Reproduce “Evolve Where It Matters: Coverage-Guided Self-Evolution of Agent Skills” 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
Boyin Tan, Jiadong Zhang, Xiaosong Ma, Chun Jason Xue, Xiaowei Huang, Youcheng Sun
SkillPlumber uses instruction-level coverage to guide revisions of natural-language agent skills. A model-based judge labels instructions as followed, not followed, or not covered. On Opus 4.8 in Claude Code, it passed 146 of 207 tasks within five revisions, versus 128 for SkillRevise. In a GPT-5.5 ablation, removing coverage reduced recoveries from 24 to 19; model-generated labels can be wrong.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.