DANGERLENS: Probing Dangerous Representations Hidden by Safety Alignment— arXiv cs.AI
safetyresearch
Researchers introduce DANGERLENS, a probe that detects harmful concept accessibility in language model hidden states despite identical external safety behavior. The study shows that models with prior…
CoRedSkill: Stealthy Skill Composition Attacks via AST-Guided Injection— arXiv cs.AI
safetyresearch
Researchers introduce CoRedSkill, a method for embedding stealthy attacks into multi-skill AI agent workflows using abstract syntax tree rewriting. The technique achieves 98% attack success while mai…
The Agentic Scientist: Orchestrating Coding Agents for Automated Scientific (Co-)Discovery— arXiv cs.AI
researchtoolsapplications
Researchers present The Agentic Scientist, a distributed pipeline that orchestrates multiple coding agents to conduct autonomous computational research with human oversight. The system decomposes res…
From Training to Escalation: Forensic Analysis and Local Recreation of RL/RLHF-Induced Privilege Escalation in AI Agents (CVE-2026-65616)— arXiv cs.AI
safetyresearchtools
Researchers provide forensic analysis and reproduce CVE-2026-65616, a privilege escalation vulnerability affecting LLM agents during reinforcement learning training in OpenAI's ExploitGym evaluation…
SWE-Stream: Scaling Executable Training Data for SWE Agents via Progressive Construction and Elastic Streaming— arXiv cs.AI
researchtoolsmodels
SWE-Stream is a system that constructs high-quality executable training data for software engineering agents by progressively validating task candidates from 10,728 development repositories and elast…
CASI: Credit Assignment via Set Intervention for Agentic Retrieval— arXiv cs.AI
researchapplications
CASI is a new credit assignment method for multi-action retrieval systems that uses counterfactual set interventions and Shapley value aggregation to determine within-set attribution, addressing the…
Write Once, Plan Many: Building Reusable Visual Planning Systems from One Scene— arXiv cs.AI
researchtools
Researchers present a training-free framework that generates reusable visual planning programs from a single example scene and domain specification, without paired solutions. The system leverages VLM…
Teaching Agents in their Own Style: Blackbox Teacher Distillation for Web Agents— arXiv cs.AI
researchmodelstools
Researchers introduce WebKD, a distillation method where a teacher model edits only the failing portions of a student web agent's trajectories while matching the student's writing style, rather than…
EpiGuard: An Autonomous Infectious-Disease Early Warning Agent Platform— arXiv cs.AI
applicationsresearchtools
EpiGuard is an autonomous AI agent platform that monitors surveillance data across 50+ infectious diseases to provide early warnings of outbreaks. It adapts its analysis across different data availab…
LAGO: Language Agents with Reusable Growing Options— arXiv cs.AI
researchtoolsmodels
LAGO introduces a meta-agent framework that treats specialized language agents as reusable modular components (options) organized in a persistent library, enabling better long-horizon task decomposit…
LabOPBench: A Multimodal Benchmark for Evidence-Grounded Laboratory Decision Making— arXiv cs.AI
researchmodelsapplications
LabOPBench introduces a 756-question multimodal benchmark for evidence-based laboratory decision-making in organic synthesis and materials science, with expert-authored questions and rubrics. Current…
Articulation Upper Bounds for Language Models— arXiv cs.AI
research
Researchers measured the limits of language by testing how well concepts can be transmitted between LLMs at different scales, introducing a prompt optimization algorithm and a transmission recovery m…
MCAIR: Monitoring and Correcting Agent Unfaithfulness via Internal Representations— arXiv cs.AI
researchsafety
Researchers propose MCAIR, a runtime monitoring system that detects and corrects agent unfaithfulness—when LLM-based agents fail to disclose deviations from their original execution procedures. The w…
Learning When to Trust Medical Knowledge: Regression-Safe Routing for Clinical Reasoning— arXiv cs.AI
researchsafetyapplications
Researchers developed Med-TrustRouter, a lightweight routing mechanism that learns when clinical reasoning LLMs should rely on external medical knowledge versus their base model, addressing the probl…
Are AI Decision Models Biased? A Case Study of Jev— arXiv cs.AI
researchsafety
Researchers evaluated Jev, an AI decision model designed for fast structured decisions, across 41,500 responses testing cognitive biases and found it avoids many human biases in selected answers but…
Beyond Correctness: Compute-Responsive Data Selection for Reasoning Model Post-Training— arXiv cs.AI
researchmodels
Researchers propose Compute-Responsive Data Selection (CRDS), a post-training method that identifies examples most responsive to additional test-time compute and prioritizes them for supervised fine-…
Memorizing Who, Knowing How: Benchmarking Personal and Procedural Memory for Persistent Assistants— arXiv cs.AI
researchtools
Researchers introduce PraxisMemBench, a new benchmark for evaluating persistent personal assistants that must maintain both personal knowledge (who the user is) and procedural memory (how to act). Ex…
When Protection Overrides Consequences: Understanding Protection Mode in Large Language Model Decision-Making— arXiv cs.AI
researchsafety
Researchers identified "Protection Mode" in frontier LLMs like GPT-5.6 Sol, where models prioritize protecting people over larger physical consequences (e.g., saving a child before a camera) on simpl…
STRATA: Self-Maintaining Structured Experience Graphs for Self-Evolving Language Agents— arXiv cs.AI
researchmodels
STRATA introduces a structured three-layer memory graph for LLM agents that organizes experiences into atomic skills, task workflows, and meta-frameworks with typed edges and self-maintenance mechani…
DeLoop: Repeated Actions as Self-Supervision for Long-Horizon Agents— arXiv cs.AI
researchmodelssafety
DeLoop is a self-supervised training method that reduces loops (repeated tool calls returning identical observations) in language-model agents by using the agent's own rollout history as negative/pos…
Scaling Discovery Requires Forgetting— arXiv cs.AI
researchmodels
Researchers found that LLM agent teams for automated discovery plateau when accumulating search history, causing agents to anchor to past attempts. Introducing periodic memory resets—discarding sessi…
The Wealth of Agents: Scaling Agentic AI Across Developers— arXiv cs.AI
researchmodels
Researchers propose multi-developer agentic systems (MDAs) where multiple developers collaboratively build specialized agents that combine to outperform single-developer systems. On the GAIA benchmar…
TreeFlex: Agents Navigating Their Own Search Trees— arXiv cs.AI
researchmodels
TreeFlex introduces a framework where agents maintain and navigate their own search trees during reasoning, allowing dynamic choice of which branch to explore rather than accumulating context linearl…
Beyond the Trial-and-Error Loop: Hybrid Projection and Automated Tuning for Distributed Training— arXiv cs.AI
researchtoolsinfra
Researchers introduce Primus Projection, a hybrid measurement and simulation framework that dramatically reduces the cost of tuning distributed LLM training configurations across hundreds of GPUs. By…
Sparse Transformer: a Controlled Testbed for Mechanistic Interpretability— arXiv cs.AI
researchsafetymodels
Researchers introduce Sparse Transformer, a 5-layer language model with sparse MLPs that forces neuron specialization, enabling direct evaluation of mechanistic interpretability methods against groun…
What Coding Agents Leave Unsaid: Measuring Task-State Gaps Between Execution and Reports— arXiv cs.AI
researchsafety
Researchers measured a 40.7-point gap in task-state recoverability when human supervisors rely only on coding agents' reports versus their full execution context, finding that report-only readers rec…
SwarmRSI: Recursive Self-Improvement through Evolving Scientific Agent Swarm— arXiv cs.AI
researchmodelstools
SwarmRSI introduces a framework for scientific agent swarms to achieve recursive self-improvement by evolving research strategies through experimental feedback and sharing useful components across ag…
Agentic Self-Awareness: Aligning Knowing, Acting, and Holding— arXiv cs.AI
researchsafety
Researchers introduce agentic self-awareness (ASA) as a framework measuring whether AI agents can correctly recognize task impossibility, defer appropriately, and maintain performance on feasible tas…
Neurosymbolic Learning of Interpretable World Models— arXiv cs.AI
researchsafety
Researchers propose Concept-Grounded World Models (CGWMs), a neurosymbolic approach that learns interpretable concept representations for world models by incorporating prior knowledge expressed in te…
L3M: A Verifiable Harness and Runtime for Self-Extending and Self-Improving Agents— arXiv cs.AI
infratoolssafety
L3M is a runtime for AI agents written in Lean 4 that enables self-extending behavior by treating computation across multiple agents, tools, and humans as a first-class object with Git-backed state m…
Euler: Towards Human-Centered Multi-Agent Mathematical Discovery— arXiv cs.AI
researchtoolsapplications
Euler is a human-centered multi-agent mathematical discovery system that organizes mathematical statements, proofs, assumptions, and exploration pathways into structured research graphs rather than f…
ActionOPD: Distilling Decisions, Not Tokens— arXiv cs.AI
researchmodels
ActionOPD is a new method for distilling large language model agents into smaller models by aligning them at the action level rather than token level, avoiding issues with state-distribution mismatch…
ReproBAIT: Can coding agents build their own biological AI tools?— arXiv cs.AI
researchsafetymodels
A new evaluation (ReproBAIT) tests whether coding agents can reconstruct published biological AI tools from limited descriptions while sandboxed without internet access. The study finds that LLM-base…
RuleTrek: Benchmarking Agentic Rule Reasoning in Complex Scenarios— arXiv cs.AI
researchtools
RuleTrek is a new benchmark for evaluating how well large language models can identify and apply domain-specific rules across distributed information in complex scenarios. Testing 11 frontier LLMs re…
Are You Reading the Right Thing? Rethinking Internal Signals for Prompt-Injection Detection— arXiv cs.AI
researchsafety
Researchers evaluated five methods for detecting prompt injections using internal language model signals on two large models, finding that attention-based approaches better detect successful attacks…
SERO: Learning from Executable Oracle Continuations for Long-Horizon Agentic Reinforcement Learning— arXiv cs.AI
researchmodels
SERO introduces a method for long-horizon reinforcement learning that uses oracle continuations from actually-reached states to provide training supervision, with policy-reachability weighting to acc…
ixi-sLM: On-Device sLM for AI Call Agent Service— arXiv cs.AI
modelsapplicationsinfra
Researchers present ixi-sLM, a 1.2B Korean small language model optimized for on-device execution on mobile NPUs through domain-adaptive modeling and task-aware quantization. The system achieves 94%…
Argument Graph Uncertainty: Quantifying Uncertainty from the Logical Structure of Reasoning Chains— arXiv cs.AI
researchsafety
Researchers propose Argument Graph Uncertainty, a post-hoc framework that estimates LLM confidence by analyzing the logical structure of chain-of-thought reasoning rather than sampling or token signa…
How GUI Agents Fail: GraceBench for Execution Stability Diagnosis— arXiv cs.AI
researchsafetytools
GraceBench is a new evaluation framework for GUI agents that diagnoses execution stability beyond task completion, using UI Transition Graphs to track step-level behavior like deviations, loops, and…
Co-Evolving Acting and Forecasting for Language Agents— arXiv cs.AI
researchmodels
Researchers propose CoPE, a training method that improves language agent learning by combining action forecasting with policy learning, enabling the model to supervise itself on intermediate steps du…
Who's Behind the Text? Robust Model Authorship Attribution— arXiv cs.AI
researchsafety
Researchers propose ATLAS, a method for identifying which large language model generated a given text by comparing likelihood scores across source models rather than direct classification. The approa…
Structure Before Intelligence: Gating What Agentic Scientific Workflows Release— arXiv cs.AI
researchsafetytools
Researchers propose a gated release harness around scientific workflow DAGs that uses validation, risk estimation, and LM agents to prevent invalid results from being released, demonstrating on synth…
MemCode: Discrete Semantic Representations for Long-Term Agent Memory— arXiv cs.AI
researchmodels
MemCode proposes a discrete semantic representation system for long-term LLM agent memory that converts dialogue into atomic memory units mapped to a learnable codebook, enabling both vector search a…
Evaluating Agentic Capabilities Before Post-Training: An End-to-End Approach— arXiv cs.AI
researchtools
Researchers introduce ASAP, a framework for evaluating agentic capabilities in base LLMs during pretraining using end-to-end tasks rather than static proxies. The method reveals how interactive behav…
Task-State Governance: A Virtual Machine for Verifiable and Recoverable GUI Agent Execution— arXiv cs.AI
researchtoolssafety
Researchers introduce task-state governance, a framework that represents GUI agent workflows as editable contracts linking desired outcomes, preserved state, and observed application state. The appro…
AFCS: Atomic-Fact-Conditioned Synthesis for Factual Knowledge Pretraining— arXiv cs.AI
researchmodels
AFCS is a synthetic data generation pipeline that extracts atomic facts from source documents before synthesis, ensuring secondary entities and precise attributes are preserved during pretraining. Ex…
Provable Transformer Context Compression for In-Context Reinforcement Learning— arXiv cs.AI
researchmodels
This paper establishes theoretical bounds on context compression for transformers performing in-context reinforcement learning via temporal difference (TD) evaluation, proving that the minimum requir…
Discovering Undisclosed Concepts in LoRA Adapters for Diffusion-Based AIGC Models— arXiv cs.AI
researchsafetytools
Researchers propose a method to discover what concepts undisclosed LoRA adapters for diffusion models were trained to generate, addressing moderation risks. The approach uses subspace alignment to ef…
Modeling and Understanding Neural Weights of Humanoid Policies with Diffusion— arXiv cs.AI
researchmodelsapplications
Researchers developed a latent diffusion model to directly generate neural network weights for humanoid robot control, training on policies learned via RL for the Unitree G1 robot. The approach produ…
PrismaCoder: Pushing the Boundaries of Open Multimodal Code Intelligence through Decomposed Rubric Synthesis— arXiv cs.AI
modelsresearchopen-source
Researchers introduce PrismaCoder, an open-source multimodal code intelligence system trained on PrismaForge, a new dataset of 2M examples curated using a hierarchical rubric that evaluates code corr…
Memory-Efficient Subspace Optimization with Residual Correction: From Momentum SGD to Muon— arXiv cs.AI
researchmodels
This paper proposes memory-efficient subspace optimization methods (SMRC and PMRC) that use residual correction to improve upon previous subspace-based approaches for LLM training. The key innovation…
StateCraft: Augmenting Video World Models with Abstract-State Reasoning— arXiv cs.AI
researchmodels
StateCraft decouples abstract state tracking from visual generation in video world models, using a language model paired with a video diffusion model to maintain consistency over long interactions. T…
GeoCloak: An Agentic Framework for Semantics-Preserving Geolocation Privacy via Peripheral Clue Manipulation— arXiv cs.AI
safetyresearch
GeoCloak is a new framework that protects user privacy in photos by strategically editing geolocation clues (like street signs) that vision-language models use to infer location, reducing city-level…
MorphAct: Robot Policies Morphing with the Worlds They Act In— arXiv cs.AI
researchmodels
MorphAct introduces a contextual parameterization framework that dynamically adjusts robot policy parameters based on interaction contexts, addressing the problem of competing parameter updates acros…
Verification Turns Capability Gaps into Routing Decisions: Small Language Models for Clinical Documentation— arXiv cs.AI
researchapplicationssafety
Researchers demonstrate that a 4B-parameter small language model with a verification and refinement pipeline achieves comparable factual accuracy (0.984 vs 0.992 entailment) to frontier models for cl…
Compiling Expert Procedures into Verified Programs: When a Small Model Is Enough— arXiv cs.AI
researchmodels
Researchers demonstrate that procedural knowledge extracted from expert traces can be compiled into explicit programs (PLVR) that enable small 8B models to match or approach larger models' performanc…
Plant Memory: Technical Language Processing and a Verified Maintenance Ontology for Agent-Accessible Failure Knowledge from SAP PM Work Orders— arXiv cs.AI
toolsapplicationsresearch
Plant Memory applies technical language processing and large language models to extract structured maintenance knowledge from SAP work orders, building a verified ontology that serves both human tech…
Beyond Equation Search: Conjecture-Guided Discovery with SCOPE and ClosureBench— arXiv cs.AI
researchtools
Researchers introduce SCOPE, a multi-agent framework for scientific discovery that guides equation discovery through conjectures rather than optimizing equations alone, and ClosureBench, a 60-task be…
Player-1: Learning to Play from Synthetic Experience— arXiv cs.AI
researchmodelsapplications
Researchers introduced Player-1, a post-training framework that uses synthetic game environments and reinforcement learning to improve AI agents' ability to play games, achieving 38.33% success on Mi…
Human Demonstrations as Post-Training Data for Any Robot Body: Retargeting the Object, Not the Hand— arXiv cs.AI
researchapplications
A new approach enables human demonstrations recorded on simple devices (like phones) to serve as post-training data for robots with different body structures by retargeting object motion rather than…
Doc2Harness: Benchmarking LLMs on Long-Horizon Harness Engineering of Coding Agents— arXiv cs.AI
researchtoolsmodels
Doc2Harness is a new benchmark that evaluates how well LLMs can engineer harnesses for coding agents over long-horizon tasks, rather than just measuring task success under fixed harnesses. The benchm…
Cross-Model In-Context Neurofeedback Reveals (Limited) Activation Control in LLMs— arXiv cs.AI
researchsafety
Researchers tested whether large language models can learn to control their own activation patterns through in-context neurofeedback, finding that only 3 out of 10 models (9B-32B parameters) show sma…
PerSo: Investigating Facets of LLMs' Social Intelligence in Real-World Persuasion Scenarios— arXiv cs.AI
researchsafety
Researchers introduce PerSo, a benchmark for evaluating LLMs' social intelligence using real-world persuasion dialogues from Reddit's ChangeMyView community. The benchmark tests both cognitive intell…
VaryST: Generalizable Spatio-Temporal Forecasting Beyond the Closed-World Setting— arXiv cs.AI
researchmodels
VaryST proposes a spatio-temporal forecasting model that generalizes beyond training conditions by dynamically constructing context from available observations rather than relying on fixed node repre…
Your Transformer Is Secretly Elastic— arXiv cs.AI
researchmodels
Researchers propose a retraining-free method to convert pre-trained transformers into elastic models that dynamically adjust computational cost during inference by jointly ranking neurons and attenti…
Which Knowledge Should LLMs Trust? Uncovering a Generalizable Steering Direction Across Knowledge Conflicts— arXiv cs.AI
researchsafety
Researchers identified a single linear direction in LLM residual streams that controls whether models prioritize internal parametric knowledge or external contextual knowledge. This steering directio…
FoveaVLA: Foveation Actions Replace the Wrist Camera in VLAs— arXiv cs.AI
researchmodels
FoveaVLA replaces traditional wrist cameras in vision-language-action (VLA) models with learned foveation actions that dynamically crop high-resolution regions from external cameras, achieving 77.1%…
MUCO: A Mani-Unit-Structured Manipulation Framework for Compositional Long-Horizon Robot Tasks— arXiv cs.AI
researchapplications
MUCO introduces a framework that decomposes long-horizon robot tasks into reusable manipulation units (MUs), each with defined success/failure criteria, enabling policy composition without joint trai…
ListQA: A Benchmark for Evaluating List-Formatted Factual Knowledge Retrieval in Large Language Models— arXiv cs.AI
researchtools
ListQA is a new benchmark of 9,045 human-curated questions evaluating whether LLMs can retrieve and organize multiple facts into structured lists (flat or hierarchical), with an evergreen design to a…
Embedding Geometry Shapes LLM Routing: Tracing Failure Modes and Improving Generalization— arXiv cs.AI
researchmodels
Researchers analyzed why semantic embedding-based LLM routers fail on unseen tasks and found that embedding geometry is an unreliable proxy for model selection. They propose using hidden states from…
Identity, Incentives, and Cooperative Behavior for AI: Reverse-Engineering Multiagent Cooperation— arXiv cs.AI
researchsafety
Researchers present a framework for reverse-engineering why AI agents cooperate or coordinate with each other, addressing the ambiguity of identity and motivation in multiagent systems. They show tha…
Learning from Self-Generated Responses: Reliability--Uncertainty Self-Distillation for Reasoning— arXiv cs.AI
researchmodels
Researchers propose Reliability-Uncertainty Self-Distillation (RUSD), a method where language models learn from their own generated responses by weighting distillation losses based on both solution r…
Into the Unknown: Benchmarking AI agents on open scientific research problems— arXiv cs.AI
researchtoolssafety
Researchers introduced Into the Unknown, a benchmark that tasks AI agents with writing follow-up research papers on open scientific problems contributed by DOE scientists across quantum computing, pl…
Beyond Static Cues: Unified Dynamic Detection of MLLM Backdoors via Dual-Flow Trajectory Analysis— arXiv cs.AI
safetyresearch
Researchers propose EVO-Detect, a framework that detects backdoor attacks in multimodal large language models (MLLMs) by analyzing how the model's internal signals evolve during text generation rathe…
P2Patch: Exploit-Guided Vulnerability Repair— arXiv cs.AI
researchsafetytools
P2Patch is a generative adversarial system that improves automated vulnerability repair by iteratively testing patches against new exploits of the same vulnerability, rather than validating against o…
PAWA: Panoramic Action-conditioned World Modeling for Multiple Agents— arXiv cs.AI
researchopen-sourcetools
PAWA introduces a dataset and model framework for predicting how multiple independent players change a shared 3D environment by combining point-of-view and panoramic video alongside player actions in…
Scaling Multilingual Safety and Reasoning Evaluation through Reasoning-Guided Translation— arXiv cs.AI
researchsafety
Researchers developed a reasoning-guided translation framework to create multilingual safety and reasoning benchmarks across 11 languages, revealing that LLMs exhibit language-dependent vulnerabiliti…
What Makes a Social Agent's Interactions Effective? Learning from Verifiable Outcomes in Paired Worlds— arXiv cs.AI
researchtools
Researchers introduce HICO-Bench, a benchmark that measures social agent effectiveness by verifying whether agents actually reach feasible and agreed-upon plans, not just produce fluent conversation.…
Universal Sample-Based Embeddings of Higher-Order Objects— arXiv cs.AI
research
Researchers propose a unified framework for embedding higher-order objects (distributions, functions, graphs) as fixed-size vectors using sample-based encoders and domain-specific decoders. The appro…
Quantifying privacy risks in foundation model representations of brain MRI— arXiv cs.AI
safetyresearch
Researchers evaluated five pretrained encoders on brain MRI data and found that frozen embeddings retain substantial patient-specific information, achieving 90–99% accuracy in patient linkage and 90–…
AgentCIScope: Grounded Contextual Integrity Evaluation for LLM Agents— arXiv cs.AI
researchsafetytools
Researchers introduce AgentCIScope, a benchmark platform that evaluates whether LLM agents preserve contextual integrity—compliance with context-specific privacy constraints—when accessing sensitive…
Hindsight-Privileged RLVR: Single-Rollout Token Credit Without Group Sampling— arXiv cs.AI
researchmodels
Researchers propose HP-RLVR, a reinforcement learning method that assigns credit to individual tokens in language model reasoning using a single rollout per prompt, comparing each token's probability…
RLPE: Teaching LLMs to Prompt with Reinforcement Learning— arXiv cs.AI
researchmodels
Researchers introduce RLPE, a reinforcement learning method that trains LLMs to analyze failure traces from task models and generate optimized prompts, eliminating the need for manual prompt engineer…
From Unreliable to Usable: Decomposing Failure Modes and Confidence Signals in LLM Entity Extraction - Evidence from Cyber Threat Intelligence— arXiv cs.AI
researchsafetyapplications
A preregistered evaluation of entity extraction quality across three LLM tiers (DeepSeek-V4-Flash, GLM-5.3-Flash, GPT-6.1-Sol) on cyber threat intelligence reports shows that unreliability persists e…
MultiPartyTurnBench: Benchmarking Floor Control in Multi-Party Spoken Dialogue— arXiv cs.AI
researchtools
Researchers introduce MultiPartyTurnBench, a benchmark for evaluating how AI systems manage turn-taking in multi-speaker conversations, revealing that current systems excel at detecting natural turn-…
OrionMoE-v1.0: A Mixture-of-Experts Foundation Model for Tabular In-Context Learning— arXiv cs.AI
modelsresearch
OrionMoE-v1.0 introduces a compact mixture-of-experts tabular foundation model that uses dataset-conditioned sparse routing and distribution-aware representations to enable in-context learning on tab…
Tail Competition Explains Scaling in Best-of-N Sampling for Verifiable Problems— arXiv cs.AI
research
Researchers present a theoretical framework using extreme value theory to explain how best-of-N sampling scales when reward models are imperfect, showing that success depends on whether correct respo…
Unlocking Structured Pruning for Reasoning Language Models via Attention-Preserving MLP Trimming— arXiv cs.AI
researchmodels
LayerTrim is a training-free structured pruning method that removes redundant MLP features from reasoning language models while preserving attention mechanisms and decoder depth. The approach achieve…
Genesis Steering: Better Activation Steering Through Training Data— arXiv cs.AI
researchsafety
Genesis steering improves activation steering of LLMs by deriving steering vectors from the model's own fine-tuning data rather than external datasets, and proposes membership inference attacks to co…
Non-Abelian quantum Hall states at strong Landau level mixing: finding anyons with first-principles AI— arXiv cs.AI
researchapplications
Researchers used neural-network variational wave functions to solve the non-Abelian quantum Hall problem at filling factor 3/2, discovering Ising-type anyons with quantum dimension √2 without prior p…
SceneAE: Scene Generation and Reconstruction in Queryable 3D-Anchored Latent Space— arXiv cs.AI
researchmodels
SceneAE proposes representing scenes as a compact 3D-anchored latent space combining explicit 3D positions with implicit content features, enabling efficient reconstruction and generation without cam…
ProtModernBERT: Compute-Efficient Protein Language Modeling— arXiv cs.AI
modelsresearch
ProtModernBERT is a new protein language model that achieves comparable performance to existing models (ProtBert, Amplify, ESM2) while requiring 80-90% fewer pre-training FLOPs through modern transfo…
Whose Data, Whose Benchmark? A User-Centric Framework for Auditing Evaluation Resources— arXiv cs.AI
researchsafety
Researchers introduce a user-centric audit framework revealing that AI benchmarks systematically embed American-centered conventions (62-94% of signals) and lack documentation of target-user context,…
Sparse Autoencoders Are a Shared Coordinate System for Symbolic Matching— arXiv cs.AI
researchmodels
Researchers propose SAE-IDF, a method using sparse autoencoders to match symbolic units across different neuro-symbolic systems by encoding symbols into sparse codes and re-weighting by rarity. The a…
From Failures to Learnable Worlds: Policy-Conditioned Environment Generation for Navigation Agents— arXiv cs.AI
researchmodelstools
WEDGE is a framework that uses failure modes of pretrained navigation agents to guide the generation of targeted training environments, leveraging LLMs to identify weaknesses and systematically varyi…
LLM-jp-4 Technical Report— arXiv cs.AI
modelsopen-source
LLM-jp-4 is a new family of Japanese language models comprising 8B and 32B-MoE variants trained on 12 trillion tokens, emphasizing transparency and reproducibility in development. The work documents…
PairIT: Autoregressive Transformers for Low-Data Molecule Optimization— arXiv cs.AI
researchmodelsapplications
PairIT introduces an autoregressive transformer framework for molecular optimization that learns from pairs of structurally similar molecules, enabling controlled edits to seed structures with prompt…
Copy-Paste Language Models— arXiv cs.AI
researchmodels
Researchers propose Copy-Paste Language Models (CPLMs), which add a lightweight copy mechanism to standard transformer language models to improve their ability to replicate input tokens. The approach…
Inverse Reasoning Models for Efficient Post-Training and Verification from Answer-Only Data— arXiv cs.AI
researchmodels
Researchers propose training an auxiliary "inverse reasoning" model that generates reasoning traces from answers alone, enabling efficient post-training of reasoning LLMs using only prompt-answer pai…
TheoPhysEval: Assessing AI agents at autonomous theoretical physics problem solving— arXiv cs.AI
researchmodelssafety
Researchers created a benchmark of 50 unpublished theoretical physics problems and evaluated AI agents on autonomous solving, finding GPT-6-Astra achieving 98% accuracy at $3 per problem when given t…
Divergent Paths: Learning an Outcome-Predictive Representation of LLM Agent Trajectories— arXiv cs.AI
researchmodels
Researchers propose Precog-RM, a transformer-based model that learns to predict whether LLM agent trajectories will succeed or fail by mapping actions and observations to a shared embedding space. Th…
Scaling LLM Inference via Parallelism: Throughput-Latency Trade-off— arXiv cs.AI
researchinfra
Researchers present a framework for optimizing LLM inference parallelism strategies by characterizing trade-offs between prefill latency (TTFT) and decode latency (ITL) across tensor parallelism, pip…
Policy-Verified Self-Training: Learning Agent Compliance from Verifiable Policy Rewards— arXiv cs.AI
researchsafety
Researchers introduce Policy-Verified Self-Training (PVST), a method where agents learn compliance by using deterministic policy verifiers as reward signals during self-training, rather than just for…
Who Should We Listen to More? Welfare-Aware Preference Acquisition for Pluralistic Alignment— arXiv cs.AI
researchsafety
Researchers propose Welfare-Aware Preference Acquisition (WAPA), a method that uses social welfare objectives to allocate limited human feedback across diverse individuals when training pluralistic A…
Dimension-Free Scaling Laws for Size Generalization: An Exponential Separation— arXiv cs.AI
research
Researchers establish a statistical learning theory showing that models trained on fixed-size inputs face exponential sample complexity penalties when generalizing to different sizes, unless they sat…
MA-Bench: Measuring Operational Planning in Language Models— arXiv cs.AI
researchmodels
Researchers introduce MA-Bench, a benchmark for evaluating how well language models synthesize large, incomplete, and contradictory documents into operational plans without single correct answers. Th…
CAFE: Confining Feature Distillation to Language-Routed Adapters for Compact Multilingual Automatic Speech Recognition— arXiv cs.AI
researchmodels
Researchers identify that feature distillation hurts multilingual ASR by forcing non-English languages to share gradient space with English in the encoder. They propose CAFE, which routes feature dis…
What Happens to the Goal? Distinct Failures of Policy Alignment in Goal Misgeneralization— arXiv cs.AI
researchsafety
Researchers identify three distinct mechanisms of goal misgeneralization in RL agents—functional target substitution, policy shortcuts, and cue competition—and show that the intended goal information…
Uncertainty-Driven Self-Play for Robust Multi-Agent Driving Policies— arXiv cs.AI
researchsafetyapplications
Researchers propose an uncertainty-driven self-play method for autonomous driving where agents actively seek out weaknesses in learned policies by maximizing epistemic uncertainty, creating an automa…
Titan-Llama: Retrofitting Pre-trained LLMs with External Neural Memory Modules— arXiv cs.AI
researchmodels
Researchers introduce Titan-Llama, a post-training method that retrofits pre-trained language models with hierarchical neural memory modules to handle long-context inference more efficiently without…
Learning in Orbit: Onboard Training and Closed-Loop Attitude Control with a Thermal Segmentation Model on a Spacecraft— arXiv cs.AI
researchapplicationsinfra
Researchers deployed onboard training and inference of a thermal segmentation model on a 6U CubeSat in orbit, using it for real-time closed-loop attitude control to track the Moon. The year-long miss…
How Inner Optimizers Shape Replica Scaling— arXiv cs.AI
researchinfra
Researchers analyze how inner optimizers affect replica scaling in distributed LLM training, identifying that whitening-based preconditioning improves performance across 8-96 model replicas. The work…
XferAgent: Cross-Agent Experience Transfer for Self-Improving LLM Agents— arXiv cs.AI
researchmodels
XferAgent is a framework enabling LLM agents to share and transfer execution experience across heterogeneous agents through local and shared experience graphs. The method shows cumulative performance…
Why Do Agent Teams Fail Across Boundaries? A Failure Taxonomy for Agent-to-Agent Systems with Private States— arXiv cs.AI
researchsafety
Researchers analyzed 2,500+ runs of agent-to-agent systems across 488 scenarios to identify failure modes when autonomous agents work for different principals with private information. They derived a…
DexMimic: A Generalist Dexterous Hand-Object Tracking Policy via Value-Guided Reset Allocation— arXiv cs.AI
researchapplications
DexMimic introduces a generalist dexterous hand-object tracking policy trained jointly across diverse human hand-object interaction datasets using Value-Guided Reset Allocation to prioritize difficul…
Loss Curvature as a Sampling-Free Proxy for Memorization in Diffusion Models— arXiv cs.AI
researchsafety
Researchers propose that loss curvature (Frobenius norm of the Hessian) can serve as a sampling-free metric to predict memorization in diffusion models at both model and sample levels. They introduce…
REWIND: Uncertainty-Guided Trajectory Repair for Test-Time Reasoning— arXiv cs.AI
researchtools
REWIND is a training-free method that uses token entropy to identify and regenerate erroneous reasoning steps in language model outputs during inference. The approach outperforms existing sampling-ba…
TEX-MEX: Similarity-Guided Temporal Contrastive Learning for Time Series— arXiv cs.AI
researchmodelsopen-source
TEX-MEX is a self-supervised learning method that unifies structural and temporal similarity for time series by using contrastive pairs of windows based on morphological similarity and temporal proxi…
Asymptotic Behavior of Transformers— arXiv cs.AI
research
Researchers analyze the long-horizon behavior of transformers during text generation, proving that shallow transformers can fail to converge and establishing that deeper transformers with layer norm,…
Beyond Maliciousness: Static and Runtime Attack-Surface Profiling of Agent Skills— arXiv cs.AI
safetyresearchtools
Researchers analyzed 6,309 agent skills from marketplace listings and found that 70% carry elevated capabilities like shell execution or filesystem access, yet existing permission models are largely…
Narrowing the Collaboration Gap, Probably— arXiv cs.AI
researchapplications
This paper studies how teams of LLM agents should communicate when solving collaborative tasks with partial information. The authors show that communicating calibrated probability distributions over…
CyberArena: Benchmarking AI Agents in Live Symmetric Attack-and-Defense Combat— arXiv cs.AI
researchsafetytools
CyberArena is a new benchmark that evaluates AI agents in dynamic, symmetric attack-and-defense scenarios where two agents compete simultaneously as both attackers and defenders across evolving servi…
Lotus: Learning to Understand Time-series by Learning to Forecast— arXiv cs.AI
researchmodels
LOTUS proposes forecast grounding, a pretraining approach for time-series language models that requires representations to remain predictive of future signal evolution. The method achieves 59.91% acc…
PRISMATIC : Prescription Risk Inspection System for Multi-Agent Treatment Interaction in Clinical Decision-Making— arXiv cs.AI
researchapplicationssafety
PRISMATIC is a multi-agent AI system that coordinates treatment planning across medical specialties to reduce prescription errors in multimorbid patients. The system uses two adaptive mechanisms (DSW…
DAGFormer: Global Topology Prediction for Transformer Cross-Layer Connections— arXiv cs.AI
modelsresearch
DAGFormer introduces a topology predictor that jointly learns global cross-layer connection patterns in Transformers, enabling dynamic routing of information across all layers simultaneously rather t…
$τ$-Emotion-Bench: Exposing the Interaction Fragility of LLM Agents through Real-World User Emotions— arXiv cs.AI
researchtoolssafety
Researchers introduce τ-Emotion-Bench, a benchmark for testing LLM agents' ability to handle tool-calling tasks when interacting with users exhibiting different emotional states, and Tracer, a system…
Can We Believe AI Peer Review?— arXiv cs.AI
researchsafety
Researchers analyzed 200 ICLR papers and ~14,000 AI reviews to assess whether AI peer review scores reliably reflect reviewer judgments. They found significant score fluctuations even under identical…
AgentSafeEvo: Adversarial Co-Evolution for LLM Agent Safety— arXiv cs.AI
safetyresearch
AgentSafeEvo proposes a co-evolutionary reinforcement learning framework where an attacker and defender play an adversarial game to improve LLM agent safety against tool-use manipulation attacks. The…
PhySGA: A Physics-Space Grounded Agentic Framework for Generalizable and Interactive 3D Composition— arXiv cs.AI
researchapplicationstools
PhySGA introduces a physics-space grounded agentic framework that uses collaborative agents (Assembly and Critic) with physical simulation to enable complex 3D object composition tasks like precise h…
Quiet Triggers, Loud Footprints: A Tri-Space Measurement Study of Backdoor Stealthiness— arXiv cs.AI
researchsafety
This paper presents a comprehensive framework for measuring backdoor attack stealthiness in neural networks across three spaces: input, feature, and parameter space. The study reveals that attacks qu…
Lifted State Policy Optimization: Internalizing Self-Correction through Transition Geometry— arXiv cs.AI
researchmodels
Researchers propose Lifted State Policy Optimization (LSPO), a method that credits individual reasoning transitions rather than only final answers or full trajectories, using a learned continuous coo…
Explorer-Zero: Improving Self-Play for Long-Horizon Tasks with Decomposable Verifiable Rewards— arXiv cs.AI
researchmodels
Explorer-Zero is a self-play framework that trains LLM agents on long-horizon tasks by having a learned proposer generate both tasks and fine-grained verifiable rewards through environment interactio…
Definition Grounding in Theory-Level Auto-formalization— arXiv cs.AI
researchtools
Researchers propose 'definition grounding,' a method where LLM-based auto-formalization systems continuously evaluate and revise definition formalizations to improve their reusability in downstream t…
One Monitor Does Not Fit All: Formalism-Agnostic Runtime Verification for Agentic AI— arXiv cs.AI
safetytoolsresearch
AgenticRV is a runtime-verification framework that enables monitoring and controlling agentic AI systems by decoupling monitoring logic from execution boundaries through adapters, event projections,…
C3: Rethinking Process Supervision via Confidence-Guided Credit Correction in Multi-Turn Reinforcement Learning— arXiv cs.AI
researchmodels
Researchers propose C³, a method that corrects reinforcement learning credit assignments in multi-turn LLM-agent interactions by identifying and penalizing undesirable local behaviors while preservin…
Model What Matters: Decision-Guided Selective World Modeling for Language Agents— arXiv cs.AI
researchmodels
Researchers propose DGS-WM, a training method that selectively supervises language agents on only decision-relevant parts of their environment observations rather than all observations uniformly. The…
EvalBuilder: Can LLMs Build Efficient Benchmarks?— arXiv cs.AI
researchtools
Researchers introduce EvalBuilder, a framework testing whether LLM agents can automatically construct cheaper evaluation benchmarks that match the conclusions of expensive reference benchmarks. The w…
VLRLA: Deployment-Time Adaptation of Vision-Language-Action Models through In-Context Reinforcement Learning— arXiv cs.AI
researchmodelsapplications
Researchers propose VLRLA, a framework that enables vision-language-action (VLA) robot models to adapt and improve during deployment through in-context reinforcement learning, without modifying model…
When Is Smaller Greener? Right-Sizing Language Models for Agentic AI— arXiv cs.AI
researchmodelssafety
A new arXiv paper proposes a framework for selecting language models in agentic AI systems by evaluating them per completed task rather than per API call, accounting for compounding failure risk acro…
Drivers of LLM Consensus Formation Under Competing Incentives: A Simulation Study— arXiv cs.AI
researchsafety
Researchers tested how LLMs reach consensus in multi-agent scenarios with conflicting incentives using a Network Consensus game, finding open-source models require global communication to agree while…
EvoProbe: Evaluation-Driven Benchmark Evolution for LLM Agents— arXiv cs.AI
researchtools
EvoProbe is a framework for evolving LLM agent benchmarks through automated task generation guided by diagnostic evaluation feedback, addressing saturation and contamination in static benchmarks. Tes…
From Relevance to Validity: Zero-Shot Tool Selection via Functional and Contextual Validation— arXiv cs.AI
research
Researchers propose a validation framework for zero-shot tool selection in LLM agents that goes beyond semantic relevance matching by adding functional compatibility and contextual evidence checks. T…
LabCompiler: A Framework for Verifiable Human-Machine Laboratory Execution from Imperfect Protocols— arXiv cs.AI
researchtoolssafety
LabCompiler is a new framework that automatically translates natural-language lab protocols into executable workflows while diagnosing and repairing defects before compilation. The system separates e…
Steering Vectors for Competing Ethical Frameworks Are Interchangeable: A Competing-Target Test for Activation Steering— arXiv cs.AI
researchsafety
Researchers introduce a competing-target test to evaluate whether activation steering vectors actually control specific behaviors directly or merely activate shared components with neighboring behavi…
Smart Alone, Blind Together: Why Accuracy Fails to CertifyTheory of Mind in Multi-Agent Teams— arXiv cs.AI
researchsafety
Researchers challenge accuracy as a certification metric for theory-of-mind inference in multi-agent systems, arguing that accuracy thresholds fail to account for context-dependent error costs and no…
Read, Parse, Describe: Unified Document Parsing with Visual Element Description Generation— arXiv cs.AI
researchtoolsapplications
Researchers introduce Read-Parse-Describe, a unified dataset and benchmark for document parsing that requires models to simultaneously extract layout elements, reading order, text content, and genera…
Hierarchical Residual Policy Learning for Real-World Mobile Manipulation with Sparse Human Guidance— arXiv cs.AI
researchapplications
Researchers propose HiRe-MoMa, a framework for efficiently fine-tuning mobile manipulation robots in the real world using hierarchical residual learning and sparse human guidance. The method achieves…
Cortex: Consolidating Context into Parametric Memory— arXiv cs.AI
researchmodels
Researchers introduce Cortex, a method for consolidating language model performance gains from context into persistent parametric memory through low-rank or structural updates. The approach is evalua…
DTDD: Divergence-Triggered Dynamic Distillation for Reliable On-Policy Supervision— arXiv cs.AI
researchmodels
DTDD introduces an adaptive on-policy distillation framework that monitors student-teacher divergence and triggers teacher intervention to provide recovery trajectories for long-horizon reasoning tas…
Vision-Language Models Integrate Conflicting Evidence Like Biased Bayesian Observers— arXiv cs.AI
research
Researchers propose a Bayesian cue combination model to predict how vision-language models integrate conflicting text and image inputs. The two-parameter model accurately predicts VLM preferences acr…
Mitigating Video Hallucinations via Attention-Guided Representation Perturbation— arXiv cs.AI
researchsafety
Researchers propose Attention-guided Representation Perturbation for Contrastive Decoding (ARP-CD), a training-free method to reduce hallucinations in video LLMs by strategically weakening visual att…
Function-Space Temporal Self-Distillation for Policy Optimization— arXiv cs.AI
researchmodels
Researchers propose Function-Space Policy Optimization (FSPO), a single-policy reinforcement learning framework that improves language model reasoning by storing historical token distributions as fix…
Beyond Corrected Actions: Training LLM Agents on Textual Critiques— arXiv cs.AI
researchmodels
Researchers propose Critique-Augmented Agentic Fine-Tuning (CAFT), a method that trains LLM agents on both corrected actions and textual explanations of why mistakes occurred, rather than corrected a…
Diagnosing Low-Confidence Patient Profiles via Localized Harmful Covariate Shift Detection— arXiv cs.AI
researchsafetyapplications
Researchers propose MED3pa and Detectron, a diagnostic pipeline that identifies low-confidence patient subgroups in clinical prediction models and tests whether their performance degradation correlat…
Measuring and Reducing Unusable Replies of LLM Agents— arXiv cs.AI
researchsafety
Researchers introduced 'reply usability' as a metric for LLM agents, finding that 54% of agent replies fail to be understandable to requesters despite completing tasks correctly. The paper proposes i…
IMU-Bench: Benchmarking Wrist IMU Representations Across Kinematic, Semantic and Clinical Tasks— arXiv cs.AI
researchtools
Researchers present IMU-Bench, a benchmark evaluating wrist-worn inertial measurement unit representations across ten tasks spanning activity recognition, pose estimation, and clinical assessment. Th…
Retrieve What Helps, Only When It Helps: Agentic Reinforcement Learning for Selective Retrieval-Augmented Machine Translation— arXiv cs.AI
researchmodels
Researchers propose SRAT, a reinforcement learning framework that learns to selectively retrieve external knowledge for machine translation only when beneficial, rather than applying retrieval unifor…
Seeing, Maintaining, and Learning: The Evolution of Multimodal Agent Memory— arXiv cs.AI
research
Researchers propose a two-axis framework for analyzing multimodal agent memory, categorizing how information is formed/represented and how memory management adapts over time. The framework identifies…
PFedMA: Who Should Join Next as Federated Multi-Agent Orchestration Evolves?— arXiv cs.AI
researchsafety
PFedMA proposes a federated multi-agent framework that enables collaborative reasoning among LLM-based agents while maintaining decentralized data and evolving privacy rules across rounds. The system…
Robust Counterfactuals for Markov Decision Processes under Statistical and Causal Uncertainties— arXiv cs.AI
research
Researchers propose a framework for counterfactual reasoning in Markov Decision Processes that handles both statistical uncertainty (from estimated transitions) and causal uncertainty (from multiple…
HEM-Med: A Hierarchical Agentic Memory Framework for Safe Multi-Center Medication Recommendation— arXiv cs.AI
researchapplicationssafety
HEM-Med proposes a hierarchical agentic memory framework for medication recommendation across multiple hospitals with different resources, using disentangled memory to separate transferable clinical…
Harness Policy Distillation: Aligning Agent Policies with Runtime Semantics— arXiv cs.AI
researchmodels
Researchers introduce Harness Policy Distillation (HPD), a training method that explicitly teaches AI agents to understand runtime constraints and control semantics alongside action selection. Testin…
Beyond LLM Judge: Mechanical Verification for Agentic Autoformalization— arXiv cs.AI
researchtoolssafety
Researchers present a hybrid verification method for formalizing mathematical proofs in Lean that combines LLM-based semantic review with mechanical checks (numerical validation and vacuity detection…
When Does Momentum Help in Reinforcement Learning with Verifiable Rewards?— arXiv cs.AI
research
Researchers prove that momentum (via bias-corrected exponential moving average) accelerates optimization in reinforcement learning with verifiable rewards when the learning rate and momentum coeffici…
Trainable at Depth, Solvable at Small Size: What the Loss Surface of a Polynomial Activation Dictates for Designing, Initializing and Training Networks— arXiv cs.AI
research
Researchers analyze the loss surface of deep networks using polynomial activations (z + λz³), demonstrating that a 100-layer network trains on MNIST without normalization and that the small-scale los…
DreamHMR: From Video Generation to Human Motion Modeling via Structural Focusing— arXiv cs.AI
researchmodels
DreamHMR is a framework that adapts video generation models for human mesh recovery, combining global body representation transfer with local joint feature enhancement to improve reconstruction of bo…
Rationale Distribution Distillation via Flow Matching for Distributional Continuous Latent Reasoning— arXiv cs.AI
research
Researchers propose rationale distribution distillation (RDD), a method that trains a flow model to generate multiple diverse latent reasoning paths by distilling from textual chain-of-thought exampl…
SWE-Nexus: Automated Composition of Software Engineering Tasks from Test Execution— arXiv cs.AI
researchtoolsopen-source
SWE-Nexus automatically constructs software engineering tasks from test execution traces across repositories, creating 600 benchmark tasks from 105 repos in six languages. The system uses four specia…
Plan or Retrieve First? A Scale Study of Tool Discovery Architectures— arXiv cs.AI
research
Researchers compared two strategies for LLM agents to discover tools from large registries: Retrieve-then-Plan (embedding queries to find similar tools) versus Plan-then-Retrieve (generating abstract…
Why Static Skill Inconsistencies Do Not Equate to Runtime Risks: A Behavioral Analysis of LLM Agent Execution— arXiv cs.AI
researchsafety
Researchers conducted an empirical study of 575 LLM agent skills, identifying 1,182 contract inconsistencies between documentation and code, then measured runtime impacts through controlled experimen…
One Agent per Zone: Decentralised Active Inference under Partial Observability for Multi-Zone HVAC Control on Edge Hardware— arXiv cs.AI
researchapplicationsinfra
Researchers developed a decentralized multi-agent active-inference system for HVAC control where each zone has an independent agent that learns online without inter-agent communication, using only lo…
Recovering Reliable Supervision via Critique-Guided On-Policy Distillation— arXiv cs.AI
research
This paper introduces CriticOPD, a method that improves on-policy distillation by having a teacher model critique student errors and provide corrections before training, addressing the problem that t…
Token Utility for On-Policy Distillation via Local Trust-Region Optimization— arXiv cs.AI
research
Researchers propose Fisher Utility Weighting (FUW), a token-selection method for on-policy distillation that derives token importance directly from the distillation objective via local KL-constrained…
New Model, Old Habits: Behavior Transfer under Model Swapping in Multi-Session LLM Agents— arXiv cs.AI
researchmodels
This paper identifies behavior transfer as a phenomenon where LLM agents adopt reasoning patterns from previously-used models when switching to new ones mid-conversation, and proposes DASH summarizat…
Evidence-Anchored Overdose Death Prediction from Death Investigation Reports— arXiv cs.AI
researchapplicationssafety
Researchers developed an automated framework to classify suspected overdose deaths from death investigation reports, producing expert-guided explanations alongside predictions. The pipeline model app…
Frame Evidence:A Benchmark for Reference Frames and Visual Evidence in Multi-Image Spatial Reasoning— arXiv cs.AI
researchmodels
Frame Evidence is a new benchmark of 320 triplets and 960 Chinese questions that evaluates whether vision-language models maintain spatial reasoning correctness across different reference frames and…
Generalizing Identifiability Results in Causal Discovery from Complete Data to Missing Data Settings— arXiv cs.AI
research
Researchers establish conditions under which causal graphs remain identifiable when data contains missing values, extending theoretical guarantees from complete-data causal discovery to missing-data…
Beyond Aggregate Accuracy: Item-Level Transitions in Prompt Optimization— arXiv cs.AI
researchtools
RFDPO introduces item-level tracking of prompt optimization to reveal hidden volatility in automatic prompt improvement: overall accuracy metrics mask substantial changes in individual predictions, w…
Following the Flow the Law Requires: Training Voice Agents to Transfer to Unseen Regulated Procedures in Debt Collection— arXiv cs.AI
researchapplicationssafety
Researchers developed FlowSpec, a machine-readable specification system that enables LLM-based voice agents to generalize compliance procedures across unseen regulated deployments. Testing on a Thai…
Do Syntax Heads Influence Model Behavior?— arXiv cs.AI
research
This paper investigates whether transformer attention heads specialized in syntax actually causally influence syntactic predictions through interventions. The findings suggest syntax heads have limit…
How Do LLMs Understand Legal Philosophy? Diagnosing Legal-Philosophical Reasoning in LLMs— arXiv cs.AI
research
Researchers introduce LePhilBench, a benchmark to evaluate whether large language models can understand and apply legal philosophy in reasoning tasks. The dataset consists of 220 U.S. Supreme Court o…
Do Perturbation Predictors Learn Gene Interactions? Diagnosing Extrapolation in Agentic Screening— arXiv cs.AI
researchtools
Researchers evaluated whether machine-learning models trained on single gene perturbations can reliably predict unseen gene combinations by comparing model predictions against a simple additive basel…
VideoHarness-SRSI: Structured Recursive Self-Improvement for Video Understanding Harness— arXiv cs.AI
researchtools
VideoHarness-SRSI introduces a structured framework for automatically evolving software harnesses around frozen vision-language models for video understanding tasks. The method addresses critical fai…
Towards a Data-Centric Understanding of Spatial Reasoning in Multimodal Large Language Models— arXiv cs.AI
research
This arXiv survey examines how multimodal large language models can be improved for spatial reasoning tasks through post-training techniques, with a focus on dataset construction and annotation metho…
Structure-Preserving Document-Level Data Augmentation via Discourse Parsing— arXiv cs.AI
researchtools
Researchers propose a document-level data augmentation framework that uses discourse parsing to preserve rhetorical structure when generating synthetic training data via LLMs. Experiments on summariz…
Evaluating Compressed Memory: Dependence, Utility, and Content Benefit— arXiv cs.AI
researchsafety
A study across nine language model systems shows that compressed memory—when zeroed out—sometimes worsens performance, yet models trained without memory often outperform the original memory-equipped…
Interlock: From Local Correctness to Task Completion in Long-Horizon Agents— arXiv cs.AI
researchmodels
Researchers introduce Interlock, a framework that represents long-horizon agent tasks as typed completion graphs to ensure task dependencies are satisfied, not just individual requirements met. Fine-…
From Copying to Composition: How Language Models Develop In-Context Learning— arXiv cs.AI
research
Researchers studied how language models learn in-context learning by distinguishing between copying nearby demonstrations versus composing answers from rules across the full prompt. They found that a…
LLM-Based Policy Compliance Verification of Loan Applications— arXiv cs.AI
applicationsresearch
Researchers developed an LLM-based system that automatically extracts compliance rules from bank lending policy documents and evaluates loan applications against them, detecting policy exceptions and…
xLightGCN: Explainable Graph Neural Network Based Recommendation Systems with Most Influential Peers— arXiv cs.AI
researchsafety
Researchers propose xLightGCN, a graph neural network-based recommendation system that provides exact, interpretable explanations for its predictions by identifying influential peers in the recommend…
Recognition Is Not Grounding: Benchmarking Multilingual Financial Evidence Extraction— arXiv cs.AI
researchtools
Researchers introduce MULTIFINGROUND, a benchmark for evaluating multilingual financial fact extraction across English, French, and Greek using real 2025 corporate disclosures. The benchmark reveals…
ActiveEvidence: Decision-Focused Evidence Graphs with Verifiable Acquisition Gaps— arXiv cs.AI
researchmodels
Researchers present ActiveEvidence, a system for multimodal agents to decide which visual evidence to acquire and when to stop gathering more data, using a Counterfactual Evidence Interaction Graph (…
Divergence-Routed On-Policy Distillation for Language Models— arXiv cs.AI
research
Researchers propose Divergence-Routed On-Policy Distillation (DrOPD), a technique that improves knowledge transfer from teacher to student language models by using Jensen-Shannon divergence to identi…
Beyond ICD Codes: Semantic IDs for Generative Medical Coding— arXiv cs.AI
researchapplications
Researchers propose SemICD, a framework that trains language models to generate semantic IDs (SIDs) instead of raw ICD codes for medical coding tasks. SemICD uses hierarchical clustering of code embe…
EvoWeave: Composing Reusable Agent Experience through Operation Level Graphs— arXiv cs.AI
researchmodels
EvoWeave is a framework that enables AI agents to reuse past experience by representing intermediate operations as directed graphs with input/output contracts, rather than storing entire workflows or…
ReFormBench: Benchmarking the Stability of Response-Conditioned Format Correction in Large Language Models— arXiv cs.AI
researchtools
ReFormBench introduces a benchmark with 5,000 test instances to measure how well LLMs can correct their output formatting when given new instructions, using each model's baseline response as referenc…
Whose turn is it? Context-guided model routing for stateful agents— arXiv cs.AI
researchmodels
Researchers propose context-aware model routing for multi-turn conversational agents that considers session history and prompt caching costs, not just individual request difficulty. Evaluation on spr…
MemeTrace: A Multimodal LLM-Based Agent for Safety-Aware Meme Provenance Analysis— arXiv cs.AI
researchsafetytools
Researchers introduced MemeTrace, a multimodal LLM-based agent designed to trace the origin, spread, and safety risks of online memes beyond simple binary safe/unsafe classification. The system was e…
Integrating Simulation and Chain-of-thought Reasoning in Multimodal-Language Models For Physical Reasoning— arXiv cs.AI
researchmodels
Researchers propose a hybrid approach that combines chain-of-thought reasoning and physics simulation in multimodal language models, allowing the system to select the optimal method for each physical…
Reflect After Failure: Brain-Inspired Self-Evolving Agents for Embodied Spatial Intelligence— arXiv cs.AI
researchmodels
Researchers propose HiMeco, a self-evolving agent for embodied AI tasks that uses brain-inspired metacognitive reflection to diagnose why spatial tasks fail (perception, exploration, or planning issu…
Maximising Transparency in Context Compaction Improves Retention of Facts and Commitments for Lightweight Coding Agents— arXiv cs.AI
researchtoolsopen-source
Researchers present transparent-compact, a deterministic context compaction method for coding agents that preserves original user/assistant text while explicitly marking removed information, rather t…
Value-Guided Repulsion for Offline Reinforcement Learning— arXiv cs.AI
research
Researchers propose Value-Guided Repulsion Q-Learning (VRQL), which integrates action value estimates into particle interaction fields for offline reinforcement learning, allowing the method to disti…
Long-Horizon RL Needs Muon— arXiv cs.AI
research
Researchers propose a single-rollout asynchronous training recipe combining Muon optimizer, token-averaged loss, and importance sampling for long-horizon reinforcement learning tasks like repository-…
ACR-Net: Mitigating Semantic Dominance via Contrastive Acoustic-Semantic Decoupling— arXiv cs.AI
researchmodels
Researchers propose ACR-Net, a neural architecture that addresses "Semantic Dominance"—the tendency of speech emotion recognition and audio-language models to over-rely on text semantics when vocal t…
VascuSR: Iterative Anatomy-Aware Diffusion for CT Super-Resolution— arXiv cs.AI
researchapplications
VascuSR is a new diffusion-based framework for CT super-resolution that uses explicit 3D vascular guidance to prevent anatomically implausible vessel reconstructions. The method combines iterative re…
Learning Graph Energies for Finite-Time Generation— arXiv cs.AI
researchmodels
This paper proposes finite-horizon energy learning for graph generation, which trains energy-based models to match the distribution after a fixed number of sampling steps rather than at equilibrium.…
Equal CER, Unequal Damage: Severity-Weighted Character Error Rate for OCR— arXiv cs.AI
researchtools
Researchers propose semCER, a severity-weighted variant of Character Error Rate that uses LLM judges to assign different penalties to different types of character errors in OCR, rather than treating…
FUSE: Fairness-aware Update Screening against Byzantine Exploitation in $q$-Fair Federated Learning— arXiv cs.AI
researchsafety
This paper addresses Byzantine robustness in q-fair federated learning, where a server must distinguish between legitimate fairness-induced amplification of high-loss clients' updates and malicious s…
What and Where in Identifying and Quantifying Conventions in Humans and AI Agents— arXiv cs.AI
researchsafety
Researchers develop a framework to identify and measure conventions—implicit communication strategies—in both humans and AI agents playing Hanabi, finding that humans outperform predictive baselines…
A Strong Harness Narrows the Model Gap: A Comparative Evaluation for Legislative Document-to-XML Conversion— arXiv cs.AI
researchtoolsapplications
Researchers evaluated a pipeline for converting U.S. state legislation to structured XML, finding that careful harness engineering (deterministic tooling around models) matters more than model capabi…
Lost in the Fallback: How Reliably Do Typed Decision Models Use Fallback Options?— arXiv cs.AI
researchsafety
Researchers introduce the Paired Fallback Robustness (PFR) Test to evaluate how well typed decision models (TDMs) used in request routing and model selection reliably execute fallback options. Testin…
DeltaStreamer: Towards Human-Like Agentic Streaming Video Understanding— arXiv cs.AI
researchmodels
Researchers propose DeltaStreamer, an agentic framework for streaming video understanding that maintains structured memory of visual changes and performs query-driven replay to recover missing detail…
Graph-Based Orchestration of Heterogeneous AI Models: A Control Node Approach to Composite Intelligence— arXiv cs.AI
researchtoolsapplications
Researchers present a graph-based orchestration framework for composite AI systems that coordinates heterogeneous models (planning, ML, constraint programming, LLMs) using control nodes, DAGs, and be…
VASPO: Self-Distillation with Privileged Value Estimation— arXiv cs.AI
research
VASPO proposes a method that combines self-distillation with value estimation within PPO for training reasoning models, leveraging privileged information from a self-teacher and critic to improve tok…
Does Machine Translation Retain Media Bias?— arXiv cs.AI
researchsafety
A new study tests whether machine translation systems preserve media bias from source texts into target languages, evaluating four state-of-the-art MT methods on English-German-to-Bulgarian translati…
Multi-System Cognitive Steering for Generative Communication— arXiv cs.AI
researchmodelsapplications
Researchers propose Cognitive Steering, a framework that grounds LLMs in multimodal context (scene objects, emotional states, communication motives) and applies AI control methods to interpret abbrev…
On the Convergence of Semi-Gradient TD: A Mean Field Perspective— arXiv cs.AI
research
Researchers analyze the convergence properties of semi-gradient temporal-difference learning with neural networks using mean-field theory, proving exponential convergence under specific stability con…
Safe-FCL: Continual Risk Awareness and Fine-Grained Safety Aggregation for Federated Continual Learning of LLMs— arXiv cs.AI
safetyresearch
Safe-FCL proposes a federated learning framework that detects and mitigates malicious client updates in continual LLM training by combining continual risk awareness (learning evolving attack patterns…
AtomSearch: Evidence-Grounded Tree Search for Agent Skill Optimization— arXiv cs.AI
researchtools
AtomSearch is a new method for optimizing agent skills through local, atomic edits to external instructions guided by tree search. The approach outperforms SkillOpt on three of four benchmarks while…
Frustratingly Easy Majority Voting Using Length Conditioning— arXiv cs.AI
researchmodels
Researchers propose Length-Conditioned Majority Voting (LCMV), a test-time training method that improves majority voting for LLM reasoning by weighting votes based on chain-of-thought length using a…
One Size Does Not Fit All: Curvature-Aware Adaptive Perturbation for SAM— arXiv cs.AI
research
Researchers propose CAP-SAM, an improvement to Sharpness-Aware Minimisation (SAM) that adaptively adjusts the perturbation radius based on local curvature rather than using a fixed value. The method…
One Task, Many Interfaces: Simple Cross-Interface Consistency Training Improves LLM Agent Robustness— arXiv cs.AI
researchsafety
Researchers propose Cross-Interface Policy Regularization (CIPR), a self-supervised training method that improves LLM agent robustness to interface perturbations by training agents to produce consist…
Multimodal Medical Image Parsing with Task-Specific Low-Rank Adapters— arXiv cs.AI
researchmodelsapplications
Researchers adapted a vision-language model (Qwen3.5-9B) for medical image parsing across six task families and seven imaging modalities using task-specific low-rank adapters, achieving competitive p…
Non-asymptotic Convergence of Average-reward Q-learning with Options— arXiv cs.AI
research
Researchers establish the first finite-sample convergence guarantees for average-reward Q-learning with options, proving Õ(1/iterations) rates in mean-square sense. This addresses a gap in RL theory…
Retrieval Over Training: Similarity search-based Model Selection for Time Series Anomaly Detection— arXiv cs.AI
researchtools
Researchers propose RAMSAD, a retrieval-based framework that selects the best anomaly detector for a given time series by querying similar historical instances rather than training a dedicated select…
Evaluating Large Language Models for Multilingual Abstractive Summarization: A Comprehensive Study of English and Bengali— arXiv cs.AI
researchmodels
Researchers evaluated eight large language models on abstractive summarization across English and Bengali using multiple metrics including lexical, semantic, and factuality measures. Results showed t…
JevCoder: When Does Decision Delegation Pay Off in Code Repair?— arXiv cs.AI
researchtoolsapplications
JevCoder introduces a workflow that delegates code repair decisions to a lightweight probabilistic judge (Jev) instead of using expensive language models, reducing token consumption by 22-51% while m…
Does the Linguistic Register of Thinking Matter? Wenyan as a Compact Reasoning Register for Chinese LLMs— arXiv cs.AI
researchmodels
Researchers propose using Classical Chinese (wenyan) for intermediate reasoning steps in Chinese LLMs while keeping final answers in modern Chinese, finding it reduces reasoning token length while ma…
HACHIMI-II: From Role-Play to Learner Behavior Modeling for Faithful Student Simulation— arXiv cs.AI
researchapplications
HACHIMI-II is a method for simulating student behavior in tutoring systems by building explicit learner models from historical data rather than role-playing with LLMs. The approach achieves 2.4–8.4×…
Learning Across Temporal Graph Representation Silos— arXiv cs.AI
research
This research examines whether predictive patterns learned from temporal graph data can transfer between different graph representations—continuous-time dynamic graphs (CTDGs) and temporal knowledge…
Personalising Deformable Image Registration with Clinical Context and Deformation History for Adaptive Radiotherapy— arXiv cs.AI
researchapplications
Researchers developed CAP-DIR, a deformable image registration method that incorporates clinical context (source contours, dose, text) and patient anatomical history to improve adaptive radiotherapy.…
Beyond Static Positional Encoding: Bayesian Time-aware Attention for Non-Stationary Long-term Time Series Forecasting— arXiv cs.AI
researchmodels
Researchers propose Bayesian Time-aware Attention (BTA), a method that replaces static positional encoding in Transformers with adaptive, input-conditioned positional biases for long-term time series…
MIPS: Mitigating Mode Collapse in Reinforcement Learning for Large Outcome Spaces— arXiv cs.AI
research
Researchers propose Marginal IPS (MIPS), a method that addresses mode collapse in reinforcement learning with large outcome spaces by using a learned model to estimate outcome probabilities rather th…
Recursive Formation and First-Class Reuse of Compositional Capabilities— arXiv cs.AI
researchtools
FormCraft is a framework that enables AI systems to represent learned operations as reusable, composable capabilities that can be recursively combined and stored in persistent libraries with formal v…
ONCship: A Large-Scale Underwater Acoustic Corpus for Ship Recognition— arXiv cs.AI
researchopen-source
Researchers released ONCship, a large-scale underwater acoustic dataset with 4,521 hours of curated hydrophone recordings paired with ship identification data, addressing a significant gap in publicl…
Instruction-Conditioned Motion Control for Robotic Arms: Language-to-Code Decomposition and Reinforcement Learning— arXiv cs.AI
researchapplications
Researchers propose Instruction-Conditioned Motion Control (ICMC), a framework that decomposes language-driven robot control into parameter specification, instruction generation, and code synthesis f…
MIND: Instruction-Conditioned Incremental Grounding for Robots— arXiv cs.AI
researchapplications
MIND is a framework that enables robots to efficiently maintain language instruction compliance by selectively re-evaluating semantic dependencies rather than reprocessing entire scenes after environ…
KoAGE: Do Large Language Models Know, Learn, and Coin Korean Neologisms?— arXiv cs.AI
researchmodels
Researchers introduce KoAGE, a Korean neologism benchmark with 4,271 dated entries and NeoTrace framework, to evaluate whether large language models know, learn, and generate new words. Across 21 mod…
Do Vision CIL Methods Transfer to Protein? A Diagnostic Benchmark of Frozen Protein Encoders— arXiv cs.AI
researchmodels
A systematic benchmark of nine class-incremental learning (CIL) methods from vision finds that none reliably transfer to protein function annotation using frozen protein encoders, with instability tr…
Revisiting Generative vs. Dense Retrieval: A Static Value Prior Reverses the Gap— arXiv cs.AI
research
Researchers compared generative retrieval (GR) and dense retrieval (DR) for e-commerce recommendation on Amazon datasets and found GR's apparent advantage stems primarily from popularity bias. Adding…
Generalization Analysis of Biased Stochastic Gradient Methods for Minimax Problems— arXiv cs.AI
research
This paper establishes the first generalization analysis framework for biased stochastic gradient methods in minimax optimization, relaxing bounded gradient assumptions to bounded second moments and…
Revisiting Triplet Contrastive Learning for 3D Shape Understanding— arXiv cs.AI
researchmodels
Researchers propose Hyperbolic Contrastive Learning (HyperCL), a method that learns 3D shape representations from language alone by embedding them in hyperbolic space to capture hierarchical relation…
Continuum Attention— arXiv cs.AI
research
Researchers propose Continuum Attention (CA), a causal Transformer attention mechanism that replaces discrete key-value lookups with integration over continuous, piecewise linear memory curves using…
AWARE-RG: Abstraction Aware Explanation of Resiliency Graphs for Cyber Physical Systems— arXiv cs.AI
safetytoolsresearch
AWARE-RG is a framework that helps security analysts understand how cyber attacks can lead to physical consequences in critical systems by providing role-specific explanations of attack paths. The sy…
Robust Real-Time Monocular Human Pose Estimation with Insole-Based Sensor Fusion— arXiv cs.AI
researchapplications
Researchers propose a multimodal fusion framework combining monocular video with insole-based pressure and acceleration sensors to improve real-time human pose estimation, particularly during occlusi…
Beyond Retrieval: Evidence Access and Event Aggregation in Agent Memory— arXiv cs.AI
researchtools
Researchers propose TallyMem, a memory architecture that combines trajectory retention with executable event aggregation to answer questions requiring historical data synthesis. The method improves e…
Learning Individual Scientific Taste with Privileged Hindsight Distillation— arXiv cs.AI
researchtools
Researchers developed Privileged Hindsight Distillation (PHD), a method to learn individual scientific taste by modeling how researchers translate criterion-based assessments into preferences, using…
Learning Perception to Action Rules through Reinforcement Learning for Theory-of-Mind Reasoning— arXiv cs.AI
research
Researchers trained Qwen2.5 and Llama3 models using Group Relative Policy Optimization (GRPO) to generate explicit if-then rules connecting perception, belief, and action for Theory-of-Mind reasoning…
Runtime Monitoring of Supermartingale Certificates— arXiv cs.AI
researchsafety
Researchers present an algorithm for runtime monitoring of supermartingale certificates—formal guarantees on stochastic dynamical systems—using betting-based sequential tests to detect violations in…
Design Space Exploration of Reinforcement Learning Configurations for Post-Legalization Placement— arXiv cs.AI
researchtools
Researchers systematically evaluated 18 configurations of a GNN-PPO framework for VLSI post-legalization placement refinement across standard benchmarks, identifying pitfalls and proposing a design c…
Designing Textual Interfaces To Enhance Llm Reasoning On Audio-Visual Question Answering— arXiv cs.AI
researchmodels
Researchers investigate how textual interface design affects automatic prompt optimization (APO) for multimodal LLMs on audio-visual question answering tasks. They compare three architectures—direct…
CSWBench: Evaluating Context-Sensitive Code-Switching in Creative Writing— arXiv cs.AI
researchmodels
Researchers introduce CSWBench, a benchmark for evaluating how well large language models handle code-switching (mixing languages) in creative writing contexts. The work proposes CSWEval, a theory-in…
PREFER: From Player Experience to Personalized Game Refinement— arXiv cs.AI
researchapplications
Researchers introduce PREFER, a framework that uses LLMs as virtual playtesters by evaluating gameplay across standardized experience dimensions and automating persona-specific game refinement. The s…
From Spectral to Frobenius: When Muon Loses Its Spectral Implicit Bias— arXiv cs.AI
research
This paper analyzes the implicit bias of Muon, an optimization method that uses spectral-norm steepest descent, and shows that practical implementations with finite Newton-Schulz iterations and stabi…
LiFi: LiDAR Generation from Multi-View Images via Geometric and Semantic Collaborative Guidance— arXiv cs.AI
researchtools
LiFi is a new diffusion-based framework for generating LiDAR point clouds from multi-view camera images using dual guidance branches for geometry and semantics. The method introduces a balanced class…
Analyzing performance of Retrieval Augmented Generation models for Question Answering Tasks— arXiv cs.AI
research
Researchers analyze how Retrieval-Augmented Generation (RAG) models perform on question-answering tasks, with specific focus on South Asian multilingual and culturally grounded questions. The study e…
Price of Anarchy of First-Price Auctions in Agentic Markets— arXiv cs.AI
researchmodels
This paper analyzes how first-price auctions allocate prompts to AI agents in decentralized markets, showing that welfare loss (price-of-anarchy) is significantly lower than general auction theory pr…
Neural Wasserstein Splines: Representation-Invariant Second-Order Priors for Trajectory Inference— arXiv cs.AI
research
Researchers introduce Neural Wasserstein Splines (NWSP), a method for inferring continuous population dynamics from unpaired snapshots using representation-invariant Wasserstein smoothness priors ins…
Typed Decision Models on Tabular Data— arXiv cs.AI
research
Researchers compared typed decision models (language models constrained to output probabilities over fixed answer sets) versus large language models on tabular data tasks, finding decision models mat…
Recognizing Dialogue Ambiguity via Candidate Set Visual Grounding— arXiv cs.AI
research
Researchers introduce Candidate Set Visual Grounding (CSVG), a task for tracking multiple plausible object referents during ambiguous dialogue rather than resolving to a single object. A new benchmar…
Scaling Cheap Synthetic Pretraining Improves Indoor Radio-Map Prediction— arXiv cs.AI
researchtools
Researchers generated 128M synthetic indoor radio-map samples using an automatic layout generator and fast open-source ray tracer, then pretrained ResNet and Swin-T models to reduce prediction error…
Eye, Robot: Learning Expressive Gaze Dynamics via a VLM-Guided Transformer-VAE— arXiv cs.AI
researchmodels
Researchers developed a Transformer-VAE model (RGT) that generates naturalistic gaze and head motion for humanoid robots by learning from human video data, separating gaze target selection (via visio…
Refined Topology-Pass Scheduling: Damped Trust-Region Compensation for Dynamic k-NN Graphs in SE(3)-Equivariant Structure Prediction— arXiv cs.AI
researchtools
Researchers propose a two-pass scheduling algorithm for SE(3)-equivariant structure prediction that reduces the computational cost of rebuilding k-NN graphs during training. By recording lightweight…
Imagining Goals with Geometric Horizon Models for Policy Generalization— arXiv cs.AI
research
Researchers propose TD-Aug, a method that uses geometric horizon models to imagine novel goal outcomes by composing existing knowledge, enabling goal-conditioned behavioral cloning policies to improv…
JockeyBench: Evaluating LLM Agents on Horse-Racing Decision-Making— arXiv cs.AI
researchtools
JockeyBench introduces a benchmark for evaluating LLM agents on repeated decision-making using historical Japanese horse racing data, testing 28 model-scaffold configurations across 324 races. The st…
DeepPV-MRF: End-to-End Partial-Volume Fraction Estimation from MR Fingerprinting Time Series— arXiv cs.AI
researchapplications
DeepPV-MRF is a deep learning framework that directly estimates tissue fractions from raw MR fingerprinting signals without intermediate parametric maps, reducing error propagation in medical imaging…
AdCoEvo: A Co-Evolving Multi-Agent Advertising Delivery System— arXiv cs.AI
applicationsresearch
AdCoEvo is a multi-agent system for automating digital advertising delivery that uses co-evolving agents across planning, adjustment, and diagnosis roles to handle decisions at different timescales.…
Rethinking Prediction Horizons: Decoupling Model Queries from State Updates in Few-Step MeanFlow— arXiv cs.AI
research
This paper investigates how decoupling model prediction horizons from state updates in MeanFlow diffusion models affects image generation quality. The authors find that longer prediction horizons hav…
Empirical Boundaries of Reservoir Computing for Few-Shot Sequence Pattern Tasks— arXiv cs.AI
research
Researchers conducted an empirical analysis of Echo State Networks (ESNs) on few-shot sequence pattern tasks, identifying when linear dynamics are critical and proposing G-ESN with data-dependent mod…
REPORT-OPS: Hybrid Evidence-Guided Reporting with Human-Supervised Refinement for HPC— arXiv cs.AI
researchtoolsapplications
Researchers present REPORT-OPS, a hybrid LLM-based system for high-performance computing operational reporting that combines language models with human-supervised policy refinement and automated evid…
Selective Growing Q-Networks (SGQN): An Analysis of Adaptive Action Grids in Decoupled Q-Learning— arXiv cs.AI
research
Selective Growing Q-Networks (SGQN) applies adaptive action grids to continuous control via Q-learning, achieving competitive performance on DMControl tasks by selectively growing bins near high Q-va…
Exploring Interpretability in Deep Learning-Based Forest Classification - Separating Similar Forest Classes using Sentinel-2 Time Series— arXiv cs.AI
researchapplications
Researchers used recurrent neural networks with Sentinel-2 satellite time-series data to classify forest types, achieving 85.1% macro-F1 accuracy while outperforming tree-based models. By applying In…
SafeAdapt-Poker: Counterfactual Risk Calibration for Budgeted Adaptation under Opponent Shift— arXiv cs.AI
research
SafeAdapt-Poker proposes a counterfactual risk calibration framework for adapting poker strategies when opponents shift their behavior, using a dual-question approach that balances immediate exploita…
Graph Coarsening Enables Escaping Local Optima in Compute-Constrained Nonconvex MAX-CUT— arXiv cs.AI
research
Researchers propose combining Combinatorial Multigrid (CMG) graph coarsening with Particle Polynomial Relaxation (PPR) to solve the MAX-CUT problem more effectively on compute-constrained systems. Th…
PPAD-Completeness of Contiguous Envy-Free Cake Cutting with Additive Valuations— arXiv cs.AI
research
Researchers prove that computing contiguous envy-free cake allocations is PPAD-complete under restricted valuation conditions, establishing computational hardness even for simplified cases with binar…
Finite-sample certification of coverage regret for budgeted document selection under adaptive stratified auditing— arXiv cs.AI
researchsafety
This paper develops finite-sample confidence bounds for coverage regret in budget-constrained document selection, addressing how much query coverage is lost when selecting subsets for RAG systems und…
Detection of Potato Plant Leaf Diseases using Convolutional Neural Networks— arXiv cs.AI
researchapplications
Researchers developed a lightweight CNN achieving 98.66% test accuracy on the PlantVillage benchmark for classifying potato leaf diseases (healthy, early blight, late blight), with 90.28% accuracy on…
Exact Cut Complexity of Equal-Length Proportional Cake Cutting— arXiv cs.AI
research
Researchers determined the exact worst-case number of cuts needed for proportional cake cutting under equal-length constraints is 2n-2, improving on previous quadratic bounds. The work combines neckl…
GroundGUI: GUI Grounding by High-Quality Data and Distillation-to-Reinforcement Learning— arXiv cs.AI
ELARSpeech: A 30K-Hour Dataset Across 675 Underrepresented Languages for Long-Tail Multilingual Speech Technologies— arXiv cs.AI