Repo-To-Skill: Distilling GitHub Repositories Into AI4AI Skills
The paper claims an agent can turn ML repositories into compact, verified skills and use them to improve research benchmarks under a fixed setup.
The authors introduce DisCo, a research agent that distills operational know-how from GitHub repositories and task needs. Its broader run produced AREX-Skill Library, described as 5,000+ verified skills from 1,000 widely used ML repositories across 20 areas. With GPT-5.5, the same harness, and the same execution budget, the skill-equipped agent outperformed the no-skill version on MLE-bench, PaperBench, FrontierCS, and PassNet. ArXiv · AI/CL/LG's note
The authors introduce DisCo, a research agent that distills operational know-how from GitHub repositories and task needs. Its broader run produced AREX-Skill Library, described as 5,000+ verified skills from 1,000 widely used ML repositories across 20 areas. With GPT-5.5, the same harness, and the same execution budget, the skill-equipped agent outperformed the no-skill version on MLE-bench, PaperBench, FrontierCS, and PassNet. ArXiv · AI/CL/LG's note
score 5