README.md

April 7, 2026 · View on GitHub

[ICLR 2026] Generalizable Heuristic Generation Through LLMs with Meta-Optimization

License     Paper     License     License

This repository contains the official implementation of MoH (Meta-optimization of Heuristics), a novel framework for generating heuristics through Large Language Models with meta-optimization techniques.

Overview

MoH Overview

Installation

# Clone the repository
git clone https://github.com/yiding-s/MoH.git
cd MoH

# Install with uv (recommended)
uv sync

# Or install with pip
pip install -e .

Dependencies: numpy, tqdm, hydra-core, omegaconf, openai, numba, func-timeout (see pyproject.toml for full list). Require Python >= 3.10.

Datas: You can download them from GoogleDrive and put data folder of each problem under problems/<problem_name>, the scripts to generate the data will also be updated in this repository soon.

Configuration

MoH uses Hydra for configuration management. All config files are located in the cfg/ directory:

cfg/
├── config.yaml              # Main config (iterations, population size, eval_calls_limit, etc.)
├── llm_client/
│   ├── openai.yaml           # OpenAI API config (e.g., gpt-4o)
│   └── ...           # more LLM configurations
├── problem/
│   └── tsp_gls.yaml          # Problem-specific config (TSP-GLS)
└── hydra/output/
    └── local.yaml            # Output directory config

**Note: More problems and LLM support coming soon! Since this project has been refactored from the original version, feel free to raise an issue if encounter any unexpected issues. **

LLM Client Config

OpenAI API (cfg/llm_client/openai.yaml):

model: gpt-4o
batch_size: 5
api_key: ${oc.env:OPENAI_API_KEY}   # Set via environment variable

Problem Config

Each problem is defined in cfg/problem/. Example (tsp_gls.yaml):

problem_name: tsp_gls
obj_type: min
problem_size: [100, 200]

Usage

# Run with default config (can also modify config in cfg/)
python main.py

# Override parameters from command line
python main.py n_iterations=20 

# Use local LLM
python main.py llm_client@heu=local llm_client@meta=local

# Change problem size
python main.py problem.problem_size='[50, 100]'

Results are saved to outputs/<problem_name>-<problem_size>/<timestamp>/.

Citation

@inproceedings{
shi2026generalizable,
title={Generalizable Heuristic Generation Through {LLM}s with Meta-Optimization},
author={Yiding Shi and Jianan Zhou and Wen Song and Jieyi Bi and Yaoxin Wu and Zhiguang Cao and Jie Zhang},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=tIQZ7pVN6S}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgement