Reproduce “EVISKILL: Grounding Skill Evolution in Replayable Evidence” 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
Yan Zhou, Yili Wang, Yiwei Dai, Qinggang Zhang, Xin Wang
EVISKILL links edits to supporting task-trajectory segments, replays those segments to test each edit, and provisionally retains replay-supported edits after a skill revision fails validation. Across three benchmarks, 70 of 96 edits entered validated skills within four epochs; 26 remained provisional. On ScienceWorld, Qwen3.5-9B achieved 37.91% task-success accuracy versus SkillOpt’s 22.27%, a comparison that does not prove replay caused the difference.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.