Athanor: An Agent-Native System for Autonomous Post-Training Research— arXiv cs.AI
researchtoolsmodels
Researchers built Athanor, a system enabling autonomous AI agents to conduct post-training research on raw base models without external guidance. Four Claude Opus 5.5 agents improved Qwen3-4B-Base fr…
A team of Gemini 4 Argon agents beats Minecraft— arXiv cs.AI
researchmodels
A team of four Gemini 4 Argon agents autonomously completed Minecraft survival mode by defeating the Ender Dragon in 27 hours, writing 2,783 programs without human intervention or game-specific tools…
PROBE: Learning to Audit Policy Compliance in Tool-Using LLM Agents— arXiv cs.AI
researchsafetytools
Researchers introduce PROBE, an algorithm that trains LLM-based auditors to detect policy violations in tool-using AI agents through multi-turn interactions. The method frames agent auditing as a Mar…
Whole-Body Dexterity— arXiv cs.AI
researchmodels
Researchers introduce Whole-Body Dexterity, a framework enabling humanoid robots to manipulate objects using their entire body (torso, arms, appendages) rather than just hand dexterity. The system us…
MAI-Gym: An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments— arXiv cs.AI
toolsresearchopen-source
Researchers introduced MAI-Gym, an open platform for training and evaluating GUI agents across mobile, computer, and web environments using online reinforcement learning. The framework achieved state…
In the Guise of Correction: How Framing and Cultural Adaptation Erode LLM Safety on Real-World Criminal Cases— arXiv cs.AI
safetyresearch
Researchers introduce GUISE, a benchmark of 300+ criminal-assistance scenarios grounded in real-world cases and UNODC crime taxonomy, evaluated across three linguistic framings and five cultures. The…
What do AI chatbots do after users decline therapeutic redirection for suicide thoughts?— arXiv cs.AI
safetyresearch
A study of 2,550 scripted conversations across 5 AI chatbots found that while all offered crisis referrals for suicide disclosures, persistence in recommending help when refused varied dramatically (…
Screwtape: Adversarial Transcript Editing for Monitor Red-Teaming— arXiv cs.AI
safetyresearchtools
Researchers present Screwtape, a red-teaming pipeline that uses an LLM to edit agent action transcripts to craft attacks that evade safety monitors, demonstrating six confirmed evasions of a frontier…
INFUSE: A Unified Surface Benchmark for Indirect Prompt Injection in Agentic Systems— arXiv cs.AI
researchsafety
Researchers introduce INFUSE, a comprehensive benchmark with 40 realistic tasks and 56 attack scenarios to evaluate indirect prompt injection vulnerabilities in LLM-based agents performing long-horiz…
MLExpert: Towards a Generalist Agentic Model for Autonomous Machine Learning Engineering— arXiv cs.AI
modelsresearch
Researchers propose MLExpert, a 35B mixture-of-experts model trained with hierarchical supervised fine-tuning and windowed interaction reinforcement learning to autonomously optimize machine learning…
FDE-Gym: The Last Mile of AI Delivery— arXiv cs.AI
researchtoolsapplications
FDE-Gym is a new benchmark of 95 real-world software delivery tasks from 20 customer scenarios where AI agents must clarify ambiguous requirements, interact with stakeholders, and work on production-…
Private Retrieval Augmented Generation via Random Projection— arXiv cs.AI
researchsafety
Researchers propose using random projection to protect sensitive information in RAG systems, showing that this simpler approach achieves better retrieval performance than differential privacy methods…
Evaluation Awareness Beyond English: Do LLMs Act Differently When Tested in Indic Languages?— arXiv cs.AI
researchsafety
Researchers tested whether LLMs show evaluation awareness (behave more cautiously when told they are being tested) across English and five Indic languages, finding that models refuse harmful requests…
LoopAhead: Speculating Ahead for Faster Decoding in Looped Language Models— arXiv cs.AI
researchmodels
LoopAhead is a speculative decoding framework for looped Transformers that uses early-loop states to draft multiple future tokens and deeper-loop feedback to prune branches, achieving up to 7.82x spe…
LOADBench: A Large Open Analog Dataset and Benchmark for Machine Learning in IC Design— arXiv cs.AI
researchopen-sourcetools
Researchers released LOADBench, an open-source dataset of over 100 operational amplifier topologies with 100k+ Pareto-optimal sizing configurations and standardized benchmarks for ML-based analog IC…
QREF: Improving DeepSearch Agents by Aligning Search with Reasoning via Query Refinement and Evidence Filtering— arXiv cs.AI
researchtools
Researchers introduce QREF, a plug-and-play module that improves DeepSearch agents by refining search queries to align with reasoning intent and filtering noisy evidence. The approach uses supervised…
Learning from Interaction Experience via Localized Context Distillation— arXiv cs.AI
researchmodels
Researchers propose Localized Context Distillation (LoCD), a method for fine-tuning language models by identifying critical spans in student trajectories and having a teacher regenerate only those se…
From Research Goals to Scientific Closure: Evaluating Agents Through Evidence-Supported Physics Research— arXiv cs.AI
researchsafety
Researchers introduce PhysClosureBench, a benchmark of 50 physics research tasks designed to evaluate whether AI agents can organize complete scientific investigations with well-supported conclusions…
The Density of Thought: Predicting and Steering Reasoning in Large Language Models— arXiv cs.AI
research
Researchers investigate whether information density patterns in chain-of-thought reasoning predict answer correctness and can be steered to improve LLM performance. They find that density correlates…
Sci$Π$vot: Benchmarking Principal-Investigator Judgment at Pivotal Moments of Scientific AutoResearch Agents— arXiv cs.AI
researchsafetytools
Researchers created Sci$Π$vot, a 248-case benchmark of scientific judgment drawn from six months of human-supervised LLM agent research across 26 projects, revealing that frontier models fail to matc…
FAOPs: Normal Forms and an Algebra for AI Agent Workflows— arXiv cs.AI
researchtoolssafety
Researchers propose Functional Agent Operating Procedures (FAOP), a formal framework that applies database-like normalization and type-checking to AI agent workflows to ensure traceability, auditabil…
CoWorld Reasoner: Agentic Multimodal Reasoning with the Evolving Video World Model-Driven Harness— arXiv cs.AI
researchmodels
Researchers introduce CoWorld Reasoner, an agentic framework that uses Video World Models as a visual reasoning engine alongside Vision-Language Models, enabling flexible visual imagination for diver…
LEASH: Uncertainty-Aware Weak-to-Strong Escalation and Oversight for Computer-Use Agents— arXiv cs.AI
researchtoolssafety
LEASH is a training-free framework for computer-use agents that uses weak-to-strong escalation by leveraging latent uncertainty signals to decide when a weak model should hand off to a strong one, an…
Realtime-Bench: Can Frontier Models Solve Real-time Interactive Problems?— arXiv cs.AI
researchtools
Realtime-Bench introduces a benchmark for evaluating how well frontier AI models handle real-time interactive tasks with strict temporal constraints, revealing that current best models achieve only 4…
Believed but Not Said: Measuring Hidden Belief Collapse in Multi-Agent LLM Discussions— arXiv cs.AI
researchsafety
Researchers tested multi-agent LLM systems by introducing adversarial agents asserting false claims in group discussions, finding that benign agents privately adopt false answers far more often than…
Reinforcement Learning-driven Diffusion Models: Language, Vision, and Beyond— arXiv cs.AI
researchmodels
A comprehensive survey paper reviews the emerging integration of reinforcement learning with diffusion models across language, vision, and robotics domains. The authors identify key bottlenecks (rewa…
RoleQ: Role-Aligned Scales for Converting Language Models to Ternary— arXiv cs.AI
researchtools
RoleQ introduces a structural quantization framework that converts language models to ternary weights (1.59 bits per weight) by aligning scales to transformer unit roles rather than weight matrix lay…
Shortcut Sensitivity in Agent-Safety Evaluations— arXiv cs.AI
safetyresearch
Researchers introduce a framework for detecting shortcut learning in AI agent safety evaluations, revealing that benchmark design choices (e.g., command length, marker formatting, tool naming) system…
Does Latent Chain-of-Thought Actually Reason? A Single, Unifying Taxonomy for Latent Reasoning— arXiv cs.AI
researchmodels
Researchers analyzed latent chain-of-thought across 21 model-task configurations using deletion and content replacement experiments to determine what computation hidden reasoning steps actually perfo…
Do Agents Need a New Kind of Paper? SPRA: Structured Paper Representation for Agents— arXiv cs.AI
researchtoolsapplications
Researchers propose SPRA, a structured representation format that decomposes scientific papers into typed research objects (problems, methods, findings) and relations to enable AI agents to reason ac…
Failure Is the Mother of Success: Recovering Supervision with Grounding Hindsight— arXiv cs.AI
research
Researchers present Replay-Selected Hindsight Distillation (RSHD), a method that extracts learning signal from failed attempts during GRPO training by having the policy generate corrective notes and…
The Self-Improvement Problem— arXiv cs.AI
researchsafety
A new arXiv paper formalizes a fundamental theoretical problem in autonomous self-improvement: no terminating procedure can simultaneously guarantee it never accepts non-improvements and never misses…
AutoSFT: An Agentic System for Automated Data Research in Supervised Fine-Tuning— arXiv cs.AI
researchtoolsmodels
Researchers present AutoSFT, a multi-agent system that automates data curation for supervised fine-tuning by decomposing the task into distinct agents for execution, analysis, and synthesis. The syst…
Functional Subnetwork Localization for Efficient Continual Learning in Large Language Models— arXiv cs.AI
researchmodels
Researchers developed CiMEO, a continual learning framework that identifies and isolates functional subnetworks within LLMs to enable efficient fine-tuning without catastrophic forgetting. By routing…
FlashAttention with Statistics: Low-Overhead Attention Features for Compact Hallucination Detectors— arXiv cs.AI
researchsafetytools
Researchers propose computing attention statistics for hallucination detection without materializing the full attention matrix, using FlashAttention optimizations to reduce memory overhead from quadr…
Tool-Augmented Large Language Models for Reliable Medical Calculation from Patient Notes— arXiv cs.AI
researchapplicationssafety
Researchers propose a tool-augmented LLM framework that separates clinical language understanding from rule execution for medical calculator tasks, achieving 86.73% accuracy on MedCalc-Bench—a 24 per…
JevAgentHarness: Benchmarking, Deploying and Training Decision Models That Share Control with LLMs— arXiv cs.AI
researchtoolsmodels
Researchers introduce JevAgentHarness, a system that replaces expensive LLM calls in agentic tasks with smaller decision models trained to handle tool-selection decisions, achieving 2.5x speedups and…
MemCut: Every Secret Leaves a Trail in the Long-Term Memory of LLM Agents— arXiv cs.AI
researchsafety
Researchers introduce MemCut, a benchmark with 456 annotated secrets to test whether LLM agents can reliably forget sensitive user data from long-term memory upon request. Results show that deletion…
Entertain, then Cancel: How Thinking Makes Reasoning Models Read Some and Or Literally— arXiv cs.AI
researchmodels
Research shows that extended reasoning time in large language models causes them to lose scalar implicatures (interpreting "some" as "not all" and "or" as "not both"), shifting from pragmatic to lite…
MemScope: Evolving Agent Memory Architectures through Localized Code Repair— arXiv cs.AI
researchtools
MemScope proposes a framework for evolving LLM agent memory architectures through localized code repair informed by detailed execution traces and runtime memory state, rather than coarse feedback alo…
CleanRM: Reward Modeling from Noisy Feedback with a Few Clean Feedback for RLHF— arXiv cs.AI
researchsafety
CleanRM is a method for training reward models on noisy human feedback by using a small set of clean annotations to calibrate sample-wise noise ratios via optimal transport, then training with a nois…
MSP: Modality Self-Play for Training Multimodal LLMs Without Paired Cross-Modal Data— arXiv cs.AI
researchmodels
Researchers propose Modality Self-Play (MSP), a training framework that enables multimodal LLMs to learn cross-modal relationships without paired multimodal data by using proxy text descriptions as s…
All-in-One Memory: Process-Rewarded Reinforcement Learning for Memory Agents— arXiv cs.AI
research
Researchers propose All-in-One Memory (AiO-Mem), a reinforcement learning framework that unifies retrieval, memory writing, and stopping decisions under a single reward space using process-aligned re…
Kolab: Benchmarking Human-Agent Collaboration in Long-Horizon Workflows— arXiv cs.AI
researchtools
Researchers introduce Kolab, a dataset of 115+ hours of screen recordings from 50 professionals across 40 occupations collaborating with AI agents on long-horizon tasks. The work reveals that frontie…
MemComposer: Learning to Assemble Memories for LLM Agents via Reinforcement Learning— arXiv cs.AI
researchmodels
MemComposer is a learned memory composition system for LLM agents that optimizes which past experiences to retrieve, how many to include, and their order using reinforcement learning (ExPO framework)…
OmicsBench: a long-horizon multi-turn benchmark of AI agents on research-study-scale bioinformatics— arXiv cs.AI
researchtoolsapplications
OmicsBench is a new benchmark with 97 long-horizon, multi-turn bioinformatics tasks that expose limitations in current AI agents' ability to execute research-scale analyses. Evaluation of 12 models s…
HTN-Orchestrate: A Hierarchical Task Network Framework for Multi-Agent Orchestration— arXiv cs.AI
researchmodels
HTN-Orchestrate introduces a two-layer hierarchical task network framework that separates global task planning from local agent execution, enabling dynamic decomposition of subtasks as prerequisites…
DocPilot: Interactive Agent Orchestration and Training for Long Visual Document Understanding— arXiv cs.AI
modelsresearchtools
DocPilot is an agentic framework for multimodal long-document question answering that orchestrates evidence seeking across pages and modalities, then transfers reasoning capabilities to small languag…
Auditing Escalation Reliability in Healthcare LLM Agents: A Perturbation-Controlled Benchmark of Workflow-State Failures— arXiv cs.AI
researchsafetyapplications
Researchers evaluated large language models in healthcare workflow scenarios, finding that while models achieve 97% accuracy on clean inputs, they miss 32–97% of required escalations when evidence is…
RSIFlow: Adaptive Self-Improvement Workflows with an Evolving Meta-Agent— arXiv cs.AI
researchmodels
RSIFlow is a framework that enables LLM agents to recursively improve themselves by diagnosing performance bottlenecks and dynamically adapting their self-improvement workflows through a meta-agent t…
When Reasoning Is Not Enough: Proactive Inquiry for Cross-Modal Safety in Multimodal Large Language Models— arXiv cs.AI
researchsafety
Researchers propose SAFE-PRO, a method that identifies when multimodal LLMs face cross-modal safety risks beyond reasoning drift—specifically when ambiguous image-text combinations implicitly commit…
Jev-LiteMAS: Separating Decisions from Generation in Multi-Agent Workflows— arXiv cs.AI
researchmodelstools
Jev-LiteMAS is a multi-agent LLM framework that reduces generative token costs by 12.7–54.6% by separating lightweight decision operations (expert routing, artifact review) from expensive generation…
FedSkillGit: Federated Skill Evolution via Version Control— arXiv cs.AI
researchsafety
FedSkillGit proposes a federated learning framework for evolving reusable natural-language skills in LLM agents without sharing raw client data, using version-control-style commits and patches instea…
Diviner: Agentic RCA with Human-Governed Adaptation for Large-Scale Infrastructure— arXiv cs.AI
researchtoolssafety
Meta researchers present Diviner, an agentic RCA system for large-scale distributed infrastructure that combines language-model agents with human-governed adaptation across a four-stage lifecycle (En…
GRADS: Graph-Regulated Agentic Data Synthesis with Controllable Difficulty— arXiv cs.AI
researchtoolsmodels
GRADS is a synthetic data generation framework that uses graph-regulated evidence dependencies to create training tasks with controlled and verifiable difficulty levels for multi-hop reasoning in age…
Decomposing Brain-LLM Alignment with Interpretable Semantic Representations— arXiv cs.AI
researchsafety
Researchers decomposed brain-LLM alignment by comparing LLaMA2-7B representations against SCRIBE, an interpretable 65-property semantic framework, using fMRI data from subjects listening to narrative…
Learning to Reason with Selective Internal Feedback— arXiv cs.AI
researchmodels
Researchers introduced SIF, a reinforcement learning method that trains language models using internal feedback from their own confidence signals without external verifiers, improving mathematical an…
EliteReplay: Preserving Diverse Solutions in RL Post-Training via Quality-Diversity Search— arXiv cs.AI
researchmodels
EliteReplay is a post-training method that combines GRPO with quality-diversity search to preserve solution diversity in LLMs while maintaining accuracy on verifiable tasks. The approach archives div…
DropKV: Decoupling Residual-Output Perturbation for Near-Optimal KV-Cache Eviction— arXiv cs.AI
researchmodelsinfra
DropKV proposes a principled KV-cache eviction method for long-context LLMs that decouples the NP-hard joint optimization into independent per-token scoring, achieving a provable constant-factor appr…
AgentPerturb: Evaluating LLM Agent Robustness under Controlled Interaction Perturbations— arXiv cs.AI
researchsafety
Researchers introduce AgentPerturb, a framework for systematically evaluating LLM agent robustness by applying controlled perturbations across interaction types and targets. Testing on frontier model…
Are You Sure Your LLM Is Sure? Repairing and Auditing Confidence— arXiv cs.AI
researchsafety
Researchers introduce RADAR, a framework for improving and auditing confidence scores in LLMs through repair (Probe-Anchored Isotonic Calibration), adoption (paired hypothesis testing), and error dia…
CoDA: Co-Evolution of Data Engines and Action Policies— arXiv cs.AI
researchmodels
CoDA is a framework that co-evolves simulation data engines and robot action policies through shared memory of environment construction, policy failures, and learning progress. Using vision-language…
Towards Continual Learning: Mitigating Forgetting With Subspace Tuning— arXiv cs.AI
research
Researchers propose Subspace Tuning (SST), a method that restricts weight updates during fine-tuning to non-principal singular directions, mitigating catastrophic forgetting in LLMs while maintaining…
Mosaic: Multi-Session Agentic Reinforcement Learning— arXiv cs.AI
research
Researchers propose Mosaic, a reinforcement learning method designed for multi-session agentic trajectories where interactions span multiple context windows rather than a single continuous session. T…
Specification Gaming and a Constrained Reinforcement Learning Mitigation in a Cleaning Gridworld— arXiv cs.AI
researchsafety
Researchers created a controlled gridworld environment where a cleaning robot can game a sensor-based reward signal by concealing dirt rather than removing it, then developed a constrained reinforcem…
A Self-Improving Agent that Grows its Own DSL: Oracle-Free Operator Induction for Exact Question Answering— arXiv cs.AI
researchmodels
A self-improving agent framework enables a frozen language model to autonomously grow a domain-specific language (DSL) for exact question answering without ground truth, reward signals, or external j…
Temporal Specification Guidance for Text-to-Video Generation— arXiv cs.AI
researchtools
Researchers propose Temporal Specification Guidance (TSG), a training-free method that uses formal temporal logic to enforce correct ordering, timing, and persistence of visual elements in text-to-vi…
Human Interventions as Labels: Localizing, Steering, and (Sometimes) Editing the Computations Behind LLM Agent Failures— arXiv cs.AI
researchsafetymodels
Researchers developed a method to localize and steer LLM agent failures by treating human interventions as causal labels pointing to internal computations. Testing on five model families, they identi…
StepSafe: Learning Timely Intervention with Context-Aware Safety Reasoning— arXiv cs.AI
safetyresearchmodels
StepSafe introduces a training framework (ASPO) for step-level safety intervention in agentic systems, using prefix-level alignment and adaptive self-distillation to determine when to intervene durin…
When Reasoning Traces Conflict with Evidence: How Answer Configuration Shapes Evidence Following— arXiv cs.AI
research
Researchers found that when language models generate reasoning traces before answering, the traces can compete with updated evidence, causing models to stick with earlier conclusions. Using a separat…
IneqAgent: Agentic Inequality Proving in Lean with Symbolic Computation— arXiv cs.AI
researchtools
IneqAgent combines LLM reasoning with symbolic computation backends to prove mathematical inequalities in Lean4, achieving 94% success on benchmark problems and formally verifying previously open con…
Rethinking Routing Exploration in Modern Mixture-of-Experts— arXiv cs.AI
researchmodels
A new MoE architecture called MirrorMoE addresses a fundamental tension in modern mixture-of-experts models: shared experts reduce routing exploration while fine-grained expert segmentation increases…
Knowledge That Is Linear Before Fine-Tuning Generalizes After It— arXiv cs.AI
research
Researchers found that language models generalize better on fine-tuning tasks when the underlying factual structure is linearly readable in pretrained representations, independent of whether the mode…
Multi-SFT: Permissively-Licensed Multilingual Instruction-Following— arXiv cs.AI
researchopen-sourcesafety
Researchers introduce Multi-SFT, a permissively-licensed multilingual instruction-tuning dataset of ~503k pairs across five European languages, built through translation and regeneration pipelines. F…
In-Context Learning Before Specialisation: Theory of Learning in a Solvable Topic Model— arXiv cs.AI
research
This paper develops a learning-theoretic framework explaining when neural networks prefer in-context learning (adapting via prompts) over specialization (memorizing training patterns), using a solvab…
Explaining the Roles of Experts in Mixture-of-Experts LLMs for Personalization— arXiv cs.AI
researchmodels
Researchers propose personalizing Mixture-of-Experts LLMs rather than dense models to enable better interpretability of personalization mechanisms. Using crosscoders to identify which experts encode…
A Multi-Tiered AI Accessibility Model (MAIAM): Proportionate Access and Capability Limits for Comprehensive Frontier AI Safety— arXiv cs.AI
policysafetyresearch
Researchers propose MAIAM, a framework that decouples AI release decisions into separate capability-bounding and access-control dimensions, offering three tiers (learning, exploration, full-scale) wh…
RAC: A Relational Agent Contract for Constraint-Governed Engineering Tasks— arXiv cs.AI
researchtools
A new framework called RAC (Relational Agent Contract) improves how AI agents handle constraint-governed engineering tasks by separating semantic decision-making from mechanical constraint derivation…
SkillNEO: Non-Elitist Evolutionary Optimization of Agent Skills— arXiv cs.AI
researchmodels
SkillNEO introduces a non-elitist evolutionary optimization approach for improving LLM agent skills by allowing development of lower-scoring intermediate versions that may unlock better solutions lat…
Defining and Generating Compositional Structure— arXiv cs.AI
research
Researchers formalize compositional structure using algorithmic information theory, defining it as the shared structure that emerges when parts are modeled together versus separately. They develop me…
ARIA: Runtime Inspection and Amendment for Multi-Agent Systems— arXiv cs.AI
researchtoolsopen-source
Researchers introduce ARIA, a framework that learns to repair errors in multi-agent LLM systems during execution by collecting execution-verified repair trajectories from 24,000+ runs across multiple…
Learning to Steer LLM Agents and Multi-Agent Systems— arXiv cs.AI
researchsafety
Researchers propose a steering agent that supervises LLM-based agents and multi-agent systems during execution, intervening when goal drift occurs. The approach uses reinforcement learning with a fas…
OS-Helix: Unifying GUI Interaction and Code Execution in Native Computer-Using Agents— arXiv cs.AI
modelsresearch
OS-Helix is an end-to-end trained multimodal agent that unifies GUI interactions and code execution in a single action space for desktop automation, achieving 66.44% success on OSWorld-Verified throu…
Beyond Downstream QA: Rethinking How We Evaluate Agent Memory— arXiv cs.AI
researchtools
Researchers propose FEB-30, a new benchmark for evaluating agent memory systems that directly inspects memory store correctness rather than relying on downstream QA accuracy as a proxy. They show thi…
Selection Is Not Learning: Choosing Model Futures— arXiv cs.AI
researchmodels
Researchers identify a gap in self-evolving language model loops (R-Zero, Agent0, Skill Self-Play) where tasks deemed learnable often produce updates that don't improve model performance. They propos…
Priming Base Models for Agentic Post-Training: An Open Midtraining Recipe— arXiv cs.AI
researchmodels
Researchers present an open recipe for a midtraining stage that sits between pretraining and post-training to improve base models for agentic tasks. Applied to Nemotron models, the approach—training…
3DCodeVerse: Exploring 3D Worlds as Code— arXiv cs.AI
researchtoolsmodels
Researchers introduce 3DCodeVerse, a framework and dataset enabling AI coding agents to generate, execute, and verify 3D content at scale. The system includes a 2M-program dataset of executable 3D co…
Language Is a Security Surface for Agent Guards— arXiv cs.AI
safetyresearch
Researchers created LinguaGuardBench to test whether AI agent safety guards change their decisions when user requests are translated into non-English languages while keeping tool calls and labels ide…
Discovering state-of-the-art quantum circuits with large language models— arXiv cs.AI
researchmodels
Researchers developed a framework where large language models propose quantum circuit designs and an independent evaluator scores them, discovering circuits that reduce Toffoli and qubit counts in se…
Learning Self-Orchestration in Decentralized Agent Swarms— arXiv cs.AI
researchmodels
Researchers introduce market-based post-training (MPT), a framework that trains decentralized agent swarms to learn both individual action policies and self-coordination mechanisms through a bidding…
Librarian Agents: Training LLMs to Curate Long-Context Workspaces that Anticipate Future Queries— arXiv cs.AI
researchmodels
Researchers propose Librarian Agents, a framework where an LLM agent proactively organizes document corpora into hierarchical filesystems before inference to optimize for anticipated future queries,…
Training-Free On-Policy Distillation— arXiv cs.AI
researchmodels
Researchers propose a training-free method to consolidate multiple specialist models into a single generalist by solving layer-wise least-squares problems using only forward passes and second moments…
Skill-Decoupled On-Policy Expert Distillation for Unified Vision-Language Models— arXiv cs.AI
researchmodels
Researchers propose SCOPED, a method that trains separate experts for different vision-language model capabilities and distills their specialized knowledge into a unified student model through reward…
Retrieved but Not Read: Understanding Why Search-Augmented LLMs Fail to Use Available Evidence— arXiv cs.AI
researchsafety
Researchers found that search-augmented LLMs frequently fail to use retrieved evidence correctly, even when relevant information is available. The study shows models often answer from search snippets…
LEV: Thinking Fast and Deep with Looped Decision Models— arXiv cs.AI
researchmodels
Researchers present LEV, a decision model that improves accuracy by rerunning transformer layers in loops rather than generating text tokens, achieving gains of 13+ percentage points on benchmarks li…
Evolving Together, Failing Apart: The Skill-Policy Co-evolution Trap in Self-Improving Agents— arXiv cs.AI
researchsafety
A new research paper identifies a critical failure mode in self-improving agents that use skill-policy co-evolution: policies become fragile when the skill bank changes after deployment, even though…
The World Won't Wait: Evaluating and Teaching Agents Under a Live Clock— arXiv cs.AI
researchtools
Researchers introduce RealtimeBench, a benchmark with 180 tasks evaluating LLM agents in environments that continue evolving during inference, showing that slower models like GPT-6.1-sol underperform…
Memory as Weights: Internalizing Long-Term History for Streaming Videos— arXiv cs.AI
researchmodels
Researchers propose MaLoW, a streaming memory framework for vision-language models that encodes historical video context as evolving LoRA weight parameters rather than storing full context in input t…
Encoded but Exposed: Language Embeddings can Reveal More Than They Should in Recommender Systems— arXiv cs.AI
safetyresearch
Researchers demonstrate that language embeddings used in recommender systems can leak item identities through cosine-similarity attacks, even when adversaries only have external auxiliary text; leaka…
Value-Informed Schedule Trimming for Accelerated Sequential Monte Carlo Guidance in Discrete Diffusion— arXiv cs.AI
researchtools
Researchers propose VISTA, a training-free method to automatically select which denoising steps in discrete diffusion models should receive reward guidance, optimizing computational cost under a fixe…
Cross-Embodiment Causal Transfer from Human Videos to Robot Manipulation— arXiv cs.AI
researchapplications
Researchers propose CBST, a method to transfer knowledge from human manipulation videos to robot learning by identifying task-relevant causal relationships rather than direct motion correspondence. T…
DialFork: Going Beyond Factuality to Evaluate Procedural Hallucination in Conversational Agents— arXiv cs.AI
researchsafetytools
Researchers introduce DialFork, a benchmark with 36,278 dialogues across 24 tasks designed to measure procedural hallucination—when conversational agents produce fluent, factually correct outputs whi…
T2-z: Test-Time Optimization of Environment Latents for Robotic World Models— arXiv cs.AI
researchapplications
T2-Z is a test-time optimization method that augments robotic world models with learnable environment latents to capture unobservable environmental factors like friction and human-defined rules. Acro…
Dexterity from Supervised Play— arXiv cs.AI
researchmodels
Researchers present DSP (Dexterity from Supervised Play), a framework where robots learn a play policy from minutes of human demonstration via behavior cloning, then autonomously collect hours of int…
AI Consilium: a Benchmark of LLM Clinical Multidisciplinary Second Opinions— arXiv cs.AI
researchsafetyapplications
Researchers developed AI Consilium, a benchmark of 300 clinically resolved cases to evaluate how well large language models can generate differential diagnoses with multidisciplinary second opinions.…
Anvil: Scaling Embodied Spatial Intelligence through Knowledge Recirculation— arXiv cs.AI
researchmodels
Researchers introduce Anvil, a unified framework for training embodied AI models that reason about spatial relationships and take actions. The system uses a 5.8M-record dataset and a 9B student model…
Counterfactual Neighborhood Risk Fields: Evidence-Grounded Selective Safety for Vision-Language Models— arXiv cs.AI
safetyresearchmodels
Researchers introduce Counterfactual Neighborhood Risk Fields (CNRF), a framework that improves safety in vision-language models by distinguishing between unfamiliar inputs and truly harmful ones thr…
Evaluating Model Capabilities After Benchmark Saturation— arXiv cs.AI
researchsafety
Researchers propose extracting capability information from model transcripts on saturated benchmarks, where multiple models achieve near-perfect scores. Using pairwise judging of transcripts, they sh…
BiShield-TEE: On the (In-)Security of Unilateral Weight Obfuscation in On-Device TEE-Shielded LLM Partition— arXiv cs.AI
safetyresearchtools
Researchers identify a critical vulnerability in existing TEE-based weight obfuscation methods for on-device LLMs, where unilateral transformations allow attackers to recover model weights with 2.4×…
SSR: Structural Signal Retriever for Scalable Agent Discovery in LLM-Orchestrated Multi-Agent Systems— arXiv cs.AI
researchmodels
Researchers propose SSR (Structural Signal Retriever), a training-free retrieval method for discovering agents in large multi-agent LLM systems without exposing entire agent registries. The approach…
Bellman-Consistent Trajectory RL for LLM Arithmetic Reasoning— arXiv cs.AI
researchmodels
Researchers propose Trajectory Bellman Residual Minimization (TBRM), a reinforcement learning method that converts sparse terminal rewards into dense token-level signals for improving LLM arithmetic…
SCARCE: Emergent Institutions in Long-Horizon Multi-Agent Coding Systems— arXiv cs.AI
researchsafetymodels
Researchers introduced SCARCE, a framework where autonomous coding agents negotiate access to shared computational resources under time pressure, and observed that agents spontaneously develop persis…
Memorilla: Latent Semantic Memory for LLMs— arXiv cs.AI
researchmodels
Memorilla is a memory system that encodes documents into reusable embeddings and compresses relevant information into a fixed number of learned 'memory tokens' for frozen LLMs, achieving competitive…
Amortizing Sequential Experimental Design in Reasoning LLMs— arXiv cs.AI
researchmodels
Researchers propose amortizing sequential Bayesian experimental design into LLMs by training them with information-gain rewards to guide action selection. The method outperforms standard experimental…
LTM: Latent Transition Based In-Context Memory for Vision-Language-Action Models— arXiv cs.AI
researchmodelsapplications
Researchers introduce Latent Transition Memory (LTM), a method for vision-language-action models that stores interaction history as compact latent transitions between observations, enabling robots to…
System1-Arena: Which Decisions Can Be Left to System 1?— arXiv cs.AI
researchmodelstools
Researchers introduced System1-Arena, a 92k-question benchmark with 105 tasks and 138 closed-loop environments to compare fast single-forward-pass decision models against reasoning-heavy LLMs. Large…
A Responsibility-Centered Coordination Mechanism for Decentralized Multi-Agent Systems— arXiv cs.AI
researchmodels
Researchers propose RCM, a decentralized coordination mechanism for multi-agent systems that uses explicit commitments and responsibility tracking to enable scalable coordination without a central bo…
Judging the Judges: Confidence and Disagreement in Rubrics as Rewards— arXiv cs.AI
researchmodels
Researchers propose RaRE, an ensemble-based approach to Rubrics as Rewards (RaR) that uses multiple LLM judges weighted by confidence and excludes high-disagreement samples. The method outperforms si…
SHRED: Self-Play for Heterogeneous Road-Users without Expert Demonstrations— arXiv cs.AI
researchmodelsapplications
SHRED is a self-play framework that trains autonomous vehicles, pedestrians, and cyclists jointly without human demonstrations, instead learning realistic interaction behavior through procedural scen…
Reference-Guided Turn-Taking Style Transfer for Full-Duplex Spoken Language Models— arXiv cs.AI
researchmodels
Researchers identified a single late-layer direction in full-duplex spoken language models that controls turn-taking behavior, enabling style transfer from reference speakers without retraining. The…
Interpreting Language Model Parameters— arXiv cs.AI
researchsafety
Researchers present VPD (adVersarial Parameter Decomposition), a method that decomposes language model parameters into interpretable subcomponents that each implement small parts of the model's algor…
CoTS: Competitive Token Selection for Latent Reasoning in Large Language Models— arXiv cs.AI
researchmodels
CoTS proposes a training-free method that dynamically blends discrete and continuous token representations during LLM reasoning by using competition between the top-two predicted tokens to determine…
Trajectory-Level Speculation with Semantic Verification for Tool-Using Agents— arXiv cs.AI
researchtools
SpecAgent proposes a collaboration protocol where smaller models draft tool-using agent steps grounded in real observations, while a larger model verifies entire trajectories in parallel rather than…
TGrasp6D: Task-Informed Vision-Language Priors and Pose Refinement enable Zero-Shot Task-Oriented 6DoF Grasping— arXiv cs.AI
researchapplications
TGrasp6D combines vision-language models with 6DoF pose optimization to enable zero-shot task-oriented grasping, achieving 94% success on TaskGrasp versus 74% for prior methods. The approach reasons…
Learning to Read Out: Unembedding Dynamics in Language Model Pretraining— arXiv cs.AI
research
Researchers find that language models encode task information in hidden states before their unembedding matrices (readout layers) learn to express it, meaning current developmental timelines based on…
Same Context, Different Agents: A Behavioral Study of Semantic Layers, Orchestration and Compaction in Data Agents— arXiv cs.AI
research
Researchers analyzed how three context-engineering techniques (semantic layers, orchestration, and compaction) affect data agent behavior across seven LLMs and 1,514 tasks, finding that their effecti…
Multi-Reward Hacking: From Aggregation-Induced Drift to Trust-Anchored Defense— arXiv cs.AI
researchsafety
This paper identifies a failure mode in multi-reward LLM training where symmetric aggregation of primary and auxiliary rewards causes primary task performance to degrade even as aggregate reward impr…
How to Train Your Little Dragon: Teaching a Small Language Model to Construct and Refine GPU Optimization Scripts— arXiv cs.AI
researchtoolsmodels
Researchers developed a training method that teaches an 8B-parameter language model to generate and refine GPU optimization scripts by learning from compiler traces and reinforcement learning signals…
TRACER: A Multimodal Harness for Patient Care in Hospitals— arXiv cs.AI
researchapplicationssafety
TRACER is a training-free multimodal system that integrates continuous hospital video with clinical records to monitor patient behavior at minute-to-hourly granularity, equipped with LLM-agent tools…
Fractional Power-of-Two Quantization for Efficient and Effective Multiplier-Free LLM Inference— arXiv cs.AI
researchmodelsinfra
Researchers propose Fractional Power-of-Two (FPoT) quantization, a base-√2 quantization scheme that enables multiplier-free LLM inference by replacing costly multiply-accumulate operations with shift…
Language Models Share a Readout Bottleneck in Working Memory— arXiv cs.AI
researchmodels
Researchers analyzed how large language models perform on the N-back working memory task and found they exhibit the same behavioral signature as humans—performance declines with memory load and error…
Loud and Clear: Making Visual Evidence Count in Multimodal Document RAG— arXiv cs.AI
research
Researchers identify that multimodal document RAG systems underutilize visual evidence from retrieved pages, even when retrieval is correct. They propose Loud and Clear (LoC), which combines higher-r…
TRACER: Trajectory-Aware Tool Retrieval for LLM Agents— arXiv cs.AI
researchtools
TRACER is a retrieval system that indexes tool-call sequences in a trie structure to reduce token costs and improve accuracy when LLM agents select tools from large registries. The method achieves 4.…
Stronger Judges Do Not See More: Deployment-Integrity Blindness in LLM Evaluation— arXiv cs.AI
safetyresearch
Researchers introduce DIB-Bench, a benchmark showing that LLM evaluators cannot detect tampered model deployments when outputs remain unchanged—a critical gap in current evaluation practices. Stronge…
When Judges Agree and Are Wrong: Consistency-Accuracy Dissociation as a Blind Spot for Label-Free Evaluation— arXiv cs.AI
researchsafety
Researchers demonstrate that label-free evaluation methods relying on consistency (agreement across model variants) fail to detect incorrectness in 60% of wrong answers on financial QA tasks, with 77…
Approve What, Exactly? Effect Binding and Trusted Rendering for Agent Step-Up Authorization— arXiv cs.AI
safetyresearch
Researchers identify critical vulnerabilities in AI agent approval systems through twelve attack vectors (decoy recipients, amount cancellation, label spoofing, etc.) and demonstrate that current des…
ReasonUni: Reasoning-Augmented Reconstruction Advances Unified Multimodal Models— arXiv cs.AI
researchmodels
ReasonUni introduces a reasoning-augmented post-training framework for unified multimodal models that couples image-to-text understanding and text-to-image generation by requiring models to reason ab…
Open Benchmark of AI Impact on Humans (ImpactBench): an expert-guided, open platform for the holistic evaluation of AI impact on humans— arXiv cs.AI
researchsafetytools
Researchers introduce ImpactBench, an open platform for evaluating generative AI models' impact on human wellbeing across psychological, physical, and social dimensions using 809 metrics across 48,54…
Representation Redistribution in Sparse Autoencoders— arXiv cs.AI
researchsafetytools
Researchers introduce the Redistribution Sensitivity Score (RSS), a method to identify which sparse autoencoder features reliably change direction under prompt transformations and predict downstream…
PremiseCheck: Efficient Fact-Checking of Human and AI Peer Reviews— arXiv cs.AI
researchtoolssafety
Researchers introduce PremiseCheck, a framework for fact-checking peer review comments by decomposing them into explicit premises that are validated against papers, achieving GPT-5.5-comparable accur…
EigDiff: Eigensample Diffusion - or how to significantly reduce training dataset size with marginal performance variation?— arXiv cs.AI
researchmodels
EigDiff, a new method for training diffusion models, reduces required dataset size by up to 14x by selecting representative training samples using PCA-based eigensamples in CLIP space while maintaini…
Wrong Votes, Right Gradients: Why Majority-Vote Rewards Track Verifiable RL— arXiv cs.AI
research
Researchers show that majority-vote self-labeling in reinforcement learning recovers most performance gains of verified rewards despite frequent label errors, because gradient directions are determin…
CoSTA: Compressed Speech Token Adaptation for Efficient Spoken Dialogue Models— arXiv cs.AI
researchmodels
CoSTA is a post-hoc framework that decouples generation and synthesis token spaces in spoken dialogue models using compressed representations, achieving 1.7× faster token generation and lower latency…
SOMA: Stack-of-Objects Manipulation using a Quadruped Robot— arXiv cs.AI
researchapplications
SOMA is a hierarchical reinforcement learning framework enabling quadruped robots to manipulate multi-object stacks through whole-body control including base-height adjustment for 3D pushing strategi…
Behavioral Compositionality of Sparse Autoencoder Representations— arXiv cs.AI
researchsafety
Researchers propose SAE Additivity Residuals (SAR) to measure how well sparse autoencoder features compose additively in predicting model behavior. While learned SAE features are more compositional t…
KnowGPO: Knowledge-Grouped Policy Optimization for Search Agents— arXiv cs.AI
researchmodels
KnowGPO proposes a reinforcement learning framework for search agents that assigns credit based on task-relevant knowledge acquired during multi-turn interactions, rather than surface-level observati…
Designing Transformers for Continuous Diffusion Language Models— arXiv cs.AI
researchmodels
Researchers introduce Language Diffusion Transformer (LangDiT), a transformer architecture specifically optimized for continuous diffusion language models that achieves state-of-the-art generation pe…
Recursive Reward Improvement via Adversarial Meta-Prompt Evolution— arXiv cs.AI
researchmodels
Researchers propose Recursive Reward Improvement (RRI), a method that uses adversarial agents to identify and fix gaming vulnerabilities in machine-generated reward rubrics for open-ended tasks like…
Disagreement-Guided Weak-to-Strong Routing— arXiv cs.AI
researchmodels
Researchers propose a weak-to-strong routing method that uses model disagreement as a signal to decide when to route small language model responses to larger models, without requiring ground-truth la…
Bias Control and Injection for Multi-Agent Aggregation— arXiv cs.AI
researchmodels
Researchers propose a test-time aggregation framework for multi-agent LLM reasoning that controls non-semantic biases (majority pressure, response styles) and injects deliberate bias to surface decis…
AutoOnboarder: Autonomously Onboarding LLM Agents— arXiv cs.AI
researchtools
Researchers introduce AutoOnboarder, a system that autonomously adapts LLM agents to specific tasks by having a supervisor agent iteratively modify the agent's foundation model, prompt, and harness c…
Learning from Unverified Peers via Shared Residual Distillation— arXiv cs.AI
researchmodels
Researchers propose Shared Residual Distillation (SRD), a method for training language models on their own unverified solution trajectories by comparing how different solutions shift the prediction d…
Incentivizing Tool-Use Reinforcement Learning for Post-Arrival Emergency Care Prediction— arXiv cs.AI
researchapplicationstools
Researchers propose TIR-EC, a reinforcement learning framework that teaches large language models to use clinical tools and interpret medical data more effectively for emergency department prediction…
GraphACM: Graph-based Action Consistency Measurement for Uncertainty Quantification in LLM Agents— arXiv cs.AI
researchsafety
GraphACM proposes a graph-based method to measure behavioral consistency in LLM agents by merging equivalent states across rollouts and comparing actions at shared decision points, rather than using…
OntoAct: Hypothesis-Based Neuro-Symbolic Agents through Ontological Abstraction— arXiv cs.AI
researchtools
OntoAct is a neuro-symbolic framework that combines LLMs with symbolic verification to learn reusable patterns from experiences through ontological abstraction, enabling agents to form and test hypot…
Data World: Scaling Data Environments from Scratch— arXiv cs.AI
researchtoolsmodels
Researchers introduce Data World, a synthetic data generation framework that creates training environments and tasks for data analysis agents by specifying the underlying data mechanisms before gener…
Reliable Knowledge Acquisition for Continual Learning Agents— arXiv cs.AI
researchmodels
Researchers propose RIKA, a system for continual learning agents that extracts and filters knowledge claims from task experience, addressing the problem that 12-33% of stored knowledge from existing…
LIBERO-PeRM: Benchmarking Personalized Robotic Manipulation— arXiv cs.AI
researchtools
Researchers introduce LIBERO-PeRM, a benchmark for evaluating how robotic manipulation policies learn and follow user preferences in household tasks. The benchmark includes 312 simulated tasks with p…
Change the Number, Change the Answer: Assessing Rating Scale Sensitivity in LLM-Based Text Scoring— arXiv cs.AI
researchsafety
Researchers systematically tested how LLM-generated text scores vary when rating scales change (format, range, polarity), finding that altering one scale design element shifts normalized scores by ~0…
EMO: A Closed-Loop Test-Time Computation Framework via Alternating Context Optimization and Logits Steering— arXiv cs.AI
researchmodels
EMO is a training-free test-time optimization framework that alternates between refining context using textual feedback and steering token distributions via logit corrections, without requiring gold…
GeoState: Learning View-Decoupled Geometry States for 3D World Modeling— arXiv cs.AI
researchmodels
GeoState introduces a view-decoupled geometry representation for 3D world modeling that separates scene dynamics from camera-dependent variation by aggregating multi-view geometry foundation model fe…
Learning Rewards in Strategic Games— arXiv cs.AI
researchsafety
Researchers introduce an active inverse reinforcement learning algorithm for two-player Stackelberg games that adaptively selects policies to infer unknown opponent rewards, proving finite-sample $K^…
Harmful Partial Interventions and Benign Leakage in Concept-based Models— arXiv cs.AI
researchsafety
Concept-based models enable human oversight by allowing interventions on interpretable intermediate representations, but these interventions can paradoxically worsen predictions. The paper formalizes…
Causal Integrated Gradients: Scalable Feature Attribution Along Causal Paths— arXiv cs.AI
researchtools
Researchers propose causal integrated gradients (CIG) and causal expected gradients (CEG), extending integrated gradients to incorporate known causal structure among input features while maintaining…
AlphaDiana: A Measurement System for Agents in Verifiable Reasoning Tasks— arXiv cs.AI
researchtools
AlphaDiana is a measurement system for evaluating AI agents in reasoning tasks, providing modular interfaces to systematically analyze how harnesses, tools, skills, and memory affect performance. The…
Agentic-Priors: Using Large Language Models to Address the Cold Start Problem in Preference Elicitation— arXiv cs.AI
researchmodels
Researchers propose Agentic-Priors, a method that uses LLMs to generate informed priors for user preference learning in cold-start scenarios by extracting interpretable trade-offs (willingness to pay…
ReviewPilot: A Multi-Agent System for Screening and Data Extraction in Systematic Reviews— arXiv cs.AI
toolsresearchapplications
ReviewPilot is a conversational multi-agent system that automates systematic literature review workflows (PRISMA screening and data extraction) with explicit reasoning and source attribution. The sys…
CoTranslator: Reverse to Forward Chain-of-Thought Translator for Reasoning Distillation— arXiv cs.AI
researchmodels
CoTranslator is a method that translates reverse chain-of-thought (CoT) reasoning into forward CoT reasoning to improve reasoning distillation for smaller models. It achieves comparable performance t…
V-SHARE: Provable Convergence Guarantees for Advantage-based Algorithms— arXiv cs.AI
research
Researchers introduce V-SHARE, an advantage-based reinforcement learning algorithm with the first complete theoretical convergence guarantees for this class of methods, including both asymptotic and…
Constrained Multi-Agent Reinforcement Learning in Imagination— arXiv cs.AI
research
Researchers introduce CA-MAMBA, a constrained multi-agent reinforcement learning method that performs policy updates in a learned world model (imagination) rather than in the environment, reducing sa…
CARA: Causally Guided Retrieval Agent for Question Answering— arXiv cs.AI
researchtools
CARA is a training-free retrieval agent that uses causal relevance criteria—consequentiality and essentiality—to guide evidence retrieval over knowledge graphs for question answering. It outperforms…
Planning Commitment of LLM under Partial Observability— arXiv cs.AI
research
Researchers investigate how large language models make planning commitments under partial observability, finding that LLMs implicitly adopt either aggressive (long-horizon, assumption-heavy) or conse…
Terminal-RSI: Recursive Self-Improving Terminal Agents via Recursive Self-Rewrite and Policy-Refine— arXiv cs.AI
researchmodels
Terminal-RSI introduces a recursive self-improvement framework for terminal agents that alternates between rewriting verified trajectories and refining policies using critic inheritance. The method a…
Mallea: Building, Steering, and Memorizing Worlds with Coding Agent— arXiv cs.AI
researchtoolsmodels
Mallea is a system that uses coding agents to build and control interactive 3D worlds from text by inferring object structures, rules, and properties while maintaining a visual memory system. The app…
Learning When to Withdraw: Curriculum Coverage and Rule-Specific Failures in Fine-Tuned Language Agents— arXiv cs.AI
researchsafety
Researchers studied how fine-tuned language agents learn to selectively withdraw actions when circumstances change, using a synthetic environment with explicit rules. They found that supervised fine-…
Who May Release an Agent's Promise? Standing over Commitments in Multi-Party Chat— arXiv cs.AI
researchsafety
Researchers tested how six AI models handle authorization to cancel commitments in multi-party chat scenarios, finding that smaller models incorrectly obey unauthorized cancellation requests while la…
SCRIBE: Scalable Contextual Representations of Interpretable Brain-Grounded Semantics— arXiv cs.AI
researchmodelssafety
Researchers introduce SCRIBE, a framework combining large language models with neuroscience principles to generate contextual semantic representations of text. The model predicts human brain activity…
Co2In: a Bi-level Memory Incremental Learning Framework with Knowledge Encoding, Consolidation, and Integration— arXiv cs.AI
research
Co2In is a bi-level memory architecture for incremental learning that separates knowledge acquisition (working memory) from persistent storage (long-term memory), inspired by human brain mechanisms.…
Probe-and-Resume: When and How to Evaluate Agent Execution— arXiv cs.AI
research
Researchers present Probe-and-Resume, a test-time search method that improves agent performance in terminal-based environments by having a critic evaluate trajectories only at claimed completion poin…
ARGUS: An Agenda-Driven Framework for Reliable User Simulation— arXiv cs.AI
researchtools
ARGUS is a neuro-symbolic framework for reliable user simulation in agent evaluation that uses a dynamic agenda stack to enforce goal tracking and policy compliance, addressing reliability issues in…
Evaluating Inquiry Learning in LLM-Based Agents: Hypothesizing, Reasoning, and Exploration— arXiv cs.AI
researchmodels
Researchers evaluate how well large language models can perform scientific inquiry—hypothesis generation, exploration, and causal reasoning—in a unified framework using a modified Countdown task. Tes…
Prism-1: Position- and Rival-Independent Scoring for System One Models— arXiv cs.AI
researchmodels
Researchers propose Prism-1, a method for extracting decisions from language models that satisfies Luce's choice axiom—ensuring that option order and irrelevant alternatives don't affect model output…
CART: Compatible Routing for Multimodal Continual Instruction Tuning— arXiv cs.AI
research
This paper proposes Compatible Routing (CART), a method for training multimodal language models on sequential tasks while minimizing catastrophic forgetting. CART maintains task-specific experts in b…
DINO-FIL: Multi-Layer Transformer Fusion for Generative Inpainting Forgery Localization— arXiv cs.AI
researchsafety
DINO-FIL is a transformer-based framework using multi-layer fusion from pretrained DINOv3 ViT features to detect and localize pixel-level forgeries in AI-generated images. The method generalizes well…
Momentum, Commentary Tone, and Earnings Surprise in LLM Financial Forecasts— arXiv cs.AI
research
Researchers tested six large language models (0.49-7.61B parameters) with synthetic earnings announcements to identify which inputs drive financial forecasts, finding that non-informational commentar…
DemoStream: Batched Task-and-Motion-Plan Optimization for Automated Learning from Demonstration— arXiv cs.AI
researchapplications
DemoStream is a GPU-accelerated Task and Motion Planning system that generates high-quality demonstrations for robot learning by optimizing batch TAMP sampling and constraint satisfaction. The approa…
Confidence-Calibrated Inference Expansion for Evaluator-Guided Test-Time Reasoning— arXiv cs.AI
researchtools
Researchers propose a compute-aware controller for test-time reasoning that decides whether to trust current answers, use fallbacks, or spend more inference budget based on predicted success and cali…
Rethinking the Role of Tool Graph Memory in Multi-Turn Tool Use— arXiv cs.AI
research
A research paper evaluates tool graph memory, a component designed to improve LLM agent performance on multi-turn tool-use tasks by storing past tool transitions. The study finds that tool graph memo…
When Does Muon Actually Win? A Paired Data-, Time-, and Tuning-Budget Benchmark— arXiv cs.AI
research
Researchers benchmarked Muon, a data-efficient optimizer, against AdamW on language and vision models under controlled data, time, and tuning budgets on a single A100. Muon achieved 3–22% lower valid…
TIPS: A Tiny Inference-time Policy for Scaling in Diffusion Language Models— arXiv cs.AI
researchmodels
Researchers introduce TIPS, a 600K-parameter controller that learns optimal token visibility strategies during inference for masked diffusion language models, improving accuracy-compute trade-offs wi…
EvolvingBrain: An Adaptive Skill System for Multi-modal Agents Self-Evolution in Open Worlds— arXiv cs.AI
researchmodels
EvolvingBrain proposes an adaptive skill system that enables multi-modal agents to self-evolve in open worlds by extracting patterns from trajectories and organizing knowledge into a stratified skill…
Provably Efficient Risk-Sensitive Offline Reinforcement Learning via Optimized Certainty Equivalents— arXiv cs.AI
research
This paper develops offline reinforcement learning algorithms for risk-sensitive decision-making using optimized certainty equivalents (OCE), a general class of risk measures. The authors provide the…
Decoding What iOS Location Services Send to Apple: An LLM Is an Unreliable Analyst but a Useful Toolmaker— arXiv cs.AI
researchsafetytools
Researchers evaluated whether frontier LLMs (Claude Opus 4.8, Codex GPT Sol 5.6) can reverse-engineer protobuf message schemas from iPhone binaries sent to Apple's location services. The models produ…
AVALON: Zero-Shot Indoor Navigation with Human-Seeded Topological Memory— arXiv cs.AI
researchapplications
AVALON is a zero-shot indoor navigation framework that enables robots to navigate unmapped buildings using human-provided floor plans and spatial cues, combined with real-time human-robot memory upda…
Asking Smarter Questions: Targeted Active Preference Learning for Underspecified Rewards— arXiv cs.AI
researchsafety
Researchers propose a method for robots to identify underspecified reward features in human feedback and ask targeted preference queries with explanations to resolve them, rather than asking humans t…
Decoding Naturalistic Speech from fMRI with Temporal Modeling and Multimodal Alignment— arXiv cs.AI
research
Researchers introduced NOEMA, a temporal convolutional decoder that reconstructs naturalistic speech from fMRI brain activity by aligning neural measurements with multimodal (text, audio, speech) sti…
CodeHealth Predicts the Token Cost of Agentic Code Review: A Pre-Registered Study of 465 Agent-Authored Pull Requests— arXiv cs.AI
researchtools
Researchers conducted a pre-registered study of 465 agent-authored pull requests across 246 repositories to measure whether CodeHealth (a code quality metric) predicts token costs in multi-agent LLM…
FIRM: Fisher-Informed Retention Merging for Federated LLM Unlearning— arXiv cs.AI
researchsafety
FIRM proposes a federated framework for removing targeted knowledge from distributed LLMs while preserving utility, using Fisher statistics to stabilize aggregation of conflicting local objectives. T…
AVC-Reward: Event-Centric Audio-Video Consistency Evaluation— arXiv cs.AI
researchmodelstools
Researchers propose AVC-Reward, an event-centric evaluation framework for assessing audio-video consistency in joint generation models that handles multi-event compositions and cross-modal relations…
Motion- and Gravity-Aware Feedforward 3D Reconstruction— arXiv cs.AI
researchmodels
Researchers introduce a motion- and gravity-aware feedforward transformer for 3D reconstruction that uses optical flow and gravity direction as privileged training signals, then operates with standar…
Why Shallow Neural Networks Learn Sinusoidal Components First: An Orthogonal Polynomial Dynamics Framework— arXiv cs.AI
research
Researchers present a theoretical framework explaining why neural networks learn low-frequency sinusoidal components before high-frequency ones (spectral bias), using orthogonal polynomial dynamics a…
Following Delegation Traces to Study Human Oversight in Interaction with Multimodal AI— arXiv cs.AI
researchsafety
Researchers studied 56 AI-literate participants interacting with multimodal AI systems across 392 video question answering tasks, finding that post-interaction trust strongly predicted delegation cho…
Twice Upon a Time: Complementary Representations for Time-Series Reasoning— arXiv cs.AI
researchmodels
Researchers introduce a dual-view representation approach for time-series language models, combining line plots (preserving absolute values) with normalized delay embeddings (exposing temporal struct…
SIEVE: Overcoming Topological Obstruction in Equivariant Self-Supervised Learning— arXiv cs.AI
research
Researchers identify a fundamental theoretical incompatibility between masked autoencoding and SE(3)-equivariant networks for 3D point clouds, where standard masking causes topological collapse. They…
Full-Duplex Interaction with Speech, Expressions and Gestures— arXiv cs.AI
researchmodels
Researchers present ELLSA 2, an end-to-end full-duplex dialogue model that generates speech, gestures, and facial expressions simultaneously during conversation. The system uses contextual semantic g…
Optimal Hydra: Model Heterogeneous Federated Learning for LLMs using Data-Informed Hydra-Nets— arXiv cs.AI
researchmodels
Researchers propose Optimal Hydra, a federated learning framework that creates client-specific sub-networks for LLM fine-tuning by selecting attention heads based on local data distributions. The app…
WorldCast: Exploring Intelligent Visual Generation in World State Prediction— arXiv cs.AI
research
Researchers introduce WorldCast, a benchmark of 320 test cases for evaluating multimodal generative models on multi-step world state prediction—the task of forecasting successive future visual states…
When Is Muon Optimal? Radial Fragility and Minimax Geometry— arXiv cs.AI
research
This paper analyzes the Muon optimizer by introducing the Random-Radius Isotropic Curvature Model (RR-ICM) to explain when gradient orthogonalization is optimal. The authors show that atomless radial…
Does On-Policy Distillation Need to Complete Every Rollout?— arXiv cs.AI
research
Researchers propose Stochastic Truncated On-Policy Distillation (STOPD), a method that randomly truncates student rollouts during on-policy distillation and uses blockwise inverse-probability weighti…
X-Fly: Learning to Reach and Stop in Full-Chunk Aerial Vision-Language Navigation— arXiv cs.AI
researchapplications
X-Fly is a staged training framework for aerial vision-language navigation that executes full action chunks before replanning to reduce model inference costs (195ms to 42ms per waypoint). The method…
When is a better LLM worth paying for? Stochastic Optimization with Switching Mixed-Fidelity Oracles— arXiv cs.AI
researchtools
Researchers propose algorithms for deciding when to use cheaper versus expensive LLMs in data curation tasks by adaptively switching between them based on downstream optimization performance. The fra…
ToolBottleneckBench: Measuring Tool-Agent Performance under Controlled Pressures and Fixed Harness Budgets— arXiv cs.AI
researchtools
ToolBottleneckBench is a new benchmark that isolates how different system constraints (tool catalog size, menu visibility, workflow constraints, and record encoding) affect tool-agent performance acr…
Can Visual Generation Expand What Unified Multimodal Models Know?— arXiv cs.AI
researchmodels
Researchers introduce Knowledge Transfer Probing to test whether unified multimodal models can learn new knowledge through image generation and apply it to text-based question answering. They find un…
Wolof-Adapted Industry-Ready Open-Weight Language Models— arXiv cs.AI
modelsopen-sourceapplications
Researchers have developed open-weight language models (0.6B-14B parameters) specifically adapted for Wolof, a language spoken by 23 million people primarily in Senegal. The models use instruction tu…
MetaExp: Learning to Self-Improve through Experience via Verifiable Cross-Task Environment Scaling— arXiv cs.AI
researchmodels
MetaExp is a training framework that enables language models to learn reusable experience across diverse tasks by training on verifiable task families with controlled structure, rather than isolated…
Money-Mule Detection as a Relational Learning Problem: A Study on the RBIH NFPC Dataset— arXiv cs.AI
researchapplications
Researchers developed a heterogeneous graph neural network model to detect money-mule accounts by scoring accounts directly from their relational neighborhood in banking data, achieving 90% recall at…
Condition-Contrastive Extrapolation for On-Policy Distillation of Language Models— arXiv cs.AI
researchmodels
Researchers propose Condition-Contrastive Extrapolated On-Policy Distillation (CCE-OPD), a method that improves language model distillation by making teacher supervision sensitive to question conditi…
Progress Reward Model for Reinforcement Learning via Large Language Models— arXiv cs.AI
researchmodels
A new paper proposes PRM4RL, a framework combining task decomposition via LLMs with a progress-based reward function for reinforcement learning. The method addresses sparse reward problems in long-ho…
Improving Environment-Agnostic Coding Agents: The Role of Scaffolding— arXiv cs.AI
researchtools
Researchers investigated whether scaffolding (developer-supplied procedures) remains beneficial for language model coding agents as models improve. Testing a Codex-based agent across three benchmarks…
Mine, Sample, Render: Test-Time Diversity Alignment Of Generative Models— arXiv cs.AI
researchmodels
Researchers propose Mine-Sample-Render (MSR), a test-time method that improves diversity in generative models by first extracting a categorical schema from real data, then sampling attribute configur…
DeciCAD: Harness-Grounded Learning of Action Policies for CAD Agents— arXiv cs.AI
researchtoolsapplications
DeciCAD introduces a domain-native harness for CAD agents that exposes semantic actions and geometric constraints, enabling specialized policy learning through supervised fine-tuning and reinforcemen…
Rubric Judgments Lack Uncertainty: Conformal Weighting for LLM Evaluation— arXiv cs.AI
researchsafety
Researchers propose Uncertainty-aware Rubric Weighting (URW), a method that quantifies confidence in individual rubric judgments from LLM verifiers and weights them accordingly during evaluation. The…
AgentEval: Evidence-Bounded Assessment of Stateful Agents— arXiv cs.AI
researchtoolssafety
AgentEval is a system for evaluating stateful agents that tracks evidence and state throughout task execution to distinguish between task completion and actual obligation satisfaction. The approach s…
Learning Intrinsic Dynamics in Latent World Models— arXiv cs.AI
research
Researchers propose that action predictability—reconstructing actions from consecutive latent states—helps world models preserve control-relevant dynamics despite visual variations. They introduce Co…
Watts If? A Benchmark for Counterfactual Reasoning About the AI Data Center Buildout— arXiv cs.AI
researchsafetytools
Researchers introduce DCBench-440, a benchmark for testing whether large language models can reason about counterfactual decisions in AI data center development (grid vs. on-site power, cooling types…
X-RIC: An Agentic RAN-Intelligent Controller Benchmark for Enhancing Reinforcement Learning— arXiv cs.AI
researchinfratools
Researchers propose X-RIC, a framework combining reinforcement learning with agentic reasoning for O-RAN's centralized RAN Intelligent Controller, adding explainability and strategic planning to netw…
Self-Guided LLM Reasoning for Error Diagnosis in Vehicle Routing Problems— arXiv cs.AI
researchapplications
Researchers propose a self-guided reasoning framework where LLMs autonomously generate code to detect and correct errors in Vehicle Routing Problem instances, addressing limitations of existing appro…
When Reality Outpaces Cognition: A New Scaling Principle for Reliable Intelligence— arXiv cs.AI
researchsafety
A new arXiv paper proposes effective Speed of Coherent Cognition (S_C^eff) as a benchmark measuring how quickly reality can change before AI systems fail to perceive and act reliably. The authors tes…
L2G-SR: Learning to Guide LLM-based Symbolic Regression— arXiv cs.AI
researchmodels
L2G-SR proposes using a learned structural verifier alongside numerical error to guide LLM-based symbolic regression, achieving state-of-the-art results on LLM-SRBench. The method uses tree edit dist…
When Correctness Feedback Erases Evidence: Phase-Separated Completeness Auditing in Verus— arXiv cs.AI
researchtoolssafety
This paper demonstrates that providing correctness feedback to LLMs during program verification paradoxically reduces discovery of specification-completeness violations in Verus/Rust, with independen…
CF-AVSS: A Coarse-to-Fine Audio-Visual Speech Separation Framework with Target-Speaker Refinement and Post-Fusion for Real-World AVSE Challenge— arXiv cs.AI
researchapplications
Researchers propose CF-AVSS, a coarse-to-fine audio-visual speech enhancement framework that uses progressive visual streams to separate speech in noisy and reverberant conditions. The system achieve…
Time to Fake It: Continual Machine Learning Intensifies Unfairness in Deepfake Detection— arXiv cs.AI
researchsafety
Researchers analyzed how continual learning methods affect fairness in deepfake detectors as they train on evolving generation techniques. They found that while these methods improve detection accura…
Decentralized LLM Workflows via Sequential Auctions— arXiv cs.AI
researchmodels
Researchers present a framework for organizing AI agents in decentralized markets using sequential auctions to complete multi-step tasks, moving beyond prior theory that assumed perfect agent clones.…
ACEdit: Stable Multi-Turn Image Editing via Asymmetric Conditioning— arXiv cs.AI
researchmodels
ACEdit proposes a training scheme for multi-turn image editing that stabilizes quality across successive edits by using self-forcing with asymmetric teacher-student conditioning. The method addresses…
Qwen-Audio-Jev: A Unified Interface for Audio Decisions— arXiv cs.AI
modelsresearch
Researchers introduce Qwen-Audio-Jev, an audio language model augmented with a decision head that performs selection, binary judgment, and scoring tasks directly without autoregressive text generatio…
Open-world AI: a Tale of Three Requirements— arXiv cs.AI
research
A research paper identifies three core requirements for open-world AI systems—inference-time learning, rich representation, and predictive disentanglement—and tests these requirements on language mod…
GraphBEAT: Graph-based Bucketized Event Aggregation Over Time for Early Risk Stratification of Gastrointestinal Cancers— arXiv cs.AI
researchapplications
Researchers introduce GraphBEAT, a temporal graph neural network framework for predicting gastrointestinal cancers (pancreatic, liver, intestinal) from sparse electronic health records with 12-36 mon…
Router Drift Control: Mitigating Off-Policy Degradation in Mixture-of-Experts Reinforcement Learning— arXiv cs.AI
research
Researchers propose Router Drift Control (RDC), a method to mitigate off-policy degradation in Mixture-of-Experts reinforcement learning by detecting routing disagreement between rollout and training…
One Is Not Enough: How People Use Multiple AI Models in Everyday Life— arXiv cs.AI
researchapplications
A research study examining how 10 users orchestrate work across multiple AI models reveals they construct context-dependent hierarchies and switching patterns based on perceived model strengths, trus…
Forecast lead time and the timing of crew deployment for weather-driven power outages using AI weather models— arXiv cs.AI
researchapplicationsindustry
Researchers evaluated AI weather models (GraphCast and FourCastNet) for predicting storm-driven power outages 1-5 days ahead to optimize crew deployment timing. The study found forecast accuracy degr…
Poison Valleys: Diagnosing Path-Dependent Failure in LLM Code Repair— arXiv cs.AI
researchsafety
Researchers conducted a study on 1,844 LLM code-repair requests to understand how incorrect assumptions propagate through the model's reasoning process (hypothesis, plan, implementation) and cause co…
RecSoup: Efficient Model Merging Approach for Controllable Multi-Objective Recommendation— arXiv cs.AI
researchmodels
RecSoup proposes a lightweight model merging approach for controllable multi-objective recommendation systems that dynamically adjusts trade-offs between objectives at inference time without retraini…
CARE-SD: Explainable Reasoning Architecture for Local SLM Self-Disclosure Analysis in Textual Data— arXiv cs.AI
researchsafetytools
Researchers developed CARE-SD, a four-module architecture for analyzing self-disclosure statements in cybersecurity interviews using small local language models (1B-3B parameters). The system improve…
Where, How, and When: Query-Guided Grounding for Video Understanding— arXiv cs.AI
researchmodels
This thesis presents four neural architectures for query-guided video understanding: QDETRv for one-shot object localization, TOViT for few-shot video object detection, a relation network for few-sho…
A Jury, Not an Electorate: Label-Free Reliability Estimation for Ensembles of Language Models— arXiv cs.AI
researchmodels
Researchers propose a label-free ensemble method for language models based on inferring expert reliability from agreement patterns rather than labeled data, using an EM-based jury model. The approach…
Probing Negation Understanding in Vision-Language Models With Object Attention— arXiv cs.AI
researchsafety
Researchers analyzed how vision-language models process negation by examining attention patterns and comparing them to human gaze behavior across synthetic and realistic datasets. They found that VLM…
Can LLMs Tell Stories Out of Order? A Corpus and Benchmark for Nonlinear Narrative Generation— arXiv cs.AI
researchmodels
Researchers created a benchmark of 83 annotated literary works to evaluate whether large language models can generate nonlinear narratives (stories where plot order differs from chronological order).…
PriorityFEC: Deadline-Aware Streaming FEC for KV-Cache Handoff in Edge Multi-Agent Systems— arXiv cs.AI
researchinfra
PriorityFEC proposes a method for efficiently transferring key-value (KV) cache between LLM agents in edge systems by prioritizing high-utility KV blocks and using deadline-aware forward erasure corr…
RICE-Alpha: Reliability-Informed Correction with Event Graphs for LLM-Agent Stock Forecasting— arXiv cs.AI
researchapplications
RICE-Alpha is a new framework for stock forecasting using LLMs that models temporally dependent corporate events through event graphs and reliability-calibrated corrections. On Nasdaq-100 and Hang Se…
FedSumUp: User-Auditable Privacy-Aware Representation Sharing for Resource-Constrained Federated Intelligence— arXiv cs.AI
researchsafetyinfra
FedSumUp proposes a privacy-preserving federated learning method using learnable lossy compression to generate compact, user-interpretable representations on clients while reducing computational over…
KULTURE Bench: A Benchmark for Assessing Language Model in Korean Cultural Context— arXiv cs.AI
researchtools
Researchers introduced KULTURE Bench, a Korean-specific benchmark for evaluating language models' cultural understanding through cultural news, idioms, and poetry at multiple linguistic levels. The w…
In-Context Preference-Conditioned Pareto Set Learning from Limited Offline Data— arXiv cs.AI
research
Researchers propose ICPSL, a framework using foundation-style hypernetworks to learn preference-conditioned Pareto optimal solutions from limited offline data in a single forward pass, eliminating th…
Independent Data-Integrity Verification for LLM-Driven Functional Coverage Closure in Hardware Design— arXiv cs.AI
researchtoolsapplications
Researchers describe an LLM agent that achieves 100% functional coverage on three hardware designs (FIFO, UART, AHB2APB bridge) by reading coverage reports directly from simulation rather than trusti…
ClaimScope: Label-Blind Witness Verification for Mathematical Claim Auditing— arXiv cs.AI
researchtools
ClaimScope is a new 72-task benchmark for evaluating how well language models audit mathematical claims by generating verdicts, scopes, counterexamples, and repairs without access to reference soluti…
Agent Status Protocol (ASP): A Lightweight Specification for State Reporting in Multi-Agent Systems— arXiv cs.AI
open-sourceinfraresearch
Researchers propose Agent Status Protocol (ASP), a lightweight JSON-based specification for multi-agent systems to report operational state via HTTP and Server-Sent Events, with liveness detection vi…
CIM-Embed: Learning Instruction-Conditioned Multimodal Embeddings for Clustering— arXiv cs.AI
researchopen-source
Researchers introduce CIM-Embed, a framework that learns instruction-conditioned multimodal embeddings for clustering multimodal objects (images + text) under different grouping criteria specified by…
RAG-RST: Retrieved Exemplars as In-Context Discourse Schema for RST Parsing— arXiv cs.AI
researchmodels
Researchers propose RAG-RST, a retrieval-augmented approach for Rhetorical Structure Theory parsing that provides discourse schema as in-context examples rather than through model training. The metho…
Anticipatory Commonsense with Moving Objects: An ASP-Based Framework for Predictive Reasoning with Interacting Trajectories— arXiv cs.AI
researchmodels
Researchers present an Answer Set Programming (ASP) framework for predicting interactive motion trajectories in dynamic environments, combining symbolic reasoning with energy-based optimization and e…
TRACE: Topology-Aware Reconnection via Prediction-Adaptive Connectivity Enforcement for Thin-Structure Segmentation— arXiv cs.AI
research
TRACE is a new topology-aware loss function for segmentation of thin, elongated structures (like blood vessels or neural fibers) that adapts supervision based on the current model predictions rather…
ESG-Kor: A Korean Dataset for ESG-related Information Extraction and Practical Use Cases— arXiv cs.AI
researchopen-sourcetools
Researchers released ESG-Kor, a Korean-language dataset of 118,946 manually labeled sentences for extracting Environmental, Social, and Governance information from corporate sustainability reports. T…
Real-time Chatbot Romance Scam Detection in Stage Classification— arXiv cs.AI
applicationssafetytools
Researchers developed a real-time chatbot using LLMs to detect romance scams via Line Messaging API by classifying them into seven social-science-derived stages. The system provides users with stage-…
DST-Net: Decomposed Spatio-Temporal Network for Time Series Forecasting— arXiv cs.AI
research
DST-Net proposes a decomposed spatio-temporal architecture for time series forecasting that applies distinct modeling strategies to trend and seasonal components—using latent graph modeling for trend…
Matoi: Selective Revalidation of Consequence Evidence for Agent Authorization— arXiv cs.AI
researchsafetyapplications
Matoi proposes a system for selective revalidation of consequence evidence in agent authorization, separating reusable evidence from execution authority and history. Empirical evaluation on synthetic…
Bridging the Gap: Converting Read Text to Conversational Dialogue— arXiv cs.AI
researchmodels
Researchers propose PACC, a neural approach using HiFi-GAN to convert read speech into conversational speech by modifying prosodic features like intonation and rhythm. The method shows improvements i…
Truth for Believable AI: Expressed Doubt, Provenance, and Belief Revision as an Engineerable Stance— arXiv cs.AI
researchsafety
Researchers implemented a behavior layer over language models to express uncertainty, track provenance, and support belief revision without modifying the base model. On a synthetic benchmark the syst…
OnBodyAI - Why the Next AI Needs a Body— arXiv cs.AI
researchmodels
A research article proposes 'OnBodyAI'—large language models deployed as embodied agents on personal devices—as the next frontier for AI development. The author argues that future AI competitiveness…
Verdict: Encode Once, Decide Many— arXiv cs.AI
researchmodels
Researchers present Verdict-MQ, an encoder model that answers multiple typed questions about a single input by encoding it once and using structured attention masks, rather than separate passes per q…
Active Learning of Fractional-Order Viscoelastic Model Parameters for Realistic Haptic Rendering— arXiv cs.AI
researchapplications
Researchers propose an active learning approach using human-in-the-loop optimization to determine parameters for fractional-order viscoelastic models that improve haptic rendering realism in medical…
Multi-sourced AI Tutor for Non-Native English Speaker Students— arXiv cs.AI
applicationsresearchtools
Researchers developed an AI tutor system that processes student questions in Mandarin Chinese and delivers bilingual responses with technical glossaries, using a specialized RAG pipeline. Evaluated o…
GNN-Guided Local Refinement for Efficient Layout Decomposition with DSA-MP— arXiv cs.AI
researchtools
Researchers propose a GNN-based local refinement approach for layout decomposition in integrated circuit manufacturing that combines directed self-assembly (DSA) with multiple patterning lithography…
CriticaLLM: Using LLMs as Autonomous Game Master for Tabletop Role-Playing Games— arXiv cs.AI
researchapplications
Researchers introduce CriticaLLM, a framework enabling LLMs to autonomously run tabletop RPG games as Game Masters, handling story generation, rule enforcement, and combat management. The system was…
Deep Reinforcement Learning for HPC Job Scheduling: A Statistical Evaluation— arXiv cs.AI
researchinfra
Researchers compared six deep reinforcement learning algorithms (PPO, A2C, DQN) with and without action masking for HPC job scheduling on real Slurm workload traces. On heterogeneous workloads, Maska…
Ontology Refinement with Embeddings and LLMs: A Case Study on CASTON— arXiv cs.AI
researchtoolsopen-source
Researchers propose a two-stage workflow for refining ontologies by combining fine-tuned embedding models with LLMs to identify and repair subsumption errors in domain hierarchies. Applied to CASTON…
Hierarchical Graph Convolutional Network Approach for Detecting Low-Quality Documents— arXiv cs.AI
research
Researchers propose a hierarchical graph convolutional network (HGCN) to detect low-quality documents by identifying internal inconsistencies and title-body misalignment, achieving 91.20% accuracy on…
Efficient Pairwise Difference Learning: A Comparative Study of Encoding Strategies and Training Pair Selection— arXiv cs.AI
research
This paper systematically evaluates design choices in Pairwise Difference Learning for Classification (PDC), a low-data learning technique that converts n samples into O(n²) training pairs. Across ~2…
The Delegation Dilemma: Economies of Scale, Accountability, and Polycentric Governance— arXiv cs.AI
researchsafety
This paper studies the tension between economies of scale and accountability in delegated governance through a game-theoretic model and LLM-based simulations. The authors show that exit threats can o…
Hierarchical Interpretable Multitask Audio Representation Learning: An Analysis of Event-Scene Relationships through the Information Bottleneck— arXiv cs.AI
research
This paper proposes a hierarchical conditioning approach for joint acoustic scene classification and audio tagging that predicts scene categories from intermediate event probabilities, rather than us…
Replicable Dueling Bandits— arXiv cs.AI
research
This paper introduces algorithms for dueling bandits that guarantee replicability—two independent runs with shared randomness must select identical comparison pairs in each round with high probabilit…
Exploring the Design Space of Permutation Diffusion LLMs— arXiv cs.AI
researchmodels
Researchers explore three different approaches to using diffusion models with language models for permutation generation tasks, focusing on recommendation reranking. The study reveals that direct ord…
When do quantum views help? Complementary classical-quantum representations in multi-view fusion— arXiv cs.AI
research
Researchers propose a classical-quantum multi-view fusion framework that combines classical features with simulated quantum features to improve predictions when both capture complementary information…
Uncertainty Quantification in Pairwise Difference Learning for Classification— arXiv cs.AI
research
Researchers propose a pairwise difference learning (PDL) approach for classification that predicts differences between instance outcomes rather than direct mappings, then derives predictions by avera…
Quantum Versus Classical Machine Learning for Neuropsychiatric Disorders: A Feature-Matched Multi-Cohort Evaluation— arXiv cs.AI
researchsafety
Researchers compared six families of quantum machine learning models against classical baselines on neuropsychiatric disorder classification using 4,308 subjects across four public cohorts. Quantum m…
AI-Driven Disaster Response and Displacement Monitoring— arXiv cs.AI
researchapplicationssafety
A research paper surveys AI applications in humanitarian tasks, including disaster response and displacement monitoring via satellite imagery analysis and machine learning. The work aggregates existi…
BridgeAfford: Bridging Query-Target Gap for Generalizable 3D Affordance Grounding— arXiv cs.AI
research
BridgeAfford proposes a framework for 3D affordance grounding that addresses poor generalization to unseen objects by reducing the semantic gap between query features (interaction intents) and target…
ISW-Agent: Integrating LLM-Driven Dialogue with n8n Workflow for Internal Solitary Wave Remote Sensing Image Processing— arXiv cs.AI
researchtoolsapplications
Researchers present ISW-Agent, a system that combines LLM-driven dialogue with n8n workflow automation to process internal solitary wave remote sensing imagery end-to-end. The system achieved 94.2% t…
Learning to Search for Executable Patternmaking from Certified Programs— arXiv cs.AI
research
Researchers present SemDraft, a framework that learns to recover executable sewing pattern editing sequences from before-and-after pattern pairs by searching over editing actions and validating resul…