§ Agents & Tool Use

Anthropic donates Agent Skills standard to Agentic AI Foundation under Linux Foundation, joining Model Context Protocol as open community standard.

Proposes Memory-Enhanced Test-Time Adaptation (METTA) for meta-RL using dual-timescale memory to improve fast task adaptation from limited interactions.

Introduces EpiSEM-Bench, executable benchmark with 300 task pairs measuring unsafe execution when tool-using LLM agents face missing safety-relevant facts.

LLM agent debugs FineWeb2 Slovak corpus via tool-use auditing, finding 1.4M machine-generated documents (5.32%) for $2.21-$2.84 per run.

Modal CEO Bernhardsson analyzes why AI compute infrastructure defaults to rectangular resource allocation, examining elastic GPUs and sandbox economics for agent deployment.

Framework for auditability of LLM-generated metadata in enterprises, tracking human stewardship decisions to prevent error propagation across systems.

Twin-world credit assignment isolates policy decisions from random sampling noise in episodic reinforcement learning with deterministic components.

Small open-weight LLM trained via RL to select tools from corpora of 10k+, matching proprietary model performance on ToolBench.

Framework for coding agents to dynamically select skill granularity (primitive/macro/hybrid) at runtime, enabling fallback recovery strategies.

WholeBodyMem: perceptual memory module for humanoid loco-manipulation via object tracking and cross-attention action expert.

Network topology significantly affects consensus formation speed and stability in LLM agent populations across naming game tasks.

New benchmark for video-grounded research agents that synthesize multimodal evidence and generate interactive reports linked to video moments.

AI agent via MCP for structural design translation from natural language to domain-specific 3D models using COMPAS API.

Active Preference Acquisition (APA) framework for web agents to identify high-value user preference questions via Prospective Personalization Value.

Multi-agent LLM consensus study: larger models absorb opinions of smaller models regardless of task design across seven model families.

General-purpose AI agents autonomously perform 3D medical image annotation via nnU-Net integration, requiring no model retraining.

Dual Exam benchmark isolates multimodal agent action-selection from world-model state-evolution via fixed mechanisms across 15 domains and 1,488 test items.

SCOPE compiles natural-language policies into neurosymbolic programs for stable, efficient LLM agent governance over multi-turn trajectories.

Study compares semantic retrieval and LLM-based strategies for tool routing in robotic systems, addressing capability selection for compound natural-language commands.

Environment Duels benchmark evaluates frontier LLMs on creating and solving interactive environments; reveals distinct authoring vs. solving abilities.

Quant-Arena benchmark evaluates LLM agents on end-to-end quantitative trading research with A-share market data and out-of-sample testing.

PAPAV framework unifies multimodal embodied agents across computer-use and robot-use via five capability functions: perceive, plan, act, process feedback, evolve.

ASTEROID learns test-time exploration via asymmetric on-policy distillation, enabling agents to infer hidden task-relevant information from privileged experts.

Circuits of LLM Classifiers organize multiple LLM predictors into adaptive query sequences to maximize accuracy within budget, treating models as noisy measurement instruments.

Models multi-agent coordination as graph-structured block optimization, proving optimal fractional coloring preserves agent edits' combined gains.

Integrates charts as epistemic actions inside ReAct-loop data agents to improve reasoning over structured data beyond text-only analysis.

HGM-Dual improves agent harness optimization for long-horizon planning by incorporating lineage-aware guidance, extending self-improvement beyond coding tasks.

Behavioral Localization identifies trajectory positions where soft skills have concentrated impact, reducing computational costs for tool-using agents.

Study reveals silent information loss in document ingestion layers for enterprise agents, causing confident errors from formatting data loss in PDF/Word conversion.

GuideEvolve framework adds expert guidance and 3Ts mechanisms (think fast/slow, throwback, termination) to self-evolving agents.

Agent Behavior Maps profile LLM-harness combinations via action log-likelihoods on reference trajectories, enabling performance prediction across 100 LLMs.

MHG-Mem, a multimodal hypergraph memory for long-term agent recall that preserves visual evidence while tracking entity state evolution across sessions.

CORD, a conflict-aware optimization framework for joint RL and self-distillation in agents, uses nearest-point projection to resolve gradient cancellation between task and teacher objectives.

ToM2Vec learns continuous theory-of-mind tokens in latent space for multi-agent LLM coordination under uncertainty.

FrontierCS 2 benchmark of 48 open-ended CS R&D tasks measures LLM agent contributions in working research environments.

Claude Mods reimagines agent interface design by shifting control from machine-facing harness to human-facing Attention Interface paradigm.

OpenAI launches Patch the Planet, an agent-based platform coordinating security researchers and maintainers to identify and address critical open-source vulnerabilities.

Tenex discusses outcome-based pricing and agent-driven SDLC tooling for enterprise AI adoption at Fortune 2000 scale.

OpenAI, PerturbAI, and Scripps Research deployed distributed AI agents to identify novel biological leads from millions of brain cells.

Cohere Rerank 5 improves passage retrieval for RAG and agentic systems with better token efficiency.

GBClaw agentic AI framework automates grain boundary discovery in materials science via autonomous experimentation and mechanistic inference.

Multi-agent framework using property graphs for autonomous root-cause analysis in distributed microservice logs.

Agent S4 meta-agent autonomously designs evolving computer-use interfaces tailored to task workflows.

Study reveals on-policy distillation fails beyond compounding error; proposes teacher repair via selective intervention in student trajectories.

SceneRig reconstructs simulation-ready 3D scenes from single images using agentic physics grounding, enabling scalable scene generation for robotics.

Sage-A1.5 35B MoE agentic model enables adaptive task composition, verification, and scientific discovery via subagent orchestration.

Cache engineering for LLM agents optimizes KV storage and prefilling costs across long trajectories without invalidating cached states.

OptScientist framework discovers symbolic optimization laws from data using neural inference and LLM-agentic exploration.

AnomalyAgent ensemble framework uses agentic fusion of vision-language models for anomaly detection with adaptive weighting.

OptionWorld benchmark reveals LLMs fail at multi-day task completion, with 35.5% of failures due to lost context before deadline.