Repo-To-Skill Distills 1,000 GitHub Repos Into 5,000 Agent Skills
The top paper on Hugging Face's daily board — 454 upvotes, a huge number for that page — is Repo-To-Skill from the Beijing Academy of Artificial Intelligence (arXiv 2609.02749). The premise cuts at something real: ML agents have planning and execution, but they lack the operational know-how that makes methods actually work in practice. That knowledge exists. It's sitting in GitHub repos, encoded in the code and configs of people who fought the problems already.
So BAAI built DisCo, an agent that mines repositories and distills them into reusable skills, two ways: task-agnostic (turn a widely-used ML repo into general-purpose skills) and task-oriented (generate skills for a specific research problem). The output is the AREX-Skill Library: 5,000+ verified skills from 1,000 repos, organized into 20 areas and 178 capability families. The name is no accident — this plugs into AREX, BAAI's self-grading research agent we covered in July (https://clauday.com/article/d49ff227-5478-491d-b271-5aca599f737b).
The numbers are startling: a GPT-5.5-powered research agent with these skills improves 134.3% on MLE-bench and 34.4% on PaperBench. Skills aren't seasoning; they roughly doubled benchmark performance while the model stayed identical.
Step back and the shape is familiar: this is what mattpocock/skills and the whole skills-repo boom do by hand, done autonomously at library scale. The skills-as-standard thread now has an industrial supply side. Paper at https://arxiv.org/abs/2609.02749.
← Back to all articles
So BAAI built DisCo, an agent that mines repositories and distills them into reusable skills, two ways: task-agnostic (turn a widely-used ML repo into general-purpose skills) and task-oriented (generate skills for a specific research problem). The output is the AREX-Skill Library: 5,000+ verified skills from 1,000 repos, organized into 20 areas and 178 capability families. The name is no accident — this plugs into AREX, BAAI's self-grading research agent we covered in July (https://clauday.com/article/d49ff227-5478-491d-b271-5aca599f737b).
The numbers are startling: a GPT-5.5-powered research agent with these skills improves 134.3% on MLE-bench and 34.4% on PaperBench. Skills aren't seasoning; they roughly doubled benchmark performance while the model stayed identical.
Step back and the shape is familiar: this is what mattpocock/skills and the whole skills-repo boom do by hand, done autonomously at library scale. The skills-as-standard thread now has an industrial supply side. Paper at https://arxiv.org/abs/2609.02749.
Comments