VectorSpaceLab · 2026-09-02 · notable
Repo-To-Skill — 5,000 verified skills distilled from 1,000 ML repos
Repo-To-Skill introduces DisCo, an agent that turns GitHub repositories into reusable skills for ML research agents. The released AREX-Skill library holds 5,000+ verified skills from 1,000+ repos and lifts MLE-bench scores by 134.3%.
DisCo distils GitHub repos into verified, executable skills that ML research agents load on demand.
Key specs
| Skills | 5,000+ |
|---|---|
| Repos distilled | 1,000+ |
| Mle bench gain | +134.3% |
What is it?
Repo-To-Skill goes after what its authors call operational knowledge — 'the know-how that separates knowing a method from making it work'. The paper introduces DisCo, an agent that reads a repository and its paper and writes out reusable skills. The team also released AREX-Skill, a library of more than 5,000 verified skills distilled from over 1,000 widely used machine-learning repositories, spread across 20 areas and 178 package families.
How does it work?
Distillation runs in four stages — scoping, evidence grounding, graph construction and verification — so each skill is checked by running it, not just written down. A skill is self-contained, carrying its scope, routing information, workflows, validation steps and recovery guidance. At use time the library narrows a request by progressive disclosure, from area to family to repository to workflow, before loading any detail, which keeps the agent's context small.
Why does it matter?
Research agents normally rediscover a library's quirks by trial and error on every new task. With the distilled skills attached, the paper reports gains of 134.3% on MLE-bench, 34.4% on PaperBench, 9.2% on FrontierCS and 14.0% on PassNet over the same agent without them. The repository is Apache-2.0, but each skill carries its own licence in its SKILL.md file and that per-skill licence takes precedence.
Who is it for?
agent and AutoML researchers
Try it
npm install -g --ignore-scripts @arex-skill/disco