Repo-To-Skill: Distilling GitHub Repositories Into AI4AI Skills

TL;DR - Repo-To-Skill introduces DisCo, an autonomous ML research agent that distills operational knowledge from repositories into compact, verified skills. Adding these skills substantially improves research-agent performance without changing the model backbone, harness, or execution budget.

• DisCo supports task-agnostic distillation of widely used repositories and task-oriented skill creation for specific research problems.
• The resulting AREX-Skill Library contains over 5,000 verified skills from 1,000 ML repositories, spanning 20 areas and 178 capability families.
• With GPT-5.5 and other experimental conditions fixed, skills improved scores by 134.3% on MLE-bench and 34.4% on PaperBench.
• The same setup also produced gains of 9.2% on FrontierCS and 14.0% on PassNet, indicating that reusable operating context can reduce repeated implementation work.