AI/ML arXiv cs.AI

EvoMaster: A Foundational Evolving Agent Framework for Agentic Science at Scale

EvoMaster is a self-evolving agent framework for scientific discovery that enables agents to iteratively refine hypotheses and accumulate knowledge.

AI/ML arXiv cs.AI

LiteResearcher: A Scalable Agentic RL Training Framework for Deep Research Agent

LiteResearcher is a scalable RL training framework that uses a virtual world to train deep research agents more efficiently and effectively.

AI/ML arXiv cs.AI

Are Performance-Optimization Benchmarks Reliably Measuring Coding Agents?

An audit of performance-optimization benchmarks for coding agents (GSO, SWE-Perf, SWE-fficiency) reveals significant instabilities and scoring flaws that may misrepresent agent progress.

AI/ML arXiv cs.AI

FurnitureVLA: Learning Long-Horizon Bimanual Furniture Assembly with Vision-Language-Action Model

Researchers introduce FurnitureVLA, a Vision-Language-Action model designed for long-horizon bimanual furniture assembly using a progress-enhanced VLA and VR teleoperation for data collection.

AI/ML arXiv cs.AI

The State-Prediction Separation Hypothesis

The state-prediction separation hypothesis proposes a Transformer variant that disentangles token prediction from state storage, leading to better data and compute efficiency.

AI/ML arXiv cs.AI

Language-Critique Imitation Learning from Suboptimal Demonstrations

A new language-critique framework for imitation learning uses structured natural language feedback instead of scalar signals to learn robust policies from suboptimal demonstrations.

AI/ML arXiv cs.AI

Measuring the Gap Between Human and LLM Research Ideas

A study on research ideation finds that LLM-generated ideas are more narrow and systematically different from human-generated ideas, often focusing on synthesis rather than diverse framing.

AI/ML arXiv cs.AI

Unexplainability of Artificial Intelligence Judgments and Functional Implementation in Kant's Perspective

This paper applies Kant's theory of judgment to AI, arguing that the Softmax function reframes AI judgments as mere possibility judgments, making true functional implementation unexplainable.

AI/ML arXiv cs.AI

From Silos to Systems: Process-Oriented Hazard Analysis for AI Systems

The authors present PHASE, a guideline for process-oriented hazard analysis for AI systems based on the System Theoretic Process Analysis (STPA) framework to identify system-level risks.

AI/ML arXiv cs.AI

Selective Expert Guidance for Effective and Diverse Exploration in Reinforcement Learning of LLMs

MENTOR is introduced as a framework for Reinforcement Learning with Verifiable Rewards (RLVR) that provides expert guidance only at critical decision points to improve exploration diversity.

AI/ML arXiv cs.AI

Large language models replicate and predict human cooperation across experiments in game theory

Research shows that specific LLMs, particularly Llama, can replicate human cooperation patterns in game-theoretic experiments with high fidelity, suggesting a mechanistic basis for human-like behavior.

AI/ML arXiv cs.AI

CoT-X: An Adaptive Framework for Cross-Model Chain-of-Thought Transfer and Optimization

CoT-X is an adaptive framework that compresses Chain-of-Thought reasoning traces via semantic segmentation, reducing inference overhead while maintaining accuracy across different LLM scales.

AI/ML arXiv cs.AI

Muon as a Residual Connection

Proposes a mechanistic interpretation of the Muon optimizer as an implicit residual connection that balances local gradient descent with representation preservation.

AI/ML arXiv cs.AI

Autonomous Scientific Discovery via Iterative Meta-Reflection

Introduces DiscoPER, an LLM-powered framework for autonomous scientific discovery that uses iterative meta-reflection and code execution to explore datasets.

AI/ML arXiv cs.AI

Skills Are Not Islands: Measuring Dependency and Risk in Agent Skill Supply Chains

Introduces Agent Skill Supply Chains (ASSCs) and the SkillDepAnalyzer tool to manage and audit dependencies in LLM agent skills, improving security and provenance.

AI/ML arXiv cs.AI

Sequentially-Controlled Interactive Multi-Particle Flow-Maps for Online Feedback-Driven Search

Presents IMPFM, a framework for sample-efficient online feedback-driven search using multi-particle flow-maps to prevent mode collapse and reward over-optimization.

AI/ML arXiv cs.AI

Adversarial Pragmatics for AI Safety Evaluation: A Benchmark for Instruction Conflict, Embedded Commands, and Policy Ambiguity

Introduces a benchmark for adversarial pragmatics to evaluate LLM behavior under instruction conflict, policy ambiguity, and embedded commands for better AI safety audits.

AI/ML arXiv cs.AI

Diffusion-GR2: Diffusion Generative Reasoning Re-ranker

Introduces Diffusion-GR2, which converts autoregressive reasoning re-rankers into block-diffusion models to significantly increase throughput while maintaining accuracy.

AI/ML arXiv cs.AI

Right in the Right Way: LM Training with Verifiable Rewards and Human Demonstrations

Proposes an adversarial generator-discriminator framework to combine verifiable rewards with human demonstrations, reducing reward hacking and improving output style in RL training.

AI/ML arXiv cs.AI

World from Motion: Generative Dynamic Gaussian Reconstruction from Monocular Video

Presents 'World from Motion', a method for generating high-quality dynamic 3D Gaussian representations from monocular videos by distilling video model generations.