Tag
The paper proposes 'Space', a skill-guided adaptive action chunking method for long-horizon LLM agents, improving success rates by 7.0%–31.3% and reducing LLM decision rounds by up to 78.9%.
DisCo is a research agent that distills operational knowledge from GitHub repositories into reusable skills, significantly improving autonomous ML research performance across benchmarks like MLE-bench.
Search2Skill is a framework that trains LLM agents to identify capability gaps, search external sources, and distill retrieved knowledge into reusable skills using rubric-based reinforcement learning, beating baselines on expert-domain benchmarks.
SKILL-KD is a contrastive skill distillation framework that improves LLM agents by distilling actionable discrepancies between teacher and student trajectories into textual skill patches, with drift-aware consolidation to iteratively refine skills.
This paper presents the first systematic exploration of filesystem-based memory for LLM agents, formalizing roles of management, search, and execution agents around a shared memory store. It finds that organization primarily reduces retrieval cost but does not yet improve answer quality, and that tooling choices affect store shape as much as model selection.
BrowserBC is an open-source system that converts recorded human browser interactions into reusable agent skills, enabling skills to be used in Claude Desktop and Claude Code with just one recording.
OPID is a framework that extracts dense token-level supervision from completed on-policy trajectories for reinforcement learning of language agents, using hierarchical skills (episode-level and step-level) to improve sample efficiency and robustness.
OPID proposes an on-policy skill distillation framework that extracts dense hindsight supervision from completed trajectories, combining outcome-based RL with token-level self-distillation to improve language agent training efficiency and performance on multi-turn tasks.
EmoDistill is an offline framework that distills emotional negotiation skills into language model agents using Implicit Q-Learning for emotion selection and LoRA-based supervised fine-tuning and judge policy optimization for emotion expression, achieving higher utility in adversarial negotiations.
PANDO is a web agent framework that improves efficiency through online skill distillation, reducing token usage by 58-61% while outperforming baselines on VisualWebArena tasks.
A single CLAUDE.md file became GitHub's top-trending repo with 44k weekly stars by distilling Andrej Karpathy's LLM coding advice into four principles.