Reproduce “SkillEvolver: Skill Learning as a Meta-Skill” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Genrui Zhang, Erle Zhu, Jinfeng Zhou, Caiyan Jia, Hongning Wang
SkillEvolver introduces a meta-skill framework enabling agents to iteratively acquire and refine reusable procedural knowledge, or 'skills,' directly from deployment-time trials without retraining underlying large language models. This approach achieved an average Avg@5 mean of 56.87% on SkillsBench, outperforming human-curated skills by 13.3 percentage points, while reducing downstream agent token usage by 19.4%.
Key insights
- Online skill learning can be framed as an 'artifact-level adaptation' process, iteratively refining reusable procedural artifacts rather than solely updating model parameters.
- Deployment-grounded feedback, derived from how separate agents utilize and fail with candidate skills, is crucial for identifying and correcting practical failure modes like under-abstraction or 'silent bypass.'
- A lightweight, plug-and-play meta-skill can orchestrate the entire lifecycle of skill acquisition and refinement, enabling online learning with a limited number of trials.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.