Repo-To-Skill: Distilling GitHub Repositories Into AI4AI Skills
DisCo turns ML repositories into verified agent skills, and the paper reports large benchmark gains under a fixed agent setup.
The authors describe “operational knowledge” as the missing layer between knowing a method and making it work. Their system distills GitHub repositories into compact skills, both broadly across popular ML repos and for specific tasks. The resulting AREX-Skill Library contains more than 5,000 verified skills from 1,000 repositories, grouped into 20 areas and 178 capability families. With the same GPT-5.5 backbone, harness, and execution budget, the skill-equipped agent scores 134.3% higher on MLE-bench and posts smaller gains on PaperBench, FrontierCS, and PassNet. HF Daily Papers' note
The authors describe “operational knowledge” as the missing layer between knowing a method and making it work. Their system distills GitHub repositories into compact skills, both broadly across popular ML repos and for specific tasks. The resulting AREX-Skill Library contains more than 5,000 verified skills from 1,000 repositories, grouped into 20 areas and 178 capability families. With the same GPT-5.5 backbone, harness, and execution budget, the skill-equipped agent scores 134.3% higher on MLE-bench and posts smaller gains on PaperBench, FrontierCS, and PassNet. HF Daily Papers' note
score 5