FunSearch is a framework for discovering mathematical solutions through program search guided by large language models (LLMs). It enables researchers and practitioners to leverage LLMs for automated mathematical discovery across various combinatorial problems, such as constructing large cap sets and admissible sets, and solving bin packing challenges. The project facilitates reproducible research by providing implementations and datasets associated with a Nature publication.

How It Works

The core of FunSearch involves an evolutionary algorithm combined with code manipulation routines. It uses LLMs to generate candidate programs (solutions) and a sandbox environment to evaluate them. The system iteratively refines these programs based on performance metrics, demonstrating a novel approach to automated mathematical discovery by integrating symbolic reasoning with the generative capabilities of LLMs.

Quick Start & Requirements

No installation is required; all functionality is accessible via Google Colab notebooks. Specific notebooks are provided for cap_set, admissible_set, bin_packing, cyclic_graphs, and corner_free_sets. Users will need a Google account to run these notebooks.

Highlighted Details

• Accompanies a publication in Nature (Romera-Paredes et al., 2023), validating its research significance.
• Demonstrates AI-driven discovery for complex combinatorial problems including cap sets, admissible sets, bin packing heuristics, and independent sets in cyclic graphs.
• Provides numerical datasets for discovered mathematical structures for convenience and further analysis.

Maintenance & Community

The repository is maintained by Google DeepMind. No specific community channels (like Discord or Slack) or detailed roadmap information are provided in the README.

Licensing & Compatibility

Software components are licensed under the Apache License, Version 2.0 (Apache 2.0), which is permissive for commercial use. Other materials are under the Creative Commons Attribution 4.0 International License (CC-BY).

Limitations & Caveats

The provided implementation directory is a component of the full FunSearch pipeline and notably excludes the language models, code execution sandbox, and distributed infrastructure. Users must provide or integrate these components themselves. The code appears to be research-oriented and may require adaptation for production environments.