Papers

New on arXiv

Wed 9 Sep 2026

402 papers

OMP-MoE: Efficient Expert Pruning for Mixture-of-Experts LLMs via Orthogonal Matching PursuitMixture-of-experts → MoE large language models
Harnessing Image Question Dependence for Better VLM Test-time Reinforcement LearningMultimodal LLMs / vision-language models → Embodied and 3D vision-language models
LLaDA-UI: Bringing Block-wise Diffusion to Vision-Language GUI AgentsLLM agents and tool use → GUI and mobile agents
Fixed State, Long Reach: What a Constant-Size Cache Buys Block Diffusion at ScaleDiffusion models → Discrete and text diffusion +1
SearchAtlas: Analyzing Agentic Search Strategies via Evidential Query GraphsLLM agents and tool use → Web, search, and research agents
ReactHuman: A Physics-Grounded Benchmark for Human-Like Reactive Decision-Making in Embodied Multimodal LLMsMultimodal LLMs / vision-language models → Embodied and 3D vision-language models
DriftNet: A Dual-Head Trajectory Transformer for Detecting and Localizing Prompt Injection in LLM AgentsLLM agents and tool use → Agent threats, guardrails and governance +1
Certifying Lower Bounds for Risk-Sensitive Reinforcement Learning under Adversarial State PerturbationsDeep reinforcement learning methods → Adversarial robustness and RL in finance
Flow Duality and Source Geometry for Categorical GenerationDiffusion models → Flow matching, time series and science
Evaluation of Vision-Language Models Across Diverse Coastal EnvironmentsMultimodal LLMs / vision-language models → Embodied and 3D vision-language models
No-Box Vulnerability Analysis: Description-only Detection of Indirect Prompt Injection Vulnerabilities in MCP ServersLLM safety and security → Prompt injection and agent security
Studying Without a Syllabus: Task-Agnostic Environment PreprocessingLLM agents and tool use → General LLM agent frameworks and benchmarks
Tapes Together Strong: The Co-evolution of Computation and CooperationLLM agents and tool use → Social simulation and multi-agent debate
BodyCam-VQA: Enhanced Body-Worn Camera Video Captioning via Multimodal Reasoning and Probe Question GenerationLLM reasoning → Multimodal reasoning
ExaServe: Large-Scale LLM Serving on Exascale HPC SystemsEfficient LLM inference and serving → Serving systems and request scheduling
From Connectivity to Rewards: Dense Reward Learning with Directed State GraphsDeep reinforcement learning methods → Deep RL algorithms and representations
Big Enough to Break Out: Tracking the Rising Capability of LLM Penetration-Testing AgentsLLM agents and tool use → Agent threats, guardrails and governance
A Bellman Optimality Equation for PlasticityDeep reinforcement learning methods → Deep RL algorithms and representations
When Synthetic Data Hurts: On Catastrophic Forgetting in Skill Retrieval for LLM AgentsLLM agents and tool use → Agent skill learning and evolution +1
Think Before You Link: Rarity, Reasoning, and Retrieval in Multilingual Entity LinkingLLM reasoning → Multimodal reasoning
The Truth Was Never Gone: Perfect Aliasing in Compliant-Context Truth ProbesMechanistic interpretability → Probing deception, bias and beliefs
AcFlow: Controlling Text-to-Image Diffusion Transformers via Learned Conditional Activation FlowDiffusion models → Image generation, text-to-image and editing
Architecting the Secure AI-SOC: A Neurosymbolic Framework for Pipeline Integrity and Threat MitigationLLM safety and security → Prompt injection and agent security
GEOSTEER: Geodesic Optimization for Activation Steering in Large Language ModelsMechanistic interpretability → Activation steering methods
Understanding LoRA Rank Trade-offs in Diffusion Model Fine-TuningDiffusion models → Diffusion foundations and sampling
Programmable World ModelWorld models → Video-generation world models and simulators
Show-Harness: Just a VLM Agent Can Play RobotsMultimodal LLMs / vision-language models → Embodied and 3D vision-language models +1
Avatar: Toward Autonomous End-to-End Orchestration of Scientific Workflows using LLMsLLM agents and tool use → Software engineering coding agents
DUET-DINO: Simultaneous Cross-View World Modeling for Latent Planning in Robot ManipulationWorld models → Embodied and robot world models
Semigroup-JEPA: Latent Dynamics Consistency for Zero-Shot Physics GeneralizationWorld models → Latent predictive world models
Advanced Brain Tissue Imaging with Data-Consistent Diffusion Priors in Laminographic X-Ray NanoimagingDiffusion models → Medical imaging and segmentation
JarvisGUI: Towards Cross-Device GUI Agents with Dynamic Task CompositionLLM agents and tool use → GUI and mobile agents
Building Multilingual Bridges: Data Mixing as the Pillar of Generalization for In-Language ReasoningLLM reasoning → Reasoning benchmarks and evaluation
ConvMem: Convolutional Memory for Long-Context ReasoningLLM reasoning → Test-time compute and efficient thinking
Forgetting Only What Matters: Layer-Selective Unlearning toward Robust LLMsLLM safety and security → Machine unlearning
Glyph: A Multi-Strategy Agentic System for Column Description and Sensitivity-Ontology Tagging of Enterprise Data CatalogsLLM agents and tool use → Web, search, and research agents
TrajMark: Ownership Attribution and Segment-Level Tamper Localization for Coding-Agent TrajectoriesLLM agents and tool use → Agent threats, guardrails and governance +1
Frequency-Conditioned Flow Matching for Vision-Language-Action ModelsVision-language-action models → Efficient VLA inference and adaptation +1
Searching for New Physics with Reinforcement LearningDeep reinforcement learning methods → RL for energy, networks and operations
Data-Driven Risk Fields for Safer End-to-End Autonomous DrivingEnd-to-end autonomous driving → End-to-end system architecture
PACE: Perceived-Latency-Aware Cascading Service Routing and Filler Control for QoE-Efficient Retrieval-Augmented Dialogue ServingRetrieval-augmented generation → Conversational and speech RAG
SceneHI: High-Resolution 3D-Consistent Scene Texturing with Controllable IlluminationDiffusion models → 3D, point cloud and scene generation
Beyond One-Size-Fits-All: Sample-Adaptive Strategy Routing for Vision Token Pruning in MLLMsMultimodal LLMs / vision-language models → Visual token pruning and efficiency +1
From Symbolic Perception to Logical Deduction: A Framework for Guiding Language Models in Geometric ReasoningLLM reasoning → Spatial and embodied reasoning
Geometry Without Coordinates: LiDAR Diffusion as a 3D Feature BridgeDiffusion models → 3D, point cloud and scene generation
On-Policy Distillation for Vision-Language Model Adaptation, an Effective Paradigm on Low-Quality Multimodal DataMultimodal LLMs / vision-language models → MLLM training and visual instruction tuning +1
Decoupled Self-Forcing Distillation for Streaming Talking Head GenerationDiffusion models → Audio, speech and music
A Dominant Diffuse Phase in the Sparse Autoencoder Phase DiagramMechanistic interpretability → Sparse autoencoder training and evaluation
TRACE: Trajectory-robust Admission with Evidence Ordering for Efficient GUI AgentsLLM agents and tool use → GUI and mobile agents
Training Trajectories Determine Circuit Removability in Annealable Soft-Prior TransformersMechanistic interpretability → Circuits and mechanistic analysis
KVShareArena: KV-Cache Reuse Across Contexts and Model CheckpointsEfficient LLM inference and serving → KV cache optimization
A-JIT: Agentic Just-In-Time Software ConstructionLLM agents and tool use → Software engineering coding agents
RoboFolDeX: A Physical-World Benchmark for Long-Horizon Robotic Manipulation of Deformable ObjectsWorld models → Embodied and robot world models
LiteRAG: Cost-Efficient Graph-Based Retrieval-Augmented GenerationRetrieval-augmented generation → Knowledge-graph and structured-knowledge RAG
The Answer Path and the Grounding Instruction in LLM Question Answering over Knowledge GraphsRetrieval-augmented generation → Knowledge-graph and structured-knowledge RAG
UOT-Gap: A Variational Principle for the Modality Gap in Vision-Language Models via Unbalanced Optimal TransportMultimodal LLMs / vision-language models → MLLM training and visual instruction tuning
Why Sample What You Can Enumerate? Exact Policy Optimization for Genomic Tool SelectionLLM reasoning → Reinforcement learning for reasoning
Beyond Surface Imitation: Contrastive Modeling for Reasoning Path Alignment in Multimodal In-Context LearningLLM reasoning → Multimodal reasoning +1
From Retrieval to Weights: Parametric Individualization of Small Language Models with Individual Text CorporaRetrieval-augmented generation → RAG robustness and adaptive retrieval
CompassOPD: Cross-Family On-Policy Distillation via Within-Family Likelihood ShiftsKnowledge distillation → On-policy and self-distillation for LLMs
Agent-Based ML-LLM Fusion with Self-Optimizing Prompts for Plateau Weather AlertsLLM agents and tool use → Financial trading and forecasting agents
Data-Centric Post-Training for Financial Reasoning: Mining, Distillation, and Verifiable LearningLLM reasoning → Reasoning for agents, planning and code
RAP: Research Attention Prediction Reveals Target-Conditioned Evidence Acquisition BiasesLLM agents and tool use → Web, search, and research agents
Direct Diversity Optimization for Diverse Successful Trajectories in Preference Post-TrainingLLM agents and tool use → Reinforcement learning for LLM agents
Guaranteeing Faithful Evidence Extraction in Speculative Retrieval-Augmented GenerationRetrieval-augmented generation → RAG robustness and adaptive retrieval
Belief-State Engine: Augmenting LLMs for Principled Planning Under Partial ObservabilityLLM agents and tool use → General LLM agent frameworks and benchmarks
VLX-VR: An Agentic-Aware Video Reasoning ModelLLM reasoning → Multimodal reasoning +1
Optimal Value Inference for Reinforcement LearningDeep reinforcement learning methods → RL theory and sample complexity
Beyond Verified Answers: Solver-Informed Self-Distillation for Bootstrapping Operations Research Language ModelsSynthetic data and self-improvement for LLMs → Self-rewarding alignment and self-training +1
HaWMPO: Hallucination-Aware World Model-based Policy Optimization for Generalist Robot PolicyWorld models → Embodied and robot world models +2
Structural Process Supervision for Latent Chain-of-Thought ReasoningLLM reasoning → Prompting and chain-of-thought reasoning
Time-Frequency Geometric Cross-Attention for Chunked Vision-Language-Action ModelsVision-language-action models → Efficient VLA inference and adaptation
Interpreting Object-Dependent Concept Brittleness in Text-to-Image Diffusion ModelsDiffusion models → Image generation, text-to-image and editing
FlowCPO: A Unified Divergence View of Preference Alignment for Flow ModelsDiffusion models → Diffusion and flow policies for robotics +1
Contrastive Projection: Reading Transformer Internals by Differencing Logit LensesMechanistic interpretability → Circuits and mechanistic analysis
StreetDiff: Multi-view Street Scenes Generation via Cross-view Consistent Multi-view Stable Diffusion with Structure PromptsDiffusion models → Diffusion and flow policies for robotics
From Pixels to Hierarchical Sequences: Quadtree Mask Encoding for Vision-Language Binary Change DetectionMultimodal LLMs / vision-language models → Embodied and 3D vision-language models
AgentAudit: An Open, Extensible Framework for Full-Lifecycle Trust Evaluation of AI AgentsLLM agents and tool use → Agent threats, guardrails and governance
Keep Evaluation Fair: Detecting Data Leakage in Code Generation Benchmarks via Membership Inference AttacksLLM safety and security → Machine unlearning
Pretraining and Distillation Matter More Than Architecture Family for Label-Free Single-Cell ClassificationKnowledge distillation → Distillation for medical imaging
The Era by Eon Benchmark: A Generated Enterprise Estate with Exact Ground Truth for Benchmarking LLM AgentsLLM agents and tool use → General LLM agent frameworks and benchmarks
S3-Bench: Evaluating Speech Interaction Models as Scientific Voice AssistantsMultimodal LLMs / vision-language models → MLLM training and visual instruction tuning
Subgroup Membership Inference Audits of Differentially Private Synthetic TextLLM safety and security → Membership inference attacks
HyperTrace: Hypothesis-Based Preference Tracing for Online LLM PersonalizationLLM alignment / RLHF / preference optimisation → Human values and AI feedback alignment
CS-Guard: Benchmarking LLM Guardrails for Code Generation SecurityLLM safety and security → Jailbreaking
How Fragile Is Safety Alignment at Frontier Scale? A Single-Direction Attack on a 320B MoELLM alignment / RLHF / preference optimisation → Safety alignment and jailbreak attacks +2
LogiScope-VQA: Benchmarking Vision-Language Models for Logistics Hazard Identification in Industrial ScenariosLLM reasoning → Chain-of-thought reasoning for driving +1
ROAM: Robust Organization of Atomic Memories for Agents through Semantic RelationsLLM agents and tool use → Agent memory systems
Proof-Carrying Cognition: Closing the Verification Gap with Reality-Settled RewardLLM reasoning → Reinforcement learning for reasoning +1
XAgent: eXecution-guided Agentic AI for Effective Localization and Resolution of GitHub IssuesLLM agents and tool use → Software engineering coding agents
Fine-Tuning a KV Cache Concatenation-Aware Model or Recomputing KV Caches? Why Not Both?Retrieval-augmented generation → Generation and context integration +1
CARRE: Counterfactual Action Retrieval and Reason Evaluation for Explainable Churn PrescriptionLLM reasoning → Reasoning for agents, planning and code
SocialRL: Refining LLMs' Social Intelligence through Multi-turn Reinforcement Learning and Reward DesignLLM alignment / RLHF / preference optimisation → Reward modeling and RLHF
LexAgentHallu: A Hierarchical Benchmark for Profiling Hallucinations in Legal AgentsLLM agents and tool use → Medical and scientific domain agents
Epoch: Compiling Diffusion Blocks for Sparse MoE ServingMixture-of-experts → Vision mixture-of-experts +1
Which Tokens Should SFT Actually Learn? A Token-Trimming Perspective on Mathematical ReasoningLLM reasoning → Test-time compute and efficient thinking
CT-SAFR: Safe and Interpretable Chain-of-Thought Reasoning for Autonomous Robots: A Multi-Layered Verification Framework for Trustworthy AI-Driven Robotic Decision MakingLLM reasoning → Spatial and embodied reasoning
ALIGN-HOLD: Experience Alignment for Real-Time Hold Control in Large-Scale Ride-Hailing Matching at DiDiLLM alignment / RLHF / preference optimisation → Reward modeling and RLHF
SEA-SpeechBench: A Large-Scale Multitask Benchmark for Speech Understanding Across Southeast AsiaMultimodal LLMs / vision-language models → MLLM training and visual instruction tuning
PELM: Power Efficient On-Device LLM Inference with Speculative Decoding and Dynamic Voltage Frequency ScalingEfficient LLM inference and serving → Serving systems and request scheduling

All 402