2393 papers
This paper proposes SToRM, a novel framework that employs a lightweight importance predictor, supervised training with pseudo-labels, and an anchor-context merging module to significantly reduce visual token redundancy in multi-modal LLMs for autonomous driving, achieving up to 30x computational savings while maintaining end-to-end performance comparable to using all tokens.
This paper introduces Agent-guided Policy Search (AGPS), a framework that replaces human supervisors with a multimodal agent acting as a semantic world model to provide precise corrective guidance, thereby significantly improving sample efficiency and scalability in robotic reinforcement learning compared to traditional Human-in-the-Loop methods.
This paper introduces M2RL, a comprehensive study comparing mixed multi-task training versus separate training with model merging for multi-domain Reinforcement Learning with Verifiable Rewards (RLVR), revealing that reasoning-intensive domains exhibit synergistic effects with minimal interference and providing mechanistic insights through extensive experiments.
The paper introduces SkillsBench, a comprehensive benchmark demonstrating that while curated agent skills significantly boost LLM performance across diverse domains—often allowing smaller models to match larger ones—self-generated skills offer no benefit and effects vary widely by task.
This paper proposes a geometric taxonomy of LLM hallucinations into three distinct types (unfaithfulness, confabulation, and factual error) and introduces corresponding detection metrics, the Semantic Grounding Index and Directional Grounding Index, which effectively identify unfaithful and confabulated outputs while revealing that apparent signals for factual errors in existing benchmarks often stem from stylistic annotation confounds rather than genuine geometric distinctions.
TrasMuon is a novel optimizer that enhances the stability and convergence of Muon-style methods by integrating global RMS calibration and energy-based trust-region clipping to preserve near-isometric geometry while mitigating sensitivity to step-size hyperparameters and high-energy outliers.
This paper demonstrates that a lightweight, automated AI pipeline integrating next-generation large language models with citation-based verification can successfully generate and solve sophisticated, research-grade mathematical problems, including previously unpublished questions, with verified results and open-sourced tools.
This paper introduces the Mean Velocity Policy (MVP), a novel one-step generative policy that employs an Instantaneous Velocity Constraint (IVC) to theoretically guarantee high expressiveness while achieving state-of-the-art performance and significantly faster training and inference speeds on challenging robotic manipulation tasks compared to existing flow-based baselines.
The paper introduces Pawsterior, a variational flow-matching framework that enhances simulation-based inference by incorporating geometric confinement for structured domains and enabling the handling of discrete latent structures, thereby improving posterior fidelity and expanding applicability to complex physical systems.
This paper introduces XTF, an explainable framework that improves LLM fine-tuning performance by decomposing token contributions into reasoning importance, knowledge novelty, and task relevance to identify and mask noisy tokens, achieving up to a 13.7% improvement across math, code, and medical tasks.
LongAudio-RAG is a hybrid edge-cloud framework that enables precise, low-hallucination question answering over multi-hour audio streams by converting recordings into timestamped event records for SQL-based retrieval, which then grounds Large Language Model responses in structured evidence rather than raw audio.
CogitoRAG is a novel Retrieval-Augmented Generation framework inspired by human episodic memory that enhances complex reasoning and reduces hallucinations by extracting semantic gists into a multi-dimensional knowledge graph, utilizing query decomposition and entity diffusion for associative retrieval, and employing a fusion-based reranking algorithm to deliver high-density evidence.
This paper introduces CondMedQA, the first benchmark for conditional biomedical question answering, and proposes Condition-Gated Reasoning (CGR), a framework that constructs condition-aware knowledge graphs to dynamically prune reasoning paths based on patient-specific factors, thereby improving the reliability of medical decision-making.
This paper introduces a framework for operational certification of split conformal predictors that moves beyond marginal coverage by providing finite-sample guarantees for critical deployment metrics like commitment frequency and error exposure through Small-Sample Beta Correction, an independent audit-based auditing protocol, and a geometric analysis of Pareto trade-offs.
This paper introduces ABD, a benchmark for default-exception abduction in finite first-order worlds that evaluates ten frontier LLMs on their ability to generate sparse, satisfiability-restoring formulas across three observation regimes, revealing that while models achieve high validity, they struggle with parsimony and exhibit distinct generalization failures.
This paper introduces INDUCTION, a benchmark designed to evaluate the ability of AI models to synthesize compact, generalizable first-order logic formulas that explain target predicates across small finite relational worlds, revealing distinct performance patterns and generalization strategies among recent elite models.
This paper establishes a comprehensive multi-KPI benchmark for Multi-Agent Reinforcement Learning in urban energy management using the CityLearn environment, demonstrating that Decentralized Training with Decentralized Execution (DTDE) consistently outperforms Centralized Training with Decentralized Execution (CTDE) in both average and worst-case performance while offering greater resilience and sustainability.
The paper introduces MrBERT, a family of efficient, open-source multilingual encoders built on the ModernBERT architecture that achieves state-of-the-art performance in specific languages and specialized domains while leveraging Matryoshka Representation Learning to reduce inference and storage costs.
This paper introduces ARLArena, a unified framework that systematically analyzes training instability in agentic reinforcement learning to derive SAMPO, a stable optimization method that ensures consistent performance across diverse agentic tasks.
CryoNet.Refine is a novel one-step diffusion model that automates and accelerates the refinement of atomic structures against cryo-EM density maps, outperforming traditional tools like Phenix in both model-map correlation and geometric quality while supporting diverse protein and nucleic acid complexes.