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

Repo-To-Skill. Distilling GitHub Repositories Into AI4AI Skills. By Jianlyu Chen and ten others, from the Beijing Academy of Artificial Intelligence, the University of Science and Technology of China, and two other institutions. Published on the second of September. Its code repository is worked out: about eight thousand source files, with a readme and a licence.

Autonomous agents are starting to carry out machine-learning research from end to end. They pair a model backbone with a harness that handles planning, execution, memory and verification. That architecture still leaves something outside the agent, and the authors give it a name: operational knowledge. It is the domain-specific know-how that separates knowing a method from making it work. Knowing what a technique is and getting it to run properly on a real task are different kinds of knowledge, and only the first tends to come with the backbone.

The field does have this knowledge. It lives in repositories and in papers. But it is written there for human readers, and in forms far too large to load in the middle of a task. The argument is that once it is distilled into compact, verified skills, it can be reused across tasks instead of being rediscovered on every run.

DisCo is a research agent built around skills: it creates them, and it uses them while doing research. Its distillation comes in two complementary forms. The task-agnostic form condenses the field's widely used repositories into reusable skills, ahead of any particular task. The task-oriented form produces the skills that a concrete task calls for, when that task arrives.

Applied across the open ecosystem, the task-agnostic form yields the AREX-Skill Library. It holds more than five thousand verified skills, distilled from one thousand widely used machine-learning repositories. Those skills are organised into twenty areas and one hundred and seventy-eight capability families, so that the know-how scattered across individual codebases ends up in a structure an agent can draw from.

The comparison is set up to isolate the skills. The backbone, GPT-5.5, is held fixed, as are the research harness and the downstream execution budget. The only difference is whether the agent has skills. With them, it scores one hundred and thirty-four point three percent higher on MLE-bench, and thirty-four point four percent higher on PaperBench.

On FrontierCS the gain is nine point two percent, and on PassNet it is fourteen percent. The authors attribute these gains to adding distilled operating context under that fixed setup, rather than to a stronger model or a larger budget.

The paper's claim is about where the missing capability lives. The agent did not need a better backbone or a different harness. It needed the practical knowledge that already sits in the field's repositories, put into a form small enough and verified enough to be used mid-task. Distil that once, and every later run can start from it rather than rediscover it.