Reproduce “Skill Issue: Lessons from Optimizing Repository SKILLs for Coding Agents” 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
Mykhailo Kozyrev, Andrei Kozyrev, Anton Podkopaev
The authors construct repository-derived tasks by reversing merged pull requests at a frozen repository state, then compare an agent using a candidate SKILL with an empty-seed run. Across three Kotlin repositories, GEPA averages 4.9 percentage points above the seed and SkillOpt 0.1 points above. Neither gain is statistically significant; held-out sets are too small to distinguish modest effects from rerun variation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.