Reproduce “EvoGuard: Mitigating Degradation in Online Skill Evolution” 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

Yanli Wang, Zichen Bai, Tianyue Jiang, Kunming Shao, Junhao Shi, Xueqin Liu, Jinguo Sun, Muye Li and 7 more

EvoGuard trains a lightweight selector to choose between an existing skill and a proposed rewrite, using offline skill utilities to teach it which version performs better. At deployment, the selector compares only the two skill texts, without extra environment rollouts. Across ALFWorld settings, it averages 56.5% success, versus 45.5% for accepting every rewrite; it cannot repair skills or guarantee improvement.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.