AREX-Skill: A Skill Library for Automated Machine Learning and Auto-Research

Introduction

AREX-Skill is an innovative project by VectorSpaceLab that transforms the vast knowledge embedded in GitHub repositories and research papers into executable skills for coding agents and autonomous machine learning research. This library provides over 5,000 verified, executable skills distilled from more than 1,000 popular ML repositories, enabling agents like Codex, Claude Code, and Pi to perform complex tasks more efficiently. It aims to bridge the gap between human-readable research knowledge and agent-usable operational intelligence, significantly accelerating auto-research across multidisciplinary machine learning studies.

Why Use and Benefits

Traditional research knowledge, while abundant, is primarily structured for human consumption. Agents often struggle to navigate papers, repositories, and blogs to extract applicable workflows, debug issues, and validate results. AREX-Skill addresses this by converting "operating knowledge" into self-contained, agent-readable units called AREX Skills. Each skill captures when a capability applies, what to run, how to validate it, and how to recover from failures, using the open Agent Skills format.

The benefits are clearly demonstrated in auto-research benchmarks:

Significant Performance Gains: Benchmarks show Codex with AREX-Skill achieves substantial improvements, such as a 134.3% gain in "Any Medal rate (%)" on MLE-bench and a 34.4% gain in "Replication score" on PaperBench.
Enhanced Task Handling: Skills help agents avoid expensive unguided trial-and-error and recover from near-failure states, especially on difficult tasks.
Productive Resource Allocation: A relevant skill graph allows agents to reach useful solution spaces earlier, dedicating more budget to experiments and validation.

The library covers a wide array of domains, including ML engineering, LLMs, computer vision, data science, and scientific computing, making it a versatile tool for various autonomous research applications.

Installation

To get started with AREX-Skill, you first need to install DisCo, the command-line interface and runtime.

Install DisCo: DisCo requires Node.js >=22.19.0. You can use the managed curl installer on macOS, Linux, WSL, or Git Bash:
```
curl -fsSL https://github.com/VectorSpaceLab/AREX-Skill/releases/latest/download/install-disco.sh | sh
```
Alternatively, install directly with npm:
```
npm install -g --ignore-scripts @arex-skill/disco
```
Configure your model provider (e.g., OpenAI, Anthropic, Gemini) on first run or via environment variables.
Install the library and start Researcher mode:
```
disco repo-skills install
disco
```
DisCo's default Researcher mode will load and route the AREX-Skill Library. You can then try concrete tasks at the prompt.
Import selected skills into another coding agent (Optional): To import skills for agents like Codex or Claude Code, use the creator mode: For Codex:
```
disco --creator -p "/skill:import-repo-skills-to-agent import vllm and sglang to ~/.agents"
```
For Claude Code:
```
disco --creator -p "/skill:import-repo-skills-to-agent import vllm and sglang to ~/.claude"
```
Remember to restart your target agent after importing new skills.

Examples

AREX-Skill can be applied to diverse machine learning workflows. Here are a couple of examples:

High-throughput inference comparison: You can use vLLM and SGLang skills to guide a controlled serving comparison:

Compare vLLM and SGLang on this model and workload. Tune both under identical hardware and memory constraints, report verified throughput, and preserve the commands and measurements needed to reproduce the comparison.

Protein structure modeling: AlphaFold2 skills provide operational guidance for protein-structure modeling workflows:

Use the installed AlphaFold2 skills to set up and verify this protein-structure modeling workflow. Start with a tiny synthetic input, check the sequence/MSA shapes and dependencies, run the relevant model path, and report the commands and checks needed to reproduce the result. Do not treat untrained outputs as scientific predictions.

The library also includes capabilities from FAISS, Unsloth, Diffusers, LeRobot, AlphaFold2, and many more, accessible through the repository catalog.