AIAny
Icon for item

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

Distills operational know‑how from ML GitHub repositories into compact, verified 'skills' that research agents can load and reuse. Produces a skill format (SKILL.md, references, scripts), the AREX‑Skill Library (5,000+ skills from 1,000 repos), and demonstrates sizable benchmark gains when agents use skills.

Introduction

Most ML research agents combine a reasoning backbone with an orchestration harness, yet still fail because the field’s operational know‑how — the pragmatic, repository‑level details that make methods actually work — is missing from an agent’s context. This paper’s core insight is simple but consequential: extract and verify that know‑how into compact, discoverable skills so agents stop rediscovering the same engineering tricks and misconfigurations across tasks. The authors implement this idea end‑to‑end (DisCo) and produce a large, verified skill corpus (AREX‑Skill Library) to quantify the effect.

Key Findings
  • Distillation at scale: task‑agnostic distillation across 1,000 widely used ML repositories yields the AREX‑Skill Library with 5,000+ verified skills organized into 20 areas and 178 capability families — so what? agents can hold many skills yet read only the few relevant to a task, keeping context small.
  • Task‑oriented distillation: on‑demand skill creation tailors operational knowledge to a concrete research problem, reducing wasted trial‑and‑error during experiments.
  • Verified impact: holding model backbone (GPT‑5.5), harness, and execution budget constant, adding skills raises scores by large margins on multiple benchmarks (e.g., +134.3% MLE‑bench, +34.4% PaperBench) — so what? the gains come from supplying execution‑ready context rather than larger models or more compute.
  • Skill design: each skill packages a short SKILL.md (interface), references (substrate), and scripts (execution), enabling progressive disclosure and automated checks so agents can safely consume thousands of skills.
Who it's for and tradeoffs

Great fit if you build or evaluate autonomous research agents, maintain agent knowledge bases, or need repeatable, repository‑level operating context for ML experiments. The approach accelerates agent workflows and reduces wasted runs by encoding common APIs, pitfalls, and pipeline settings. Look elsewhere if your tasks require always‑fresh repository content without a maintenance pipeline: skills can go stale across releases, and safe ingestion requires auditing (security and supply‑chain concerns). Building and verifying large skill corpora also demands upfront engineering and verification budgets.

Where It Fits

This work sits between model improvements and orchestration engineering: instead of changing the reasoning backbone or the harness, it supplies the missing operational layer that maps methods and tools to runnable procedures. It complements retrieval/RAG approaches by packaging executable and verified operational context rather than raw docs.

How It Works (brief)

Distillation pipelines parse repos, extract candidate operational items, generate concise SKILL.md interfaces backed by references and scripts, and apply automated verification and repair. Distillation runs both ahead‑of‑time (task‑agnostic library) and on demand (task‑oriented skills), with a router that selects relevant skill graphs at runtime.

Information

  • Websitearxiv.org
  • OrganizationsBeijing Academy of Artificial Intelligence, University of Science and Technology of China, Renmin University of China, Hong Kong Polytechnic University
  • AuthorsJianlyu Chen, Yuyang Hu, Hongjin Qian, Jiawei Liu, Wenqing Wei, Xiaolong Chen, Defu Lian, Zhicheng Dou, Chaozhuo Li, Qiwei Ye
  • Published date2026/09/02

Categories

More Items

Predicts an LLM agent's final success or failure from partial execution traces and halts runs when outcomes are confident to save per-task compute. Uses LightGBM success/failure classifiers on behavioral, textual, and reference features; cuts 13–26% steps and up to 44% input tokens on benchmarks.

Autonomous multimodal GUI agent that executes natural-language interface tasks across mobile apps, web domains, and desktop OS. Expands environment coverage (170+ multilingual apps, 4,000+ web domains), uses function-grounded task generation and keypoint-based multi-model verification to produce reliable RL rewards for real-world deployment.

Turns each research paper into a training environment to generate verifiable research plans by synthesizing questions from goals/background and deriving evaluation criteria from methods/experiments. Key features: four-stage extraction that reduces criterion leakage to 3.7%, a two-stage rubric-centered training (self-distillation then GRPO), and the PaperGym-20k corpus with two held-out benchmarks.