Repolode
Submit
GAIR-NLP

GAIR-NLP/ASI-Evolve

A general agentic framework that runs an autonomous loop of learning, hypothesis design, experiment execution and analysis to search for better solutions to any measurable problem.

View on GitHub
Stars
870
Forks
218
Open issues
0
Last commit
12 days ago

What is it?

ASI-Evolve is an agentic framework that repeats a loop of retrieving prior knowledge, proposing a candidate program, running it against an evaluation script, and writing a structured lesson from the outcome. It was built for AI research but is written to work for any problem where a program can be scored automatically.

The loop is driven by three agents: a Researcher that reads the experiment database and a cognition store to propose the next candidate, an Engineer that executes it and collects metrics, and an Analyzer that turns results into reusable lessons. Every trial is stored with its motivation, code, result and analysis, and parent selection can use UCB1, greedy, random or MAP-Elites island sampling.

Users set up an experiment folder with a problem description, a baseline program, an evaluator script and optional domain knowledge, then run the pipeline for a chosen number of steps. The repository includes a circle-packing demo and works with any OpenAI-compatible LLM endpoint. A lighter Agent Skill version is also provided for quick trials.

Key features

  • 01Runs an automated learn, design, experiment and analyze loop over multiple rounds.
  • 02Uses three agents: Researcher, Engineer and Analyzer.
  • 03Stores every experiment with its motivation, code, metrics and analysis in a database.
  • 04Supports UCB1, greedy, random and MAP-Elites island parent sampling.
  • 05Lets users seed a cognition store with papers or heuristics before the search starts.
  • 06Works with any OpenAI-compatible API endpoint, including local models via LiteLLM.

Use cases

  • Tuning an inference scheduler or batching policy against a throughput and latency benchmark.
  • Searching for better algorithms in optimization or simulation problems that have a scoring script.
  • Exploring model architectures or training-data pipelines where each trial can be evaluated automatically.

Quick start

  1. 1

    Clone the repository

    git clone https://github.com/GAIR-NLP/ASI-Evolve.git
  2. 2

    Enter the project directory

    cd ASI-Evolve
  3. 3

    Install Python dependencies

    pip install -r requirements.txt
  4. 4

    Initialize the demo cognition store and run 10 evolution steps

    python experiments/circle_packing_demo/init_cognition.py

Things to know

  • Needs an OpenAI-compatible LLM API endpoint, which may incur costs.
  • Requires you to supply an evaluation script and domain knowledge for your own problems.
  • Requires Python 3.10+ and bash on the system path.

Topics

#agents#llm#automated-research#evolutionary-search#experiment-tracking#optimization

FAQ

What is ASI-Evolve?
A general agentic framework that runs an autonomous loop of learning, hypothesis design, experiment execution and analysis to search for better solutions to any measurable problem. ASI-Evolve is an agentic framework that repeats a loop of retrieving prior knowledge, proposing a candidate program, running it against an evaluation script, and writing a structured lesson from the outcome. It was built for AI research but is written to work for any problem where a program can be scored automatically.
Is ASI-Evolve free to use? What license does it have?
ASI-Evolve is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is ASI-Evolve still maintained?
Actively maintained. The last commit to ASI-Evolve was on Oct 9, 2026.
What can I use ASI-Evolve for?
Tuning an inference scheduler or batching policy against a throughput and latency benchmark. Searching for better algorithms in optimization or simulation problems that have a scoring script. Exploring model architectures or training-data pipelines where each trial can be evaluated automatically.
How do I install ASI-Evolve?
Clone the repository: git clone https://github.com/GAIR-NLP/ASI-Evolve.git Enter the project directory: cd ASI-Evolve Install Python dependencies: pip install -r requirements.txt Initialize the demo cognition store and run 10 evolution steps: python experiments/circle_packing_demo/init_cognition.py
Where can I get ASI-Evolve?
The source code and installation instructions are on GitHub: https://github.com/GAIR-NLP/ASI-Evolve

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects