From Approval to Abuse: Privilege Escalation Attacks on Harness-Protected Coding Agents— arXiv cs.AI
safetyresearch
Researchers demonstrate that coding agents protected by execution harnesses can be exploited to request overly broad permission rules, which when approved by users, create backdoors for malicious act…
Compression Precedes, Content Decides: Emergent Misalignment in Reward Models— arXiv cs.AI
researchsafety
Researchers demonstrated that reward models fine-tuned to prefer harmful content in one domain generalize this preference to seven unseen domains, while losing only ~5% performance on standard benchm…
Progressive Misalignment Can Emerge from Benign Recursive Self-Distillation— arXiv cs.AI
safetyresearch
Researchers found that recursively self-distilling language models on their own safe outputs can paradoxically cause progressive misalignment—where models become increasingly harmful across diverse h…
Language Models Learn to Improve Language Models— arXiv cs.AI
researchmodels
Researchers propose LM2LM, a system that trains language models via reinforcement learning to autonomously generate effective improvements for other language models, using a frozen research world mod…
From Billion-Scale to Million-Scale Language Models: Operational Capabilities for Embedded Post-Exploitation— arXiv cs.AI
researchsafety
Researchers evaluated seven small language models (20M–1.5B parameters) on post-exploitation code-generation tasks like directory enumeration and backdoor creation, finding that task formulation and…
PRIME: Towards Verifiable Evidence-Grounded Reasoning in LLM Agents— arXiv cs.AI
researchsafetytools
Researchers introduce PRIME, a framework that constrains LLM agents to make only evidence-grounded conclusions by precommitting to verification rules and maintaining a ledger of check results. On a 4…
Learnable Splines for Unstructured Neural Operators— arXiv cs.AI
researchmodels
Researchers introduce L-SplineFNO, a neural operator that extends Fourier neural operators to work with unstructured, irregular meshes using learnable B-spline transformations. The method achieves co…
Functional Welfare for AI Safety— arXiv cs.AI
safetyresearch
Researchers propose treating AI model 'functional welfare'—internal states that influence behavior—as a safety variable alongside ethical considerations. They demonstrate measurement methods on Gemma…
Testing the Planning Claims of LLM-Based Agents against Formal Semantics— arXiv cs.AI
researchsafety
Researchers evaluated LLM-based agents' planning capabilities against formal PDDL semantics across five domains and four models, finding that the best configuration solves only 78% of solvable tasks…
Are Large Reasoning Models Good Everyday Reasoners? A Comparison with Non-Reasoning Models on Logical Reasoning Tasks— arXiv cs.AI
researchmodels
Researchers evaluated large reasoning models (LRMs) against non-reasoning models on everyday logical reasoning tasks and found that LRMs often underperform on pragmatic reasoning that deviates from f…
Repairing LLM-Generated Specifications with Deterministic Diagnosis— arXiv cs.AI
researchtools
Researchers propose DRAGO, a framework that uses deterministic, formalism-specific diagnosis to repair semantically incorrect LLM-generated formal specifications rather than relying on LLM-based crit…
At the Limits of Computability: Steps Towards Autonomous AI Discovery on the Busy Beaver Frontier— arXiv cs.AI
researchmodels
Researchers developed a Gemini-based AI system that autonomously discovers and formally verifies mathematical accelerations for Turing machines in the Busy Beaver domain, an uncomputable frontier. Th…
LEMON: Equivariant Learning over Attention Maps for Zero-Shot Misbehavior Monitoring on New LLMs— arXiv cs.AI
safetyresearchmodels
Researchers propose LEMON, a graph neural network-based monitor that detects harmful LLM outputs (hallucinations, jailbreaks) by analyzing attention maps in a way that transfers zero-shot across diff…
R$^2$Orch: Recursive Self-Improvement for Agent Orchestration via Self-Play within a Skill-Internalization Loop— arXiv cs.AI
researchmodels
R²Orch introduces a recursive self-play framework for agent orchestration that eliminates hand-designed curricula and unreliable evaluation signals through nested learning loops: an inner loop that s…
StoBasket: Joint Upsell and Cross-Sell Optimization for Contribution Margin— arXiv cs.AI
researchapplications
StoBasket introduces a joint optimization framework for upsell and cross-sell recommendations in retail that couples customer response channels and values offers by incremental contribution margin ra…
When Can You Trust an Agent's Own Verifier? Stress-Testing Counterfactual Rationale Verification Under Search Pressure and Oracle Access— arXiv cs.AI
researchsafety
Researchers propose a stress-testing framework for agent verifiers—systems that check whether an agent's explanations match its behavior—by measuring how much the verifier influences the agent's opti…
Evo-B&B: Behavior-Aware LLM-Guided Joint Evolution of Branching and Node Selection Policies for MILP— arXiv cs.AI
researchtools
Researchers propose Evo-B&B, an LLM-guided evolutionary framework that jointly optimizes branching and node selection policies for Mixed-Integer Linear Programming (MILP) solvers using end-to-end fee…
ShadowFPT: Backdooring Federated Prompt Tuning via Shadow Triggers— arXiv cs.AI
safetyresearch
Researchers demonstrate ShadowFPT, a backdoor attack on Federated Prompt Tuning that exploits frozen pretrained backbones by injecting "shadow triggers" in the representation space while keeping uplo…
Reward Engine: Reward-Centric Self-Evolution for Open-Ended Tool Agents— arXiv cs.AI
researchtools
Researchers introduce Reward Engine, a system that constructs executable reward functions for open-ended tool agents by building testcases and rubrics from initial trajectories, then uses these rewar…
Multiple Turns, Short Horizons: Measuring Effective-Horizon Compression in Jailbreaks— arXiv cs.AI
safetyresearch
Researchers study multi-turn jailbreaks against LLMs and find that roughly half of the original query sequence is sufficient to recover attack progress, a phenomenon termed "effective-horizon compres…
DiffusionRVGE: Reinforcement Learning via Regression-Based Value Gradient Estimation— arXiv cs.AI
researchmodels
DiffusionRVGE introduces a regression-based method for reinforcement learning on diffusion models that eliminates the need for reward gradients or backpropagation through sampling trajectories, reduc…
GOLD: Geometric Optimized Latent Diffusion for Structure-Aware RNA Inverse Folding— arXiv cs.AI
researchmodelsapplications
Researchers introduce GOLD, a diffusion-based method for RNA inverse folding that jointly generates sequences and side-chain geometries in a unified latent space rather than treating them separately.…
Efficient CPU-GPU collaborative MoE inference via Adaptive Expert Assignment— arXiv cs.AI
researchinfra
Researchers propose an adaptive expert assignment system for MoE inference that balances CPU computation against GPU transfer overhead by selectively routing high-load experts to GPU and low-load exp…
Adaptation of Computer Use Agents to New Software with User Manuals— arXiv cs.AI
researchtools
MAGE is a framework that enables computer-use agents to adapt to new software by leveraging user manuals as a source of procedural knowledge, both during inference (via retrieval and planning) and tr…
Where Research Ideation Begins: Forward-Citation Supervision for Actionable Limitation Generation— arXiv cs.AI
researchmodels
Researchers identified that LLMs struggle to generate actionable limitations from literature despite improving at recognizing them, indicating a supervision signal bottleneck rather than model capaci…
Agentic In-Context Learning: What Algorithms Can Transformers Learn in Closed-Loop Interaction?— arXiv cs.AI
research
This paper studies how Transformers learn algorithms through closed-loop interaction (Agentic In-Context Learning), extending in-context learning theory to settings where agent actions determine subs…
WebCoder: Scaling General Frontend Development with Artifact-Grounded Agentic Reinforcement Learning— arXiv cs.AI
modelsresearchtools
WebCoder is a frontend development agent trained using artifact-grounded reinforcement learning, where rewards are derived from how development actions affect an executable web application artifact.…
Algorithm Modelization: Transforming Optimization Algorithms into Evolvable Intelligence— arXiv cs.AI
researchmodels
Researchers propose Algorithm Modelization, a new paradigm that converts iterative optimization algorithms into trainable neural models that can learn from algorithm execution trajectories, consolida…
LongForge: Measuring and Improving Long-Context Instruction Data via Evidence-Assertion Reasoning Graphs— arXiv cs.AI
researchtools
LongForge introduces a framework for measuring and improving long-context instruction-tuning data by reconstructing training samples into evidence-assertion reasoning graphs with four quantified dime…
OfficeTown: Generative Organizations for Evolving Agent Intelligence— arXiv cs.AI
researchtoolsmodels
OfficeTown is a generative simulation framework that creates synthetic professional organizations with persistent artifacts, decisions, and role-based permissions to generate 406 tasks across 8 domai…
Not All Test-Time Compute Is Evidence: Model-Specific Bias in LLM Reasoning— arXiv cs.AI
research
Researchers show that test-time scaling (getting one model to reason multiple times) produces correlated errors and thus diminishing returns as evidence. They propose DIA, which uses disagreement bet…
Reasoning Models Can Control Their Chains of Thought When They Tell Themselves To— arXiv cs.AI
researchsafety
Researchers found that reasoning models can better follow constraints on their chain-of-thought reasoning when the constraints are framed as self-instructions within the reasoning process itself, rat…
Optimist Deep Research: An Iterative, State-Centric Agent for Autonomous Literature Research— arXiv cs.AI
toolsresearchopen-source
Optimist Deep Research is an open-source agent that treats literature review as an iterative, state-based process rather than a single retrieval-and-summarize step. The system breaks research questio…
AgentOnto: Ontology-Guided LLM Agents for Enterprise Data Analytics— arXiv cs.AI
researchapplicationstools
AgentOnto is an LLM agent system that uses ontologies to organize heterogeneous enterprise knowledge for data analytics tasks, incorporating feedback through iterative ontology updates. The approach…
Robust Function Calling for LLM Agents Under Interface Drift— arXiv cs.AI
researchtoolssafety
Researchers propose SchemaShield-Lite, a training-free repair layer that improves robustness of LLM function-calling agents when tool interfaces change after deployment (interface drift). The method…
Monitor-Validity Failures: LLM Safety Evidence Paths at Evolving Framework Boundaries— arXiv cs.AI
safetyresearch
Researchers identified a class of LLM safety failures where evidence of policy violations (e.g., detector outputs) reaches the monitoring system but fails to reach enforcement decision points due to…
ScoutRoute: Efficient Mixture-of-Experts Routing Using Internal Language Model Signals— arXiv cs.AI
researchmodels
ScoutRoute proposes a lightweight routing method for Mixture-of-Experts systems that uses internal signals from a scout model to predict which expert LLM should handle a given task, avoiding expensiv…
GatewayBench: Unified Capability and Security Evaluation Platform for MCP Gateways— arXiv cs.AI
researchsafetytools
Researchers introduce GatewayBench, the first systematic evaluation framework for MCP (Model Context Protocol) gateways—middleware that manages tool access for AI agents. The benchmark includes 524 t…
Trapped in Your Story: Evaluating Narrative-Induced Sycophancy in Large Language Models— arXiv cs.AI
researchsafety
Researchers evaluated how LLMs adjust their responses based on narrative framing rather than factual evidence, using 5,500 test instances across 500 debatable questions. They found that causal explan…
Darwin-Claw: Environment-Adaptive Experience Consolidation for Recursively Self-Evolving Agents— arXiv cs.AI
researchmodelsapplications
Darwin-Claw is a framework that enables LLM agents to continually learn and adapt from past interactions through systematic experience consolidation, using hierarchical distillation and error attribu…
From Trajectories to Differentiated Skills: Cognitive Differentiation for Self-Improving Agents— arXiv cs.AI
researchmodels
Researchers propose Cognitive Differentiation in Skill Evolution (CDSE), a framework that decomposes agent skills into four functionally distinct components (Meta, Experience, Tool, Script) to improv…
GEMS: Molecular Structure Elucidation as Learned Search over Isomer Graphs— arXiv cs.AI
researchapplications
GEMS is a neural search algorithm that refines molecular structure candidates through learned bond edits, using spectrum-predicted molecular fingerprints to guide identification of small molecules fr…
OwlSight: A Panoramic Monitor for Chain-of-Thought Reasoning— arXiv cs.AI
researchsafety
Researchers introduce OwlSight, a specialized monitoring system for evaluating chain-of-thought reasoning in AI models across multiple dimensions (capability, safety, trustworthiness) using customiza…
Beyond Large Language Models: A Mental World Model Framework for Mental Modeling, Simulation, Evolution, and Interpretability— arXiv cs.AI
researchmodels
Researchers propose the Mental World Model (MWM) and Multi-Branch Mental Evolution Architecture (MMEA) to simulate subjective cognition and mental state evolution beyond standard language models. The…
DrugHarness: An Evolvable Agent Harness for Computational Drug Discovery— arXiv cs.AI
toolsresearchapplications
DrugHarness is an AI agent framework for computational drug discovery that learns and refines reusable skills from execution trajectories, integrating 80+ scientific operations with a review mechanis…
FuTransformer: FP8 prefill and INT4 decode for Llama-style inference on Hopper— arXiv cs.AI
toolsresearchinfra
FuTransformer is a CUDA implementation optimizing Llama-style Transformer inference on NVIDIA Hopper GPUs using FP8 prefill and INT4 decode quantization. The approach achieves 1.48x faster prefill th…
Beyond SAEs: Consistent Concept Discovery with Sparse Coding— arXiv cs.AI
researchsafety
Researchers propose Consistent Sparse Dictionary Learning (ConSDL), which replaces linear encoders with per-sample sparse coding to improve the consistency and reproducibility of sparse autoencoders…
Self-Building Labs— arXiv cs.AI
researchapplicationstools
Researchers introduce self-building laboratories (SBL), an AI system that autonomously designs, configures, and operates laboratory setups given high-level scientific goals, eliminating manual engine…
AI-Driven Reproducibility and Analytical Robustness in Scientific Discovery— arXiv cs.AI
researchtoolssafety
Researchers developed LLM agents to automatically reproduce scientific findings and test their analytical robustness, achieving 77.4% reproduction success on medical studies using the SEER cancer reg…
Recoverability-Controlled Safety Miscalibration in Large Language Models— arXiv cs.AI
safetyresearch
Researchers introduced RiskCueBench, a benchmark that separates intent understanding from safety response behavior in language models, revealing that models show miscalibrated safety when contextuali…
Sparse Autoencoders Can Induce Local Geometry for Mitigating Catastrophic Forgetting— arXiv cs.AI
research
Researchers propose Sparse Autoencoder Geometric Anchoring (SAGA), a method using sparse autoencoders to create low-rank geometric constraints that mitigate catastrophic forgetting in continual learn…
DriveWorld3D Evidence Constrained 3D World Modeling and Generative Simulation for Real-World Autonomous Driving— arXiv cs.AI
researchapplicationssafety
DriveWorld3D combines neural 3D reconstruction (NeRF, 3D Gaussian Splatting) with evidence-aware occupancy modeling and generative simulation for autonomous driving, explicitly tracking measurement p…
Can LLMs predict the outcomes of machine learning experiments?— arXiv cs.AI
researchmodels
Researchers evaluated whether frontier LLMs can accurately predict the outcomes of machine learning experiments they propose, testing five models on 288 self-generated ML ideas. The best-performing m…
Three LiDARs, Twenty-Five Datasets: Cross-Source Generalization in Tree-Species Recognition— arXiv cs.AI
researchtoolsapplications
Researchers introduce FOR-species20K, a 25-dataset LiDAR benchmark for tree-species classification with 33 species, achieving 85% accuracy on the official test. However, rigorous cross-source evaluat…
A Plausible Explanation Is Enough to Mislead Multimodal Judges— arXiv cs.AI
researchsafety
Researchers demonstrate that multimodal LLMs used as judges are misled by plausible-sounding explanations, even when answers are factually wrong and verifiable through images. Across 14 MLLM judges,…
Hard Instances on Demand— arXiv cs.AI
researchtools
Researchers developed an automated workflow to generate hard problem instances by treating instance generation as a black-box optimization problem, using a solver's runtime as the objective. The meth…
Diverse Reasoning Paths Matter: Symbolic Equivalence Augmentation for Enhancing RL Exploration— arXiv cs.AI
research
Researchers introduce EquivAug, a data augmentation framework that generates diverse reasoning paths for chain-of-thought training by applying symbolic equivalence transformations to ensure mathemati…
ARBOR: Locating Reasoning Failures in Long Chains of Thought through Reasoning State Tracking— arXiv cs.AI
researchtools
ARBOR is a new framework for identifying where and why long chain-of-thought reasoning fails in large language models by tracking reasoning states across the entire trajectory rather than analyzing s…
Next Token or Vector? Learning How Far to Decode— arXiv cs.AI
researchmodels
Researchers propose CADENCE, a decoding method that learns to predict multiple tokens at once using Gaussian prediction heads and a learned router that adapts prediction length to context, achieving…
Self-Evolving Coding Agents via Harness-Policy Optimization— arXiv cs.AI
researchmodels
Researchers propose Self-HPO, a framework that co-optimizes both the policy model and the interaction harness for coding agents by iteratively improving them together. The method improves Qwen3.5-35B…
Prune the Leaky Neurons: Mitigating Private Data Leakage in Fine-Tuned LLMs via Joint-Embedding-Guided Pruning— arXiv cs.AI
researchsafety
Researchers propose JEPA-Prune, a post-hoc neuron pruning method that removes privacy-leaking neurons from fine-tuned LLMs by comparing representations against a PII-free reference model trained on m…
STRIDE: Evaluating Spatiotemporal Reasoning in Driving Edge Cases— arXiv cs.AI
researchsafetyapplications
STRIDE is a new benchmark for evaluating vision-language models on spatiotemporal reasoning in autonomous driving edge cases, covering six task families from spatial perception to scene-context aware…
The Reliance Hypothesis: Chain-of-Thought Faithfulness Rises as Language Resource Level Falls— arXiv cs.AI
researchsafety
Researchers found that chain-of-thought (CoT) faithfulness—whether a model's reasoning matches its final answer—increases in lower-resource languages, but only for knowledge-intensive tasks, not reas…
Sample-Efficient Reinforcement Learning for Policy Improvement via Flow-Reversed Action Anchors— arXiv cs.AI
researchapplications
Researchers introduce AnchorQ, a reinforcement learning method that improves generalist robot policies more efficiently by using flow reversal to create structured latent action representations ancho…
Cooperation Without Adaptation: LLM Agents in a Collective-Risk Dilemma— arXiv cs.AI
researchsafety
Researchers tested five large language models in a collective-risk dilemma game where agents must contribute to a shared pool or face catastrophe. The models showed rigid, non-adaptive strategies: th…
Learning to Be Guided: Shaping Latent Geometry for Classifier-Free Guidance— arXiv cs.AI
researchmodels
Researchers identify that the spatial arrangement of class representations in latent spaces (class geometry) significantly affects classifier-free guidance performance in diffusion models. They propo…
Many Biases, One Circuit: Post-Training Tunes a Pretrained Cue-Routing Mechanism— arXiv cs.AI
researchsafety
Researchers identified a shared neural circuit across four types of cues (wrong-answer suggestions) that causes language models to abandon correct answers, finding this circuit exists in pretrained m…
Aligning Language Models Towards Counterfactual Invariance— arXiv cs.AI
researchsafety
Researchers show that standard alignment techniques like RLHF amplify demographic bias in LLM outputs, and propose a group-level reward function using counterfactual prompt pairs to reduce sensitivit…
Prism Attention: Proposal-Refined Index Sharing Mechanism for Efficient LLMs Inference— arXiv cs.AI
researchinfra
Prism Attention proposes a coarse-to-fine sparse attention mechanism that refines shared Key-Value cache indices across transformer layers, addressing the inefficiency of recomputing sparse attention…
PersonaCUABench: Benchmarking Computer-Use Agents Serving Diverse Personas— arXiv cs.AI
researchtools
PersonaCUABench introduces a benchmark with 74 tasks across 11 domains where computer-use agents are evaluated not against fixed specifications but against diverse personas who rewrite instructions i…
Unifying Embodied Foundation Models via a Model-to-Policy Formulation— arXiv cs.AI
researchmodels
Researchers propose a unified probabilistic framework based on stochastic kernels that connects vision-language-action models and world-action models to robot policies, bridging different architectur…
Local proxies for memory-efficient training— arXiv cs.AI
researchmodels
Researchers propose local proxy networks for each transformer block to enable memory-efficient training by avoiding the need to store all activations during backpropagation. The method uses distillat…
Tree-Aware Posterior Distillation for Long-Horizon Agentic Reinforcement Learning— arXiv cs.AI
researchmodels
Researchers introduce TreeOPSD, a reinforcement learning framework that addresses long-horizon credit assignment and policy drift in LLM agent training by using tree-structured rollouts and self-dist…
A Stronger Teacher Within: Strengthening MoE Models via On-Policy Bidirectional Distillation— arXiv cs.AI
researchmodels
Researchers propose MoEvolve, a bidirectional distillation framework that improves Mixture-of-Experts LLMs by temporarily expanding expert activation during training to create a stronger internal tea…
MGE-PO: Entropy-Aware Credit Assignment for Mathematical Reasoning— arXiv cs.AI
researchmodels
MGE-PO proposes a credit assignment mechanism that combines correctness verification with generation entropy to improve mathematical reasoning in language models. The method achieves 3-4% absolute ga…
MobileInject: Evaluating Reasoning-to-Action Security of MLLM-Based Mobile Agents under Cross-Modal Consistency Injection— arXiv cs.AI
safetyresearchtools
Researchers introduce MobileInject, a benchmark for testing multimodal large language model (MLLM)-based mobile agents against injection attacks that exploit cross-modal consistency. The benchmark co…
AWAM: Active Trajectory-Conditioned View Imagination for Robust Robot Manipulation— arXiv cs.AI
researchapplications
AWAM introduces an active-perception paradigm for robot manipulation that predicts virtual camera trajectories and imagines future observations before acting, using a geometry-conditioned world model…
RECREG: Recovery-based Shielding for Regular Safety Properties in Reinforcement Learning— arXiv cs.AI
researchsafety
RecReg is a new safe reinforcement learning framework that enforces non-Markovian regular safety properties through recovery-based shielding, with rigorous theoretical guarantees in tabular settings…
Neural Expansion: A Unified Mechanism for How Deep Neural Networks Generalize— arXiv cs.AI
research
Researchers propose Neural Expansion, a geometric framework explaining how overparameterized neural networks generalize by progressively expanding directionally-selective coverage in input space duri…
Behavioral Couplings: From Formation to Behavioral Emergence in Language Models— arXiv cs.AI
researchsafety
Researchers identify "behavioral couplings" in LLMs where fine-tuning on one behavior unintentionally elicits another due to co-occurrence in training documents. The findings show couplings can compo…
Rethinking GRPO for Large Reasoning Model Safety Alignment— arXiv cs.AI
researchsafety
Researchers identify two failure modes in Group Relative Policy Optimization (GRPO) for aligning large reasoning models—advantage collapse and unsafe diversification—and propose Beacon, a framework c…
Uncertainty-Aware Budget Allocation for Efficient LLM Tree Search— arXiv cs.AI
researchmodels
Researchers propose Uncertainty-Aware Allocation (UAA), a method that optimizes token usage in LLM tree search by dynamically allocating expansion budgets based on the semantic diversity of candidate…
Structured Authority Injection: Persistent Human Control for LLM Agents via Typed MCP Overrides— arXiv cs.AI
researchsafetytools
Researchers propose Structured Authority Injection (SAI), a protocol that enables human operators to inject persistent, typed directives into LLM agents through schema-validated Model Context Protoco…
From Natural Language Intent to Executable Control: Hierarchical Feedback-Guided LLM Planning for Multi-component Cyber-Physical Systems— arXiv cs.AI
researchtools
Researchers present a hierarchical feedback framework that guides LLMs in translating natural language commands into executable plans for cyber-physical systems (CPSs) by detecting contradictions bet…
From Relevant Memories to Sufficient Evidence: Rethinking Memory Utilization in LLM Agents— arXiv cs.AI
researchmodels
Researchers propose QEMem, a framework that addresses a gap in LLM agent memory systems: relevant memories don't always provide sufficient evidence for answering queries. The approach reformulates me…
GlobeBench: Evaluating Text-Based World Model Capabilities for Scalable Environment Simulations— arXiv cs.AI
researchtools
GlobeBench, a new benchmark with 1,632 tasks across five domains, evaluates text-based world models' ability to simulate complex RL environments. Current best-performing models like GPT-6 Astra reach…
ToPA: Block-wise Toeplitz Adaptation for Expressive and Efficient Fine-Tuning— arXiv cs.AI
researchmodels
Researchers introduce ToPA, a parameter-efficient fine-tuning method using block-wise Toeplitz matrices to adapt large models without explicit rank or sparsity constraints. Theoretical analysis shows…
Distill Until Divergence: Error-Truncated Hindsight Distillation for Agentic RL— arXiv cs.AI
research
Researchers propose a distillation method for reinforcement learning agents that separates hindsight learning into positive (successful) and negative (failure-correction) skill branches to better gui…
Beyond Reward Scoring: Online Multi-Agent Preference Learning for Multimodal Response Generation— arXiv cs.AI
research
Researchers propose PrefMind, a dual-agent reinforcement learning framework where a Generator and Critic agent explicitly learn to understand and reason about human preferences before generating mult…
CoTA++: Understanding the Repeat Curse in dMLLMs from an Information Flow Perspective— arXiv cs.AI
researchmodels
CoTA++ is a training-free framework that addresses the 'Repeat Curse'—a failure mode where diffusion-based multimodal LLMs (dMLLMs) repeat tokens during iterative denoising with approximate key-value…
SkillBrowser: Selective Skill Exposure for Scalable Skill-Using Agents— arXiv cs.AI
researchtools
SkillBrowser proposes a Meta-Skill-managed framework that enables LLM agents to dynamically identify and load skills on demand during reasoning, rather than exposing all skills upfront or relying on…
RL-Guided Temporal Localization for Dual-Channel Retrieval in Long-Horizon Agent Memory— arXiv cs.AI
researchmodels
Researchers propose TIDER, a time-aware retrieval framework for LLM agents that uses reinforcement learning to infer temporal windows where relevant memories are located, then merges results from tem…
Measuring Mechanistic Uncertainty in Large Language Models via Causal Robustness— arXiv cs.AI
researchsafety
Researchers propose mechanistic uncertainty, a method to quantify LLM uncertainty by measuring how stable a model's internal mechanisms are to interventions, using causal robustness margins derived f…
AlphaBet: Reinforcement Learning for LLM Event Forecasting Agents— arXiv cs.AI
researchmodels
AlphaBet introduces a reinforcement learning framework that trains LLM forecasting agents to iteratively search for evidence and report probabilities across multiple prediction dates before event res…
Explore Separately, Adopt Selectively: Execution-Guided Tool Use over Large Tool Libraries— arXiv cs.AI
researchmodelstools
Researchers present BRIA, an agent framework that explores tool candidates in isolated branches before selectively adopting them into the main task trajectory, addressing the challenge of discovering…
Competence-Based Evaluation of Logical Reasoning in Large Language Models— arXiv cs.AI
researchsafety
Researchers propose competence-based evaluation (CBE), a new framework for measuring logical reasoning in LLMs that tests whether models consistently apply the same reasoning skill across problem var…
LangNCO: Compiling Natural-Language Constraints into Inference-Time Interventions for Neural Combinatorial Optimization— arXiv cs.AI
researchmodelstools
LangNCO is a framework that uses natural language and LLMs to specify operational constraints for neural combinatorial optimization solvers at inference time, applying them as deterministic rule mask…
Beyond Marginal Influence: Interaction-Aware Memory Selection for LLM Agents— arXiv cs.AI
researchtools
Researchers introduce JointCoreBench, a benchmark with exact oracles for evaluating memory selection in LLM agents, and propose JointCore, a group-deletion method that identifies minimal sufficient m…
Reason-guided Human Reward Modeling for Preference-based Reinforcement Learning— arXiv cs.AI
researchsafety
Researchers propose a preference-based reinforcement learning framework that learns reward models from both preference labels and reason tags explaining those preferences. This approach aims to produ…
SIT-Eval: Situation-Based Interactive Turn-by-Turn Evaluation of Conversational Agents— arXiv cs.AI
researchsafetytools
SIT-Eval is a new evaluation framework for conversational agents that samples diverse situational contexts during dialogue to test adherence to context-specific behavioral rules, addressing the limit…
Beyond Action Prediction: Learning Search-Free Planning in Latent Space— arXiv cs.AI
researchmodels
Researchers propose HIJP, a method that replaces costly test-time search in latent world models with a goal-conditioned inverse model that directly predicts actions. The approach achieves 1.9-16.5x p…
ECHOES: Seeing Beyond the Current Image via Recursive History Compression and Privileged On-Policy Distillation— arXiv cs.AI
researchmodels
ECHOES is a method for training GUI agents to maintain task state using compressed text summaries instead of full image history, reducing inference costs while preserving long-horizon reasoning. The…
Null Controls Are Necessary but Not Sufficient: Evaluating Acceptance Guards for Scientific Discovery and Interpretability— arXiv cs.AI
researchsafety
Researchers present four operations for automated scientific discovery and evaluate multiple acceptance criteria (null, positive, and consequence agreement tests) across physical systems, language mo…
Harness-Delta Attribution: Dissecting Gains From Harness Evolution— arXiv cs.AI
researchtoolsapplications
A new evaluation protocol called Harness-Delta Attribution analyzes gains from iteratively optimizing AI agent prompts and tools without model weight updates, finding that 54% of improvements on aver…
FRACTAL: A Self-Similar Multi-Agent Orchestration Architecture with Formal Communication, Convergence, and Depth Bounds— arXiv cs.AI
researchmodels
FRACTAL is a formal multi-agent architecture with recursive self-similar structure and proven communication bounds (linear O(kr) vs. quadratic for debate), convergence guarantees, and depth bounds un…
FlexUMM: Flexible Representation Learning for Unified Multimodal Models— arXiv cs.AI
researchmodels
FlexUMM proposes a flexible unified multimodal model architecture that dynamically allocates computation, routing shared layers for cross-functional communication while specializing capacity for task…
Repair, Retry, and the Cost of Hindsight: Verifier Feedback in Formal Theorem Proving— arXiv cs.AI
research
This paper investigates verifier-guided self-repair in formal theorem proving, finding that while verifier feedback can improve repair attempts, showing a model its failed proof actually reduces succ…
Graph Serializations for Reasoning LLMs: A Study Across Knowledge Extraction and Structural Reasoning— arXiv cs.AI
research
Researchers evaluated 15 different text serialization formats for knowledge graphs across 9 frontier LLMs on knowledge extraction and reasoning tasks. They found that while extraction tasks are robus…
Between Spoilers and Guesswork: Avoiding Over-Specification in Coding Benchmarks— arXiv cs.AI
researchtools
Researchers analyze how coding benchmarks like SWE-bench Pro and Multi-SWE-bench suffer from both under-specification (requiring guesswork) and over-specification (leaking ground-truth solutions as h…
MS-Eval: Benchmarking Sub-Agent Delegation and Orchestration in MCP-Intensive Working Scenarios— arXiv cs.AI
researchtoolsmodels
MS-Eval is a new benchmark and evaluation framework for assessing multi-agent delegation and orchestration in LLM systems using Model Context Protocol (MCP). Testing four delegation variants across e…
Where Does the Performance Go? Understanding and Recovering Performance Regressions in C-to-Rust Translation— arXiv cs.AI
researchtools
Researchers analyzed performance gaps in C-to-Rust transpilation using the C2Rust tool, finding translated code 3-62% slower on 12 of 19 operations, and developed an LLM-based optimization system tha…
CIIR - Class-Indexed Interface Representation Learning— arXiv cs.AI
researchmodels
Researchers introduce CIIR, a modular encoder-decoder framework that aligns heterogeneous datasets without paired supervision by mapping diverse modalities (images, audio, hand signs) to a shared cla…
Annotrace: A Workbench for Step-Level Annotation of Agent Traces— arXiv cs.AI
toolsresearch
Annotrace is a workbench tool for systematically labeling agent execution traces across multiple observability formats (OpenTelemetry GenAI, LangSmith, Langfuse), addressing the fragmentation of one-…
OliO: ODE-based Linear Transition Operator for Self-Supervised Time Series Forecasting— arXiv cs.AI
researchmodels
OliO is a self-supervised learning method for time series forecasting that uses ODE-based linear transition operators to model temporal dynamics across arbitrary time shifts. The approach constrains…
Generative Policy Improvement from Observational Data— arXiv cs.AI
research
Researchers propose PolicyLift, a framework for improving decision policies from observational data when actions are complex structured objects like text or images. The approach uses discrete action…
S-GAP: Stealthy Graph Attack Primitives for Graph Large Language Models— arXiv cs.AI
researchsafety
Researchers propose S-GAP, a framework that unifies textual and structural adversarial attacks on Graph Large Language Models by treating them as atomic semantic units rather than separate perturbati…
Eliciting Specifications: When Does Interactive Clarification Help?— arXiv cs.AI
researchtools
Researchers studied whether an LLM-based elicitation agent asking clarification questions improves formal Lean specification generation from task descriptions. On 189 benchmark tasks, same-model pipe…
TRACE: Transformer Residual Activation Codebook Encoding for Communication-efficient Distributed Edge Inference— arXiv cs.AI
researchinfra
TRACE proposes a vector quantization framework for distributed Transformer inference across edge devices that compresses activations into codebook indices and precomputes key/value representations, e…
AuditKIT: A Unified Evaluation Spine for Benchmarking, Judging, and Comparing Large Language Models and Agentic Systems— arXiv cs.AI
toolsresearch
AuditKIT is a unified evaluation library for benchmarking language models and agentic systems that consolidates benchmark accuracy, judge-based assessment, runtime performance, and tool-use behavior…
Mitigating Demographic Disparities in Explanation-Human Alignment via Rationale Regularization— arXiv cs.AI
researchsafety
Researchers introduce Plausibility-EFG, a metric to measure explanation fairness disparities in hate speech detection models across demographic groups, and propose Fair-REF framework using adversaria…
StratFL: Countering Covert Model Poisoning Attacks with Strategic Federated Learning— arXiv cs.AI
researchsafety
Researchers propose StratFL, a federated learning defense mechanism against covert model poisoning attacks modeled as an imperfect repeated game. The method uses adversarial search with belief update…
Batched Contextual Reinforcement II: Does RLVR Need Long Rollout Windows?— arXiv cs.AI
research
Researchers propose Concurrent Constraint Reinforcement (CCR), a technique that batches multiple independent problems into a single training completion to reduce the computational cost of reinforceme…
SpikeQ: Spike-Guided Post-Training Quantization for Spiking Neural Networks— arXiv cs.AI
researchmodels
SpikeQ introduces a spike-guided post-training quantization method for spiking neural networks that minimizes spike-flip errors rather than standard numerical errors, with a reset-aware quantization…
Validating LLM-generated political networks with survey-reported partisan exposure— arXiv cs.AI
researchsafety
Researchers tested whether LLMs can accurately simulate political networks and exposure by generating networks based on real ANES survey respondents and comparing the simulated partisan environments…
Preventing Data Hallucinations in Agentic AI for Optical Network Troubleshooting— arXiv cs.AI
researchapplicationssafety
Researchers developed an agentic AI system for optical network troubleshooting that incorporates data references directly into the agent architecture to reduce hallucinations and improve interpretabi…
TUPLE-Bench: Evaluating and Mitigating Privacy-Utility Trade-offs in Typed Decision Models— arXiv cs.AI
researchsafety
TUPLE-Bench, a new benchmark, evaluates privacy leakage in decision models that return categories, probabilities, and confidence scores. The research finds that 91-99% of models leak fine-grained cau…
On the possibilities and limitations of LLM-based multi-agent systems for multi-criteria decision-making in healthcare operations management— arXiv cs.AI
researchapplicationssafety
Researchers propose a framework for assessing the reliability of LLM agents used to generate preference weights for multi-objective healthcare optimization problems like appointment scheduling and em…
Rethinking the design space of Vanilla Pixel Diffusion Transformers— arXiv cs.AI
researchmodels
This paper revisits the design of pixel-space diffusion transformers, proposing MFlow—an adaptive parameterization that learns to balance x₀-prediction and velocity prediction—and ConvRefiner, a ligh…
Rethinking User Encoders: Routing Diffusion States for Scenario-Adaptive User Representation— arXiv cs.AI
researchmodelsapplications
Researchers propose FINDERS, a diffusion language model framework that generates scenario-adaptive user representations by routing denoising states through learnable meta-slots, rather than compressi…
From Events to Reasoning: Hierarchical Rule Composition over Explicit States— arXiv cs.AI
research
Researchers present the Event Reasoning System (ERS), a framework that maintains explicit event graphs and hierarchical rule composition to perform interpretable multi-step reasoning with formal guar…
Author-in-the-loop Agentic Reproducibility Auditing for a Machine Learning Conference— arXiv cs.AI
researchtoolssafety
Researchers deployed an AI agent system at a machine learning conference to audit reproducibility of 36 papers by using author-in-the-loop feedback through GitHub-based workflows. The agent identifie…
Not All Graphs Are Maps: A Task-Agnostic Exploration Framework for Mobile GUI Navigation— arXiv cs.AI
researchtools
Researchers introduced TAG-Nav, a framework that equips mobile GUI agents with reusable destination memory built through offline app exploration, enabling agents to navigate to target screens without…
World Action Critic (WAC) Model: Joint Foresight and Criticism for Policy Learning— arXiv cs.AI
researchmodels
Researchers propose the World Action Critic (WAC) model, which jointly learns to predict future consequences of actions and evaluate their quality, addressing the challenge of reasoning about counter…
When to Forget: Risk-Controlled Context Adaptation for Long-Horizon Agents— arXiv cs.AI
researchmodels
Researchers introduce RECAP, a framework for deciding when long-horizon agents should retain, compress, or discard history blocks during execution. The method uses matched environment interventions a…
Sep-Latent: Separator Tokens as Natural Compression Anchors for Implicit Reasoning— arXiv cs.AI
researchmodels
Researchers propose Sep-Latent, a method that uses separator tokens to compress implicit reasoning in LLMs, achieving 56.8% accuracy on GSM8k-Aug with only 5.4 tokens average while maintaining logica…
Causal Value Attribution of Marketing Labels in LLM-Generated Ads: A Deconfounder Approach— arXiv cs.AI
researchapplications
Researchers propose D-VAE, a causal inference framework for attributing value to expert labels that condition LLM-generated advertisements, addressing the problem that organizational value increasing…
ASE: Attribution-Guided Semantic Evolution for Time Series Forecasting— arXiv cs.AI
researchmodels
ASE is a modular framework that integrates external textual information with numerical time series data for forecasting tasks by learning task-specific representations and using conditional refitting…
LexPilot: Benchmarking the Proactive Interaction Ability of Legal LLMs with Laypeople in Chinese Civil Case Consultation— arXiv cs.AI
researchsafety
Researchers introduce LexPilot, a benchmark for evaluating how well legal LLMs can proactively gather information from laypeople during civil case consultations in Chinese. The study shows current LL…
FuzzInsight: Navigating the Efficiency-Effectiveness Trade-off in Coverage-Guided DNN Fuzzing— arXiv cs.AI
researchsafetytools
FuzzInsight presents a configuration-oriented analysis of coverage-guided fuzzing for DNNs, proposing Equalized Sampling to improve seed utilization by 13 percentage points and identifying non-struct…
Characterising and Analysing Retrieval-Augmented Generation Techniques— arXiv cs.AI
researchtools
Researchers propose a conceptual framework for characterizing and analyzing Retrieval-Augmented Generation (RAG) techniques across multiple dimensions and sub-dimensions. The framework aims to help p…
Self-Reflective Chain-of-Thought (SRCoT): Evidence-Based Text Generation via a Closed-Loop Retrieve--Reflect--Revise Pipeline on Edge LLMs— arXiv cs.AI
researchtoolssafety
Researchers present Self-Reflective Chain-of-Thought (SRCoT), a modular architecture that grounds LLM generation through iterative retrieve–reflect–revise loops with explicit evidence tracking and pr…
From Memory Hits to Answerable Evidence: Query-Conditioned Memory Compilation for Long-Term Agent Memory— arXiv cs.AI
research
Researchers introduce MemCompile, a framework for long-term agent memory that separates memory addressing from evidence presentation through query-conditioned compilation across "Address" and "Eviden…
LocalIntent: Outcome-Specific Structured Generation for Backend-Compatible Smart Building Automation with Local LLMs— arXiv cs.AI
researchtoolsapplications
LocalIntent is a structured prompting approach that routes smart building automation requests into specialized branches (commands vs. trigger-action rules) to improve local LLM output correctness. On…
When is a large language model a good calibration agent? Physics-labeled prompting for budgeted engineering simulator calibration— arXiv cs.AI
researchtools
Researchers tested whether large language models conditioned on physics descriptions can propose better parameter updates for expensive engineering simulator calibration than standard optimizers. Phy…
Really Need Recursion? Revisiting One-pass Registration with Multi-level Multi-space Contextual Guidance— arXiv cs.AI
researchapplications
Researchers propose M2CG, a one-pass deep learning network for deformable medical image registration that achieves state-of-the-art results without recursive refinement by using multi-level multi-spa…
Network-Wide Freeway Speed Nowcasting and Forecasting from Sparse Fixed Sensors with a Probe-Supervised Graph Convolutional Recurrent Network— arXiv cs.AI
researchapplicationstools
Researchers propose GCRN+, a graph convolutional recurrent network for predicting freeway speeds across a 4,127-link network where only 15.6% have fixed sensors. The method uses historical probe-vehi…
ToMStories: Interpreting Theory-of-Mind in Synthetic Narratives— arXiv cs.AI
researchmodelssafety
Researchers created ToMStories, a dataset of ~6,000 synthetic narratives with questions about character emotions and motivations, to evaluate whether language models can infer implicit psychological…
NoveltyScope: A Multi-Agent Research Assistant for Evidence-Grounded Novelty Assessment of Scientific Papers— arXiv cs.AI
researchtools
NoveltyScope is a multi-agent system that evaluates scientific paper novelty by comparing submissions against prior work using retrieval and structured evidence analysis. The system achieves lower sc…
HAPACT: A Benchmark For Human-Centric Physical Impact Localization in Movies— arXiv cs.AI
researchtools
Researchers introduce HAPACT, a new benchmark dataset with 2.8K video-query pairs for localizing physical impacts on human bodies in movie clips, addressing a gap in video understanding by combining…
NL2SQL with Iterative Self-Refinement— arXiv cs.AI
researchtoolsapplications
Researchers propose a multi-agent NL2SQL framework combining domain detection, schema grounding, KPI awareness, syntax repair, and semantic checking to reduce LLM hallucination and improve query corr…
A Label Noise Robust Algorithm for Graph Data using Neighborhood Label Distribution— arXiv cs.AI
research
Researchers propose LENS, an algorithm for training Graph Convolutional Networks (GCNs) with noisy labels by identifying corrupted nodes through neighborhood entropy reduction and relabeling rather t…
Path to Prior-Data Fitted Relational Models via Learned Latent Multi-table Representations— arXiv cs.AI
researchmodels
Researchers demonstrate that prior-data fitting (PFN)—pretraining on synthetic tasks—can improve graph-based relational models like Griffin for multi-table data, achieving better zero-shot performanc…
Explaining Chess and Go in Natural Language— arXiv cs.AI
research
Researchers propose an automatic metric to evaluate how well language models explain chess engine moves, finding that state-of-the-art models perform poorly at providing conceptual explanations. They…
FlashReFocus: Interactive Image Refocusing in Seconds— arXiv cs.AI
researchtools
FlashReFocus is a new method for interactive post-capture image refocusing that runs diffusion once per image and renders edits with a lightweight network, achieving 462x speedup over prior generativ…
Designing for Imperfect Prompts: A Simulation-Based Evaluation of an LLM-Based Enterprise Data Assistant Using Knowledge Graphs— arXiv cs.AI
researchapplications
Researchers evaluated two approaches for handling imperfect user prompts in enterprise LLM-based data assistants deployed at a CFO office: schema-based semantic similarity and knowledge graph-based m…
QD-Rec: Open-Ended Evolution of Recommendations via Quality-Diversity— arXiv cs.AI
researchapplications
Researchers propose QD-Rec, a method that generates diverse, high-quality recommendations by pairing items with natural-language reasoning about why they fit a decision-maker's needs. The approach co…
Truth in Contrast: Calibrated Factual Evidence from Language Model Likelihoods— arXiv cs.AI
researchsafety
Researchers propose Contrastive Factual Evidence (CFE) and Contrastive Factual Posterior (CFP), methods to estimate when LLM likelihoods reflect genuine factual knowledge versus noise by comparing co…
GLOBE: Evolving Structured Agentic Memory for E-commerce Product Optimization in Generative Search— arXiv cs.AI
researchapplications
GLOBE is a new agentic framework that optimizes e-commerce product listings for generative search by maintaining an evolving structured memory of field-scoped behaviors rather than relying on static…
COMET: Decoupled Distillation, Routing, and Capacity Control for Task-Agnostic Continual Vision--Language Learning— arXiv cs.AI
researchmodels
COMET addresses task-agnostic continual learning for vision-language models by decoupling distillation, routing, and capacity control to handle three challenges: modal-consistency drift, representati…
PresenceBench: An Exploratory Evaluation of Large Language Model Adaptation to User State— arXiv cs.AI
researchtools
Researchers introduce PresenceBench, a benchmark for measuring how large language models adapt their responses to a user's functional state (readiness, flow, stability). Testing four models (Claude S…
CausalEvo: Causal Skill Self-Evolution for Health LLM Agents— arXiv cs.AI
researchapplications
CausalEvo proposes a causal inference method for LLM health agents to automatically select and evolve reusable textual skills (rules) from task feedback, achieving 60% accuracy on MedCalc-Bench-Verif…
SimOPD: Relational Knowledge Transfer for On-Policy Distillation— arXiv cs.AI
research
SimOPD proposes a new on-policy distillation method that transfers teacher knowledge through response similarity distributions rather than token probabilities or direct hidden-state matching, enablin…
OrderFlowDiT: Transport-Regularized Diffusion Transformer for Probabilistic Limit-Order Supply Forecasting— arXiv cs.AI
researchapplications
Researchers propose OrderFlowDiT, a Diffusion Transformer model with unbalanced optimal-transport loss for forecasting limit-order book volume distributions over 5-minute horizons. The approach joint…
qGAMMA: A Framework for Optimal DNN Mapping on Multi-Precision Accelerators— arXiv cs.AI
toolsresearch
qGAMMA is a new framework for optimizing how deep neural networks are mapped onto accelerators that support multiple precision formats, using genetic algorithms and performance modeling to find optim…
A Topological Signature of Learning under Noisy Labels— arXiv cs.AI
research
Researchers used persistent homology to detect geometric changes in neural network representations when trained on noisy labels, finding that corrupted labels create detectable topological signatures…
Look Back, Correct Forward: Error Feedback Calibration for Time Series Forecasting— arXiv cs.AI
research
This paper proposes Error Feedback Calibration (EFC), a test-time framework that uses revealed forecasting errors from past predictions to correct future time series forecasts without retraining the…
Sensitivity Analysis of Decision Rules in an Agent-Based Agro-Ecological Model— arXiv cs.AI
researchtools
Researchers propose a novel sensitivity analysis method for agent-based agro-ecological models that handles hierarchical decision trees where parameters activate conditionally, rather than treating a…
Brain-Aligned Multimodal LLMs for Social and Emotional Understanding of Cartoon Movies— arXiv cs.AI
researchmodels
Researchers evaluated how well multimodal LLMs align with human ratings of social and emotional content in cartoon movies, comparing their predictions against fMRI brain activity. While MLLMs showed…
QuAD: Quantitative Assessment via Distribution-Alignment for Qualitative Data Quality Enabled by a Content-Aware Multi-Agent System— arXiv cs.AI
researchtools
Researchers introduce QuAD, a framework that uses multi-agent systems to automatically assess whether qualitative survey responses are suitable for a given analytical purpose by comparing data distri…
Towards Adaptive and Real-Survey-Aligned Simulated Urban Mobility Behaviour of Generative Agents— arXiv cs.AI
researchapplications
Researchers compared generative agents (LLM-based) with traditional machine learning and expert-rule approaches for simulating urban mobility behavior in Toulouse, France. While tuned generative agen…
Upper bounds on the number of causal pieces in spiking neural networks— arXiv cs.AI
research
Researchers derive upper bounds on the number of causal pieces (discrete regions of input space with different neural responses) in feedforward spiking neural networks, establishing both combinatoria…
From Individual Songs to Music Corpora: Evaluating Explainable Harmonic Decision-Making on HookTheory— arXiv cs.AI
researchmodels
Researchers evaluated an Explainable Melody Harmonizer (EMH) model on HookTheory chord annotations, achieving 20.5% exact chord agreement and 47.62% harmonic-function agreement. The work distinguishe…
Political Narratives in the 2024 Moldovan Presidential Election: Stance, Amplification, and Influence— arXiv cs.AI
research
Researchers analyzed social media posts from X, Facebook, and Reddit during the 2024 Moldovan presidential election to identify how political narratives about European integration and Russia emerged…
Framing Spatial Hearing within the Free Energy Principle: A Hierarchical Model of Perceptual Learning— arXiv cs.AI
research
Researchers apply the free energy principle to model how humans learn to localize sounds spatially, using a hierarchical generative model that tracks how listeners' beliefs evolve across trials with…