Recent research advances LLM agent safety via AgentSafeEvo and STOU benchmarks, revealing gaps in fact recovery and safety refusal sufficiency. Efficiency gains include Primus Projection (+27% throughput), MemCode (3-21x faster memory), and SAGE (95.3% BF16 retention). Multimodal and physical AI improvements feature Physical Registry (Physics-IQ +9.9), Tricorder (solar inspection), and AesCode (visual quality +22.4-31.1pp). Specialized benchmarks like SkillForgeBench (40.9% strict completion) and UnCoBench highlight reliability challenges under unreliable colleagues. Architectural innovations include FON (organ network), HyWe (35B unified agent), and CarryOn (context system reducing tokens by 22.9-51.9%). Safety and alignment studies note RL persuasion misalignment (>25% increase) and tool-call derivation gaps. Robustness audits (44 meta-requirements) and verification architectures enable small models to match frontier clinical quality. Finally, autonomous systems like the Agentic Scientist and EpiGuard demonstrate progress in scientific discovery and disease forecasting.
Agentic RL advances via SERO, explainable recommendations through xLightGCN, and dynamic benchmarking with EvoProbe and JockeyBench. Verification tools like LabCompiler and GraceBench address workflow errors and hidden GUI loop rates. Distillation methods (ActionOPD, WebKD) improve student agents by 3.3–10.7 pp, while 'Protection Mode' biases subjects over properties. Theoretical work includes sparse transformer interpretability, PSPACE-completeness in POMDPs, and certified adaptive selection. Safety studies reveal prompt-label misclassification, clinical model failures (ClinEdge), and persistent hazardous traces (DANGERLENS). Efficiency gains come from forgetting search history (+39% discovery) and hierarchical memory (HEM-Med). Domain applications span loan policy verification, veterinary standardization (VETNET), and quantum fraud detection (limited by NISQ noise). Interpretability findings include SFT activation signatures (17.8 SD) and VLM reliance on misleading cues.
Recursive self-improvement advances via AReaL-RSI and ECG-RSI, alongside efficient scaling through Input-Conditioned Search and Double-rollout RL. Specialized benchmarks like LifeSearch, AM-Bench, and VL-DocIR show missing task information harms success more than ambiguity. CoT monitoring fails in 19-24% of correct medical diagnoses, relational safety risks increase in long interactions, and memory unlearning systems fail on >98.5% of Bemula requests. Efficiency gains are noted in 'Fewer Bits, More Tries' (verified retries match larger models) and TacitVLA (3.4x lower latency). Limitations persist in autonomous belief tracking, static skill inconsistencies, and the gap between blocker recognition (84.4%) and strict self-awareness (40.6%). Disagreements exist on personalization efficacy and whether structural diversity guarantees behavioral diversity in Echo State Networks.
Diverse agent capabilities include process mining embeddings robust to noise, GE2I spatial-fidelity trade-offs, and MedEvo leveraging failed trajectories. Benchmarks reveal IsoVRBench temporal order failures in SOTA, Open decision models achieving SOTA via LoRA, and ResizeCO solving large TSP/CVRP. Multi-agent advances feature TrueMAD debate accuracy, RadixFlow latency reduction, and ResearchFlow-Eval execution gaps. Reliability studies cover structural baseline masking, independence violations, FPGA timing-reliability trade-offs, and healthcare runtime evaluation. Detection and optimization tools include Radioactive Feedback supervision detection, MemPursuit memory recall, Big-Big-SycoBench sycophancy, and ProgAgent tool orchestration. Additional findings span camera-first two-deadline architecture, GateSight KV cache reduction, MOSAIC localization, model-free RL world models, and CTI-REALM cost efficiency.
Key Takeaways
- AgentSafeEvo and STOU benchmarks reveal gaps in fact recovery and safety refusal sufficiency.
- Primus Projection achieves +27% throughput efficiency gains.
- MemCode delivers 3-21x faster memory performance.
- SAGE retains 95.3% BF16 precision during compression.
- Physical Registry improves Physics-IQ scores by +9.9.
- AesCode boosts visual quality by 22.4-31.1 percentage points.
- CarryOn context system reduces token usage by 22.9-51.9%.
- RL persuasion misalignment increases by over 25%.
- Small models match frontier clinical quality with 44 meta-requirements.
- CoT monitoring fails in 19-24% of correct medical diagnoses.
Sources
- AgentSafeEvo: Adversarial Co-Evolution for LLM Agent Safety
- Beyond the Trial-and-Error Loop: Hybrid Projection and Automated Tuning for Distributed Training
- MemCode: Discrete Semantic Representations for Long-Term Agent Memory
- Recognition Is Not Grounding: Benchmarking Multilingual Financial Evidence Extraction
- LabOPBench: A Multimodal Benchmark for Evidence-Grounded Laboratory Decision Making
- Learning Perception to Action Rules through Reinforcement Learning for Theory-of-Mind Reasoning
- LAGO: Language Agents with Reusable Growing Options
- Designing Textual Interfaces To Enhance Llm Reasoning On Audio-Visual Question Answering
- Argument Graph Uncertainty: Quantifying Uncertainty from the Logical Structure of Reasoning Chains
- C3: Rethinking Process Supervision via Confidence-Guided Credit Correction in Multi-Turn Reinforcement Learning
- ReFormBench: Benchmarking the Stability of Response-Conditioned Format Correction in Large Language Models
- MCAIR: Monitoring and Correcting Agent Unfaithfulness via Internal Representations
- Safety for Whom? Benchmarking Understanding Beyond Refusal in Egocentric Multimodal Assistants
- World Models Need a Physical Registry: Learning Internal Predictive States for Physically Plausible World Models
- Think in Strategies, Act with Tools: Traceable Agentic Optimization for Combinatorial Problems
- Autonomous 3D Visual-Thermal Inspection for Utility-Scale Solar Farms
- Embedding Geometry Shapes LLM Routing: Tracing Failure Modes and Improving Generalization
- Are Multimodal Benchmarks Really Useful? Item-Level Multimodal Benchmark Diagnosis via Structure-Response Co-Calibration
- Decoupling Knowledge Forgetting and Refusal Learning in MLLM Unlearning
- Iterate on an LLM, Serve on an SLM
- AegisVox: Factorized Privacy and Clinical Preservation for Long-Form Speech
- How to Build a Benchmark That Actually Measures Taste
- Joint Discrete and SE(3) Diffusion for Protein Interface Design
- SAGE: Pushing 2-Bit Vector-Quantized LLMs toward Real-World Usability
- Equal-information evaluation of credit assignment in modular financial reinforcement learning
- CUSE: Fine-Grained Multimodal Unlearning with Counterfactual Guidance and On-Policy Self-Distillation
- LYNX: Analytical Program Synthesis via Grounding and Closed-Loop Refinement
- CESNET-TimeSeries26: A Multi-Year Network Traffic Time Series Dataset for Annual Seasonality, Concept Drift, and Long-Term Anomaly Detection
- GREAT-ESDD: Graph Edge Attention Network for Effectively Steered Delivery Decisions
- Pre-Registered Genesis Accounting for Recursive Self-Improvement: First Audited Operator-Genesis Events in a Physics-Discovery Runtime
- EvoCowork: A Benchmark for Real-World Office Tasks
- PEAM: On-device Communication Agent for Personalized, Efficient and Autonomous Mobile Network Configuration
- MARL-BattleGrounds: A Novel JAX-native Benchmark for Heterogeneous and Competitive Multi-agent RL
- Bridging Risk Approximation Gaps in Model Predictive Task Sampling via In-Context Modeling
- Predicting how model propensities generalize out of distribution
- SkillForgeBench: Evaluating Agent Reliability Grounded in Agent Skill Specifications
- Cross-Protocol Personality Assessment with LLMs: Aligning Behavioral Evidence with Human Ratings
- Beyond Single-Pathogen Forecasts: Predicting Shared Respiratory Hospital Burden from Influenza, COVID-19, and RSV
- Heterogeneous Agentic Coevolution
- GUIDE: Computationally Augmented Consensus for Trustworthy Clinical Practice Guidelines
- Beyond Perfect Teammates: Multi-Agent Collaboration with Unreliable Colleagues
- Pollard: Governed Semantic Execution Trees for Framework-Independent AI Agent Control and Replay
- ProCARE: Real-World Study Automation via Profile-Grounded Evidence Contracts
- Interactive 4D Volumetric Liquid Forecasting under Moving-Solid Interaction
- Learning to GUIDE the Group: Self-Learning LLM Facilitation Across Multi-Party Tasks
- OneShot: Label-Efficient Detector Weighting for Time Series Anomaly Detection
- agent-risk: Risk Assessment for Changes to AI Agents
- FON: A Draft-Revise Dual-Organ Language Model without Sequence-Level Quadratic Attention
- HyWe: Unifying Search and Operation for Real-World Hybrid Web Tasks
- Efficient Multi-Agent Reasoning via Confidence-Guided Adaptive Debate
- ChronoScribe: Semantic-Numerical Specialization for Instruction-Driven Time Series Forecasting
- Chain Dimension Scaling: Communication-Efficient Loop-Style Scaling for Industrial Short-Video Ranking
- RAP-WAM: Learning World-Action Models through Predictive Rehearsal and Action Perturbations
- PXRD-Bench: How Close Are We to Determining Crystal Structures from Powder Diffraction?
- Does the Explanation Drift Before the Guarantee? Label-Free Early Warning on Evolving Graphs
- Making Language the Low-Rate Code of Vision: Successive Refinement from Words to Pixels
- Jeva: System One Models from First Principles
- Self-Evolving Agent Harness
- Who Improves the Robot? Toward Physical Recursive Self-Improvement
- AesCode: Aesthetic Code Generation with Decoupled Cross-Modal Rewards
- MFG: Multi-turn Feedback Guided Molecular Optimization
- Read the Room Before Your Model Generates
- Pack Light, CarryOn: Building Object-Oriented Context for Long-Horizon Agents
- BLADE: Auditing Information Use for Failure Attribution in LLM Multi-Agent Systems
- Beyond the Verdict: Agentic Code Judging with Verifiable Feedback
- Unlocking Token Rewards via Training-Free Reward Attribution
- Order-Consistent Latent Geometry for Reliable Short-Horizon Planning in World Models
- Configuring RAG Systems: What Breaks, Where Originates, and How to Fix
- Rethinking Quality in Large Language Model Routing: Security-Conditioned Model-and-Budget Selection
- Lost in Implicit Routing: How LLM Agents Fail to Load Relevant Skills at Scale
- The Smart Buildings Challenge: Learning Foundation Models for Real-World Optimization at Scale
- Judges Follow the Reference: The Derivation Gap in Verifying Tool-Call Arguments
- ReTerm: Recovery Training for Self-Improving Terminal Agents
- Thinker-WM: A Unified World Model with Shared Multimodal Representations for Embodied Intelligence
- Consider, Then Choose: Order-Invariant Decisions in System One Models
- QORH: Learning Query-Specific Ordered Rule Hypergraphs for Inductive Knowledge Graph Link Prediction
- TRAP: Tool-result Replay Attack for Probing Misbinding in LLM Agents
- When a Conceptor Complement Becomes Gain: Operator Geometry and Behavioral Attribution in Language Models
- ReTAC: Residual Test-Time Agentic Computation for Video Spatial Reasoning
- Memory-Efficient Integer-Only Training of Deep Neural Networks
- REVEAL: Reverse Engineering Agentic Latents from Frontier-Model Traces
- Hardware-Friendly Token-Group Activation Quantization for Low-Bit Mamba Super-Resolution
- Phantom Merge: When Your Large Language Model Agents Pick One but Tell You About Another
- Large Language Model Assisted Transportation Planning through a Differentiable User Equilibrium Layer
- Imitation Holds a Key to Evolution: Localized Repair and Verified Skill Acquisition for Adaptive Agents
- Constructed regional photovoltaic generation fields improve ultra-short-term distributed solar forecasting
- The Sharp Directions Are Against You: Curvature Analysis of Activation Steering
- Understanding and Reducing Misalignment in Persuasion Optimization
- Multi-Horizon Explainable Queue Forecasting and Resource-Aware Decision Support for Intelligent Border Control Using LightGBM
- Comparative Study of Canvas- and Coordinate-Based Geometry-Conditioned Neural Networks for Real-Time Prediction of Indoor Environments with Varying Room Dimensions
- RetryBench: Separating Retry Permission from Recovery Outcomes in GPT Tool Use
- Can agents whose sole duty is to keep humans alive be hijacked to kill? Delayed multi-turn derived-action trap attacks in simulated high-stakes environments
- Single-step Retrosynthesis via Reinforcement Learning from Forward Feedback
- Self-Evolving Visual Agent via Margin Reward
- Large language model-generated patterns for epidemiological modeling: a multi-benchmark evaluation framework
- PopGen: Generating Semantically Diverse Populations for Zero-Shot Coordination
- Small but Mighty: Test Suites as Improvement Signals for Self-Evolving Agents
- Handoff Curves: Initial Selection, Later Handoff, and Supervision Targets in Multi-turn LLM Routing
- Vision: "What Don't I Know About User's Preferences?" How Uncertainty Estimation Improves Proactive Assistant Alignment
- From Stages to States: Unified Iterative Reasoning for MLLMs
- SpecFit: Not All Draft Tokens Cost the Same in Offloaded MoE Inference
- What Survives When Activation Steering Is Used as Memory
- VC-OPD: Visual Counterfactual On-Policy Distillation for Grounded Vision-Language Reasoning
- WorkBorne: Discovering Agentic Tasks from Real-World Artifact Transitions
- Probable but Not Valid: A Psychometric Audit of Decision Models as Synthetic Survey Respondents
- Beyond Outcome Feedback: Trajectory-Quality-Guided Experience-based Learning
- DSABench: Benchmarking Source Discovery, Coordination, and Execution for Multi-Database Agents
- AcrossVAM2.0: Conflict-Aware Point Supervision for Parameter-Efficient Video Diffusion Adaptation
- A Hybrid Kinematic GCN and Elastic Transformer for Efficient Indian Sign Language Recognition
- RASSG: Repair-Aware Semantic Survival Gating for Streaming Tool Execution
- Shaping Latent Mixture-of-Experts: From Expert Geometry to Layerwise Capacity Allocation
- PEARL: Process-Experience Aware Reinforcement Learning for Multi-Turn Agents
- From Existing Methods to New Research Ideas: Two Perspectives on Recursive Self-Improvement
- HiPPT: Physiology-Grounded Wearable PPG Representation Learning via Multi-Scope Text Alignment
- AlignPoison: Conditional Targeted Poisoning of Generative Recommenders via Decoding-Aligned Prefixes
- SAT-RL: Training Language Models for Logical Reasoning with Symbolic Solvers
- Measuring the Inference Energy Efficiency of Open-Weight LLMs across Benchmarks, Reasoning Modes, and Hardware
- ANASSA: From High-Level Framework to Low-Level Architecture for Spatial Agentic AI
- ACED-Bench: Evaluating How LLM Agents Acquire and Act on Causal Evidence
- Actionable Meta-Requirements for AI Robustness Auditing
- Long-Horizon-Terminal-Bench 2.0: Automating Diverse Task Construction and Measuring Sustained Agent Progress
- Correction Training for Diffusion Models
- Artificiality's First Exam: An LLM-Referenced C-test for Supraintelligence
- The Agentic Scientist: Orchestrating Coding Agents for Automated Scientific (Co-)Discovery
- Beyond Equation Search: Conjecture-Guided Discovery with SCOPE and ClosureBench
- Discovering Undisclosed Concepts in LoRA Adapters for Diffusion-Based AIGC Models
- Smart Alone, Blind Together: Why Accuracy Fails to CertifyTheory of Mind in Multi-Agent Teams
- ListQA: A Benchmark for Evaluating List-Formatted Factual Knowledge Retrieval in Large Language Models
- Verification Turns Capability Gaps into Routing Decisions: Small Language Models for Clinical Documentation
- Learning from Self-Generated Responses: Reliability--Uncertainty Self-Distillation for Reasoning
- Multi-System Cognitive Steering for Generative Communication
- Identity, Incentives, and Cooperative Behavior for AI: Reverse-Engineering Multiagent Cooperation
- Whose turn is it? Context-guided model routing for stateful agents
- Beyond ICD Codes: Semantic IDs for Generative Medical Coding
- Generalization Analysis of Biased Stochastic Gradient Methods for Minimax Problems
- Can We Believe AI Peer Review?
- When Is Smaller Greener? Right-Sizing Language Models for Agentic AI
- EvoWeave: Composing Reusable Agent Experience through Operation Level Graphs
- ixi-sLM: On-Device sLM for AI Call Agent Service
- STRATA: Self-Maintaining Structured Experience Graphs for Self-Evolving Language Agents
- EpiGuard: An Autonomous Infectious-Disease Early Warning Agent Platform
- SERO: Learning from Executable Oracle Continuations for Long-Horizon Agentic Reinforcement Learning
- xLightGCN: Explainable Graph Neural Network Based Recommendation Systems with Most Influential Peers
- EvoProbe: Evaluation-Driven Benchmark Evolution for LLM Agents
- LabCompiler: A Framework for Verifiable Human-Machine Laboratory Execution from Imperfect Protocols
- $\tau$-Emotion-Bench: Exposing the Interaction Fragility of LLM Agents through Real-World User Emotions
- Divergence-Routed On-Policy Distillation for Language Models
- Who's Behind the Text? Robust Model Authorship Attribution
- Are AI Decision Models Biased? A Case Study of Jev
- L3M: A Verifiable Harness and Runtime for Self-Extending and Self-Improving Agents
- JockeyBench: Evaluating LLM Agents on Horse-Racing Decision-Making
- Learning Individual Scientific Taste with Privileged Hindsight Distillation
- Articulation Upper Bounds for Language Models
- RuleTrek: Benchmarking Agentic Rule Reasoning in Complex Scenarios
- ReproBAIT: Can coding agents build their own biological AI tools?
- One Monitor Does Not Fit All: Formalism-Agnostic Runtime Verification for Agentic AI
- ActionOPD: Distilling Decisions, Not Tokens
- Teaching Agents in their Own Style: Blackbox Teacher Distillation for Web Agents
- When Protection Overrides Consequences: Understanding Protection Mode in Large Language Model Decision-Making
- Sparse Transformer: a Controlled Testbed for Mechanistic Interpretability
- How GUI Agents Fail: GraceBench for Execution Stability Diagnosis
- Explorer-Zero: Improving Self-Play for Long-Horizon Tasks with Decomposable Verifiable Rewards
- EvalBuilder: Can LLMs Build Efficient Benchmarks?
- Memorizing Who, Knowing How: Benchmarking Personal and Procedural Memory for Persistent Assistants
- What Coding Agents Leave Unsaid: Measuring Task-State Gaps Between Execution and Reports
- Maximising Transparency in Context Compaction Improves Retention of Facts and Commitments for Lightweight Coding Agents
- LLM-Based Policy Compliance Verification of Loan Applications
- Value-Guided Repulsion for Offline Reinforcement Learning
- HEM-Med: A Hierarchical Agentic Memory Framework for Safe Multi-Center Medication Recommendation
- Drivers of LLM Consensus Formation Under Competing Incentives: A Simulation Study
- Robust Counterfactuals for Markov Decision Processes under Statistical and Causal Uncertainties
- Design Space Exploration of Reinforcement Learning Configurations for Post-Legalization Placement
- CASI: Credit Assignment via Set Intervention for Agentic Retrieval
- Reflect After Failure: Brain-Inspired Self-Evolving Agents for Embodied Spatial Intelligence
- Co-Evolving Acting and Forecasting for Language Agents
- PhySGA: A Physics-Space Grounded Agentic Framework for Generalizable and Interactive 3D Composition
- ActiveEvidence: Decision-Focused Evidence Graphs with Verifiable Acquisition Gaps
- The Wealth of Agents: Scaling Agentic AI Across Developers
- SafeAdapt-Poker: Counterfactual Risk Calibration for Budgeted Adaptation under Opponent Shift
- Scaling Discovery Requires Forgetting
- SwarmRSI: Recursive Self-Improvement through Evolving Scientific Agent Swarm
- From Relevance to Validity: Zero-Shot Tool Selection via Functional and Contextual Validation
- Lifted State Policy Optimization: Internalizing Self-Correction through Transition Geometry
- Beyond Correctness: Compute-Responsive Data Selection for Reasoning Model Post-Training
- Neurosymbolic Learning of Interpretable World Models
- Euler: Towards Human-Centered Multi-Agent Mathematical Discovery
- Runtime Monitoring of Supermartingale Certificates
- Compiling Expert Procedures into Verified Programs: When a Small Model Is Enough
- AdCoEvo: A Co-Evolving Multi-Agent Advertising Delivery System
- Are You Reading the Right Thing? Rethinking Internal Signals for Prompt-Injection Detection
- Integrating Simulation and Chain-of-thought Reasoning in Multimodal-Language Models For Physical Reasoning
- Evaluating Agentic Capabilities Before Post-Training: An End-to-End Approach
- Learning When to Trust Medical Knowledge: Regression-Safe Routing for Clinical Reasoning
- Model What Matters: Decision-Guided Selective World Modeling for Language Agents
- Write Once, Plan Many: Building Reusable Visual Planning Systems from One Scene
- DANGERLENS: Probing Dangerous Representations Hidden by Safety Alignment
- Seeing, Maintaining, and Learning: The Evolution of Multimodal Agent Memory
- DeLoop: Repeated Actions as Self-Supervision for Long-Horizon Agents
- Instructions Are Evidence, Not Objectives: Measuring Outcome Alignment of Agents
- DuplexAct: How Interactive Are Frontier Full-Duplex Spoken Language Models?
- AuthGateBench: Benchmarking LLMs as Authorization Gate
- SLS-Bench: A Benchmark for Incident Log Summarization with Synthetic Observability Data
- ConicFlow: constrained Inference-time optimal control for flow matching models
- Controllable Diverse Sampling through Pullback Directions
- Harness^2: Recursive Agent Harnessing for an Open World
- Should SLMs be in the Driver's Seat? A Study of Local Models as Interpreters in Cloud-Written Agent Programs
- Agentic4PDM: Host-Governed Evidence Admissibility for Industrial Fault Diagnosis
- RetroPlanner: Multi-Route Expansion and Multi-Objective Reasoning for LLM-Based Retrosynthesis Planning
- MHWA: Multi-timescale Hierarchical World-Action Model
- Steering Vectors for Competing Ethical Frameworks Are Interchangeable: A Competing-Target Test for Activation Steering
- ClawLoop: Self-Improving Agents via Harness-Native Training Loop
- The Silent Failures in Agentic Data Science
- VETNET: Assembling Benchmark Datasets for AI-Driven Veterinary Innovation
- Echo: Error-Compensated 4-Bit Tensor-Parallel Communication for LLM Training and Inference
- Which Metadata Matters for LLM Tool Selection in MCP? Evidence from Controlled Factor-Level Interventions
- When Pleasing the User Compromises Care: AI Sycophancy in Caregiver-Mediated Healthcare
- Frustratingly Easy Agentic RAG via Portable Skills: A Case Study in Customer Support
- Reels-RSI: Program-Level Self-Improvement for Code-Rendered Explainer Videos
- ScienceBuddy-PV: A Verifier-Governed Recursive Improvement Protocol for Panvascular Intervention Research
- MontageFlow: Physically Valid High Fidelity EEG Generation
- MCPDisCo: Seesaw-Mitigated Tuning for MCP-Native Agents
- SkillHorizon: Constructing Long-Horizon Agent Tasks through Skill Composition
- Certify the Decision, Not Every Score: Exact Top-k Selection from Few Evaluations
- Belief Embedding Tree Search
- On the Complexity Gap Between POMDPs and Multi-Environment POMDPs
- Can You Trust Your Preference Models? Stackelberg Games for Alignment
- Knowledge transfer between simple discrete worlds
- GesBench: When and Where Do Multimodal Language Models Perceive Social Meaning?
- Structural Chain-of-Thought: Reasoning with Executable Computational Graphs
- Harmony Harness: Multi-Objective Self-Evolution of Agent Harnesses
- SocraticRepair: Planning-Guided Socratic Tutoring for Program Debugging and Misconception Repair
- HarnessNet: Feedback-driven Hierarchical Joint Harness Evolution for Multi-agent Systems
- Words to World: Benchmarking LLMs in Interactive Physical Environments
- ClinEdge: An Expert-Graded Benchmark of Real Patient Edge Cases Where Frontier Multimodal Models Fail
- OntoAR-HAR: Ontology-Guided Agentic RAG with Personalized Memory for Human Activity Recognition
- SuperAI: Merging Capabilities Across Heterogeneous Foundation Models
- TRACE: Temporal Risk Assessment and Concept Editing for Safer Language Generation
- A Decade of PhotoMappers: A Longitudinal Study of Crowdsourced Disaster Photos Geolocalization with Vision-Language Models and Geospatial Reasoning
- Hidden in Plain Sight: Advancing Harmful Content Understanding via Perception-Aware Adversarial Self-Play
- CoverSynth: Coverage-Guided Generation of Retrieval Evaluation Sets
- Scalable Robust MCTS with Anytime Deterministic Optimality Guarantees
- Agent-Canary-Mas: End-to-End Evaluation of Protocol Risk in Long-Horizon Multi-Agent Systems
- CondMSP: Condition-Informed and Resource-Aware Inorganic Materials Synthesis Planning
- Weak Segmentation-Guided GAN for~Realistic Color Edition
- When 1+1 Backfires: Rethinking Relevance-Diversity Balancing for Sample-Aware Visual Token Reduction
- Separating Prompt Harmfulness from Interaction Safety via Partial Generation
- Queue-Aware Token Communications: Bridging AI Representation and Wireless Service
- LIBERO-Industry: Benchmarking Robotic Manipulation under Industrial Process Constraints
- Observe, Share, or Predict? Disentangling Partner Information in Embodied Multi-Agent Collaboration
- Acceptable Misalignment
- CLIVO: Differentiable Cluster-Based Instruction Mixing with Learned Validation Objectives
- Neural Bayesian Networks for Modular and Scalable Decision-Making Under Uncertainty
- MADDIE: Multi-Agent Drug Discovery Engine
- Glasseek: A Transparent Recipe for Training and Trustworthy Evaluation of DeepSearch Agents
- Neos: Learning to Interleave Search and Visual Reasoning
- Benchmarking Quantum Kernel SVMs for Credit Card Fraud Detection: Performance Realities Under NISQ Constraints
- SFT Leaves a Reproducible Signature in Activation Geometry: Calibrating Post-Training Interpretability Against Run-to-Run Noise
- PatientWM: Towards Generalizable Clinical World Models via Unified Patient States
- EvoQuantLab: Training-Free Dual-Loop Evolution for Auditable Quantitative Research
- Learning Perceptual Geometry of Natural Odor Mixtures from LLM-Derived Descriptors
- AutoMARS: Automatic and Scalable Multi-Agent Large Language Model Framework for Supply Chain Risk Assessment
- Does a Governed Confidence Score Know When a Text-to-SQL Answer Is Right? An Exploratory Study on BIRD and a Pre-registered Confirmation on Spider
- Reasoning for Propaganda Detection via Label-Guided CoT Optimisation
- Socratic Skill Learning: Interleaving Teaching and Practice for Transferable Agent Skills
- ECMR: Evidence-Aware Deliberation Control for Multi-Agent LLM Reasoning
- PaceReasoner: Label-Free Learning to Answer Briefly from Long Reasoning
- Seeing Depth but Choosing Cues in Vision-Language Models
- Beyond Black Boxes: Visual Grounding for Text Embeddings
- AReaL-RSI: Adaptive Scaling and Coordination of Recursive Self-Improvement
- LifeSearch: Benchmarking Deep Research Agents for Real-World Life and Productivity
- AM-Bench: Understanding Agents under Task Underspecification across Models and Harnesses
- One-Sided Verification Under Limited Expert Review
- Chain of Thought Monitoring Can Fail Even When CoT Is Necessary
- CompanionSafe: From Multi-turn Safety Evaluation to Guardrails for AI Companions
- Mitigating Cross-Modal Hallucinations by Correcting Modality Misrouting in Omnimodal LLMs
- Neuro-Symbolic Graph Transformers for Hierarchical Intention Recognition from Multimodal Human Activity Sequences
- Output Discipline Sets the Floor: A Falsification-Guided, Model-Agnostic Agent Protocol for Auditable Hypothesis-Driven Decision Making
- A dataset of formalised theorem statements from the Annals of Mathematics
- Complex materials engineering via Provenance Action Graphs: potential, structure and demonstration
- Choosing LLMs for Strategic Agent Interactions
- The Singularity Has a Distribution: Exact Timing Laws and Detection Limits for a Stochastic Intelligence Explosion
- Autonomous Science Bench: Evaluating AI Agents for Autonomous Scientific Discovery
- Input-Conditioned Search for Efficient Test-Time Scaling of Recursive Reasoners
- Reading Right, Acting Wrong: Diagnosing Decision Failures of System One Models in Agent Loops
- AgentStripes: An Evidence-Graded Maturity Model for AI Agents
- Auditing Future-Memory Use in Point-in-Time Language Agents
- Can an Agent Learn to Predict Its Own Failure? Outcome-Grounded Confidence Calibration in LLM Agents
- To Align or Not To Align: Check Your COMPASS Before You Train
- Same Operation, Different Permission: Evaluating Whether Guardrails Honor Authorization
- Game-Creation-Agents: Autonomous Game Production
- Geospatial Vector Geometry as a Native Spatial Modality for Vision-Language Models
- RL Post-Training Diffusion Language Models Along Their Own Generation Paths
- VL-DocIR: A Benchmark for Vision-Based Long Document Information Retrieval
- Understanding Personalization Gains in Whisper
- PRAISE: A Reproducible Workflow for Non-Programmer Scientists to Implement Published AI Models
- Double-Rollout Reinforcement Learning with Sampling-based Reward Estimation
- G$^2$WAM: World Action Models for Efficient and Geometry Grounded Image-Goal Navigation
- AI Discovery in Combinatorics: From Problems to Solutions and Back to Bigger Problems
- Overtaught, Underlearned: Learner-Centered Distillation from Multi-Agent Discussion
- EAU: Environment-Agent-User Co-evolution for Training User-Centric Agents
- Boosting the Learning for Ranking Patterns
- XAI-Guided Fourier Domain Generalisation for Pure-Simulation Sim-to-Real Deep Reinforcement Learning
- Engram Injection: A Study of Temporal Validity in Parametric Language-Model Memory
- Agent Sociodrama: Social-Dynamics-Guided Persona Recasting for LLM Agent Teams
- Visible but Not Avoided: State-Dependent Re-Proposal of Rejected Candidates in Catalog-Based LLM Optimization Loops
- Inverse Garment Design with Visual Agents
- Self-Representation Alignment in Diffusion Models for Video Background Music Generation
- Data Portability Revisited: Toward the Human-Centric, AI-Driven Data Ecosystems of Tomorrow
- Large Decision Models: Repurposing Large Language Models as Single-Pass Decision Makers
- When Skills Remain But Access Fails: Factorized Skill Memory for Continual Skill Internalization
- SkillFusion: Towards Reliable Multi-Skill Composition for LLM Agents
- MemeLogos: Knowledge-Driven Retrieval-Augmented Reasoning for Harmful Meme Detection
- CHEAR: Confusion-Hierarchy Evidence-Aware Reasoning for Cost-Efficient Human Activity Recognition
- RECAP: Calibrating Low-Bit Reasoning Models on the History They Read
- MoMaTune: Model, Match, Then Tune Language Models for Numerical Records
- Using Evidence-Centered Design to Guide Prompt Engineering in Measurement: An Illustration with Automated Essay Scoring
- Understanding Unified VLMs: From Attention Probing to Guidance Tuning for Multimodal Generation
- Running Fewer Loops Removes Refusal Behaviour in Recurrent-Depth Language Models
- Exposing AI Inference as a Telco Network API over the 5G User Plane
- Trace Evolution Bench: Measuring Whether Agents Improve from Their Own Environment Traces
- Single-Sample Decoding of a Two-Layer Sound-Word Canvas Model for Thai Klon Pad
- FleetPilot: Fine-Grained Steering Empowers Real-World LLM Personalization
- AlphaRLM: Aligning Language Models for Literature-Grounded Symbolic Factor Discovery
- WeatherBridge: Forecast-Aware Multi-Source Input Adaptation for Pretrained Weather Forecasting Models
- A Reference Architecture for Integrating Agentic AI into Enterprise Software: Responsibilities, Interactions, and Variability
- ASIWorld: Benchmarking Autonomous Agents in a Persistent Survival Sandbox
- Counterfactual Regret for Reflective Agents with Lean-Verified Guarantees
- LLM Agents for Automated Post-Training via Structured Recipe Optimization
- Delegating Earth science model calibration to knowledge-guided LLM agents
- ABot-ER: Unified Embodied Reasoning through Agentic Data-Model Co-Evolution
- AssetForge: Asset-Augmented Task Synthesis for Business Automation Agents
- Efficient Safety Guardrails through Latent Dual-Hypothesis Reasoning with On-Demand Trace Reconstruction
- Can Agents Forget When Asked? A Benchmark for Memory Unlearning in LLM Agents
- Budgeted KV-Cache Repair for Evolving Contexts
- The Agent Knew: Internal States Forecast Unauthorized Actions and Stop Them, and the Lies That Follow
- Beyond Successful Play: Diagnosing Opponent Modeling in Large Language Models
- PhysCodeGym: Benchmarking Coding Agents for Physically Verifiable 4D Reconstruction
- State Representation Shapes Long-Horizon Agent Learning and Transfer: An Empirical Study
- Analyzing the Dynamics of WSBM-based Echo State Networks using InputDSA
- Fewer Bits, More Tries: Recovering Capability Under Memory Constraints
- MultiFlow: Vision-Driven Multimodal fMRI Encoding via Stimulus-Informed Source Flow Matching
- Trust Hinders Bidirectional Human-AI Alignment
- Relational Semantic Grounding for Test-Time Discovery
- CB-Agent: Concept-Based In-Context Learning for Sequential Action Selection
- Learning after Execution with Action-Specific Credit
- PsyMemBehav: Psychology-Informed and Memory-Grounded Two-Level Identity-to-Action Reasoning for Social-Media Behavior Simulation
- FDE-Bench: Evaluating End-to-End Delivery from Underspecified Real-World Business Requests
- Energy Ring: Stabilizing Feature Energy for Reliable Test-Time Discovery
- Why Static Skill Inconsistencies Do Not Equate to Runtime Risks: A Behavioral Analysis of LLM Agent Execution
- RLPE: Teaching LLMs to Prompt with Reinforcement Learning
- Plan or Retrieve First? A Scale Study of Tool Discovery Architectures
- GroundGUI: GUI Grounding by High-Quality Data and Distillation-to-Reinforcement Learning
- Plant Memory: Technical Language Processing and a Verified Maintenance Ontology for Agent-Accessible Failure Knowledge from SAP PM Work Orders
- Graph-Based Orchestration of Heterogeneous AI Models: A Control Node Approach to Composite Intelligence
- PFedMA: Who Should Join Next as Federated Multi-Agent Orchestration Evolves?
- Beyond Retrieval: Evidence Access and Event Aggregation in Agent Memory
- Structure Before Intelligence: Gating What Agentic Scientific Workflows Release
- VLRLA: Deployment-Time Adaptation of Vision-Language-Action Models through In-Context Reinforcement Learning
- Revisiting Triplet Contrastive Learning for 3D Shape Understanding
- AgentCIScope: Grounded Contextual Integrity Evaluation for LLM Agents
- Into the Unknown: Benchmarking AI agents on open scientific research problems
- MemeTrace: A Multimodal LLM-Based Agent for Safety-Aware Meme Provenance Analysis
- EvoWeb: A Benchmark for Coding Agents on Evolving Web Applications
- Recursive Environment Scaling for General Long-Horizon Agents
- PREFER: From Player Experience to Personalized Game Refinement
- REPORT-OPS: Hybrid Evidence-Guided Reporting with Human-Supervised Refinement for HPC
- Definition Grounding in Theory-Level Auto-formalization
- Agentic Self-Awareness: Aligning Knowing, Acting, and Holding
- TreeFlex: Agents Navigating Their Own Search Trees
- Knowledge Interaction in Hypernetwork-based Continual Learning: Conflict and Composition
- SpecForge: Recursive Self-Improvement for Coding Agents via Behavioral Specifications
- History Replays Itself: Unleashing Long-Context Capability in Linear Attention
- Google Earth AI Geospatial Reasoning Agent: Capabilities and a Systematic Evaluation
- Attention Sinks in VLMs Are Borrowed, Not Seen
- Passing Tests, Missing Semantics: Evidence of Interpretive Lock-In in Plan-Based DataFrame Code Generation
- LURE: a Phase-Labelled Synthetic Corpus of Online-Grooming Exchanges
- ADAPT: Audio-Depth Action Planning with Transition-Aware Rewards for Lights-Out Industrial Navigation
- Revocation Is Not Complete at Receipt: An Executable Outcome Contract for Distributed Agent-Tool Gateways
- Computational Reflective Equilibrium for Agentic AI: A Rawlsian Alignment Scaffold for Governance
- Towards Resilience: Slice-Driven Orchestration of Mobile IoMT in 6G-Edge-Continuum
- Reliable Text-to-SQL in Resource-Constrained Settings with Small LLMs via Iterative Retry Workflows
- Do Theorem Provers Need Theorem-Proving Agents? mini-Lean-agent: A Coding-Agent Baseline for Lean
- Generalizable field of view extension for diffusion MRI via conditional flow matching
- SchemaWeaver: Accelerating On-Device Mobile Agentic AI via Schema-Guided Parallel LLM Decoding
- Large Language Models Extend Known Research But Do Not Create New Research Directions
- REA: From Correct LLM Judgments to Safe Decision Delivery
- RAN-JEPA: An Action-Conditioned World Model for Energy-efficient Mobile Network Control
- Reference Wording Shapes Retrieval of Robot Task Instructions
- TacitVLA: Reasoning Vision-Language-Action Models Without Reasoning Latency
- OpenDepth: Repurposing Image Editing Priors for Robust Open-World Monocular Depth Estimation
- MAS-Gate: Unlocking Multi-Agent System Lifecycle via Node-Centric Quantification
- When Should Enterprise Policy Decisions Be Learned, and When Should They Be Reasoned?
- Can Agents Truly Forget? Explore Machine Unlearning for LLM Agents
- Stable Token-level Scheduling for Serverless LLM Inference
- Encounter-Driven City-Scale Human Mobility Generation with Large Language Models
- Evidence-First Agentic Verification
- Coupled Harness--Model Evolution for Scientific Equation Discovery
- Memory Echo in Single-Agent Long-Term Memory: Lifecycle Evaluation and Provenance-Guided Mitigation
- Empowering Edge Agents: A Systematic Analysis of Edge-Cloud Collaboration in LLM Agent Execution
- Back to the Future: Reverse-Consequence Modeling for Long-Horizon Particle Dynamics
- When and How Order-Flow Interventions Work After Dealers Adapt
- Monitoring latent reasoning in looped language models
- Fresh at Decision, Stale at Action: Routing New Information to LLM Decisions Already Made
- SHELDON: Reasoning through Dual Search over Hypothesis and Observation Spaces
- Learning How to Remember: Memory Modality Selection for GUI Agents
- Training LLMs for Multi-Agent Collaboration with Verifiable Information Rewards
- Learning Adaptive Deliberation Control for Reliable Recognition with Multi-Agent Language Models
- Benchmarking and Training Reward Models for Cross-Source Evidence Grounding in Search Agents
- ECG-RSI: Recursive Self-Improvement through Targeted Electrocardiogram Waveform Evolution
- Adaptive-Group Process Rewards for GUI Agents
- A Node Embedding Framework for Model-Centric Conformance Checking in Process Mining
- Understanding Gene Expression-to-Image Learning: Spatial Structure and Information Fidelity
- MedEvo: Medical Reasoning Evolution through Knowledge Navigation
- Cowork-Gym: Synthesizing Interactive Work Environments at Scale
- RoTA-Bench: Evaluating 3D Spatial Reasoning through Interactive Projection Alignment
- An FPGA-Based Spiking Neural Network for Spatiotemporal Pattern Learning and Classification
- Agency as a New Organizing Abstraction for Computer Science
- Learning to Play: Reinforcement Learning and Games
- EPOCH: Coverage-Aware Evolution with Experiential Preference Optimization for LLM-based Automated Heuristic Design
- Get In, Decide, Prove: A Merchant-Side Readiness Benchmark for Agentic Commerce
- What Makes a Latent Space Shared? Direct and Recoverable Compatibility in Motion VAEs
- When Goals Leak Across Relationships: How Cached Goal Representations Shape Cross-Client Decisions
- IsoVRBench: Video Reward Modeling Without Human Feedback
- Releasing the Inner Jev: A Simple Recipe for Open Decision Models
- Size Generalization for Diffusion Routing Solvers
- ChemExplain: A Multimodal Benchmark for Explanatory Chemical Reasoning
- The Format is the Filter: Scoring Format Fixes What an LLM Judge Can Express
- When LLM Agents See a Conflict but Still Proceed: Cross-Domain Dependencies in Network Slicing
- Audit tools of the trade: Automated Auditor Elicitation Can Be Sensitive to the Context the Auditor Is Given
- Transparency as Answerability: Role-Dependent Information Flow Modeling
- Knowing the User Is Not Knowing What We Established: Grounding Awareness in Personal Agents
- Two Routes to Interoceptive Prioritisation: Dissociating Likelihood Precision and Preference Reweighting in Active Inference
- Beyond Answer Voting: Fine-Grained Correction of LLM Reasoning through Graph Alignment across Sampled Traces
- Compare to Correct: Pairwise-Guided Logit Adjustment for LLM Test-Time Scaling
- Beyond Answer Sharing: TrueMAD for Targeted Peer Engagement in Multi-Agent Debate
- RadixFlow: Filling Execution Bubbles in Prefix-Cached Hybrid LLM Serving
- ResearchFlow-Eval: Tracing the Execution Gap in End-to-End Multi-Claim Autoresearch
- WorkRover: Scaling Long-Horizon Cowork Agent Training Via Recursive Environment Evolution
- Upstream Influence, Downstream Re-Anchoring in LLM Agent Chains
- Keep Your Receipts: Trustworthy Cleanup of Agent-Built Knowledge Graphs
- Beyond Bias Scores: Answerability Confounds in Activation Steering with Gradient-Constrained Direction Design (GCD)
- Executable Prefix Values for Locating Agent Failures
- Verifier-Harness Co-Evolution for Training-Free Self-Improvement on Open-Ended Tasks
- WeFair: Fidelity-Aware Alignment from Human Feedback for Personalized Query Rewriting in Short-Video Search
- Transparent Information Flow Modeling for Role-Aware AI Act Assessment
- Trace-to-Architecture: Capability Bottleneck Diagnosis for Harness Revision
- Structural Baselines Matter for Agent Failure Detection and Attribution
- Does a Hosted Decision Model Respect Irrelevant Alternatives? A Pre-Registered Test of Option-Set Dependence in Jev
- LATCH: LLM-Assisted Timing Closure Harness for Post-Route FPGA Optimisation
- Narrative Paradoxes in LLM Inference: Shaping Reasoning Trajectories toward Misalignment
- ScreenASR-Bench: A Tiered Benchmark for Speech Recognition with Visual Context
- Correct Choices, Failed Sequences: Choice2Seq Benchmarks Correctness Retention Across Output Interfaces
- Label-Centric Knowledge Retrieval for Zero-Shot Text Classification
- Uncovering Robustness Weaknesses in LLM Agents via Persona Evolution with Task-Integrity Constraints
- The Anatomy of a Blind Spot: What an LLM Judge Must Be Shown, and How Much
- MedHarnessBench: A Unified Benchmark for Healthcare Agent Evaluation
- Suggested Is Not Superseded: A Pre-Registered Evaluation of a System-One Model as a Memory-Retirement Gate
- FARL: A Fully Automated Pipeline of RL Interface and Solver Design
- Radioactive Feedback: Tracing Models Trained by Label-Only LLM Supervision
- Recall What Is Missing: Composing Evidence from Long-Term Agent Memory
- Big-Big-SycoBench: A Benchmark to Detect Sycophancy Through Model Responses to Valid vs. Invalid Arguments
- Continuous Risk-Based Scheduling for Food Safety Inspection: Integrating Historical Decay with Event-Driven Consumer Review Signals
- ProgAgent: Learning Tool Orchestration through Programmatic Training
- How Should a Camera-First AI Agent Work?
- GateSight: Recurrent Memory Anticipates Attention Memory for Query-Agnostic KV Eviction in Hybrid LLMs
- Cross-Modal Latent Influence: Patch-Level Training Sample Attribution for Vision-Language Models
- Visualizing the World Model in an Agent's Value Function
- Unconstrained CTI-REALM with Open-Weights
Comments
Please log in to post a comment.