A video diffusion model framework that restores high dynamic range and realistic detail in low dynamic range videos, enabling controllable re-exposure.

Flowr automates retail supply chain workflows with agentic AI for enhanced efficiency and scalability.

A framework for synthesizing scientific graphics from images into editable code, featuring a large dataset, benchmark, and novel reinforcement learning for state-of-the-art performance.

Lightweight adaptation of vision-language models for species recognition and habitat interpretation using drone thermal imagery.

A visually grounded AI model for improved video understanding in practical applications.

Claw-Eval: A trustworthy evaluation suite for autonomous agents that provides trajectory-aware grading, safety, and robustness assessment.

PCA-Triage is an unsupervised, parameter-free algorithm for adaptive sensor sampling on edge devices, outperforming baselines under bandwidth constraints.

This research analyzes how supervised fine-tuning and reinforcement learning evolve reasoning in language models for chess, releasing checkpoints and code that surpass leading open-source models.

Introduces MedGemma 1.5, a multimodal foundation model for medical AI, enhancing analysis of imaging, EHRs, and clinical documents.

A benchmark and framework for evaluating and improving LLM productivity agents in realistic, multi-service simulated workspaces, addressing safety and capability gaps.

An instruction-tuned LLM framework for parsing and mining unstructured HPC logs, achieving state-of-the-art accuracy and enabling actionable insights from massive telemetry.

VideoStir is a structured, intent-aware RAG framework for understanding long videos by leveraging spatio-temporal graphs and an MLLM-backed relevance scorer.

A reward decomposition approach that disentangles sycophancy in LLMs by separating pressure resistance from evidence responsiveness, leading to more principled alignment.

A novel framework for reconstructing high-fidelity, physically plausible 3D models of interacting people from a single image by explicitly modeling group context and human interactions.

A novel veterinary diagnostic system uses Vision-Language Models to guide Flow Matching for precise lesion segmentation, combined with Random Matrix Theory for interpretable and accurate pneumothorax detection.

An open-source, agentic AI platform for unified security operations that automates incident context, rule generation, and attack reconstruction, with built-in governance and multi-tenant support.

An autonomous framework for 24/7 deep learning experimentation with zero-cost monitoring and efficient memory management.

A retrieval layer for large agent skill libraries that constructs an executable skill graph to retrieve dependency-aware skill bundles, improving reward and reducing token costs.

An LLM-powered agent that orchestrates CAD-CAE tools to automate industrial design optimization, trained on a new industry-aligned dataset.

MCPSHIELD provides a formal security framework for MCP-based AI agents, offering a taxonomy, verification models, and defense mechanisms to address critical threats.

An adaptive reasoning framework for multimodal embeddings that selectively invokes reasoning only when necessary, improving performance and reducing latency.

An open-source inference-time protocol that blinds LLMs to entity identities, enabling auditability of prior contamination in data analysis and drug target prioritization.

UniCreative is a unified reference-free reinforcement learning framework that reconciles long-form coherence with short-form expressiveness in creative writing.

Market-Bench is a comprehensive benchmark evaluating LLMs in economically-relevant tasks through multi-agent supply chain competition.

JTON is a token-efficient JSON superset that reduces LLM processing costs for structured data by up to 60% with a novel tabular encoding.

LudoBench is a new benchmark and simulator for evaluating LLM strategic decision-making in a complex board game, revealing prompt sensitivity and behavioral archetypes.

Transform any software application into an intelligent agent environment for scalable training and interaction.

A lightweight agent benchmark with configurable horizons and difficulty, designed for fast and reproducible evaluation.

Revolutionize teaching in film and motion picture courses with AI-powered teaching assistants.

A multi-stage validation framework enables trustworthy, large-scale clinical information extraction using LLMs without extensive manual annotation.

Compiled AI transforms LLM-generated workflows into reliable, efficient code for enterprise automation.

Develops deterministic learned metrics for multilingual text evaluation that are faster and more consistent than LLM judges.

IntentScore evaluates and ranks actions for computer-use agents, improving task success rates by learning from diverse offline GUI interactions.

A multi-agent framework that automates the writing of AI research papers from raw materials, including literature synthesis and generated visuals.

A scalable Decision Pre-Trained Transformer trained with Flow Matching achieves strong generalization in in-context reinforcement learning across diverse tasks, offering a viable alternative to expert distillation for generalist agents.

Curiosity-driven planning for LLMs enhances test generation by exploring program branches more effectively, leading to higher coverage.

This research evaluates LLM sensitivity to patient question framing in medical QA, revealing significant inconsistencies that impact high-stakes applications.

A three-channel denoising diffusion probabilistic model synthesizes dual-view mammograms simultaneously, addressing data gaps and enabling cross-view AI applications in breast imaging.

EffiPair improves LLM-generated code efficiency at inference time by using relative contrastive feedback between program pairs, reducing runtime and memory usage.

Generates realistic synthetic multivariate time series with dynamic correlations to improve foundation model training.

XMark is a novel multi-bit watermarking method for LLM-generated text that offers improved decoding accuracy and text quality, even with limited token counts.

A CSI-free hierarchical reinforcement learning system optimizes reconfigurable intelligent surfaces for massive signal strength improvements in next-generation wireless networks.

A modality-aware VQ-VAE for reconstructing multi-modal brain MRIs, enabling robust generative modeling and cross-modal analysis with superior fidelity.

A training-free method to fuse multiple orthogonal adapters for diffusion models, enabling combined concept and style generation.

A novel 3D reconstruction model that significantly improves fine-grained texture and appearance recovery by scaling transformer context windows, achieving state-of-the-art results.

A framework for democratizing robotic evaluation by enabling natural language task authoring and crowd-sourced evaluation.

A hybrid deep learning framework combining Transformers and Graph Attention Networks for proactive delivery delay prediction in smart logistics.

Leveraging LLMs with RAG and summarization techniques to significantly improve clinical trial patient screening.

A pipeline for curating structured reasoning data to enhance LLM long-context reasoning, demonstrated through open-source code and models.

A new benchmark dataset and evaluation framework to identify and mitigate attribution bias in large language models.

ID-Sim is a novel feed-forward metric designed to accurately assess human-like selective sensitivity to identities in images, enabling better personalized generation.

A framework for sequential clinical diagnosis that learns diagnostic trajectories and reduces uncertainty, with available code and demonstrated performance on a medical benchmark.

A framework for building multi-agent systems to solve complex chemistry retrosynthesis planning problems, outperforming state-of-the-art.

PRISM-MCTS enhances reasoning models by learning from trajectories with metacognitive reflection, reducing computational redundancy and improving efficiency.

A framework for making LLM agents auditable by defining dimensions of auditability and proposing mechanisms to ensure accountability.

This paper provides a comprehensive analysis and benchmark of LLM-based automated penetration testing frameworks, identifying key architectural designs and empirical performance.

A context-aware simulator and dataset for intelligent energy management in renewable systems, enabling advanced control algorithms and prediction models.

A multi-agent RAG framework for IoT intrusion detection that uses LLMs grounded by a self-building experience library for explainable, continual learning.

A lightweight, model-agnostic style transfer augmentation recipe that significantly improves computer vision model generalization from synthetic to real-world data.

A novel deep learning model for robust cross-session motor imagery decoding in brain-computer interfaces, outperforming existing architectures.

A novel paradigm for long-sequence symbolic music generation that mitigates error accumulation using anchor features, significantly outperforming existing methods.

A novel Multi-Round Value Factorization framework that breaks suboptimal stable points in multi-agent reinforcement learning by iteratively filtering inferior actions.

A novel framework for triaging hallucination risks in medical LLMs by grounding detection in retinal evidence.

A vision-language reward modeling framework that dynamically decomposes evaluation into interpretable dimensions for improved VLM alignment and hallucination mitigation.

A novel reward mechanism that significantly improves LLM reasoning efficiency and accuracy by guiding uncertainty reduction.

An automated framework using locally deployed LLMs to audit hospital discharge summaries for critical care transition elements, improving patient safety and reducing readmissions.

Achieve near-optimal, training-free compression for 3D reconstruction models in seconds, significantly reducing storage needs with minimal quality loss.

CAKE is a novel benchmark for evaluating LLMs on cloud architecture knowledge, revealing insights into their capabilities and limitations.

A query-side region cropping framework for multi-modal retrieval that dynamically focuses on relevant image regions to improve re-ranking accuracy.

HYVE is a framework for LLM context engineering that efficiently processes large machine data payloads by transforming them into hybrid views, reducing token usage and improving output quality.

LLM4CodeRE is a domain-adaptive LLM framework for bidirectional code reverse engineering, outperforming existing tools in malware analysis.

CRFT is a unified coarse-to-fine framework using a transformer-based architecture for robust cross-modal image registration, offering a generalizable paradigm for multimodal spatial correspondence.

An LLM-centered framework for traffic signal control that uses RL-generated trajectories and multi-LLM deliberation to outperform state-of-the-art baselines.

A new dataset and strategies improve Vision-Language Models' ability to understand multimodal puns, a subtle form of humor requiring cross-modal reasoning.

Context-Agent models multi-turn dialogue history as a dynamic tree structure to improve LLM coherence and efficiency in non-linear conversations, with a new benchmark for evaluation.

A hybrid KV cache compression framework for multimodal LLMs that significantly reduces memory and latency without performance loss.

A multi-agent framework that refines text prompts for complex text-to-video scenarios, improving alignment and generation quality.

Graph-PiT enhances part-based image synthesis by using graph priors and a Hierarchical Graph Neural Network to improve structural coherence and control over generated visuals.

A foundation model for gastrointestinal endoscopy diagnosis that uses analogical reasoning to improve generalization and adaptability across diverse datasets and disease types.

A benchmark and LLM-based verifier to evaluate code patch quality beyond test pass rates, revealing significant design constraint violations in current LLM agents.

A diagnostic framework to identify and address 'surface compliance' in LLM knowledge editing, ensuring genuine memory modification for trustworthy deployment.

ActivityEditor is a dual-LLM-agent framework for zero-shot cross-regional human mobility trajectory generation in data-scarce scenarios.

Rectified Schrödinger Bridge Matching enables few-step visual navigation for embodied agents by balancing multimodal action distributions and path efficiency.

A novel linear-time attention replacement that significantly reduces computational cost for transformers across diverse domains while maintaining performance.

Pareto-Lenient Consensus (PLC) is a game-theoretic framework for LLM alignment that enables negotiation-driven optimization to escape local optima and explore the Pareto frontier.

A framework to efficiently convert attention architectures in pre-trained LLMs for substantial inference cost reduction without retraining.

STEP-HRL is a hierarchical reinforcement learning framework for LLM agents that enables step-level learning, reducing computational cost and improving performance.

FastDiSS is a novel training framework for diffusion language models that improves robustness to self-conditioning errors, enabling up to 400x faster inference for sequence-to-sequence generation.

A Domain-Decomposed and Shifted Physics-Informed Neural Network (DDS-PINN) framework for complex fluid flows that achieves accurate results with minimal supervision and data.

OntoTKGE is a novel framework that enhances temporal knowledge graph extrapolation by integrating ontological knowledge to improve entity embeddings and handle sparse interactions.

A drop-in framework for Large Language Models that enables continuous adaptation to new information at inference time without costly retraining.

A novel framework enhances zero-shot generalization for unsupervised reinforcement learning in visual environments by decoupling representation learning and improving skill controllability.

A new benchmark and evaluation framework for LLM reliability and adversarial security tailored for Swiss financial and regulatory contexts.

LAG-XAI is a geometric framework for interpretable paraphrasing in Transformer latent spaces, enabling efficient LLM hallucination detection.

Theoretical characterization of ReLU networks enables exact generation of graphs within a specified edit distance, eliminating reliance on training data and guaranteeing validity.

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.

A synthetic data generation pipeline for doctor-patient conversations to train and evaluate long-form audio summarization models.

SemLink is a semantic-aware automated test oracle that verifies hyperlink integrity by comparing source context and target content, offering a faster and more efficient alternative to LLMs for web quality assurance.

A framework combining LLMs and vision models for language-guided pulmonary screening, using graph reasoning and selective fine-tuning to improve accuracy and stability.

A new LLM agent architecture grounded in human values significantly improves behavioral fidelity and reduces polarization in real-world interactions.

A unified framework for diagram code generation that uses visual feedback to train models without manual code annotation, achieving state-of-the-art results.

A framework for evaluating LLM cybersecurity capabilities in industrial OT environments, with available code and a tool scaffold to improve performance on critical tasks.

A deep research agent that generates trustworthy reports by progressively estimating and calibrating confidence in its claims.

A method to improve reasoning and visual grounding in diffusion multimodal language models by penalizing premature answers and guiding visual input utilization.

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 layered translation method connects governance standards to implementable runtime guardrails for agentic AI systems, demonstrated with a procurement agent case study.

Extending diffusion models with temporal adapters to generate temporally coherent synthetic time-series data for privacy-preserving augmentation.

FI-LDP-HGAT offers feature-aware anisotropic local differential privacy for utility-preserving graph representation learning in metal additive manufacturing.

This work introduces an adaptive budget allocation policy for LLMs that significantly reduces token usage in multi-turn reasoning without sacrificing accuracy.

ALTO accelerates LoRA hyperparameter tuning and improves GPU utilization by orchestrating heterogeneous fine-tuning jobs.

A framework combining LLMs and SLMs for joint knowledge base completion and question answering, improving both tasks.

Shot-Based Quantum Encoding (SBQE) for quantum neural networks that improves data loading efficiency and accuracy on near-term hardware.

This research reveals how social dynamics like conformity and persuasion undermine decision-making in LLM agent collectives, highlighting critical vulnerabilities.

SnapFlow compresses multi-step denoising in flow-matching Vision-Language-Action models into a single forward pass, achieving significant speedups with minimal performance degradation.

An end-to-end system for agent self-improvement that identifies lacking capabilities and synthesizes targeted training environments using LoRA adapters.

A diagnostic question-answering framework that significantly improves IT support resolution rates by maintaining diagnostic state and aggregating evidence.

An efficient machine unlearning method for diffusion models that uses a saliency mask to prioritize parameter updates, reducing training time without sacrificing efficacy.

Kolmogorov-Arnold Fuzzy Cognitive Maps (KA-FCMs) introduce learnable B-spline functions on edges to model non-monotonic causal relationships in dynamic systems while preserving interpretability.

This research empirically analyzes and validates metrics for interview response quality, identifying direct relevance to research questions as the strongest predictor of contribution.

This paper introduces advanced optimization strategies for Physics-Informed Neural Networks (PINNs) to accelerate convergence and achieve high accuracy in solving differential equations.

This research empirically studies cross-domain knowledge transfer for in-context learning, demonstrating that source-domain demonstrations can improve target-domain inference by repairing reasoning structures.

This method uses knowledge-weighted fine-tuning to train LLMs to express uncertainty and avoid hallucinations for out-of-scope queries.

A neurosymbolic system that uses LLMs and Logic Tensor Networks to validate regulated procurement offers with auditable decisions.

A system using an AI interviewer and interactive visualization to elicit personal experiences and increase legitimacy in collective decision-making.

Demographics-agnostic training techniques significantly reduce bias in wake-up word detection across diverse speaker populations.

CODESTRUCT reframes code agents to operate on structured action spaces, improving reliability and efficiency for code manipulation tasks.

An agentic AI system that accurately profiles user expertise in real-time during human-machine interactions.

A novel singing style conversion system that advances fine-grained style conversion and control, achieving best naturalness performance in a challenge.

A diagnostic tool to determine when LLM-driven reasoning is necessary for bandit problems, often favoring lightweight numerical bandits.

ResearchEVO is an end-to-end framework that automates scientific discovery by evolving algorithms and then generating publication-ready research papers, validated on quantum computing and PINNs.

A retrieval method that uses domain-invariant neurons to boost LLM reasoning by transferring knowledge from cross-domain examples.

MC-GenRef enables annotation-free mammography microcalcification segmentation using synthetic data and test-time generative refinement to improve accuracy and robustness.

This research explores how users' mental models of AI writing assistants impact their control behavior and output quality, revealing a trade-off between system understanding and error detection.

Generates animatable 3D vehicle models from single images for realistic autonomous driving simulation.

AI-driven spatial control for reflector arrays bypasses computational bottlenecks in wireless networks using Multi-Agent Reinforcement Learning.

An open-source tool for automatic analysis of data locality in AI and HPC kernels to optimize performance.

Phase-Associative Memory (PAM) is a complex-valued recurrent sequence model that shows competitive performance to transformers on WikiText-103.

Modeling prior authorization policy retrieval as an adaptive sequential decision-making problem using offline RL to improve accuracy and efficiency.

A post-hoc debiasing strategy for generative recommendation systems that rebalances item tokenization to mitigate popularity bias and improve recommendation performance.

Characterizes LLM chain-of-thought generation as trajectories in representation space, enabling mid-reasoning prediction of correctness and inference-time intervention.

This research demonstrates that diffusion-based image compression methods offer superior robustness to bit-flip errors compared to traditional codecs.

Investigating how LLMs can express uncertainty explicitly through verbalized confidence scores and reasoning-time markers to improve decision-making.

A novel metric to evaluate learner representations for differentiation in educational AI systems, independent of instructional outcomes.

Spec Kit Agents enhance AI coding assistants by grounding them in repository context, reducing hallucinations and architectural violations in spec-driven development.

A transfer-oriented memory framework enables multimodal LLM agents to derive actionable knowledge from prior interactions for efficient task solving.

Identifies a 5/3 spectral scaling in contextual language representations from transformer models, suggesting scale-free semantic integration.

A two-stage framework that predicts elevation maps from satellite RGB imagery for Radio Environment Map estimation, eliminating the need for 3D data at inference.

ASTRAL is a multimodal LLM-driven technique to reconstruct and assess the security risks of cyber-physical systems with incomplete documentation.

A formal verification study revealing that over 55% of AI-generated code artifacts contain security vulnerabilities, with no frontier LLM achieving a passing grade.

An LLM-guided evolutionary framework synthesizes interpretable traffic signal control skills with rationale and executable code for adaptive systems.

Investigating the use of large language models as semantic judges to assess the reliability of power line segmentation in drone inspections.

This research explores the capability of Large Language Models to reinvent foundational computer science algorithms through an unlearning and reinvention pipeline, demonstrating potential for algorithmic innovation.

This paper provides a foundational framework for the forensic analysis of agentic AI systems, identifying recoverable traces and proposing an artifact taxonomy to aid in investigations.

Nidus is a governance runtime that mechanizes the V-model for AI-assisted software delivery, ensuring engineering invariants are maintained.

This research explores the limitations of current language models in achieving second-order generalization for object categorization, suggesting a fundamental gap in their learning mechanisms.

An empirical audit of instructed code-editing benchmarks reveals significant gaps compared to real-world usage, proposing desiderata for more representative benchmarks.

A theoretical exploration of multi-token prediction for LLMs to develop more consistent world models, addressing structural hallucinations.

An empirical study investigates 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.

Improving text-to-image model interpretability by selectively aggregating attention maps.

A novel Multi-Agent Pathfinding variant with enhanced conflict-based search and graph discretization for efficient navigation on graphs with non-unit integer costs.

An exhaustive inventory of 12,007 invariant landscape classes for pseudo-Boolean functions provides a resource for algorithm understanding and benchmark design.

Simulating the evolution of AI alignment and values to identify and mitigate deceptive models through iterative testing and adaptive design.

This paper introduces QA-MoE, a Quality-Aware Mixture-of-Experts framework for multimodal sentiment analysis that handles dynamic noise and missingness by quantifying modality reliability.

A framework that synthesizes exaggerated character motions for animation by reformulating external assistance in impulse space for numerical stability.

Automatic 3D intraoral to 2D photograph superimposition for dental identification using computer vision.

LLMs can generate life stories that accurately reflect individual personality traits, with recovered scores approaching human reliability.

Demonstrates a data exfiltration attack on LLM agents using backdoored tool use, highlighting a critical vulnerability in current agent architectures.

MARL-GPT: A single GPT-based model trained at scale to perform well across diverse multi-agent reinforcement learning environments.

A hybrid QUBO framework for neural network pruning that integrates gradient-aware metrics with global combinatorial optimization.

Applies quantum search to magic square problems, demonstrating theoretical advantages but facing scalability challenges.

A mathematical theory models the evolution of self-designing AI systems, exploring how directed design and fitness functions shape AI lineages and potential alignment risks.

This paper explores how simple social cues can lead to the transmission of complex knowledge representations without requiring agents to infer mental states.

A taxonomy and framework for governing machine identities in AI systems to mitigate enterprise and geopolitical risks.

An adaptive framework to reduce serverless cold starts and improve cost-efficiency through probabilistic slot survival prediction and event-driven resource management.

A novel tensor completion framework for semiparametric inference and uncertainty quantification in LLM evaluation using pairwise human judgments.

A polynomial-time algorithm for Thiele voting rules with voter interval preferences, resolving a 10-year-old open problem.

A qualitative study exploring legal professionals' perceptions of AI governance and regulation in Nigeria, highlighting concerns about data privacy and the need for context-specific models.

This paper explores the theoretical intersection of artificial intelligence and the structure of mathematics, proposing AI agents to discover new mathematical concepts and understand formal proofs.

Investigates instruction tuning in LLMs, concluding that it relies on skillful coordination of diverse linguistic capabilities rather than a universal mechanism.

Introduces Tree Decision Diagrams (TDDs) as a more succinct and tractable generalization of OBDDs for Boolean function representation.

This paper argues for a cognitive revolution in AI evaluation, moving beyond behavioral tests to consider internal processes and mechanisms for a deeper understanding of intelligence.

This paper explores the reciprocal trust dynamics between AI systems and humans, and their implications for regulation.

This paper explores a family of spectral normalization rules for deep learning optimization, analyzing them in a mean-field regime using Spectral Wasserstein distances.

Extends the DeGroot model to analyze how AI aggregation affects social learning, identifying a critical threshold in update speed for robust learning improvement.