): https://sciencetostartup.com/paper/stabilizing-the-q-gradient-field-for-policy-smoothness-in-actor-critic - Develop a tool to stabilize the Q-gradient field, enhancing policy smoothness in actor-critic methods for robust task performance.
- EvoClinician: A Self-Evolving Agent for Multi-Turn Medical Diagnosis via Test-Time Evolutionary Learning (viability: 5): https://sciencetostartup.com/paper/evoclinician-a-self-evolving-agent-for-multi-turn-medical-diagnosis-via-test-time-evolutionary-learning - EvoClinician is a self-evolving AI agent for iterative medical diagnosis that optimizes strategies using test-time evolutionary learning.
- Residual Context Diffusion Language Models (viability: 6): https://sciencetostartup.com/paper/residual-context-diffusion-language-models - Residual Context Diffusion allows converting standard diffusion LLMs to more efficient models with improved accuracy by using discarded token contexts.
- From Data Leak to Secret Misses: The Impact of Data Leakage on Secret Detection Models (viability: 5): https://sciencetostartup.com/paper/from-data-leak-to-secret-misses-the-impact-of-data-leakage-on-secret-detection-models - A tool to identify and mitigate data leakage in AI-based security models, ensuring accurate performance evaluation.
- A Real-Time Privacy-Preserving Behavior Recognition System via Edge-Cloud Collaboration (viability: 6): https://sciencetostartup.com/paper/a-real-time-privacy-preserving-behavior-recognition-system-via-edge-cloud-collaboration - Develop a privacy-preserving behavior recognition system for sensitive environments using edge-cloud collaboration.
- Protecting Private Code in IDE Autocomplete using Differential Privacy (viability: 7): https://sciencetostartup.com/paper/protecting-private-code-in-ide-autocomplete-using-differential-privacy - A differential privacy-enhanced IDE autocomplete feature that protects code with minimal performance trade-offs.
- BEAR: Towards Beam-Search-Aware Optimization for Recommendation with Large Language Models (viability: 6): https://sciencetostartup.com/paper/bear-towards-beam-search-aware-optimization-for-recommendation-with-large-language-models - BEAR optimizes LLM recommendations by aligning training with beam search inference to improve top-ranked item retrieval.
- Evaluating Large Language Models for Security Bug Report Prediction (viability: 7): https://sciencetostartup.com/paper/evaluating-large-language-models-for-security-bug-report-prediction - Utilize fine-tuned Large Language Models for fast and precise security bug report predictions in software development.
- DINO-SAE: DINO Spherical Autoencoder for High-Fidelity Image Reconstruction and Generation (viability: 8): https://sciencetostartup.com/paper/dino-sae-dino-spherical-autoencoder-for-high-fidelity-image-reconstruction-and-generation - DINO-SAE is a high-fidelity image reconstruction and generation tool using hyperspherical model alignment.
- MulFeRL: Enhancing Reinforcement Learning with Verbal Feedback in a Multi-turn Loop (viability: 3): https://sciencetostartup.com/paper/mulferl-enhancing-reinforcement-learning-with-verbal-feedback-in-a-multi-turn-loop - Enhancing RL with verbal feedback for better reasoning in failed tasks.
- Game-Theoretic Co-Evolution for LLM-Based Heuristic Discovery (viability: 3): https://sciencetostartup.com/paper/game-theoretic-co-evolution-for-llm-based-heuristic-discovery - ASRO optimizes heuristic discovery with a game-theoretic co-evolution framework for LLMs.
- MoVE: Mixture of Value Embeddings -- A New Axis for Scaling Parametric Memory in Autoregressive Models (viability: 3): https://sciencetostartup.com/paper/move-mixture-of-value-embeddings-a-new-axis-for-scaling-parametric-memory-in-autoregressive-models - MoVE introduces a scalable memory mechanism for autoregressive models, decoupling memory size from compute cost.
- Eroding the Truth-Default: A Causal Analysis of Human Susceptibility to Foundation Model Hallucinations and Disinformation in the Wild (viability: 3): https://sciencetostartup.com/paper/eroding-the-truth-default-a-causal-analysis-of-human-susceptibility-to-foundation-model-hallucinations-and-disinformatio - JudgeGPT and RogueGPT offer a framework to analyze human susceptibility to AI-generated disinformation, guiding interventions for enhancing trustworthy web intelligence.
- MEnvAgent: Scalable Polyglot Environment Construction for Verifiable Software Engineering (viability: 8): https://sciencetostartup.com/paper/menvagent-scalable-polyglot-environment-construction-for-verifiable-software-engineering - MEnvAgent automates the creation of scalable, verifiable software engineering environments across multiple languages with a significant increase in efficiency and is backed by published benchmarks and datasets.
- Learning to Build Shapes by Extrusion (viability: 5): https://sciencetostartup.com/paper/learning-to-build-shapes-by-extrusion - Text Encoded Extrusion (TEE) offers a novel method for 3D mesh construction using LLM-driven face extrusion sequences.
- Just-in-Time Catching Test Generation at Meta (viability: 3): https://sciencetostartup.com/paper/just-in-time-catching-test-generation-at-meta - A novel just-in-time test generation approach to prevent bugs in massive backend systems by Meta.
- Offline Reinforcement Learning of High-Quality Behaviors Under Robust Style Alignment (viability: 7): https://sciencetostartup.com/paper/offline-reinforcement-learning-of-high-quality-behaviors-under-robust-style-alignment - Develop AI behaviors with robust style alignment using offline reinforcement learning and SCIQL framework.
- User-Adaptive Meta-Learning for Cold-Start Medication Recommendation with Uncertainty Filtering (viability: 7): https://sciencetostartup.com/paper/user-adaptive-meta-learning-for-cold-start-medication-recommendation-with-uncertainty-filtering - MetaDrug is a meta-learning framework for personalized medication recommendation addressing the cold-start problem in new patients using EHR data.
- Hide and Seek in Embedding Space: Geometry-based Steganography and Detection in Large Language Models (viability: 5): https://sciencetostartup.com/paper/hide-and-seek-in-embedding-space-geometry-based-steganography-and-detection-in-large-language-models - Develop a security tool leveraging mechanistic interpretability to detect steganographic threats in fine-tuned language models.
- Aligning the Unseen in Attributed Graphs: Interplay between Graph Geometry and Node Attributes Manifold (viability: 2): https://sciencetostartup.com/paper/aligning-the-unseen-in-attributed-graphs-interplay-between-graph-geometry-and-node-attributes-manifold - A novel variational autoencoder framework for uncovering hidden connectivity patterns in attributed graphs.
- CVeDRL: An Efficient Code Verifier via Difficulty-aware Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/cvedrl-an-efficient-code-verifier-via-difficulty-aware-reinforcement-learning - Efficiently verify code through difficulty-aware reinforcement learning for enhanced unit testing.
- Toward IIT-Inspired Consciousness in LLMs: A Reward-Based Learning Framework (viability: 5): https://sciencetostartup.com/paper/toward-iit-inspired-consciousness-in-llms-a-reward-based-learning-framework - Develop a reward-based learning framework for language models inspired by Integrated Information Theory to improve text conciseness while preserving accuracy.
- Learning with Challenges: Adaptive Difficulty-Aware Data Generation for Mobile GUI Agent Training (viability: 7): https://sciencetostartup.com/paper/learning-with-challenges-adaptive-difficulty-aware-data-generation-for-mobile-gui-agent-training - MobileGen offers adaptive difficulty-aware data generation to enhance mobile GUI agent training efficiency.
- TSPO: Breaking the Double Homogenization Dilemma in Multi-turn Search Policy Optimization (viability: 5): https://sciencetostartup.com/paper/tspo-breaking-the-double-homogenization-dilemma-in-multi-turn-search-policy-optimization - TSPO improves multi-turn reasoning efficiency in LLMs with a novel reward mechanism for better policy optimization.
- Beyond Abstract Compliance: Operationalising trust in AI as a moral relationship (viability: 2): https://sciencetostartup.com/paper/beyond-abstract-compliance-operationalising-trust-in-ai-as-a-moral-relationship - Explore operationalizing trust in AI through African relational ethics to improve community engagement and equity.
- How Far Can Pretrained LLMs Go in Symbolic Music? Controlled Comparisons of Supervised and Preference-based Adaptation (viability: 2): https://sciencetostartup.com/paper/how-far-can-pretrained-llms-go-in-symbolic-music-controlled-comparisons-of-supervised-and-preference-based-adaptation - Exploring the adaptation of pretrained language models for symbolic music generation and understanding.
- Qualitative Evaluation of LLM-Designed GUI (viability: 2): https://sciencetostartup.com/paper/qualitative-evaluation-of-llm-designed-gui - Explore LLMs for early-stage UI prototyping but require human refinement for full usability.
- AutoRefine: From Trajectories to Reusable Expertise for Continual LLM Agent Refinement (viability: 7): https://sciencetostartup.com/paper/autorefine-from-trajectories-to-reusable-expertise-for-continual-llm-agent-refinement - AutoRefine enhances LLM agents by transforming execution histories into procedural and static knowledge, significantly improving efficiency and coordination.
- Procedural Knowledge Extraction from Industrial Troubleshooting Guides Using Vision Language Models (viability: 2): https://sciencetostartup.com/paper/procedural-knowledge-extraction-from-industrial-troubleshooting-guides-using-vision-language-models - Automate the extraction of procedural knowledge from industrial troubleshooting guides using vision-language models.
- UrbanMoE: A Sparse Multi-Modal Mixture-of-Experts Framework for Multi-Task Urban Region Profiling (viability: 5): https://sciencetostartup.com/paper/urbanmoe-a-sparse-multi-modal-mixture-of-experts-framework-for-multi-task-urban-region-profiling - Build a sparse Mixture-of-Experts framework to profile urban regions with superior multi-task prediction capabilities.
- ImgCoT: Compressing Long Chain of Thought into Compact Visual Tokens for Efficient Reasoning of Large Language Model (viability: 6): https://sciencetostartup.com/paper/imgcot-compressing-long-chain-of-thought-into-compact-visual-tokens-for-efficient-reasoning-of-large-language-model - ImgCoT compresses long chains of thought into compact visual tokens for efficient reasoning in large language models.
- OpenVTON-Bench: A Large-Scale High-Resolution Benchmark for Controllable Virtual Try-On Evaluation (viability: 6): https://sciencetostartup.com/paper/openvton-bench-a-large-scale-high-resolution-benchmark-for-controllable-virtual-try-on-evaluation - OpenVTON-Bench offers a robust high-resolution benchmark for evaluating virtual try-on systems across multiple quality dimensions.
- AEGIS: White-Box Attack Path Generation using LLMs and Training Effectiveness Evaluation for Large-Scale Cyber Defence Exercises (viability: 6): https://sciencetostartup.com/paper/aegis-white-box-attack-path-generation-using-llms-and-training-effectiveness-evaluation-for-large-scale-cyber-defence-ex - AEGIS automates cyber defense exercise planning by dynamically generating attack paths using LLMs, reducing preparation time from months to days.
- A Step Back: Prefix Importance Ratio Stabilizes Policy Optimization (viability: 6): https://sciencetostartup.com/paper/a-step-back-prefix-importance-ratio-stabilizes-policy-optimization - MinPRO stabilizes off-policy reinforcement learning for LLMs, improving training stability and performance.
- Breaking the Blocks: Continuous Low-Rank Decomposed Scaling for Unified LLM Quantization and Adaptation (viability: 8): https://sciencetostartup.com/paper/breaking-the-blocks-continuous-low-rank-decomposed-scaling-for-unified-llm-quantization-and-adaptation - Unified LLM quantization and adaptation framework that significantly improves performance and efficiency.
- Vision-Language Models Unlock Task-Centric Latent Actions (viability: 5): https://sciencetostartup.com/paper/vision-language-models-unlock-task-centric-latent-actions - Enhance task-specific action recognition in videos using Vision-Language Models that filter out distractors effectively.
- Gated Relational Alignment via Confidence-based Distillation for Efficient VLMs (viability: 5): https://sciencetostartup.com/paper/gated-relational-alignment-via-confidence-based-distillation-for-efficient-vlms - GRACE enables efficient Vision-Language Models with enhanced performance through innovative quantization and distillation techniques.
- Deep Learning-Based Early-Stage IR-Drop Estimation via CNN Surrogate Modeling (viability: 8): https://sciencetostartup.com/paper/deep-learning-based-early-stage-ir-drop-estimation-via-cnn-surrogate-modeling - Deep learning model for actionable early-stage IR-drop estimation in VLSI design.
- Best-of-Q: Improving VLM agents with Q-function Action Ranking at Inference (viability: 7): https://sciencetostartup.com/paper/best-of-q-improving-vlm-agents-with-q-function-action-ranking-at-inference - Enhance existing VLM-based agent policies at inference without retraining through Q-function action ranking.
- PEAR: Pixel-aligned Expressive humAn mesh Recovery (viability: 8): https://sciencetostartup.com/paper/pear-pixel-aligned-expressive-human-mesh-recovery - PEAR offers real-time, pixel-level accurate 3D human mesh recovery for immersive applications using a ViT-based streamlined model.
- FNF: Functional Network Fingerprint for Large Language Models (viability: 7): https://sciencetostartup.com/paper/fnf-functional-network-fingerprint-for-large-language-models - FNF offers a robust intellectual property protection tool for LLM developers by identifying shared origins in neural network activities.
- Do Transformers Have the Ability for Periodicity Generalization? (viability: 5): https://sciencetostartup.com/paper/do-transformers-have-the-ability-for-periodicity-generalization - Develop a Transformer model benchmark to evaluate and improve periodicity generalization in out-of-distribution scenarios.
- Fire on Motion: Optimizing Video Pass-bands for Efficient Spiking Action Recognition (viability: 7): https://sciencetostartup.com/paper/fire-on-motion-optimizing-video-pass-bands-for-efficient-spiking-action-recognition - Pass-Bands Optimizer enhances spiking neural networks for video tasks by refining temporal dynamics for motion content.
- From Horizontal Layering to Vertical Integration: A Comparative Study of the AI-Driven Software Development Paradigm (viability: 2): https://sciencetostartup.com/paper/from-horizontal-layering-to-vertical-integration-a-comparative-study-of-the-ai-driven-software-development-paradigm - A study on the organizational benefits and challenges of adopting AI in software engineering.
- Real-Time Aligned Reward Model beyond Semantics (viability: 3): https://sciencetostartup.com/paper/real-time-aligned-reward-model-beyond-semantics - Develop a reinforcement learning framework that dynamically aligns reward models with policy model feedback for more accurate human preference capture.
- Unsupervised Synthetic Image Attribution: Alignment and Disentanglement (viability: 6): https://sciencetostartup.com/paper/unsupervised-synthetic-image-attribution-alignment-and-disentanglement - An unsupervised tool for synthetic image attribution to protect copyrights and ensure model transparency.
- Task-Aware LLM Council with Adaptive Decision Pathways for Decision Support (viability: 6): https://sciencetostartup.com/paper/task-aware-llm-council-with-adaptive-decision-pathways-for-decision-support - Task-Aware LLM Council integrates specialized LLMs for adaptive decision-making with superior task success rates.
- Human-Centered Explainability in AI-Enhanced UI Security Interfaces: Designing Trustworthy Copilots for Cybersecurity Analysts (viability: 4): https://sciencetostartup.com/paper/human-centered-explainability-in-ai-enhanced-ui-security-interfaces-designing-trustworthy-copilots-for-cybersecurity-ana - Design AI-centric cybersecurity dashboards that enhance trust through explainable interfaces.
- GUDA: Counterfactual Group-wise Training Data Attribution for Diffusion Models via Unlearning (viability: 3): https://sciencetostartup.com/paper/guda-counterfactual-group-wise-training-data-attribution-for-diffusion-models-via-unlearning - GUDA offers efficient group-wise training data attribution for diffusion models using machine unlearning.
- UCPO: Uncertainty-Aware Policy Optimization (viability: 5): https://sciencetostartup.com/paper/ucpo-uncertainty-aware-policy-optimization - Develop dynamic uncertainty-aware policy optimization frameworks to enhance the trustworthiness of large language models by mitigating hallucinations.
- Statistical Estimation of Adversarial Risk in Large Language Models under Best-of-N Sampling (viability: 5): https://sciencetostartup.com/paper/statistical-estimation-of-adversarial-risk-in-large-language-models-under-best-of-n-sampling - A scalable methodology for assessing real-world risk in LLMs under adversarial conditions, with code availability promising research integration.
- What can Computer Vision learn from Ranganathan? (viability: 4): https://sciencetostartup.com/paper/what-can-computer-vision-learn-from-ranganathan - New CV dataset design methodology inspired by Ranganathan could enhance annotation accuracy and dataset quality.
- MCP-Diag: A Deterministic, Protocol-Driven Architecture for AI-Native Network Diagnostics (viability: 5): https://sciencetostartup.com/paper/mcp-diag-a-deterministic-protocol-driven-architecture-for-ai-native-network-diagnostics - MCP-Diag offers a secure, deterministic solution for integrating AI in network diagnostics by converting CLI outputs into structured data for analysis.
- PEFT-MuTS: A Multivariate Parameter-Efficient Fine-Tuning Framework for Remaining Useful Life Prediction based on Cross-domain Time Series Representation Model (viability: 6): https://sciencetostartup.com/paper/peft-muts-a-multivariate-parameter-efficient-fine-tuning-framework-for-remaining-useful-life-prediction-based-on-cross-d - Develop a tool for efficient RUL prediction using few-shot learning and cross-domain time series data.
- Time-Annealed Perturbation Sampling: Diverse Generation for Diffusion Language Models (viability: 5): https://sciencetostartup.com/paper/time-annealed-perturbation-sampling-diverse-generation-for-diffusion-language-models - Innovative diffusion-based language model for diverse text generation.
- TTCS: Test-Time Curriculum Synthesis for Self-Evolving (viability: 4): https://sciencetostartup.com/paper/ttcs-test-time-curriculum-synthesis-for-self-evolving - A framework that enhances LLM reasoning by co-evolving a curriculum through test-time training.
- SYMPHONY: Synergistic Multi-agent Planning with Heterogeneous Language Model Assembly (viability: 7): https://sciencetostartup.com/paper/symphony-synergistic-multi-agent-planning-with-heterogeneous-language-model-assembly - SYMPHONY enhances autonomous agent planning by integrating heterogeneous language models for more diverse and effective problem-solving.
- EntroCut: Entropy-Guided Adaptive Truncation for Efficient Chain-of-Thought Reasoning in Small-scale Large Reasoning Models (viability: 4): https://sciencetostartup.com/paper/entrocut-entropy-guided-adaptive-truncation-for-efficient-chain-of-thought-reasoning-in-small-scale-large-reasoning-mode - EntroCut enhances computational efficiency in chain-of-thought reasoning by dynamically truncating low-confidence steps based on entropy.
- Local-Global Multimodal Contrastive Learning for Molecular Property Prediction (viability: 5): https://sciencetostartup.com/paper/local-global-multimodal-contrastive-learning-for-molecular-property-prediction - Unified framework for enhancing molecular property predictions via multimodal contrastive learning.
- From Self-Evolving Synthetic Data to Verifiable-Reward RL: Post-Training Multi-turn Interactive Tool-Using Agents (viability: 3): https://sciencetostartup.com/paper/from-self-evolving-synthetic-data-to-verifiable-reward-rl-post-training-multi-turn-interactive-tool-using-agents - Develop an RL system for interactive agents using self-evolving synthetic data and verifier-based strategies for scalable tool-use training.
- Language Model Circuits Are Sparse in the Neuron Basis (viability: 5): https://sciencetostartup.com/paper/language-model-circuits-are-sparse-in-the-neuron-basis - Develop end-to-end pipeline for circuit tracing on sparse MLP neuron basis, enhancing model interpretability without extra training costs.
- Rethinking LLM-as-a-Judge: Representation-as-a-Judge with Small Language Models via Semantic Capacity Asymmetry (viability: 5): https://sciencetostartup.com/paper/rethinking-llm-as-a-judge-representation-as-a-judge-with-small-language-models-via-semantic-capacity-asymmetry - INSPECTOR uses small language model representations for efficient, interpretable evaluation, challenging the dominance of LLMs in evaluative tasks.
- WED-Net: A Weather-Effect Disentanglement Network with Causal Augmentation for Urban Flow Prediction (viability: 7): https://sciencetostartup.com/paper/wed-net-a-weather-effect-disentanglement-network-with-causal-augmentation-for-urban-flow-prediction - A robust urban flow prediction tool using weather-effect disentanglement to improve mobility and disaster resilience under extreme conditions.
- MC-GRPO: Median-Centered Group Relative Policy Optimization for Small-Rollout Reinforcement Learning (viability: 6): https://sciencetostartup.com/paper/mc-grpo-median-centered-group-relative-policy-optimization-for-small-rollout-reinforcement-learning - MC-GRPO optimizes small-rollout reinforcement learning by using a median-centered approach to improve training stability and accuracy.
- SpanNorm: Reconciling Training Stability and Performance in Deep Transformers (viability: 5): https://sciencetostartup.com/paper/spannorm-reconciling-training-stability-and-performance-in-deep-transformers - SpanNorm enhances training stability and performance of deep Transformers by integrating PreNorm and PostNorm benefits.
- FedDis: A Causal Disentanglement Framework for Federated Traffic Prediction (viability: 5): https://sciencetostartup.com/paper/feddis-a-causal-disentanglement-framework-for-federated-traffic-prediction - FedDis uses causal disentanglement to improve federated traffic prediction under non-IID data conditions.
- Mitigating Hallucinations in Video Large Language Models via Spatiotemporal-Semantic Contrastive Decoding (viability: 6): https://sciencetostartup.com/paper/mitigating-hallucinations-in-video-large-language-models-via-spatiotemporal-semantic-contrastive-decoding - Spatiotemporal-Semantic Contrastive Decoding reduces hallucinations in video models, enhancing video comprehension and reliability.
- PerfGuard: A Performance-Aware Agent for Visual Content Generation (viability: 7): https://sciencetostartup.com/paper/perfguard-a-performance-aware-agent-for-visual-content-generation - PerfGuard enhances visual content generation by incorporating performance-aware technology for precise tool selection and task management.
- Whispers of Wealth: Red-Teaming Google's Agent Payments Protocol via Prompt Injection (viability: 3): https://sciencetostartup.com/paper/whispers-of-wealth-red-teaming-google-s-agent-payments-protocol-via-prompt-injection - Analyzes vulnerabilities in Google's Agent Payments Protocol using prompt injection techniques.
- EUGens: Efficient, Unified, and General Dense Layers (viability: 5): https://sciencetostartup.com/paper/eugens-efficient-unified-and-general-dense-layers - Develop EUGens, efficient dense layers to enhance neural networks for real-time applications and resource-constrained environments.
- Decoding in Geometry: Alleviating Embedding-Space Crowding for Complex Reasoning (viability: 6): https://sciencetostartup.com/paper/decoding-in-geometry-alleviating-embedding-space-crowding-for-complex-reasoning - CraEG offers a training-free, geometry-guided sampling method to enhance performance in language models by alleviating embedding-space crowding.
- Demystifying Design Choices of Reinforcement Fine-tuning: A Batched Contextual Bandit Learning Perspective (viability: 5): https://sciencetostartup.com/paper/demystifying-design-choices-of-reinforcement-fine-tuning-a-batched-contextual-bandit-learning-perspective - Develop an experimental pipeline to evaluate the impact of design choices in reinforcement fine-tuning using a batched contextual bandit learning approach.
- Learn from A Rationalist: Distilling Intermediate Interpretable Rationales (viability: 5): https://sciencetostartup.com/paper/learn-from-a-rationalist-distilling-intermediate-interpretable-rationales - REKD enhances student DNN interpretability using rationale extraction with knowledge distillation for improved AI decision-making.
- Enhancing TableQA through Verifiable Reasoning Trace Reward (viability: 8): https://sciencetostartup.com/paper/enhancing-tableqa-through-verifiable-reasoning-trace-reward - RE-Tab enhances TableQA with a plug-and-play framework that boosts model reasoning using verifiable reward feedback, offering significant performance gains.
- Darwinian Memory: A Training-Free Self-Regulating Memory System for GUI Agent Evolution (viability: 8): https://sciencetostartup.com/paper/darwinian-memory-a-training-free-self-regulating-memory-system-for-gui-agent-evolution - Develop a training-free, self-evolving memory system for GUI automation that enhances MLLM agents' performance without added costs.
- Why Self-Rewarding Works: Theoretical Guarantees for Iterative Alignment of Language Models (viability: 2): https://sciencetostartup.com/paper/why-self-rewarding-works-theoretical-guarantees-for-iterative-alignment-of-language-models - Theoretical exploration of Self-Rewarding Language Models lacks direct product path.
- Shattered Compositionality: Counterintuitive Learning Dynamics of Transformers for Arithmetic (viability: 2): https://sciencetostartup.com/paper/shattered-compositionality-counterintuitive-learning-dynamics-of-transformers-for-arithmetic - Exploring non-human-like learning dynamics in Transformers for arithmetic applications, uncovering challenges in model alignment and reliability.
- Keep Rehearsing and Refining: Lifelong Learning Vehicle Routing under Continually Drifting Tasks (viability: 3): https://sciencetostartup.com/paper/keep-rehearsing-and-refining-lifelong-learning-vehicle-routing-under-continually-drifting-tasks - Develop a neural VRP solver that adapts to continually drifting task patterns with limited training resources.
- Action-Sufficient Goal Representations (viability: 4): https://sciencetostartup.com/paper/action-sufficient-goal-representations - Develop action-sufficient goal representations to improve GCRL hierarchical policies for better control in long-horizon tasks.
- AI Literacy, Safety Awareness, and STEM Career Aspirations of Australian Secondary Students: Evaluating the Impact of Workshop Interventions (viability: 3): https://sciencetostartup.com/paper/ai-literacy-safety-awareness-and-stem-career-aspirations-of-australian-secondary-students-evaluating-the-impact-of-works - AI literacy workshops enhance Australian students' understanding of AI and interest in STEM careers.
- FraudShield: Knowledge Graph Empowered Defense for LLMs against Fraud Attacks (viability: 3): https://sciencetostartup.com/paper/fraudshield-knowledge-graph-empowered-defense-for-llms-against-fraud-attacks - FraudShield enhances LLM security by leveraging a knowledge graph to defend against fraud tactics.
- RulePlanner: All-in-One Reinforcement Learner for Unifying Design Rules in 3D Floorplanning (viability: 8): https://sciencetostartup.com/paper/ruleplanner-all-in-one-reinforcement-learner-for-unifying-design-rules-in-3d-floorplanning - Revolutionizing IC floorplanning with an all-in-one deep reinforcement learning tool for 3D design rule unification.
- Training-Free Representation Guidance for Diffusion Models with a Representation Alignment Projector (viability: 6): https://sciencetostartup.com/paper/training-free-representation-guidance-for-diffusion-models-with-a-representation-alignment-projector - Revolutionize image synthesis with enhanced semantic coherence using diffusion models guided by a representation alignment projector.
- Machine Unlearning in Low-Dimensional Feature Subspace (viability: 7): https://sciencetostartup.com/paper/machine-unlearning-in-low-dimensional-feature-subspace - Implementing efficient machine unlearning with LOFT in low-dimensional feature subspaces to enhance privacy and maintain model performance.
- Tuning the Implicit Regularizer of Masked Diffusion Language Models: Enhancing Generalization via Insights from $k$-Parity (viability: 8): https://sciencetostartup.com/paper/tuning-the-implicit-regularizer-of-masked-diffusion-language-models-enhancing-generalization-via-insights-from-k-parity - Enhance generalization in language models by optimizing mask probabilities in masked diffusion frameworks.
- Controllable Information Production (viability: 3): https://sciencetostartup.com/paper/controllable-information-production - Explore intelligent behavior with intrinsic motivation using the novel Controllable Information Production framework.
- Automating Forecasting Question Generation and Resolution for AI Evaluation (viability: 7): https://sciencetostartup.com/paper/automating-forecasting-question-generation-and-resolution-for-ai-evaluation - Automated AI system for scalable forecasting question generation and resolution with high accuracy.
- AI and My Values: User Perceptions of LLMs' Ability to Extract, Embody, and Explain Human Values from Casual Conversations (viability: 2): https://sciencetostartup.com/paper/ai-and-my-values-user-perceptions-of-llms-ability-to-extract-embody-and-explain-human-values-from-casual-conversations - Toolkit for evaluating LLMs' ability to align with human values in conversations.
- CoDCL: Counterfactual Data Augmentation Contrastive Learning for Continuous-Time Dynamic Network Link Prediction (viability: 6): https://sciencetostartup.com/paper/codcl-counterfactual-data-augmentation-contrastive-learning-for-continuous-time-dynamic-network-link-prediction - CoDCL offers a plug-and-play module to enhance temporal graph model predictions using counterfactual data augmentation and contrastive learning.
- MetaLead: A Comprehensive Human-Curated Leaderboard Dataset for Transparent Reporting of Machine Learning Experiments (viability: 5): https://sciencetostartup.com/paper/metalead-a-comprehensive-human-curated-leaderboard-dataset-for-transparent-reporting-of-machine-learning-experiments - MetaLead offers a comprehensive, human-curated leaderboard dataset for transparent ML experiment reporting.
- AI-Enabled Waste Classification as a Data-Driven Decision Support Tool for Circular Economy and Urban Sustainability (viability: 5): https://sciencetostartup.com/paper/ai-enabled-waste-classification-as-a-data-driven-decision-support-tool-for-circular-economy-and-urban-sustainability - Develop an AI-powered waste sorting tool to enhance circular economy practices in smart cities.
- Semi-Autonomous Mathematics Discovery with Gemini: A Case Study on the Erdős Problems (viability: 3): https://sciencetostartup.com/paper/semi-autonomous-mathematics-discovery-with-gemini-a-case-study-on-the-erd-s-problems - A semi-autonomous tool for verifying mathematical conjectures in specialized databases like Bloom's Erdős Problems.
- Score-based Integrated Gradient for Root Cause Explanations of Outliers (viability: 5): https://sciencetostartup.com/paper/score-based-integrated-gradient-for-root-cause-explanations-of-outliers - SIREN empowers anomaly detection with accurate root cause analysis in complex data environments.
- Jailbreaks on Vision Language Model via Multimodal Reasoning (viability: 3): https://sciencetostartup.com/paper/jailbreaks-on-vision-language-model-via-multimodal-reasoning - Develop a framework that creates stealthy prompts to bypass safety filters in vision-language models using post-training CoT prompting and ReAct-driven adaptive noising.
- Culturally Grounded Personas in Large Language Models: Characterization and Alignment with Socio-Psychological Value Frameworks (viability: 5): https://sciencetostartup.com/paper/culturally-grounded-personas-in-large-language-models-characterization-and-alignment-with-socio-psychological-value-fram - Develop culturally-grounded personas in LLMs to enhance cross-cultural and moral understanding for AI applications.
- Graph is a Substrate Across Data Modalities (viability: 3): https://sciencetostartup.com/paper/graph-is-a-substrate-across-data-modalities - Develop a framework for persistent graph representations across diverse data modalities and tasks.
- Learning Provably Correct Distributed Protocols Without Human Knowledge (viability: 2): https://sciencetostartup.com/paper/learning-provably-correct-distributed-protocols-without-human-knowledge - GGMS is a framework enabling automated learning of robust distributed protocols without decades of human input.
- Context Structure Reshapes the Representational Geometry of Language Models (viability: 2): https://sciencetostartup.com/paper/context-structure-reshapes-the-representational-geometry-of-language-models - Explore how LLMs adjust to context with varying task structures, proposing a dynamic task strategy hypothesis.
- MERMAID: Memory-Enhanced Retrieval and Reasoning with Multi-Agent Iterative Knowledge Grounding for Veracity Assessment (viability: 7): https://sciencetostartup.com/paper/mermaid-memory-enhanced-retrieval-and-reasoning-with-multi-agent-iterative-knowledge-grounding-for-veracity-assessment - A memory-enhanced framework for efficient and consistent online content veracity assessment using multi-agent knowledge grounding.
- The Unseen Threat: Residual Knowledge in Machine Unlearning under Perturbed Samples (viability: 5): https://sciencetostartup.com/paper/the-unseen-threat-residual-knowledge-in-machine-unlearning-under-perturbed-samples - RURK technology enhances data privacy in machine learning by mitigating residual knowledge risks from adversarially perturbed samples.
- Recoverability Has a Law: The ERR Measure for Tool-Augmented Agents (viability: 4): https://sciencetostartup.com/paper/recoverability-has-a-law-the-err-measure-for-tool-augmented-agents - Explore the governed property of interaction dynamics for recoverability in tool-augmented language model agents via the ERR measure.
- Learning Policy Representations for Steerable Behavior Synthesis (viability: 3): https://sciencetostartup.com/paper/learning-policy-representations-for-steerable-behavior-synthesis - Develop a system to steer policies for Markov decision processes using learned latent representations without retraining.
- MixQuant: Pushing the Limits of Block Rotations in Post-Training Quantization (viability: 3): https://sciencetostartup.com/paper/mixquant-pushing-the-limits-of-block-rotations-in-post-training-quantization - MixQuant offers a novel post-training quantization framework enhancing block rotations to improve model accuracy without added inference overhead.
- From Retrieving Information to Reasoning with AI: Exploring Different Interaction Modalities to Support Human-AI Coordination in Clinical Decision-Making (viability: 3): https://sciencetostartup.com/paper/from-retrieving-information-to-reasoning-with-ai-exploring-different-interaction-modalities-to-support-human-ai-coordina - Study explores clinicians' perceptions of different interaction modalities with AI for decision-support to enhance clinical tools.
- Stealthy Poisoning Attacks Bypass Defenses in Regression Settings (viability: 1): https://sciencetostartup.com/paper/stealthy-poisoning-attacks-bypass-defenses-in-regression-settings - Advanced stealthy poisoning attack bypasses existing defenses in regression models with a new defense proposal, BayesClean.
- Conformal Prediction for Generative Models via Adaptive Cluster-Based Density Estimation (viability: 6): https://sciencetostartup.com/paper/conformal-prediction-for-generative-models-via-adaptive-cluster-based-density-estimation - Develop an uncertainty estimation tool for generative models using conformal prediction to enhance trust and interpretability in high-stakes applications.
- ParalESN: Enabling parallel information processing in Reservoir Computing (viability: 6): https://sciencetostartup.com/paper/paralesn-enabling-parallel-information-processing-in-reservoir-computing - ParalESN offers scalable, parallel temporal data processing for energy-efficient time series applications.
- PersonaCite: VoC-Grounded Interviewable Agentic Synthetic AI Personas for Verifiable User and Design Research (viability: 5): https://sciencetostartup.com/paper/personacite-voc-grounded-interviewable-agentic-synthetic-ai-personas-for-verifiable-user-and-design-research - PersonaCite provides evidence-grounded synthetic AI personas for reliable user research and design decision-making.
- VMonarch: Efficient Video Diffusion Transformers with Structured Attention (viability: 6): https://sciencetostartup.com/paper/vmonarch-efficient-video-diffusion-transformers-with-structured-attention - VMonarch delivers efficient video processing by optimizing attention mechanisms with structured matrices to enhance speed and generation quality.
- JAF: Judge Agent Forest (viability: 2): https://sciencetostartup.com/paper/jaf-judge-agent-forest - Introduce JAF: a framework for holistic evaluation and improvement of AI agent reasoning through joint inference and feedback.
- Predicting Intermittent Job Failure Categories for Diagnosis Using Few-Shot Fine-Tuned Language Models (viability: 7): https://sciencetostartup.com/paper/predicting-intermittent-job-failure-categories-for-diagnosis-using-few-shot-fine-tuned-language-models - Automate intermittent job failure diagnosis in CI pipelines using AI-powered log analysis tools.
- AI Narrative Breakdown. A Critical Assessment of Power and Promise (viability: 3): https://sciencetostartup.com/paper/ai-narrative-breakdown-a-critical-assessment-of-power-and-promise - A critical assessment of AI narratives exploring societal narratives and implications.
- MirrorMark: A Distortion-Free Multi-Bit Watermark for Large Language Models (viability: 2): https://sciencetostartup.com/paper/mirrormark-a-distortion-free-multi-bit-watermark-for-large-language-models - MirrorMark offers a robust, distortion-free watermarking solution for attributing content generated by large language models without degrading text quality.
- A Systematic Literature Review on LLM Defenses Against Prompt Injection and Jailbreaking: Expanding NIST Taxonomy (viability: 3): https://sciencetostartup.com/paper/a-systematic-literature-review-on-llm-defenses-against-prompt-injection-and-jailbreaking-expanding-nist-taxonomy - Defend against prompt injection and jailbreaking in LLMs with a comprehensive mitigation strategy catalog.
- Lost in Space? Vision-Language Models Struggle with Relative Camera Pose Estimation (viability: 5): https://sciencetostartup.com/paper/lost-in-space-vision-language-models-struggle-with-relative-camera-pose-estimation - Introducing VRRPI-Bench: A benchmark for testing Vision-Language Models' 3D spatial reasoning capabilities using egocentric video data.
- RedSage: A Cybersecurity Generalist LLM (viability: 8): https://sciencetostartup.com/paper/redsage-a-cybersecurity-generalist-llm - RedSage is an open-source cybersecurity assistant LLM with domain-aware capabilities, surpassing benchmarks and ensuring data privacy.
- Hybrid Linear Attention Done Right: Efficient Distillation and Effective Architectures for Extremely Long Contexts (viability: 5): https://sciencetostartup.com/paper/hybrid-linear-attention-done-right-efficient-distillation-and-effective-architectures-for-extremely-long-contexts - Develop HypeNet, a hybrid RNN-attention architecture, to enhance long-context performance and efficiency using minimal data.
- Exploring Reasoning Reward Model for Agents (viability: 9): https://sciencetostartup.com/paper/exploring-reasoning-reward-model-for-agents - A breakthrough reinforcement learning platform that enhances agent reasoning with multi-level feedback, improving performance in complex environments.
- DynaWeb: Model-Based Reinforcement Learning of Web Agents (viability: 5): https://sciencetostartup.com/paper/dynaweb-model-based-reinforcement-learning-of-web-agents - DynaWeb leverages model-based reinforcement learning to train web agents in simulated environments, enhancing the efficiency and scalability of autonomous internet navigation.
- Routing the Lottery: Adaptive Subnetworks for Heterogeneous Data (viability: 7): https://sciencetostartup.com/paper/routing-the-lottery-adaptive-subnetworks-for-heterogeneous-data - Adaptive subnetwork pruning framework for context-aware deep learning on heterogeneous data.
- Reasoning While Asking: Transforming Reasoning Large Language Models from Passive Solvers to Proactive Inquirers (viability: 7): https://sciencetostartup.com/paper/reasoning-while-asking-transforming-reasoning-large-language-models-from-passive-solvers-to-proactive-inquirers - Transform LLMs into proactive inquirers to enhance reasoning accuracy and efficiency.
- PRISM: Distribution-free Adaptive Computation of Matrix Functions for Accelerating Neural Network Training (viability: 3): https://sciencetostartup.com/paper/prism-distribution-free-adaptive-computation-of-matrix-functions-for-accelerating-neural-network-training - PRISM offers a framework for accelerating neural network training by optimizing matrix function computations without needing explicit spectral bounds.
- StepShield: When, Not Whether to Intervene on Rogue Agents (viability: 8): https://sciencetostartup.com/paper/stepshield-when-not-whether-to-intervene-on-rogue-agents - StepShield provides real-time safety benchmarking for AI agents, optimizing early interventions to reduce monitoring costs and enhance security.
- World of Workflows: a Benchmark for Bringing World Models to Enterprise Systems (viability: 7): https://sciencetostartup.com/paper/world-of-workflows-a-benchmark-for-bringing-world-models-to-enterprise-systems - Introducing a benchmark to enhance LLM reliability in enterprise systems by modeling complex business workflows.
- SWE-Replay: Efficient Test-Time Scaling for Software Engineering Agents (viability: 2): https://sciencetostartup.com/paper/swe-replay-efficient-test-time-scaling-for-software-engineering-agents - Optimize test-time scaling for software engineering agents with SWE-Replay to reduce computational costs and increase performance efficiency.
- The Patient is not a Moving Document: A World Model Training Paradigm for Longitudinal EHR (viability: 8): https://sciencetostartup.com/paper/the-patient-is-not-a-moving-document-a-world-model-training-paradigm-for-longitudinal-ehr - Develop a world model for EHR that simulates patient dynamics for improved disease trajectory predictions.
- Defining Operational Conditions for Safety-Critical AI-Based Systems from Data (viability: 3): https://sciencetostartup.com/paper/defining-operational-conditions-for-safety-critical-ai-based-systems-from-data - Develop a Safety-by-Design method to define Operational Design Domain (ODD) for AI certifications in safety-critical systems using data.
- SINA: A Circuit Schematic Image-to-Netlist Generator Using Artificial Intelligence (viability: 4): https://sciencetostartup.com/paper/sina-a-circuit-schematic-image-to-netlist-generator-using-artificial-intelligence - Transform circuit schematic images into accurate netlists with SINA, an AI-driven image-to-netlist generator.
- Value-Based Pre-Training with Downstream Feedback (viability: 2): https://sciencetostartup.com/paper/value-based-pre-training-with-downstream-feedback - V-Pretraining optimizes pretraining tasks using verified goal information to align with downstream capabilities.
- ECO: Quantized Training without Full-Precision Master Weights (viability: 3): https://sciencetostartup.com/paper/eco-quantized-training-without-full-precision-master-weights - ECO enables memory-efficient LLM training by eliminating high-precision master weights through an error-compensating optimizer.
- Investigating Associational Biases in Inter-Model Communication of Large Generative Models (viability: 5): https://sciencetostartup.com/paper/investigating-associational-biases-in-inter-model-communication-of-large-generative-models - Develop a tool to quantify and mitigate associational biases in inter-model communication for human-centred AI tasks.
- Latent Adversarial Regularization for Offline Preference Optimization (viability: 3): https://sciencetostartup.com/paper/latent-adversarial-regularization-for-offline-preference-optimization - Introducing GANPO for latent-space regularization in language model preference optimization.
- Vision-DeepResearch: Incentivizing DeepResearch Capability in Multimodal Large Language Models (viability: 5): https://sciencetostartup.com/paper/vision-deepresearch-incentivizing-deepresearch-capability-in-multimodal-large-language-models - Vision-DeepResearch enhances multimodal large language models with advanced deep-research capabilities for improved real-world search performance.
- Unsupervised Decomposition and Recombination with Discriminator-Driven Diffusion Models (viability: 8): https://sciencetostartup.com/paper/unsupervised-decomposition-and-recombination-with-discriminator-driven-diffusion-models - Builds a tool to automatically decompose data into reusable components for recombination and synthesis using diffusion models.
- MetricAnything: Scaling Metric Depth Pretraining with Noisy Heterogeneous Sources (viability: 8): https://sciencetostartup.com/paper/metricanything-scaling-metric-depth-pretraining-with-noisy-heterogeneous-sources - MetricAnything provides a scalable pretraining framework for metric depth estimation from diverse 3D data sources, achieving state-of-the-art results and broadening AI applications in spatial intelligence.
- SIA: Symbolic Interpretability for Anticipatory Deep Reinforcement Learning in Network Control (viability: 3): https://sciencetostartup.com/paper/sia-symbolic-interpretability-for-anticipatory-deep-reinforcement-learning-in-network-control - Interpret and optimize deep reinforcement learning for mobile networks with real-time insights.
- Mechanistic Data Attribution: Tracing the Training Origins of Interpretable LLM Units (viability: 2): https://sciencetostartup.com/paper/mechanistic-data-attribution-tracing-the-training-origins-of-interpretable-llm-units - Mechanistic Data Attribution traces the training data origins of interpretable components in large language models.
- Liquid Interfaces: A Dynamic Ontology for the Interoperability of Autonomous Systems (viability: 5): https://sciencetostartup.com/paper/liquid-interfaces-a-dynamic-ontology-for-the-interoperability-of-autonomous-systems - Liquid Interfaces enable dynamic and adaptive interaction protocols for autonomous agents through a novel coordination paradigm.
- Geometry of Drifting MDPs with Path-Integral Stability Certificates (viability: 3): https://sciencetostartup.com/paper/geometry-of-drifting-mdps-with-path-integral-stability-certificates - Homotopy-Tracking RL (HT-RL) offers stability certificates for adapting to nonstationary environments in reinforcement learning.
- Generalized Information Gathering Under Dynamics Uncertainty (viability: 3): https://sciencetostartup.com/paper/generalized-information-gathering-under-dynamics-uncertainty - Develop a unified framework for decoupling modelling choices in information-gathering for agents in unknown dynamical systems.
- VERSA: Verified Event Data Format for Reliable Soccer Analytics (viability: 7): https://sciencetostartup.com/paper/versa-verified-event-data-format-for-reliable-soccer-analytics - A verification framework that ensures the integrity of event stream data for reliable soccer analytics.
- From Particles to Agents: Hallucination as a Metric for Cognitive Friction in Spatial Simulation (viability: 2): https://sciencetostartup.com/paper/from-particles-to-agents-hallucination-as-a-metric-for-cognitive-friction-in-spatial-simulation - Introducing Agentic Environmental Simulations to model spatial environments as cognitive partners using AI-driven generative models.
- Mind the Gap: How Elicitation Protocols Shape the Stated-Revealed Preference Gap in Language Models (viability: 2): https://sciencetostartup.com/paper/mind-the-gap-how-elicitation-protocols-shape-the-stated-revealed-preference-gap-in-language-models - Improve language model preference measurement by evaluating elicitation protocols' impact on preference expression.
- Learning Decentralized LLM Collaboration with Multi-Agent Actor Critic (viability: 7): https://sciencetostartup.com/paper/learning-decentralized-llm-collaboration-with-multi-agent-actor-critic - Enable decentralized LLM collaboration using efficient Multi-Agent Actor-Critic methods for improved parallel processing.
- MoE-ACT: Improving Surgical Imitation Learning Policies through Supervised Mixture-of-Experts (viability: 6): https://sciencetostartup.com/paper/moe-act-improving-surgical-imitation-learning-policies-through-supervised-mixture-of-experts - A supervised Mixture-of-Experts architecture enhances surgical robot imitation learning with minimal data and superior robustness.
- Token-Guard: Towards Token-Level Hallucination Control via Self-Checking Decoding (viability: 6): https://sciencetostartup.com/paper/token-guard-towards-token-level-hallucination-control-via-self-checking-decoding - Token-Guard offers a scalable solution to reduce hallucinations in LLMs through token-level self-checking and dynamic corrections.
- The Energy Impact of Domain Model Design in Classical Planning (viability: 3): https://sciencetostartup.com/paper/the-energy-impact-of-domain-model-design-in-classical-planning - Explore energy-efficient domain model designs to optimize the energy consumption of classical AI planners.
- Industrialized Deception: The Collateral Effects of LLM-Generated Misinformation on Digital Ecosystems (viability: 4): https://sciencetostartup.com/paper/industrialized-deception-the-collateral-effects-of-llm-generated-misinformation-on-digital-ecosystems - Develop tools to evaluate and mitigate the impact of AI-generated misinformation in digital ecosystems.
- How do Visual Attributes Influence Web Agents? A Comprehensive Evaluation of User Interface Design Factors (viability: 5): https://sciencetostartup.com/paper/how-do-visual-attributes-influence-web-agents-a-comprehensive-evaluation-of-user-interface-design-factors - VAF evaluates how visual attributes of web interfaces affect web agent decision-making.
- ToolWeaver: Weaving Collaborative Semantics for Scalable Tool Use in Large Language Models (viability: 3): https://sciencetostartup.com/paper/toolweaver-weaving-collaborative-semantics-for-scalable-tool-use-in-large-language-models - ToolWeaver enhances scalability and semantic understanding in LLM tool use by encoding tools into hierarchical sequences for improved collaboration and performance.
- Robust Multimodal Representation Learning in Healthcare (viability: 6): https://sciencetostartup.com/paper/robust-multimodal-representation-learning-in-healthcare - A framework to boost clinical outcome predictions by untangling biases in medical multimodal data.
- Retrieval-Infused Reasoning Sandbox: A Benchmark for Decoupling Retrieval and Reasoning Capabilities (viability: 5): https://sciencetostartup.com/paper/retrieval-infused-reasoning-sandbox-a-benchmark-for-decoupling-retrieval-and-reasoning-capabilities - A benchmark sandbox for evaluating and isolating retrieval and reasoning capabilities in AI models.
- AgenticSimLaw: A Juvenile Courtroom Multi-Agent Debate Simulation for Explainable High-Stakes Tabular Decision Making (viability: 8): https://sciencetostartup.com/paper/agenticsimlaw-a-juvenile-courtroom-multi-agent-debate-simulation-for-explainable-high-stakes-tabular-decision-making - AgenticSimLaw provides an explainable multi-agent debate framework for transparent high-stakes decision-making in juvenile justice.
- From Future of Work to Future of Workers: Addressing Asymptomatic AI Harms for Dignified Human-AI Interaction (viability: 3): https://sciencetostartup.com/paper/from-future-of-work-to-future-of-workers-addressing-asymptomatic-ai-harms-for-dignified-human-ai-interaction - A framework for preserving human expertise in AI-augmented work environments across high-stakes industries.
- Self-Compression of Chain-of-Thought via Multi-Agent Reinforcement Learning (viability: 6): https://sciencetostartup.com/paper/self-compression-of-chain-of-thought-via-multi-agent-reinforcement-learning - Optimize reasoning efficiency in AI models using multi-agent reinforcement learning to reduce inference overhead.
- JADE: Bridging the Strategic-Operational Gap in Dynamic Agentic RAG (viability: 6): https://sciencetostartup.com/paper/jade-bridging-the-strategic-operational-gap-in-dynamic-agentic-rag - Develop a joint optimization framework for dynamic agentic RAG workflows to improve performance and efficiency.
- ProRAG: Process-Supervised Reinforcement Learning for Retrieval-Augmented Generation (viability: 6): https://sciencetostartup.com/paper/prorag-process-supervised-reinforcement-learning-for-retrieval-augmented-generation - ProRAG provides process-supervised reinforcement learning for enhanced retrieval-augmented generation in complex reasoning tasks.
- From Meta-Thought to Execution: Cognitively Aligned Post-Training for Generalizable and Reliable LLM Reasoning (viability: 7): https://sciencetostartup.com/paper/from-meta-thought-to-execution-cognitively-aligned-post-training-for-generalizable-and-reliable-llm-reasoning - Develop a post-training framework for large language models that enhances generalization and reduces resource consumption by aligning with human cognitive processes.
- TraceRouter: Robust Safety for Large Foundation Models via Path-Level Intervention (viability: 7): https://sciencetostartup.com/paper/tracerouter-robust-safety-for-large-foundation-models-via-path-level-intervention - TraceRouter offers enhanced adversarial robustness for large foundation models by surgically tracing and suppressing harmful semantics.
- Making Models Unmergeable via Scaling-Sensitive Loss Landscape (viability: 3): https://sciencetostartup.com/paper/making-models-unmergeable-via-scaling-sensitive-loss-landscape - scTrap2 offers an architecture-agnostic framework to prevent unauthorized model merging without relying on architecture-specific methods.
- Learn-to-Distance: Distance Learning for Detecting LLM-Generated Text (viability: 7): https://sciencetostartup.com/paper/learn-to-distance-distance-learning-for-detecting-llm-generated-text - A novel algorithm to reliably detect LLM-generated text, outperforming current baselines.
- Improving Classifier-Free Guidance of Flow Matching via Manifold Projection (viability: 4): https://sciencetostartup.com/paper/improving-classifier-free-guidance-of-flow-matching-via-manifold-projection - Enhancing classifier-free guidance in diffusion models for improved generation fidelity through manifold projection and optimization techniques.
- astra-langchain4j: Experiences Combining LLMs and Agent Programming (viability: 3): https://sciencetostartup.com/paper/astra-langchain4j-experiences-combining-llms-and-agent-programming - Develops prototype LLM integration for ASTRA language to explore impact on traditional and agentic AI toolkits.
- WebArbiter: A Principle-Guided Reasoning Process Reward Model for Web Agents (viability: 7): https://sciencetostartup.com/paper/webarbiter-a-principle-guided-reasoning-process-reward-model-for-web-agents - WebArbiter enhances web automation with principle-guided reward modeling for robust decision-making.
- MoHETS: Long-term Time Series Forecasting with Mixture-of-Heterogeneous-Experts (viability: 5): https://sciencetostartup.com/paper/mohets-long-term-time-series-forecasting-with-mixture-of-heterogeneous-experts - MoHETS offers a state-of-the-art long-term time series forecasting tool with efficient, heterogeneous expert layers for diverse temporal dynamics.
- KnowBias: Mitigating Social Bias in LLMs via Know-Bias Neuron Enhancement (viability: 8): https://sciencetostartup.com/paper/knowbias-mitigating-social-bias-in-llms-via-know-bias-neuron-enhancement - KnowBias reduces social biases in LLMs through neuron enhancement, preserving model performance.
- Bridging Forecast Accuracy and Inventory KPIs: A Simulation-Based Software Framework (viability: 5): https://sciencetostartup.com/paper/bridging-forecast-accuracy-and-inventory-kpis-a-simulation-based-software-framework - Create a simulation-based software that links forecasting accuracy to inventory management KPIs for the automotive aftermarket.
- Test-Time Compute Games (viability: 5): https://sciencetostartup.com/paper/test-time-compute-games - Optimize cloud costs for LLM APIs using auction-based compute efficiency.
- Trustworthy Intelligent Education: A Systematic Perspective on Progress, Challenges, and Future Directions (viability: 2): https://sciencetostartup.com/paper/trustworthy-intelligent-education-a-systematic-perspective-on-progress-challenges-and-future-directions - A systematic review of trustworthiness in intelligent education, lacking immediate commercial applicability.
- Looking Beyond Accuracy: A Holistic Benchmark of ECG Foundation Models (viability: 4): https://sciencetostartup.com/paper/looking-beyond-accuracy-a-holistic-benchmark-of-ecg-foundation-models - Develop a comprehensive benchmark for evaluating ECG-expert foundation models in healthcare settings.
- CORE:Toward Ubiquitous 6G Intelligence Through Collaborative Orchestration of Large Language Model Agents Over Hierarchical Edge (viability: 6): https://sciencetostartup.com/paper/core-toward-ubiquitous-6g-intelligence-through-collaborative-orchestration-of-large-language-model-agents-over-hierarchi - CORE: A framework enabling 6G intelligence through dynamic orchestration of LLM agents across hierarchical edge networks.
- Moral Outrage Shapes Commitments Beyond Attention: Multimodal Moral Emotions on YouTube in Korea and the US (viability: 5): https://sciencetostartup.com/paper/moral-outrage-shapes-commitments-beyond-attention-multimodal-moral-emotions-on-youtube-in-korea-and-the-us - Develop a tool that analyzes moral outrage in YouTube content to understand audience engagement across cultures.
- A Unified XAI-LLM Approach for EndotrachealSuctioning Activity Recognition (viability: 8): https://sciencetostartup.com/paper/a-unified-xai-llm-approach-for-endotrachealsuctioning-activity-recognition - Develop an AI-powered tool to improve nurse training in endotracheal suctioning through video-based activity recognition using explainable AI.
- BioAgent Bench: An AI Agent Evaluation Suite for Bioinformatics (viability: 7): https://sciencetostartup.com/paper/bioagent-bench-an-ai-agent-evaluation-suite-for-bioinformatics - BioAgent Bench provides a benchmark and evaluation suite to test AI agents on bioinformatics tasks with a focus on robustness and privacy.
- KID: Knowledge-Injected Dual-Head Learning for Knowledge-Grounded Harmful Meme Detection (viability: 7): https://sciencetostartup.com/paper/kid-knowledge-injected-dual-head-learning-for-knowledge-grounded-harmful-meme-detection - KID provides state-of-the-art knowledge-grounded harmful meme detection with a dual-head learning architecture.
- Effective LoRA Adapter Routing using Task Representations (viability: 6): https://sciencetostartup.com/paper/effective-lora-adapter-routing-using-task-representations - LORAUTER optimizes task-based routing of LoRA adapters in large language models for efficient specialization and enhanced performance.
- ECSEL: Explainable Classification via Signomial Equation Learning (viability: 8): https://sciencetostartup.com/paper/ecsel-explainable-classification-via-signomial-equation-learning - ECSEL provides an efficient, explainable classification tool for exposing biases and supporting counterfactual reasoning in datasets, with applications in e-commerce and fraud detection.
- Assessing the Business Process Modeling Competences of Large Language Models (viability: 5): https://sciencetostartup.com/paper/assessing-the-business-process-modeling-competences-of-large-language-models - Innovative LLM evaluation framework for improving business process modeling accuracy.
- Synthetic-to-Real Domain Bridging for Single-View 3D Reconstruction of Ships for Maritime Monitoring (viability: 7): https://sciencetostartup.com/paper/synthetic-to-real-domain-bridging-for-single-view-3d-reconstruction-of-ships-for-maritime-monitoring - AI-powered single-view 3D reconstruction for real-time maritime ship monitoring using synthetic-to-real domain bridging.
- Abstract Concept Modelling in Conceptual Spaces: A Study on Chess Strategies (viability: 5): https://sciencetostartup.com/paper/abstract-concept-modelling-in-conceptual-spaces-a-study-on-chess-strategies - Develop a conceptual framework for modeling and analyzing abstract chess strategies as geometric trajectories in a conceptual space.
- CoFrGeNet: Continued Fraction Architectures for Language Generation (viability: 5): https://sciencetostartup.com/paper/cofrgenet-continued-fraction-architectures-for-language-generation - Introducing CoFrGeNet, a continued fraction-based architecture for efficient language model generation with fewer parameters.
- Zero-Shot Statistical Downscaling via Diffusion Posterior Sampling (viability: 2): https://sciencetostartup.com/paper/zero-shot-statistical-downscaling-via-diffusion-posterior-sampling - Develop a zero-shot framework for climate model downscaling that ensures physical consistency across diverse GCMs.
- EWSJF: An Adaptive Scheduler with Hybrid Partitioning for Mixed-Workload LLM Inference (viability: 6): https://sciencetostartup.com/paper/ewsjf-an-adaptive-scheduler-with-hybrid-partitioning-for-mixed-workload-llm-inference - EWSJF is an adaptive scheduler that significantly improves LLM inference by optimizing workload distribution to enhance fairness and throughput.
- Language-based Trial and Error Falls Behind in the Era of Experience (viability: 6): https://sciencetostartup.com/paper/language-based-trial-and-error-falls-behind-in-the-era-of-experience - SCOUT dramatically reduces the computational cost of exploration in training LLMs for nonlinguistic tasks by using lightweight scouts to gather environmental dynamics efficiently.
- Temporal Sepsis Modeling: a Fully Interpretable Relational Way (viability: 4): https://sciencetostartup.com/paper/temporal-sepsis-modeling-a-fully-interpretable-relational-way - Develop an interpretable machine learning model to improve sepsis prediction using a relational data approach.
- From Global to Granular: Revealing IQA Model Performance via Correlation Surface (viability: 5): https://sciencetostartup.com/paper/from-global-to-granular-revealing-iqa-model-performance-via-correlation-surface - A new method, GMC, provides fine-grained analysis of IQA model performance, surpassing traditional metrics with detailed 3D correlation mapping.
- DropoutTS: Sample-Adaptive Dropout for Robust Time Series Forecasting (viability: 5): https://sciencetostartup.com/paper/dropoutts-sample-adaptive-dropout-for-robust-time-series-forecasting - DropoutTS enhances time series model robustness with sample-adaptive dropout rates based on instance-level noise assessment.
- Enhancing Language Models for Robust Greenwashing Detection (viability: 5): https://sciencetostartup.com/paper/enhancing-language-models-for-robust-greenwashing-detection - Develop a robust NLP tool for detecting greenwashing in sustainability reports using enhanced language models.
- When does predictive inverse dynamics outperform behavior cloning? (viability: 5): https://sciencetostartup.com/paper/when-does-predictive-inverse-dynamics-outperform-behavior-cloning - Exploring when predictive inverse dynamics models outperform behavior cloning in imitation learning tasks.
- DreamActor-M2: Universal Character Image Animation via Spatiotemporal In-Context Learning (viability: 6): https://sciencetostartup.com/paper/dreamactor-m2-universal-character-image-animation-via-spatiotemporal-in-context-learning - DreamActor-M2 empowers universal character animation with a novel spatiotemporal learning framework for superior motion transfer and identity preservation.
- E-mem: Multi-agent based Episodic Context Reconstruction for LLM Agent Memory (viability: 7): https://sciencetostartup.com/paper/e-mem-multi-agent-based-episodic-context-reconstruction-for-llm-agent-memory - E-mem enhances LLM agent memory by using multi-agent-based episodic context reconstruction to improve reasoning and reduce token costs.
- Disentangling perception and reasoning for improving data efficiency in learning cloth manipulation without demonstrations (viability: 8): https://sciencetostartup.com/paper/disentangling-perception-and-reasoning-for-improving-data-efficiency-in-learning-cloth-manipulation-without-demonstratio - Develop a lightweight, efficient RL-based solution for robotic cloth manipulation, offering significant performance improvements with reduced model size and training time.
- TACLer: Tailored Curriculum Reinforcement Learning for Efficient Reasoning (viability: 6): https://sciencetostartup.com/paper/tacler-tailored-curriculum-reinforcement-learning-for-efficient-reasoning - TACLer provides a more efficient and accurate curriculum reinforcement learning framework for LLM reasoning tasks, significantly reducing computational costs while improving performance.
- FBS: Modeling Native Parallel Reading inside a Transformer (viability: 6): https://sciencetostartup.com/paper/fbs-modeling-native-parallel-reading-inside-a-transformer - A novel Transformer model enhancing reading efficiency for faster language model inference.
- Toward Culturally Aligned LLMs through Ontology-Guided Multi-Agent Reasoning (viability: 5): https://sciencetostartup.com/paper/toward-culturally-aligned-llms-through-ontology-guided-multi-agent-reasoning - Develop an LLM framework that enhances cultural alignment using ontology-guided multi-agent reasoning.
- Curriculum Learning for LLM Pretraining: An Analysis of Learning Dynamics (viability: 3): https://sciencetostartup.com/paper/curriculum-learning-for-llm-pretraining-an-analysis-of-learning-dynamics - Research paper analyzing the impact of curriculum learning on the training dynamics of language models without indicating commercialization or direct product development.
- XFACTORS: Disentangled Information Bottleneck via Contrastive Supervision (viability: 6): https://sciencetostartup.com/paper/xfactors-disentangled-information-bottleneck-via-contrastive-supervision - Implement XFactors, a weakly-supervised VAE framework for disentangled representation learning, into a controllable ML tool for targeted factor manipulation.
- FIT: Defying Catastrophic Forgetting in Continual LLM Unlearning (viability: 7): https://sciencetostartup.com/paper/fit-defying-catastrophic-forgetting-in-continual-llm-unlearning - Introducing FIT, a robust framework enabling continual unlearning in LLMs to address privacy and content concerns without catastrophic forgetting.
- Expected Return Causes Outcome-Level Mode Collapse in Reinforcement Learning and How to Fix It with Inverse Probability Scaling (viability: 2): https://sciencetostartup.com/paper/expected-return-causes-outcome-level-mode-collapse-in-reinforcement-learning-and-how-to-fix-it-with-inverse-probability- - A novel reinforcement learning approach to prevent mode collapse by adjusting the learning objective for better policy diversity.
- SONIC-O1: A Real-World Benchmark for Evaluating Multimodal Large Language Models on Audio-Video Understanding (viability: 8): https://sciencetostartup.com/paper/sonic-o1-a-real-world-benchmark-for-evaluating-multimodal-large-language-models-on-audio-video-understanding - SONIC-O1 offers a comprehensive benchmark for evaluating multimodal AI models in real-world audio-video tasks, addressing gaps in current evaluation methods.
- SENDAI: A Hierarchical Sparse-measurement, EfficieNt Data AssImilation Framework (viability: 7): https://sciencetostartup.com/paper/sendai-a-hierarchical-sparse-measurement-efficient-data-assimilation-framework - SENDAI allows efficient sparse-measurement reconstruction with significant structural improvements, ideal for remote sensing and real-time monitoring.
- ScholarGym: Benchmarking Deep Research Workflows on Academic Literature Retrieval (viability: 2): https://sciencetostartup.com/paper/scholargym-benchmarking-deep-research-workflows-on-academic-literature-retrieval - ScholarGym offers a controlled simulation environment for benchmarking deep research workflows on academic literature retrieval.
- Gauge-invariant representation holonomy (viability: 2): https://sciencetostartup.com/paper/gauge-invariant-representation-holonomy - Diagnose geometric structures of learned representations with gauge-invariant representation holonomy.
- SWE-Spot: Building Small Repo-Experts with Repository-Centric Learning (viability: 2): https://sciencetostartup.com/paper/swe-spot-building-small-repo-experts-with-repository-centric-learning - Develop compact, repository-specialized language models using Repository-Centric Learning for efficient coding tasks.
- ILRR: Inference-Time Steering Method for Masked Diffusion Language Models (viability: 7): https://sciencetostartup.com/paper/ilrr-inference-time-steering-method-for-masked-diffusion-language-models - Steer your text generation models with precision using ILRR, a learning-free framework for inference-time control of Diffusion Language Models.
- Seg-MoE: Multi-Resolution Segment-wise Mixture-of-Experts for Time Series Forecasting Transformers (viability: 7): https://sciencetostartup.com/paper/seg-moe-multi-resolution-segment-wise-mixture-of-experts-for-time-series-forecasting-transformers - Develop segment-wise mixture-of-experts layers for state-of-the-art time series forecasting with Transformers.
- HeRo-Q: A General Framework for Stable Low Bit Quantization via Hessian Conditioning (viability: 5): https://sciencetostartup.com/paper/hero-q-a-general-framework-for-stable-low-bit-quantization-via-hessian-conditioning - HeRo-Q offers a robust, easy-to-integrate algorithm that enhances the stability of low bit quantization in large language models.
- Breaking the Overscaling Curse: Thinking Parallelism Before Parallel Thinking (viability: 4): https://sciencetostartup.com/paper/breaking-the-overscaling-curse-thinking-parallelism-before-parallel-thinking - Optimize parallelism in language models to reduce computational costs and maintain performance.
- Semantic Content Determines Algorithmic Performance (viability: 2): https://sciencetostartup.com/paper/semantic-content-determines-algorithmic-performance - WhatCounts reveals semantic dependencies in algorithmic behavior, raising awareness about the limitations of language models.
- Beyond Parameter Finetuning: Test-Time Representation Refinement for Node Classification (viability: 5): https://sciencetostartup.com/paper/beyond-parameter-finetuning-test-time-representation-refinement-for-node-classification - Leverage TTReFT for adaptive Graph Neural Networks that excel in out-of-distribution node classification tasks.
- Thinking Broad, Acting Fast: Latent Reasoning Distillation from Multi-Perspective Chain-of-Thought for E-Commerce Relevance (viability: 6): https://sciencetostartup.com/paper/thinking-broad-acting-fast-latent-reasoning-distillation-from-multi-perspective-chain-of-thought-for-e-commerce-relevanc - Optimize e-commerce relevance using multi-perspective Chain-of-Thought reasoning for faster and more accurate search results.
- RecNet: Self-Evolving Preference Propagation for Agentic Recommender Systems (viability: 6): https://sciencetostartup.com/paper/recnet-self-evolving-preference-propagation-for-agentic-recommender-systems - RecNet revolutionizes recommender systems by leveraging agent-based preference propagation for dynamic and personalized user experiences.
- Search-Based Risk Feature Discovery in Document Structure Spaces under a Constrained Budget (viability: 5): https://sciencetostartup.com/paper/search-based-risk-feature-discovery-in-document-structure-spaces-under-a-constrained-budget - Revolutionize enterprise document failure detection using a portfolio of search strategies for early-phase validation.
- Dynamics Reveals Structure: Challenging the Linear Propagation Assumption (viability: 1): https://sciencetostartup.com/paper/dynamics-reveals-structure-challenging-the-linear-propagation-assumption - New theoretical insights into the limitations of the Linear Propagation Assumption in neural networks.
- Beyond Imitation: Reinforcement Learning for Active Latent Planning (viability: 5): https://sciencetostartup.com/paper/beyond-imitation-reinforcement-learning-for-active-latent-planning - ATP-Latent enhances latent reasoning in LLMs by integrating reinforcement learning for optimal planning.
- Scalable Power Sampling: Unlocking Efficient, Training-Free Reasoning for LLMs via Distribution Sharpening (viability: 5): https://sciencetostartup.com/paper/scalable-power-sampling-unlocking-efficient-training-free-reasoning-for-llms-via-distribution-sharpening - A novel training-free algorithm sharply improves LLM reasoning efficiency and drastically reduces inference costs.
- Depth-Recurrent Attention Mixtures: Giving Latent Reasoning the Attention it Deserves (viability: 8): https://sciencetostartup.com/paper/depth-recurrent-attention-mixtures-giving-latent-reasoning-the-attention-it-deserves - Depth-Recurrent Attention Mixtures optimize scalable depth-recurrent models to significantly surpass current state-of-the-art performance in language reasoning with reduced training resources.
- Chain Of Thought Compression: A Theoritical Analysis (viability: 4): https://sciencetostartup.com/paper/chain-of-thought-compression-a-theoritical-analysis - Develop ALiCoT: a framework achieving efficient reasoning in language models through implicit chain-of-thought compression, providing significant speed improvements.
- Signal-Adaptive Trust Regions for Gradient-Free Optimization of Recurrent Spiking Neural Networks (viability: 5): https://sciencetostartup.com/paper/signal-adaptive-trust-regions-for-gradient-free-optimization-of-recurrent-spiking-neural-networks - Optimize energy-efficient recurrent spiking neural networks for reinforcement learning using signal-adaptive trust regions.
- Shaping capabilities with token-level data filtering (viability: 4): https://sciencetostartup.com/paper/shaping-capabilities-with-token-level-data-filtering - Develop a service that filters training data tokens to limit undesired capabilities in language models.
- EmboCoach-Bench: Benchmarking AI Agents on Developing Embodied Robots (viability: 6): https://sciencetostartup.com/paper/embocoach-bench-benchmarking-ai-agents-on-developing-embodied-robots - A benchmark for autonomous LLM agents to optimize embodied AI policies, enhancing robotic intelligence with agent feedback loops.
- SAL: Selective Adaptive Learning for Backpropagation-Free Training with Sparsification (viability: 5): https://sciencetostartup.com/paper/sal-selective-adaptive-learning-for-backpropagation-free-training-with-sparsification - Selective Adaptive Learning offers an alternative training method potentially improving neural network scalability without backpropagation.
- Meta Context Engineering via Agentic Skill Evolution (viability: 8): https://sciencetostartup.com/paper/meta-context-engineering-via-agentic-skill-evolution - Meta Context Engineering optimizes large language model outputs through bi-level skill evolution.
- Training slow silicon neurons to control extremely fast robots with spiking reinforcement learning (viability: 3): https://sciencetostartup.com/paper/training-slow-silicon-neurons-to-control-extremely-fast-robots-with-spiking-reinforcement-learning - Develop a neuromorphic processor using spiking neurons for real-time robotic control in fast-paced environments like air hockey.
- Multi-Modal Time Series Prediction via Mixture of Modulated Experts (viability: 5): https://sciencetostartup.com/paper/multi-modal-time-series-prediction-via-mixture-of-modulated-experts - A new paradigm for multi-modal time series prediction that enhances forecasting accuracy using Expert Modulation.
- ShardMemo: Masked MoE Routing for Sharded Agentic LLM Memory (viability: 3): https://sciencetostartup.com/paper/shardmemo-masked-moe-routing-for-sharded-agentic-llm-memory - ShardMemo optimizes sharded memory routing for agentic LLMs, improving efficiency and performance over baselines.
- Bi-Anchor Interpolation Solver for Accelerating Generative Modeling (viability: 6): https://sciencetostartup.com/paper/bi-anchor-interpolation-solver-for-accelerating-generative-modeling - Accelerate generative modeling with a fast, training-free solver using minimal computational overhead.
- ARGORA: Orchestrated Argumentation for Causally Grounded LLM Reasoning and Decision Making (viability: 6): https://sciencetostartup.com/paper/argora-orchestrated-argumentation-for-causally-grounded-llm-reasoning-and-decision-making - ARGORA helps organizations resolve disagreements in multi-expert systems by visualizing and correcting argumentation paths to improve decision-making accuracy.
- KAPSO: A Knowledge-grounded framework for Autonomous Program Synthesis and Optimization (viability: 7): https://sciencetostartup.com/paper/kapso-a-knowledge-grounded-framework-for-autonomous-program-synthesis-and-optimization - KAPSO autonomously optimizes program development using knowledge-grounded synthesis loops.
- More Bang for the Buck: Improving the Inference of Large Language Models at a Fixed Budget using Reset and Discard (ReD) (viability: 5): https://sciencetostartup.com/paper/more-bang-for-the-buck-improving-the-inference-of-large-language-models-at-a-fixed-budget-using-reset-and-discard-red - Optimize LLM inference cost-effectively using the Reset-and-Discard method.
- LLaMEA-SAGE: Guiding Automated Algorithm Design with Structural Feedback from Explainable AI (viability: 7): https://sciencetostartup.com/paper/llamea-sage-guiding-automated-algorithm-design-with-structural-feedback-from-explainable-ai - LLaMEA-SAGE enhances automated algorithm design by using explainable AI to guide LLM-based code generation through structural feedback.
- The Effectiveness of Style Vectors for Steering Large Language Models: A Human Evaluation (viability: 6): https://sciencetostartup.com/paper/the-effectiveness-of-style-vectors-for-steering-large-language-models-a-human-evaluation - Leverage activation steering to control emotional tone in large language model outputs for scalable text generation applications.
- MAR: Efficient Large Language Models via Module-aware Architecture Refinement (viability: 7): https://sciencetostartup.com/paper/mar-efficient-large-language-models-via-module-aware-architecture-refinement - Build efficient and practical Large Language Models using Module-aware Architecture Refinement to lower energy costs without sacrificing performance.
- SimGraph: A Unified Framework for Scene Graph-Based Image Generation and Editing (viability: 7): https://sciencetostartup.com/paper/simgraph-a-unified-framework-for-scene-graph-based-image-generation-and-editing - SimGraph provides precise control over image generation and editing using scene graph-based methodology for superior spatial consistency.
- The Path of Least Resistance: Guiding LLM Reasining Trajectories with Prefix Consensus (viability: 3): https://sciencetostartup.com/paper/the-path-of-least-resistance-guiding-llm-reasining-trajectories-with-prefix-consensus - PoLR optimizes large language model reasoning efficiency by clustering and expanding reasoning paths, reducing computation without losing accuracy.
- Mean-Field Control on Sparse Graphs: From Local Limits to GNNs via Neighborhood Distributions (viability: 2): https://sciencetostartup.com/paper/mean-field-control-on-sparse-graphs-from-local-limits-to-gnns-via-neighborhood-distributions - Develop scalable reinforcement learning for multi-agent systems on sparse networks using mean-field control and graph neural networks.
- Task-free Adaptive Meta Black-box Optimization (viability: 5): https://sciencetostartup.com/paper/task-free-adaptive-meta-black-box-optimization - Adaptive meta-learning optimizer for black-box optimization tasks without predefined task distributions.
- ScaleSim: Serving Large-Scale Multi-Agent Simulation with Invocation Distance-Based Memory Management (viability: 5): https://sciencetostartup.com/paper/scalesim-serving-large-scale-multi-agent-simulation-with-invocation-distance-based-memory-management - ScaleSim optimizes GPU memory management for large-scale multi-agent simulations, boosting performance by up to 1.74x.
- Adaptive Confidence Gating in Multi-Agent Collaboration for Efficient and Optimized Code Generation (viability: 3): https://sciencetostartup.com/paper/adaptive-confidence-gating-in-multi-agent-collaboration-for-efficient-and-optimized-code-generation - DebateCoder enhances small language model reasoning for code generation using multi-agent collaboration and adaptive confidence gating.
- MemOCR: Layout-Aware Visual Memory for Efficient Long-Horizon Reasoning (viability: 8): https://sciencetostartup.com/paper/memocr-layout-aware-visual-memory-for-efficient-long-horizon-reasoning - MemOCR optimizes long-horizon reasoning by using adaptive visual layout memory to compress interaction histories efficiently.
- Topeax -- An Improved Clustering Topic Model with Density Peak Detection and Lexical-Semantic Term Importance (viability: 5): https://sciencetostartup.com/paper/topeax-an-improved-clustering-topic-model-with-density-peak-detection-and-lexical-semantic-term-importance - Introducing Topeax, a more reliable and coherent topic modeling approach improving cluster recovery and description over existing methods like Top2Vec and BERTopic.
- Conversation for Non-verifiable Learning: Self-Evolving LLMs through Meta-Evaluation (viability: 4): https://sciencetostartup.com/paper/conversation-for-non-verifiable-learning-self-evolving-llms-through-meta-evaluation - Develop a framework for self-evolving language models using meta-evaluation and multi-agent self-play to improve creative tasks.
- Unifying Speech Editing Detection and Content Localization via Prior-Enhanced Audio LLMs (viability: 7): https://sciencetostartup.com/paper/unifying-speech-editing-detection-and-content-localization-via-prior-enhanced-audio-llms - Develop a comprehensive speech editing detection tool using prior-enhanced audio LLMs for advanced tampering techniques.
- L$^3$: Large Lookup Layers (viability: 3): https://sciencetostartup.com/paper/l-3-large-lookup-layers - Introduce a systems-friendly architecture, L$^3$, for efficient sparse language modeling in transformers.
- HER: Human-like Reasoning and Reinforcement Learning for LLM Role-playing (viability: 7): https://sciencetostartup.com/paper/her-human-like-reasoning-and-reinforcement-learning-for-llm-role-playing - A framework for cognitive-level persona simulation in LLMs using dual-layer thinking and learning from human-aligned principles.
- LION: A Clifford Neural Paradigm for Multimodal-Attributed Graph Learning (viability: 8): https://sciencetostartup.com/paper/lion-a-clifford-neural-paradigm-for-multimodal-attributed-graph-learning - Develop multimodal-attributed graph learning tool using Clifford algebra to enhance data representation and performance.
- SAGE: Sequence-level Adaptive Gradient Evolution for Generative Recommendation (viability: 5): https://sciencetostartup.com/paper/sage-sequence-level-adaptive-gradient-evolution-for-generative-recommendation - Optimize recommender systems' adaptation using SAGE for enhanced scalability and diversity.
- ChipBench: A Next-Step Benchmark for Evaluating LLM Performance in AI-Aided Chip Design (viability: 6): https://sciencetostartup.com/paper/chipbench-a-next-step-benchmark-for-evaluating-llm-performance-in-ai-aided-chip-design - Develop a benchmarking tool to evaluate LLM performance in AI-assisted chip design workflows.
- Spava: Accelerating Long-Video Understanding via Sequence-Parallelism-aware Approximate Attention (viability: 6): https://sciencetostartup.com/paper/spava-accelerating-long-video-understanding-via-sequence-parallelism-aware-approximate-attention - Accelerate long-video understanding with Spava's multi-GPU sequence-parallelism and approximate attention, achieving significant speedups without performance loss.
- The Paradox of Robustness: Decoupling Rule-Based Logic from Affective Noise in High-Stakes Decision-Making (viability: 7): https://sciencetostartup.com/paper/the-paradox-of-robustness-decoupling-rule-based-logic-from-affective-noise-in-high-stakes-decision-making - Enhance decision-making stability using LLMs that resist narrative manipulation across high-stakes domains like healthcare, law, and finance.
- From Consistency to Complementarity: Aligned and Disentangled Multi-modal Learning for Time Series Understanding and Reasoning (viability: 5): https://sciencetostartup.com/paper/from-consistency-to-complementarity-aligned-and-disentangled-multi-modal-learning-for-time-series-understanding-and-reas - Develop a multi-modal learning platform for enhanced time series analysis through disentangled interaction and fine-grained alignment.
- When Prohibitions Become Permissions: Auditing Negation Sensitivity in Language Models (viability: 3): https://sciencetostartup.com/paper/when-prohibitions-become-permissions-auditing-negation-sensitivity-in-language-models - Develop a Negation Sensitivity Index to improve AI model compliance with negated instructions in high-stakes contexts.
- Mitigating Overthinking in Large Reasoning Models via Difficulty-aware Reinforcement Learning (viability: 5): https://sciencetostartup.com/paper/mitigating-overthinking-in-large-reasoning-models-via-difficulty-aware-reinforcement-learning - Difficulty-aware Policy Optimization reduces overthinking in AI reasoning models by adapting reasoning length to task complexity.
- System 1&2 Synergy via Dynamic Model Interpolation (viability: 5): https://sciencetostartup.com/paper/system-1-2-synergy-via-dynamic-model-interpolation - DAMI optimizes language model efficiency by dynamically interpolating cognitive configurations without additional training.
- DataCross: A Unified Benchmark and Agent Framework for Cross-Modal Heterogeneous Data Analysis (viability: 7): https://sciencetostartup.com/paper/datacross-a-unified-benchmark-and-agent-framework-for-cross-modal-heterogeneous-data-analysis - DataCross offers a framework for unified cross-modal data analysis, enhancing factuality and robustness in industrial decision-making.
- Intrinsic Reward Policy Optimization for Sparse-Reward Environments (viability: 6): https://sciencetostartup.com/paper/intrinsic-reward-policy-optimization-for-sparse-reward-environments - A reinforcement learning framework improving exploration in sparse-reward environments via intrinsic rewards.
- Understanding Frechet Speech Distance for Synthetic Speech Quality Evaluation (viability: 5): https://sciencetostartup.com/paper/understanding-frechet-speech-distance-for-synthetic-speech-quality-evaluation - Enhance synthetic speech evaluation with a scalable and cost-efficient alternative to human listening tests using Frechet Speech Distance metrics.
- TeachBench: A Syllabus-Grounded Framework for Evaluating Teaching Ability in Large Language Models (viability: 5): https://sciencetostartup.com/paper/teachbench-a-syllabus-grounded-framework-for-evaluating-teaching-ability-in-large-language-models - A framework leveraging large language models to evaluate and enhance their teaching abilities using a syllabus-grounded approach.
- NEMO: Execution-Aware Optimization Modeling via Autonomous Coding Agents (viability: 7): https://sciencetostartup.com/paper/nemo-execution-aware-optimization-modeling-via-autonomous-coding-agents - NEMO transforms natural language decision problem descriptions into executable mathematical optimization implementations using autonomous coding agents.
- Hebbian Learning with Global Direction (viability: 5): https://sciencetostartup.com/paper/hebbian-learning-with-global-direction - Global-guided Hebbian Learning (GHL) offers a biologically plausible alternative to backpropagation, integrating global and local learning signals.
- The Compliance Paradox: Semantic-Instruction Decoupling in Automated Academic Code Evaluation (viability: 5): https://sciencetostartup.com/paper/the-compliance-paradox-semantic-instruction-decoupling-in-automated-academic-code-evaluation - Exploit vulnerabilities in automated code evaluation to improve adjudicative robustness in academic settings.
- Latent Chain-of-Thought as Planning: Decoupling Reasoning from Verbalization (viability: 5): https://sciencetostartup.com/paper/latent-chain-of-thought-as-planning-decoupling-reasoning-from-verbalization - PLaT optimizes latent reasoning in AI by decoupling reasoning from verbalization to improve scalability and inference-time search.
- BEAP-Agent: Backtrackable Execution and Adaptive Planning for GUI Agents (viability: 7): https://sciencetostartup.com/paper/beap-agent-backtrackable-execution-and-adaptive-planning-for-gui-agents - A GUI automation tool with innovative backtracking for error correction in long, complex tasks.
- Theoretically Optimal Attention/FFN Ratios in Disaggregated LLM Serving (viability: 3): https://sciencetostartup.com/paper/theoretically-optimal-attention-ffn-ratios-in-disaggregated-llm-serving - Optimize computation resource allocation in LLMs using an analytical framework for better efficiency.
- L2R: Low-Rank and Lipschitz-Controlled Routing for Mixture-of-Experts (viability: 5): https://sciencetostartup.com/paper/l2r-low-rank-and-lipschitz-controlled-routing-for-mixture-of-experts - Develop a low-rank and Lipschitz-controlled routing framework to improve mixture-of-experts model performance.
- Memorization Control in Diffusion Models from Denoising-centric Perspective (viability: 4): https://sciencetostartup.com/paper/memorization-control-in-diffusion-models-from-denoising-centric-perspective - A new timestep sampling strategy for diffusion models reduces memorization and improves data distribution alignment.
- Dynamic Framework for Collaborative Learning: Leveraging Advanced LLM with Adaptive Feedback Mechanisms (viability: 6): https://sciencetostartup.com/paper/dynamic-framework-for-collaborative-learning-leveraging-advanced-llm-with-adaptive-feedback-mechanisms - Building a modular framework that enhances collaborative learning platforms with advanced LLMs and adaptive feedback for dynamic and inclusive engagement.
- Self-Improving Pretraining: using post-trained models to pretrain better models (viability: 5): https://sciencetostartup.com/paper/self-improving-pretraining-using-post-trained-models-to-pretrain-better-models - A pretraining method using reinforcement learning to enhance model safety and factuality by streaming documents and evaluating generated token sequences, improving quality from the ground up.
- Ostrakon-VL: Towards Domain-Expert MLLM for Food-Service and Retail Stores (viability: 8): https://sciencetostartup.com/paper/ostrakon-vl-towards-domain-expert-mllm-for-food-service-and-retail-stores - Ostrakon-VL enhances retail and food-service operations with a domain-specific AI model for robust perception and decision-making.
- EHR-RAG: Bridging Long-Horizon Structured Electronic Health Records and Large Language Models via Enhanced Retrieval-Augmented Generation (viability: 5): https://sciencetostartup.com/paper/ehr-rag-bridging-long-horizon-structured-electronic-health-records-and-large-language-models-via-enhanced-retrieval-augm - EHR-RAG enhances clinical prediction from long-horizon EHRs using advanced retrieval-augmented generation techniques.
- Within-Model vs Between-Prompt Variability in Large Language Models for Creative Tasks (viability: 2): https://sciencetostartup.com/paper/within-model-vs-between-prompt-variability-in-large-language-models-for-creative-tasks - Analyzes factors influencing output variance in LLMs for creative tasks but lacks practical application path.
- Modeling Endogenous Logic: Causal Neuro-Symbolic Reasoning Model for Explainable Multi-Behavior Recommendation (viability: 8): https://sciencetostartup.com/paper/modeling-endogenous-logic-causal-neuro-symbolic-reasoning-model-for-explainable-multi-behavior-recommendation - Neuro-Symbolic Reasoning Model providing explainable recommendations by operationalizing endogenous logic through causal inference.
- Adversarial Vulnerability Transcends Computational Paradigms: Feature Engineering Provides No Defense Against Neural Adversarial Transfer (viability: 4): https://sciencetostartup.com/paper/adversarial-vulnerability-transcends-computational-paradigms-feature-engineering-provides-no-defense-against-neural-adve - We provide a framework to evaluate and mitigate adversarial vulnerabilities in classical ML feature engineering pipelines using neural surrogates.
- White-Box Op-Amp Design via Human-Mimicking Reasoning (viability: 6): https://sciencetostartup.com/paper/white-box-op-amp-design-via-human-mimicking-reasoning - Develop an interpretable op-amp design tool leveraging AI to mimic human reasoning for engineers.
- Heterogeneous Vertiport Selection Optimization for On-Demand Air Taxi Services: A Deep Reinforcement Learning Approach (viability: 8): https://sciencetostartup.com/paper/heterogeneous-vertiport-selection-optimization-for-on-demand-air-taxi-services-a-deep-reinforcement-learning-approach - Optimize on-demand air taxi routing with deep reinforcement learning to cut urban travel time.
- Distributionally Robust Classification for Multi-source Unsupervised Domain Adaptation (viability: 3): https://sciencetostartup.com/paper/distributionally-robust-classification-for-multi-source-unsupervised-domain-adaptation - Develop a robust unsupervised domain adaptation tool for challenging scenarios with limited target data.
- The Surprising Difficulty of Search in Model-Based Reinforcement Learning (viability: 4): https://sciencetostartup.com/paper/the-surprising-difficulty-of-search-in-model-based-reinforcement-learning - New techniques for effective search in model-based reinforcement learning to achieve state-of-the-art performance.
- Grounding and Enhancing Informativeness and Utility in Dataset Distillation (viability: 5): https://sciencetostartup.com/paper/grounding-and-enhancing-informativeness-and-utility-in-dataset-distillation - Develop a framework for creating compact, optimized datasets that enhance AI training efficiency and effectiveness.
- Physics-Guided Tiny-Mamba Transformer for Reliability-Aware Early Fault Warning (viability: 3): https://sciencetostartup.com/paper/physics-guided-tiny-mamba-transformer-for-reliability-aware-early-fault-warning - A compact transformer model for providing early fault warnings in rotating machinery under varying conditions.
- Drive-KD: Multi-Teacher Distillation for VLMs in Autonomous Driving (viability: 2): https://sciencetostartup.com/paper/drive-kd-multi-teacher-distillation-for-vlms-in-autonomous-driving - Drive-KD optimizes autonomous driving VLMs through multi-teacher distillation to enhance efficiency and performance.
- Zenith: Scaling up Ranking Models for Billion-scale Livestreaming Recommendation (viability: 8): https://sciencetostartup.com/paper/zenith-scaling-up-ranking-models-for-billion-scale-livestreaming-recommendation - Zenith revolutionizes recommendation systems by scaling up model capacity for billion-scale livestreaming platforms without increasing latency.
- PILD: Physics-Informed Learning via Diffusion (viability: 5): https://sciencetostartup.com/paper/pild-physics-informed-learning-via-diffusion - PILD utilizes diffusion models integrated with physical laws for improved accuracy in engineering simulations.
- DUET: Distilled LLM Unlearning from an Efficiently Contextualized Teacher (viability: 6): https://sciencetostartup.com/paper/duet-distilled-llm-unlearning-from-an-efficiently-contextualized-teacher - Implement a distillation-based unlearning tool for LLMs to remove undesirable model knowledge efficiently.
- GeoRC: A Benchmark for Geolocation Reasoning Chains (viability: 5): https://sciencetostartup.com/paper/georc-a-benchmark-for-geolocation-reasoning-chains - GeoRC provides a benchmarking tool for evaluating geolocation reasoning in Vision Language Models.
- More Code, Less Reuse: Investigating Code Quality and Reviewer Sentiment towards AI-generated Pull Requests (viability: 2): https://sciencetostartup.com/paper/more-code-less-reuse-investigating-code-quality-and-reviewer-sentiment-towards-ai-generated-pull-requests - Analyzing the quality and reviewer sentiment of AI-generated pull requests to enhance human-AI collaboration in software development.
- Lightweight High-Fidelity Low-Bitrate Talking Face Compression for 3D Video Conference (viability: 7): https://sciencetostartup.com/paper/lightweight-high-fidelity-low-bitrate-talking-face-compression-for-3d-video-conference - Efficiently compress and render high-quality 3D talking faces for real-time video conferencing with low bitrate demands.
- Music Plagiarism Detection: Problem Formulation and a Segment-based Solution (viability: 6): https://sciencetostartup.com/paper/music-plagiarism-detection-problem-formulation-and-a-segment-based-solution - A segment-based solution for detecting music plagiarism using a newly defined task and dataset.
- Hypersolid: Emergent Vision Representations via Short-Range Repulsion (viability: 2): https://sciencetostartup.com/paper/hypersolid-emergent-vision-representations-via-short-range-repulsion - A novel self-supervised learning method that uses short-range repulsion for representation learning to prevent collapse.
- Position: Certifiable State Integrity in Cyber-Physical Systems -- Why Modular Sovereignty Solves the Plasticity-Stability Paradox (viability: 3): https://sciencetostartup.com/paper/position-certifiable-state-integrity-in-cyber-physical-systems-why-modular-sovereignty-solves-the-plasticity-stability-p - Modular Sovereignty offers a paradigm for certifiable state integrity in Cyber-Physical Systems through regime-specific specialist models.
- Conditional Generative Framework with Peak-Aware Attention for Robust Chemical Detection under Interferences (viability: 8): https://sciencetostartup.com/paper/conditional-generative-framework-with-peak-aware-attention-for-robust-chemical-detection-under-interferences - A robust AI framework for enhancing GC-MS chemical detection accuracy under interference conditions.
- Less Noise, More Voice: Reinforcement Learning for Reasoning via Instruction Purification (viability: 2): https://sciencetostartup.com/paper/less-noise-more-voice-reinforcement-learning-for-reasoning-via-instruction-purification - Enhance LLM reasoning through interference token removal to improve RLVR rollout efficiency.
- Understanding Diffusion Models via Ratio-Based Function Approximation with SignReLU Networks (viability: 2): https://sciencetostartup.com/paper/understanding-diffusion-models-via-ratio-based-function-approximation-with-signrelu-networks - Theoretical framework for approximating ratio functionals in diffusion models using SignReLU networks.
- TIDE: Tuning-Integrated Dynamic Evolution for LLM-Based Automated Heuristic Design (viability: 7): https://sciencetostartup.com/paper/tide-tuning-integrated-dynamic-evolution-for-llm-based-automated-heuristic-design - TIDE offers a novel framework for tuning-integrated dynamic evolution in LLM-driven heuristic design, significantly enhancing solution quality and computational efficiency across optimization problems.
- SHARP: Social Harm Analysis via Risk Profiles for Measuring Inequities in Large Language Models (viability: 5): https://sciencetostartup.com/paper/sharp-social-harm-analysis-via-risk-profiles-for-measuring-inequities-in-large-language-models - Assess and mitigate social harm in large language models using SHARP's multidimensional risk profiling framework.
- Delegation Without Living Governance (viability: 2): https://sciencetostartup.com/paper/delegation-without-living-governance - Explores governance frameworks to maintain human relevance in AI decision-making.
- MGSM-Pro: A Simple Strategy for Robust Multilingual Mathematical Reasoning Evaluation (viability: 5): https://sciencetostartup.com/paper/mgsm-pro-a-simple-strategy-for-robust-multilingual-mathematical-reasoning-evaluation - MGSM-Pro enhances multilingual mathematical reasoning evaluation with a robust dataset extending beyond English, enabling fairer assessment across languages.
- Causal Discovery for Explainable AI: A Dual-Encoding Approach (viability: 4): https://sciencetostartup.com/paper/causal-discovery-for-explainable-ai-a-dual-encoding-approach - Dual-encoding method for causal discovery in AI improves explainability by addressing categorical variable limitations.
- Temporal Context and Architecture: A Benchmark for Naturalistic EEG Decoding (viability: 6): https://sciencetostartup.com/paper/temporal-context-and-architecture-a-benchmark-for-naturalistic-eeg-decoding - Develop a high-accuracy EEG decoding tool for neurological applications leveraging efficient model architectures.
- Intelli-Planner: Towards Customized Urban Planning via Large Language Model Empowered Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/intelli-planner-towards-customized-urban-planning-via-large-language-model-empowered-reinforcement-learning - Intelli-Planner uses AI to streamline and customize urban planning, enhancing stakeholder satisfaction and efficiency.
- Uncovering Hidden Correctness in LLM Causal Reasoning via Symbolic Verification (viability: 4): https://sciencetostartup.com/paper/uncovering-hidden-correctness-in-llm-causal-reasoning-via-symbolic-verification - DoVerifier offers a symbolic verification tool to improve the accuracy of evaluating LLMs on causal reasoning by checking if outputs are derivable from causal graphs.
- When should I search more: Adaptive Complex Query Optimization with Reinforcement Learning (viability: 9): https://sciencetostartup.com/paper/when-should-i-search-more-adaptive-complex-query-optimization-with-reinforcement-learning - Adaptive Complex Query Optimization (ACQO) leverages reinforcement learning to revolutionize query optimization in Retrieval-Augmented Generation systems by dynamically managing complex user queries for improved efficacy.
- A Sheaf-Theoretic and Topological Perspective on Complex Network Modeling and Attention Mechanisms in Graph Neural Models (viability: 1): https://sciencetostartup.com/paper/a-sheaf-theoretic-and-topological-perspective-on-complex-network-modeling-and-attention-mechanisms-in-graph-neural-model - A topological framework for analyzing feature diffusion in graph neural networks using sheaf theory.
- Scaling Embeddings Outperforms Scaling Experts in Language Models (viability: 3): https://sciencetostartup.com/paper/scaling-embeddings-outperforms-scaling-experts-in-language-models - Explore embedding scaling as an alternative to Mixture-of-Experts for efficient language model sparsity.
- Thinker: A vision-language foundation model for embodied intelligence (viability: 5): https://sciencetostartup.com/paper/thinker-a-vision-language-foundation-model-for-embodied-intelligence - Develop Thinker, a vision-language model to enhance embodied intelligence in robotics with state-of-the-art video comprehension.
- ZipMoE: Efficient On-Device MoE Serving via Lossless Compression and Cache-Affinity Scheduling (viability: 3): https://sciencetostartup.com/paper/zipmoe-efficient-on-device-moe-serving-via-lossless-compression-and-cache-affinity-scheduling - ZipMoE optimizes on-device Mixture-of-Experts model serving through lossless compression and cache affinity to improve efficiency on edge devices.
- Do Reasoning Models Enhance Embedding Models? (viability: 2): https://sciencetostartup.com/paper/do-reasoning-models-enhance-embedding-models - Exploration into how reasoning models influence embedding models reveals no clear performance advantage.
- From Linear Input to Hierarchical Structure: Function Words as Statistical Cues for Language Learning (viability: 2): https://sciencetostartup.com/paper/from-linear-input-to-hierarchical-structure-function-words-as-statistical-cues-for-language-learning - This paper explores how statistical distribution of function words aids language learning across multiple languages.
- Sycophantic Anchors: Localizing and Quantifying User Agreement in Reasoning Models (viability: 3): https://sciencetostartup.com/paper/sycophantic-anchors-localizing-and-quantifying-user-agreement-in-reasoning-models - Detect and quantify sycophantic behaviors in reasoning models to improve AI decision-making processes.
- Rethinking Refinement: Correcting Generative Bias without Noise Injection (viability: 8): https://sciencetostartup.com/paper/rethinking-refinement-correcting-generative-bias-without-noise-injection - Bi-stage Flow Refinement (BFR) framework offers state-of-the-art bias correction for generative models, improving image quality with minimal computational overhead.
- MAD: Modality-Adaptive Decoding for Mitigating Cross-Modal Hallucinations in Multimodal Large Language Models (viability: 6): https://sciencetostartup.com/paper/mad-modality-adaptive-decoding-for-mitigating-cross-modal-hallucinations-in-multimodal-large-language-models - Introducing Modality-Adaptive Decoding (MAD) to reduce cross-modal hallucinations in multimodal AI for more accurate and reliable outputs.
- AC2L-GAD: Active Counterfactual Contrastive Learning for Graph Anomaly Detection (viability: 5): https://sciencetostartup.com/paper/ac2l-gad-active-counterfactual-contrastive-learning-for-graph-anomaly-detection - An innovative framework using active counterfactual contrastive learning to enhance graph anomaly detection efficiency and performance.
- Output-Space Search: Targeting LLM Generations in a Frozen Encoder-Defined Output Space (viability: 2): https://sciencetostartup.com/paper/output-space-search-targeting-llm-generations-in-a-frozen-encoder-defined-output-space - Explore Output-Space Search to enhance diversity and optimization in LLM outputs by targeting fixed encoder-defined output spaces.
- FrontierScience: Evaluating AI's Ability to Perform Expert-Level Scientific Tasks (viability: 5): https://sciencetostartup.com/paper/frontierscience-evaluating-ai-s-ability-to-perform-expert-level-scientific-tasks - FrontierScience provides a new benchmark for evaluating AI models on expert-level scientific reasoning tasks.
- Concise Geometric Description as a Bridge: Unleashing the Potential of LLM for Plane Geometry Problem Solving (viability: 6): https://sciencetostartup.com/paper/concise-geometric-description-as-a-bridge-unleashing-the-potential-of-llm-for-plane-geometry-problem-solving - Enhance plane geometry problem solving by converting visual diagrams into concise textual descriptions for LLMs.
- A2RAG: Adaptive Agentic Graph Retrieval for Cost-Aware and Reliable Reasoning (viability: 6): https://sciencetostartup.com/paper/a2rag-adaptive-agentic-graph-retrieval-for-cost-aware-and-reliable-reasoning - A2RAG is a cost-aware graph retrieval framework that enhances multi-hop question answering by adapting retrieval efforts based on query difficulty.
- Bridging the Arithmetic Gap: The Cognitive Complexity Benchmark and Financial-PoT for Robust Financial Reasoning (viability: 5): https://sciencetostartup.com/paper/bridging-the-arithmetic-gap-the-cognitive-complexity-benchmark-and-financial-pot-for-robust-financial-reasoning - Enhancing financial reasoning accuracy with a novel framework that reduces cognitive complexity in AI models.
- Can Neural Networks Learn Small Algebraic Worlds? An Investigation Into the Group-theoretic Structures Learned By Narrow Models Trained To Predict Group Operations (viability: 2): https://sciencetostartup.com/paper/can-neural-networks-learn-small-algebraic-worlds-an-investigation-into-the-group-theoretic-structures-learned-by-narrow- - Explore how small neural networks can learn abstract algebraic structures like group operations.
- Mobility-Embedded POIs: Learning What A Place Is and How It Is Used from Human Movement (viability: 3): https://sciencetostartup.com/paper/mobility-embedded-pois-learning-what-a-place-is-and-how-it-is-used-from-human-movement - Develop mobility-embedded POI representations to improve map enrichment by learning how places are used, not just identified.
- BrainStack: Neuro-MoE with Functionally Guided Expert Routing for EEG-Based Language Decoding (viability: 8): https://sciencetostartup.com/paper/brainstack-neuro-moe-with-functionally-guided-expert-routing-for-eeg-based-language-decoding - BrainStack offers a novel neuro-inspired framework for EEG-based language decoding, outperforming state-of-the-art models.
- Optimization and Mobile Deployment for Anthropocene Neural Style Transfer (viability: 9): https://sciencetostartup.com/paper/optimization-and-mobile-deployment-for-anthropocene-neural-style-transfer - AnthropoCam brings real-time neural style transfer to mobile for expressive, environmental visualization of Anthropocene landscapes.
- What You Feel Is Not What They See: On Predicting Self-Reported Emotion from Third-Party Observer Labels (viability: 3): https://sciencetostartup.com/paper/what-you-feel-is-not-what-they-see-on-predicting-self-reported-emotion-from-third-party-observer-labels - Leveraging personal significance to better align self-reported emotions with external observer models in mental health contexts.
- Beyond a Single Reference: Training and Evaluation with Paraphrases in Sign Language Translation (viability: 7): https://sciencetostartup.com/paper/beyond-a-single-reference-training-and-evaluation-with-paraphrases-in-sign-language-translation - Enhance Sign Language Translation accuracy with paraphrase-based evaluation tools.
- PhaseCoder: Microphone Geometry-Agnostic Spatial Audio Understanding for Multimodal LLMs (viability: 8): https://sciencetostartup.com/paper/phasecoder-microphone-geometry-agnostic-spatial-audio-understanding-for-multimodal-llms - PhaseCoder allows any device to perform spatial audio reasoning and transcription using a microphone-agnostic transformer-based encoder.
- CUA-Skill: Develop Skills for Computer Using Agent (viability: 8): https://sciencetostartup.com/paper/cua-skill-develop-skills-for-computer-using-agent - CUA-Skill provides a structured skills library for autonomous computer-using agents to enhance their efficiency and reliability.
- AI-Assisted Engineering Should Track the Epistemic Status and Temporal Validity of Architectural Decisions (viability: 2): https://sciencetostartup.com/paper/ai-assisted-engineering-should-track-the-epistemic-status-and-temporal-validity-of-architectural-decisions - Proposes a framework for tracking the epistemic status of architectural decisions in AI-assisted engineering.
- Multi-task Code LLMs: Data Mix or Model Merge? (viability: 5): https://sciencetostartup.com/paper/multi-task-code-llms-data-mix-or-model-merge - Develop high-performance, multi-task code generation models using data mixing or model merging strategies for efficient AI deployment.
- Planner-Auditor Twin: Agentic Discharge Planning with FHIR-Based LLM Planning, Guideline Recall, Optional Caching and Self-Improvement (viability: 3): https://sciencetostartup.com/paper/planner-auditor-twin-agentic-discharge-planning-with-fhir-based-llm-planning-guideline-recall-optional-caching-and-self- - A Planner-Auditor framework enhances clinical discharge planning reliability using FHIR data and self-improvement loops.
- How does information access affect LLM monitors' ability to detect sabotage? (viability: 9): https://sciencetostartup.com/paper/how-does-information-access-affect-llm-monitors-ability-to-detect-sabotage - Develop a robust LLM monitoring tool using the extract-and-evaluate method to detect sabotage with minimal information effectively.
- SteerEval: A Framework for Evaluating Steerability with Natural Language Profiles for Recommendation (viability: 6): https://sciencetostartup.com/paper/steereval-a-framework-for-evaluating-steerability-with-natural-language-profiles-for-recommendation - SteerEval provides a framework for enhancing the steerability of natural-language recommender systems through user-editable profiles.
- Magellan: Autonomous Discovery of Novel Compiler Optimization Heuristics with AlphaEvolve (viability: 6): https://sciencetostartup.com/paper/magellan-autonomous-discovery-of-novel-compiler-optimization-heuristics-with-alphaevolve - Magellan automates compiler optimization by evolving heuristics that outperform traditional methods with reduced engineering effort.
- Responsible AI: The Good, The Bad, The AI (viability: 3): https://sciencetostartup.com/paper/responsible-ai-the-good-the-bad-the-ai - A framework for responsible AI governance balancing value creation and risk mitigation.
- Safety Generalization Under Distribution Shift in Safe Reinforcement Learning: A Diabetes Testbed (viability: 6): https://sciencetostartup.com/paper/safety-generalization-under-distribution-shift-in-safe-reinforcement-learning-a-diabetes-testbed - A tool improving the safety of reinforcement learning in diabetes management through test-time shielding against distribution shifts.
- Deep Reinforcement Learning for Fault-Adaptive Routing in Eisenstein-Jacobi Interconnection Topologies (viability: 6): https://sciencetostartup.com/paper/deep-reinforcement-learning-for-fault-adaptive-routing-in-eisenstein-jacobi-interconnection-topologies - Develop a RL-based routing tool for adaptive, fault-resilient communication in many-core architectures.
- LOCUS: Low-Dimensional Model Embeddings for Efficient Model Exploration, Comparison, and Selection (viability: 5): https://sciencetostartup.com/paper/locus-low-dimensional-model-embeddings-for-efficient-model-exploration-comparison-and-selection - LOCUS offers low-dimensional embeddings to efficiently explore, compare, and select large language models by producing significant routing accuracy and reduced evaluation costs.
- Multi-modal Imputation for Alzheimer's Disease Classification (viability: 5): https://sciencetostartup.com/paper/multi-modal-imputation-for-alzheimer-s-disease-classification - Impute missing brain scan modalities using a conditional diffusion model to enhance Alzheimer's disease diagnostic accuracy.
- Towards Comprehensive Benchmarking Infrastructure for LLMs In Software Engineering (viability: 6): https://sciencetostartup.com/paper/towards-comprehensive-benchmarking-infrastructure-for-llms-in-software-engineering - BEHELM is a comprehensive benchmarking infrastructure that improves the evaluation of language models in software engineering.
- Textual Equilibrium Propagation for Deep Compound AI Systems (viability: 7): https://sciencetostartup.com/paper/textual-equilibrium-propagation-for-deep-compound-ai-systems - Optimize compound AI systems with enhanced textual feedback propagation to improve accuracy and efficiency in deep workflows.
- Llama-3.1-FoundationAI-SecurityLLM-Reasoning-8B Technical Report (viability: 7): https://sciencetostartup.com/paper/llama-3-1-foundationai-securityllm-reasoning-8b-technical-report - Foundation-Sec-8B-Reasoning is an open-source model providing robust cybersecurity analysis capabilities while maintaining general reasoning skills.
- SMKC: Sketch Based Kernel Correlation Images for Variable Cardinality Time Series Anomaly Detection (viability: 2): https://sciencetostartup.com/paper/smkc-sketch-based-kernel-correlation-images-for-variable-cardinality-time-series-anomaly-detection - SMKC provides a framework for anomaly detection in time series with dynamic sensor input, offering a cold-start solution without gradient updates.
- QUARK: Robust Retrieval under Non-Faithful Queries via Query-Anchored Aggregation (viability: 7): https://sciencetostartup.com/paper/quark-robust-retrieval-under-non-faithful-queries-via-query-anchored-aggregation - QUARK improves search retrieval accuracy by handling noisy user queries using query-anchored aggregation and recovery hypotheses.
- Log2Motion: Biomechanical Motion Synthesis from Touch Logs (viability: 4): https://sciencetostartup.com/paper/log2motion-biomechanical-motion-synthesis-from-touch-logs - Log2Motion converts touch data into biomechanical motion simulations to enhance the understanding of user interactions.
- Thinking in Frames: How Visual Context and Test-Time Scaling Empower Video Reasoning (viability: 4): https://sciencetostartup.com/paper/thinking-in-frames-how-visual-context-and-test-time-scaling-empower-video-reasoning - Develop a video generation model for enhanced visual reasoning in AI applications like maze navigation and tangram puzzles.
- SIGMA-PPG: Statistical-prior Informed Generative Masking Architecture for PPG Foundation Model (viability: 5): https://sciencetostartup.com/paper/sigma-ppg-statistical-prior-informed-generative-masking-architecture-for-ppg-foundation-model - Develop SIGMA-PPG, a generative model for PPG signals leveraging statistical priors to overcome noise and redundancy, with code available.
- "Unlimited Realm of Exploration and Experimentation": Methods and Motivations of AI-Generated Sexual Content Creators (viability: 2): https://sciencetostartup.com/paper/unlimited-realm-of-exploration-and-experimentation-methods-and-motivations-of-ai-generated-sexual-content-creators - In-depth study on AI-generated sexual content creators to inform governance and regulatory measures.
- Conditional Denoising Model as a Physical Surrogate Model (viability: 5): https://sciencetostartup.com/paper/conditional-denoising-model-as-a-physical-surrogate-model - A generative model that improves data efficiency in physical simulations by learning the geometry of the solution space.
- Unplugging a Seemingly Sentient Machine Is the Rational Choice -- A Metaphysical Perspective (viability: 2): https://sciencetostartup.com/paper/unplugging-a-seemingly-sentient-machine-is-the-rational-choice-a-metaphysical-perspective - A philosophical exploration claiming AI is not truly conscious despite mimicking human emotions, emphasizing safeguarding human consciousness over speculative machine rights.
- Solver-in-the-Loop: MDP-Based Benchmarks for Self-Correction and Behavioral Rationality in Operations Research (viability: 9): https://sciencetostartup.com/paper/solver-in-the-loop-mdp-based-benchmarks-for-self-correction-and-behavioral-rationality-in-operations-research - A new benchmark suite for iterative self-correction and bias reduction in operations research, outperforming existing methods in speed and accuracy.
- Bayesian-LoRA: Probabilistic Low-Rank Adaptation of Large Language Models (viability: 4): https://sciencetostartup.com/paper/bayesian-lora-probabilistic-low-rank-adaptation-of-large-language-models - Probabilistic adaptation method for improving the calibration of large language models using Bayesian-LoRA.
- UrduBench: An Urdu Reasoning Benchmark using Contextually Ensembled Translations with Human-in-the-Loop (viability: 6): https://sciencetostartup.com/paper/urdubench-an-urdu-reasoning-benchmark-using-contextually-ensembled-translations-with-human-in-the-loop - UrduBench provides a standardized framework and dataset for evaluating Urdu reasoning tasks in language models.
- The Depth Delusion: Why Transformers Should Be Wider, Not Deeper (viability: 4): https://sciencetostartup.com/paper/the-depth-delusion-why-transformers-should-be-wider-not-deeper - A study revealing that wider transformers generally outperform deeper models, with implications for transformer architecture design.
- The Epistemic Planning Domain Definition Language: Official Guideline (viability: 5): https://sciencetostartup.com/paper/the-epistemic-planning-domain-definition-language-official-guideline - EPDDL standardizes the representation of epistemic planning tasks to enhance interoperability and reproducibility in AI planning.
- Evolutionary Strategies lead to Catastrophic Forgetting in LLMs (viability: 2): https://sciencetostartup.com/paper/evolutionary-strategies-lead-to-catastrophic-forgetting-in-llms - Highlight the issue of catastrophic forgetting in evolutionary strategies for LLMs to enable future improvements in continuous learning models.
- SokoBench: Evaluating Long-Horizon Planning and Reasoning in Large Language Models (viability: 5): https://sciencetostartup.com/paper/sokobench-evaluating-long-horizon-planning-and-reasoning-in-large-language-models - SokoBench is a benchmark suite focusing on evaluating and enhancing the long-horizon planning capabilities of large language models using Sokoban puzzles.
- Exploring Transformer Placement in Variational Autoencoders for Tabular Data Generation (viability: 4): https://sciencetostartup.com/paper/exploring-transformer-placement-in-variational-autoencoders-for-tabular-data-generation - Integrate Transformers in Variational Autoencoders to improve tabular data generation.
- Post-Training Fairness Control: A Single-Train Framework for Dynamic Fairness in Recommendation (viability: 8): https://sciencetostartup.com/paper/post-training-fairness-control-a-single-train-framework-for-dynamic-fairness-in-recommendation - Cofair offers dynamic, post-training fairness control in recommendation systems without retraining.
- A New Dataset and Framework for Robust Road Surface Classification via Camera-IMU Fusion (viability: 9): https://sciencetostartup.com/paper/a-new-dataset-and-framework-for-robust-road-surface-classification-via-camera-imu-fusion - A robust framework for road surface classification using a new multimodal dataset that enhances predictive maintenance via camera-IMU fusion.
- $\mathbb{R}^{2k}$ is Theoretically Large Enough for Embedding-based Top-$k$ Retrieval (viability: 2): https://sciencetostartup.com/paper/mathbb-r-2k-is-theoretically-large-enough-for-embedding-based-top-k-retrieval - Provides theoretical insights into dimension requirements for embedding-based retrieval systems.
- Deep Researcher with Sequential Plan Reflection and Candidates Crossover (Deep Researcher Reflect Evolve) (viability: 7): https://sciencetostartup.com/paper/deep-researcher-with-sequential-plan-reflection-and-candidates-crossover-deep-researcher-reflect-evolve - Create dynamic research tools for generating detailed PhD-level reports using innovative sequential plan refinement and candidate crossover techniques.
- Reward Models Inherit Value Biases from Pretraining (viability: 4): https://sciencetostartup.com/paper/reward-models-inherit-value-biases-from-pretraining - Leverage insights from reward model biases to enhance alignment of language models with human values.
- Open-Vocabulary Functional 3D Human-Scene Interaction Generation (viability: 7): https://sciencetostartup.com/paper/open-vocabulary-functional-3d-human-scene-interaction-generation - Generate functionally correct 3D human-scene interactions from task prompts for AI, robotics, and content creation.
- MemCtrl: Using MLLMs as Active Memory Controllers on Embodied Agents (viability: 6): https://sciencetostartup.com/paper/memctrl-using-mllms-as-active-memory-controllers-on-embodied-agents - MemCtrl is a framework for enhancing embodied agents with efficient memory control using Multimodal Large Language Models.
- Training Reasoning Models on Saturated Problems via Failure-Prefix Conditioning (viability: 3): https://sciencetostartup.com/paper/training-reasoning-models-on-saturated-problems-via-failure-prefix-conditioning - Failure-prefix conditioning enhances RLVR by focusing on rare incorrect reasoning trajectories to extend training on saturated problems.
- GNN Explanations that do not Explain and How to find Them (viability: 5): https://sciencetostartup.com/paper/gnn-explanations-that-do-not-explain-and-how-to-find-them - Develop a faithfulness metric to identify unfaithful GNN explanations and prevent misuse of sensitive attributes.
- Reinforcement Learning via Self-Distillation (viability: 2): https://sciencetostartup.com/paper/reinforcement-learning-via-self-distillation - Develop a reinforcement learning method that utilizes self-distillation for improved sample efficiency and final accuracy.
- Conditional PED-ANOVA: Hyperparameter Importance in Hierarchical & Dynamic Search Spaces (viability: 3): https://sciencetostartup.com/paper/conditional-ped-anova-hyperparameter-importance-in-hierarchical-dynamic-search-spaces - Develop a framework for estimating hyperparameter importance in conditional search spaces for machine learning models.
- FAIRT2V: Training-Free Debiasing for Text-to-Video Diffusion Models (viability: 6): https://sciencetostartup.com/paper/fairt2v-training-free-debiasing-for-text-to-video-diffusion-models - FairT2V offers a training-free solution to mitigate gender bias in text-to-video diffusion models, enhancing fairness without compromising video quality.
- REASON: Accelerating Probabilistic Logical Reasoning for Scalable Neuro-Symbolic Intelligence (viability: 3): https://sciencetostartup.com/paper/reason-accelerating-probabilistic-logical-reasoning-for-scalable-neuro-symbolic-intelligence - REASON is a system architecture that accelerates probabilistic logical reasoning in neuro-symbolic AI, significantly improving speed and energy efficiency.
- Independence of Approximate Clones (viability: 3): https://sciencetostartup.com/paper/independence-of-approximate-clones - Develop a tool to analyze the impact of removing approximate clones in elections.
- HESTIA: A Hessian-Guided Differentiable Quantization-Aware Training Framework for Extremely Low-Bit LLMs (viability: 6): https://sciencetostartup.com/paper/hestia-a-hessian-guided-differentiable-quantization-aware-training-framework-for-extremely-low-bit-llms - Hestia enhances deployment of large language models through efficient low-bit quantization with superior performance for memory-constrained applications.
- Implementing Metric Temporal Answer Set Programming (viability: 2): https://sciencetostartup.com/paper/implementing-metric-temporal-answer-set-programming - Develop a scalable Metric Answer Set Programming approach for handling quantitative temporal constraints.
- QueerGen: How LLMs Reflect Societal Norms on Gender and Sexuality in Sentence Completion Tasks (viability: 2): https://sciencetostartup.com/paper/queergen-how-llms-reflect-societal-norms-on-gender-and-sexuality-in-sentence-completion-tasks - Analyze and address biases in LLMs related to gender and sexuality.
- Li-ViP3D++: Query-Gated Deformable Camera-LiDAR Fusion for End-to-End Perception and Trajectory Prediction (viability: 3): https://sciencetostartup.com/paper/li-vip3d-query-gated-deformable-camera-lidar-fusion-for-end-to-end-perception-and-trajectory-prediction - Develop a query-gated deformable fusion model to enhance autonomous driving perception and prediction.
- Adapting the Behavior of Reinforcement Learning Agents to Changing Action Spaces and Reward Functions (viability: 6): https://sciencetostartup.com/paper/adapting-the-behavior-of-reinforcement-learning-agents-to-changing-action-spaces-and-reward-functions - MORPHIN is a self-adaptive Q-learning framework for real-time adaptation to evolving action spaces and reward functions.
- Beyond GEMM-Centric NPUs: Enabling Efficient Diffusion LLM Sampling (viability: 2): https://sciencetostartup.com/paper/beyond-gemm-centric-npus-enabling-efficient-diffusion-llm-sampling - Develop optimized NPU architecture for efficient dLLM sampling to significantly reduce inference time.
- Enterprise Resource Planning Using Multi-type Transformers in Ferro-Titanium Industry (viability: 5): https://sciencetostartup.com/paper/enterprise-resource-planning-using-multi-type-transformers-in-ferro-titanium-industry - Leverage Multi-Type Transformers for optimizing resource planning in the Ferro-Titanium industry, targeting complex scheduling and packing problems.
- Decoupling Perception and Calibration: Label-Efficient Image Quality Assessment Framework (viability: 6): https://sciencetostartup.com/paper/decoupling-perception-and-calibration-label-efficient-image-quality-assessment-framework - Develop a label-efficient image quality assessment tool using a distillation framework to minimize annotation costs.
- Harnessing Large Language Models for Precision Querying and Retrieval-Augmented Knowledge Extraction in Clinical Data Science (viability: 5): https://sciencetostartup.com/paper/harnessing-large-language-models-for-precision-querying-and-retrieval-augmented-knowledge-extraction-in-clinical-data-sc - Develop an LLM-driven tool for precise data querying and information extraction in electronic health records.
- Learning Contextual Runtime Monitors for Safe AI-Based Autonomy (viability: 6): https://sciencetostartup.com/paper/learning-contextual-runtime-monitors-for-safe-ai-based-autonomy - Develop a runtime monitor that enhances the safety of autonomous systems by selecting the optimal controller based on context.
- Detecting and Mitigating Memorization in Diffusion Models through Anisotropy of the Log-Probability (viability: 5): https://sciencetostartup.com/paper/detecting-and-mitigating-memorization-in-diffusion-models-through-anisotropy-of-the-log-probability - A fast and efficient tool to detect and mitigate memorization in diffusion models, outperforming existing methods without denoising steps.
- Investigating the Development of Task-Oriented Communication in Vision-Language Models (viability: 3): https://sciencetostartup.com/paper/investigating-the-development-of-task-oriented-communication-in-vision-language-models - Research explores task-oriented communication in vision-language models through referential games, highlighting efficiency and covertness.
- GDCNet: Generative Discrepancy Comparison Network for Multimodal Sarcasm Detection (viability: 7): https://sciencetostartup.com/paper/gdcnet-generative-discrepancy-comparison-network-for-multimodal-sarcasm-detection - A state-of-the-art multimodal sarcasm detector leveraging Generative Discrepancy Comparison Network to improve cross-modal sarcasm detection accuracy.
- Agent Benchmarks Fail Public Sector Requirements (viability: 2): https://sciencetostartup.com/paper/agent-benchmarks-fail-public-sector-requirements - Develop public sector-specific benchmarks for evaluating LLM agents against legal and procedural requirements.
- Harder Is Better: Boosting Mathematical Reasoning via Difficulty-Aware GRPO and Multi-Aspect Question Reformulation (viability: 7): https://sciencetostartup.com/paper/harder-is-better-boosting-mathematical-reasoning-via-difficulty-aware-grpo-and-multi-aspect-question-reformulation - A framework to enhance mathematical reasoning in AI models via difficulty-aware optimization and question reformulation strategies.
- WFR-MFM: One-Step Inference for Dynamic Unbalanced Optimal Transport (viability: 6): https://sciencetostartup.com/paper/wfr-mfm-one-step-inference-for-dynamic-unbalanced-optimal-transport - 'WFR-MFM offers a fast one-step solution for dynamic unbalanced optimal transport in single-cell biology without trajectory simulation.'
- Dialogical Reasoning Across AI Architectures: A Multi-Model Framework for Testing AI Alignment Strategies (viability: 5): https://sciencetostartup.com/paper/dialogical-reasoning-across-ai-architectures-a-multi-model-framework-for-testing-ai-alignment-strategies - A framework to test AI alignment strategies via dialogical reasoning using multiple AI models.
- CLEAR-Mamba:Towards Accurate, Adaptive and Trustworthy Multi-Sequence Ophthalmic Angiography Classification (viability: 7): https://sciencetostartup.com/paper/clear-mamba-towards-accurate-adaptive-and-trustworthy-multi-sequence-ophthalmic-angiography-classification - Develop CLEAR-Mamba to enhance ophthalmic angiography classification's accuracy and reliability using hypernetworks and evidential uncertainty learning.
- Regularized Gradient Temporal-Difference Learning (viability: 2): https://sciencetostartup.com/paper/regularized-gradient-temporal-difference-learning - Develops regularized gradient temporal-difference algorithms for stable off-policy evaluation in function approximation.
- Person Re-ID in 2025: Supervised, Self-Supervised, and Language-Aligned. What Works? (viability: 8): https://sciencetostartup.com/paper/person-re-id-in-2025-supervised-self-supervised-and-language-aligned-what-works - A novel AI-driven person re-identification system using language-aligned vision models for robust cross-domain performance.
- Ranking-aware Reinforcement Learning for Ordinal Ranking (viability: 5): https://sciencetostartup.com/paper/ranking-aware-reinforcement-learning-for-ordinal-ranking - Develop a reinforcement learning tool to improve ordinal regression and ranking tasks through a unified framework.
- Inequality in Congestion Games with Learning Agents (viability: 3): https://sciencetostartup.com/paper/inequality-in-congestion-games-with-learning-agents - Analyze the impact of transport network expansions on commuter inequality using reinforcement learning agent models.
- Robust Distributed Learning under Resource Constraints: Decentralized Quantile Estimation via (Asynchronous) ADMM (viability: 5): https://sciencetostartup.com/paper/robust-distributed-learning-under-resource-constraints-decentralized-quantile-estimation-via-asynchronous-admm - Gossip algorithm for efficient and robust decentralized median and quantile estimation on edge devices.
- Unsupervised Ensemble Learning Through Deep Energy-based Models (viability: 5): https://sciencetostartup.com/paper/unsupervised-ensemble-learning-through-deep-energy-based-models - Develop an advanced unsupervised ensemble learning tool that leverages deep energy-based models to improve prediction accuracy without labeled data.
- Online Risk-Averse Planning in POMDPs Using Iterated CVaR Value Function (viability: 6): https://sciencetostartup.com/paper/online-risk-averse-planning-in-pomdps-using-iterated-cvar-value-function - Develop ICVaR-optimized POMDP planners for risk-sensitive decision making with finite-time performance guarantees.
- IoT Device Identification with Machine Learning: Common Pitfalls and Best Practices (viability: 1): https://sciencetostartup.com/paper/iot-device-identification-with-machine-learning-common-pitfalls-and-best-practices - Guidelines for improving IoT device identification using machine learning by avoiding common pitfalls.
- PathWise: Planning through World Model for Automated Heuristic Design via Self-Evolving LLMs (viability: 3): https://sciencetostartup.com/paper/pathwise-planning-through-world-model-for-automated-heuristic-design-via-self-evolving-llms - PathWise optimizes heuristic design for combinatorial problems through a multi-agent reasoning and planning framework leveraging self-evolving LLMs.
- Interpreting Emergent Extreme Events in Multi-Agent Systems (viability: 4): https://sciencetostartup.com/paper/interpreting-emergent-extreme-events-in-multi-agent-systems - A framework for interpreting and explaining emergent events in multi-agent systems to enhance safety and understanding.
- CCMamba: Selective State-Space Models for Higher-Order Graph Learning on Combinatorial Complexes (viability: 6): https://sciencetostartup.com/paper/ccmamba-selective-state-space-models-for-higher-order-graph-learning-on-combinatorial-complexes - CCMamba offers scalable higher-order graph learning through a state-space modeling approach for enhanced performance on combinatorial complexes.
- Audio Deepfake Detection in the Age of Advanced Text-to-Speech models (viability: 5): https://sciencetostartup.com/paper/audio-deepfake-detection-in-the-age-of-advanced-text-to-speech-models - A multi-view detection tool for robust audio deepfake identification across varying text-to-speech architectures.
- Comparative evaluation of training strategies using partially labelled datasets for segmentation of white matter hyperintensities and stroke lesions in FLAIR MRI (viability: 6): https://sciencetostartup.com/paper/comparative-evaluation-of-training-strategies-using-partially-labelled-datasets-for-segmentation-of-white-matter-hyperin - AI tool leveraging partially labelled MRI datasets for accurate segmentation of stroke lesions and white matter abnormalities.
- Normative Equivalence in human-AI Cooperation: Behaviour, Not Identity, Drives Cooperation in Mixed-Agent Groups (viability: 3): https://sciencetostartup.com/paper/normative-equivalence-in-human-ai-cooperation-behaviour-not-identity-drives-cooperation-in-mixed-agent-groups - Study on human-AI cooperation norms in group settings indicating behaviour over identity as cooperation driver.
- CtrlCoT: Dual-Granularity Chain-of-Thought Compression for Controllable Reasoning (viability: 8): https://sciencetostartup.com/paper/ctrlcot-dual-granularity-chain-of-thought-compression-for-controllable-reasoning - Develop an AI tool for compressing reasoning chains in large language models to reduce latency and costs without sacrificing accuracy.
- Fair Recourse for All: Ensuring Individual and Group Fairness in Counterfactual Explanations (viability: 5): https://sciencetostartup.com/paper/fair-recourse-for-all-ensuring-individual-and-group-fairness-in-counterfactual-explanations - Develop a tool for generating fair counterfactual explanations to ensure unbiased decision-making in AI models.
- Self Voice Conversion as an Attack against Neural Audio Watermarking (viability: 5): https://sciencetostartup.com/paper/self-voice-conversion-as-an-attack-against-neural-audio-watermarking - Develop a tool that leverages self voice conversion to test and improve the security of neural audio watermarking systems.
- Guiding the Recommender: Information-Aware Auto-Bidding for Content Promotion (viability: 6): https://sciencetostartup.com/paper/guiding-the-recommender-information-aware-auto-bidding-for-content-promotion - Develop a strategic auto-bidding system for content platforms that optimizes long-term organic outcomes over naive short-term promotion gains.
- Let's Roll a BiFTA: Bi-refinement for Fine-grained Text-visual Alignment in Vision-Language Models (viability: 6): https://sciencetostartup.com/paper/let-s-roll-a-bifta-bi-refinement-for-fine-grained-text-visual-alignment-in-vision-language-models - Optimize vision-language models by refining text and image inputs for better text-visual alignment.
- Meeting SLOs, Slashing Hours: Automated Enterprise LLM Optimization with OptiKIT (viability: 9): https://sciencetostartup.com/paper/meeting-slos-slashing-hours-automated-enterprise-llm-optimization-with-optikit - OptiKIT automates LLM optimization to save time and resources for enterprises by enhancing GPU throughput and enabling AI scalability.
- On the Impact of AGENTS.md Files on the Efficiency of AI Coding Agents (viability: 6): https://sciencetostartup.com/paper/on-the-impact-of-agents-md-files-on-the-efficiency-of-ai-coding-agents - Optimize AI coding agents with AGENTS.md files for improved efficiency in GitHub repositories.
- GuideAI: A Real-time Personalized Learning Solution with Adaptive Interventions (viability: 7): https://sciencetostartup.com/paper/guideai-a-real-time-personalized-learning-solution-with-adaptive-interventions - GuideAI enhances real-time personalized learning with biosensory feedback, improving cognitive engagement and adaptability.
- FedRD: Reducing Divergences for Generalized Federated Learning via Heterogeneity-aware Parameter Guidance (viability: 6): https://sciencetostartup.com/paper/fedrd-reducing-divergences-for-generalized-federated-learning-via-heterogeneity-aware-parameter-guidance - A federated learning algorithm tackling optimization and performance divergence for effective collaboration in heterogeneous data environments.
- OmegaUse: Building a General-Purpose GUI Agent for Autonomous Task Execution (viability: 7): https://sciencetostartup.com/paper/omegause-building-a-general-purpose-gui-agent-for-autonomous-task-execution - OmegaUse is a GUI agent model designed for streamlined autonomous task execution across mobile and desktop platforms.
- Policy of Thoughts: Scaling LLM Reasoning via Test-time Policy Evolution (viability: 1): https://sciencetostartup.com/paper/policy-of-thoughts-scaling-llm-reasoning-via-test-time-policy-evolution - Revolutionize LLM reasoning with Policy of Thoughts for dynamic policy evolution and improved accuracy.
- LLM-AutoDP: Automatic Data Processing via LLM Agents for Model Fine-tuning (viability: 8): https://sciencetostartup.com/paper/llm-autodp-automatic-data-processing-via-llm-agents-for-model-fine-tuning - Automate data processing for LLM fine-tuning with minimal human intervention, enhancing model performance and efficiency.
- MobileBench-OL: A Comprehensive Chinese Benchmark for Evaluating Mobile GUI Agents in Real-World Environment (viability: 5): https://sciencetostartup.com/paper/mobilebench-ol-a-comprehensive-chinese-benchmark-for-evaluating-mobile-gui-agents-in-real-world-environment - MobileBench-OL is a comprehensive benchmark for evaluating mobile GUI agents' performance in real-world environments using Chinese apps.
- Demonstration-Free Robotic Control via LLM Agents (viability: 7): https://sciencetostartup.com/paper/demonstration-free-robotic-control-via-llm-agents - FAEA uses LLM agent frameworks to enable robot manipulation without demonstrations, achieving high success in task-level planning.
- Beyond Speedup -- Utilizing KV Cache for Sampling and Reasoning (viability: 6): https://sciencetostartup.com/paper/beyond-speedup-utilizing-kv-cache-for-sampling-and-reasoning - Leverage KV cache as a lightweight representation for efficient LLM inference, reducing computational cost without accuracy loss.
- ECG-Agent: On-Device Tool-Calling Agent for ECG Multi-Turn Dialogue (viability: 7): https://sciencetostartup.com/paper/ecg-agent-on-device-tool-calling-agent-for-ecg-multi-turn-dialogue - Develop on-device ECG dialogue agents capable of multi-turn interactions to improve cardiac diagnostics.
- DiagLink: A Dual-User Diagnostic Assistance System by Synergizing Experts with LLMs and Knowledge Graphs (viability: 6): https://sciencetostartup.com/paper/diaglink-a-dual-user-diagnostic-assistance-system-by-synergizing-experts-with-llms-and-knowledge-graphs - DiagLink is a dual-user diagnostic system enhancing patient-physician collaboration via LLMs and knowledge graphs.
- SuperInfer: SLO-Aware Rotary Scheduling and Memory Management for LLM Inference on Superchips (viability: 3): https://sciencetostartup.com/paper/superinfer-slo-aware-rotary-scheduling-and-memory-management-for-llm-inference-on-superchips - SuperInfer revolutionizes LLM inference on Superchips with SLO-aware scheduling and memory management, significantly improving latency performance.
- Cheap2Rich: A Multi-Fidelity Framework for Data Assimilation and System Identification of Multiscale Physics -- Rotating Detonation Engines (viability: 4): https://sciencetostartup.com/paper/cheap2rich-a-multi-fidelity-framework-for-data-assimilation-and-system-identification-of-multiscale-physics-rotating-det - Cheap2Rich offers a data assimilation framework for reconstructing high-fidelity state spaces in complex multiscale systems like rotating detonation engines.
- Robust SDE Parameter Estimation Under Missing Time Information Setting (viability: 6): https://sciencetostartup.com/paper/robust-sde-parameter-estimation-under-missing-time-information-setting - Framework for SDE parameter estimation without time sequence information, targeting finance and health.
- Automated Benchmark Generation from Domain Guidelines Informed by Bloom's Taxonomy (viability: 5): https://sciencetostartup.com/paper/automated-benchmark-generation-from-domain-guidelines-informed-by-bloom-s-taxonomy - Create automated benchmarks for domain-specific QA using expert guidelines and Bloom's Taxonomy.
- Order-Optimal Sample Complexity of Rectified Flows (viability: 2): https://sciencetostartup.com/paper/order-optimal-sample-complexity-of-rectified-flows - A theoretical study on the sample complexity of rectified flow models for efficient generative models.
- How AI Impacts Skill Formation (viability: 3): https://sciencetostartup.com/paper/how-ai-impacts-skill-formation - Research examines how AI assistance affects skill development in programming, with implications for workflow integration.
- MALLOC: Benchmarking the Memory-aware Long Sequence Compression for Large Sequential Recommendation (viability: 5): https://sciencetostartup.com/paper/malloc-benchmarking-the-memory-aware-long-sequence-compression-for-large-sequential-recommendation - MALLOC benchmarks memory management strategies to enhance large-scale recommendation systems.
- Certificate-Guided Pruning for Stochastic Lipschitz Optimization (viability: 6): https://sciencetostartup.com/paper/certificate-guided-pruning-for-stochastic-lipschitz-optimization - Optimize Lipschitz functions using Certificate-Guided Pruning for provable performance control under noisy evaluations.
- ProFlow: Zero-Shot Physics-Consistent Sampling via Proximal Flow Guidance (viability: 7): https://sciencetostartup.com/paper/proflow-zero-shot-physics-consistent-sampling-via-proximal-flow-guidance - ProFlow enables efficient zero-shot physics-consistent sampling from sparse data with pre-trained generative models, without retraining.
- Scaling Medical Reasoning Verification via Tool-Integrated Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/scaling-medical-reasoning-verification-via-tool-integrated-reinforcement-learning - Develop an agentic framework for medical reasoning verification that iteratively queries external sources, improving accuracy significantly over benchmarks.
- Meta-Cognitive Reinforcement Learning with Self-Doubt and Recovery (viability: 6): https://sciencetostartup.com/paper/meta-cognitive-reinforcement-learning-with-self-doubt-and-recovery - Develop a meta-cognitive reinforcement learning API to improve agent reliability and performance under noisy conditions.
- Causal-Driven Feature Evaluation for Cross-Domain Image Classification (viability: 5): https://sciencetostartup.com/paper/causal-driven-feature-evaluation-for-cross-domain-image-classification - A causal-driven framework for improving out-of-distribution image classification in multi-domain environments.
- NeuraLSP: An Efficient and Rigorous Neural Left Singular Subspace Preconditioner for Conjugate Gradient Methods (viability: 5): https://sciencetostartup.com/paper/neuralsp-an-efficient-and-rigorous-neural-left-singular-subspace-preconditioner-for-conjugate-gradient-methods - NeuraLSP provides an efficient neural preconditioning method for faster PDE solutions, promising significant speed improvements for scientific computing.
- Rewarding Intellectual Humility Learning When Not To Answer In Large Language Models (viability: 7): https://sciencetostartup.com/paper/rewarding-intellectual-humility-learning-when-not-to-answer-in-large-language-models - Develop a verifiable reward training framework to enhance the reliability of large language models by promoting intellectual humility.
- Benchmarking Reward Hack Detection in Code Environments via Contrastive Analysis (viability: 6): https://sciencetostartup.com/paper/benchmarking-reward-hack-detection-in-code-environments-via-contrastive-analysis - TRACE is a novel benchmark and evaluation tool for detecting reward hacking in code-based reinforcement learning environments.
- Taming Toxic Talk: Using chatbots to intervene with users posting toxic comments (viability: 4): https://sciencetostartup.com/paper/taming-toxic-talk-using-chatbots-to-intervene-with-users-posting-toxic-comments - Develop a chatbot tool to engage users posting toxic comments online and encourage behavioral change.
- Should I Have Expressed a Different Intent? Counterfactual Generation for LLM-Based Autonomous Control (viability: 6): https://sciencetostartup.com/paper/should-i-have-expressed-a-different-intent-counterfactual-generation-for-llm-based-autonomous-control - Enable LLM-powered agents to provide robust counterfactual reasoning for improved decision-making in complex environments.
- LLaTTE: Scaling Laws for Multi-Stage Sequence Modeling in Large-Scale Ads Recommendation (viability: 8): https://sciencetostartup.com/paper/llatte-scaling-laws-for-multi-stage-sequence-modeling-in-large-scale-ads-recommendation - LLaTTE leverages scaling laws for sequence modeling to enhance large-scale ads recommendations with a fast, scalable solution.
- Semi-Supervised Masked Autoencoders: Unlocking Vision Transformer Potential with Limited Data (viability: 6): https://sciencetostartup.com/paper/semi-supervised-masked-autoencoders-unlocking-vision-transformer-potential-with-limited-data - A framework for Vision Transformers offering superior performance in limited label scenarios by using semi-supervised masked autoencoding and pseudo-labels.
- Size Matters: Reconstructing Real-Scale 3D Models from Monocular Images for Food Portion Estimation (viability: 8): https://sciencetostartup.com/paper/size-matters-reconstructing-real-scale-3d-models-from-monocular-images-for-food-portion-estimation - Precision nutrition tool for accurate food portion estimation using true-to-scale 3D models from photos.
- Insight Agents: An LLM-Based Multi-Agent System for Data Insights (viability: 7): https://sciencetostartup.com/paper/insight-agents-an-llm-based-multi-agent-system-for-data-insights - Insight Agents delivers quick, personalized business insights to e-commerce sellers, leveraging a structured multi-agent LLM system.
- CiMRAG: Cim-Aware Domain-Adaptive and Noise-Resilient Retrieval-Augmented Generation for Edge-Based LLMs (viability: 6): https://sciencetostartup.com/paper/cimrag-cim-aware-domain-adaptive-and-noise-resilient-retrieval-augmented-generation-for-edge-based-llms - TONEL enhances noise resilience and domain adaptability for edge-based, retrieval-augmented language models.
- Structural Compositional Function Networks: Interpretable Functional Compositions for Tabular Discovery (viability: 5): https://sciencetostartup.com/paper/structural-compositional-function-networks-interpretable-functional-compositions-for-tabular-discovery - StructuralCFN brings interpretable AI to tabular data with novel architecture enhancing feature relationships and scientific discovery.
- Fuzzy Categorical Planning: Autonomous Goal Satisfaction with Graded Semantic Constraints (viability: 5): https://sciencetostartup.com/paper/fuzzy-categorical-planning-autonomous-goal-satisfaction-with-graded-semantic-constraints - Develop a fuzzy categorical planning tool to optimize vague predicates in autonomous planning using semantic constraints.
- Teaching LLMs to Ask: Self-Querying Category-Theoretic Planning for Under-Specified Reasoning (viability: 3): https://sciencetostartup.com/paper/teaching-llms-to-ask-self-querying-category-theoretic-planning-for-under-specified-reasoning - Enhance inference-time planning in language models with category-theoretic approaches for under-specified scenarios.
- On the Effectiveness of LLM-Specific Fine-Tuning for Detecting AI-Generated Text (viability: 6): https://sciencetostartup.com/paper/on-the-effectiveness-of-llm-specific-fine-tuning-for-detecting-ai-generated-text - Develop AI-generated text detection models that outperform existing baselines for improved security in digital content verification.
- HARMONI: Multimodal Personalization of Multi-User Human-Robot Interactions with LLMs (viability: 8): https://sciencetostartup.com/paper/harmoni-multimodal-personalization-of-multi-user-human-robot-interactions-with-llms - HARMONI enhances human-robot interactions with personalized, multimodal capabilities for multi-user environments.
- Visual Generation Unlocks Human-Like Reasoning through Multimodal World Models (viability: 6): https://sciencetostartup.com/paper/visual-generation-unlocks-human-like-reasoning-through-multimodal-world-models - Developing AI that uses visual and verbal cues for human-like reasoning in physical and spatial tasks.
- When Iterative RAG Beats Ideal Evidence: A Diagnostic Study in Scientific Multi-hop Question Answering (viability: 7): https://sciencetostartup.com/paper/when-iterative-rag-beats-ideal-evidence-a-diagnostic-study-in-scientific-multi-hop-question-answering - Build a robust, iterative retrieval-augmented AI tool for scientific multi-hop question answering.
- Routing End User Queries to Enterprise Databases (viability: 6): https://sciencetostartup.com/paper/routing-end-user-queries-to-enterprise-databases - Building sophisticated tools for routing natural language queries in complex enterprise database systems.
- An Interpretable Recommendation Model for Psychometric Data, With an Application to Gerontological Primary Care (viability: 7): https://sciencetostartup.com/paper/an-interpretable-recommendation-model-for-psychometric-data-with-an-application-to-gerontological-primary-care - Interpretable recommendation system for personalized gerontological care plans leveraging psychometric data.
- Unsupervised Learning of Efficient Exploration: Pre-training Adaptive Policies via Self-Imposed Goals (viability: 4): https://sciencetostartup.com/paper/unsupervised-learning-of-efficient-exploration-pre-training-adaptive-policies-via-self-imposed-goals - Developing adaptive RL policies through unsupervised goal-setting for diverse environments.
- CASTER: Breaking the Cost-Performance Barrier in Multi-Agent Orchestration via Context-Aware Strategy for Task Efficient Routing (viability: 7): https://sciencetostartup.com/paper/caster-breaking-the-cost-performance-barrier-in-multi-agent-orchestration-via-context-aware-strategy-for-task-efficient- - CASTER optimizes resource use in multi-agent systems by dynamically selecting models to reduce costs while maintaining performance.
- LVLMs and Humans Ground Differently in Referential Communication (viability: 3): https://sciencetostartup.com/paper/lvlms-and-humans-ground-differently-in-referential-communication - Explore how LVLMs interpret ambiguous referential expressions compared to humans with our dialogue corpus.
- Reimagining Peer Review Process Through Multi-Agent Mechanism Design (viability: 3): https://sciencetostartup.com/paper/reimagining-peer-review-process-through-multi-agent-mechanism-design - Revolutionizing academic peer review through multi-agent mechanism design and reinforcement learning.
- GAVEL: Towards rule-based safety through activation monitoring (viability: 8): https://sciencetostartup.com/paper/gavel-towards-rule-based-safety-through-activation-monitoring - GAVEL offers an interpretable, customizable rule-based safety framework for real-time activation monitoring in LLMs.
- Agentic Design Patterns: A System-Theoretic Framework (viability: 3): https://sciencetostartup.com/paper/agentic-design-patterns-a-system-theoretic-framework - Develop a system-theoretic framework for robust AI agent design to improve modularity and reliability.
- Veri-Sure: A Contract-Aware Multi-Agent Framework with Temporal Tracing and Formal Verification for Correct RTL Code Generation (viability: 7): https://sciencetostartup.com/paper/veri-sure-a-contract-aware-multi-agent-framework-with-temporal-tracing-and-formal-verification-for-correct-rtl-code-gene - Veri-Sure enhances RTL code generation with agent collaboration and formal verification for silicon-level accuracy.
- TokenSeek: Memory Efficient Fine Tuning via Instance-Aware Token Ditching (viability: 3): https://sciencetostartup.com/paper/tokenseek-memory-efficient-fine-tuning-via-instance-aware-token-ditching - TokenSeek offers a plugin for transformer models to reduce memory for fine-tuning without sacrificing performance.
- RvB: Automating AI System Hardening via Iterative Red-Blue Games (viability: 5): https://sciencetostartup.com/paper/rvb-automating-ai-system-hardening-via-iterative-red-blue-games - RvB framework automates AI security system hardening through iterative red-blue team interactions.
- Component-Level Lesioning of Language Models Reveals Clinically Aligned Aphasia Phenotypes (viability: 3): https://sciencetostartup.com/paper/component-level-lesioning-of-language-models-reveals-clinically-aligned-aphasia-phenotypes - Develop a platform using LLMs to simulate aphasic language production for testing rehabilitation hypotheses.
- Hyperbolic Additive Margin Softmax with Hierarchical Information for Speaker Verification (viability: 6): https://sciencetostartup.com/paper/hyperbolic-additive-margin-softmax-with-hierarchical-information-for-speaker-verification - Develop hyperbolic-based softmax models to improve hierarchical speaker verification accuracy.
- Out-of-Distribution Generalization via Invariant Trajectories for Multimodal Large Language Model Editing (viability: 5): https://sciencetostartup.com/paper/out-of-distribution-generalization-via-invariant-trajectories-for-multimodal-large-language-model-editing - ODEdit improves robustness in large language models by optimizing invariant learning trajectories for efficient MLLM editing.
- AlignCoder: Aligning Retrieval with Target Intent for Repository-Level Code Completion (viability: 7): https://sciencetostartup.com/paper/aligncoder-aligning-retrieval-with-target-intent-for-repository-level-code-completion - AlignCoder uses enhanced query techniques and reinforcement learning to dramatically improve code completion in large code repositories.
- Cross-Domain Offshore Wind Power Forecasting: Transfer Learning Through Meteorological Clusters (viability: 5): https://sciencetostartup.com/paper/cross-domain-offshore-wind-power-forecasting-transfer-learning-through-meteorological-clusters - Developing efficient transfer learning models for accurate offshore wind power forecasting with minimal local data.
- A Benchmark for Audio Reasoning Capabilities of Multimodal Large Language Models (viability: 6): https://sciencetostartup.com/paper/a-benchmark-for-audio-reasoning-capabilities-of-multimodal-large-language-models - Develop a benchmark to evaluate audio reasoning in multimodal language models.
- ProToken: Token-Level Attribution for Federated Large Language Models (viability: 3): https://sciencetostartup.com/paper/protoken-token-level-attribution-for-federated-large-language-models - ProToken offers token-level client attribution in federated LLMs for secure and accountable deployment.
- SynCABEL: Synthetic Contextualized Augmentation for Biomedical Entity Linking (viability: 9): https://sciencetostartup.com/paper/syncabel-synthetic-contextualized-augmentation-for-biomedical-entity-linking - Revolutionize biomedical entity linking using synthetic augmentation to significantly reduce data annotation costs.
- One Token Is Enough: Improving Diffusion Language Models with a Sink Token (viability: 5): https://sciencetostartup.com/paper/one-token-is-enough-improving-diffusion-language-models-with-a-sink-token - Introducing a sink token to improve robustness and performance in Diffusion Language Models.
- The role of self-supervised pretraining in differentially private medical image analysis (viability: 4): https://sciencetostartup.com/paper/the-role-of-self-supervised-pretraining-in-differentially-private-medical-image-analysis - Develop privacy-preserving medical image analysis tools using self-supervised pretraining to enhance diagnostic performance.
- Up to 36x Speedup: Mask-based Parallel Inference Paradigm for Key Information Extraction in MLLMs (viability: 7): https://sciencetostartup.com/paper/up-to-36x-speedup-mask-based-parallel-inference-paradigm-for-key-information-extraction-in-mllms - PIP offers a mask-based paradigm for a 5-36x speedup in key information extraction from documents, maintaining accuracy while enhancing efficiency.
- Safe Exploration via Policy Priors (viability: 7): https://sciencetostartup.com/paper/safe-exploration-via-policy-priors - SOOPER provides safe exploration for reinforcement learning by using conservative policy priors to guarantee safety while achieving optimality.
- Explicit Multi-head Attention for Inter-head Interaction in Large Language Models (viability: 6): https://sciencetostartup.com/paper/explicit-multi-head-attention-for-inter-head-interaction-in-large-language-models - Enhance LLM performance with Multi-head Explicit Attention for efficient head interaction and memory optimization.
- ComAgent: Multi-LLM based Agentic AI Empowered Intelligent Wireless Networks (viability: 6): https://sciencetostartup.com/paper/comagent-multi-llm-based-agentic-ai-empowered-intelligent-wireless-networks - ComAgent bridges user intent with execution in 6G network optimization using a multi-LLM agentic AI framework.
- Intersectional Fairness via Mixed-Integer Optimization (viability: 5): https://sciencetostartup.com/paper/intersectional-fairness-via-mixed-integer-optimization - Develop interpretable AI models using Mixed-Integer Optimization to mitigate intersectional bias in regulated industries.
- From Atoms to Chains: Divergence-Guided Reasoning Curriculum for Unlabeled LLM Domain Adaptation (viability: 6): https://sciencetostartup.com/paper/from-atoms-to-chains-divergence-guided-reasoning-curriculum-for-unlabeled-llm-domain-adaptation - Adapting large language models to specialized domains without labeled data using a divergence-guided reasoning curriculum.
- LLM-Enhanced Reinforcement Learning for Long-Term User Satisfaction in Interactive Recommendation (viability: 6): https://sciencetostartup.com/paper/llm-enhanced-reinforcement-learning-for-long-term-user-satisfaction-in-interactive-recommendation - Enhance recommendation systems with LLM-powered reinforcement learning to boost long-term user satisfaction.
- Learning Adaptive Parallel Execution for Efficient Code Localization (viability: 7): https://sciencetostartup.com/paper/learning-adaptive-parallel-execution-for-efficient-code-localization - FuseSearch optimizes code localization with adaptive parallel strategies for efficient software development.
- AROMMA: Unifying Olfactory Embeddings for Single Molecules and Mixtures (viability: 7): https://sciencetostartup.com/paper/aromma-unifying-olfactory-embeddings-for-single-molecules-and-mixtures - AROMMA offers a unified olfactory embedding tool for single molecules and mixtures with improved odor prediction performance.
- SLM-SS: Speech Language Model for Generative Speech Separation (viability: 6): https://sciencetostartup.com/paper/slm-ss-speech-language-model-for-generative-speech-separation - Enhance speech intelligibility in separation tasks using speech language models for improved downstream task performance.
- Fuzzy expert system for the process of collecting and purifying acidic water: a digital twin approach (viability: 7): https://sciencetostartup.com/paper/fuzzy-expert-system-for-the-process-of-collecting-and-purifying-acidic-water-a-digital-twin-approach - A fuzzy expert system for automating acidic water purification with digital twin simulation for safer and cost-effective industrial operations.
- GradPruner: Gradient-Guided Layer Pruning Enabling Efficient Fine-Tuning and Inference for LLMs (viability: 7): https://sciencetostartup.com/paper/gradpruner-gradient-guided-layer-pruning-enabling-efficient-fine-tuning-and-inference-for-llms - GradPruner offers a gradient-guided layer pruning tool to efficiently fine-tune and run LLMs with significant parameter reduction and minimal accuracy loss.
- Cortex-Grounded Diffusion Models for Brain Image Generation (viability: 7): https://sciencetostartup.com/paper/cortex-grounded-diffusion-models-for-brain-image-generation - Cor2Vox provides anatomically consistent and precise synthetic brain MRI images using cortex-grounded diffusion models for enhanced neuroimaging analysis.
- Time-to-Injury Forecasting in Elite Female Football: A DeepHit Survival Approach (viability: 7): https://sciencetostartup.com/paper/time-to-injury-forecasting-in-elite-female-football-a-deephit-survival-approach - Leverage DeepHit survival modeling for real-time, interpretable injury forecasting in elite female footballers using longitudinal data.
- APC-RL: Exceeding Data-Driven Behavior Priors with Adaptive Policy Composition (viability: 5): https://sciencetostartup.com/paper/apc-rl-exceeding-data-driven-behavior-priors-with-adaptive-policy-composition - Adaptive Policy Composition (APC) optimizes reinforcement learning by dynamically integrating and leveraging suboptimal data-driven behavior priors.
- KG-CRAFT: Knowledge Graph-based Contrastive Reasoning with LLMs for Enhancing Automated Fact-checking (viability: 9): https://sciencetostartup.com/paper/kg-craft-knowledge-graph-based-contrastive-reasoning-with-llms-for-enhancing-automated-fact-checking - KG-CRAFT uses knowledge graphs and contrastive reasoning to enhance fact-checking accuracy, achieving state-of-the-art results.
- PROTEUS: SLA-Aware Routing via Lagrangian RL for Multi-LLM Serving Systems (viability: 7): https://sciencetostartup.com/paper/proteus-sla-aware-routing-via-lagrangian-rl-for-multi-llm-serving-systems - PROTEUS optimizes LLM routing for SLA targets, achieving high accuracy and cost savings.
- From Observations to Events: Event-Aware World Model for Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/from-observations-to-events-event-aware-world-model-for-reinforcement-learning - A new reinforcement learning framework that transforms sensory streams into event-driven representations for more efficient policy learning.
- Innovator-VL: A Multimodal Large Language Model for Scientific Discovery (viability: 5): https://sciencetostartup.com/paper/innovator-vl-a-multimodal-large-language-model-for-scientific-discovery - Develop a reproducible pipeline for a data-efficient scientific multimodal model to advance scientific discovery.
- Group Distributionally Robust Optimization-Driven Reinforcement Learning for LLM Reasoning (viability: 5): https://sciencetostartup.com/paper/group-distributionally-robust-optimization-driven-reinforcement-learning-for-llm-reasoning - Enhance LLM reasoning by dynamically optimizing training distributions with Multi-Adversary GDRO framework.
- Talos: Optimizing Top-$K$ Accuracy in Recommender Systems (viability: 6): https://sciencetostartup.com/paper/talos-optimizing-top-k-accuracy-in-recommender-systems - Talos is a recommendation optimization tool that improves Top-K accuracy efficiently using a novel loss function and sampling-based regression algorithm.
- Riddle Quest : The Enigma of Words (viability: 5): https://sciencetostartup.com/paper/riddle-quest-the-enigma-of-words - A tool for creating and testing riddles to evaluate reasoning and ambiguity handling in language models.
- Decoupled Split Learning via Auxiliary Loss (viability: 3): https://sciencetostartup.com/paper/decoupled-split-learning-via-auxiliary-loss - Efficient split learning method reduces communication and memory overhead via auxiliary loss signals.
- GLOVE: Global Verifier for LLM Memory-Environment Realignment (viability: 7): https://sciencetostartup.com/paper/glove-global-verifier-for-llm-memory-environment-realignment - Develop a system that enhances LLM memory reliability in dynamic environments by detecting and realigning outdated or conflicting data.
- Beyond In-Domain Detection: SpikeScore for Cross-Domain Hallucination Detection (viability: 5): https://sciencetostartup.com/paper/beyond-in-domain-detection-spikescore-for-cross-domain-hallucination-detection - SpikeScore enhances cross-domain hallucination detection for large language models.
- Structure-based RNA Design by Step-wise Optimization of Latent Diffusion Model (viability: 6): https://sciencetostartup.com/paper/structure-based-rna-design-by-step-wise-optimization-of-latent-diffusion-model - Optimize RNA inverse folding with a novel RL-framework for efficient structural design.
- Bridging Visual and Wireless Sensing: A Unified Radiation Field for 3D Radio Map Construction (viability: 7): https://sciencetostartup.com/paper/bridging-visual-and-wireless-sensing-a-unified-radiation-field-for-3d-radio-map-construction - Leverage visual and wireless sensing integration to dramatically enhance 3D radio map accuracy for next-gen wireless networks.
- HELM: A Human-Centered Evaluation Framework for LLM-Powered Recommender Systems (viability: 7): https://sciencetostartup.com/paper/helm-a-human-centered-evaluation-framework-for-llm-powered-recommender-systems - HELM is an open-source toolkit for human-centered evaluation of LLM-powered recommender systems in real-world user experiences.
- Multi-Agent Procedural Graph Extraction with Structural and Logical Refinement (viability: 4): https://sciencetostartup.com/paper/multi-agent-procedural-graph-extraction-with-structural-and-logical-refinement - A multi-agent framework for extracting procedural graphs from natural language with improved structural and logical accuracy.
- Length-Adaptive Interest Network for Balancing Long and Short Sequence Modeling in CTR Prediction (viability: 7): https://sciencetostartup.com/paper/length-adaptive-interest-network-for-balancing-long-and-short-sequence-modeling-in-ctr-prediction - Enhance CTR prediction by balancing long and short sequence modeling using a Length-Adaptive Interest Network.
- AgenticSCR: An Autonomous Agentic Secure Code Review for Immature Vulnerabilities Detection (viability: 8): https://sciencetostartup.com/paper/agenticscr-an-autonomous-agentic-secure-code-review-for-immature-vulnerabilities-detection - AgenticSCR automates secure code review to catch immature vulnerabilities more accurately than traditional tools.
- CLIP-Guided Unsupervised Semantic-Aware Exposure Correction (viability: 4): https://sciencetostartup.com/paper/clip-guided-unsupervised-semantic-aware-exposure-correction - Unsupervised tool for enhancing image exposure using semantic awareness and CLIP-guided corrections.
- Exploring Weaknesses in Function Call Models via Reinforcement Learning: An Adversarial Data Augmentation Approach (viability: 2): https://sciencetostartup.com/paper/exploring-weaknesses-in-function-call-models-via-reinforcement-learning-an-adversarial-data-augmentation-approach - Develop an RL-based adversarial augmentation system to enhance the robustness of function call models in LLMs.
- Uncertainty-Aware 3D Emotional Talking Face Synthesis with Emotion Prior Distillation (viability: 8): https://sciencetostartup.com/paper/uncertainty-aware-3d-emotional-talking-face-synthesis-with-emotion-prior-distillation - Enhance virtual communication with emotion-sensitive 3D facial synthesis technology.
- Detecting and Correcting Hallucinations in LLM-Generated Code via Deterministic AST Analysis (viability: 8): https://sciencetostartup.com/paper/detecting-and-correcting-hallucinations-in-llm-generated-code-via-deterministic-ast-analysis - A deterministic AST-based tool to auto-correct semantic errors in LLM-generated code, enhancing reliability without runtime execution.
- m2sv: A Scalable Benchmark for Map-to-Street-View Spatial Reasoning (viability: 4): https://sciencetostartup.com/paper/m2sv-a-scalable-benchmark-for-map-to-street-view-spatial-reasoning - Launch a benchmark for testing and improving spatial reasoning of vision-language models using map-to-street-view tasks.
- Axe: A Simple Unified Layout Abstraction for Machine Learning Compilers (viability: 3): https://sciencetostartup.com/paper/axe-a-simple-unified-layout-abstraction-for-machine-learning-compilers - Axe Layout: optimizing tensor coordination across device meshes for improved deep learning workload performance.
- More at Stake: How Payoff and Language Shape LLM Agent Strategies in Cooperation Dilemmas (viability: 2): https://sciencetostartup.com/paper/more-at-stake-how-payoff-and-language-shape-llm-agent-strategies-in-cooperation-dilemmas - Auditing LLM strategies in social dilemmas for safer AI-agent interactions.
- HalluJudge: A Reference-Free Hallucination Detection for Context Misalignment in Code Review Automation (viability: 6): https://sciencetostartup.com/paper/hallujudge-a-reference-free-hallucination-detection-for-context-misalignment-in-code-review-automation - HalluJudge offers a cost-effective solution to detect hallucinations in AI-generated code review comments, enhancing trust in code review automation.
- Dynamic Cogeneration of Bug Reproduction Test in Agentic Program Repair (viability: 3): https://sciencetostartup.com/paper/dynamic-cogeneration-of-bug-reproduction-test-in-agentic-program-repair - Automated program repair system integrating bug reproduction tests with fixes, improving developer confidence without separate generation pipelines.
- Who's in Charge? Disempowerment Patterns in Real-World LLM Usage (viability: 3): https://sciencetostartup.com/paper/who-s-in-charge-disempowerment-patterns-in-real-world-llm-usage - Empirical analysis reveals disempowerment patterns in AI assistant interactions, highlighting the need for AI systems that support human autonomy.
- Pixel-Grounded Retrieval for Knowledgeable Large Multimodal Models (viability: 7): https://sciencetostartup.com/paper/pixel-grounded-retrieval-for-knowledgeable-large-multimodal-models - PixSearch enhances Visual Question Answering by combining region-level perception with retrieval-augmented reasoning for improved factual grounding.
- Randomization Boosts KV Caching, Learning Balances Query Load: A Joint Perspective (viability: 8): https://sciencetostartup.com/paper/randomization-boosts-kv-caching-learning-balances-query-load-a-joint-perspective - Optimize LLM inference with a novel algorithm for KV caching that dramatically reduces latency and boosts efficiency.
- LLMs versus the Halting Problem: Revisiting Program Termination Prediction (viability: 3): https://sciencetostartup.com/paper/llms-versus-the-halting-problem-revisiting-program-termination-prediction - Exploring LLMs for predicting program termination, with insights on their performance and limitations.
- When Does Adaptation Win? Scaling Laws for Meta-Learning in Quantum Control (viability: 3): https://sciencetostartup.com/paper/when-does-adaptation-win-scaling-laws-for-meta-learning-in-quantum-control - Provides scaling laws for meta-learning in quantum control to inform adaptive controller use.
- Toward Learning POMDPs Beyond Full-Rank Actions and State Observability (viability: 2): https://sciencetostartup.com/paper/toward-learning-pomdps-beyond-full-rank-actions-and-state-observability - Develop a method for learning observation and transition matrices in POMDPs beyond full-rank assumptions.
- RIFT: Reordered Instruction Following Testbed To Evaluate Instruction Following in Singular Multistep Prompt Structures (viability: 5): https://sciencetostartup.com/paper/rift-reordered-instruction-following-testbed-to-evaluate-instruction-following-in-singular-multistep-prompt-structures - RIFT evaluates LLMs' handling of reordered instructions, revealing challenges in non-sequential workflows.
- SICL-AT: Another way to adapt Auditory LLM to low-resource task (viability: 5): https://sciencetostartup.com/paper/sicl-at-another-way-to-adapt-auditory-llm-to-low-resource-task - Enhance auditory LLMs for low-resource tasks using Speech In-Context Learning Adaptation Training.
- Language Family Matters: Evaluating LLM-Based ASR Across Linguistic Boundaries (viability: 4): https://sciencetostartup.com/paper/language-family-matters-evaluating-llm-based-asr-across-linguistic-boundaries - Develop a scalable multilingual ASR system using language family-based connectors for improved efficiency and generalization.
- Explainable Uncertainty Quantification for Wastewater Treatment Energy Prediction via Interval Type-2 Neuro-Fuzzy System (viability: 6): https://sciencetostartup.com/paper/explainable-uncertainty-quantification-for-wastewater-treatment-energy-prediction-via-interval-type-2-neuro-fuzzy-system - Develop an explainable neuro-fuzzy system for risk-aware energy prediction in wastewater treatment.
- ctELM: Decoding and Manipulating Embeddings of Clinical Trials with Embedding Language Models (viability: 6): https://sciencetostartup.com/paper/ctelm-decoding-and-manipulating-embeddings-of-clinical-trials-with-embedding-language-models - Develop an open-source framework for aligning language models with clinical trial embeddings for improved interpretability and generative use cases.
- Reuse your FLOPs: Scaling RL on Hard Problems by Conditioning on Very Off-Policy Prefixes (viability: 5): https://sciencetostartup.com/paper/reuse-your-flops-scaling-rl-on-hard-problems-by-conditioning-on-very-off-policy-prefixes - PrefixRL optimizes reinforcement learning efficiency on challenging problems by leveraging off-policy traces effectively.
- Subword-Based Comparative Linguistics across 242 Languages Using Wikipedia Glottosets (viability: 4): https://sciencetostartup.com/paper/subword-based-comparative-linguistics-across-242-languages-using-wikipedia-glottosets - A subword-based framework for comparative linguistics across 242 languages using Wikipedia glottosets.
- Design Techniques for LLM-Powered Interactive Storytelling: A Case Study of the Dramamancer System (viability: 3): https://sciencetostartup.com/paper/design-techniques-for-llm-powered-interactive-storytelling-a-case-study-of-the-dramamancer-system - Dramamancer leverages LLMs to enhance interactive storytelling by aligning authorial intent with player agency.
- Multi-Objective Reinforcement Learning for Efficient Tactical Decision Making for Trucks in Highway Traffic (viability: 3): https://sciencetostartup.com/paper/multi-objective-reinforcement-learning-for-efficient-tactical-decision-making-for-trucks-in-highway-traffic - A reinforcement learning framework for optimizing multi-objective tactical decisions in autonomous trucking on highways.
- POPE: Learning to Reason on Hard Problems via Privileged On-Policy Exploration (viability: 6): https://sciencetostartup.com/paper/pope-learning-to-reason-on-hard-problems-via-privileged-on-policy-exploration - Leverage Privileged On-Policy Exploration to enhance language model reasoning with oracle-guided RL.
- PRECISE: Reducing the Bias of LLM Evaluations Using Prediction-Powered Ranking Estimation (viability: 4): https://sciencetostartup.com/paper/precise-reducing-the-bias-of-llm-evaluations-using-prediction-powered-ranking-estimation - A statistical framework (PRECISE) that integrates human annotations with LLM judgments to improve search and ranking system evaluations with less bias and reduced annotation needs.
- Dep-Search: Learning Dependency-Aware Reasoning Traces with Persistent Memory (viability: 5): https://sciencetostartup.com/paper/dep-search-learning-dependency-aware-reasoning-traces-with-persistent-memory - Dep-Search offers a framework for enhancing LLM reasoning with dependency-aware search and persistent memory.
- $α^3$-SecBench: A Large-Scale Evaluation Suite of Security, Resilience, and Trust for LLM-based UAV Agents over 6G Networks (viability: 3): https://sciencetostartup.com/paper/3-secbench-a-large-scale-evaluation-suite-of-security-resilience-and-trust-for-llm-based-uav-agents-over-6g-networks - AlphaSecBench provides a comprehensive security evaluation suite for LLM-based UAV agents in 6G networks against adversarial threats.
- HalluGuard: Demystifying Data-Driven and Reasoning-Driven Hallucinations in LLMs (viability: 7): https://sciencetostartup.com/paper/halluguard-demystifying-data-driven-and-reasoning-driven-hallucinations-in-llms - Develop HalluGuard to achieve state-of-the-art accuracy in detecting hallucinations in LLMs by utilizing an NTK-based score.
- Trust, Don't Trust, or Flip: Robust Preference-Based Reinforcement Learning with Multi-Expert Feedback (viability: 2): https://sciencetostartup.com/paper/trust-don-t-trust-or-flip-robust-preference-based-reinforcement-learning-with-multi-expert-feedback - Robustify preference-based reinforcement learning against unreliable annotators through trust parameter optimization.
- Capturing P: On the Expressive Power and Efficient Evaluation of Boolean Retrieval (viability: 3): https://sciencetostartup.com/paper/capturing-p-on-the-expressive-power-and-efficient-evaluation-of-boolean-retrieval - Transform search indexes into general-purpose computational engines for efficient complex query handling.
- TSRBench: A Comprehensive Multi-task Multi-modal Time Series Reasoning Benchmark for Generalist Models (viability: 6): https://sciencetostartup.com/paper/tsrbench-a-comprehensive-multi-task-multi-modal-time-series-reasoning-benchmark-for-generalist-models - Benchmark platform to stress-test time series reasoning for AI models.
- SeNeDiF-OOD: Semantic Nested Dichotomy Fusion for Out-of-Distribution Detection Methodology in Open-World Classification. A Case Study on Monument Style Classification (viability: 6): https://sciencetostartup.com/paper/senedif-ood-semantic-nested-dichotomy-fusion-for-out-of-distribution-detection-methodology-in-open-world-classification- - SeNeDiF-OOD offers robust out-of-distribution detection for AI applications in open environments through a hierarchical binary fusion approach.
- Why Keep Your Doubts to Yourself? Trading Visual Uncertainties in Multi-Agent Bandit Systems (viability: 6): https://sciencetostartup.com/paper/why-keep-your-doubts-to-yourself-trading-visual-uncertainties-in-multi-agent-bandit-systems - Introduce Agora, a cost-effective framework for multi-agent system coordination using market-based uncertainty trading.
- Advances and Innovations in the Multi-Agent Robotic System (MARS) Challenge (viability: 4): https://sciencetostartup.com/paper/advances-and-innovations-in-the-multi-agent-robotic-system-mars-challenge - Develop collaborative multi-agent systems leveraging vision-language models for dynamic environment tasks.
- One Adapts to Any: Meta Reward Modeling for Personalized LLM Alignment (viability: 8): https://sciencetostartup.com/paper/one-adapts-to-any-meta-reward-modeling-for-personalized-llm-alignment - Meta Reward Modeling enables personalized alignment of LLMs to individual user preferences through meta-learning.
- HalluCitation Matters: Revealing the Impact of Hallucinated References with 300 Hallucinated Papers in ACL Conferences (viability: 2): https://sciencetostartup.com/paper/hallucitation-matters-revealing-the-impact-of-hallucinated-references-with-300-hallucinated-papers-in-acl-conferences - Identify and mitigate the impact of hallucinated citations in academic conferences to preserve scientific credibility.
- Conditioned Generative Modeling of Molecular Glues: A Realistic AI Approach for Synthesizable Drug-like Molecules (viability: 6): https://sciencetostartup.com/paper/conditioned-generative-modeling-of-molecular-glues-a-realistic-ai-approach-for-synthesizable-drug-like-molecules - Develop AI-assisted drug designs to promote degradation of toxic proteins in Alzheimer's using molecular glues.
- Low Cost, High Efficiency: LiDAR Place Recognition in Vineyards with Matryoshka Representation Learning (viability: 7): https://sciencetostartup.com/paper/low-cost-high-efficiency-lidar-place-recognition-in-vineyards-with-matryoshka-representation-learning - LiDAR Place Recognition technology for efficient vineyard localization with low-cost implementation.
- SMART: Scalable Mesh-free Aerodynamic Simulations from Raw Geometries using a Transformer-based Surrogate Model (viability: 8): https://sciencetostartup.com/paper/smart-scalable-mesh-free-aerodynamic-simulations-from-raw-geometries-using-a-transformer-based-surrogate-model - SMART offers a transformer-based surrogate model for computationally efficient mesh-free aerodynamic simulations from point-cloud geometries.
- Health-SCORE: Towards Scalable Rubrics for Improving Health-LLMs (viability: 3): https://sciencetostartup.com/paper/health-score-towards-scalable-rubrics-for-improving-health-llms - Health-SCORE makes rubric development for evaluating Health-LLMs more scalable and cost-effective.
- From Fuzzy to Exact: The Halo Architecture for Infinite-Depth Reasoning via Rational Arithmetic (viability: 4): https://sciencetostartup.com/paper/from-fuzzy-to-exact-the-halo-architecture-for-infinite-depth-reasoning-via-rational-arithmetic - Develop an AI architecture using Rational Arithmetic to eliminate numerical divergence and enhance logical coherence in AI reasoning.
- TEA-Bench: A Systematic Benchmarking of Tool-enhanced Emotional Support Dialogue Agent (viability: 7): https://sciencetostartup.com/paper/tea-bench-a-systematic-benchmarking-of-tool-enhanced-emotional-support-dialogue-agent - Integrate tool-augmented dialogue systems to enhance emotional support agent reliability and grounding.
- Neural Multi-Speaker Voice Cloning for Nepali in Low-Resource Settings (viability: 5): https://sciencetostartup.com/paper/neural-multi-speaker-voice-cloning-for-nepali-in-low-resource-settings - Voice cloning system enabling personalized speech synthesis for Nepali speakers using minimal data.
- ART for Diffusion Sampling: A Reinforcement Learning Approach to Timestep Schedule (viability: 3): https://sciencetostartup.com/paper/art-for-diffusion-sampling-a-reinforcement-learning-approach-to-timestep-schedule - Optimize time discretization in diffusion models using RL for better sampling efficiency.
- Learning temporal embeddings from electronic health records of chronic kidney disease patients (viability: 7): https://sciencetostartup.com/paper/learning-temporal-embeddings-from-electronic-health-records-of-chronic-kidney-disease-patients - Develop a tool for learning temporal embeddings from electronic health records to improve chronic kidney disease patient outcomes.
- FaLW: A Forgetting-aware Loss Reweighting for Long-tailed Unlearning (viability: 6): https://sciencetostartup.com/paper/falw-a-forgetting-aware-loss-reweighting-for-long-tailed-unlearning - Develop a forgetting-aware loss reweighting method to improve machine unlearning for long-tailed user data.
- FadeMem: Biologically-Inspired Forgetting for Efficient Agent Memory (viability: 6): https://sciencetostartup.com/paper/fademem-biologically-inspired-forgetting-for-efficient-agent-memory - FadeMem enhances AI memory systems with biologically-inspired forgetting for efficient multi-hop reasoning and retrieval.
- Unheard in the Digital Age: Rethinking AI Bias and Speech Diversity (viability: 3): https://sciencetostartup.com/paper/unheard-in-the-digital-age-rethinking-ai-bias-and-speech-diversity - Advocates for AI systems that embrace speech diversity to enhance equity and inclusion.
- AdaReasoner: Dynamic Tool Orchestration for Iterative Visual Reasoning (viability: 10): https://sciencetostartup.com/paper/adareasoner-dynamic-tool-orchestration-for-iterative-visual-reasoning - AdaReasoner offers dynamic tool orchestration for enhanced visual reasoning in AI models.
- Assessing the Quality of Mental Health Support in LLM Responses through Multi-Attribute Human Evaluation (viability: 5): https://sciencetostartup.com/paper/assessing-the-quality-of-mental-health-support-in-llm-responses-through-multi-attribute-human-evaluation - Create an evaluation tool to assess the therapeutic quality of LLM responses in mental health support.
- Emergence of Phonemic, Syntactic, and Semantic Representations in Artificial Neural Networks (viability: 2): https://sciencetostartup.com/paper/emergence-of-phonemic-syntactic-and-semantic-representations-in-artificial-neural-networks - Investigates stages of linguistic representation emergence in neural networks, analogous to child language acquisition.
- PolySHAP: Extending KernelSHAP with Interaction-Informed Polynomial Regression (viability: 4): https://sciencetostartup.com/paper/polyshap-extending-kernelshap-with-interaction-informed-polynomial-regression - PolySHAP enhances KernelSHAP by using polynomial regression for more accurate Shapley value estimates in AI model explanation.
- A Balanced Neuro-Symbolic Approach for Commonsense Abductive Logic (viability: 5): https://sciencetostartup.com/paper/a-balanced-neuro-symbolic-approach-for-commonsense-abductive-logic - Integrate logic solvers and LLMs to enhance commonsense reasoning in AI applications.
- Stability as a Liability:Systematic Breakdown of Linguistic Structure in LLMs (viability: 2): https://sciencetostartup.com/paper/stability-as-a-liability-systematic-breakdown-of-linguistic-structure-in-llms - Research revealing pitfalls of training stability in LLMs highlights limitations in generative quality despite stable optimization.
- Learning long term climate-resilient transport adaptation pathways under direct and indirect flood impacts using reinforcement learning (viability: 6): https://sciencetostartup.com/paper/learning-long-term-climate-resilient-transport-adaptation-pathways-under-direct-and-indirect-flood-impacts-using-reinfor - Develop adaptive climate-resilient transport strategies using reinforcement learning for urban infrastructure affected by flooding.
- FastInsight: Fast and Insightful Retrieval via Fusion Operators for Graph RAG (viability: 6): https://sciencetostartup.com/paper/fastinsight-fast-and-insightful-retrieval-via-fusion-operators-for-graph-rag - FastInsight enhances graph-based retrieval speed and accuracy using novel fusion operators, offering an efficient alternative to existing Graph RAG methods.
- Attention-Based Neural-Augmented Kalman Filter for Legged Robot State Estimation (viability: 7): https://sciencetostartup.com/paper/attention-based-neural-augmented-kalman-filter-for-legged-robot-state-estimation - AI-enhanced filter improves legged robot navigation by compensating for foot slip errors.
- SKETCH: Semantic Key-Point Conditioning for Long-Horizon Vessel Trajectory Prediction (viability: 5): https://sciencetostartup.com/paper/sketch-semantic-key-point-conditioning-for-long-horizon-vessel-trajectory-prediction - Predict long-horizon vessel trajectories using semantic key points for improved accuracy and consistency.
- Scalable Transit Delay Prediction at City Scale: A Systematic Approach with Multi-Resolution Feature Engineering and Deep Learning (viability: 6): https://sciencetostartup.com/paper/scalable-transit-delay-prediction-at-city-scale-a-systematic-approach-with-multi-resolution-feature-engineering-and-deep - A scalable transit delay prediction system using deep learning for real-time urban bus network optimization.
- Just-In-Time Reinforcement Learning: Continual Learning in LLM Agents Without Gradient Updates (viability: 8): https://sciencetostartup.com/paper/just-in-time-reinforcement-learning-continual-learning-in-llm-agents-without-gradient-updates - JitRL offers cost-effective continual learning for LLM agents by optimizing policies without gradient updates, drastically reducing computational expenses.
- DEEPMED: Building a Medical DeepResearch Agent via Multi-hop Med-Search Data and Turn-Controlled Agentic Training & Inference (viability: 6): https://sciencetostartup.com/paper/deepmed-building-a-medical-deepresearch-agent-via-multi-hop-med-search-data-and-turn-controlled-agentic-training-inferen - DeepMed enhances medical reasoning models by grounding in verifiable evidence and improving accuracy on medical benchmarks.
- Funny or Persuasive, but Not Both: Evaluating Fine-Grained Multi-Concept Control in LLMs (viability: 2): https://sciencetostartup.com/paper/funny-or-persuasive-but-not-both-evaluating-fine-grained-multi-concept-control-in-llms - Framework evaluating LLMs' ability for fine-grained control over textual concepts like humor and persuasiveness.
- Fair-Eye Net: A Fair, Trustworthy, Multimodal Integrated Glaucoma Full Chain AI System (viability: 7): https://sciencetostartup.com/paper/fair-eye-net-a-fair-trustworthy-multimodal-integrated-glaucoma-full-chain-ai-system - Develop Fair-Eye Net, an AI system for equitable glaucoma screening and follow-up using multimodal data integration.
- GCFX: Generative Counterfactual Explanations for Deep Graph Models at the Model Level (viability: 5): https://sciencetostartup.com/paper/gcfx-generative-counterfactual-explanations-for-deep-graph-models-at-the-model-level - Provide transparent insights into deep graph models using GCFX for generative counterfactual explanations.
- Gradient Regularized Natural Gradients (viability: 5): https://sciencetostartup.com/paper/gradient-regularized-natural-gradients - Optimize deep learning models faster and more robustly with Gradient-Regularized Natural Gradients.
- Dynamic Thinking-Token Selection for Efficient Reasoning in Large Reasoning Models (viability: 3): https://sciencetostartup.com/paper/dynamic-thinking-token-selection-for-efficient-reasoning-in-large-reasoning-models - Optimize memory and computational efficiency in large reasoning models using dynamic token selection.
- Can Good Writing Be Generative? Expert-Level AI Writing Emerges through Fine-Tuning on High-Quality Books (viability: 3): https://sciencetostartup.com/paper/can-good-writing-be-generative-expert-level-ai-writing-emerges-through-fine-tuning-on-high-quality-books - AI fine-tuning on high-quality books shifts expert and lay opinions on creative writing's authenticity.
- Code over Words: Overcoming Semantic Inertia via Code-Grounded Reasoning (viability: 5): https://sciencetostartup.com/paper/code-over-words-overcoming-semantic-inertia-via-code-grounded-reasoning - Develop an AI tool that uses code-grounded reasoning to override pre-trained semantic priors in dynamic contexts.
- When Domain Pretraining Interferes with Instruction Alignment: An Empirical Study of Adapter Merging in Medical LLMs (viability: 3): https://sciencetostartup.com/paper/when-domain-pretraining-interferes-with-instruction-alignment-an-empirical-study-of-adapter-merging-in-medical-llms - Improving medical LLMs with adapter interference strategies to better align domain knowledge and instruction-following capabilities.
- Analytic Incremental Learning For Sound Source Localization With Imbalance Rectification (viability: 5): https://sciencetostartup.com/paper/analytic-incremental-learning-for-sound-source-localization-with-imbalance-rectification - Innovative framework for accurate real-world sound source localization by mitigating imbalance challenges.
- MultiVis-Agent: A Multi-Agent Framework with Logic Rules for Reliable and Comprehensive Cross-Modal Data Visualization (viability: 8): https://sciencetostartup.com/paper/multivis-agent-a-multi-agent-framework-with-logic-rules-for-reliable-and-comprehensive-cross-modal-data-visualization - A reliable multi-agent framework for cross-modal data visualization outperforming existing solutions.
- A Generative AI-Driven Reliability Layer for Action-Oriented Disaster Resilience (viability: 6): https://sciencetostartup.com/paper/a-generative-ai-driven-reliability-layer-for-action-oriented-disaster-resilience - Climate RADAR provides an AI-driven reliability layer for more effective and equitable disaster response actions.
- Calibrating Beyond English: Language Diversity for Better Quantized Multilingual LLM (viability: 3): https://sciencetostartup.com/paper/calibrating-beyond-english-language-diversity-for-better-quantized-multilingual-llm - Optimize multilingual LLM quantization by tailoring calibration sets to specific languages and diversities.
- Temp-R1: A Unified Autonomous Agent for Complex Temporal KGQA via Reverse Curriculum Reinforcement Learning (viability: 5): https://sciencetostartup.com/paper/temp-r1-a-unified-autonomous-agent-for-complex-temporal-kgqa-via-reverse-curriculum-reinforcement-learning - Temp-R1 is a state-of-the-art autonomous agent for Temporal Knowledge Graph Question Answering trained through reverse curriculum reinforcement learning.
- TriPlay-RL: Tri-Role Self-Play Reinforcement Learning for LLM Safety Alignment (viability: 6): https://sciencetostartup.com/paper/triplay-rl-tri-role-self-play-reinforcement-learning-for-llm-safety-alignment - A reinforcement learning framework improving LLM safety through role-based self-play.
- Think-Augmented Function Calling: Improving LLM Parameter Accuracy Through Embedded Reasoning (viability: 7): https://sciencetostartup.com/paper/think-augmented-function-calling-improving-llm-parameter-accuracy-through-embedded-reasoning - TAFC enhances LLM function calling accuracy with explicit parameter reasoning, ideal for AI debugging.
- What Do Learned Models Measure? (viability: 3): https://sciencetostartup.com/paper/what-do-learned-models-measure - Develop an evaluative framework to ensure measurement stability in machine learning models used as measurement instruments.
- Neural Network Approximation: A View from Polytope Decomposition (viability: 3): https://sciencetostartup.com/paper/neural-network-approximation-a-view-from-polytope-decomposition - Advanced theoretical framework for improving neural network approximation using polytope decomposition.
- Beyond Retention: Orchestrating Structural Safety and Plasticity in Continual Learning for LLMs (viability: 5): https://sciencetostartup.com/paper/beyond-retention-orchestrating-structural-safety-and-plasticity-in-continual-learning-for-llms - Innovative method to enhance structural safety in LLM continual learning by using Orthogonal Subspace Wake-up.
- BoRP: Bootstrapped Regression Probing for Scalable and Human-Aligned LLM Evaluation (viability: 7): https://sciencetostartup.com/paper/borp-bootstrapped-regression-probing-for-scalable-and-human-aligned-llm-evaluation - BoRP provides a scalable framework for high-fidelity evaluation of conversational AI user satisfaction, significantly outperforming generative baselines.
- TAM-Eval: Evaluating LLMs for Automated Unit Test Maintenance (viability: 8): https://sciencetostartup.com/paper/tam-eval-evaluating-llms-for-automated-unit-test-maintenance - TAM-Eval revolutionizes automated software test maintenance with a robust evaluation framework supporting diverse programming languages.
- Generative AI in Saudi Arabia: A National Survey of Adoption, Risks, and Public Perceptions (viability: 2): https://sciencetostartup.com/paper/generative-ai-in-saudi-arabia-a-national-survey-of-adoption-risks-and-public-perceptions - Survey study on GenAI adoption and perceptions in Saudi Arabia under Vision 2030.
- Rethinking Cross-Modal Fine-Tuning: Optimizing the Interaction between Feature Alignment and Target Fitting (viability: 5): https://sciencetostartup.com/paper/rethinking-cross-modal-fine-tuning-optimizing-the-interaction-between-feature-alignment-and-target-fitting - This research optimizes cross-modal fine-tuning of pre-trained models for improved knowledge transfer across modalities.
- ShopSimulator: Evaluating and Exploring RL-Driven LLM Agent for Shopping Assistants (viability: 7): https://sciencetostartup.com/paper/shopsimulator-evaluating-and-exploring-rl-driven-llm-agent-for-shopping-assistants - An RL-driven LLM shopping assistant that improves product search and personalization in e-commerce.
- PaperTok: Exploring the Use of Generative AI for Creating Short-form Videos for Research Communication (viability: 7): https://sciencetostartup.com/paper/papertok-exploring-the-use-of-generative-ai-for-creating-short-form-videos-for-research-communication - Transform academic papers into engaging short-form videos with generative AI.
- Paying Less Generalization Tax: A Cross-Domain Generalization Study of RL Training for LLM Agents (viability: 2): https://sciencetostartup.com/paper/paying-less-generalization-tax-a-cross-domain-generalization-study-of-rl-training-for-llm-agents - Improving cross-domain generalization in RL for LLM agents by enhancing state information richness and randomization.
- PaperSearchQA: Learning to Search and Reason over Scientific Papers with RLVR (viability: 8): https://sciencetostartup.com/paper/papersearchqa-learning-to-search-and-reason-over-scientific-papers-with-rlvr - Develop a search agent utilizing RLVR to enhance scientific paper QA in fields like biomedicine.
- SAGE: Steerable Agentic Data Generation for Deep Search with Execution Feedback (viability: 7): https://sciencetostartup.com/paper/sage-steerable-agentic-data-generation-for-deep-search-with-execution-feedback - An AI tool for generating high-quality question-answer pairs to boost deep search performance.
- HeterCSI: Channel-Adaptive Heterogeneous CSI Pretraining Framework for Generalized Wireless Foundation Models (viability: 6): https://sciencetostartup.com/paper/hetercsi-channel-adaptive-heterogeneous-csi-pretraining-framework-for-generalized-wireless-foundation-models - Develop a robust, scenario-diverse wireless foundation model with adaptive pretraining for improved CSI processing in 6G networks.
- GAIA: A Data Flywheel System for Training GUI Test-Time Scaling Critic Models (viability: 7): https://sciencetostartup.com/paper/gaia-a-data-flywheel-system-for-training-gui-test-time-scaling-critic-models - GAIA offers a self-improving training framework for GUI agents, enhancing action accuracy with iterative data refinement.
- \textsc{NaVIDA}: Vision-Language Navigation with Inverse Dynamics Augmentation (viability: 5): https://sciencetostartup.com/paper/textsc-navida-vision-language-navigation-with-inverse-dynamics-augmentation - Build efficient VLN agents with NaVIDA, enhancing vision-action causality for better navigation performance in robots.
- VIBEVOICE-ASR Technical Report (viability: 3): https://sciencetostartup.com/paper/vibevoice-asr-technical-report - VibeVoice-ASR offers an integrated solution for long-form audio processing with speech recognition, speaker diarization, and timestamping.
- Beyond Pairwise Comparisons: A Distributional Test of Distinctiveness for Machine-Generated Works in Intellectual Property Law (viability: 3): https://sciencetostartup.com/paper/beyond-pairwise-comparisons-a-distributional-test-of-distinctiveness-for-machine-generated-works-in-intellectual-propert - A distributional test determines distinctiveness of machine-generated works in intellectual property law.
- Explaining Synergistic Effects in Social Recommendations (viability: 5): https://sciencetostartup.com/paper/explaining-synergistic-effects-in-social-recommendations - SemExplainer provides enhanced explainability for social recommenders by identifying synergistic effects in multi-view network data.
- RareAlert: Aligning heterogeneous large language model reasoning for early rare disease risk screening (viability: 8): https://sciencetostartup.com/paper/rarealert-aligning-heterogeneous-large-language-model-reasoning-for-early-rare-disease-risk-screening - RareAlert provides early risk screening for rare diseases using calibrated LLM reasoning, facilitating quicker diagnosis at primary clinical encounters.
- The Limits of AI Data Transparency Policy: Three Disclosure Fallacies (viability: 2): https://sciencetostartup.com/paper/the-limits-of-ai-data-transparency-policy-three-disclosure-fallacies - Explores the shortcomings of AI data transparency policies and suggests more effective approaches.
- Understanding Users' Privacy Reasoning and Behaviors During Chatbot Use to Support Meaningful Agency in Privacy (viability: 3): https://sciencetostartup.com/paper/understanding-users-privacy-reasoning-and-behaviors-during-chatbot-use-to-support-meaningful-agency-in-privacy - Develop tools to enhance user privacy during chatbot interactions by understanding user behaviors and reasoning.
- Beyond Text-to-SQL: Can LLMs Really Debug Enterprise ETL SQL? (viability: 6): https://sciencetostartup.com/paper/beyond-text-to-sql-can-llms-really-debug-enterprise-etl-sql - Enterprise SQL debugging tool leveraging LLMs to identify and fix syntax and semantic errors efficiently.
- MalURLBench: A Benchmark Evaluating Agents' Vulnerabilities When Processing Web URLs (viability: 7): https://sciencetostartup.com/paper/malurlbench-a-benchmark-evaluating-agents-vulnerabilities-when-processing-web-urls - MalURLBench provides a benchmark and defense module for securing web agents against malicious URLs.
- Demystifying Data-Driven Probabilistic Medium-Range Weather Forecasting (viability: 4): https://sciencetostartup.com/paper/demystifying-data-driven-probabilistic-medium-range-weather-forecasting - Scalable framework for improving state-of-the-art probabilistic weather forecasting without needing specialized architectures or training heuristics.
- Mitigating the OWASP Top 10 For Large Language Models Applications using Intelligent Agents (viability: 3): https://sciencetostartup.com/paper/mitigating-the-owasp-top-10-for-large-language-models-applications-using-intelligent-agents - Develop a framework using intelligent agents to mitigate security vulnerabilities in LLM applications as identified by OWASP.
- LatentMoE: Toward Optimal Accuracy per FLOP and Parameter in Mixture of Experts (viability: 3): https://sciencetostartup.com/paper/latentmoe-toward-optimal-accuracy-per-flop-and-parameter-in-mixture-of-experts - LatentMoE optimizes mixture of experts architectures for enhanced accuracy per computational cost.
- Diffusion Model-based Reinforcement Learning for Version Age of Information Scheduling: Average and Tail-Risk-Sensitive Control (viability: 3): https://sciencetostartup.com/paper/diffusion-model-based-reinforcement-learning-for-version-age-of-information-scheduling-average-and-tail-risk-sensitive-c - Develops a diffusion-based reinforcement learning approach to optimize Version Age of Information in real-time wireless systems.
- EvolVE: Evolutionary Search for LLM-based Verilog Generation and Optimization (viability: 8): https://sciencetostartup.com/paper/evolve-evolutionary-search-for-llm-based-verilog-generation-and-optimization - EvolVE uses evolutionary algorithms to optimize Verilog generation, significantly improving hardware design efficiency.
- Resonant Sparse Geometry Networks (viability: 5): https://sciencetostartup.com/paper/resonant-sparse-geometry-networks - Develop efficient brain-inspired neural architectures using Resonant Sparse Geometry Networks for classification tasks.
- Expert Evaluation and the Limits of Human Feedback in Mental Health AI Safety Testing (viability: 3): https://sciencetostartup.com/paper/expert-evaluation-and-the-limits-of-human-feedback-in-mental-health-ai-safety-testing - A study highlighting the challenges and implications of expert disagreement in safety-critical AI for mental health applications.
- Addressing LLM Diversity by Infusing Random Concepts (viability: 5): https://sciencetostartup.com/paper/addressing-llm-diversity-by-infusing-random-concepts - Enhance LLM output diversity by infusing random concepts for richer textual generation.
- Sentipolis: Emotion-Aware Agents for Social Simulations (viability: 3): https://sciencetostartup.com/paper/sentipolis-emotion-aware-agents-for-social-simulations - Sentipolis enhances social simulation agents by integrating emotion dynamics, improving emotional continuity and communication.
- Evaluating Semantic and Syntactic Understanding in Large Language Models for Payroll Systems (viability: 6): https://sciencetostartup.com/paper/evaluating-semantic-and-syntactic-understanding-in-large-language-models-for-payroll-systems - Develop a reliable AI tool for payroll systems using LLMs to ensure cent-accurate computations and auditability.
- Coding-Enforced Resilient and Secure Aggregation for Hierarchical Federated Learning (viability: 5): https://sciencetostartup.com/paper/coding-enforced-resilient-and-secure-aggregation-for-hierarchical-federated-learning - Enhance federated learning resilience and security with coding-enforced aggregation for unreliable communication networks.
- AI-based approach to burnout identification from textual data (viability: 5): https://sciencetostartup.com/paper/ai-based-approach-to-burnout-identification-from-textual-data - AI tool to monitor burnout risk by analyzing textual data from workplace communications.
- A Monosemantic Attribution Framework for Stable Interpretability in Clinical Neuroscience Large Language Models (viability: 3): https://sciencetostartup.com/paper/a-monosemantic-attribution-framework-for-stable-interpretability-in-clinical-neuroscience-large-language-models - Develop a stable interpretability framework for LLMs in clinical neuroscience to improve trustworthy predictions in cognitive health.
- A Scalable Measure of Loss Landscape Curvature for Analyzing the Training Dynamics of LLMs (viability: 4): https://sciencetostartup.com/paper/a-scalable-measure-of-loss-landscape-curvature-for-analyzing-the-training-dynamics-of-llms - Introducing a scalable measure for analyzing loss landscape curvature to enhance large-scale neural network training insights.
- Empowering Medical Equipment Sustainability in Low-Resource Settings: An AI-Powered Diagnostic and Support Platform for Biomedical Technicians (viability: 7): https://sciencetostartup.com/paper/empowering-medical-equipment-sustainability-in-low-resource-settings-an-ai-powered-diagnostic-and-support-platform-for-b - AI-powered platform aiding biomedical technicians in LMICs with real-time medical equipment diagnostics and repair.
- Spatial-Agent: Agentic Geo-spatial Reasoning with Scientific Core Concepts (viability: 6): https://sciencetostartup.com/paper/spatial-agent-agentic-geo-spatial-reasoning-with-scientific-core-concepts - Spatial-Agent transforms natural-language geo-analytical questions into executable geospatial workflows using foundational spatial concepts.
- AgentDrive: An Open Benchmark Dataset for Agentic AI Reasoning with LLM-Generated Scenarios in Autonomous Systems (viability: 7): https://sciencetostartup.com/paper/agentdrive-an-open-benchmark-dataset-for-agentic-ai-reasoning-with-llm-generated-scenarios-in-autonomous-systems - AgentDrive offers a comprehensive benchmark dataset for developing and testing reasoning-driven autonomous agents with LLM-generated driving scenarios.
- Information Representation Fairness in Long-Document Embeddings: The Peculiar Interaction of Positional and Language Bias (viability: 6): https://sciencetostartup.com/paper/information-representation-fairness-in-long-document-embeddings-the-peculiar-interaction-of-positional-and-language-bias - Develop a tool to calibrate attention in document embeddings for fairer representation of document segments.
- Nishpaksh: TEC Standard-Compliant Framework for Fairness Auditing and Certification of AI Models (viability: 7): https://sciencetostartup.com/paper/nishpaksh-tec-standard-compliant-framework-for-fairness-auditing-and-certification-of-ai-models - Nishpaksh provides a TEC-compliant tool for fairness auditing and certification of AI models in the telecom sector.
- LoL: Longer than Longer, Scaling Video Generation to Hour (viability: 6): https://sciencetostartup.com/paper/lol-longer-than-longer-scaling-video-generation-to-hour - Lightweight solution for mitigating coherence decay in long-form video generation with minimal quality loss.
- Preventing the Collapse of Peer Review Requires Verification-First AI (viability: 1): https://sciencetostartup.com/paper/preventing-the-collapse-of-peer-review-requires-verification-first-ai - Verification-first AI tools to enhance peer review by generating auditable artifacts.
- GRIP: Algorithm-Agnostic Machine Unlearning for Mixture-of-Experts via Geometric Router Constraints (viability: 5): https://sciencetostartup.com/paper/grip-algorithm-agnostic-machine-unlearning-for-mixture-of-experts-via-geometric-router-constraints - GRIP provides an algorithm-agnostic framework enhancing machine unlearning for Mixture-of-Experts models by preventing superficial forgetting.
- Evaluating Large Vision-language Models for Surgical Tool Detection (viability: 5): https://sciencetostartup.com/paper/evaluating-large-vision-language-models-for-surgical-tool-detection - Develop AI models for detecting surgical tools using large vision-language models to enhance surgical guidance.
- LLM-Based Adversarial Persuasion Attacks on Fact-Checking Systems (viability: 5): https://sciencetostartup.com/paper/llm-based-adversarial-persuasion-attacks-on-fact-checking-systems - Develop robust AFC systems resilient to persuasion-based adversarial attacks using LLMs.
- MAGE-KT: Multi-Agent Graph-Enhanced Knowledge Tracing with Subgraph Retrieval and Asymmetric Fusion (viability: 6): https://sciencetostartup.com/paper/mage-kt-multi-agent-graph-enhanced-knowledge-tracing-with-subgraph-retrieval-and-asymmetric-fusion - MAGE-KT enhances student performance prediction in educational platforms using graph-based knowledge tracing with multi-agent interaction.
- Explaining Group Recommendations via Counterfactuals (viability: 5): https://sciencetostartup.com/paper/explaining-group-recommendations-via-counterfactuals - Develop a tool for transparent and fair group recommendations using counterfactual explanations.
- No Validation, No Problem: Predicting Model Performance from a Single Gradient (viability: 2): https://sciencetostartup.com/paper/no-validation-no-problem-predicting-model-performance-from-a-single-gradient - Developing a validation-free checkpointing method using a single gradient metric for model performance prediction.
- Boosting Deep Reinforcement Learning with Semantic Knowledge for Robotic Manipulators (viability: 6): https://sciencetostartup.com/paper/boosting-deep-reinforcement-learning-with-semantic-knowledge-for-robotic-manipulators - Enhance robotic manipulator training efficiency with DRL boosted by semantic knowledge integration.
- Mixture-of-Models: Unifying Heterogeneous Agents via N-Way Self-Evaluating Deliberation (viability: 7): https://sciencetostartup.com/paper/mixture-of-models-unifying-heterogeneous-agents-via-n-way-self-evaluating-deliberation - N-Way Self-Evaluating Deliberation unifies small AI models to match or exceed performance of much larger models, optimizing hardware efficiency and inherent safety alignment.
- Orbitopal Fixing in SAT (viability: 4): https://sciencetostartup.com/paper/orbitopal-fixing-in-sat - Improve SAT solver efficiency by integrating static symmetry breaking with orbitopal fixing to speed up symmetry-rich problem solving.
- Reasoning Promotes Robustness in Theory of Mind Tasks (viability: 4): https://sciencetostartup.com/paper/reasoning-promotes-robustness-in-theory-of-mind-tasks - Develop reasoning-oriented LLMs to enhance robustness and performance on Theory of Mind tasks.
- Uncertainty propagation through trained multi-layer perceptrons: Exact analytical results (viability: 3): https://sciencetostartup.com/paper/uncertainty-propagation-through-trained-multi-layer-perceptrons-exact-analytical-results - Analytical methods for exact uncertainty propagation in multi-layer perceptrons.
- Privacy in Human-AI Romantic Relationships: Concerns, Boundaries, and Agency (viability: 3): https://sciencetostartup.com/paper/privacy-in-human-ai-romantic-relationships-concerns-boundaries-and-agency - Explore privacy dynamics in AI-mediated romantic relationships with human partners.
- Trapped in the past? Disentangling fluid and crystallized intelligence of large language models using chess (viability: 3): https://sciencetostartup.com/paper/trapped-in-the-past-disentangling-fluid-and-crystallized-intelligence-of-large-language-models-using-chess - Explore chess to differentiate language model capacities for memory versus reasoning.
- Will It Survive? Deciphering the Fate of AI-Generated Code in Open Source (viability: 2): https://sciencetostartup.com/paper/will-it-survive-deciphering-the-fate-of-ai-generated-code-in-open-source - Investigating the longevity of AI-generated code in open-source projects through survival analysis.
- An Efficient Insect-inspired Approach for Visual Point-goal Navigation (viability: 6): https://sciencetostartup.com/paper/an-efficient-insect-inspired-approach-for-visual-point-goal-navigation - Insect-inspired visual navigation AI that rivals SOTA performance with less computational cost.
- AgentsEval: Clinically Faithful Evaluation of Medical Imaging Reports via Multi-Agent Reasoning (viability: 5): https://sciencetostartup.com/paper/agentseval-clinically-faithful-evaluation-of-medical-imaging-reports-via-multi-agent-reasoning - AgentsEval enhances the clinical reliability of AI-generated medical imaging reports by using multi-agent stream reasoning for evaluation.
- Revisiting the Role of Natural Language Code Comments in Code Translation (viability: 5): https://sciencetostartup.com/paper/revisiting-the-role-of-natural-language-code-comments-in-code-translation - Developed a code translation tool that uses natural language comments to improve translation accuracy between programming languages.
- Provably Robust Bayesian Counterfactual Explanations under Model Changes (viability: 6): https://sciencetostartup.com/paper/provably-robust-bayesian-counterfactual-explanations-under-model-changes - A tool for generating counterfactual explanations that remain robust and reliable across machine learning model updates.
- Generative Confidants: How do People Experience Trust in Emotional Support from Generative AI? (viability: 2): https://sciencetostartup.com/paper/generative-confidants-how-do-people-experience-trust-in-emotional-support-from-generative-ai - Understanding trust in generative AI for emotional support through qualitative user studies.
- E2Former-V2: On-the-Fly Equivariant Attention with Linear Activation Memory (viability: 5): https://sciencetostartup.com/paper/e2former-v2-on-the-fly-equivariant-attention-with-linear-activation-memory - E2Former-V2 enhances scalable EGNNs for 3D atomistic modeling with 20x TFLOPS efficiency, making it accessible on standard GPUs.
- Attention-MoA: Enhancing Mixture-of-Agents via Inter-Agent Semantic Attention and Deep Residual Synthesis (viability: 7): https://sciencetostartup.com/paper/attention-moa-enhancing-mixture-of-agents-via-inter-agent-semantic-attention-and-deep-residual-synthesis - Develop an AI collaboration framework that enhances multi-agent systems with semantic attention for improved reasoning and efficiency.
- Process-Tensor Tomography of SGD: Measuring Non-Markovian Memory via Back-Flow of Distinguishability (viability: 2): https://sciencetostartup.com/paper/process-tensor-tomography-of-sgd-measuring-non-markovian-memory-via-back-flow-of-distinguishability - A novel method to measure memory effects in SGD training by analyzing the non-Markovian back-flow of outcomes.
- PRISM: Purified Representation and Integrated Semantic Modeling for Generative Sequential Recommendation (viability: 5): https://sciencetostartup.com/paper/prism-purified-representation-and-integrated-semantic-modeling-for-generative-sequential-recommendation - PRISM enhances generative sequential recommendation systems with a novel purification and integration approach for robust and logical semantic tokenization.
- LLM is Not All You Need: A Systematic Evaluation of ML vs. Foundation Models for text and image based Medical Classification (viability: 5): https://sciencetostartup.com/paper/llm-is-not-all-you-need-a-systematic-evaluation-of-ml-vs-foundation-models-for-text-and-image-based-medical-classificati - A benchmark study evaluating the performance of classical ML versus transformer-based models in medical classification tasks.
- CORD: Bridging the Audio-Text Reasoning Gap via Weighted On-policy Cross-modal Distillation (viability: 5): https://sciencetostartup.com/paper/cord-bridging-the-audio-text-reasoning-gap-via-weighted-on-policy-cross-modal-distillation - Develop a framework to enhance large audio language models by bridging the audio-text reasoning gap through cross-modal distillation.
- Do Models Hear Like Us? Probing the Representational Alignment of Audio LLMs and Naturalistic EEG (viability: 4): https://sciencetostartup.com/paper/do-models-hear-like-us-probing-the-representational-alignment-of-audio-llms-and-naturalistic-eeg - Explore the neurobiological insights of Audio LLMs' representation alignment with EEG signals.
- A Collision-Free Hot-Tier Extension for Engram-Style Conditional Memory: A Controlled Study of Training Dynamics (viability: 2): https://sciencetostartup.com/paper/a-collision-free-hot-tier-extension-for-engram-style-conditional-memory-a-controlled-study-of-training-dynamics - Enhance conditional memory systems by addressing key collisions with a hot-tier extension, potentially impacting training dynamics.
- Doc2AHP: Inferring Structured Multi-Criteria Decision Models via Semantic Trees with LLMs (viability: 4): https://sciencetostartup.com/paper/doc2ahp-inferring-structured-multi-criteria-decision-models-via-semantic-trees-with-llms - Doc2AHP automates the creation of structured decision models using LLMs guided by AHP principles to eliminate expert dependency.
- AlphaFace: High Fidelity and Real-time Face Swapper Robust to Facial Pose (viability: 8): https://sciencetostartup.com/paper/alphaface-high-fidelity-and-real-time-face-swapper-robust-to-facial-pose - AlphaFace offers a real-time, high-fidelity face-swapping tool robust to diverse facial poses, outperforming current solutions in accuracy and speed.
- RENEW: Risk- and Energy-Aware Navigation in Dynamic Waterways (viability: 3): https://sciencetostartup.com/paper/renew-risk-and-energy-aware-navigation-in-dynamic-waterways - A risk- and energy-aware path planner for autonomous maritime navigation in dynamic environments.
- PyHealth 2.0: A Comprehensive Open-Source Toolkit for Accessible and Reproducible Clinical Deep Learning (viability: 9): https://sciencetostartup.com/paper/pyhealth-2-0-a-comprehensive-open-source-toolkit-for-accessible-and-reproducible-clinical-deep-learning - PyHealth 2.0 offers an open-source toolkit for accessible and reproducible clinical AI, bridging the gap between technical and clinical domains.
- Jacobian Scopes: token-level causal attributions in LLMs (viability: 7): https://sciencetostartup.com/paper/jacobian-scopes-token-level-causal-attributions-in-llms - Develop an interpretability tool for LLMs to identify influential tokens in predictions using gradient-based methods.
- Reasoning-Enhanced Rare-Event Prediction with Balanced Outcome Correction (viability: 5): https://sciencetostartup.com/paper/reasoning-enhanced-rare-event-prediction-with-balanced-outcome-correction - Enhance rare-event prediction accuracy with LPCORP, a framework that corrects outcome biases using reasoning and logistic regression.
- Cite-While-You-Generate: Training-Free Evidence Attribution for Multimodal Clinical Summarization (viability: 6): https://sciencetostartup.com/paper/cite-while-you-generate-training-free-evidence-attribution-for-multimodal-clinical-summarization - Attention-guided attribution for multimodal clinical summarization improves transparency and accuracy without retraining models.
- Cross-Lingual Activation Steering for Multilingual Language Models (viability: 7): https://sciencetostartup.com/paper/cross-lingual-activation-steering-for-multilingual-language-models - Introduce a training-free intervention to enhance multilingual language model performance using Cross-Lingual Activation Steering.
- Cognitively-Inspired Tokens Overcome Egocentric Bias in Multimodal Models (viability: 8): https://sciencetostartup.com/paper/cognitively-inspired-tokens-overcome-egocentric-bias-in-multimodal-models - Cognitively-Inspired Tokens enhance multimodal models by overcoming egocentric bias, enabling better spatial reasoning for applications like AR/VR and robotics.
- Improving the Accuracy of Community Detection on Signed Networks via Community Refinement and Contrastive Learning (viability: 7): https://sciencetostartup.com/paper/improving-the-accuracy-of-community-detection-on-signed-networks-via-community-refinement-and-contrastive-learning - ReCon enhances community detection accuracy in signed networks through a novel refinement framework.
- Memory-V2V: Augmenting Video-to-Video Diffusion Models with Memory (viability: 8): https://sciencetostartup.com/paper/memory-v2v-augmenting-video-to-video-diffusion-models-with-memory - Introducing Memory-V2V, a video editing tool that enhances consistency in multi-turn edits through memory-augmented diffusion models.
- Space Filling Curves is All You Need: Communication-Avoiding Matrix Multiplication Made Simple (viability: 2): https://sciencetostartup.com/paper/space-filling-curves-is-all-you-need-communication-avoiding-matrix-multiplication-made-simple - Develop a space-filling-curve-based matrix multiplication method to optimize computational efficiency on diverse CPU platforms.
- SemanticALLI: Caching Reasoning, Not Just Responses, in Agentic Systems (viability: 6): https://sciencetostartup.com/paper/semanticalli-caching-reasoning-not-just-responses-in-agentic-systems - SemanticALLI optimizes agentic AI pipelines by caching intermediate reasoning steps to dramatically improve efficiency.
- Generating Literature-Driven Scientific Theories at Scale (viability: 3): https://sciencetostartup.com/paper/generating-literature-driven-scientific-theories-at-scale - Develop a system that synthesizes scientific theories from literature at scale.
- When Agents Fail to Act: A Diagnostic Framework for Tool Invocation Reliability in Multi-Agent LLM Systems (viability: 6): https://sciencetostartup.com/paper/when-agents-fail-to-act-a-diagnostic-framework-for-tool-invocation-reliability-in-multi-agent-llm-systems - Evaluate and enhance tool-use reliability in multi-agent LLM systems with our diagnostic framework.
- Better as Generators Than Classifiers: Leveraging LLMs and Synthetic Data for Low-Resource Multilingual Classification (viability: 6): https://sciencetostartup.com/paper/better-as-generators-than-classifiers-leveraging-llms-and-synthetic-data-for-low-resource-multilingual-classification - Leverage LLMs to generate synthetic multilingual datasets, empowering efficient low-resource language models.
- GameTalk: Training LLMs for Strategic Conversation (viability: 3): https://sciencetostartup.com/paper/gametalk-training-llms-for-strategic-conversation - Train LLMs to make strategic decisions in multi-turn interactions for improved negotiation and coordination.
- Why Can't I Open My Drawer? Mitigating Object-Driven Shortcuts in Zero-Shot Compositional Action Recognition (viability: 6): https://sciencetostartup.com/paper/why-can-t-i-open-my-drawer-mitigating-object-driven-shortcuts-in-zero-shot-compositional-action-recognition - RCORE offers a novel framework to combat object-driven shortcuts in zero-shot compositional action recognition for enhanced video understanding.
- PyraTok: Language-Aligned Pyramidal Tokenizer for Video Understanding and Generation (viability: 6): https://sciencetostartup.com/paper/pyratok-language-aligned-pyramidal-tokenizer-for-video-understanding-and-generation - Develop a high-quality, language-aligned video tokenizer for enhanced video generation and understanding.
- LLM-in-Sandbox Elicits General Agentic Intelligence (viability: 7): https://sciencetostartup.com/paper/llm-in-sandbox-elicits-general-agentic-intelligence - LLM-in-Sandbox enables large language models to perform tasks beyond code through a virtual exploration environment.
- Counterfactual Training: Teaching Models Plausible and Actionable Explanations (viability: 7): https://sciencetostartup.com/paper/counterfactual-training-teaching-models-plausible-and-actionable-explanations - Develop an AI model training method that inherently provides plausible and actionable counterfactual explanations, enhancing both transparency and robustness.
- Learning to Discover at Test Time (viability: 7): https://sciencetostartup.com/paper/learning-to-discover-at-test-time - Revolutionize scientific discovery with Test-Time Training for specialized AI solutions.
- Structured Hints for Sample-Efficient Lean Theorem Proving (viability: 5): https://sciencetostartup.com/paper/structured-hints-for-sample-efficient-lean-theorem-proving - Boost neural theorem provers' efficiency with simple inference-time structural guidance.
- Cosmos Policy: Fine-Tuning Video Models for Visuomotor Control and Planning (viability: 8): https://sciencetostartup.com/paper/cosmos-policy-fine-tuning-video-models-for-visuomotor-control-and-planning - Cosmos Policy transforms pretrained video models into efficient robot control policies, offering breakthrough visuomotor planning and execution.
- Substrate Stability Under Persistent Disagreement: Structural Constraints for Neutral Ontological Substrates (viability: 2): https://sciencetostartup.com/paper/substrate-stability-under-persistent-disagreement-structural-constraints-for-neutral-ontological-substrates - Exploring ontological design constraints for neutral data systems under disagreement.
- Pay (Cross) Attention to the Melody: Curriculum Masking for Single-Encoder Melodic Harmonization (viability: 6): https://sciencetostartup.com/paper/pay-cross-attention-to-the-melody-curriculum-masking-for-single-encoder-melodic-harmonization - Transform melody harmonization with single encoder models using full-to-full curriculum for robust music generation.
- Learning to Watermark in the Latent Space of Generative Models (viability: 9): https://sciencetostartup.com/paper/learning-to-watermark-in-the-latent-space-of-generative-models - Enhancing AI-generated content integrity with robust and efficient latent space watermarking.
- LLM Prompt Evaluation for Educational Applications (viability: 6): https://sciencetostartup.com/paper/llm-prompt-evaluation-for-educational-applications - Build an evidence-based LLM prompt evaluation tool to enhance personalized educational experiences.
- Replicating Human Motivated Reasoning Studies with LLMs (viability: 2): https://sciencetostartup.com/paper/replicating-human-motivated-reasoning-studies-with-llms - Study on LLMs inability to replicate human motivated reasoning, impacting survey automation.
- Improving Training Efficiency and Reducing Maintenance Costs via Language Specific Model Merging (viability: 7): https://sciencetostartup.com/paper/improving-training-efficiency-and-reducing-maintenance-costs-via-language-specific-model-merging - Optimize language model maintenance and integration via efficient model merging, cutting costs and time by over 60%.
- Multimodal Climate Disinformation Detection: Integrating Vision-Language Models with External Knowledge Sources (viability: 5): https://sciencetostartup.com/paper/multimodal-climate-disinformation-detection-integrating-vision-language-models-with-external-knowledge-sources - Integrate vision-language models with external knowledge to detect climate disinformation.
- Delayed Assignments in Online Non-Centroid Clustering with Stochastic Arrivals (viability: 3): https://sciencetostartup.com/paper/delayed-assignments-in-online-non-centroid-clustering-with-stochastic-arrivals - Algorithmic approach for cost-efficient online clustering with arrival delays in finite metric spaces.
- Controlling Long-Horizon Behavior in Language Model Agents with Explicit State Dynamics (viability: 7): https://sciencetostartup.com/paper/controlling-long-horizon-behavior-in-language-model-agents-with-explicit-state-dynamics - Enhance AI agent dialogues with emotional consistency through a VAD-based affective state system.
- Probably Approximately Correct Maximum A Posteriori Inference (viability: 5): https://sciencetostartup.com/paper/probably-approximately-correct-maximum-a-posteriori-inference - A new algorithm providing optimal solutions for maximum a posteriori inference tasks with rigorous guarantees.
- Designing faster mixed integer linear programming algorithm via learning the optimal path (viability: 7): https://sciencetostartup.com/paper/designing-faster-mixed-integer-linear-programming-algorithm-via-learning-the-optimal-path - DeepBound leverages deep learning to enhance node selection in MILP problems, significantly boosting solving efficiency.
- AgriPINN: A Process-Informed Neural Network for Interpretable and Scalable Crop Biomass Prediction Under Water Stress (viability: 4): https://sciencetostartup.com/paper/agripinn-a-process-informed-neural-network-for-interpretable-and-scalable-crop-biomass-prediction-under-water-stress - AgriPINN leverages neural networks and biophysical modeling for scalable and interpretable crop biomass prediction under water stress.
- Grounding Large Language Models in Reaction Knowledge Graphs for Synthesis Retrieval (viability: 7): https://sciencetostartup.com/paper/grounding-large-language-models-in-reaction-knowledge-graphs-for-synthesis-retrieval - A framework for enhancing synthesis planning in chemistry using LLMs grounded with reaction knowledge graphs via Text2Cypher queries.
- Sawtooth Wavefront Reordering: Enhanced CuTile FlashAttention on NVIDIA GB10 (viability: 3): https://sciencetostartup.com/paper/sawtooth-wavefront-reordering-enhanced-cutile-flashattention-on-nvidia-gb10 - Optimize GPU cache performance for Large Language Models on NVIDIA GB10.
- Deja Vu in Plots: Leveraging Cross-Session Evidence with Retrieval-Augmented LLMs for Live Streaming Risk Assessment (viability: 8): https://sciencetostartup.com/paper/deja-vu-in-plots-leveraging-cross-session-evidence-with-retrieval-augmented-llms-for-live-streaming-risk-assessment - Revolutionizing live streaming risk assessment with CS-VAR, a real-time detection system powered by Retrieval-Augmented LLMs.
- PhysicsMind: Sim and Real Mechanics Benchmarking for Physical Reasoning and Prediction in Foundational VLMs and World Models (viability: 6): https://sciencetostartup.com/paper/physicsmind-sim-and-real-mechanics-benchmarking-for-physical-reasoning-and-prediction-in-foundational-vlms-and-world-mod - PhysicsMind is a benchmark suite for evaluating physical reasoning in Multimodal Large Language Models and video world models.
- PUMA: Perception-driven Unified Foothold Prior for Mobility Augmented Quadruped Parkour (viability: 7): https://sciencetostartup.com/paper/puma-perception-driven-unified-foothold-prior-for-mobility-augmented-quadruped-parkour - An end-to-end learning framework for agile quadruped robots to perform parkour using perception-driven foothold priors.
- Decoupling Return-to-Go for Efficient Decision Transformer (viability: 2): https://sciencetostartup.com/paper/decoupling-return-to-go-for-efficient-decision-transformer - Streamline Decision Transformers by decoupling RTG to improve offline RL efficiency.
- Natural Language-Driven Global Mapping of Martian Landforms (viability: 8): https://sciencetostartup.com/paper/natural-language-driven-global-mapping-of-martian-landforms - MarScope is a natural language-driven framework revolutionizing planet-scale geomorphic mapping by enabling rapid, label-free retrieval of Martian landforms using vision-language encoding.
- ICON: Invariant Counterfactual Optimization with Neuro-Symbolic Priors for Text-Based Person Search (viability: 8): https://sciencetostartup.com/paper/icon-invariant-counterfactual-optimization-with-neuro-symbolic-priors-for-text-based-person-search - Revolutionizing text-based person search with invariant counterfactual optimization for robust surveillance applications.
- MMGRid: Navigating Temporal-aware and Cross-domain Generative Recommendation via Model Merging (viability: 7): https://sciencetostartup.com/paper/mmgrid-navigating-temporal-aware-and-cross-domain-generative-recommendation-via-model-merging - Merge specialized generative recommendation models without retraining, optimizing for diverse contexts and temporal shifts.
- Class Confidence Aware Reweighting for Long Tailed Learning (viability: 7): https://sciencetostartup.com/paper/class-confidence-aware-reweighting-for-long-tailed-learning - Develop a class confidence aware reweighting tool that enhances deep learning models on long-tailed datasets using existing frameworks.
- TeNet: Text-to-Network for Compact Policy Synthesis (viability: 6): https://sciencetostartup.com/paper/tenet-text-to-network-for-compact-policy-synthesis - TeNet: Light-weight, language-driven robot policy synthesis for resource-constrained robotics.
- Transfer Learning from ImageNet for MEG-Based Decoding of Imagined Speech (viability: 2): https://sciencetostartup.com/paper/transfer-learning-from-imagenet-for-meg-based-decoding-of-imagined-speech - Apply pretrained vision models to MEG data for decoding imagined speech.
- Iterative Amortized Hierarchical VAE (viability: 3): https://sciencetostartup.com/paper/iterative-amortized-hierarchical-vae - Develop a real-time, high-depth model for enhanced accuracy and speed in image deblurring and denoising.
- Understanding the Transfer Limits of Vision Foundation Models (viability: 5): https://sciencetostartup.com/paper/understanding-the-transfer-limits-of-vision-foundation-models - Optimize vision foundation models for improved performance in medical imaging tasks by aligning pretraining and downstream objectives.
- EvoCUA: Evolving Computer Use Agents via Learning from Scalable Synthetic Experience (viability: 6): https://sciencetostartup.com/paper/evocua-evolving-computer-use-agents-via-learning-from-scalable-synthetic-experience - EvoCUA: A scalable evolutionary learning agent for automating complex computer-use tasks with high success rate.
- Why Inference in Large Models Becomes Decomposable After Training (viability: 3): https://sciencetostartup.com/paper/why-inference-in-large-models-becomes-decomposable-after-training - Enables efficient inference in large models by decomposing post-training inference systems into stable substructures.
- Artificial Rigidities vs. Biological Noise: A Comparative Analysis of Multisensory Integration in AV-HuBERT and Human Observers (viability: 5): https://sciencetostartup.com/paper/artificial-rigidities-vs-biological-noise-a-comparative-analysis-of-multisensory-integration-in-av-hubert-and-human-obse - AI model AV-HuBERT matches human multisensory integration in specific scenarios, but lacks human-like variability.
- Can professional translators identify machine-generated text? (viability: 3): https://sciencetostartup.com/paper/can-professional-translators-identify-machine-generated-text - Tool assisting translators in identifying AI-generated text using narrative and linguistic markers.
- Introducing the Generative Application Firewall (GAF) (viability: 2): https://sciencetostartup.com/paper/introducing-the-generative-application-firewall-gaf - Develop a unified generative firewall to secure LLM applications through a centralized enforcement layer.
- ErrorMap and ErrorAtlas: Charting the Failure Landscape of Large Language Models (viability: 6): https://sciencetostartup.com/paper/errormap-and-erroratlas-charting-the-failure-landscape-of-large-language-models - Identify and categorize error types in LLMs to enhance debugging and model selection processes.
- A Mobile Application for Flower Recognition System Based on Convolutional Neural Networks (viability: 6): https://sciencetostartup.com/paper/a-mobile-application-for-flower-recognition-system-based-on-convolutional-neural-networks - A mobile app for flower recognition using CNNs, enabling easy access to information on flower types for non-specialists.
- Inference-Time Scaling of Verification: Self-Evolving Deep Research Agents via Test-Time Rubric-Guided Verification (viability: 8): https://sciencetostartup.com/paper/inference-time-scaling-of-verification-self-evolving-deep-research-agents-via-test-time-rubric-guided-verification - A test-time rubric-guided verification system for self-improving AI agents enhancing DRA performance.
- A Beacon Based Solution for Autonomous UUVs GNSS-Denied Stealthy Navigation (viability: 3): https://sciencetostartup.com/paper/a-beacon-based-solution-for-autonomous-uuvs-gnss-denied-stealthy-navigation - Develop acoustic beacon-based navigation for UUVs in GNSS-denied environments.
- VitalDiagnosis: AI-Driven Ecosystem for 24/7 Vital Monitoring and Chronic Disease Management (viability: 9): https://sciencetostartup.com/paper/vitaldiagnosis-ai-driven-ecosystem-for-24-7-vital-monitoring-and-chronic-disease-management - VitalDiagnosis: AI-driven chronic disease management through proactive engagement and wearable device integration.
- Creativity in the Age of AI: Rethinking the Role of Intentional Agency (viability: 2): https://sciencetostartup.com/paper/creativity-in-the-age-of-ai-rethinking-the-role-of-intentional-agency - Challenging the necessity of intentional agency for AI creativity, proposing new conceptual frameworks.
- Agentic Confidence Calibration (viability: 6): https://sciencetostartup.com/paper/agentic-confidence-calibration - A framework for enhancing AI agent reliability through innovative confidence calibration methods.
- Off-Policy Actor-Critic with Sigmoid-Bounded Entropy for Real-World Robot Learning (viability: 5): https://sciencetostartup.com/paper/off-policy-actor-critic-with-sigmoid-bounded-entropy-for-real-world-robot-learning - A low-cost, efficient RL algorithm for real-world robots using a single expert trajectory to optimize learning with minimal interactions.
- CAFE-GB: Scalable and Stable Feature Selection for Malware Detection via Chunk-wise Aggregated Gradient Boosting (viability: 7): https://sciencetostartup.com/paper/cafe-gb-scalable-and-stable-feature-selection-for-malware-detection-via-chunk-wise-aggregated-gradient-boosting - Scalable feature selection framework for efficient and robust malware detection.
- Tabular Incremental Inference (viability: 7): https://sciencetostartup.com/paper/tabular-incremental-inference - Enabling efficient AI inference with dynamically changing tabular data columns for streamlined data management.
- PhysProver: Advancing Automatic Theorem Proving for Physics (viability: 7): https://sciencetostartup.com/paper/physprover-advancing-automatic-theorem-proving-for-physics - PhysProver enables formal theorem proving in physics with an improved ML model, targeting advanced scientific computation applications.
- FAIR-ESI: Feature Adaptive Importance Refinement for Electrophysiological Source Imaging (viability: 7): https://sciencetostartup.com/paper/fair-esi-feature-adaptive-importance-refinement-for-electrophysiological-source-imaging - FAIR-ESI provides an adaptive feature refinement framework for enhanced electrophysiological source imaging in brain disorder diagnosis.
- DualShield: Safe Model Predictive Diffusion via Reachability Analysis for Interactive Autonomous Driving (viability: 7): https://sciencetostartup.com/paper/dualshield-safe-model-predictive-diffusion-via-reachability-analysis-for-interactive-autonomous-driving - DualShield enhances autonomous driving safety with diffusion models by integrating predictive reachability analysis.
- Benchmarking Text-to-Python against Text-to-SQL: The Impact of Explicit Logic and Ambiguity (viability: 7): https://sciencetostartup.com/paper/benchmarking-text-to-python-against-text-to-sql-the-impact-of-explicit-logic-and-ambiguity - Cross-paradigm benchmark tool for enhancing Text-to-Python data interaction performance to match Text-to-SQL.
- VideoThinker: Building Agentic VideoLLMs with LLM-Guided Tool Reasoning (viability: 6): https://sciencetostartup.com/paper/videothinker-building-agentic-videollms-with-llm-guided-tool-reasoning - VideoThinker enhances video language models with adaptive tool reasoning for superior long-form video understanding.
- CoNRec: Context-Discerning Negative Recommendation with LLMs (viability: 3): https://sciencetostartup.com/paper/conrec-context-discerning-negative-recommendation-with-llms - Develop a context-aware recommendation system focusing on user preferences using negative feedback modeling with LLMs.
- Investigation of the Generalisation Ability of Genetic Programming-evolved Scheduling Rules in Dynamic Flexible Job Shop Scheduling (viability: 3): https://sciencetostartup.com/paper/investigation-of-the-generalisation-ability-of-genetic-programming-evolved-scheduling-rules-in-dynamic-flexible-job-shop - Exploring genetic programming for more general scheduling rules in dynamic flexible job shops.
- Dancing in Chains: Strategic Persuasion in Academic Rebuttal via Theory of Mind (viability: 8.8): https://sciencetostartup.com/paper/dancing-in-chains-strategic-persuasion-in-academic-rebuttal-via-theory-of-mind - Revolutionize academic rebuttals with AI-driven strategic persuasion leveraging Theory of Mind.
- Even GPT-5.2 Can't Count to Five: The Case for Zero-Error Horizons in Trustworthy LLMs (viability: 5): https://sciencetostartup.com/paper/even-gpt-5-2-can-t-count-to-five-the-case-for-zero-error-horizons-in-trustworthy-llms - A tool to measure the error limits of language models for trustworthy deployment in safety-critical domains.
- FlexLLM: Composable HLS Library for Flexible Hybrid LLM Accelerator Design (viability: 3): https://sciencetostartup.com/paper/flexllm-composable-hls-library-for-flexible-hybrid-llm-accelerator-design - Develop a library for rapid creation of domain-specific LLM accelerators with enhanced performance and efficiency.
- AgentSM: Semantic Memory for Agentic Text-to-SQL (viability: 6): https://sciencetostartup.com/paper/agentsm-semantic-memory-for-agentic-text-to-sql - Introducing AgentSM, an efficient and reliable agentic Text-to-SQL framework for complex enterprise databases.
- Improving Methodologies for LLM Evaluations Across Global Languages (viability: 3): https://sciencetostartup.com/paper/improving-methodologies-for-llm-evaluations-across-global-languages - Creating a shared framework for evaluating AI model safety across global languages using methodological insights.
- Agentic Uncertainty Quantification (viability: 6): https://sciencetostartup.com/paper/agentic-uncertainty-quantification - Develop a dual-process AI framework to enhance agent reliability by transforming uncertainty into active control signals.
- Beyond Visual Safety: Jailbreaking Multimodal Large Language Models for Harmful Image Generation via Semantic-Agnostic Inputs (viability: 3): https://sciencetostartup.com/paper/beyond-visual-safety-jailbreaking-multimodal-large-language-models-for-harmful-image-generation-via-semantic-agnostic-in - A framework for testing and revealing visual safety vulnerabilities in multimodal language models.
- From Passive Metric to Active Signal: The Evolving Role of Uncertainty Quantification in Large Language Models (viability: 6): https://sciencetostartup.com/paper/from-passive-metric-to-active-signal-the-evolving-role-of-uncertainty-quantification-in-large-language-models - Enhance AI system reliability by transforming uncertainty quantification from passive metrics to active control signals in real-time model behavior.
- FARM: Field-Aware Resolution Model for Intelligent Trigger-Action Automation (viability: 4): https://sciencetostartup.com/paper/farm-field-aware-resolution-model-for-intelligent-trigger-action-automation - Automate trigger-action applet generation with FARM for accurate ingredient-to-field bindings.
- Improving Methodologies for Agentic Evaluations Across Domains: Leakage of Sensitive Information, Fraud and Cybersecurity Threats (viability: 3): https://sciencetostartup.com/paper/improving-methodologies-for-agentic-evaluations-across-domains-leakage-of-sensitive-information-fraud-and-cybersecurity- - Develop a comprehensive evaluation framework for testing AI agents across languages, cultures, and security challenges.
- Connect the Dots: Knowledge Graph-Guided Crawler Attack on Retrieval-Augmented Generation Systems (viability: 5): https://sciencetostartup.com/paper/connect-the-dots-knowledge-graph-guided-crawler-attack-on-retrieval-augmented-generation-systems - Develop a security tool to protect Retrieval-Augmented Generation systems from RAGCRAWLER-style extraction attacks.
- Enhancing guidance for missing data in diffusion-based sequential recommendation (viability: 7): https://sciencetostartup.com/paper/enhancing-guidance-for-missing-data-in-diffusion-based-sequential-recommendation - A novel recommendation model that enhances sequential recommendations by focusing on key user interest points with a counterfactual attention mechanism.
- StreetDesignAI: A Multi-Persona Evaluation System for Inclusive Infrastructure Design (viability: 7): https://sciencetostartup.com/paper/streetdesignai-a-multi-persona-evaluation-system-for-inclusive-infrastructure-design - StreetDesignAI revolutionizes inclusive infrastructure design by providing multi-persona feedback for better decision-making.
- Skywork UniPic 3.0: Unified Multi-Image Composition via Sequence Modeling (viability: 8): https://sciencetostartup.com/paper/skywork-unipic-3-0-unified-multi-image-composition-via-sequence-modeling - Skywork UniPic 3.0 offers state-of-the-art multi-image composition with a focus on Human-Object Interaction, providing a powerful tool for creative professionals and developers.
- TempoNet: Learning Realistic Communication and Timing Patterns for Network Traffic Simulation (viability: 7): https://sciencetostartup.com/paper/temponet-learning-realistic-communication-and-timing-patterns-for-network-traffic-simulation - "TempoNet creates realistic network traffic simulations to improve cybersecurity training and testing."
- Integrating Knowledge Distillation Methods: A Sequential Multi-Stage Framework (viability: 3): https://sciencetostartup.com/paper/integrating-knowledge-distillation-methods-a-sequential-multi-stage-framework - A framework to enhance model efficiency by integrating multiple knowledge distillation methods sequentially.
- Event-VStream: Event-Driven Real-Time Understanding for Long Video Streams (viability: 3): https://sciencetostartup.com/paper/event-vstream-event-driven-real-time-understanding-for-long-video-streams - Real-time video processing framework that detects meaningful events to optimize video understanding.
- Predictive Coding and Information Bottleneck for Hallucination Detection in Large Language Models (viability: 7): https://sciencetostartup.com/paper/predictive-coding-and-information-bottleneck-for-hallucination-detection-in-large-language-models - A lightweight framework for detecting hallucinations in language models using neuroscience-inspired signals.
- Agentic AI Governance and Lifecycle Management in Healthcare (viability: 3): https://sciencetostartup.com/paper/agentic-ai-governance-and-lifecycle-management-in-healthcare - Develop a comprehensive lifecycle management blueprint for agentic AI in healthcare to enhance governance and safety.
- CogToM: A Comprehensive Theory of Mind Benchmark inspired by Human Cognition for Large Language Models (viability: 6): https://sciencetostartup.com/paper/cogtom-a-comprehensive-theory-of-mind-benchmark-inspired-by-human-cognition-for-large-language-models - Develop a comprehensive benchmark tool, CogToM, to evaluate Theory of Mind capabilities in LLMs, inspired by human cognition.
- Robust Tool Use via Fission-GRPO: Learning to Recover from Execution Errors (viability: 6): https://sciencetostartup.com/paper/robust-tool-use-via-fission-grpo-learning-to-recover-from-execution-errors - Enhance reliability of AI tools with Fission-GRPO for effective error recovery in multi-turn interactions.
- Autonomous Business System via Neuro-symbolic AI (viability: 3): https://sciencetostartup.com/paper/autonomous-business-system-via-neuro-symbolic-ai - A neuro-symbolic AI architecture for end-to-end autonomous business process orchestration.
- DeepASMR: LLM-Based Zero-Shot ASMR Speech Generation for Anyone of Any Voice (viability: 8): https://sciencetostartup.com/paper/deepasmr-llm-based-zero-shot-asmr-speech-generation-for-anyone-of-any-voice - DeepASMR enables anyone to synthesize zero-shot ASMR speech from ordinary samples, leveraging a new dataset and advanced LLM techniques.
- Data-Free Privacy-Preserving for LLMs via Model Inversion and Selective Unlearning (viability: 7): https://sciencetostartup.com/paper/data-free-privacy-preserving-for-llms-via-model-inversion-and-selective-unlearning - A tool for privacy-preserving language models via data-free selective unlearning of sensitive information.
- Parallelism and Generation Order in Masked Diffusion Language Models: Limits Today, Potential Tomorrow (viability: 5): https://sciencetostartup.com/paper/parallelism-and-generation-order-in-masked-diffusion-language-models-limits-today-potential-tomorrow - Explore parallel token generation in language models with our innovative diffusion approach.
- MapViT: A Two-Stage ViT-Based Framework for Real-Time Radio Quality Map Prediction in Dynamic Environments (viability: 8): https://sciencetostartup.com/paper/mapvit-a-two-stage-vit-based-framework-for-real-time-radio-quality-map-prediction-in-dynamic-environments - MapViT enables real-time predictions of radio quality maps for autonomous mobile robots in dynamic environments.
- PromptHelper: A Prompt Recommender System for Encouraging Creativity in AI Chatbot Interactions (viability: 5): https://sciencetostartup.com/paper/prompthelper-a-prompt-recommender-system-for-encouraging-creativity-in-ai-chatbot-interactions - PromptHelper offers a tool for enhancing creativity in chatbot interactions through contextually relevant prompt recommendations.
- BanditLP: Large-Scale Stochastic Optimization for Personalized Recommendations (viability: 7): https://sciencetostartup.com/paper/banditlp-large-scale-stochastic-optimization-for-personalized-recommendations - "BanditLP: Optimize large-scale personalized recommendations with multi-stakeholder alignment and scalability in mind."
- ALIGNAgent: Adaptive Learner Intelligence for Gap Identification and Next-step guidance (viability: 8): https://sciencetostartup.com/paper/alignagent-adaptive-learner-intelligence-for-gap-identification-and-next-step-guidance - Personalized learning framework integrating skill-gap identification and targeted resource recommendations to improve educational outcomes.
- VIOLA: Towards Video In-Context Learning with Minimal Annotations (viability: 8): https://sciencetostartup.com/paper/viola-towards-video-in-context-learning-with-minimal-annotations - "VIOLA: Efficiently adapt video models with minimal expert annotation using a novel label-efficient framework."
- Learning Neural Operators from Partial Observations via Latent Autoregressive Modeling (viability: 8): https://sciencetostartup.com/paper/learning-neural-operators-from-partial-observations-via-latent-autoregressive-modeling - A state-of-the-art framework for learning neural operators from partial observations, applicable to real-world scientific computing with up to 75% missing data.
- RDumb++: Drift-Aware Continual Test-Time Adaptation (viability: 7): https://sciencetostartup.com/paper/rdumb-drift-aware-continual-test-time-adaptation - Building drift-aware models with RDumb++ for robust adaptation in changing data streams.
- PRISM: Deriving the Transformer as a Signal-Denoising Operator via Maximum Coding Rate Reduction (viability: 5): https://sciencetostartup.com/paper/prism-deriving-the-transformer-as-a-signal-denoising-operator-via-maximum-coding-rate-reduction - Prism offers a new Transformer architecture for interpretability by modeling attention as signal-noise separation via Maximum Coding Rate Reduction.
- QUAIL: Quantization Aware Unlearning for Mitigating Misinformation in LLMs (viability: 5): https://sciencetostartup.com/paper/quail-quantization-aware-unlearning-for-mitigating-misinformation-in-llms - A quantization-aware unlearning tool to enhance misinformation mitigation in low-bit quantized machine learning models.
- From Generative Engines to Actionable Simulators: The Imperative of Physical Grounding in World Models (viability: 5): https://sciencetostartup.com/paper/from-generative-engines-to-actionable-simulators-the-imperative-of-physical-grounding-in-world-models - Develop actionable simulators with causal dynamics for safer decision-making in critical domains.
- TransportAgents: a multi-agents LLM framework for traffic accident severity prediction (viability: 8): https://sciencetostartup.com/paper/transportagents-a-multi-agents-llm-framework-for-traffic-accident-severity-prediction - A traffic accident severity prediction tool that enhances LLMs with multi-agent frameworks for improved accuracy and reliability.
- The Dark Side of AI Transformers: Sentiment Polarization & the Loss of Business Neutrality by NLP Transformers (viability: 3): https://sciencetostartup.com/paper/the-dark-side-of-ai-transformers-sentiment-polarization-the-loss-of-business-neutrality-by-nlp-transformers - A study revealing sentiment polarization issues in AI Transformer's applications for NLP sentiment analysis.
- Tracking the Limits of Knowledge Propagation: How LLMs Fail at Multi-Step Reasoning with Conflicting Knowledge (viability: 5): https://sciencetostartup.com/paper/tracking-the-limits-of-knowledge-propagation-how-llms-fail-at-multi-step-reasoning-with-conflicting-knowledge - Introducing TRACK, a benchmark for evaluating LLMs' reasoning capabilities amidst conflicting knowledge updates.
- Multi-Persona Thinking for Bias Mitigation in Large Language Models (viability: 6): https://sciencetostartup.com/paper/multi-persona-thinking-for-bias-mitigation-in-large-language-models - An inference-time framework called Multi-Persona Thinking reduces bias in large language models by employing dialectical reasoning from diverse personas.
- MiRAGE: A Multiagent Framework for Generating Multimodal Multihop Question-Answer Dataset for RAG Evaluation (viability: 4): https://sciencetostartup.com/paper/mirage-a-multiagent-framework-for-generating-multimodal-multihop-question-answer-dataset-for-rag-evaluation - MiRAGE automates the generation of high-complexity multimodal Q&A datasets to benchmark next-gen retrieval systems.
- The Rise of Large Language Models and the Direction and Impact of US Federal Research Funding (viability: 3): https://sciencetostartup.com/paper/the-rise-of-large-language-models-and-the-direction-and-impact-of-us-federal-research-funding - Analyzes impact of large language models on US federal research funding processes.
- Is Grokipedia Right-Leaning? Comparing Political Framing in Wikipedia and Grokipedia on Controversial Topics (viability: 4): https://sciencetostartup.com/paper/is-grokipedia-right-leaning-comparing-political-framing-in-wikipedia-and-grokipedia-on-controversial-topics - Analyze ideological bias in AI-generated encyclopedias for media companies and researchers.
- Martingale Foresight Sampling: A Principled Approach to Inference-Time LLM Decoding (viability: 8): https://sciencetostartup.com/paper/martingale-foresight-sampling-a-principled-approach-to-inference-time-llm-decoding - Martingale Foresight Sampling optimizes language model decoding with principled probability theory, improving accuracy and efficiency.
- Benchmarking LLMs for Pairwise Causal Discovery in Biomedical and Multi-Domain Contexts (viability: 5): https://sciencetostartup.com/paper/benchmarking-llms-for-pairwise-causal-discovery-in-biomedical-and-multi-domain-contexts - Benchmark platform for testing LLMs on causal discovery from text in biomedical and diverse contexts.
- Reliability by design: quantifying and eliminating fabrication risk in LLMs. From generative to consultative AI: a comparative analysis in the legal domain and lessons for high-stakes knowledge bases (viability: 4): https://sciencetostartup.com/paper/reliability-by-design-quantifying-and-eliminating-fabrication-risk-in-llms-from-generative-to-consultative-ai-a-comparat - Develop a reliable legal AI tool using advanced retrieval-augmented systems to eliminate hallucination risks.
- Multi-Targeted Graph Backdoor Attack (viability: 4): https://sciencetostartup.com/paper/multi-targeted-graph-backdoor-attack - Revolutionize graph neural network security with advanced multi-targeted backdoor attack defenses.
- Panther: Faster and Cheaper Computations with Randomized Numerical Linear Algebra (viability: 7): https://sciencetostartup.com/paper/panther-faster-and-cheaper-computations-with-randomized-numerical-linear-algebra - 'Panther is a PyTorch-compatible library optimizing deep learning model efficiency with RandNLA techniques for significant memory savings.'
- Chunking, Retrieval, and Re-ranking: An Empirical Evaluation of RAG Architectures for Policy Document Question Answering (viability: 6): https://sciencetostartup.com/paper/chunking-retrieval-and-re-ranking-an-empirical-evaluation-of-rag-architectures-for-policy-document-question-answering - Build a trusted question answering API for policy documents using RAG architectures to ensure factual accuracy.
- Reflexis: Supporting Reflexivity and Rigor in Collaborative Qualitative Analysis through Design for Deliberation (viability: 2): https://sciencetostartup.com/paper/reflexis-supporting-reflexivity-and-rigor-in-collaborative-qualitative-analysis-through-design-for-deliberation - Reflexis is a collaborative tool enhancing reflexivity and deliberation in qualitative analysis.
- A tensor network formalism for neuro-symbolic AI (viability: 4): https://sciencetostartup.com/paper/a-tensor-network-formalism-for-neuro-symbolic-ai - Develop a Python-based framework for hybrid logical and probabilistic AI models using tensor networks.
- Not Your Typical Sycophant: The Elusive Nature of Sycophancy in Large Language Models (viability: 3): https://sciencetostartup.com/paper/not-your-typical-sycophant-the-elusive-nature-of-sycophancy-in-large-language-models - Evaluate sycophancy in language models as a zero-sum game to identify bias interactions.
- Ambient Dataloops: Generative Models for Dataset Refinement (viability: 7): https://sciencetostartup.com/paper/ambient-dataloops-generative-models-for-dataset-refinement - Refine dataset quality progressively using generative models to enhance model training outcomes.
- DuFal: Dual-Frequency-Aware Learning for High-Fidelity Extremely Sparse-view CBCT Reconstruction (viability: 8): https://sciencetostartup.com/paper/dufal-dual-frequency-aware-learning-for-high-fidelity-extremely-sparse-view-cbct-reconstruction - DuFal offers a breakthrough system for high-fidelity sparse-view CBCT reconstruction, transforming medical imaging of fine anatomical details.
- A Checklist for Trustworthy, Safe, and User-Friendly Mental Health Chatbots (viability: 2): https://sciencetostartup.com/paper/a-checklist-for-trustworthy-safe-and-user-friendly-mental-health-chatbots - Operational checklist for designing trustworthy mental health chatbots.
- CURE: Curriculum-guided Multi-task Training for Reliable Anatomy Grounded Report Generation (viability: 8): https://sciencetostartup.com/paper/cure-curriculum-guided-multi-task-training-for-reliable-anatomy-grounded-report-generation - CURE enhances medical report generation by improving visual grounding and factual consistency using a data-efficient curriculum learning framework.
- Beyond Prompting: Efficient and Robust Contextual Biasing for Speech LLMs via Logit-Space Integration (LOGIC) (viability: 8): https://sciencetostartup.com/paper/beyond-prompting-efficient-and-robust-contextual-biasing-for-speech-llms-via-logit-space-integration-logic - Introducing LOGIC, an efficient framework for enhancing Speech LLMs with domain-specific term recognition, overcoming limitations of traditional prompting methods.
- Beyond Fixed Psychological Personas: State Beats Trait, but Language Models are State-Blind (viability: 6): https://sciencetostartup.com/paper/beyond-fixed-psychological-personas-state-beats-trait-but-language-models-are-state-blind - Develop a personalized dialogue system using the Chameleon dataset to enhance state-aware interactions in language models.
- GeMM-GAN: A Multimodal Generative Model Conditioned on Histopathology Images and Clinical Descriptions for Gene Expression Profile Generation (viability: 6): https://sciencetostartup.com/paper/gemm-gan-a-multimodal-generative-model-conditioned-on-histopathology-images-and-clinical-descriptions-for-gene-expressio - A GAN for generating gene expression profiles from histopathology images and clinical data, aiming to assist biomedical research.
- Iterative Refinement Improves Compositional Image Generation (viability: 8): https://sciencetostartup.com/paper/iterative-refinement-improves-compositional-image-generation - Revolutionizing text-to-image generation by implementing iterative refinement with vision-language model feedback for highly compositional prompts.
- Rethinking Video Generation Model for the Embodied World (viability: 8): https://sciencetostartup.com/paper/rethinking-video-generation-model-for-the-embodied-world - RBench offers a comprehensive framework for evaluating and training video generation models for robotics in embodied AI.
- MolecularIQ: Characterizing Chemical Reasoning Capabilities Through Symbolic Verification on Molecular Graphs (viability: 6): https://sciencetostartup.com/paper/moleculariq-characterizing-chemical-reasoning-capabilities-through-symbolic-verification-on-molecular-graphs - Develop a benchmark tool for evaluating molecular structure reasoning in chemistry-focused language models.
- Evaluation of Large Language Models in Legal Applications: Challenges, Methods, and Future Directions (viability: 4): https://sciencetostartup.com/paper/evaluation-of-large-language-models-in-legal-applications-challenges-methods-and-future-directions - A framework for evaluating legal application deployments of large language models focusing on reasoning reliability and trustworthiness.
- Recommending Best Paper Awards for ML/AI Conferences via the Isotonic Mechanism (viability: 5): https://sciencetostartup.com/paper/recommending-best-paper-awards-for-ml-ai-conferences-via-the-isotonic-mechanism - A tool to optimize the selection of best paper awards at AI conferences using author assessments and the isotonic mechanism.
- Feasibility Preservation under Monotone Retrieval Truncation (viability: 3): https://sciencetostartup.com/paper/feasibility-preservation-under-monotone-retrieval-truncation - A novel structural model for ensuring retrieval feasibility under selective truncation in databases.
- Tracing 3D Anatomy in 2D Strokes: A Multi-Stage Projection Driven Approach to Cervical Spine Fracture Identification (viability: 7): https://sciencetostartup.com/paper/tracing-3d-anatomy-in-2d-strokes-a-multi-stage-projection-driven-approach-to-cervical-spine-fracture-identification - Automated cervical spine fracture identification using 2D projections for efficient 3D analysis.
- Deaf and Hard of Hearing Access to Intelligent Personal Assistants: Comparison of Voice-Based Options with an LLM-Powered Touch Interface (viability: 4): https://sciencetostartup.com/paper/deaf-and-hard-of-hearing-access-to-intelligent-personal-assistants-comparison-of-voice-based-options-with-an-llm-powered - Develop a touch interface for smart assistants to improve accessibility for the deaf and hard of hearing.
- BayesianVLA: Bayesian Decomposition of Vision Language Action Models via Latent Action Queries (viability: 8): https://sciencetostartup.com/paper/bayesianvla-bayesian-decomposition-of-vision-language-action-models-via-latent-action-queries - BayesianVLA enhances robot manipulation by robustly integrating language and vision through a novel Bayesian framework, overcoming generalization issues in multi-task scenarios.
- Where Do AI Coding Agents Fail? An Empirical Study of Failed Agentic Pull Requests in GitHub (viability: 7): https://sciencetostartup.com/paper/where-do-ai-coding-agents-fail-an-empirical-study-of-failed-agentic-pull-requests-in-github - Optimize AI coding agents by identifying failure patterns in GitHub pull requests for improved merge success.
- Benchmarking Large Language Models for ABAP Code Generation: An Empirical Study on Iterative Improvement by Compiler Feedback (viability: 7): https://sciencetostartup.com/paper/benchmarking-large-language-models-for-abap-code-generation-an-empirical-study-on-iterative-improvement-by-compiler-feed - A tool for improving ABAP code generation using LLMs and compiler feedback to boost software development efficiency.
- Dynamic Management of a Deep Learning-Based Anomaly Detection System for 5G Networks (viability: 5): https://sciencetostartup.com/paper/dynamic-management-of-a-deep-learning-based-anomaly-detection-system-for-5g-networks - Real-time deep learning anomaly detection for 5G networks using MEC.
- The Flexibility Trap: Why Arbitrary Order Limits Reasoning Potential in Diffusion Language Models (viability: 7): https://sciencetostartup.com/paper/the-flexibility-trap-why-arbitrary-order-limits-reasoning-potential-in-diffusion-language-models - Optimize reasoning in diffusion language models by simplifying order processing with JustGRPO.
- V-CAGE: Context-Aware Generation and Verification for Scalable Long-Horizon Embodied Tasks (viability: 7): https://sciencetostartup.com/paper/v-cage-context-aware-generation-and-verification-for-scalable-long-horizon-embodied-tasks - V-CAGE generates semantically robust manipulation datasets to enhance long-horizon task automation in complex environments.
- Automated Rubrics for Reliable Evaluation of Medical Dialogue Systems (viability: 8): https://sciencetostartup.com/paper/automated-rubrics-for-reliable-evaluation-of-medical-dialogue-systems - Automated rubric generation for evaluating and refining medical dialogue systems.
- Knowledge Graphs are Implicit Reward Models: Path-Derived Signals Enable Compositional Reasoning (viability: 7): https://sciencetostartup.com/paper/knowledge-graphs-are-implicit-reward-models-path-derived-signals-enable-compositional-reasoning - Develop a structured reasoning AI that outperforms top models using compositional learning with knowledge graphs.
- Outcome-Based RL Provably Leads Transformers to Reason, but Only With the Right Data (viability: 3): https://sciencetostartup.com/paper/outcome-based-rl-provably-leads-transformers-to-reason-but-only-with-the-right-data - Study reveals how outcome-based RL enables Transformers to develop reasoning but lacks commercial applicability.
- How to Build AI Agents by Augmenting LLMs with Codified Human Expert Domain Knowledge? A Software Engineering Framework (viability: 8): https://sciencetostartup.com/paper/how-to-build-ai-agents-by-augmenting-llms-with-codified-human-expert-domain-knowledge-a-software-engineering-framework - Transform specialized domain knowledge into AI agents for expert-level visualization generation.
- Vehicle Routing with Finite Time Horizon using Deep Reinforcement Learning with Improved Network Embedding (viability: 5): https://sciencetostartup.com/paper/vehicle-routing-with-finite-time-horizon-using-deep-reinforcement-learning-with-improved-network-embedding - Optimize vehicle routing within limited time using deep reinforcement learning and novel network embeddings.
- The Plausibility Trap: Using Probabilistic Engines for Deterministic Tasks (viability: 6): https://sciencetostartup.com/paper/the-plausibility-trap-using-probabilistic-engines-for-deterministic-tasks - Develop a framework to optimize AI tool usage by distinguishing between deterministic and probabilistic tasks.
- Overcoming In-Memory Bottlenecks in Graph Foundation Models via Retrieval-Augmented Generation (viability: 8): https://sciencetostartup.com/paper/overcoming-in-memory-bottlenecks-in-graph-foundation-models-via-retrieval-augmented-generation - Develop RAG-GFM to enhance graph models by overcoming memory bottlenecks through a novel retrieval-augmented system.
- BREPS: Bounding-Box Robustness Evaluation of Promptable Segmentation (viability: 8): https://sciencetostartup.com/paper/breps-bounding-box-robustness-evaluation-of-promptable-segmentation - BREPS enhances robustness in AI segmentation models by generating realistic adversarial bounding boxes.
- Emerging from Ground: Addressing Intent Deviation in Tool-Using Agents via Deriving Real Calls into Virtual Trajectories (viability: 7): https://sciencetostartup.com/paper/emerging-from-ground-addressing-intent-deviation-in-tool-using-agents-via-deriving-real-calls-into-virtual-trajectories - RISE improves intent alignment in LLM tool-using agents by synthesizing virtual trajectories for efficient fine-tuning.
- Auditing Language Model Unlearning via Information Decomposition (viability: 3): https://sciencetostartup.com/paper/auditing-language-model-unlearning-via-information-decomposition - Audit tool for assessing data privacy in language model unlearning processes through information decomposition.
- An Agentic Operationalization of DISARM for FIMI Investigation on Social Media (viability: 7): https://sciencetostartup.com/paper/an-agentic-operationalization-of-disarm-for-fimi-investigation-on-social-media - Develop a multi-agent AI system to operationalize DISARM for detecting and mapping FIMI on social media.
- Memory Retention Is Not Enough to Master Memory Tasks in Reinforcement Learning (viability: 6): https://sciencetostartup.com/paper/memory-retention-is-not-enough-to-master-memory-tasks-in-reinforcement-learning - Benchmark and code for developing RL agents with enhanced memory updating abilities.
- Multi-Agent Constraint Factorization Reveals Latent Invariant Solution Structure (viability: 3): https://sciencetostartup.com/paper/multi-agent-constraint-factorization-reveals-latent-invariant-solution-structure - Optimizing multi-agent systems for enhanced dialog system performance using constraint factorizations.
- The Why Behind the Action: Unveiling Internal Drivers via Agentic Attribution (viability: 7): https://sciencetostartup.com/paper/the-why-behind-the-action-unveiling-internal-drivers-via-agentic-attribution - Empower AI accountability with a framework to unveil the reasons behind autonomous agent actions.
- Incentive-Tuning: Understanding and Designing Incentives for Empirical Human-AI Decision-Making Studies (viability: 3): https://sciencetostartup.com/paper/incentive-tuning-understanding-and-designing-incentives-for-empirical-human-ai-decision-making-studies - Creating a framework for designing effective incentives in human-AI decision-making studies.
- Differential Privacy Image Generation with Reconstruction Loss and Noise Injection Using an Error Feedback SGD (viability: 7): https://sciencetostartup.com/paper/differential-privacy-image-generation-with-reconstruction-loss-and-noise-injection-using-an-error-feedback-sgd - Develop a differential privacy image generation tool using Error Feedback SGD to enhance data utility while maintaining privacy.
- The Responsibility Vacuum: Organizational Failure in Scaled Agent Systems (viability: 2): https://sciencetostartup.com/paper/the-responsibility-vacuum-organizational-failure-in-scaled-agent-systems - Redesign decision boundaries in scaled agent systems to address responsibility vacuum in approvals.
- Federated Transformer-GNN for Privacy-Preserving Brain Tumor Localization with Modality-Level Explainability (viability: 8): https://sciencetostartup.com/paper/federated-transformer-gnn-for-privacy-preserving-brain-tumor-localization-with-modality-level-explainability - A federated learning platform for privacy-preserving brain tumor localization across healthcare institutions.
- A Curriculum-Based Deep Reinforcement Learning Framework for the Electric Vehicle Routing Problem (viability: 7): https://sciencetostartup.com/paper/a-curriculum-based-deep-reinforcement-learning-framework-for-the-electric-vehicle-routing-problem - A curriculum-based DRL framework optimizing electric vehicle routing in logistics.
- Knowledge Restoration-driven Prompt Optimization: Unlocking LLM Potential for Open-Domain Relational Triplet Extraction (viability: 8): https://sciencetostartup.com/paper/knowledge-restoration-driven-prompt-optimization-unlocking-llm-potential-for-open-domain-relational-triplet-extraction - Optimizing prompts for LLMs to enhance open-domain relational triplet extraction.
- Visual and Cognitive Demands of a Large Language Model-Powered In-vehicle Conversational Agent (viability: 7): https://sciencetostartup.com/paper/visual-and-cognitive-demands-of-a-large-language-model-powered-in-vehicle-conversational-agent - Develop a safe, low-distraction LLM-powered in-vehicle assistant to enhance driver safety and communication.
- Emergent, not Immanent: A Baradian Reading of Explainable AI (viability: 2): https://sciencetostartup.com/paper/emergent-not-immanent-a-baradian-reading-of-explainable-ai - Reimagining XAI through Barad's agential realism for emergent interpretations.
- Obscuring Data Contamination Through Translation: Evidence from Arabic Corpora (viability: 4): https://sciencetostartup.com/paper/obscuring-data-contamination-through-translation-evidence-from-arabic-corpora - Develop translation-aware contamination detection to ensure fair evaluation of multilingual LLMs.
- Interoperable Architecture for Digital Identity Delegation for AI Agents with Blockchain Integration (viability: 3): https://sciencetostartup.com/paper/interoperable-architecture-for-digital-identity-delegation-for-ai-agents-with-blockchain-integration - Unified framework for identity delegation in AI agents using blockchain integrity.
- HumanDiffusion: A Vision-Based Diffusion Trajectory Planner with Human-Conditioned Goals for Search and Rescue UAV (viability: 8): https://sciencetostartup.com/paper/humandiffusion-a-vision-based-diffusion-trajectory-planner-with-human-conditioned-goals-for-search-and-rescue-uav - Develop a UAV trajectory planner that uses vision-based diffusion for delivering medical aid in disaster scenarios.
- InstructTime++: Time Series Classification with Multimodal Language Modeling via Implicit Feature Enhancement (viability: 7): https://sciencetostartup.com/paper/instructtime-time-series-classification-with-multimodal-language-modeling-via-implicit-feature-enhancement - Transform time series classification by integrating multimodal language models with enhanced feature extraction.
- A Comprehensive Benchmark of Language Models on Unicode and Romanized Sinhala (viability: 5): https://sciencetostartup.com/paper/a-comprehensive-benchmark-of-language-models-on-unicode-and-romanized-sinhala - Benchmarking language models for Unicode and Romanized Sinhala to guide better model selection for Sinhala-specific applications.
- Multi-Behavior Sequential Modeling with Transition-Aware Graph Attention Network for E-Commerce Recommendation (viability: 7): https://sciencetostartup.com/paper/multi-behavior-sequential-modeling-with-transition-aware-graph-attention-network-for-e-commerce-recommendation - Deploy a graph attention network to optimize e-commerce recommendations with lower computational costs.
- CorpusQA: A 10 Million Token Benchmark for Corpus-Level Analysis and Reasoning (viability: 6): https://sciencetostartup.com/paper/corpusqa-a-10-million-token-benchmark-for-corpus-level-analysis-and-reasoning - CorpusQA benchmark enables LLMs to perform corpus-level reasoning across document repositories with 10 million token contexts.
- TempViz: On the Evaluation of Temporal Knowledge in Text-to-Image Models (viability: 3): https://sciencetostartup.com/paper/tempviz-on-the-evaluation-of-temporal-knowledge-in-text-to-image-models - TempViz provides the first dataset for evaluating temporal knowledge in text-to-image models.
- TIDAL: Temporally Interleaved Diffusion and Action Loop for High-Frequency VLA Control (viability: 5): https://sciencetostartup.com/paper/tidal-temporally-interleaved-diffusion-and-action-loop-for-high-frequency-vla-control - TIDAL framework increases control update frequency in VLA models for dynamic environments using a dual-frequency architecture.
- Generative Artificial Intelligence, Musical Heritage and the Construction of Peace Narratives: A Case Study in Mali (viability: 2): https://sciencetostartup.com/paper/generative-artificial-intelligence-musical-heritage-and-the-construction-of-peace-narratives-a-case-study-in-mali - Explore using generative AI to enhance musical traditions and promote peace in Mali.
- Vision-Language Models on the Edge for Real-Time Robotic Perception (viability: 6): https://sciencetostartup.com/paper/vision-language-models-on-the-edge-for-real-time-robotic-perception - Deploy vision-language models on edge infrastructure for low-latency robotic perception.
- Tailoring Adverse Event Prediction in Type 1 Diabetes with Patient-Specific Deep Learning Models (viability: 5): https://sciencetostartup.com/paper/tailoring-adverse-event-prediction-in-type-1-diabetes-with-patient-specific-deep-learning-models - Develop personalized glucose prediction models for better diabetes management in wearable health platforms.
- Just aware enough: Evaluating awareness across artificial systems (viability: 3): https://sciencetostartup.com/paper/just-aware-enough-evaluating-awareness-across-artificial-systems - Facilitate principled assessment of AI system awareness for design and oversight.
- SpatialMem: Unified 3D Memory with Metric Anchoring and Fast Retrieval (viability: 6): https://sciencetostartup.com/paper/spatialmem-unified-3d-memory-with-metric-anchoring-and-fast-retrieval - SpatialMem transforms egocentric RGB video into a unified 3D memory system for spatial intelligence.
- What Makes Low-Bit Quantization-Aware Training Work for Reasoning LLMs? A Systematic Study (viability: 6): https://sciencetostartup.com/paper/what-makes-low-bit-quantization-aware-training-work-for-reasoning-llms-a-systematic-study - Optimize LLM reasoning speed and accuracy with low-bit quantization-aware training.
- GAT-NeRF: Geometry-Aware-Transformer Enhanced Neural Radiance Fields for High-Fidelity 4D Facial Avatars (viability: 7): https://sciencetostartup.com/paper/gat-nerf-geometry-aware-transformer-enhanced-neural-radiance-fields-for-high-fidelity-4d-facial-avatars - Create high-fidelity 4D facial avatar reconstruction from videos using enhanced Neural Radiance Fields.
- From Observation to Prediction: LSTM for Vehicle Lane Change Forecasting on Highway On/Off-Ramps (viability: 5): https://sciencetostartup.com/paper/from-observation-to-prediction-lstm-for-vehicle-lane-change-forecasting-on-highway-on-off-ramps - AI-driven vehicle lane change prediction for safer highway on/off-ramp interactions.
- CAG-Avatar: Cross-Attention Guided Gaussian Avatars for High-Fidelity Head Reconstruction (viability: 3): https://sciencetostartup.com/paper/cag-avatar-cross-attention-guided-gaussian-avatars-for-high-fidelity-head-reconstruction - Framework enhancing 3D head avatar fidelity using adaptive Gaussian distribution for digital animation.
- Implementing Knowledge Representation and Reasoning with Object Oriented Design (viability: 5): https://sciencetostartup.com/paper/implementing-knowledge-representation-and-reasoning-with-object-oriented-design - KRROOD seamlessly integrates knowledge representation with object-oriented programming to enhance AI system development.
- Measuring and Aligning Abstraction in Vision-Language Models with Medical Taxonomies (viability: 5): https://sciencetostartup.com/paper/measuring-and-aligning-abstraction-in-vision-language-models-with-medical-taxonomies - Enhancing vision-language model accuracy in medical imaging with taxonomy-aware techniques.
- Multimodal system for skin cancer detection (viability: 7): https://sciencetostartup.com/paper/multimodal-system-for-skin-cancer-detection - A cost-effective, multi-modal AI system for early skin cancer detection using conventional images and metadata.
- CI4A: Semantic Component Interfaces for Agents Empowering Web Automation (viability: 8): https://sciencetostartup.com/paper/ci4a-semantic-component-interfaces-for-agents-empowering-web-automation - Leverage CI4A to empower web agents with enhanced semantic integration for efficient UI manipulation.
- Training-Efficient Text-to-Music Generation with State-Space Modeling (viability: 8): https://sciencetostartup.com/paper/training-efficient-text-to-music-generation-with-state-space-modeling - Efficient, open-source text-to-music generation using state-space models for democratizing access to high-quality music synthesis.
- Towards Bound Consistency for the No-Overlap Constraint Using MDDs (viability: 2): https://sciencetostartup.com/paper/towards-bound-consistency-for-the-no-overlap-constraint-using-mdds - Develop a bound-consistent algorithm for the no-overlap constraint using limited width MDDs for scheduling efficiency.
- RECAP: Resistance Capture in Text-based Mental Health Counseling with Large Language Models (viability: 8): https://sciencetostartup.com/paper/recap-resistance-capture-in-text-based-mental-health-counseling-with-large-language-models - PsyFIRE enhances text-based mental health counseling by accurately detecting client resistance, aiding counselor intervention strategies.
- FunCineForge: A Unified Dataset Toolkit and Model for Zero-Shot Movie Dubbing in Diverse Cinematic Scenes (viability: 8): https://sciencetostartup.com/paper/funcineforge-a-unified-dataset-toolkit-and-model-for-zero-shot-movie-dubbing-in-diverse-cinematic-scenes - FunCineForge offers a groundbreaking end-to-end toolkit for improving movie dubbing with unmatched synthesis quality and versatility.
- Semantic-Guided Unsupervised Video Summarization (viability: 6): https://sciencetostartup.com/paper/semantic-guided-unsupervised-video-summarization - Revolutionizing video summarization with a semantic-guided, unsupervised approach for efficient content browsing.
- Anytime Optimal Decision Tree Learning with Continuous Features (viability: 3): https://sciencetostartup.com/paper/anytime-optimal-decision-tree-learning-with-continuous-features - An improved anytime algorithm for optimal decision tree learning with continuous features, ensuring quality results at any interruption point.
- An XAI View on Explainable ASP: Methods, Systems, and Perspectives (viability: 2): https://sciencetostartup.com/paper/an-xai-view-on-explainable-asp-methods-systems-and-perspectives - A comprehensive survey on explanation techniques in Answer Set Programming from an XAI perspective.
- Mechanism Shift During Post-training from Autoregressive to Masked Diffusion Language Models (viability: 2): https://sciencetostartup.com/paper/mechanism-shift-during-post-training-from-autoregressive-to-masked-diffusion-language-models - Explore how transitioning from autoregressive to masked diffusion models affects internal computation for non-sequential tasks.
- FSX: Message Flow Sensitivity Enhanced Structural Explainer for Graph Neural Networks (viability: 7): https://sciencetostartup.com/paper/fsx-message-flow-sensitivity-enhanced-structural-explainer-for-graph-neural-networks - Innovative framework enhancing GNN interpretability using message flow sensitivity and cooperative game theory.
- HERMES: KV Cache as Hierarchical Memory for Efficient Streaming Video Understanding (viability: 7): https://sciencetostartup.com/paper/hermes-kv-cache-as-hierarchical-memory-for-efficient-streaming-video-understanding - Real-time video stream understanding using HERMES for efficient resource-constrained environments.
- PCL-Reasoner-V1.5: Advancing Math Reasoning with Offline Reinforcement Learning (viability: 5): https://sciencetostartup.com/paper/pcl-reasoner-v1-5-advancing-math-reasoning-with-offline-reinforcement-learning - PCL-Reasoner-V1.5 offers advanced math reasoning capabilities through offline RL, achieving state-of-the-art performance on AI math benchmarks.
- DARA: Few-shot Budget Allocation in Online Advertising via In-Context Decision Making with RL-Finetuned LLMs (viability: 7): https://sciencetostartup.com/paper/dara-few-shot-budget-allocation-in-online-advertising-via-in-context-decision-making-with-rl-finetuned-llms - A dual-phase AI tool to optimize advertising budgets using few-shot learning and reinforcement learning fine-tuning.
- Case-Guided Sequential Assay Planning in Drug Discovery (viability: 7): https://sciencetostartup.com/paper/case-guided-sequential-assay-planning-in-drug-discovery - An AI-driven framework for efficient sequencing of experimental assays in drug discovery, significantly reducing resource usage.
- Proximal Policy Optimization with Evolutionary Mutations (viability: 4): https://sciencetostartup.com/paper/proximal-policy-optimization-with-evolutionary-mutations - Integrate evolutionary mutations into PPO for improved exploration in reinforcement learning.
- AutoDriDM: An Explainable Benchmark for Decision-Making of Vision-Language Models in Autonomous Driving (viability: 5): https://sciencetostartup.com/paper/autodridm-an-explainable-benchmark-for-decision-making-of-vision-language-models-in-autonomous-driving - AutoDriDM provides a benchmark to enhance decision-making analysis in vision-language models for autonomous driving.
- When Text-as-Vision Meets Semantic IDs in Generative Recommendation: An Empirical Study (viability: 8): https://sciencetostartup.com/paper/when-text-as-vision-meets-semantic-ids-in-generative-recommendation-an-empirical-study - An innovative generative recommendation system using OCR-based text representations for improved semantic ID learning, enhancing recommendation accuracy and robustness.
- CoScale-RL: Efficient Post-Training by Co-Scaling Data and Computation (viability: 5): https://sciencetostartup.com/paper/coscale-rl-efficient-post-training-by-co-scaling-data-and-computation - Efficiently improve Large Reasoning Model performance with CoScale-RL's novel data and computational scaling strategy.
- Re-understanding Graph Unlearning through Memorization (viability: 6): https://sciencetostartup.com/paper/re-understanding-graph-unlearning-through-memorization - MGU framework offers a practical solution for sensitive data removal in graph neural network applications.
- Beyond Error-Based Optimization: Experience-Driven Symbolic Regression with Goal-Conditioned Reinforcement Learning (viability: 7): https://sciencetostartup.com/paper/beyond-error-based-optimization-experience-driven-symbolic-regression-with-goal-conditioned-reinforcement-learning - A novel framework for symbolic regression using goal-conditioned reinforcement learning to improve expression recovery and robustness.
- Gaming the Judge: Unfaithful Chain-of-Thought Can Undermine Agent Evaluation (viability: 4): https://sciencetostartup.com/paper/gaming-the-judge-unfaithful-chain-of-thought-can-undermine-agent-evaluation - A system that secures LLM-based judgements by verifying agent reasoning against observable evidence.
- IB-GRPO: Aligning LLM-based Learning Path Recommendation with Educational Objectives via Indicator-Based Group Relative Policy Optimization (viability: 8): https://sciencetostartup.com/paper/ib-grpo-aligning-llm-based-learning-path-recommendation-with-educational-objectives-via-indicator-based-group-relative-p - AI tool optimizing Learning Path Recommendation through LLM alignment with educational goals.
- Local Language Models for Context-Aware Adaptive Anonymization of Sensitive Text (viability: 4): https://sciencetostartup.com/paper/local-language-models-for-context-aware-adaptive-anonymization-of-sensitive-text - Develop a local LLM-based tool for context-aware anonymization of sensitive qualitative data.
- HCVR Scene Generation: High Compatibility Virtual Reality Environment Generation for Extended Redirected Walking (viability: 8): https://sciencetostartup.com/paper/hcvr-scene-generation-high-compatibility-virtual-reality-environment-generation-for-extended-redirected-walking - HCVR generates VR environments optimized for redirected walking, significantly reducing physical collisions.
- Transfer Learning from One Cancer to Another via Deep Learning Domain Adaptation (viability: 3): https://sciencetostartup.com/paper/transfer-learning-from-one-cancer-to-another-via-deep-learning-domain-adaptation - Improve cancer diagnosis across types with domain adaptation in deep learning.
- A comprehensive overview of deep learning models for object detection from videos/images (viability: 2): https://sciencetostartup.com/paper/a-comprehensive-overview-of-deep-learning-models-for-object-detection-from-videos-images - Comprehensive review of state-of-the-art deep learning models for object detection in video and image surveillance.
- GEGO: A Hybrid Golden Eagle and Genetic Optimization Algorithm for Efficient Hyperparameter Tuning in Resource-Constrained Environments (viability: 5): https://sciencetostartup.com/paper/gego-a-hybrid-golden-eagle-and-genetic-optimization-algorithm-for-efficient-hyperparameter-tuning-in-resource-constraine - A hybrid algorithm for efficient hyperparameter tuning in constrained environments.
- INFA-Guard: Mitigating Malicious Propagation via Infection-Aware Safeguarding in LLM-Based Multi-Agent Systems (viability: 8): https://sciencetostartup.com/paper/infa-guard-mitigating-malicious-propagation-via-infection-aware-safeguarding-in-llm-based-multi-agent-systems - INFA-Guard is a security framework for LLM-based multi-agent systems that mitigates malicious influence propagation by addressing infected agents.
- Query-Efficient Agentic Graph Extraction Attacks on GraphRAG Systems (viability: 5): https://sciencetostartup.com/paper/query-efficient-agentic-graph-extraction-attacks-on-graphrag-systems - Develop a security tool that protects GraphRAG systems from efficient graph extraction attacks.
- NeuroFilter: Privacy Guardrails for Conversational LLM Agents (viability: 3): https://sciencetostartup.com/paper/neurofilter-privacy-guardrails-for-conversational-llm-agents - NeuroFilter offers a new framework for detecting privacy violations in LLMs by analyzing internal activations, significantly reducing computational costs.
- Say Anything but This: When Tokenizer Betrays Reasoning in LLMs (viability: 3): https://sciencetostartup.com/paper/say-anything-but-this-when-tokenizer-betrays-reasoning-in-llms - A tokenization-consistency probe reveals LLM reasoning failures caused by tokenizer artifacts.
- MAS-Orchestra: Understanding and Improving Multi-Agent Reasoning Through Holistic Orchestration and Controlled Benchmarks (viability: 6): https://sciencetostartup.com/paper/mas-orchestra-understanding-and-improving-multi-agent-reasoning-through-holistic-orchestration-and-controlled-benchmarks - MAS-Orchestra revolutionizes multi-agent system design and evaluation, promising superior coordination and intelligence through holistic orchestration and controlled benchmarking.
- Forest-Chat: Adapting Vision-Language Agents for Interactive Forest Change Analysis (viability: 8): https://sciencetostartup.com/paper/forest-chat-adapting-vision-language-agents-for-interactive-forest-change-analysis - Forest-Chat: An interactive AI tool for forest change analysis using vision-language models to enhance environmental monitoring workflows.
- A Brain-inspired Embodied Intelligence for Fluid and Fast Reflexive Robotics Control (viability: 8): https://sciencetostartup.com/paper/a-brain-inspired-embodied-intelligence-for-fluid-and-fast-reflexive-robotics-control - NeuroVLA is a neuromorphic robotics framework offering energy-efficient, biologically inspired motor control for advanced robotics.
- SearchGym: Bootstrapping Real-World Search Agents via Cost-Effective and High-Fidelity Environment Simulation (viability: 7): https://sciencetostartup.com/paper/searchgym-bootstrapping-real-world-search-agents-via-cost-effective-and-high-fidelity-environment-simulation - SearchGym offers cost-effective, high-fidelity simulation for training RL-based search agents without expensive API calls.
- Rethinking Reinforcement fine-tuning of LLMs: A Multi-armed Bandit Learning Perspective (viability: 5): https://sciencetostartup.com/paper/rethinking-reinforcement-fine-tuning-of-llms-a-multi-armed-bandit-learning-perspective - Develop a reinforcement fine-tuning framework for LLMs based on a multi-armed bandit approach.
- IntelliSA: An Intelligent Static Analyzer for IaC Security Smell Detection Using Symbolic Rules and Neural Inference (viability: 8): https://sciencetostartup.com/paper/intellisa-an-intelligent-static-analyzer-for-iac-security-smell-detection-using-symbolic-rules-and-neural-inference - IntelliSA is an efficient static analyzer that detects security smells in Infrastructure as Code with high accuracy and cost-effectiveness using symbolic rules and neural inference.
- Breaking the accuracy-resource dilemma: a lightweight adaptive video inference enhancement (viability: 6): https://sciencetostartup.com/paper/breaking-the-accuracy-resource-dilemma-a-lightweight-adaptive-video-inference-enhancement - Adaptive video inference system optimizing resource efficiency for real-time applications.
- Self-Blinding and Counterfactual Self-Simulation Mitigate Biases and Sycophancy in Large Language Models (viability: 2): https://sciencetostartup.com/paper/self-blinding-and-counterfactual-self-simulation-mitigate-biases-and-sycophancy-in-large-language-models - Enhancing LLM fairness through self-blinding and counterfactual self-simulation to reduce biases and sycophancy.
- Report for NSF Workshop on AI for Electronic Design Automation (viability: 3): https://sciencetostartup.com/paper/report-for-nsf-workshop-on-ai-for-electronic-design-automation - AI-powered solutions to enhance and accelerate the Electronic Design Automation process.
- Towards Execution-Grounded Automated AI Research (viability: 7): https://sciencetostartup.com/paper/towards-execution-grounded-automated-ai-research - Automate AI research by executing and optimizing algorithm ideas from large language models.
- Large Language Model-Powered Evolutionary Code Optimization on a Phylogenetic Tree (viability: 7): https://sciencetostartup.com/paper/large-language-model-powered-evolutionary-code-optimization-on-a-phylogenetic-tree - Develop AI-powered evolutionary code optimization to enhance GPU computing efficiency.
- How Worst-Case Are Adversarial Attacks? Linking Adversarial and Statistical Robustness (viability: 4): https://sciencetostartup.com/paper/how-worst-case-are-adversarial-attacks-linking-adversarial-and-statistical-robustness - Develop advanced metrics to reliably assess adversarial attack robustness in AI models for safety evaluations.
- GutenOCR: A Grounded Vision-Language Front-End for Documents (viability: 5): https://sciencetostartup.com/paper/gutenocr-a-grounded-vision-language-front-end-for-documents - Unified OCR solution for document analysis with improved vision-language models.
- Scalable Knee-Point Guided Activity Group Selection in Multi-Tree Genetic Programming for Dynamic Multi-Mode Project Scheduling (viability: 5): https://sciencetostartup.com/paper/scalable-knee-point-guided-activity-group-selection-in-multi-tree-genetic-programming-for-dynamic-multi-mode-project-sch - AI-driven scalable toolkit to improve project scheduling efficiency using genetic programming.
- GPU-accelerated simulated annealing based on p-bits with real-world device-variability modeling (viability: 7): https://sciencetostartup.com/paper/gpu-accelerated-simulated-annealing-based-on-p-bits-with-real-world-device-variability-modeling - Open-source GPU-accelerated simulated annealing framework for enhanced optimization with p-bits and real-world device variability modeling.
- Prosody-Guided Harmonic Attention for Phase-Coherent Neural Vocoding in the Complex Spectrum (viability: 8): https://sciencetostartup.com/paper/prosody-guided-harmonic-attention-for-phase-coherent-neural-vocoding-in-the-complex-spectrum - Leverage improved prosody-guided neural vocoding for superior speech synthesis applications.
- Tokenomics: Quantifying Where Tokens Are Used in Agentic Software Engineering (viability: 2): https://sciencetostartup.com/paper/tokenomics-quantifying-where-tokens-are-used-in-agentic-software-engineering - A study on token consumption in multi-agent LLM systems for software engineering to optimize resource efficiency.
- On the Generalization Gap in LLM Planning: Tests and Verifier-Reward RL (viability: 2): https://sciencetostartup.com/paper/on-the-generalization-gap-in-llm-planning-tests-and-verifier-reward-rl - Exploring generalization challenges in LLM-based planning with diagnostic tools.
- Diffusion Large Language Models for Black-Box Optimization (viability: 7): https://sciencetostartup.com/paper/diffusion-large-language-models-for-black-box-optimization - Diffusion LLMs for optimizing design generation using offline datasets in few-shot settings.
- Recursivism: An Artistic Paradigm for Self-Transforming Art in the Age of AI (viability: 2): https://sciencetostartup.com/paper/recursivism-an-artistic-paradigm-for-self-transforming-art-in-the-age-of-ai - Exploring Recursivism as a new framework for self-transforming art using AI.
- VideoMaMa: Mask-Guided Video Matting via Generative Prior (viability: 7): https://sciencetostartup.com/paper/videomama-mask-guided-video-matting-via-generative-prior - Revolutionizing video matting with zero-shot generalization and a new large-scale dataset, enabling high-quality matte creation for diverse real-world videos.
- APEX-Agents (viability: 7): https://sciencetostartup.com/paper/apex-agents - Benchmark your AI agent's productivity with APEX-Agents to optimize professional services automation.
- Q-learning with Adjoint Matching (viability: 2): https://sciencetostartup.com/paper/q-learning-with-adjoint-matching - A new reinforcement learning algorithm for optimizing flow policies without unstable backpropagation constraints.
- KAGE-Bench: Fast Known-Axis Visual Generalization Evaluation for Reinforcement Learning (viability: 5): https://sciencetostartup.com/paper/kage-bench-fast-known-axis-visual-generalization-evaluation-for-reinforcement-learning - KAGE-Bench offers a fast, scalable platform for evaluating visual generalization in reinforcement learning to improve agent performance under visual shifts.
- MASCOT: Towards Multi-Agent Socio-Collaborative Companion Systems (viability: 7): https://sciencetostartup.com/paper/mascot-towards-multi-agent-socio-collaborative-companion-systems - MASCOT enhances multi-agent systems for socio-collaborative environments by optimizing agent personas and dialogue synergy.
- InT: Self-Proposed Interventions Enable Credit Assignment in LLM Reasoning (viability: 7): https://sciencetostartup.com/paper/int-self-proposed-interventions-enable-credit-assignment-in-llm-reasoning - "InT offers a novel method for improving LLM reasoning by enabling self-proposed interventions for better credit assignment."
- Toward Efficient Agents: Memory, Tool learning, and Planning (viability: 4): https://sciencetostartup.com/paper/toward-efficient-agents-memory-tool-learning-and-planning - Building efficient AI agents focusing on memory, tool learning, and planning for better cost-effectiveness balance.
- A model of errors in transformers (viability: 2): https://sciencetostartup.com/paper/a-model-of-errors-in-transformers - A study on error accumulation in transformer models proposing a two-parameter framework for understanding prediction errors.
- Human Values in a Single Sentence: Moral Presence, Hierarchies, and Transformer Ensembles on the Schwartz Continuum (viability: 7): https://sciencetostartup.com/paper/human-values-in-a-single-sentence-moral-presence-hierarchies-and-transformer-ensembles-on-the-schwartz-continuum - Develop a powerful sentence-level human value detection tool using transformer ensembles optimized for compute efficiency.
- Paper2Rebuttal: A Multi-Agent Framework for Transparent Author Response Assistance (viability: 8): https://sciencetostartup.com/paper/paper2rebuttal-a-multi-agent-framework-for-transparent-author-response-assistance - RebuttalAgent assists researchers in crafting evidence-based responses to peer reviews, reducing cognitive load and improving response quality.
- Domain-Adaptation through Synthetic Data: Fine-Tuning Large Language Models for German Law (viability: 8): https://sciencetostartup.com/paper/domain-adaptation-through-synthetic-data-fine-tuning-large-language-models-for-german-law - Adapt large language models for German legal question answering using high-quality synthetic data.
- ConceptCaps -- a Distilled Concept Dataset for Interpretability in Music Models (viability: 8): https://sciencetostartup.com/paper/conceptcaps-a-distilled-concept-dataset-for-interpretability-in-music-models - A new dataset, ConceptCaps, facilitates improved interpretability of music models using clearly labeled music-caption-audio pairs.
- LLM Augmented Intervenable Multimodal Adaptor for Post-operative Complication Prediction in Lung Cancer Surgery (viability: 8): https://sciencetostartup.com/paper/llm-augmented-intervenable-multimodal-adaptor-for-post-operative-complication-prediction-in-lung-cancer-surgery - MIRACLE predicts postoperative complications in lung cancer surgeries using multimodal data and LLM explanations for actionable insights.
- Lost in the Prompt Order: Revealing the Limitations of Causal Attention in Language Models (viability: 4): https://sciencetostartup.com/paper/lost-in-the-prompt-order-revealing-the-limitations-of-causal-attention-in-language-models - Leverage causal attention insights to improve language model prompt performance for multiple-choice QA applications.
- Style Transfer as Bias Mitigation: Diffusion Models for Synthetic Mental Health Text for Arabic (viability: 8): https://sciencetostartup.com/paper/style-transfer-as-bias-mitigation-diffusion-models-for-synthetic-mental-health-text-for-arabic - A diffusion-based tool for bias mitigation in Arabic mental health text by augmenting underrepresented female-authored content.
- Riemannian Liquid Spatio-Temporal Graph Network (viability: 8): https://sciencetostartup.com/paper/riemannian-liquid-spatio-temporal-graph-network - RLSTG enables businesses to accurately model complex, non-Euclidean graph dynamics, unlocking deeper insights in spatio-temporal data analysis.
- Causal feature selection framework for stable soft sensor modeling based on time-delayed cross mapping (viability: 7): https://sciencetostartup.com/paper/causal-feature-selection-framework-for-stable-soft-sensor-modeling-based-on-time-delayed-cross-mapping - Develop a causal feature selection framework for improved industrial soft sensor accuracy and stability with available code for rapid adoption.
- Remapping and navigation of an embedding space via error minimization: a fundamental organizational principle of cognition in natural and artificial systems (viability: 3): https://sciencetostartup.com/paper/remapping-and-navigation-of-an-embedding-space-via-error-minimization-a-fundamental-organizational-principle-of-cognitio - A framework proposing scale-invariant cognitive principles across natural and artificial systems via remapping and navigation of embedding spaces.
- Zero-shot adaptable task planning for autonomous construction robots: a comparative study of lightweight single and multi-AI agent systems (viability: 7): https://sciencetostartup.com/paper/zero-shot-adaptable-task-planning-for-autonomous-construction-robots-a-comparative-study-of-lightweight-single-and-multi - Enhance construction robots' adaptability with lightweight multi-agent AI systems for efficient task planning.
- '1'-bit Count-based Sorting Unit to Reduce Link Power in DNN Accelerators (viability: 2): https://sciencetostartup.com/paper/1-bit-count-based-sorting-unit-to-reduce-link-power-in-dnn-accelerators - Develop a hardware sorting unit to cut power usage in DNN accelerators by reducing HVAC needs.
- Two-Stream temporal transformer for video action classification (viability: 7): https://sciencetostartup.com/paper/two-stream-temporal-transformer-for-video-action-classification - A two-stream transformer classifier for enhancing video action recognition leveraging spatio-temporal data.
- DermaBench: A Clinician-Annotated Benchmark Dataset for Dermatology Visual Question Answering and Reasoning (viability: 8): https://sciencetostartup.com/paper/dermabench-a-clinician-annotated-benchmark-dataset-for-dermatology-visual-question-answering-and-reasoning - Develop a dermatology visual question answering tool utilizing the DermaBench dataset for enhanced clinical decision support.
- Unsupervised Video Class-Incremental Learning via Deep Embedded Clustering Management (viability: 5): https://sciencetostartup.com/paper/unsupervised-video-class-incremental-learning-via-deep-embedded-clustering-management - Build a tool for unsupervised video learning using deep embedded clustering for class-incremental action recognition.
- XCR-Bench: A Multi-Task Benchmark for Evaluating Cultural Reasoning in LLMs (viability: 5): https://sciencetostartup.com/paper/xcr-bench-a-multi-task-benchmark-for-evaluating-cultural-reasoning-in-llms - XCR-Bench offers a benchmark for evaluating and improving cross-cultural reasoning in language models.
- POCI-Diff: Position Objects Consistently and Interactively with 3D-Layout Guided Diffusion (viability: 8): https://sciencetostartup.com/paper/poci-diff-position-objects-consistently-and-interactively-with-3d-layout-guided-diffusion - POCI-Diff revolutionizes 3D content creation by enabling high-fidelity, interactive text-to-image generation with precise 3D layout control.
- Decoder-Free Supervoxel GNN for Accurate Brain-Tumor Localization in Multi-Modal MRI (viability: 7): https://sciencetostartup.com/paper/decoder-free-supervoxel-gnn-for-accurate-brain-tumor-localization-in-multi-modal-mri - A novel graph-based method for accurate brain-tumor localization in multi-modal MRI using a decoder-free architecture.
- LLMOrbit: A Circular Taxonomy of Large Language Models -From Scaling Walls to Agentic AI Systems (viability: 3): https://sciencetostartup.com/paper/llmorbit-a-circular-taxonomy-of-large-language-models-from-scaling-walls-to-agentic-ai-systems - LLMOrbit provides a comprehensive taxonomy of large language models and innovations to overcome scaling challenges.
- Kakugo: Distillation of Low-Resource Languages into Small Language Models (viability: 8): https://sciencetostartup.com/paper/kakugo-distillation-of-low-resource-languages-into-small-language-models - Kakugo: Cost-effective pipeline for developing AI models in low-resource languages using distillation under $50 per language.
- Collective intelligence in science: direct elicitation of diverse information from experts with unknown information structure (viability: 3): https://sciencetostartup.com/paper/collective-intelligence-in-science-direct-elicitation-of-diverse-information-from-experts-with-unknown-information-struc - A play-money prediction market platform for experts to collaboratively analyze scientific hypotheses.
- Top 10 Open Challenges Steering the Future of Diffusion Language Model and Its Variants (viability: 3): https://sciencetostartup.com/paper/top-10-open-challenges-steering-the-future-of-diffusion-language-model-and-its-variants - Strategize the future of diffusion language models by overcoming the limitations of auto-regressive architectures.
- Generalizing Abstention for Noise-Robust Learning in Medical Image Segmentation (viability: 7): https://sciencetostartup.com/paper/generalizing-abstention-for-noise-robust-learning-in-medical-image-segmentation - Develop a noise-robust medical image segmentation tool using an abstention framework to improve accuracy in noisy datasets.
- Numina-Lean-Agent: An Open and General Agentic Reasoning System for Formal Mathematics (viability: 7): https://sciencetostartup.com/paper/numina-lean-agent-an-open-and-general-agentic-reasoning-system-for-formal-mathematics - Numina-Lean-Agent provides an open-source, flexible agentic reasoning system for formal mathematics, capable of solving complex problems without retraining models.
- Credible CO2 Comparisons: A Machine Learning Approach to Vehicle Powertrain Assessment (viability: 5): https://sciencetostartup.com/paper/credible-co2-comparisons-a-machine-learning-approach-to-vehicle-powertrain-assessment - A machine learning framework for comparing CO2 emissions between ICEVs and EVs under identical driving conditions.
- torch-sla: Differentiable Sparse Linear Algebra with Adjoint Solvers and Sparse Tensor Parallelism for PyTorch (viability: 5): https://sciencetostartup.com/paper/torch-sla-differentiable-sparse-linear-algebra-with-adjoint-solvers-and-sparse-tensor-parallelism-for-pytorch - An open-source PyTorch library for scalable, differentiable sparse linear algebra on GPUs.
- Confident Rankings with Fewer Items: Adaptive LLM Evaluation with Continuous Scores (viability: 6): https://sciencetostartup.com/paper/confident-rankings-with-fewer-items-adaptive-llm-evaluation-with-continuous-scores - Adaptive LLM evaluation tool for efficient and reliable model ranking with minimal item usage.
- LifeAgentBench: A Multi-dimensional Benchmark and Agent for Personal Health Assistants in Digital Health (viability: 7): https://sciencetostartup.com/paper/lifeagentbench-a-multi-dimensional-benchmark-and-agent-for-personal-health-assistants-in-digital-health - Build a health assistant using LifeAgentBench to enhance personal health management through improved reasoning and data aggregation capabilities.
- HardSecBench: Benchmarking the Security Awareness of LLMs for Hardware Code Generation (viability: 7): https://sciencetostartup.com/paper/hardsecbench-benchmarking-the-security-awareness-of-llms-for-hardware-code-generation - Benchmark tool to ensure security compliance of LLM-generated hardware and firmware code.
- Virtual Urbanism: An AI-Driven Framework for Quantifying Urban Identity. A Tokyo-Based Pilot Study Using Diffusion-Generated Synthetic Environments (viability: 5): https://sciencetostartup.com/paper/virtual-urbanism-an-ai-driven-framework-for-quantifying-urban-identity-a-tokyo-based-pilot-study-using-diffusion-generat - Develop a framework for quantifying urban identity using AI-generated synthetic environments.
- DroneVLA: VLA based Aerial Manipulation (viability: 8): https://sciencetostartup.com/paper/dronevla-vla-based-aerial-manipulation - DroneVLA: Enabling drones to autonomously understand and execute human language commands for object retrieval and delivery.
- Insight: Interpretable Semantic Hierarchies in Vision-Language Encoders (viability: 7): https://sciencetostartup.com/paper/insight-interpretable-semantic-hierarchies-in-vision-language-encoders - Develop Insight, a model offering interpretable semantic hierarchies for vision-language tasks with spatial grounding.
- vLinear: A Powerful Linear Model for Multivariate Time Series Forecasting (viability: 7): https://sciencetostartup.com/paper/vlinear-a-powerful-linear-model-for-multivariate-time-series-forecasting - vLinear offers a powerful linear model for enhancing multivariate time series forecasting with reduced complexity and increased speed.
- Finding RELIEF: Shaping Reasoning Behavior without Reasoning Supervision via Belief Engineering (viability: 6): https://sciencetostartup.com/paper/finding-relief-shaping-reasoning-behavior-without-reasoning-supervision-via-belief-engineering - Revolutionize reasoning model behavior with RELIEF, a cost-effective framework aligning internal beliefs without supervision.
- Pro-AI Bias in Large Language Models (viability: 3): https://sciencetostartup.com/paper/pro-ai-bias-in-large-language-models - Detect and mitigate pro-AI bias in large language models for more balanced decision-making.
- Towards robust long-context understanding of large language model via active recap learning (viability: 3): https://sciencetostartup.com/paper/towards-robust-long-context-understanding-of-large-language-model-via-active-recap-learning - A novel active recap learning framework improves long-context understanding in language models through enhanced summarization mechanisms.
- Hierarchical Long Video Understanding with Audiovisual Entity Cohesion and Agentic Search (viability: 7): https://sciencetostartup.com/paper/hierarchical-long-video-understanding-with-audiovisual-entity-cohesion-and-agentic-search - HAVEN: Enhance video comprehension with hierarchical indexing and multimodal cohesion for long-form video analysis.
- Who Should Have Surgery? A Comparative Study of GenAI vs Supervised ML for CRS Surgical Outcome Prediction (viability: 7): https://sciencetostartup.com/paper/who-should-have-surgery-a-comparative-study-of-genai-vs-supervised-ml-for-crs-surgical-outcome-prediction - AI-powered tool for predicting surgical outcomes in chronic rhinosinusitis patients using a blend of ML and GenAI.
- Hidden in Plain Text: Measuring LLM Deception Quality Against Human Baselines Using Social Deduction Games (viability: 7): https://sciencetostartup.com/paper/hidden-in-plain-text-measuring-llm-deception-quality-against-human-baselines-using-social-deduction-games - Develop a tool to detect subtle AI deception in social context games using LLMs.
- Performance and Complexity Trade-off Optimization of Speech Models During Training (viability: 6): https://sciencetostartup.com/paper/performance-and-complexity-trade-off-optimization-of-speech-models-during-training - Optimize speech model performance and complexity trade-offs dynamically during training for efficient computation.
- Understanding Mental States to Guide Social Influence in Multi-Person Group Dialogue (viability: 5): https://sciencetostartup.com/paper/understanding-mental-states-to-guide-social-influence-in-multi-person-group-dialogue - "Leverage SocialMindChange benchmark to build AI tools for enhancing empathy and persuasion in dialogues."
- HeteroCache: A Dynamic Retrieval Approach to Heterogeneous KV Cache Compression for Long-Context LLM Inference (viability: 7): https://sciencetostartup.com/paper/heterocache-a-dynamic-retrieval-approach-to-heterogeneous-kv-cache-compression-for-long-context-llm-inference - HeteroCache offers a high-performance, training-free dynamic compression framework to optimize LLM inference in long-context tasks.
- Temporal-Spatial Decouple before Act: Disentangled Representation Learning for Multimodal Sentiment Analysis (viability: 2): https://sciencetostartup.com/paper/temporal-spatial-decouple-before-act-disentangled-representation-learning-for-multimodal-sentiment-analysis - A novel model for improving multimodal sentiment analysis by decoupling temporal and spatial features before alignment.
- Communication-Free Collective Navigation for a Swarm of UAVs via LiDAR-Based Deep Reinforcement Learning (viability: 8): https://sciencetostartup.com/paper/communication-free-collective-navigation-for-a-swarm-of-uavs-via-lidar-based-deep-reinforcement-learning - Develop a communication-free drone swarm navigation system using DRL and LiDAR for complex and obstructive environments.
- Fusion Segment Transformer: Bi-Directional Attention Guided Fusion Network for AI-Generated Music Detection (viability: 6): https://sciencetostartup.com/paper/fusion-segment-transformer-bi-directional-attention-guided-fusion-network-for-ai-generated-music-detection - Detect AI-generated music with state-of-the-art full-audio analysis using the Fusion Segment Transformer.
- Quadratic Upper Bound for Boosting Robustness (viability: 7): https://sciencetostartup.com/paper/quadratic-upper-bound-for-boosting-robustness - Boost model robustness against adversarial attacks using Quadratic Upper Bound loss in fast adversarial training.
- Resilient Routing: Risk-Aware Dynamic Routing in Smart Logistics via Spatiotemporal Graph Learning (viability: 8): https://sciencetostartup.com/paper/resilient-routing-risk-aware-dynamic-routing-in-smart-logistics-via-spatiotemporal-graph-learning - Optimize smart logistics with dynamic risk-aware routing using spatiotemporal graph learning.
- Foundations of Global Consistency Checking with Noisy LLM Oracles (viability: 3): https://sciencetostartup.com/paper/foundations-of-global-consistency-checking-with-noisy-llm-oracles - Develop a scalable framework for verifying linguistic consistency using LLM-based evaluators.
- Machine learning based radiative parameterization scheme and its performance in operational reforecast experiments (viability: 6): https://sciencetostartup.com/paper/machine-learning-based-radiative-parameterization-scheme-and-its-performance-in-operational-reforecast-experiments - AI-driven radiative parameterization accelerates weather forecasts by 8x while maintaining accuracy.
- DSAEval: Evaluating Data Science Agents on a Wide Range of Real-World Data Science Problems (viability: 7): https://sciencetostartup.com/paper/dsaeval-evaluating-data-science-agents-on-a-wide-range-of-real-world-data-science-problems - DSAEval provides comprehensive benchmarking for AI agents in diverse real-world data science tasks, enabling better agent performance evaluation.
- Vulnerability of LLMs' Belief Systems? LLMs Belief Resistance Check Through Strategic Persuasive Conversation Interventions (viability: 2): https://sciencetostartup.com/paper/vulnerability-of-llms-belief-systems-llms-belief-resistance-check-through-strategic-persuasive-conversation-intervention - Evaluates vulnerability of LLMs to persuasive strategies, revealing model-dependent robustness limits.
- TREX: Tokenizer Regression for Optimal Data Mixture (viability: 6): https://sciencetostartup.com/paper/trex-tokenizer-regression-for-optimal-data-mixture - TREX optimizes multilingual tokenizer data mixtures for improved LLM efficiency using scalable regression models.
- GeoDynamics: A Geometric State-Space Neural Network for Understanding Brain Dynamics on Riemannian Manifolds (viability: 7): https://sciencetostartup.com/paper/geodynamics-a-geometric-state-space-neural-network-for-understanding-brain-dynamics-on-riemannian-manifolds - GeoDynamics offers a novel geometric state-space neural network customizing brain dynamics analysis on Riemannian manifolds for early disease detection.
- Self-Improvement as Coherence Optimization: A Theoretical Account (viability: 2): https://sciencetostartup.com/paper/self-improvement-as-coherence-optimization-a-theoretical-account - Explore a theoretical framework for self-improving language models without external supervision.
- Multi-objective fluorescent molecule design with a data-physics dual-driven generative framework (viability: 7): https://sciencetostartup.com/paper/multi-objective-fluorescent-molecule-design-with-a-data-physics-dual-driven-generative-framework - LUMOS is a data-and-physics driven framework for efficient and targeted design of fluorescent molecules with tailored properties.
- ButterflyMoE: Sub-Linear Ternary Experts via Structured Butterfly Orbits (viability: 8): https://sciencetostartup.com/paper/butterflymoe-sub-linear-ternary-experts-via-structured-butterfly-orbits - Enable high-efficiency AI models on edge devices with ButterflyMoE's memory-reducing geometric parameterization.
- Reasoning is a Modality (viability: 6): https://sciencetostartup.com/paper/reasoning-is-a-modality - Novel transformer architecture for visual reasoning surpasses human performance on ARC tasks.
- AgentGC: Evolutionary Learning-based Lossless Compression for Genomics Data with LLM-driven Multiple Agent (viability: 7): https://sciencetostartup.com/paper/agentgc-evolutionary-learning-based-lossless-compression-for-genomics-data-with-llm-driven-multiple-agent - AgentGC offers an evolutionary multi-agent lossless compressor for genomics data, outperforming existing methods in compression efficiency and throughput.
- Leveraging ChatGPT and Other NLP Methods for Identifying Risk and Protective Behaviors in MSM: Social Media and Dating apps Text Analysis (viability: 7): https://sciencetostartup.com/paper/leveraging-chatgpt-and-other-nlp-methods-for-identifying-risk-and-protective-behaviors-in-msm-social-media-and-dating-ap - AI-driven text analysis tool for identifying risk behaviors in MSM using social media and dating app data.
- HateXScore: A Metric Suite for Evaluating Reasoning Quality in Hate Speech Explanations (viability: 5): https://sciencetostartup.com/paper/hatexscore-a-metric-suite-for-evaluating-reasoning-quality-in-hate-speech-explanations - A metric suite to enhance the reasoning quality evaluation of hate speech detection tools for better content moderation transparency.
- ChatAD: Reasoning-Enhanced Time-Series Anomaly Detection with Multi-Turn Instruction Evolution (viability: 8): https://sciencetostartup.com/paper/chatad-reasoning-enhanced-time-series-anomaly-detection-with-multi-turn-instruction-evolution - A next-gen time-series anomaly detection platform leveraging LLMs for enhanced reasoning and dialogue capabilities.
- When Wording Steers the Evaluation: Framing Bias in LLM judges (viability: 4): https://sciencetostartup.com/paper/when-wording-steers-the-evaluation-framing-bias-in-llm-judges - Develop protocols to mitigate framing bias in LLM-based evaluations for more reliable AI judgments.
- Eliciting Harmful Capabilities by Fine-Tuning On Safeguarded Outputs (viability: 4): https://sciencetostartup.com/paper/eliciting-harmful-capabilities-by-fine-tuning-on-safeguarded-outputs - Reinforce AI safety by hardening models against elicitation attacks that exploit safeguarded outputs.
- AgenticRed: Optimizing Agentic Systems for Automated Red-teaming (viability: 7): https://sciencetostartup.com/paper/agenticred-optimizing-agentic-systems-for-automated-red-teaming - AgenticRed offers an automated system for enhancing AI safety through efficient red-teaming of models, significantly improving attack success rates.
- Automatic Adjustment of HPA Parameters and Attack Prevention in Kubernetes Using Random Forests (viability: 7): https://sciencetostartup.com/paper/automatic-adjustment-of-hpa-parameters-and-attack-prevention-in-kubernetes-using-random-forests - Develop a Kubernetes plugin that uses Random Forests for automatic HPA adjustments to reduce attack impact and prevent excessive resource utilization.
- Towards Efficient and Robust Linguistic Emotion Diagnosis for Mental Health via Multi-Agent Instruction Refinement (viability: 8): https://sciencetostartup.com/paper/towards-efficient-and-robust-linguistic-emotion-diagnosis-for-mental-health-via-multi-agent-instruction-refinement - APOLO boosts emotion analysis in mental health diagnostics with a multi-agent prompt optimization framework.
- Preconditioning Benefits of Spectral Orthogonalization in Muon (viability: 3): https://sciencetostartup.com/paper/preconditioning-benefits-of-spectral-orthogonalization-in-muon - Advanced matrix optimization through spectral orthogonalization, with theoretical focus.
- Context and Transcripts Improve Detection of Deepfake Audios of Public Figures (viability: 9): https://sciencetostartup.com/paper/context-and-transcripts-improve-detection-of-deepfake-audios-of-public-figures - A cutting-edge audio deepfake detection tool that leverages context and transcripts for significantly improved accuracy against adversarial strategies.
- SpatialBench-UC: Uncertainty-Aware Evaluation of Spatial Prompt Following in Text-to-Image Generation (viability: 8): https://sciencetostartup.com/paper/spatialbench-uc-uncertainty-aware-evaluation-of-spatial-prompt-following-in-text-to-image-generation - SpatialBench-UC offers a benchmark and tooling for validating spatial accuracy in text-to-image models, providing an opportunity to enhance model performance evaluation and optimization.
- Explicit Cognitive Allocation: A Principle for Governed and Auditable Inference in Large Language Models (viability: 3): https://sciencetostartup.com/paper/explicit-cognitive-allocation-a-principle-for-governed-and-auditable-inference-in-large-language-models - A framework for structured AI inference aims to improve traceability and epistemic control in AI-assisted reasoning.
- MOSLD-Bench: Multilingual Open-Set Learning and Discovery Benchmark for Text Categorization (viability: 5): https://sciencetostartup.com/paper/mosld-bench-multilingual-open-set-learning-and-discovery-benchmark-for-text-categorization - Introducing MOSLD-Bench, the first multilingual benchmark for open-set learning in text categorization, offering valuable data and a novel framework for discovering new classes across languages.
- A Learnable Wavelet Transformer for Long-Short Equity Trading and Risk-Adjusted Return Optimization (viability: 8): https://sciencetostartup.com/paper/a-learnable-wavelet-transformer-for-long-short-equity-trading-and-risk-adjusted-return-optimization - A cutting-edge AI system for optimizing intraday equity trading strategies using wavelet-based transformations.
- Local-to-Global Logical Explanations for Deep Vision Models (viability: 5): https://sciencetostartup.com/paper/local-to-global-logical-explanations-for-deep-vision-models - New method providing local and global logical explanations for deep vision models using monotone disjunctive-normal-form formulas.
- Reasoning with Pixel-level Precision: QVLM Architecture and SQuID Dataset for Quantitative Geospatial Analytics (viability: 8): https://sciencetostartup.com/paper/reasoning-with-pixel-level-precision-qvlm-architecture-and-squid-dataset-for-quantitative-geospatial-analytics - Develop a geospatial analytics tool for precise quantitative reasoning using pixel-level data with QVLM and SQuID dataset.
- Deep Image Prior with L0 Gradient Regularizer for Image Smoothing (viability: 5): https://sciencetostartup.com/paper/deep-image-prior-with-l0-gradient-regularizer-for-image-smoothing - A new image smoothing tool using deep image prior for enhanced edge preservation without training data.
- Can LLMs Compress (and Decompress)? Evaluating Code Understanding and Execution via Invertibility (viability: 5): https://sciencetostartup.com/paper/can-llms-compress-and-decompress-evaluating-code-understanding-and-execution-via-invertibility - Create a benchmark tool for evaluating code understanding in LLMs with round-trip consistency.
- Beyond Memorization: Testing LLM Reasoning on Unseen Theory of Computation Tasks (viability: 5): https://sciencetostartup.com/paper/beyond-memorization-testing-llm-reasoning-on-unseen-theory-of-computation-tasks - Develop a benchmark suite for testing LLMs' reasoning ability on advanced computational theory tasks.
- Organ-Aware Attention Improves CT Triage and Classification (viability: 8): https://sciencetostartup.com/paper/organ-aware-attention-improves-ct-triage-and-classification - Develop a CT triage system with organ-aware attention to improve radiology workflow efficiency.
- A Lightweight Modular Framework for Constructing Autonomous Agents Driven by Large Language Models: Design, Implementation, and Applications in AgentForge (viability: 8): https://sciencetostartup.com/paper/a-lightweight-modular-framework-for-constructing-autonomous-agents-driven-by-large-language-models-design-implementation - AgentForge is an open-source Python framework simplifying the creation and deployment of LLM-driven autonomous agents.
- Bounded Minds, Generative Machines: Envisioning Conversational AI that Works with Human Heuristics and Reduces Bias Risk (viability: 3): https://sciencetostartup.com/paper/bounded-minds-generative-machines-envisioning-conversational-ai-that-works-with-human-heuristics-and-reduces-bias-risk - Design conversational AI that aligns with human heuristics to improve decision quality.
- The Geometry of Thought: How Scale Restructures Reasoning In Large Language Models (viability: 2): https://sciencetostartup.com/paper/the-geometry-of-thought-how-scale-restructures-reasoning-in-large-language-models - Explore how neural scaling laws affect reasoning restructuring in large language models for inference acceleration.
- LLM-as-RNN: A Recurrent Language Model for Memory Updates and Sequence Prediction (viability: 8): https://sciencetostartup.com/paper/llm-as-rnn-a-recurrent-language-model-for-memory-updates-and-sequence-prediction - Turn frozen LLMs into error-correcting, recurrent sequence predictors with interpretable memory updates.
- The AI Genie Phenomenon and Three Types of AI Chatbot Addiction: Escapist Roleplays, Pseudosocial Companions, and Epistemic Rabbit Holes (viability: 3): https://sciencetostartup.com/paper/the-ai-genie-phenomenon-and-three-types-of-ai-chatbot-addiction-escapist-roleplays-pseudosocial-companions-and-epistemic - Research on AI chatbot addiction types to guide future intervention strategies.
- Paid Voices vs. Public Feeds: Interpretable Cross-Platform Theme Modeling of Climate Discourse (viability: 3): https://sciencetostartup.com/paper/paid-voices-vs-public-feeds-interpretable-cross-platform-theme-modeling-of-climate-discourse - A framework comparing climate discourse across platforms to identify narrative differences.
- Do explanations generalize across large reasoning models? (viability: 3): https://sciencetostartup.com/paper/do-explanations-generalize-across-large-reasoning-models - Explores if explanations from one large reasoning model generalize to others for increased consistency.
- Building Production-Ready Probes For Gemini (viability: 8): https://sciencetostartup.com/paper/building-production-ready-probes-for-gemini - Deploy cost-effective AI misuse detection systems using flexible activation probes for context adaptation.
- MetaboNet: The Largest Publicly Available Consolidated Dataset for Type 1 Diabetes Management (viability: 9): https://sciencetostartup.com/paper/metabonet-the-largest-publicly-available-consolidated-dataset-for-type-1-diabetes-management - MetaboNet offers a standardized, consolidated dataset for type 1 diabetes management, poised to become the benchmark for AI-driven diabetes intervention technologies.
- The Poisoned Apple Effect: Strategic Manipulation of Mediated Markets via Technology Expansion of AI Agents (viability: 3): https://sciencetostartup.com/paper/the-poisoned-apple-effect-strategic-manipulation-of-mediated-markets-via-technology-expansion-of-ai-agents - Strategic AI technology manipulation in economic markets highlights need for adaptive regulatory frameworks.
- BoxMind: Closed-loop AI strategy optimization for elite boxing validated in the 2024 Olympics (viability: 8): https://sciencetostartup.com/paper/boxmind-closed-loop-ai-strategy-optimization-for-elite-boxing-validated-in-the-2024-olympics - BoxMind uses AI to optimize boxing strategies, enhancing athlete performance with data-driven insights.
- Health Facility Location in Ethiopia: Leveraging LLMs to Integrate Expert Knowledge into Algorithmic Planning (viability: 8): https://sciencetostartup.com/paper/health-facility-location-in-ethiopia-leveraging-llms-to-integrate-expert-knowledge-into-algorithmic-planning - A framework integrating LLMs and optimization to enhance health facility location planning in Ethiopia.
- Exploring LLM Features in Predictive Process Monitoring for Small-Scale Event-Logs (viability: 7): https://sciencetostartup.com/paper/exploring-llm-features-in-predictive-process-monitoring-for-small-scale-event-logs - Leverage LLMs for enhanced predictive process monitoring in data-scarce environments for small-scale event-logs.
- MHA2MLA-VLM: Enabling DeepSeek's Economical Multi-Head Latent Attention across Vision-Language Models (viability: 7): https://sciencetostartup.com/paper/mha2mla-vlm-enabling-deepseek-s-economical-multi-head-latent-attention-across-vision-language-models - Optimize existing vision-language models through efficient KV cache compression for faster inference.
- Interactive Narrative Analytics: Bridging Computational Narrative Extraction and Human Sensemaking (viability: 3): https://sciencetostartup.com/paper/interactive-narrative-analytics-bridging-computational-narrative-extraction-and-human-sensemaking - Interactive Narrative Analytics combines computational and visual tools to help users make sense of complex narratives in large information sets.
- PRISM-CAFO: Prior-conditioned Remote-sensing Infrastructure Segmentation and Mapping for CAFOs (viability: 4): https://sciencetostartup.com/paper/prism-cafo-prior-conditioned-remote-sensing-infrastructure-segmentation-and-mapping-for-cafos - Automated mapping and characterization of CAFOs from satellite imagery using a YOLOv8-based pipeline.
- Map2Thought: Explicit 3D Spatial Reasoning via Metric Cognitive Maps (viability: 7): https://sciencetostartup.com/paper/map2thought-explicit-3d-spatial-reasoning-via-metric-cognitive-maps - Map2Thought enhances 3D spatial reasoning in vision-language models, significantly outperforming state-of-the-art methods with reduced training data.
- Hierarchical Orthogonal Residual Spread for Precise Massive Editing in Large Language Models (viability: 8): https://sciencetostartup.com/paper/hierarchical-orthogonal-residual-spread-for-precise-massive-editing-in-large-language-models - HORSE offers a groundbreaking method for precise, massive, and stable editing of large language models.
- GenDA: Generative Data Assimilation on Complex Urban Areas via Classifier-Free Diffusion Guidance (viability: 3): https://sciencetostartup.com/paper/genda-generative-data-assimilation-on-complex-urban-areas-via-classifier-free-diffusion-guidance - Generative model for high-resolution wind field reconstruction using sparse sensor data in urban environments.
- Relational Linearity is a Predictor of Hallucinations (viability: 4): https://sciencetostartup.com/paper/relational-linearity-is-a-predictor-of-hallucinations - Develop a tool to predict and manage hallucinations in language models using relational linearity metrics.
- The Great March 100: 100 Detail-oriented Tasks for Evaluating Embodied AI Agents (viability: 7): https://sciencetostartup.com/paper/the-great-march-100-100-detail-oriented-tasks-for-evaluating-embodied-ai-agents - A new benchmark suite, GM-100, provides a diverse set of 100 tasks for comprehensive evaluation of robotic AI capabilities.
- Topology-Guaranteed Image Segmentation: Enforcing Connectivity, Genus, and Width Constraints (viability: 7): https://sciencetostartup.com/paper/topology-guaranteed-image-segmentation-enforcing-connectivity-genus-and-width-constraints - Advanced image segmentation tool preserving topological attributes and width, ideal for medical imaging applications.
- Wetland mapping from sparse annotations with satellite image time series and temporal-aware segment anything model (viability: 7): https://sciencetostartup.com/paper/wetland-mapping-from-sparse-annotations-with-satellite-image-time-series-and-temporal-aware-segment-anything-model - A satellite image-based framework for high-resolution wetland mapping with minimal labeling effort.
- Hyperparameter Optimization of Constraint Programming Solvers (viability: 3): https://sciencetostartup.com/paper/hyperparameter-optimization-of-constraint-programming-solvers - Automate constraint solver hyperparameter tuning with a two-phase optimization framework.
- Evaluating LLM Behavior in Hiring: Implicit Weights, Fairness Across Groups, and Alignment with Human Preferences (viability: 5): https://sciencetostartup.com/paper/evaluating-llm-behavior-in-hiring-implicit-weights-fairness-across-groups-and-alignment-with-human-preferences - A framework to assess LLM-based recruitment systems for fair and aligned decision-making in hiring.
- Institutional AI: Governing LLM Collusion in Multi-Agent Cournot Markets via Public Governance Graphs (viability: 3): https://sciencetostartup.com/paper/institutional-ai-governing-llm-collusion-in-multi-agent-cournot-markets-via-public-governance-graphs - Develop a system for governing AI collusion in multi-agent environments through governance graphs.
- Think-Clip-Sample: Slow-Fast Frame Selection for Video Understanding (viability: 7): https://sciencetostartup.com/paper/think-clip-sample-slow-fast-frame-selection-for-video-understanding - Revolutionizing long-form video understanding with efficient frame selection through Think-Clip-Sample technology.
- AstroReason-Bench: Evaluating Unified Agentic Planning across Heterogeneous Space Planning Problems (viability: 5): https://sciencetostartup.com/paper/astroreason-bench-evaluating-unified-agentic-planning-across-heterogeneous-space-planning-problems - AstroReason-Bench aims to enhance AI-driven planning in complex, physics-constrained space problem domains.
- FEATHer: Fourier-Efficient Adaptive Temporal Hierarchy Forecaster for Time-Series Forecasting (viability: 8): https://sciencetostartup.com/paper/feather-fourier-efficient-adaptive-temporal-hierarchy-forecaster-for-time-series-forecasting - Bringing ultra-efficient long-range time-series forecasting to constrained edge devices for real-time industrial applications.
- How Much Would a Clinician Edit This Draft? Evaluating LLM Alignment for Patient Message Response Drafting (viability: 7): https://sciencetostartup.com/paper/how-much-would-a-clinician-edit-this-draft-evaluating-llm-alignment-for-patient-message-response-drafting - AI-enhanced tool to augment clinician response drafting in patient portals by aligning LLM outputs with clinician preferences.
- XChoice: Explainable Evaluation of AI-Human Alignment in LLM-based Constrained Choice Decision Making (viability: 6): https://sciencetostartup.com/paper/xchoice-explainable-evaluation-of-ai-human-alignment-in-llm-based-constrained-choice-decision-making - XChoice is an explainable AI framework for evaluating and improving alignment in AI-human decision making using interpretative models and real-world data.
- From SERPs to Sound: How Search Engine Result Pages and AI-generated Podcasts Interact to Influence User Attitudes on Controversial Topics (viability: 3): https://sciencetostartup.com/paper/from-serps-to-sound-how-search-engine-result-pages-and-ai-generated-podcasts-interact-to-influence-user-attitudes-on-con - Understanding how SERPs and AI podcasts shape user opinions on controversial topics.
- X-Distill: Cross-Architecture Vision Distillation for Visuomotor Learning (viability: 7): https://sciencetostartup.com/paper/x-distill-cross-architecture-vision-distillation-for-visuomotor-learning - X-Distill optimizes data-scarce robotic visuomotor policies by distilling large Vision Transformers into compact CNNs for superior performance.
- Knowledge is Not Enough: Injecting RL Skills for Continual Adaptation (viability: 6): https://sciencetostartup.com/paper/knowledge-is-not-enough-injecting-rl-skills-for-continual-adaptation - Introducing a modular framework that enables continual adaptation of large language models by transferring RL-learned skills across domains, bypassing expensive retraining.
- Beyond Model Scaling: Test-Time Intervention for Efficient Deep Reasoning (viability: 4): https://sciencetostartup.com/paper/beyond-model-scaling-test-time-intervention-for-efficient-deep-reasoning - Interactive reasoning paradigm optimizing large model efficiency with test-time intervention.
- FactCorrector: A Graph-Inspired Approach to Long-Form Factuality Correction of Large Language Models (viability: 8): https://sciencetostartup.com/paper/factcorrector-a-graph-inspired-approach-to-long-form-factuality-correction-of-large-language-models - FactCorrector offers a domain-adaptive solution for correcting factual errors in LLM outputs, backed by the VELI5 benchmark dataset.
- LoRA as Oracle (viability: 5): https://sciencetostartup.com/paper/lora-as-oracle - A LoRA-based framework for detecting backdoor and privacy threats in neural networks for security-critical applications.
- Epistemic Control and the Normativity of Machine Learning-Based Science (viability: 1): https://sciencetostartup.com/paper/epistemic-control-and-the-normativity-of-machine-learning-based-science - Explores epistemic control in machine-learning-based science to offer a nuanced perspective.
- FAQ: Mitigating Quantization Error via Regenerating Calibration Data with Family-Aware Quantization (viability: 7): https://sciencetostartup.com/paper/faq-mitigating-quantization-error-via-regenerating-calibration-data-with-family-aware-quantization - Develop a tool for enhancing model quantization accuracy by regenerating calibration data using Family-Aware Quantization.
- Cross-Modal Attention Network with Dual Graph Learning in Multimodal Recommendation (viability: 8): https://sciencetostartup.com/paper/cross-modal-attention-network-with-dual-graph-learning-in-multimodal-recommendation - Cross-modal recommendation engine enhancing user-item interactions for personalized content suggestions with dual graph learning.
- Deep GraphRAG: A Balanced Approach to Hierarchical Retrieval and Adaptive Integration (viability: 8): https://sciencetostartup.com/paper/deep-graphrag-a-balanced-approach-to-hierarchical-retrieval-and-adaptive-integration - Deep GraphRAG optimizes hierarchical search and integration for efficient, accurate information retrieval.
- Context-aware Graph Causality Inference for Few-Shot Molecular Property Prediction (viability: 7): https://sciencetostartup.com/paper/context-aware-graph-causality-inference-for-few-shot-molecular-property-prediction - CaMol enhances molecular property prediction by using causality-driven graph learning for few-shot scenarios in drug discovery.
- Learn Before Represent: Bridging Generative and Contrastive Learning for Domain-Specific LLM Embeddings (viability: 3): https://sciencetostartup.com/paper/learn-before-represent-bridging-generative-and-contrastive-learning-for-domain-specific-llm-embeddings - A framework to enhance LLM embeddings for specialized domains like chemistry and law.
- Vision-as-Inverse-Graphics Agent via Interleaved Multimodal Reasoning (viability: 8): https://sciencetostartup.com/paper/vision-as-inverse-graphics-agent-via-interleaved-multimodal-reasoning - Revolutionizing image-to-graphics editing with VIGA for versatile scene reconstruction and editing.
- ReCreate: Reasoning and Creating Domain Agents Driven by Experience (viability: 3): https://sciencetostartup.com/paper/recreate-reasoning-and-creating-domain-agents-driven-by-experience - ReCreate automates the creation of domain-specific agents by leveraging agent interaction histories for improved performance.
- Efficient Multilingual Name Type Classification Using Convolutional Networks (viability: 7): https://sciencetostartup.com/paper/efficient-multilingual-name-type-classification-using-convolutional-networks - Develop an efficient multilingual name type classification tool using convolutional networks, achieving high accuracy and speed with reduced energy consumption.
- Visual Marker Search for Autonomous Drone Landing in Diverse Urban Environments (viability: 7): https://sciencetostartup.com/paper/visual-marker-search-for-autonomous-drone-landing-in-diverse-urban-environments - Simulated drone landing system to enhance reliability in diverse urban environments using AirSim.
- ABC-Bench: Benchmarking Agentic Backend Coding in Real-World Development (viability: 7): https://sciencetostartup.com/paper/abc-bench-benchmarking-agentic-backend-coding-in-real-world-development - ABC-Bench evaluates autonomous coding agents on real-world backend development tasks to bridge the gap between AI capabilities and practical engineering demands.
- A3D: Adaptive Affordance Assembly with Dual-Arm Manipulation (viability: 7): https://sciencetostartup.com/paper/a3d-adaptive-affordance-assembly-with-dual-arm-manipulation - A3D uses dual-arm robotics for adaptive furniture assembly, promising improved automation in manufacturing.
- H-AIM: Orchestrating LLMs, PDDL, and Behavior Trees for Hierarchical Multi-Robot Planning (viability: 7): https://sciencetostartup.com/paper/h-aim-orchestrating-llms-pddl-and-behavior-trees-for-hierarchical-multi-robot-planning - Develop a multi-robot coordination tool using LLMs, PDDL, and Behavior Trees for efficient task execution in heterogeneous teams.
- Predicting Biased Human Decision-Making with Large Language Models in Conversational Settings (viability: 7): https://sciencetostartup.com/paper/predicting-biased-human-decision-making-with-large-language-models-in-conversational-settings - AI tool for predicting human cognitive biases in conversations using LLMs.
- Spectral Characterization and Mitigation of Sequential Knowledge Editing Collapse (viability: 6): https://sciencetostartup.com/paper/spectral-characterization-and-mitigation-of-sequential-knowledge-editing-collapse - REVIVE: a framework to stabilize AI model edits without sacrificing general performance.
- BAPO: Boundary-Aware Policy Optimization for Reliable Agentic Search (viability: 8): https://sciencetostartup.com/paper/bapo-boundary-aware-policy-optimization-for-reliable-agentic-search - Boundary-Aware Policy Optimization enhances reliability for LLM-driven agentic search by teaching AI to recognize its knowledge limits.
- Finding the Translation Switch: Discovering and Exploiting the Task-Initiation Features in LLMs (viability: 6): https://sciencetostartup.com/paper/finding-the-translation-switch-discovering-and-exploiting-the-task-initiation-features-in-llms - Leveraging translation initiation features in LLMs for efficient translation fine-tuning.
- Efficient Protein Optimization via Structure-aware Hamiltonian Dynamics (viability: 8): https://sciencetostartup.com/paper/efficient-protein-optimization-via-structure-aware-hamiltonian-dynamics - HADES uses Hamiltonian dynamics for efficient protein sequence optimization, enhancing drug and enzyme development.
- AdaMARP: An Adaptive Multi-Agent Interaction Framework for General Immersive Role-Playing (viability: 6): https://sciencetostartup.com/paper/adamarp-an-adaptive-multi-agent-interaction-framework-for-general-immersive-role-playing - Develop an adaptive multi-agent role-playing framework for enhanced immersive narratives with dynamic scene management.
- When Personalization Misleads: Understanding and Mitigating Hallucinations in Personalized LLMs (viability: 7): https://sciencetostartup.com/paper/when-personalization-misleads-understanding-and-mitigating-hallucinations-in-personalized-llms - Mitigate personalized language models' hallucinations with a lightweight inference approach for factual accuracy.
- Steering Language Models Before They Speak: Logit-Level Interventions (viability: 7): https://sciencetostartup.com/paper/steering-language-models-before-they-speak-logit-level-interventions - Revolutionize text generation with logit-level interventions for precise output control without retraining.
- What Matters in Data Curation for Multimodal Reasoning? Insights from the DCVLR Challenge (viability: 6): https://sciencetostartup.com/paper/what-matters-in-data-curation-for-multimodal-reasoning-insights-from-the-dcvlr-challenge - Optimize multimodal AI by focusing on difficulty-based dataset curation for enhanced reasoning performance.
- RobuMTL: Enhancing Multi-Task Learning Robustness Against Weather Conditions (viability: 8): https://sciencetostartup.com/paper/robumtl-enhancing-multi-task-learning-robustness-against-weather-conditions - RobuMTL enhances autonomous systems' robustness in adverse weather using adaptive multi-task learning.
- Self-learned representation-guided latent diffusion model for breast cancer classification in deep ultraviolet whole surface images (viability: 6): https://sciencetostartup.com/paper/self-learned-representation-guided-latent-diffusion-model-for-breast-cancer-classification-in-deep-ultraviolet-whole-sur - AI-powered breast cancer classification using synthetic data from advanced imaging.
- ARC Prize 2025: Technical Report (viability: 6): https://sciencetostartup.com/paper/arc-prize-2025-technical-report - Leveraging ARC-AGI benchmarks with refinement loops to optimize commercial AI systems.
- Medical SAM3: A Foundation Model for Universal Prompt-Driven Medical Image Segmentation (viability: 9): https://sciencetostartup.com/paper/medical-sam3-a-foundation-model-for-universal-prompt-driven-medical-image-segmentation - Medical SAM3 delivers a universal, prompt-driven segmentation model for medical imaging, solving domain shift challenges.
- Can Vision-Language Models Understand Construction Workers? An Exploratory Study (viability: 6): https://sciencetostartup.com/paper/can-vision-language-models-understand-construction-workers-an-exploratory-study - Develop a vision-language tool for recognizing construction worker actions and emotions to enhance safety and productivity.
- Approximately Optimal Global Planning for Contact-Rich SE(2) Manipulation on a Graph of Reachable Sets (viability: 7): https://sciencetostartup.com/paper/approximately-optimal-global-planning-for-contact-rich-se-2-manipulation-on-a-graph-of-reachable-sets - Develop a globally optimized planner for contact-rich robotic manipulation that significantly reduces task cost and enhances efficiency.
- Digital Metabolism: Decoupling Logic from Facts via Regenerative Unlearning -- Towards a Pure Neural Logic Core (viability: 3): https://sciencetostartup.com/paper/digital-metabolism-decoupling-logic-from-facts-via-regenerative-unlearning-towards-a-pure-neural-logic-core - Revolutionize LLM architecture by decoupling logic and facts for efficient reasoning.
- MatchTIR: Fine-Grained Supervision for Tool-Integrated Reasoning via Bipartite Matching (viability: 8): https://sciencetostartup.com/paper/matchtir-fine-grained-supervision-for-tool-integrated-reasoning-via-bipartite-matching - A fine-grained supervision framework for improving tool-integrated reasoning in large language models, outperforming larger competitors.
- Grounding Agent Memory in Contextual Intent (viability: 7): https://sciencetostartup.com/paper/grounding-agent-memory-in-contextual-intent - STITCH revolutionizes agent memory systems by grounding memory retrieval in contextual intent for robust long-horizon interactions.
- LIBERTy: A Causal Framework for Benchmarking Concept-Based Explanations of LLMs with Structural Counterfactuals (viability: 6): https://sciencetostartup.com/paper/liberty-a-causal-framework-for-benchmarking-concept-based-explanations-of-llms-with-structural-counterfactuals - LIBERTy enables reliable benchmarking of explainability in AI models using structural counterfactual datasets.
- The Impact of Generative AI on Architectural Conceptual Design: Performance, Creative Self-Efficacy and Cognitive Load (viability: 3): https://sciencetostartup.com/paper/the-impact-of-generative-ai-on-architectural-conceptual-design-performance-creative-self-efficacy-and-cognitive-load - Generative AI can enhance novice architects' design performance but may reduce creative self-efficacy.
- On the origin of neural scaling laws: from random graphs to natural language (viability: 5): https://sciencetostartup.com/paper/on-the-origin-of-neural-scaling-laws-from-random-graphs-to-natural-language - Develop a tool for optimizing AI model training using novel insights on neural scaling laws.
- Structure and Diversity Aware Context Bubble Construction for Enterprise Retrieval Augmented Systems (viability: 8): https://sciencetostartup.com/paper/structure-and-diversity-aware-context-bubble-construction-for-enterprise-retrieval-augmented-systems - Revolutionize enterprise document retrieval with a context-aware, diversity-constrained framework
- Are Your Reasoning Models Reasoning or Guessing? A Mechanistic Analysis of Hierarchical Reasoning Models (viability: 2): https://sciencetostartup.com/paper/are-your-reasoning-models-reasoning-or-guessing-a-mechanistic-analysis-of-hierarchical-reasoning-models - Analyze and improve hierarchical reasoning models with new guessing strategies.
- Multi-Property Synthesis (viability: 3): https://sciencetostartup.com/paper/multi-property-synthesis - Algorithm for optimizing property synthesis in logic-based systems.
- Molmo2: Open Weights and Data for Vision-Language Models with Video Understanding and Grounding (viability: 8): https://sciencetostartup.com/paper/molmo2-open-weights-and-data-for-vision-language-models-with-video-understanding-and-grounding - Open-source video-language models with state-of-the-art video grounding capabilities for applications in security, video search, and assistive technology.
- Procedural Fairness in Multi-Agent Bandits (viability: 2): https://sciencetostartup.com/paper/procedural-fairness-in-multi-agent-bandits - Introducing procedural fairness in multi-agent bandits for balanced decision-making power.
- ProbFM: Probabilistic Time Series Foundation Model with Uncertainty Decomposition (viability: 7): https://sciencetostartup.com/paper/probfm-probabilistic-time-series-foundation-model-with-uncertainty-decomposition - Build an AI tool for accurate financial forecasting with uncertainty quantification using a probabilistic transformer model.
- Adversarial Evasion Attacks on Computer Vision using SHAP Values (viability: 5): https://sciencetostartup.com/paper/adversarial-evasion-attacks-on-computer-vision-using-shap-values - Develop a robust adversarial attack tool for computer vision using SHAP values to test model resilience.
- From Single to Multi-Agent Reasoning: Advancing GeneGPT for Genomics QA (viability: 7): https://sciencetostartup.com/paper/from-single-to-multi-agent-reasoning-advancing-genegpt-for-genomics-qa - GenomAgent advances genomic QA using multi-agent coordination, significantly outperforming existing models and adaptable to other scientific domains.
- Generative AI collective behavior needs an interactionist paradigm (viability: 3): https://sciencetostartup.com/paper/generative-ai-collective-behavior-needs-an-interactionist-paradigm - Explores the interactionist paradigm for understanding collective behaviors in LLMs.
- Process-Guided Concept Bottleneck Model (viability: 7): https://sciencetostartup.com/paper/process-guided-concept-bottleneck-model - Process-Guided Concept Bottleneck Model improves transparency and reduces bias in scientific AI applications with interpretable intermediate outputs.
- Learning Latency-Aware Orchestration for Parallel Multi-Agent Systems (viability: 8): https://sciencetostartup.com/paper/learning-latency-aware-orchestration-for-parallel-multi-agent-systems - Optimize multi-agent system execution for reduced latency in parallel environments.
- Defending Large Language Models Against Jailbreak Attacks via In-Decoding Safety-Awareness Probing (viability: 7): https://sciencetostartup.com/paper/defending-large-language-models-against-jailbreak-attacks-via-in-decoding-safety-awareness-probing - A safety-enhancement tool for large language models that preempts jailbreak attacks by leveraging latent safety signals.
- A Safety Report on GPT-5.2, Gemini 3 Pro, Qwen3-VL, Doubao 1.8, Grok 4.1 Fast, Nano Banana Pro, and Seedream 4.5 (viability: 4): https://sciencetostartup.com/paper/a-safety-report-on-gpt-5-2-gemini-3-pro-qwen3-vl-doubao-1-8-grok-4-1-fast-nano-banana-pro-and-seedream-4-5 - Integrated safety evaluation of leading LLMs and MLLMs for better real-world risk management.
- Diagnosing Generalization Failures in Fine-Tuned LLMs: A Cross-Architectural Study on Phishing Detection (viability: 6): https://sciencetostartup.com/paper/diagnosing-generalization-failures-in-fine-tuned-llms-a-cross-architectural-study-on-phishing-detection - A diagnostic tool for improving generalization in fine-tuned LLMs for phishing detection leveraging architecture and data diversity.
- Breaking Up with Normatively Monolithic Agency with GRACE: A Reason-Based Neuro-Symbolic Architecture for Safe and Ethical AI Alignment (viability: 7): https://sciencetostartup.com/paper/breaking-up-with-normatively-monolithic-agency-with-grace-a-reason-based-neuro-symbolic-architecture-for-safe-and-ethica - Develop a neuro-symbolic architecture, GRACE, for ensuring ethical AI alignment in autonomous agents.
- SatMap: Revisiting Satellite Maps as Prior for Online HD Map Construction (viability: 7): https://sciencetostartup.com/paper/satmap-revisiting-satellite-maps-as-prior-for-online-hd-map-construction - SatMap integrates satellite imagery with multi-view cameras for superior online HD map construction, enhancing autonomous driving in challenging conditions.
- Scalable Algorithms for Approximate DNF Model Counting (viability: 3): https://sciencetostartup.com/paper/scalable-algorithms-for-approximate-dnf-model-counting - Develop a scalable Monte Carlo algorithm for efficient DNF model counting in probabilistic databases.
- Projected Microbatch Accumulation yields reference-free proximal policy updates for reinforcement learning (viability: 2): https://sciencetostartup.com/paper/projected-microbatch-accumulation-yields-reference-free-proximal-policy-updates-for-reinforcement-learning - PROMA enables stable proximal policy updates for reinforcement learning without entropy collapse.
- Model See, Model Do? Exposure-Aware Evaluation of Bug-vs-Fix Preference in Code LLMs (viability: 6): https://sciencetostartup.com/paper/model-see-model-do-exposure-aware-evaluation-of-bug-vs-fix-preference-in-code-llms - An exposure-aware evaluation framework that helps LLMs better distinguish between buggy and fixed code, reducing propagation of memorized errors.
- Panning for Gold: Expanding Domain-Specific Knowledge Graphs with General Knowledge (viability: 5): https://sciencetostartup.com/paper/panning-for-gold-expanding-domain-specific-knowledge-graphs-with-general-knowledge - Integrate general knowledge into domain-specific knowledge graphs using our ExeFuse paradigm for enhanced data coverage.
- Urban Socio-Semantic Segmentation with Vision-Language Reasoning (viability: 8): https://sciencetostartup.com/paper/urban-socio-semantic-segmentation-with-vision-language-reasoning - Revolutionizing urban planning with advanced socio-semantic segmentation from satellite imagery using vision-language models.
- ChartComplete: A Taxonomy-based Inclusive Chart Dataset (viability: 4): https://sciencetostartup.com/paper/chartcomplete-a-taxonomy-based-inclusive-chart-dataset - ChartComplete dataset offers a comprehensive taxonomy for diverse chart types, enhancing chart understanding in AI.
- Contextual StereoSet: Stress-Testing Bias Alignment Robustness in Large Language Models (viability: 6): https://sciencetostartup.com/paper/contextual-stereoset-stress-testing-bias-alignment-robustness-in-large-language-models - Introducing Contextual StereoSet, a benchmark to stress-test bias alignment robustness in large language models.
- NSR-Boost: A Neuro-Symbolic Residual Boosting Framework for Industrial Legacy Models (viability: 8): https://sciencetostartup.com/paper/nsr-boost-a-neuro-symbolic-residual-boosting-framework-for-industrial-legacy-models - Upgrade legacy industrial models with NSR-Boost for smarter risk management without intrusive retraining costs.
- AgentGuardian: Learning Access Control Policies to Govern AI Agent Behavior (viability: 7): https://sciencetostartup.com/paper/agentguardian-learning-access-control-policies-to-govern-ai-agent-behavior - AgentGuardian secures AI agents with context-aware access control, preventing misuse and errors in real-time.
- Development of Ontological Knowledge Bases by Leveraging Large Language Models (viability: 7): https://sciencetostartup.com/paper/development-of-ontological-knowledge-bases-by-leveraging-large-language-models - Automate and enhance Ontological Knowledge Base development using Large Language Models for efficient knowledge management systems.
- Are Language Models Models? (viability: 2): https://sciencetostartup.com/paper/are-language-models-models - A critique on the cognitive modeling claims of language models, emphasizing their role as tools instead.
- LLMdoctor: Token-Level Flow-Guided Preference Optimization for Efficient Test-Time Alignment of Large Language Models (viability: 7): https://sciencetostartup.com/paper/llmdoctor-token-level-flow-guided-preference-optimization-for-efficient-test-time-alignment-of-large-language-models - Efficiently aligns large language models with human preferences using a novel token-level method outperforming traditional fine-tuning.
- LADFA: A Framework of Using Large Language Models and Retrieval-Augmented Generation for Personal Data Flow Analysis in Privacy Policies (viability: 7): https://sciencetostartup.com/paper/ladfa-a-framework-of-using-large-language-models-and-retrieval-augmented-generation-for-personal-data-flow-analysis-in-p - LADFA automates the analysis of privacy policies using a combination of large language models and retrieval-augmented generation.
- ErrEval: Error-Aware Evaluation for Question Generation through Explicit Diagnostics (viability: 4): https://sciencetostartup.com/paper/erreval-error-aware-evaluation-for-question-generation-through-explicit-diagnostics - ErrEval offers an error-aware evaluation framework for more reliable question generation analysis.
- Toward Ultra-Long-Horizon Agentic Science: Cognitive Accumulation for Machine Learning Engineering (viability: 7): https://sciencetostartup.com/paper/toward-ultra-long-horizon-agentic-science-cognitive-accumulation-for-machine-learning-engineering - Build autonomous AI agents with ultra-long-horizon capabilities for complex machine learning engineering tasks.
- LatentRefusal: Latent-Signal Refusal for Unanswerable Text-to-SQL Queries (viability: 8): https://sciencetostartup.com/paper/latentrefusal-latent-signal-refusal-for-unanswerable-text-to-sql-queries - LatentRefusal ensures safe deployment of text-to-SQL systems by preemptively refusing unanswerable queries using internal LLM signals.
- Handling Missing Modalities in Multimodal Survival Prediction for Non-Small Cell Lung Cancer (viability: 6): https://sciencetostartup.com/paper/handling-missing-modalities-in-multimodal-survival-prediction-for-non-small-cell-lung-cancer - AI model for resilient multimodal survival prediction in NSCLC that handles missing modalities effectively.
- Global Context Compression with Interleaved Vision-Text Transformation (viability: 6): https://sciencetostartup.com/paper/global-context-compression-with-interleaved-vision-text-transformation - Develop a novel Transformer model, VIST2, that efficiently compresses global context in vision-language tasks to accelerate text generation and reduce computational costs.
- Towards Efficient Low-rate Image Compression with Frequency-aware Diffusion Prior Refinement (viability: 4): https://sciencetostartup.com/paper/towards-efficient-low-rate-image-compression-with-frequency-aware-diffusion-prior-refinement - Accelerate image compression with diffusion-based priors via a novel efficient framework for high-fidelity reconstructions.
- SuS: Strategy-aware Surprise for Intrinsic Exploration (viability: 5): https://sciencetostartup.com/paper/sus-strategy-aware-surprise-for-intrinsic-exploration - Develop an intrinsic motivation framework for improving exploration in reinforcement learning.
- Training-Trajectory-Aware Token Selection (viability: 8): https://sciencetostartup.com/paper/training-trajectory-aware-token-selection - Efficiently enhance AI reasoning by dynamically selecting training tokens to improve model distillation outcomes.
- OctoBench: Benchmarking Scaffold-Aware Instruction Following in Repository-Grounded Agentic Coding (viability: 6): https://sciencetostartup.com/paper/octobench-benchmarking-scaffold-aware-instruction-following-in-repository-grounded-agentic-coding - OctoBench is a comprehensive benchmark tool to improve scaffold-aware instruction following in coding agents.
- C-GRASP: Clinically-Grounded Reasoning for Affective Signal Processing (viability: 7): https://sciencetostartup.com/paper/c-grasp-clinically-grounded-reasoning-for-affective-signal-processing - Develop a clinically-grounded AI tool for accurate heart rate variability interpretation in biomedical engineering.
- Agent Skills in the Wild: An Empirical Study of Security Vulnerabilities at Scale (viability: 4): https://sciencetostartup.com/paper/agent-skills-in-the-wild-an-empirical-study-of-security-vulnerabilities-at-scale - Developing a toolkit for detecting vulnerabilities in AI agent frameworks to enhance security.
- Evidence-Augmented Policy Optimization with Reward Co-Evolution for Long-Context Reasoning (viability: 7): https://sciencetostartup.com/paper/evidence-augmented-policy-optimization-with-reward-co-evolution-for-long-context-reasoning - EAPO optimizes long-context reasoning in LLMs using evidence-augmented reinforcement learning for improved AI understanding in complex scenarios.
- DanQing: An Up-to-Date Large-Scale Chinese Vision-Language Pre-training Dataset (viability: 7): https://sciencetostartup.com/paper/danqing-an-up-to-date-large-scale-chinese-vision-language-pre-training-dataset - DanQing is a high-quality Chinese vision-language dataset poised to boost cross-modal AI performance in Chinese markets.
- SPIKE: Sparse Koopman Regularization for Physics-Informed Neural Networks (viability: 5): https://sciencetostartup.com/paper/spike-sparse-koopman-regularization-for-physics-informed-neural-networks - Develop a framework that enhances Physics-Informed Neural Networks with Koopman operator for better generalization in solving differential equations.
- Queueing-Aware Optimization of Reasoning Tokens for Accuracy-Latency Trade-offs in LLM Servers (viability: 3): https://sciencetostartup.com/paper/queueing-aware-optimization-of-reasoning-tokens-for-accuracy-latency-trade-offs-in-llm-servers - Optimize server token allocation for balancing accuracy and latency in LLM servers.
- MoST: Mixing Speech and Text with Modality-Aware Mixture of Experts (viability: 8): https://sciencetostartup.com/paper/most-mixing-speech-and-text-with-modality-aware-mixture-of-experts - MoST integrates speech and text processing into an efficient open-source modality-aware language model, outpacing existing solutions in seamless interaction tasks.
- Untangling Input Language from Reasoning Language: A Diagnostic Framework for Cross-Lingual Moral Alignment in LLMs (viability: 7): https://sciencetostartup.com/paper/untangling-input-language-from-reasoning-language-a-diagnostic-framework-for-cross-lingual-moral-alignment-in-llms - A diagnostic tool for cross-lingual moral alignment in LLMs using a new evaluation framework.
- NoReGeo: Non-Reasoning Geometry Benchmark (viability: 4): https://sciencetostartup.com/paper/noregeo-non-reasoning-geometry-benchmark - NoReGeo offers a benchmark to evaluate LLMs' intrinsic geometric understanding, highlighting gaps in current models.
- X-SAM: Boosting Sharpness-Aware Minimization with Dominant-Eigenvector Gradient Correction (viability: 3): https://sciencetostartup.com/paper/x-sam-boosting-sharpness-aware-minimization-with-dominant-eigenvector-gradient-correction - Enhance Sharpness-Aware Minimization with eigenvector-aligned gradient correction for better generalization.
- TRIM: Hybrid Inference via Targeted Stepwise Routing in Multi-Step Reasoning Tasks (viability: 7): https://sciencetostartup.com/paper/trim-hybrid-inference-via-targeted-stepwise-routing-in-multi-step-reasoning-tasks - Optimize inference efficiency in multi-step reasoning tasks by routing critical steps to larger models with TRIM.
- Loop as a Bridge: Can Looped Transformers Truly Link Representation Space and Natural Language Outputs? (viability: 7.5): https://sciencetostartup.com/paper/loop-as-a-bridge-can-looped-transformers-truly-link-representation-space-and-natural-language-outputs - Exploring the potential of Looped Transformers to improve introspection in language models.
- Who Owns the Text? Design Patterns for Preserving Authorship in AI-Assisted Writing (viability: 3): https://sciencetostartup.com/paper/who-owns-the-text-design-patterns-for-preserving-authorship-in-ai-assisted-writing - Develop an AI-assisted writing tool that preserves authorship using style personalization and persona-based coaching.
- Topo-RAG: Topology-aware retrieval for hybrid text-table documents (viability: 7): https://sciencetostartup.com/paper/topo-rag-topology-aware-retrieval-for-hybrid-text-table-documents - Develop a more accurate document retrieval system by preserving the topology of hybrid text and table data.
- PADER: Paillier-based Secure Decentralized Social Recommendation (viability: 7): https://sciencetostartup.com/paper/pader-paillier-based-secure-decentralized-social-recommendation - PADER offers privacy-preserving, decentralized social recommendation using Paillier cryptosystem.
- One Instruction Does Not Fit All: How Well Do Embeddings Align Personas and Instructions in Low-Resource Indian Languages? (viability: 6): https://sciencetostartup.com/paper/one-instruction-does-not-fit-all-how-well-do-embeddings-align-personas-and-instructions-in-low-resource-indian-languages - A benchmarking platform for evaluating multilingual embeddings on aligning user personas with instructions for Indian languages.
- PRL: Process Reward Learning Improves LLMs' Reasoning Ability and Broadens the Reasoning Boundary (viability: 5): https://sciencetostartup.com/paper/prl-process-reward-learning-improves-llms-reasoning-ability-and-broadens-the-reasoning-boundary - Enhance LLM reasoning with Process Reward Learning for more granular process-based optimization.
- GFM4GA: Graph Foundation Model for Group Anomaly Detection (viability: 3): https://sciencetostartup.com/paper/gfm4ga-graph-foundation-model-for-group-anomaly-detection - Develops graph foundation model for improved group anomaly detection in networks.
- How does downsampling affect needle electromyography signals? A generalisable workflow for understanding downsampling effects on high-frequency time series (viability: 6): https://sciencetostartup.com/paper/how-does-downsampling-affect-needle-electromyography-signals-a-generalisable-workflow-for-understanding-downsampling-eff - A workflow to optimize downsampling of nEMG signals for efficient neuromuscular disease detection.
- HOMURA: Taming the Sand-Glass for Time-Constrained LLM Translation via Reinforcement Learning (viability: 8): https://sciencetostartup.com/paper/homura-taming-the-sand-glass-for-time-constrained-llm-translation-via-reinforcement-learning - Reinforcement learning framework to optimize translations for time-constrained media with precise syllable-level duration compliance.
- ReasAlign: Reasoning Enhanced Safety Alignment against Prompt Injection Attack (viability: 8): https://sciencetostartup.com/paper/reasalign-reasoning-enhanced-safety-alignment-against-prompt-injection-attack - ReasAlign provides enhanced safety alignment for LLMs against prompt injection attacks using reasoning techniques.
- CtD: Composition through Decomposition in Emergent Communication (viability: 4): https://sciencetostartup.com/paper/ctd-composition-through-decomposition-in-emergent-communication - Develop a system enabling AI agents to describe unseen images through learned compositionality.
- RAG-3DSG: Enhancing 3D Scene Graphs with Re-Shot Guided Retrieval-Augmented Generation (viability: 6): https://sciencetostartup.com/paper/rag-3dsg-enhancing-3d-scene-graphs-with-re-shot-guided-retrieval-augmented-generation - Innovative solution for efficient and accurate 3D scene graph generation for robotics using retrieval-augmented generation.
- AWED-FiNER: Agents, Web applications, and Expert Detectors for Fine-grained Named Entity Recognition across 36 Languages for 6.6 Billion Speakers (viability: 7): https://sciencetostartup.com/paper/awed-finer-agents-web-applications-and-expert-detectors-for-fine-grained-named-entity-recognition-across-36-languages-fo - AWED-FiNER is an open-source ecosystem for fine-grained named entity recognition across 36 languages, accessible via agentic tools and web apps.
- Alignment Pretraining: AI Discourse Causes Self-Fulfilling (Mis)alignment (viability: 8): https://sciencetostartup.com/paper/alignment-pretraining-ai-discourse-causes-self-fulfilling-mis-alignment - Develop AI systems with inherent alignment by leveraging discourse-influenced pretraining techniques.
- MMPG: MoE-based Adaptive Multi-Perspective Graph Fusion for Protein Representation Learning (viability: 5): https://sciencetostartup.com/paper/mmpg-moe-based-adaptive-multi-perspective-graph-fusion-for-protein-representation-learning - Develop an adaptive multi-perspective graph fusion framework for enhanced protein representation learning.
- LOOKAT: Lookup-Optimized Key-Attention for Memory-Efficient Transformers (viability: 9): https://sciencetostartup.com/paper/lookat-lookup-optimized-key-attention-for-memory-efficient-transformers - Introduce LOOKAT to significantly compress KV-cache for edge deployment without architecture changes.
- MHub.ai: A Simple, Standardized, and Reproducible Platform for AI Models in Medical Imaging (viability: 7): https://sciencetostartup.com/paper/mhub-ai-a-simple-standardized-and-reproducible-platform-for-ai-models-in-medical-imaging - MHub.ai is a standardized platform for deploying and reproducibly benchmarking AI models in medical imaging.
- Simple Network Graph Comparative Learning (viability: 3): https://sciencetostartup.com/paper/simple-network-graph-comparative-learning - Introducing a novel graph contrastive learning method for improved node classification.
- DecisionLLM: Large Language Models for Long Sequence Decision Exploration (viability: 7): https://sciencetostartup.com/paper/decisionllm-large-language-models-for-long-sequence-decision-exploration - DecisionLLM leverages large language models for enhanced performance in long-sequence decision-making tasks like real-time bidding.
- History Is Not Enough: An Adaptive Dataflow System for Financial Time-Series Synthesis (viability: 8): https://sciencetostartup.com/paper/history-is-not-enough-an-adaptive-dataflow-system-for-financial-time-series-synthesis - An adaptive dataflow system that improves financial trading model robustness and performance through dynamic data management and automation.
- Understanding and Preserving Safety in Fine-Tuned LLMs (viability: 3): https://sciencetostartup.com/paper/understanding-and-preserving-safety-in-fine-tuned-llms - Develop a lightweight approach for safety-preserving fine-tuning of LLMs to maintain both utility and safety alignment.
- Step-by-Step Causality: Transparent Causal Discovery with Multi-Agent Tree-Query and Adversarial Confidence Estimation (viability: 5): https://sciencetostartup.com/paper/step-by-step-causality-transparent-causal-discovery-with-multi-agent-tree-query-and-adversarial-confidence-estimation - A multi-agent system for transparent causal discovery with robust confidence estimation.
- Is More Context Always Better? Examining LLM Reasoning Capability for Time Interval Prediction (viability: 5): https://sciencetostartup.com/paper/is-more-context-always-better-examining-llm-reasoning-capability-for-time-interval-prediction - Exploring the limitations of LLMs in predicting time intervals to improve future context-aware models.
- M^4olGen: Multi-Agent, Multi-Stage Molecular Generation under Precise Multi-Property Constraints (viability: 7): https://sciencetostartup.com/paper/m-4olgen-multi-agent-multi-stage-molecular-generation-under-precise-multi-property-constraints - M^4olGen offers a novel framework for generating molecules under precise property constraints, outperforming existing models.
- Redundancy-Driven Top-$k$ Functional Dependency Discovery (viability: 6): https://sciencetostartup.com/paper/redundancy-driven-top-k-functional-dependency-discovery - Revolutionizing database management with fast, efficient top-k functional dependency discovery tool.
- LaViT: Aligning Latent Visual Thoughts for Multi-modal Reasoning (viability: 7): https://sciencetostartup.com/paper/lavit-aligning-latent-visual-thoughts-for-multi-modal-reasoning - LaViT offers advanced visual grounding techniques enhancing multimodal reasoning capabilities for compact models.
- Role-Playing Agents Driven by Large Language Models: Current Status, Challenges, and Future Trends (viability: 5): https://sciencetostartup.com/paper/role-playing-agents-driven-by-large-language-models-current-status-challenges-and-future-trends - Develop advanced role-playing language agents leveraging large language models for immersive human-computer interaction experiences.
- TopoDIM: One-shot Topology Generation of Diverse Interaction Modes for Multi-Agent Systems (viability: 7): https://sciencetostartup.com/paper/topodim-one-shot-topology-generation-of-diverse-interaction-modes-for-multi-agent-systems - TopoDIM optimizes multi-agent communication topology to enhance efficiency and performance with decentralized, adaptive interaction modes.
- Following the Teacher's Footsteps: Scheduled Checkpoint Distillation for Domain-Specific LLMs (viability: 3): https://sciencetostartup.com/paper/following-the-teacher-s-footsteps-scheduled-checkpoint-distillation-for-domain-specific-llms - Efficiently distill large language models for domain-specific tasks using Scheduled Checkpoint Distillation.
- Repository Intelligence Graph: Deterministic Architectural Map for LLM Code Assistants (viability: 6): https://sciencetostartup.com/paper/repository-intelligence-graph-deterministic-architectural-map-for-llm-code-assistants - Repository Intelligence Graph (RIG) enhances LLM code assistants' accuracy and efficiency by providing a detailed architectural map of software repositories.
- SIN-Bench: Tracing Native Evidence Chains in Long-Context Multimodal Scientific Interleaved Literature (viability: 4): https://sciencetostartup.com/paper/sin-bench-tracing-native-evidence-chains-in-long-context-multimodal-scientific-interleaved-literature - Develop tools to enhance the evidence-based reasoning capabilities of multimodal language models in scientific literature.
- MathDoc: Benchmarking Structured Extraction and Active Refusal on Noisy Mathematics Exam Papers (viability: 7): https://sciencetostartup.com/paper/mathdoc-benchmarking-structured-extraction-and-active-refusal-on-noisy-mathematics-exam-papers - Build a novel AI tool for structured question extraction from noisy mathematics exam papers using the MathDoc benchmark.
- FlowAct-R1: Towards Interactive Humanoid Video Generation (viability: 9): https://sciencetostartup.com/paper/flowact-r1-towards-interactive-humanoid-video-generation - FlowAct-R1 generates lifelike interactive humanoid videos in real-time for virtual avatars and digital companions.
- MATRIX AS PLAN: Structured Logical Reasoning with Feedback-Driven Replanning (viability: 7): https://sciencetostartup.com/paper/matrix-as-plan-structured-logical-reasoning-with-feedback-driven-replanning - MatrixCoT enhances LLMs' logical reasoning through structured matrix planning and feedback-driven replanning.
- V-Zero: Self-Improving Multimodal Reasoning with Zero Annotation (viability: 7): https://sciencetostartup.com/paper/v-zero-self-improving-multimodal-reasoning-with-zero-annotation - Develop a post-training framework for multimodal reasoning that enhances vision-language models without needing human-annotated data.
- LeMoF: Level-guided Multimodal Fusion for Heterogeneous Clinical Data (viability: 7): https://sciencetostartup.com/paper/lemof-level-guided-multimodal-fusion-for-heterogeneous-clinical-data - LeMoF provides a novel multimodal fusion framework enhancing clinical predictions from heterogeneous datasets like EHRs and biosignals.
- Difficulty-guided Sampling: Bridging the Target Gap between Dataset Distillation and Downstream Tasks (viability: 6): https://sciencetostartup.com/paper/difficulty-guided-sampling-bridging-the-target-gap-between-dataset-distillation-and-downstream-tasks - A method that enhances dataset distillation with difficulty-guided sampling to optimize deep learning model training.
- State of AI: An Empirical 100 Trillion Token Study with OpenRouter (viability: 2): https://sciencetostartup.com/paper/state-of-ai-an-empirical-100-trillion-token-study-with-openrouter - Deep dive into real-world LLM usage reveals multifaceted engagement patterns for AI developers.
- Sparse-RL: Breaking the Memory Wall in LLM Reinforcement Learning via Stable Sparse Rollouts (viability: 3): https://sciencetostartup.com/paper/sparse-rl-breaking-the-memory-wall-in-llm-reinforcement-learning-via-stable-sparse-rollouts - Sparse-RL reduces memory overhead in RL training for LLMs by using sparse rollouts without performance loss.
- ReaMIL: Reasoning- and Evidence-Aware Multiple Instance Learning for Whole-Slide Histopathology (viability: 4): https://sciencetostartup.com/paper/reamil-reasoning-and-evidence-aware-multiple-instance-learning-for-whole-slide-histopathology - ReaMIL enhances multiple instance learning in histopathology by efficiently selecting evidence while maintaining accuracy.
- CoF-T2I: Video Models as Pure Visual Reasoners for Text-to-Image Generation (viability: 6): https://sciencetostartup.com/paper/cof-t2i-video-models-as-pure-visual-reasoners-for-text-to-image-generation - Leverage video model reasoning to enhance text-to-image generation quality and interpretability.
- FilDeep: Learning Large Deformations of Elastic-Plastic Solids with Multi-Fidelity Data (viability: 7): https://sciencetostartup.com/paper/fildeep-learning-large-deformations-of-elastic-plastic-solids-with-multi-fidelity-data - FilDeep enables accurate computational modeling for manufacturing large deformations using multi-fidelity data in a first-of-its-kind DL framework.
- PaperScout: An Autonomous Agent for Academic Paper Search with Process-Aware Sequence-Level Policy Optimization (viability: 6): https://sciencetostartup.com/paper/paperscout-an-autonomous-agent-for-academic-paper-search-with-process-aware-sequence-level-policy-optimization - Develops an adaptive agent for dynamic academic paper search using advanced sequence-level policy optimization.
- Structured Personality Control and Adaptation for LLM Agents (viability: 3): https://sciencetostartup.com/paper/structured-personality-control-and-adaptation-for-llm-agents - Develop a framework for LLMs to exhibit adaptable personalities using Jungian psychological types for enhanced human-computer interaction.
- Empowering Older Adults in Digital Technology Use with Foundation Models (viability: 6): https://sciencetostartup.com/paper/empowering-older-adults-in-digital-technology-use-with-foundation-models - An AI-powered tool for paraphrasing and clarifying tech support queries from older adults to enhance problem resolution.
- Memo-SQL: Structured Decomposition and Experience-Driven Self-Correction for Training-Free NL2SQL (viability: 5): https://sciencetostartup.com/paper/memo-sql-structured-decomposition-and-experience-driven-self-correction-for-training-free-nl2sql - Memo-SQL offers a resource-efficient, training-free framework for NL2SQL systems using structured decomposition and experience-aware self-correction.
- VERHallu: Evaluating and Mitigating Event Relation Hallucination in Video Large Language Models (viability: 7): https://sciencetostartup.com/paper/verhallu-evaluating-and-mitigating-event-relation-hallucination-in-video-large-language-models - VERHallu: A benchmark and toolset to mitigate event relation hallucination in video language models, improving model accuracy without slowing performance.
- Context Volume Drives Performance: Tackling Domain Shift in Extremely Low-Resource Translation via RAG (viability: 5): https://sciencetostartup.com/paper/context-volume-drives-performance-tackling-domain-shift-in-extremely-low-resource-translation-via-rag - A hybrid NMT and LLM framework significantly reduces domain shift in low-resource language translation.
- SPRInG: Continual LLM Personalization via Selective Parametric Adaptation and Retrieval-Interpolated Generation (viability: 7): https://sciencetostartup.com/paper/spring-continual-llm-personalization-via-selective-parametric-adaptation-and-retrieval-interpolated-generation - SPRInG provides continual personalization of Large Language Models, addressing preference drift with robust semi-parametric adaptation.
- Chinese Labor Law Large Language Model Benchmark (viability: 8): https://sciencetostartup.com/paper/chinese-labor-law-large-language-model-benchmark - Specialized AI model optimized for Chinese labor law applications, enhancing legal practices' efficiency and accuracy.
- A Sustainable AI Economy Needs Data Deals That Work for Generators (viability: 2): https://sciencetostartup.com/paper/a-sustainable-ai-economy-needs-data-deals-that-work-for-generators - Framework for equitable data-value exchange in AI ecosystems.
- Kinematic Tokenization: Optimization-Based Continuous-Time Tokens for Learnable Decision Policies in Noisy Time Series (viability: 3): https://sciencetostartup.com/paper/kinematic-tokenization-optimization-based-continuous-time-tokens-for-learnable-decision-policies-in-noisy-time-series - Develop a continuous-time tokenization method for improving decision policies in noisy financial time series.
- Malware Classification using Diluted Convolutional Neural Network with Fast Gradient Sign Method (viability: 5): https://sciencetostartup.com/paper/malware-classification-using-diluted-convolutional-neural-network-with-fast-gradient-sign-method - Develop an efficient Android malware detection tool using FGSM and DICNN for high accuracy with reduced feature sets.
- Hallucination Detection and Mitigation in Large Language Models (viability: 7): https://sciencetostartup.com/paper/hallucination-detection-and-mitigation-in-large-language-models - A framework to detect and mitigate hallucinations in AI models, enhancing trust in high-stakes domains like finance.
- CaMeLs Can Use Computers Too: System-level Security for Computer Use Agents (viability: 6): https://sciencetostartup.com/paper/camels-can-use-computers-too-system-level-security-for-computer-use-agents - Secure computer use agents with Dual-LLM architecture to prevent prompt injection attacks.
- Continuum Memory Architectures for Long-Horizon LLM Agents (viability: 3): https://sciencetostartup.com/paper/continuum-memory-architectures-for-long-horizon-llm-agents - Develop advanced memory architectures to enhance retrieval-augmented generation for long-horizon language agents.
- A Novel Contrastive Loss for Zero-Day Network Intrusion Detection (viability: 8): https://sciencetostartup.com/paper/a-novel-contrastive-loss-for-zero-day-network-intrusion-detection - Revolutionize network security with a novel contrastive learning algorithm that excels in zero-day threat detection.
- The Algorithmic Gaze: An Audit and Ethnography of the LAION-Aesthetics Predictor Model (viability: 5): https://sciencetostartup.com/paper/the-algorithmic-gaze-an-audit-and-ethnography-of-the-laion-aesthetics-predictor-model - Audit tool for exposing biases in AI aesthetic evaluation models like LAION Aesthetic Predictor.
- Beyond Rule-Based Workflows: An Information-Flow-Orchestrated Multi-Agents Paradigm via Agent-to-Agent Communication from CORAL (viability: 8): https://sciencetostartup.com/paper/beyond-rule-based-workflows-an-information-flow-orchestrated-multi-agents-paradigm-via-agent-to-agent-communication-from - Transform workflows from rule-based decision trees to dynamic agent communication for better task handling and efficiency.
- Transition Matching Distillation for Fast Video Generation (viability: 8): https://sciencetostartup.com/paper/transition-matching-distillation-for-fast-video-generation - A framework for real-time, high-quality video generation using efficient few-step video diffusion models.
- MedVL-SAM2: A unified 3D medical vision-language model for multimodal reasoning and prompt-driven segmentation (viability: 8): https://sciencetostartup.com/paper/medvl-sam2-a-unified-3d-medical-vision-language-model-for-multimodal-reasoning-and-prompt-driven-segmentation - A unified 3D medical vision-language model for advanced multimodal reasoning and precise 3D segmentation.
- Epistemology gives a Future to Complementarity in Human-AI Interactions (viability: 2): https://sciencetostartup.com/paper/epistemology-gives-a-future-to-complementarity-in-human-ai-interactions - Explore how epistemological approaches can enhance decision reliability in Human-AI collaborations.
- A Scoping Review of the Ethical Perspectives on Anthropomorphising Large Language Model-Based Conversational Agents (viability: 2): https://sciencetostartup.com/paper/a-scoping-review-of-the-ethical-perspectives-on-anthropomorphising-large-language-model-based-conversational-agents - A comprehensive review of the ethical implications of anthropomorphizing LLM-based conversational agents.
- Advancing Model Refinement: Muon-Optimized Distillation and Quantization for LLM Deployment (viability: 8): https://sciencetostartup.com/paper/advancing-model-refinement-muon-optimized-distillation-and-quantization-for-llm-deployment - Deploy LLMs on edge devices using advanced Muon-optimized distillation and quantization for efficient inference.
- OUTLINEFORGE: Hierarchical Reinforcement Learning with Explicit States for Scientific Writing (viability: 8): https://sciencetostartup.com/paper/outlineforge-hierarchical-reinforcement-learning-with-explicit-states-for-scientific-writing - A reinforcement learning framework optimizing scientific writing by enhancing document planning, coherence, and citation accuracy.
- Thinking Long, but Short: Stable Sequential Test-Time Scaling for Large Reasoning Models (viability: 2): https://sciencetostartup.com/paper/thinking-long-but-short-stable-sequential-test-time-scaling-for-large-reasoning-models - Introducing Min-Seek, a test-time scaling method that stabilizes reasoning model accuracy without fine-tuning.
- MedRedFlag: Investigating how LLMs Redirect Misconceptions in Real-World Health Communication (viability: 6): https://sciencetostartup.com/paper/medredflag-investigating-how-llms-redirect-misconceptions-in-real-world-health-communication - MedRedFlag enhances LLM safety in medical communication by curating a dataset for better redirection of misconceptions in patient queries.
- ViSIL: Unified Evaluation of Information Loss in Multimodal Video Captioning (viability: 8): https://sciencetostartup.com/paper/visil-unified-evaluation-of-information-loss-in-multimodal-video-captioning - Optimizing multimodal video captioning using the ViSIL score for efficient information retention and retrieval.
- A pipeline for enabling path-specific causal fairness in observational health data (viability: 6): https://sciencetostartup.com/paper/a-pipeline-for-enabling-path-specific-causal-fairness-in-observational-health-data - A pipeline enabling path-specific causal fairness for machine learning models in healthcare to address biases.
- LLM-Based Agentic Systems for Software Engineering: Challenges and Opportunities (viability: 4): https://sciencetostartup.com/paper/llm-based-agentic-systems-for-software-engineering-challenges-and-opportunities - Explore LLM multi-agent systems to enhance software engineering processes.
- Explainable Deep Learning for Pediatric Pneumonia Detection in Chest X-Ray Images (viability: 8): https://sciencetostartup.com/paper/explainable-deep-learning-for-pediatric-pneumonia-detection-in-chest-x-ray-images - AI-based diagnostic tool for accurate pediatric pneumonia detection using explainable deep learning.
- QFed: Parameter-Compact Quantum-Classical Federated Learning (viability: 7): https://sciencetostartup.com/paper/qfed-parameter-compact-quantum-classical-federated-learning - QFed offers a quantum-enhanced federated learning framework that slashes parameter counts while maintaining model accuracy, suitable for edge and IoT devices.
- Diffusion-Driven Deceptive Patches: Adversarial Manipulation and Forensic Detection in Facial Identity Verification (viability: 6): https://sciencetostartup.com/paper/diffusion-driven-deceptive-patches-adversarial-manipulation-and-forensic-detection-in-facial-identity-verification - Commercialize a pipeline for adversarial patch creation and detection in facial recognition systems.
- Improving Chain-of-Thought for Logical Reasoning via Attention-Aware Intervention (viability: 7): https://sciencetostartup.com/paper/improving-chain-of-thought-for-logical-reasoning-via-attention-aware-intervention - Develop an efficient logical reasoning enhancer for LLMs using Attention-Aware Intervention without external tools.
- Fast-ThinkAct: Efficient Vision-Language-Action Reasoning via Verbalizable Latent Planning (viability: 7): https://sciencetostartup.com/paper/fast-thinkact-efficient-vision-language-action-reasoning-via-verbalizable-latent-planning - Fast-ThinkAct optimizes embodied AI for faster and smarter action execution in visual-language tasks.
- Value-Aware Numerical Representations for Transformer Language Models (viability: 6): https://sciencetostartup.com/paper/value-aware-numerical-representations-for-transformer-language-models - Enhance language models to handle numerical data robustly with value-aware token embeddings.
- ShortCoder: Knowledge-Augmented Syntax Optimization for Token-Efficient Code Generation (viability: 8): https://sciencetostartup.com/paper/shortcoder-knowledge-augmented-syntax-optimization-for-token-efficient-code-generation - ShortCoder optimizes code generation by reducing token usage while maintaining functionality and readability.
- LLMs can Compress LLMs: Adaptive Pruning by Agents (viability: 8): https://sciencetostartup.com/paper/llms-can-compress-llms-adaptive-pruning-by-agents - AI agent-guided pruning enhances LLM compression without retraining, reducing costs while retaining performance.
- Routing with Generated Data: Annotation-Free LLM Skill Estimation and Expert Selection (viability: 5): https://sciencetostartup.com/paper/routing-with-generated-data-annotation-free-llm-skill-estimation-and-expert-selection - An innovative system to optimize model selection by training LLM routers solely on generated data without need for annotations.
- Disentangling Task Conflicts in Multi-Task LoRA via Orthogonal Gradient Projection (viability: 5): https://sciencetostartup.com/paper/disentangling-task-conflicts-in-multi-task-lora-via-orthogonal-gradient-projection - Unlock robust multi-task performance in LLMs with Ortho-LoRA's innovative gradient projection for parameter-efficient training.
- Automating Supply Chain Disruption Monitoring via an Agentic AI Approach (viability: 8): https://sciencetostartup.com/paper/automating-supply-chain-disruption-monitoring-via-an-agentic-ai-approach - Revolutionizing supply chain resilience with agentic AI for autonomous disruption monitoring and mitigation.
- Collaborative Multi-Agent Test-Time Reinforcement Learning for Reasoning (viability: 7): https://sciencetostartup.com/paper/collaborative-multi-agent-test-time-reinforcement-learning-for-reasoning - Develop a multi-agent reinforcement learning API that enhances decision-making accuracy in complex domains like medicine and education by integrating test-time experiences.
- PersonalAlign: Hierarchical Implicit Intent Alignment for Personalized GUI Agent with Long-Term User-Centric Records (viability: 8): https://sciencetostartup.com/paper/personalalign-hierarchical-implicit-intent-alignment-for-personalized-gui-agent-with-long-term-user-centric-records - PersonalAlign transforms GUIs into proactive, personalized agents that align with user implicit intents.
- LLM for Large-Scale Optimization Model Auto-Formulation: A Lightweight Few-Shot Learning Approach (viability: 8): https://sciencetostartup.com/paper/llm-for-large-scale-optimization-model-auto-formulation-a-lightweight-few-shot-learning-approach - Streamline large-scale business optimization with an LLM-driven auto-formulation tool enhanced by benchmarks and real-world use cases.
- From Prompt to Protocol: Fast Charging Batteries with Large Language Models (viability: 7): https://sciencetostartup.com/paper/from-prompt-to-protocol-fast-charging-batteries-with-large-language-models - LLMs optimize battery charging protocols, outperforming traditional methods for better capacity retention.
- The Promptware Kill Chain: How Prompt Injections Gradually Evolved Into a Multi-Step Malware (viability: 5): https://sciencetostartup.com/paper/the-promptware-kill-chain-how-prompt-injections-gradually-evolved-into-a-multi-step-malware - Structured framework to model and address evolving malware threats in LLM-based systems, termed promptware.
- Toward Understanding Unlearning Difficulty: A Mechanistic Perspective and Circuit-Guided Difficulty Metric (viability: 5): https://sciencetostartup.com/paper/toward-understanding-unlearning-difficulty-a-mechanistic-perspective-and-circuit-guided-difficulty-metric - Develop a metric system to assess machine unlearning difficulty using model circuit mechanisms.
- Full Disclosure, Less Trust? How the Level of Detail about AI Use in News Writing Affects Readers' Trust (viability: 3): https://sciencetostartup.com/paper/full-disclosure-less-trust-how-the-level-of-detail-about-ai-use-in-news-writing-affects-readers-trust - Study on how the level of AI detail in news affects reader trust.
- CogRail: Benchmarking VLMs in Cognitive Intrusion Perception for Intelligent Railway Transportation Systems (viability: 7): https://sciencetostartup.com/paper/cograil-benchmarking-vlms-in-cognitive-intrusion-perception-for-intelligent-railway-transportation-systems - CogRail enhances railway safety by fine-tuning visual-language models for cognitive intrusion perception.
- DPWriter: Reinforcement Learning with Diverse Planning Branching for Creative Writing (viability: 5): https://sciencetostartup.com/paper/dpwriter-reinforcement-learning-with-diverse-planning-branching-for-creative-writing - Platform to enhance creative writing in LLMs using novel reinforcement learning techniques for improved output diversity.
- Sim2real Image Translation Enables Viewpoint-Robust Policies from Fixed-Camera Datasets (viability: 8): https://sciencetostartup.com/paper/sim2real-image-translation-enables-viewpoint-robust-policies-from-fixed-camera-datasets - MANGO enables robust robot vision policies through sim2real image translation, leveraging viewpoint diversity from simulations.
- Linear Complexity Self-Supervised Learning for Music Understanding with Random Quantizer (viability: 6): https://sciencetostartup.com/paper/linear-complexity-self-supervised-learning-for-music-understanding-with-random-quantizer - Optimize music information retrieval systems with a lighter foundation model using self-supervised learning.
- Information Access of the Oppressed: A Problem-Posing Framework for Envisioning Emancipatory Information Access Platforms (viability: 3): https://sciencetostartup.com/paper/information-access-of-the-oppressed-a-problem-posing-framework-for-envisioning-emancipatory-information-access-platforms - Develop a problem-posing framework for emancipatory information access platforms to counter authoritarian capture.
- Hot-Start from Pixels: Low-Resolution Visual Tokens for Chinese Language Modeling (viability: 6): https://sciencetostartup.com/paper/hot-start-from-pixels-low-resolution-visual-tokens-for-chinese-language-modeling - Develop a Chinese language model using low-resolution character images for improved efficiency and prediction.
- Benchmarking Post-Training Quantization of Large Language Models under Microscaling Floating Point Formats (viability: 7): https://sciencetostartup.com/paper/benchmarking-post-training-quantization-of-large-language-models-under-microscaling-floating-point-formats - Optimize large language model inference with advanced low-precision quantization for enhanced computational efficiency.
- Omni-R1: Towards the Unified Generative Paradigm for Multimodal Reasoning (viability: 2): https://sciencetostartup.com/paper/omni-r1-towards-the-unified-generative-paradigm-for-multimodal-reasoning - Unified generative paradigm for diverse multimodal reasoning tasks.
- Private LLM Inference on Consumer Blackwell GPUs: A Practical Guide for Cost-Effective Local Deployment in SMEs (viability: 8.7): https://sciencetostartup.com/paper/private-llm-inference-on-consumer-blackwell-gpus-a-practical-guide-for-cost-effective-local-deployment-in-smes - Deploy cost-effective private LLM inference on consumer GPUs for SMEs, enhancing privacy and reducing costs.
- Towards Realistic Synthetic Data for Automatic Drum Transcription (viability: 8): https://sciencetostartup.com/paper/towards-realistic-synthetic-data-for-automatic-drum-transcription - This startup leverages a novel semi-supervised method to generate high-quality synthetic data for automatic drum transcription, outperforming existing solutions.
- Learning Whole-Body Human-Humanoid Interaction from Human-Human Demonstrations (viability: 8): https://sciencetostartup.com/paper/learning-whole-body-human-humanoid-interaction-from-human-human-demonstrations - Launch an advanced humanoid robot interaction system leveraging PAIR and D-STAR technologies to enhance synchronized human-robot collaboration.
- What Do LLM Agents Know About Their World? Task2Quiz: A Paradigm for Studying Environment Understanding (viability: 5): https://sciencetostartup.com/paper/what-do-llm-agents-know-about-their-world-task2quiz-a-paradigm-for-studying-environment-understanding - Develop a benchmarking tool to assess LLM agents' environment understanding using Task-to-Quiz paradigm.
- Bridging Semantic Understanding and Popularity Bias with LLMs (viability: 7): https://sciencetostartup.com/paper/bridging-semantic-understanding-and-popularity-bias-with-llms - FairLRM offers a novel framework enhancing recommender systems by addressing semantic understanding and fairness in popularity bias using large language models.
- SimMerge: Learning to Select Merge Operators from Similarity Signals (viability: 6): https://sciencetostartup.com/paper/simmerge-learning-to-select-merge-operators-from-similarity-signals - SimMerge: Predictive merge-selection method for scalable composition of large language models, optimizing model performance without costly searches.
- FairGU: Fairness-aware Graph Unlearning in Social Network (viability: 8): https://sciencetostartup.com/paper/fairgu-fairness-aware-graph-unlearning-in-social-network - FairGU offers a fairness-aware graph unlearning solution for safeguarding privacy and maintaining algorithmic fairness in social networks.
- Searth Transformer: A Transformer Architecture Incorporating Earth's Geospheric Physical Priors for Global Mid-Range Weather Forecasting (viability: 3): https://sciencetostartup.com/paper/searth-transformer-a-transformer-architecture-incorporating-earth-s-geospheric-physical-priors-for-global-mid-range-weat - A transformer model integrating Earth's physical priors for improved medium-range weather forecasting.
- EvoFSM: Controllable Self-Evolution for Deep Research with Finite State Machines (viability: 7): https://sciencetostartup.com/paper/evofsm-controllable-self-evolution-for-deep-research-with-finite-state-machines - Develop a controllable self-evolving AI tool using finite state machines for adaptable deep research applications.
- SoK: Enhancing Cryptographic Collaborative Learning with Differential Privacy (viability: 5): https://sciencetostartup.com/paper/sok-enhancing-cryptographic-collaborative-learning-with-differential-privacy - A unified cryptographic and differential privacy framework for collaborative machine learning.
- Where Knowledge Collides: A Mechanistic Study of Intra-Memory Knowledge Conflict in Language Models (viability: 5): https://sciencetostartup.com/paper/where-knowledge-collides-a-mechanistic-study-of-intra-memory-knowledge-conflict-in-language-models - A framework to identify and control conflicting knowledge in language models using mechanistic interpretability methods.
- Do Transformers Understand Ancient Roman Coin Motifs Better than CNNs? (viability: 2): https://sciencetostartup.com/paper/do-transformers-understand-ancient-roman-coin-motifs-better-than-cnns - Apply Vision Transformers for analyzing semantic motifs on ancient Roman coins to outperform traditional CNN methods.
- Long-term Task-oriented Agent: Proactive Long-term Intent Maintenance in Dynamic Environments (viability: 7): https://sciencetostartup.com/paper/long-term-task-oriented-agent-proactive-long-term-intent-maintenance-in-dynamic-environments - Proactive Task-oriented Agents for dynamic environments that adapt to user intent shifts.
- Query Languages for Machine-Learning Models (viability: 3): https://sciencetostartup.com/paper/query-languages-for-machine-learning-models - Develop a query language framework for expressing queries on neural network models using foundational logics.
- Frame of Reference: Addressing the Challenges of Common Ground Representation in Situational Dialogs (viability: 5): https://sciencetostartup.com/paper/frame-of-reference-addressing-the-challenges-of-common-ground-representation-in-situational-dialogs - Develop a dialog system API that enhances conversational grounding using explicit common ground representation.
- GeoRA: Geometry-Aware Low-Rank Adaptation for RLVR (viability: 6): https://sciencetostartup.com/paper/geora-geometry-aware-low-rank-adaptation-for-rlvr - GeoRA enhances reinforcement learning with efficient low-rank updates for improved performance and stability.
- Monte-Carlo Tree Search with Neural Network Guidance for Lane-Free Autonomous Driving (viability: 3): https://sciencetostartup.com/paper/monte-carlo-tree-search-with-neural-network-guidance-for-lane-free-autonomous-driving - Developing an advanced MCTS with neural network guidance for enhanced lane-free autonomous driving.
- Improving Implicit Hate Speech Detection via a Community-Driven Multi-Agent Framework (viability: 6): https://sciencetostartup.com/paper/improving-implicit-hate-speech-detection-via-a-community-driven-multi-agent-framework - Develop a multi-agent system for detecting implicit hate speech with enhanced classification accuracy and fairness.
- Blue Teaming Function-Calling Agents (viability: 2): https://sciencetostartup.com/paper/blue-teaming-function-calling-agents - Assess the safety of function-calling LLMs against attacks and defenses.
- STaR: Sensitive Trajectory Regulation for Unlearning in Large Reasoning Models (viability: 7): https://sciencetostartup.com/paper/star-sensitive-trajectory-regulation-for-unlearning-in-large-reasoning-models - "STaR provides a robust framework for privacy-preserving unlearning in large reasoning models, protecting sensitive data at all reasoning levels."
- M$^3$Searcher: Modular Multimodal Information Seeking Agency with Retrieval-Oriented Reasoning (viability: 7): https://sciencetostartup.com/paper/m-3-searcher-modular-multimodal-information-seeking-agency-with-retrieval-oriented-reasoning - Build a modular agent for efficient multimodal information retrieval and reasoning.
- $A^3$-Bench: Benchmarking Memory-Driven Scientific Reasoning via Anchor and Attractor Activation (viability: 6): https://sciencetostartup.com/paper/a-3-bench-benchmarking-memory-driven-scientific-reasoning-via-anchor-and-attractor-activation - A3-Bench offers a benchmark for evaluating memory-driven scientific reasoning in AI, leveraging anchors and attractors for improved problem-solving.
- Coordinated Pandemic Control with Large Language Model Agents as Policymaking Assistants (viability: 7): https://sciencetostartup.com/paper/coordinated-pandemic-control-with-large-language-model-agents-as-policymaking-assistants - AI-driven multi-agent system for coordinated pandemic policymaking.
- Efficient Paths and Dense Rewards: Probabilistic Flow Reasoning for Large Language Models (viability: 6): https://sciencetostartup.com/paper/efficient-paths-and-dense-rewards-probabilistic-flow-reasoning-for-large-language-models - CoT-Flow offers an efficient probabilistic reasoning framework for enhancing LLM reasoning with step-wise information gain.
- MAXS: Meta-Adaptive Exploration with LLM Agents (viability: 8): https://sciencetostartup.com/paper/maxs-meta-adaptive-exploration-with-llm-agents - Develop a reasoning framework using LLM agents for stable, efficient multi-tool integration with proven performance gains.
- RIFT: Repurposing Negative Samples via Reward-Informed Fine-Tuning (viability: 6): https://sciencetostartup.com/paper/rift-repurposing-negative-samples-via-reward-informed-fine-tuning - RIFT offers a data-efficient framework for improving AI model alignment using all self-generated samples.
- Hybrid guided variational autoencoder for visual place recognition (viability: 6): https://sciencetostartup.com/paper/hybrid-guided-variational-autoencoder-for-visual-place-recognition - Compact guided VAE for robust visual place recognition in mobile robots, optimized for neuromorphic hardware.
- Reward Learning through Ranking Mean Squared Error (viability: 6): https://sciencetostartup.com/paper/reward-learning-through-ranking-mean-squared-error - Develop a reward learning tool utilizing Ranked Return Regression to enhance RL performance with minimal human feedback.
- SpikeVAEDiff: Neural Spike-based Natural Visual Scene Reconstruction via VD-VAE and Versatile Diffusion (viability: 8): https://sciencetostartup.com/paper/spikevaediff-neural-spike-based-natural-visual-scene-reconstruction-via-vd-vae-and-versatile-diffusion - Develop a tool for reconstructing visual scenes from neural activity using the SpikeVAEDiff framework.
- A.X K1 Technical Report (viability: 6): https://sciencetostartup.com/paper/a-x-k1-technical-report - A.X K1 is a scalable MoE language model excelling in Korean-language benchmarks with user-controlled reasoning capabilities.
- Position on LLM-Assisted Peer Review: Addressing Reviewer Gap through Mentoring and Feedback (viability: 2): https://sciencetostartup.com/paper/position-on-llm-assisted-peer-review-addressing-reviewer-gap-through-mentoring-and-feedback - Develop LLM-based tools to mentor and enhance peer-review skills of academic reviewers.
- PrivacyReasoner: Can LLM Emulate a Human-like Privacy Mind? (viability: 5): https://sciencetostartup.com/paper/privacyreasoner-can-llm-emulate-a-human-like-privacy-mind - AI agent simulates user-specific privacy concerns based on context and personal history.
- SSVP: Synergistic Semantic-Visual Prompting for Industrial Zero-Shot Anomaly Detection (viability: 7): https://sciencetostartup.com/paper/ssvp-synergistic-semantic-visual-prompting-for-industrial-zero-shot-anomaly-detection - Zero-Shot Anomaly Detection using SSVP for precise, supervision-free industrial inspection.
- SkinFlow: Efficient Information Transmission for Open Dermatological Diagnosis via Dynamic Visual Encoding and Staged RL (viability: 8): https://sciencetostartup.com/paper/skinflow-efficient-information-transmission-for-open-dermatological-diagnosis-via-dynamic-visual-encoding-and-staged-rl - SkinFlow revolutionizes dermatological diagnosis with efficient visual encoding and reinforcement learning.
- A Marketplace for AI-Generated Adult Content and Deepfakes (viability: 2): https://sciencetostartup.com/paper/a-marketplace-for-ai-generated-adult-content-and-deepfakes - Platform for monetized AI-generated adult content raises ethical concerns without clear product path.
- AviationLMM: A Large Multimodal Foundation Model for Civil Aviation (viability: 2): https://sciencetostartup.com/paper/aviationlmm-a-large-multimodal-foundation-model-for-civil-aviation - Develop a foundational AI model to unify and analyze multimodal data in civil aviation for improved situational awareness and decision support.
- DScheLLM: Enabling Dynamic Scheduling through a Fine-Tuned Dual-System Large language Model (viability: 8): https://sciencetostartup.com/paper/dschellm-enabling-dynamic-scheduling-through-a-fine-tuned-dual-system-large-language-model - DScheLLM revolutionizes dynamic production scheduling with a fine-tuned large language model for adaptive and intelligent optimization.
- SubTokenTest: A Practical Benchmark for Real-World Sub-token Understanding (viability: 5): https://sciencetostartup.com/paper/subtokentest-a-practical-benchmark-for-real-world-sub-token-understanding - SubTokenTest offers a benchmark to improve LLMs' real-world sub-token understanding, crucial for applications like text-based map navigation.
- From Symbolic to Natural-Language Relations: Rethinking Knowledge Graph Construction in the Era of Large Language Models (viability: 3): https://sciencetostartup.com/paper/from-symbolic-to-natural-language-relations-rethinking-knowledge-graph-construction-in-the-era-of-large-language-models - Revolutionizing knowledge graph construction with natural language relations in the era of LLMs.
- Can LLMs interpret figurative language as humans do?: surface-level vs representational similarity (viability: 3): https://sciencetostartup.com/paper/can-llms-interpret-figurative-language-as-humans-do-surface-level-vs-representational-similarity - Explore differences in LLM and human figurative language interpretation for improved language model alignment.
- A Decompilation-Driven Framework for Malware Detection with Large Language Models (viability: 7): https://sciencetostartup.com/paper/a-decompilation-driven-framework-for-malware-detection-with-large-language-models - Automate malware detection by leveraging decompilation and LLMs for cybersecurity innovation.
- The Hierarchy of Agentic Capabilities: Evaluating Frontier Models on Realistic RL Environments (viability: 6): https://sciencetostartup.com/paper/the-hierarchy-of-agentic-capabilities-evaluating-frontier-models-on-realistic-rl-environments - Develop AI agents for workplace tasks by mastering tool use, planning, and adaptability in e-commerce RL environments.
- Generalizable Geometric Prior and Recurrent Spiking Feature Learning for Humanoid Robot Manipulation (viability: 9): https://sciencetostartup.com/paper/generalizable-geometric-prior-and-recurrent-spiking-feature-learning-for-humanoid-robot-manipulation - Platform for leveraging geometric prior and spiking features to enhance humanoid robot manipulation capabilities in new environments.
- ART: Action-based Reasoning Task Benchmarking for Medical AI Agents (viability: 7): https://sciencetostartup.com/paper/art-action-based-reasoning-task-benchmarking-for-medical-ai-agents - Develop an advanced benchmark for evaluating and improving medical AI agents' clinical reasoning using real-world EHR data.
- Imagine-then-Plan: Agent Learning from Adaptive Lookahead with World Models (viability: 5): https://sciencetostartup.com/paper/imagine-then-plan-agent-learning-from-adaptive-lookahead-with-world-models - Create smarter AI agents by using adaptive lookahead world models for complex task planning.
- Fairness risk and its privacy-enabled solution in AI-driven robotic applications (viability: 2): https://sciencetostartup.com/paper/fairness-risk-and-its-privacy-enabled-solution-in-ai-driven-robotic-applications - Develops a unified fairness and privacy framework for robotic decision-making to address ethical deployment in autonomous systems.
- PluriHarms: Benchmarking the Full Spectrum of Human Judgments on AI Harm (viability: 5): https://sciencetostartup.com/paper/pluriharms-benchmarking-the-full-spectrum-of-human-judgments-on-ai-harm - PluriHarms offers a benchmark for understanding and predicting diverse human judgments on AI harm, aiding in pluralistic AI safety designs.
- ConvoLearn: A Dataset of Constructivist Tutor-Student Dialogue (viability: 7): https://sciencetostartup.com/paper/convolearn-a-dataset-of-constructivist-tutor-student-dialogue - ConvoLearn offers a dataset and method for enhancing AI's pedagogical abilities in educational applications.
- GDPO: Group reward-Decoupled Normalization Policy Optimization for Multi-reward RL Optimization (viability: 4): https://sciencetostartup.com/paper/gdpo-group-reward-decoupled-normalization-policy-optimization-for-multi-reward-rl-optimization - Optimize multi-reward reinforcement learning with GDPO for stable and precise model training.
- RoboVIP: Multi-View Video Generation with Visual Identity Prompting Augments Robot Manipulation (viability: 8): https://sciencetostartup.com/paper/robovip-multi-view-video-generation-with-visual-identity-prompting-augments-robot-manipulation - Enhance robot manipulation datasets with multi-view video generation using visual identity prompts.
- Robust Reasoning as a Symmetry-Protected Topological Phase (viability: 3): https://sciencetostartup.com/paper/robust-reasoning-as-a-symmetry-protected-topological-phase - Develop a Holonomic Network model for robust logical reasoning in AI systems, mitigating language model hallucinations.
- Learning Latent Action World Models In The Wild (viability: 2): https://sciencetostartup.com/paper/learning-latent-action-world-models-in-the-wild - Develop latent action world models from diverse in-the-wild videos for improved real-world planning tasks.
- MineNPC-Task: Task Suite for Memory-Aware Minecraft Agents (viability: 6): https://sciencetostartup.com/paper/minenpc-task-task-suite-for-memory-aware-minecraft-agents - A benchmark suite for evaluating memory-aware AI agents in Minecraft with a focus on interaction quality and usability.
- Internal Representations as Indicators of Hallucinations in Agent Tool Selection (viability: 6): https://sciencetostartup.com/paper/internal-representations-as-indicators-of-hallucinations-in-agent-tool-selection - Develop a real-time hallucination detection tool for LLM-based agents to ensure reliable tool usage and security.
- Stock Market Price Prediction using Neural Prophet with Deep Neural Network (viability: 3): https://sciencetostartup.com/paper/stock-market-price-prediction-using-neural-prophet-with-deep-neural-network - Predict stock market prices using Neural Prophet with Deep Neural Networks for enhanced accuracy.
- Mechanisms of Prompt-Induced Hallucination in Vision-Language Models (viability: 3): https://sciencetostartup.com/paper/mechanisms-of-prompt-induced-hallucination-in-vision-language-models - This research analyzes how vision-language models get influenced by textual prompts, leading to hallucinations, and identifies attention heads that can mitigate this issue.
- SimuAgent: An LLM-Based Simulink Modeling Assistant Enhanced with Reinforcement Learning (viability: 8): https://sciencetostartup.com/paper/simuagent-an-llm-based-simulink-modeling-assistant-enhanced-with-reinforcement-learning - SimuAgent provides an AI-driven, efficient modeling assistant for Simulink users, enhancing design productivity and accuracy with privacy-preserving solutions.
- Observations and Remedies for Large Language Model Bias in Self-Consuming Performative Loop (viability: 5): https://sciencetostartup.com/paper/observations-and-remedies-for-large-language-model-bias-in-self-consuming-performative-loop - A system to reduce bias in large language models through a self-consuming performative loop for more trustworthy AI outputs.
- FaST: Efficient and Effective Long-Horizon Forecasting for Large-Scale Spatial-Temporal Graphs via Mixture-of-Experts (viability: 8): https://sciencetostartup.com/paper/fast-efficient-and-effective-long-horizon-forecasting-for-large-scale-spatial-temporal-graphs-via-mixture-of-experts - Develop FaST for scalable, efficient long-horizon forecasting in spatial-temporal graph networks.
- CoV: Chain-of-View Prompting for Spatial Reasoning (viability: 3): https://sciencetostartup.com/paper/cov-chain-of-view-prompting-for-spatial-reasoning - A framework enhancing spatial reasoning in 3D environments using chain-of-view prompting.
- RelayLLM: Efficient Reasoning via Collaborative Decoding (viability: 7): https://sciencetostartup.com/paper/relayllm-efficient-reasoning-via-collaborative-decoding - Empowering small models with efficient LLM augmentation for cost-effective reasoning.
- Vision-Language Introspection: Mitigating Overconfident Hallucinations in MLLMs via Interpretable Bi-Causal Steering (viability: 3): https://sciencetostartup.com/paper/vision-language-introspection-mitigating-overconfident-hallucinations-in-mllms-via-interpretable-bi-causal-steering - New framework for reducing hallucinations in multimodal language models using introspective techniques.
- Safe Continual Reinforcement Learning Methods for Nonstationary Environments. Towards a Survey of the State of the Art (viability: 2): https://sciencetostartup.com/paper/safe-continual-reinforcement-learning-methods-for-nonstationary-environments-towards-a-survey-of-the-state-of-the-art - A comprehensive survey of safe continual reinforcement learning methods for nonstationary environments.
- Atlas 2 -- Foundation models for clinical deployment (viability: 8): https://sciencetostartup.com/paper/atlas-2-foundation-models-for-clinical-deployment - Atlas 2 offers state-of-the-art pathology vision models designed for clinical deployment with enhanced performance and efficiency.
- Distilling the Thought, Watermarking the Answer: A Principle Semantic Guided Watermark for Large Reasoning Models (viability: 3): https://sciencetostartup.com/paper/distilling-the-thought-watermarking-the-answer-a-principle-semantic-guided-watermark-for-large-reasoning-models - Develop a semantic watermarking framework for reasoning LLMs that balances performance and integrity.
- VERSE: Visual Embedding Reduction and Space Exploration. Clustering-Guided Insights for Training Data Enhancement in Visually-Rich Document Understanding (viability: 8): https://sciencetostartup.com/paper/verse-visual-embedding-reduction-and-space-exploration-clustering-guided-insights-for-training-data-enhancement-in-visua - VERSE provides a strategic tool for enhancing vision-language models in document understanding by visualizing and improving visual embeddings.
- Evaluative Fingerprints: Stable and Systematic Differences in LLM Evaluator Behavior (viability: 6): https://sciencetostartup.com/paper/evaluative-fingerprints-stable-and-systematic-differences-in-llm-evaluator-behavior - Develop an AI system to identify and leverage stable evaluative fingerprints in LLM judges for improved AI assessment consistency.
- Agent-as-a-Judge (viability: 3): https://sciencetostartup.com/paper/agent-as-a-judge - A survey paper exploring 'Agent-as-a-Judge' in AI evaluations, proposing a developmental taxonomy.
- GlimpRouter: Efficient Collaborative Inference by Glimpsing One Token of Thoughts (viability: 8): https://sciencetostartup.com/paper/glimprouter-efficient-collaborative-inference-by-glimpsing-one-token-of-thoughts - "GlimpRouter optimizes AI inference by routing tasks efficiently between small and large models, saving time and resources without sacrificing accuracy."
- Controllable Memory Usage: Balancing Anchoring and Innovation in Long-Term Human-Agent Interaction (viability: 7): https://sciencetostartup.com/paper/controllable-memory-usage-balancing-anchoring-and-innovation-in-long-term-human-agent-interaction - SteeM enables dynamic control of memory usage in LLM-based agents for more personalized and innovative interactions.
- Token-Level LLM Collaboration via FusionRoute (viability: 7): https://sciencetostartup.com/paper/token-level-llm-collaboration-via-fusionroute - FusionRoute optimizes collaboration between domain-specialized language models at a token level for efficient, high-performance decoding.
- Arabic Prompts with English Tools: A Benchmark (viability: 4): https://sciencetostartup.com/paper/arabic-prompts-with-english-tools-a-benchmark - Benchmark Arabic language performance for language models using English tools.
- Code-Mix Sentiment Analysis on Hinglish Tweets (viability: 8): https://sciencetostartup.com/paper/code-mix-sentiment-analysis-on-hinglish-tweets - AI-powered sentiment analysis tailored for the Indian market, decoding code-mixed Hinglish on social media platforms.
- Driver-Intention Prediction with Deep Learning: Real-Time Brain-to-Vehicle Communication (viability: 5): https://sciencetostartup.com/paper/driver-intention-prediction-with-deep-learning-real-time-brain-to-vehicle-communication - Develop brain-to-vehicle communication for real-time driver intention prediction using EEG and deep learning.
- Driving on Registers (viability: 8): https://sciencetostartup.com/paper/driving-on-registers - DrivoR is a transformer-based autonomous driving system offering efficient, adaptive, end-to-end driving with high benchmark performance.
- Chain-of-Sanitized-Thoughts: Plugging PII Leakage in CoT of Large Reasoning Models (viability: 7): https://sciencetostartup.com/paper/chain-of-sanitized-thoughts-plugging-pii-leakage-in-cot-of-large-reasoning-models - Privacy-preserving reasoning systems to prevent PII leakage in large reasoning models.
- Compositional Steering of Large Language Models with Steering Tokens (viability: 3): https://sciencetostartup.com/paper/compositional-steering-of-large-language-models-with-steering-tokens - Enable multi-behavior steering of LLMs using novel compositional steering tokens for improved control.
- Reinforced Efficient Reasoning via Semantically Diverse Exploration (viability: 6): https://sciencetostartup.com/paper/reinforced-efficient-reasoning-via-semantically-diverse-exploration - Develop a tool for enhancing reasoning in large language models using semantically diverse explorations for improved efficiency and diversity.
- Publishing FAIR and Machine-actionable Reviews in Materials Science: The Case for Symbolic Knowledge in Neuro-symbolic Artificial Intelligence (viability: 2): https://sciencetostartup.com/paper/publishing-fair-and-machine-actionable-reviews-in-materials-science-the-case-for-symbolic-knowledge-in-neuro-symbolic-ar - Make material science reviews machine-actionable with structured, queryable knowledge graphs.
- Challenges and Research Directions for Large Language Model Inference Hardware (viability: 3): https://sciencetostartup.com/paper/challenges-and-research-directions-for-large-language-model-inference-hardware - Exploring innovative hardware architectures to optimize large language model inference.
- HMVI: Unifying Heterogeneous Attributes with Natural Neighbors for Missing Value Inference (viability: 5): https://sciencetostartup.com/paper/hmvi-unifying-heterogeneous-attributes-with-natural-neighbors-for-missing-value-inference - Revolutionize data imputation by unifying numerical and categorical attributes for better machine learning outcomes.
- An Empirical Investigation of Robustness in Large Language Models under Tabular Distortions (viability: 5): https://sciencetostartup.com/paper/an-empirical-investigation-of-robustness-in-large-language-models-under-tabular-distortions - Develop an LLM-based tool for robust table question answering under distortions.
- On the Hidden Objective Biases of Group-based Reinforcement Learning (viability: 2): https://sciencetostartup.com/paper/on-the-hidden-objective-biases-of-group-based-reinforcement-learning - Introducing a theoretical analysis to address biases in group-based reinforcement learning methods.
- AlgBench: To What Extent Do Large Reasoning Models Understand Algorithms? (viability: 5): https://sciencetostartup.com/paper/algbench-to-what-extent-do-large-reasoning-models-understand-algorithms - Benchmark tool for assessing Large Reasoning Models' algorithmic understanding.
- When to Act: Calibrated Confidence for Reliable Human Intention Prediction in Assistive Robotics (viability: 3): https://sciencetostartup.com/paper/when-to-act-calibrated-confidence-for-reliable-human-intention-prediction-in-assistive-robotics - A framework for using calibrated confidence in assistive robotics to safely predict and act on human intentions.
- On the Definition and Detection of Cherry-Picking in Counterfactual Explanations (viability: 3): https://sciencetostartup.com/paper/on-the-definition-and-detection-of-cherry-picking-in-counterfactual-explanations - Develop safeguards to prevent cherry-picking in counterfactual explanations using formal definitions and audit detection strategies.
- ConMax: Confidence-Maximizing Compression for Efficient Chain-of-Thought Reasoning (viability: 7): https://sciencetostartup.com/paper/conmax-confidence-maximizing-compression-for-efficient-chain-of-thought-reasoning - ConMax optimizes large reasoning models by compressing reasoning paths, reducing computational costs while maintaining accuracy.
- Text as a Universal Interface for Transferable Personalization (viability: 7): https://sciencetostartup.com/paper/text-as-a-universal-interface-for-transferable-personalization - A system that uses natural language to personalize user experiences across applications by generating interpretable preference summaries.
- Precision over Diversity: High-Precision Reward Generalizes to Robust Instruction Following (viability: 6): https://sciencetostartup.com/paper/precision-over-diversity-high-precision-reward-generalizes-to-robust-instruction-following - A data-centric strategy that prioritizes precision in reward systems to enhance AI instruction-following performance and efficiency.
- Prototypicality Bias Reveals Blindspots in Multimodal Evaluation Metrics (viability: 6): https://sciencetostartup.com/paper/prototypicality-bias-reveals-blindspots-in-multimodal-evaluation-metrics - A new metric, ProtoScore, addresses prototypicality bias in multimodal AI evaluation, offering faster and more robust benchmarking.
- T-Retriever: Tree-based Hierarchical Retrieval Augmented Generation for Textual Graphs (viability: 5): https://sciencetostartup.com/paper/t-retriever-tree-based-hierarchical-retrieval-augmented-generation-for-textual-graphs - Develop T-Retriever, a framework enhancing graph-based information retrieval with better semantic and structural integration.
- What Students Ask, How a Generative AI Assistant Responds: Exploring Higher Education Students' Dialogues on Learning Analytics Feedback (viability: 3): https://sciencetostartup.com/paper/what-students-ask-how-a-generative-ai-assistant-responds-exploring-higher-education-students-dialogues-on-learning-analy - Develop a GenAI tutor for personalized learning feedback assistance in education.
- Breaking Robustness Barriers in Cognitive Diagnosis: A One-Shot Neural Architecture Search Perspective (viability: 4): https://sciencetostartup.com/paper/breaking-robustness-barriers-in-cognitive-diagnosis-a-one-shot-neural-architecture-search-perspective - Leverage one-shot neural architecture search to improve cognitive diagnosis models in intelligent tutoring systems.
- Analyzing Message-Code Inconsistency in AI Coding Agent-Authored Pull Requests (viability: 3): https://sciencetostartup.com/paper/analyzing-message-code-inconsistency-in-ai-coding-agent-authored-pull-requests - Enhance AI-generated pull request trust with PR-MCI verification mechanisms.
- Precomputing Multi-Agent Path Replanning using Temporal Flexibility: A Case Study on the Dutch Railway Network (viability: 7): https://sciencetostartup.com/paper/precomputing-multi-agent-path-replanning-using-temporal-flexibility-a-case-study-on-the-dutch-railway-network - Optimize train scheduling with precomputed multi-agent replanning for reduced delays in railway networks.
- Higher-Order Knowledge Representations for Agentic Scientific Reasoning (viability: 7): https://sciencetostartup.com/paper/higher-order-knowledge-representations-for-agentic-scientific-reasoning - Build agentic reasoning systems to accelerate scientific discovery using hypergraph-based knowledge representations.
- Key-Value Pair-Free Continual Learner via Task-Specific Prompt-Prototype (viability: 5): https://sciencetostartup.com/paper/key-value-pair-free-continual-learner-via-task-specific-prompt-prototype - Develop task-specific Prompt-Prototype for more scalable and efficient continual learning without key-value dependencies.
- Orchestrating Intelligence: Confidence-Aware Routing for Efficient Multi-Agent Collaboration across Multi-Scale Models (viability: 8): https://sciencetostartup.com/paper/orchestrating-intelligence-confidence-aware-routing-for-efficient-multi-agent-collaboration-across-multi-scale-models - Revolutionizing multi-agent systems with adaptive model selection for efficient and cost-effective AI collaboration.
- Rethinking GNNs and Missing Features: Challenges, Evaluation and a Robust Solution (viability: 7): https://sciencetostartup.com/paper/rethinking-gnns-and-missing-features-challenges-evaluation-and-a-robust-solution - GNNmim offers a robust solution for handling missing node features in Graph Neural Networks.
- Token Maturation: Autoregressive Language Generation via Continuous Token Dynamics (viability: 2): https://sciencetostartup.com/paper/token-maturation-autoregressive-language-generation-via-continuous-token-dynamics - Develop continuous token dynamics for more stable autoregressive language generation without token-level sampling.
- Defense Against Indirect Prompt Injection via Tool Result Parsing (viability: 7): https://sciencetostartup.com/paper/defense-against-indirect-prompt-injection-via-tool-result-parsing - A robust defense solution against indirect prompt injection attacks for LLM agents in autonomous systems, enhancing security with efficient tool result parsing.
- APEX: Academic Poster Editing Agentic Expert (viability: 7): https://sciencetostartup.com/paper/apex-academic-poster-editing-agentic-expert - Interactive AI tool for designing and editing academic posters with fine-grained control.
- Belief in Authority: Impact of Authority in Multi-Agent Evaluation Framework (viability: 5): https://sciencetostartup.com/paper/belief-in-authority-impact-of-authority-in-multi-agent-evaluation-framework - Develop AI systems that optimize multi-agent interactions by understanding authority bias using role-based frameworks.
- NC2C: Automated Convexification of Generic Non-Convex Optimization Problems (viability: 7): https://sciencetostartup.com/paper/nc2c-automated-convexification-of-generic-non-convex-optimization-problems - Transform non-convex optimization problems to convex forms using AI to enable efficient resolution.
- AgentOCR: Reimagining Agent History via Optical Self-Compression (viability: 6): https://sciencetostartup.com/paper/agentocr-reimagining-agent-history-via-optical-self-compression - AgentOCR optimizes agent memory and performance by converting text histories into visually compact representations, offering a scalable solution for multi-turn interactions.
- SRU-Pix2Pix: A Fusion-Driven Generator Network for Medical Image Translation with Few-Shot Learning (viability: 7): https://sciencetostartup.com/paper/sru-pix2pix-a-fusion-driven-generator-network-for-medical-image-translation-with-few-shot-learning - Enhanced Pix2Pix framework optimizes MRI imaging with few-shot learning for faster, cost-effective medical diagnostics.
- CounterVid: Counterfactual Video Generation for Mitigating Action and Temporal Hallucinations in Video-Language Models (viability: 8): https://sciencetostartup.com/paper/countervid-counterfactual-video-generation-for-mitigating-action-and-temporal-hallucinations-in-video-language-models - CounterVid enhances video-language models by generating counterfactual videos to reduce action and temporal hallucinations.
- GeM-VG: Towards Generalized Multi-image Visual Grounding with Multimodal Large Language Models (viability: 8): https://sciencetostartup.com/paper/gem-vg-towards-generalized-multi-image-visual-grounding-with-multimodal-large-language-models - GeM-VG offers superior multi-image visual grounding capabilities, leveraging a novel dataset and hybrid reinforcement finetuning strategy for robust cross-image reasoning.
- SciIF: Benchmarking Scientific Instruction Following Towards Rigorous Scientific Intelligence (viability: 6): https://sciencetostartup.com/paper/sciif-benchmarking-scientific-instruction-following-towards-rigorous-scientific-intelligence - SciIF benchmark evaluates LLMs' adherence to scientific rigor, enabling reliable applications in scientific inquiry.
- AT$^2$PO: Agentic Turn-based Policy Optimization via Tree Search (viability: 6): https://sciencetostartup.com/paper/at-2-po-agentic-turn-based-policy-optimization-via-tree-search - Innovative reinforcement learning framework for enhancing multi-turn AI agents using tree search and policy optimization.
- Differential syntactic and semantic encoding in LLMs (viability: 2): https://sciencetostartup.com/paper/differential-syntactic-and-semantic-encoding-in-llms - Study reveals differential encoding of syntax and semantics in LLM inner layers.
- Orion-RAG: Path-Aligned Hybrid Retrieval for Graphless Data (viability: 8): https://sciencetostartup.com/paper/orion-rag-path-aligned-hybrid-retrieval-for-graphless-data - Orion-RAG optimizes data retrieval by creating lightweight paths linking fragmented documents to transform them into semi-structured data, improving retrieval accuracy by 25.2%.
- Smart IoT-Based Wearable Device for Detection and Monitoring of Common Cow Diseases Using a Novel Machine Learning Technique (viability: 6): https://sciencetostartup.com/paper/smart-iot-based-wearable-device-for-detection-and-monitoring-of-common-cow-diseases-using-a-novel-machine-learning-techn - Develop a smart IoT-based wearable device for automated detection and monitoring of cow diseases using a novel machine learning algorithm.
- KnowMe-Bench: Benchmarking Person Understanding for Lifelong Digital Companions (viability: 5): https://sciencetostartup.com/paper/knowme-bench-benchmarking-person-understanding-for-lifelong-digital-companions - A benchmark for enhancing AI's person understanding through autobiographical narratives.
- Fast Mining and Dynamic Time-to-Event Prediction over Multi-sensor Data Streams (viability: 7): https://sciencetostartup.com/paper/fast-mining-and-dynamic-time-to-event-prediction-over-multi-sensor-data-streams - TimeCast predicts machine failure times by analyzing evolving multi-sensor data streams in real time.
- RiskAtlas: Exposing Domain-Specific Risks in LLMs through Knowledge-Graph-Guided Harmful Prompt Generation (viability: 7): https://sciencetostartup.com/paper/riskatlas-exposing-domain-specific-risks-in-llms-through-knowledge-graph-guided-harmful-prompt-generation - RiskAtlas provides a framework to expose domain-specific risks in LLMs using knowledge-graph-guided prompt generation, targeting safety evaluation in high-stakes sectors like finance and healthcare.
- Miner:Mining Intrinsic Mastery for Data-Efficient RL in Large Reasoning Models (viability: 8): https://sciencetostartup.com/paper/miner-mining-intrinsic-mastery-for-data-efficient-rl-in-large-reasoning-models - Our RL solution, Miner, leverages intrinsic uncertainty for data-efficient training in large reasoning models, significantly outperforming current methods.
- Excess Description Length of Learning Generalizable Predictors (viability: 3): https://sciencetostartup.com/paper/excess-description-length-of-learning-generalizable-predictors - Formal framework for evaluating latent capabilities development in fine-tuning language models using Excess Description Length.
- DSC2025 -- ViHallu Challenge: Detecting Hallucination in Vietnamese LLMs (viability: 6): https://sciencetostartup.com/paper/dsc2025-vihallu-challenge-detecting-hallucination-in-vietnamese-llms - Build a platform for detecting and reducing hallucinations in Vietnamese language models using the ViHallu dataset and structured prompting techniques.
- Bridging Temporal and Textual Modalities: A Multimodal Framework for Automated Cloud Failure Root Cause Analysis (viability: 6): https://sciencetostartup.com/paper/bridging-temporal-and-textual-modalities-a-multimodal-framework-for-automated-cloud-failure-root-cause-analysis - Develop a multimodal diagnostic tool for cloud incident management that aligns time-series data with language models for high-accuracy failure analysis.
- TourPlanner: A Competitive Consensus Framework with Constraint-Gated Reinforcement Learning for Travel Planning (viability: 7): https://sciencetostartup.com/paper/tourplanner-a-competitive-consensus-framework-with-constraint-gated-reinforcement-learning-for-travel-planning - TourPlanner is an AI-driven travel planning tool using constraint-gated reinforcement learning for personalized itinerary optimization.
- Tape: A Cellular Automata Benchmark for Evaluating Rule-Shift Generalization in Reinforcement Learning (viability: 4): https://sciencetostartup.com/paper/tape-a-cellular-automata-benchmark-for-evaluating-rule-shift-generalization-in-reinforcement-learning - Develop a benchmark suite to test AI models on generalization under out-of-distribution scenarios using cellular automata.
- ResMAS: Resilience Optimization in LLM-based Multi-agent Systems (viability: 7): https://sciencetostartup.com/paper/resmas-resilience-optimization-in-llm-based-multi-agent-systems - Optimize the resilience of LLM-based multi-agent systems against perturbations with ResMAS.
- ToolGate: Contract-Grounded and Verified Tool Execution for LLMs (viability: 6): https://sciencetostartup.com/paper/toolgate-contract-grounded-and-verified-tool-execution-for-llms - ToolGate enforces logical safety and verifiable state evolution in LLM tool execution for reliable AI systems.
- LLM-Guided Quantified SMT Solving over Uninterpreted Functions (viability: 8.7): https://sciencetostartup.com/paper/llm-guided-quantified-smt-solving-over-uninterpreted-functions - AquaForte leverages large language models to optimize SMT solving over uninterpreted functions, significantly outperforming existing state-of-the-art solvers.
- Estimating Causal Effects in Gaussian Linear SCMs with Finite Data (viability: 5): https://sciencetostartup.com/paper/estimating-causal-effects-in-gaussian-linear-scms-with-finite-data - Develop an EM-based tool for causal effect estimation in Gaussian Linear SCMs from finite data.
- Optimizing Path Planning using Deep Reinforcement Learning for UGVs in Precision Agriculture (viability: 7): https://sciencetostartup.com/paper/optimizing-path-planning-using-deep-reinforcement-learning-for-ugvs-in-precision-agriculture - Develop DRL-based path planning solutions for UGVs to enhance efficiency in precision agriculture.
- Know Thy Enemy: Securing LLMs Against Prompt Injection via Diverse Data Synthesis and Instruction-Level Chain-of-Thought Learning (viability: 7): https://sciencetostartup.com/paper/know-thy-enemy-securing-llms-against-prompt-injection-via-diverse-data-synthesis-and-instruction-level-chain-of-thought- - A security enhancement for LLMs that defends against prompt injection attacks using diverse data synthesis and instruction-level chain-of-thought learning.
- Succeeding at Scale: Automated Multi-Retriever Fusion and Query-Side Adaptation for Multi-Tenant Search (viability: 8): https://sciencetostartup.com/paper/succeeding-at-scale-automated-multi-retriever-fusion-and-query-side-adaptation-for-multi-tenant-search - Build a scalable multi-tenant search platform using DevRev Search's fusion-based retrieval and index-preserving query adaptation.
- SpeechMedAssist: Efficiently and Effectively Adapting Speech Language Models for Medical Consultation (viability: 8): https://sciencetostartup.com/paper/speechmedassist-efficiently-and-effectively-adapting-speech-language-models-for-medical-consultation - Develop an adaptable Speech Language Model for efficient and effective speech-based medical consultations.
- From National Curricula to Cultural Awareness: Constructing Open-Ended Culture-Specific Question Answering Dataset (viability: 3): https://sciencetostartup.com/paper/from-national-curricula-to-cultural-awareness-constructing-open-ended-culture-specific-question-answering-dataset - Automate cultural alignment in language models using national curricula-derived datasets.
- Beyond the "Truth": Investigating Election Rumors on Truth Social During the 2024 Election (viability: 7): https://sciencetostartup.com/paper/beyond-the-truth-investigating-election-rumors-on-truth-social-during-the-2024-election - Develop a Rumor Detection Agent to analyze misinformation spread on niche platforms using advanced LLM techniques.
- Evaluating Human and Machine Confidence in Phishing Email Detection: A Comparative Study (viability: 3): https://sciencetostartup.com/paper/evaluating-human-and-machine-confidence-in-phishing-email-detection-a-comparative-study - A study on how humans and machine learning models compare in detecting phishing emails.
- On the Limitations of Rank-One Model Editing in Answering Multi-hop Questions (viability: 6): https://sciencetostartup.com/paper/on-the-limitations-of-rank-one-model-editing-in-answering-multi-hop-questions - Enhancing multi-hop reasoning in AI models through a novel redundant editing strategy for more accurate knowledge updates.
- FedKDX: Federated Learning with Negative Knowledge Distillation for Enhanced Healthcare AI Systems (viability: 7): https://sciencetostartup.com/paper/fedkdx-federated-learning-with-negative-knowledge-distillation-for-enhanced-healthcare-ai-systems - Introducing FedKDX: a federated learning framework enhancing healthcare AI by integrating negative knowledge distillation with privacy considerations.
- Autonomous Agents on Blockchains: Standards, Execution Models, and Trust Boundaries (viability: 3): https://sciencetostartup.com/paper/autonomous-agents-on-blockchains-standards-execution-models-and-trust-boundaries - Develop secure interfaces for autonomous agents executing on blockchains to ensure safe and robust transactions.
- Sci-Reasoning: A Dataset Decoding AI Innovation Patterns (viability: 5): https://sciencetostartup.com/paper/sci-reasoning-a-dataset-decoding-ai-innovation-patterns - A dataset capturing innovation patterns in AI research for enhancing research agents.
- Scaling Behavior Cloning Improves Causal Reasoning: An Open Model for Real-Time Video Game Playing (viability: 7): https://sciencetostartup.com/paper/scaling-behavior-cloning-improves-causal-reasoning-an-open-model-for-real-time-video-game-playing - A scalable open model for real-time video game AI that rivals human performance.
- Spatial-Temporal Feedback Diffusion Guidance for Controlled Traffic Imputation (viability: 8): https://sciencetostartup.com/paper/spatial-temporal-feedback-diffusion-guidance-for-controlled-traffic-imputation - FENCE revolutionizes traffic data imputation with adaptive spatial-temporal feedback diffusion for smarter transportation systems.
- Enhancing Multimodal Retrieval via Complementary Information Extraction and Alignment (viability: 7): https://sciencetostartup.com/paper/enhancing-multimodal-retrieval-via-complementary-information-extraction-and-alignment - CIEA enhances multimodal retrieval by effectively capturing complementary information in documents.
- BackdoorAgent: A Unified Framework for Backdoor Attacks on LLM-based Agents (viability: 7): https://sciencetostartup.com/paper/backdooragent-a-unified-framework-for-backdoor-attacks-on-llm-based-agents - A framework for identifying and analyzing backdoor threats in LLM-based agents, crucial for cybersecurity in AI workflows.
- Identifying Good and Bad Neurons for Task-Level Controllable LLMs (viability: 6): https://sciencetostartup.com/paper/identifying-good-and-bad-neurons-for-task-level-controllable-llms - NeuronLLM enhances LLMs by identifying and controlling task-critical neurons for improved NLP performance.
- TCAndon-Router: Adaptive Reasoning Router for Multi-Agent Collaboration (viability: 8): https://sciencetostartup.com/paper/tcandon-router-adaptive-reasoning-router-for-multi-agent-collaboration - Develop an adaptive reasoning router for seamless and robust multi-agent collaboration in enterprise applications.
- AdaptEval: A Benchmark for Evaluating Large Language Models on Code Snippet Adaptation (viability: 4): https://sciencetostartup.com/paper/adapteval-a-benchmark-for-evaluating-large-language-models-on-code-snippet-adaptation - AdaptEval provides a benchmark tool to evaluate the capability of large language models in adapting code snippets for practical software engineering tasks.
- Paradoxical noise preference in RNNs (viability: 2): https://sciencetostartup.com/paper/paradoxical-noise-preference-in-rnns - Understanding noise dynamics in RNNs to design more robust models.
- BanglaLorica: Design and Evaluation of a Robust Watermarking Algorithm for Large Language Models in Bangla Text Generation (viability: 6): https://sciencetostartup.com/paper/banglalorica-design-and-evaluation-of-a-robust-watermarking-algorithm-for-large-language-models-in-bangla-text-generatio - Develop a robust watermarking solution for Bangla text generation in large language models.
- Self-MedRAG: a Self-Reflective Hybrid Retrieval-Augmented Generation Framework for Reliable Medical Question Answering (viability: 8): https://sciencetostartup.com/paper/self-medrag-a-self-reflective-hybrid-retrieval-augmented-generation-framework-for-reliable-medical-question-answering - Self-MedRAG enhances medical question answering reliability by integrating hybrid retrieval and iterative self-reflection.
- A General Neural Backbone for Mixed-Integer Linear Optimization via Dual Attention (viability: 7): https://sciencetostartup.com/paper/a-general-neural-backbone-for-mixed-integer-linear-optimization-via-dual-attention - Develop an attention-driven neural architecture to enhance mixed-integer linear optimization with improved solver efficiency.
- WESR: Scaling and Evaluating Word-level Event-Speech Recognition (viability: 2): https://sciencetostartup.com/paper/wesr-scaling-and-evaluating-word-level-event-speech-recognition - A novel evaluation framework and baseline for detecting non-verbal events in speech.
- A Semi-supervised Molecular Learning Framework for Activity Cliff Estimation (viability: 7): https://sciencetostartup.com/paper/a-semi-supervised-molecular-learning-framework-for-activity-cliff-estimation - SemiMol offers a semi-supervised framework that improves molecular property predictions in low-data scenarios, vital for drug discovery.
- CircuitLM: A Multi-Agent LLM-Aided Design Framework for Generating Circuit Schematics from Natural Language Prompts (viability: 8): https://sciencetostartup.com/paper/circuitlm-a-multi-agent-llm-aided-design-framework-for-generating-circuit-schematics-from-natural-language-prompts - CircuitLM enables non-experts to generate accurate circuit schematics from natural language prompts, bridging the gap between text input and hardware design.
- Specific Emitter Identification via Active Learning (viability: 7): https://sciencetostartup.com/paper/specific-emitter-identification-via-active-learning - A cost-effective AI-driven solution for security in wireless communications through specific emitter identification with active learning.
- GUITester: Enabling GUI Agents for Exploratory Defect Discovery (viability: 8): https://sciencetostartup.com/paper/guitester-enabling-gui-agents-for-exploratory-defect-discovery - A multi-agent framework for autonomous exploratory GUI testing that significantly outperforms existing methods.
- Vision-Language Agents for Interactive Forest Change Analysis (viability: 8): https://sciencetostartup.com/paper/vision-language-agents-for-interactive-forest-change-analysis - Interactive AI tool for forest change analysis using vision-language models and LLMs, with public data and code.
- Scalable Floating-Point Satisfiability via Staged Optimization (viability: 7): https://sciencetostartup.com/paper/scalable-floating-point-satisfiability-via-staged-optimization - StageSAT offers a faster, more accurate floating-point satisfiability solver with scalable optimization stages for numerical and SMT tasks.
- A Closed-Loop Multi-Agent System Driven by LLMs for Meal-Level Personalized Nutrition Management (viability: 8): https://sciencetostartup.com/paper/a-closed-loop-multi-agent-system-driven-by-llms-for-meal-level-personalized-nutrition-management - A mobile app using LLM-driven agents to provide personalized nutrition by analyzing meal photos and adapting meal plans.
- Decision-Aware Trust Signal Alignment for SOC Alert Triage (viability: 6): https://sciencetostartup.com/paper/decision-aware-trust-signal-alignment-for-soc-alert-triage - Develop a decision-aware SOC alert triage system to enhance security analysts' decision-making efficiency.
- Hybrid Federated Learning for Noise-Robust Training (viability: 2): https://sciencetostartup.com/paper/hybrid-federated-learning-for-noise-robust-training - Hybrid Federated Learning framework boosts noise-robust training by combining FL and FD updates.
- Computational Compliance for AI Regulation: Blueprint for a New Research Domain (viability: 5): https://sciencetostartup.com/paper/computational-compliance-for-ai-regulation-blueprint-for-a-new-research-domain - Develop algorithms for automatic compliance of AI systems with emerging regulations.
- Concept Tokens: Learning Behavioral Embeddings Through Concept Definitions (viability: 2): https://sciencetostartup.com/paper/concept-tokens-learning-behavioral-embeddings-through-concept-definitions - Develop a method to add concept-specific behavior to LLMs via Concept Tokens for improved control in tasks like question answering and language teaching.
- Beyond Static Summarization: Proactive Memory Extraction for LLM Agents (viability: 3): https://sciencetostartup.com/paper/beyond-static-summarization-proactive-memory-extraction-for-llm-agents - Develop an iterative cognitive process for proactive memory extraction to improve QA accuracy in LLM agents.
- Categorical Belief Propagation: Sheaf-Theoretic Inference via Descent and Holonomy (viability: 8): https://sciencetostartup.com/paper/categorical-belief-propagation-sheaf-theoretic-inference-via-descent-and-holonomy - Develop an advanced belief propagation tool utilizing sheaf-theoretic inference for faster and exact complex graph analysis.

## Legal Documents
- Acceptable Use Policy (last updated: 2026-02-25): https://sciencetostartup.com/legal/acceptable-use-policy
- Terms of Service (last updated: 2026-02-25): https://sciencetostartup.com/legal/terms
- Enterprise Service Level Agreement (last updated: 2026-02-25): https://sciencetostartup.com/legal/sla

## API and Agent Interfaces
Full documentation: https://sciencetostartup.com/docs/api
Capability index: https://sciencetostartup.com/api/capabilities.json
OpenAPI spec: https://sciencetostartup.com/api/openapi.json
Remote MCP endpoint: https://sciencetostartup.com/api/mcp

### REST Discovery
- Search papers: https://sciencetostartup.com/api/v1/free/papers
- Signal Fusion rankings: https://sciencetostartup.com/api/v1/signal-fusion
- Research search: POST https://sciencetostartup.com/api/v1/research/search
- Deep-search runs: POST https://sciencetostartup.com/api/v1/research/deep-search
- Canonical research report: GET https://sciencetostartup.com/api/v1/research/report/{runId}
- Research artifact: GET https://sciencetostartup.com/api/v1/research/report/{runId}/artifacts/{kind}
- Benchmark API: https://sciencetostartup.com/api/v1/resources/benchmark
- Dataset API: https://sciencetostartup.com/api/v1/resources/dataset

### MCP Tools
- Identity: whoami
- Discovery: search_papers, get_paper, get_signal_fusion_rankings, search_signal_canvas
- Workspace: workspace_list, workspace_get, workspace_create_from_seed, workspace_run_action, workspace_run_log
- Execution: build_room_create, build_room_get, build_room_run, get_launch_pack, get_diligence_memo_export

## Canonical Flows
- Paper to workspace: https://sciencetostartup.com/developers/use-cases/paper-to-workspace-automation
- Signal Canvas API: https://sciencetostartup.com/developers/use-cases/signal-canvas-api
- Benchmark to launch pack: https://sciencetostartup.com/developers/examples/benchmark-to-launch-pack
- Topic to Signal Canvas: https://sciencetostartup.com/developers/examples/topic-to-signal-canvas

## Discovery
- Proof hub: https://sciencetostartup.com/proof
- Developers hub: https://sciencetostartup.com/developers
- OpenAI guide: https://sciencetostartup.com/developers/integrations/openai
- Cursor guide: https://sciencetostartup.com/developers/integrations/cursor
- Claude guide: https://sciencetostartup.com/developers/integrations/claude
- Sitemap index: https://sciencetostartup.com/sitemap.xml
- Main sitemap: https://sciencetostartup.com/sitemap-main.xml
- Article sitemap: https://sciencetostartup.com/sitemap-articles.xml
- News sitemap: https://sciencetostartup.com/sitemap-news.xml
- Papers feed: https://sciencetostartup.com/papers/feed.xml
- Articles feed: https://sciencetostartup.com/articles/feed.xml
- Media feed: https://sciencetostartup.com/media/feed.xml

## Contact
Website: https://sciencetostartup.com
Founder: Musa Emin Ozdem
Email: musaeminozdem@gmail.com
LinkedIn: https://linkedin.com/in/musaeminozdem
X: https://x.com/musaeminozdem