The paper presents DisCo, a research agent that distills operational knowledge from GitHub repositories into reusable AI skills using two complementary approaches: task-agnostic condensation of widely-used ML repositories and task-oriented skill generation for specific problems. The authors built the AREX-Skill Library, containing over 5,000 verified skills distilled from 1,000 ML repositories across 20 research areas. A GPT-5.5-backed agent equipped with these skills scored 134.3% higher on MLE-bench and 34.4% higher on PaperBench than without them, under the same computational budget.
