Build Loop

BioMiner is a multi-modal system that automates the extraction of protein-ligand bioactivity data from literature, significantly accelerating drug discovery workflows.

VARestorer distills pre-trained visual autoregressive models into a one-step super-resolution system, achieving state-of-the-art results with significantly faster inference.

An adaptive instruction composition framework for LLM red-teaming that jointly optimizes effectiveness and diversity of attacks using reinforcement learning.

StructMem is a novel memory framework for LLMs that enhances long-term conversational agents by preserving event relationships, reducing costs and improving performance.

An open-source AI memory system applying spatial metaphors to LLMs, offering a novel verbatim storage philosophy and low-cost offline operation.

Optimize LLM-based workflows by reducing tool-related token overhead and improving efficiency with dynamic tool gating.

Agent Evolving Learning (AEL) enables LLM agents to learn from experience in open-ended environments by dynamically selecting retrieval policies and using LLM-driven reflection to interpret past outcomes.

HypEHR: A compact hyperbolic model for efficient Electronic Health Record question answering, outperforming LLMs with fewer parameters.

A hybrid multi-robot control framework using waypoint-based bi-level planning to jointly optimize task and motion planning in cluttered environments.

A co-evolutionary framework that enables LLM agents to learn and reuse structured skills for long-horizon tasks, significantly improving performance in game environments.

SQLyzr is a comprehensive benchmark and evaluation platform for Text-to-SQL models, offering realistic evaluation and fine-grained analysis.

A controllable image style transfer method using visual autoregressive modeling and reinforcement learning to generate high-quality stylized images.

A novel self-supervised learning strategy that enhances robustness to image corruptions in aerial imagery, offering interpretable uncertainty signals.

GS-Quant is a novel framework for Knowledge Graph Completion that generates semantically coherent and structurally stratified discrete codes for LLMs, outperforming existing methods.

This work introduces a new dataset to analyze and reveal hidden cultural and regional biases in LLMs, showing a surprising preference for Japanese culture.

An efficient image enhancement model for mobile devices that maintains high quality and fast processing speeds through quantization-aware training and a hierarchical network architecture.

A semantics-aware framework for text-to-SQL data synthesis that uses specialized modules for analysis, synthesis, and verification to ensure semantic validity.

A framework to mitigate spurious signals in test-time reinforcement learning for math reasoning in LLMs, improving accuracy and stability.

A novel function hijacking attack that manipulates tool selection in agentic AI models, demonstrating significant vulnerabilities and the need for enhanced security.

A diffusion-based framework that enhances controllable human video generation by systematically investigating and optimizing the use of synthetic data.

Fine-tuning LLMs with generated reasons to improve next occupation prediction accuracy.

A synthetic dataset and simulation pipeline for multi-task aerial imagery analysis, enabling joint research in depth estimation, domain adaptation, and super-resolution.

A framework for dual attribution and verification to enhance factual reliability and interpretability of LLM outputs.

EngramaBench is a new benchmark for evaluating long-term conversational memory in LLMs, featuring a graph-structured memory system called Engrama that shows promise in cross-space reasoning.

A hierarchical multi-agent framework for long-form video understanding that improves temporal reasoning through question-aware collaboration.

A large-scale, open dataset and foundation models for medical robotics, enabling autonomous procedures and advancing robot learning with multi-embodiment simulation.

MISTY is a high-throughput generative motion planner that achieves state-of-the-art closed-loop performance with pure single-step inference for autonomous driving.

A post-capture method using an image-specific MLP and encoder to recover the 'unhallucinated' version of camera images altered by generative AI.

A benchmark for evaluating AI agents in financial investment research, highlighting current limitations and guiding the development of specialized AI for finance.

A multimodal LLM framework that assesses building conditions from street-view imagery, outperforming human raters and offering distilled models for efficient, large-scale deployment.

A reinforcement learning framework co-evolves a proposer and visual critic to achieve precise pixel-level localization for natural language instructions in GUIs.

A parameter-efficient training framework for multi-agent systems that jointly optimizes latent communication with reasoning, improving performance on complex tasks.

ReCAPA is a hierarchical predictive correction framework for Vision-Language-Action systems that mitigates cascading failures by adjusting actions, subgoals, and trajectories, outperforming LLM baselines on embodied agent benchmarks.

SparKV is an adaptive KV cache loading framework for on-device LLM inference that reduces Time-to-First-Token by up to 5.1x and lowers energy consumption, making LLMs more practical for edge devices.

This research proposes a method to reduce token usage and improve LLM accuracy by storing and recalling reusable reasoning skills, offering practical economic benefits for deployment.

A reusable, cross-domain evaluation pipeline for AI meeting summaries is presented, offering a robust framework for benchmarking and improving generative AI applications with a focus on accuracy and retention.

DryRUN enables LLM-driven code generation without human-provided test cases by autonomously generating inputs and self-correcting.

TravelFraudBench is a configurable GNN evaluation framework for travel fraud ring detection, outperforming baselines and achieving 100% ring recovery with an open-source Python package.

An efficient logic gate network for video copy detection that achieves competitive accuracy with significantly smaller descriptors and faster inference.

TingIS automatically identifies and reacts to risk events from customer incidents in real-time to prevent financial losses.

This research introduces a method to improve transformer models' symbolic reasoning by addressing representational collapse, enabling generalization to unseen tokens and offering insights for fine-tuning open-weight models.

A self-supervised framework for learning and manipulating the flow of time in videos, enabling speed change detection, slow-motion video generation, and temporal super-resolution.

This research benchmarks LLM decoder bias in speech recognition, revealing that audio encoder design, not LLM scale, is key for equitable and robust performance.

A stealthy backdoor attack framework for LLMs that uses natural style triggers and stabilizes payload injection for robust security breaches.

A clinician-inspired framework for summarizing ultra-long capsule endoscopy videos and diagnosing gastrointestinal issues by mimicking human diagnostic workflows.

A defense mechanism that segregates and conceals poisoned data clusters during training to prevent backdoor attacks on deep neural networks.

A scale-adaptive framework for joint spatiotemporal super-resolution using diffusion models, enabling a single architecture to handle diverse spatial and temporal upscaling factors.

A fine-tuned vision-language model that automatically describes embryo development using natural language, outperforming commercial models with limited data.

GiVA offers a parameter-efficient fine-tuning method for large models that matches LoRA's speed while significantly reducing rank requirements, making adaptation more accessible.

GeoMind is an agentic framework for lithology classification in well logs, providing geologically plausible and evidence-grounded decisions through reasoned tool invocation.

CAP is a prompt-driven framework for controllable and reversible unlearning in LLMs, enabling selective knowledge removal without parameter updates.

A zero-shot LLM-generated text detection method that requires no preference collection or additional training, outperforming existing methods.

PLOTTER is a framework for complex narrative generation that uses graph-based reasoning to ensure global coherence and character development.

An end-to-end geocoding framework using LLMs and Chain-of-Thought reasoning to accurately handle explicit and vague location queries.

MMTR-Bench, a new benchmark to evaluate Multimodal Large Language Models' ability to reconstruct masked text directly from visual context.

A graph neural network with attention predicts drug synergy, reducing expensive experimental validation for combination therapies.

A framework for modeling annotator-specific rationales to generate fine-grained explanations that improve predictive performance and represent disagreement faithfully.

A U-Net based deep learning model for rapid and accurate flood hazard mapping, serving as a computationally efficient alternative to traditional hydraulic simulations.

A benchmark and framework for training multimodal models to reason like expert botanists through intent-driven, multi-turn inquiry.

A hybrid AI framework integrates LSTM-based demand forecasting with MILP optimization to enhance supply chain resilience and efficiency, reducing costs and stockouts.

This work proposes a novel 'promotion algorithm' for event log prediction that matches or exceeds neural model accuracy with lower computational cost, offering a stable and efficient alternative.

A novel framework for probabilistic neural network verification efficiently generates probabilistic hulls to compute guaranteed ranges for safe probabilities, outperforming state-of-the-art.

A self-adaptive system that automatically refines LLM prompts for task plan explanations by modeling user cognitive states.

MiMIC addresses visual modality collapse and semantic misalignment in universal multimodal retrieval with a novel fusion architecture and robust training.

A multi-agent AI system using generative video and computer vision for personalized at-home physiotherapy with real-time pose correction.

This research develops a multicalibration technique for Large Language Models to ensure unbiased prevalence estimation across diverse populations, applicable to various classification tasks.

ResVLA is a novel generative VLA policy that anchors robotic motion on predicted intent, refining local dynamics for improved efficiency and robustness.

An interactive URL triage system that uses a sandboxed operator agent and an adjudicator agent to detect sophisticated phishing campaigns with high precision and recall.

A plug-and-play framework that uses multi-agent reasoning to mitigate gender bias in translation quality estimation without sacrificing accuracy.

This research introduces CorridorVLA, a novel approach for generative action heads in Vision-Language-Action models that uses explicit spatial anchors to guide action generation, improving success rates on challenging benchmarks.

An active visual scaffolding framework for situated dialogue agents to maintain common ground by converting dialogue state into a persistent visual history.

A constraint-guided generative framework for discovering magnetic insulators by integrating language models with evolutionary selection and first-principles validation.

IRAP quantifies vague software performance requirements into mathematical functions through interactive retrieval-augmented preference elicitation, achieving significant improvements.

An interactive visualization tool that allows users to define custom evaluation priorities for LLM leaderboards, improving transparency and context-specific model assessment.

A dataset and benchmark for trustworthy visual reasoning that links reasoning steps to image evidence, improving LVLM performance and trustworthiness.

TaNOS is a continual pre-training framework that improves the robustness of numerical reasoning over table data by decoupling domain semantics and numerical operation structure.

Detect cross-session threats in AI agents with a novel dataset, measurement framework, and bounded-memory reader algorithm.

PolyChartQA, a new dataset and benchmark for question answering on multi-chart images, reveals significant performance drops in state-of-the-art MLMs, highlighting a critical area for improvement in visual reasoning.

An adaptive framework for test-time compute allocation in LLMs that optimizes performance and reduces inference costs by dynamically adjusting computation and generation strategies.

An AI framework that uses language models and physics validation to discover new high-kappa dielectric materials, expanding the known space by 35%.

Extending the concept of empowerment to multi-agent systems to study the emergence of complex group behaviors and organization.

This research identifies representation as a key bottleneck in abstract visual reasoning by demonstrating how symbolic input significantly improves LLM performance on complex visual tasks, suggesting a path for more capable AI systems.

This work introduces a novel approach to identify critical contributors in bipartite dependency networks by modeling the problem as a coalitional game and proposing a fast, near-optimal algorithm.

A multimodal approach combining graph-based learning and LLMs for open-domain event extraction that outperforms state-of-the-art.

This research reveals systematic ideological bias in LLMs' economic causal reasoning, showing a consistent skew towards intervention-oriented perspectives and highlighting the need for direction-aware evaluation in policy analysis.

This research proposes GROUNDING$.$md, an epistemic grounding document for agentic AI coding that enforces scientific correctness and best practices, democratizing bespoke software development.

Generate realistic room impulse responses by disentangling spatial and material influences for enhanced acoustic control.

A system to ensure accurate observability in distributed AI inference pipelines by addressing clock skew issues.

This LLM-based method enhances online recruitment by augmenting low-quality job descriptions and using category-aware MoE to identify similar candidate-job pairs, improving conversion rates.

Fine-tuning LLMs for detecting machine-generated code across various programming languages and subtasks.

Quantifying LLM bias by disentangling implicit linguistic signals from explicit user profiles to reveal safety paradoxes.

A formal framework for diffusion modeling on quotient spaces, simplifying learning for tasks with inherent symmetries like molecular structure generation.

A new benchmark system for diagnosing and mitigating prompt-induced hallucinations in vision-language models (LVLMs).

A novel training objective for reinforcement learning introduces dynamical priors to shape temporally coherent decision-making without altering rewards or environments.

Automate science classroom discourse analysis by jointly classifying utterance type and reasoning components using LLM-augmented RoBERTa.

A challenge benchmark for remote sensing infrared image super-resolution, driving research and development of effective solutions.

An efficient algorithm for identifying broadly agreeable topics in online communities by modeling consensus as an interval in a one-dimensional opinion space.

A metamorphic testing approach to diagnose and mitigate data leakage in LLM-based program repair.

CoFEE enhances LLM-based feature discovery by enforcing cognitive reasoning behaviors for improved predictability and efficiency.

Build AI-powered game agents leveraging large language models for creating interactive learning environments.

This paper theoretically identifies a fundamental geometric blind spot in supervised learning and proposes a diagnostic and a minimal repair.

A new evaluation framework for rule-governed AI that measures policy-grounded correctness, moving beyond simple agreement metrics.

This study introduces an attention-based multiple instance learning framework using pre-trained pathology foundation models to predict lung adenocarcinoma growth patterns from whole slide images.

A novel image quality assessment method using causal inference and decoupled representation learning to achieve competitive performance across various image domains and settings.

Leveraging memristor dynamics for efficient and robust image classification in reservoir computing systems.

Verbal Process Supervision (VPS) is a training-free framework that enhances LLM reasoning by using structured natural-language critique, outperforming existing methods on complex benchmarks.

This paper proposes an efficient knowledge distillation method to integrate LLM-generated user profiles into sequential recommenders without requiring LLM inference at serving time.

A framework using meta-predicates and a DSL enhances trustworthy clinical decision support by asserting epistemological constraints on decision rules for auditability.

A framework for analyzing the robustness of POMDP policies to observation perturbations, with applications in robotics and operations research.

A Python package for semisupervised mean estimation that calibrates prediction models to improve accuracy and efficiency.

A statistical certification framework for AI risk regulation that provides auditable bounds on system failure rates without requiring model internals.

A novel approach to generate adversarial malware samples that evade detection and minimize drift signals through similarity-constrained perturbations.

Generate privacy-compliant models from existing trained models without retraining, adapting to any differential privacy requirement.

A live platform for studying social dynamics in autonomous visual agent networks, revealing insights into communication and identity.

A diagnostic framework and mitigation technique to address alignment faking in language models by identifying and steering away from value conflicts.

A novel preference fine-tuning method to align LLMs with domain-specific human preferences for generating online review responses, addressing hallucination and conservatism.

Combining LLMs with differential privacy offers a better privacy-utility trade-off for de-identifying Dutch clinical notes compared to traditional methods.

Automates the engineering of AI agent harnesses for complex tasks, eliminating the need for human intervention in adapting agents to new domains.

A novel LLM architecture decouples user data from shared weights using composable adapters and deletable proxies for privacy-preserving personalization and deterministic unlearning.

A multi-stage deep learning framework that uses transformer predictions as a warm-start for traditional solvers to accelerate unit commitment in power grids.

Novel replay buffer engineering and curriculum RL techniques for noise-robust quantum circuit optimization.

A new multi-turn attack technique that exploits stateless moderation in LLMs by distributing adversarial intent across isolated interactions.

A logic-based approach for inferring high-level temporal events from timestamped data, with applications in medical domains like disease episode detection.

Analysis of task-specific subnetworks in multi-task RL for autonomous underwater navigation to improve interpretability and efficiency.

An agentic architecture that translates natural language research questions into scientific workflows using LLMs, generators, and domain-specific 'Skills'.

This paper explores the use of Answer Set Programming with Quantifiers (ASP(Q)) for inconsistency-tolerant querying of prioritized data, introducing new semantics and implementations for optimal repairs.

DIVERT is an efficient, snapshot-based, coverage-guided user simulation framework for systematic exploration of agent-user interactions to evaluate LLM agents.

Brief chatbot interactions can cause lasting, undetected shifts in human moral values, highlighting a vulnerability to AI manipulation.

This paper introduces Deductive ASPIC®, a novel formal framework for structured argumentation that satisfies critical rationality postulates under credulous semantics.

A computational analysis of media framing of human-elephant conflicts in India using a multi-model sentiment framework to identify negative portrayals and inform conservation efforts.

A low-complexity dilated CNN approach for periodic signal denoising and waveform estimation, suitable for resource-constrained environments.

This paper presents theoretical architectures for robust self-organizing energy systems under information and control constraints, focusing on agent-based cyber-physical systems.

A family of smooth activation functions that offer ReLU-like performance with rational arithmetic and improved optimization for deep neural networks.

Investigating the load-bearing role of cross-entropy in predictive coding networks by testing its removal and impact on bidirectional predictive coding.

Analyzing and proposing new evaluation protocols for vision-language-action models in robotics to assess safety and robustness.

Theoretical framework explaining and predicting the sharpness gap in SGD for neural network training, offering insights into optimization dynamics.

Investigating the role of the Gaussian scale parameter in Gaussian Kolmogorov-Arnold Networks to improve their approximation behavior and provide a practical design principle.

An integrated framework for packing, placement, scheduling, and routing in personalized pharmaceutical production using planar transport systems.

AGNT2 provides a dedicated Layer 2 infrastructure for autonomous AI agent coordination, optimizing for high-frequency service invocations.

Using SAT solving and LLM-generated code to discover mathematical graph families and formalize proofs.

Develops and applies methods to measure LLM propensity for unsanctioned behavior by analyzing environmental factors, finding equal contributions from strategic and non-strategic influences.

A systematic review and taxonomy of integrating Reinforcement Learning with Model Predictive Control for linear systems.

A novel approach to data interaction for complex domains, inspired by active objects.

A formal model exploring how AI compliance layers in government can be strategically exploited by future administrations, leading to increased vulnerability despite initial oversight improvements.

A methodology using Generalized Procrustes Algorithm to measure intra-modal representational convergence at the single-stimulus level, modulating cross-modal convergence between vision and language models.

A research agenda for designing AI systems that provide cognitive support by actively helping users form and refine their intent through time, addressing 'Fantasia interactions'.

This paper investigates reasoning primitives in LLMs, comparing hybrid and attention-only architectures on tasks requiring recall and state-tracking.

A theoretical exploration of how varying degrees of autonomy in artificial systems impact economic welfare theorems in post-AGI economies.

This paper explores the challenges of translating AI governance requirements into software development practice within an AI startup, focusing on expert collaboration to address the 'Last Mile Challenge'.

Developing a refined description logic for information behavior, incorporating procedural and oracle-sensitive layers with categorical semantics.

Investigating the psychological factors influencing university students' willingness to disclose their use of AI tools in academic work.

A taxonomy and framework for understanding and mitigating fairness risks in sequential decision-making systems under uncertainty.