acceptodds
Desk rejected at ICLR 2027

SkillRefiner: Offline Skill Refinement from Historical Agent Traces

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Skills capture procedural knowledge that helps agents solve complex tasks. Existing skill-refinement methods improve these instructions by generating new rollouts with candidate skills, which can be costly or impractical when environments are not replayable or supervision arrives only after deployment. We introduce SkillRefiner, an offline method that refines deployed agent skills from historical execution traces and observed outcomes. SkillRefiner summarizes long trajectories, clusters recurring behaviors and failure modes, and derives targeted edits from each cluster. Successful clusters reinforce effective behavior, while failure clusters produce guardrails and procedural corrections that are checked against supporting evidence. We evaluate SkillRefiner on spreadsheet manipulation, mathematical reasoning, and production pull-request review. SkillRefiner improves over both the LLM and human authored skills across all settings while using 1.4–42× fewer refinement tokens than competing refinement methods.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Skill learning, Coding Agents, Large Language Models

Related papers

Discussion (0)

Sign in to comment.