Skip to content

XAI-liacs/LLaMEA

BranchesTags

Repository files navigation

Shows the LLaMEA logo.

LLaMEA: Large Language Model Evolutionary Algorithm

⭐ If you like this, please give the repo a star – it helps!

PyPI version Downloads per month Maintenance Python 3.10+ Test DOI Open In Colab Docs
Check our demo on Colab

The fully-open successor to Google DeepMind’s AlphaEvolve for automated algorithm discovery. First released 📅 Nov 2024 • MIT License • 100 % reproducible. Also read the Documentation. 🥈 Winner of the Silver Humies 2025 at GECCO!

LLaMEA couples large-language-model reasoning with an evolutionary loop to invent, mutate and benchmark algorithms fully autonomously.

Table of Contents

Introduction

LLaMEA (Large Language Model Evolutionary Algorithm) is an innovative framework that leverages the power of large language models (LLMs) such as GPT-4 for the automated generation and refinement of metaheuristic optimization algorithms. The framework utilizes a novel approach to evolve and optimize algorithms iteratively based on performance metrics and runtime evaluations without requiring extensive prior algorithmic knowledge. This makes LLaMEA an ideal tool for both research and practical applications in fields where optimization is crucial.

Key Features:

  • Automated Algorithm Generation: Automatically generates and refines algorithms using GPT-based or similar LLM models.
  • Performance Evaluation: Integrates seamlessly with the IOHexperimenter for real-time performance feedback, guiding the evolutionary process.
  • LLaMEA-HPO: Provides an in-the-loop hyper-parameter optimization mechanism (via SMAC) to offload numerical tuning, so that LLM queries focus on novel structural improvements.
  • Extensible & Modular: You can easily integrate additional models and evaluation tools.
  • Niching for Diversity: Fitness sharing and clearing strategies maintain a diverse set of solutions.
  • Unified Diff Mode: Evolve code through patch-based edits for efficient token usage.
  • Population Evaluation Mode: Optionally evaluate whole populations in a single call to speed up expensive fitness functions.

LLaMEA framework

Example use-cases:

  • Problem specific optimization algorithms: Easily generate and fine-tune optimization algorithms to work on your specific problem. By leveraging problem knowledge in the prompt the generated optimized algorithms can perform even better.
  • Efficient new Bayesian Optimization algorithms: Generate optimized and novel Bayesian optimization algorithms, specifically for optimizing very expensive problems such as auto-motive crash worthiness or car shape design optimization tasks.
  • Machine Learning Pipelines: Without any ML knowledge, you can use LLaMEA to generate optimized machine learning pipelines for any task. Just insert the task description and provide the dataset and evaluation metric and start LLaMEA.

🔥 News

🎁 Installation

Important

Ensure that SWIG is installed on your system by running swig --version If it is not installed, use the appropriate method for your platform:

  • macOS: brew install swig
  • Ubuntu/Linux: sudo apt-get install swig (or your distribution’s package manager)
  • Windows: Download and install from the official SWIG website and make sure to add it to your system PATH.

It is the easiest to use LLaMEA from the pypi package.

  pip install llamea

Important

The Python version must be higher or equal to 3.11. You need an OpenAI/Gemini/Ollama API key for using LLM models.

For a slimmed-down installation without bundled LLM or HPO dependencies, install the lite package:

  pip install llamea-lite

HPO features in the lite package require installing ConfigSpace separately.

You can also install the package from source using uv (0.7.19). make sure you have uv installed.

  1. Clone the repository: 
    git clone https://github.com/XAI-liacs/LLaMEA.git
cd LLaMEA
  2. Install the required dependencies via uv: 
    uv sync
  3. Optional install dev or/and example dependencies: 
    uv sync --dev --group examples

💻 Quick Start

Tip

See also the getting started demo:

Open In Colab

  1. Set up an OpenAI API key:

    • Obtain an API key from OpenAI.
    • Set the API key in your environment variables: 
      export OPENAI_API_KEY='your_api_key_here'

  2. Running an Experiment

    To run an optimization experiment using LLaMEA:

    from llamea import LLaMEA

# Define your evaluation function
def your_evaluation_function(solution):
    # Implementation of your function
    # return feedback, quality score, error information
    return "feedback for LLM", 0.1, ""

# Initialize LLaMEA with your API key and other parameters
optimizer = LLaMEA(f=your_evaluation_function, api_key="your_api_key_here")

# Run the optimizer
best_solution, best_fitness = optimizer.run()
print(f"Best Solution: {best_solution}, Fitness: {best_fitness}")

⚙️ Configuration

Key hyper-parameters of LLaMEA:

Parameter Description
n_parents, n_offspring Population sizes controlling selection and mutation
budget Number of generations to run
niching Diversity management strategy (None, "sharing", "clearing") with niche_radius and related options
evaluate_population If True, the evaluation function f operates on a list of solutions
diff_mode Requests unified diff patches instead of full code for mutations
HPO Enable in-the-loop hyper-parameter optimization
eval_timeout, max_workers, parallel_backend Control evaluation time and parallelism
adaptive_mutation, adaptive_prompt Adaptive control of mutation strength and task prompt

See the Documentation for a complete description of all parameters.

💻 Examples

Below are two example scripts from the examples directory demonstrating LLaMEA in action for black-box optimization with a BBOB (24 noiseless) function suite. One script (examples/black-box-optimization.py) runs basic LLaMEA, while the other (examples/black-box-opt-with-HPO.py) incorporates a hyper-parameter optimization pipeline—known as LLaMEA-HPO—that employs SMAC to tune the algorithm’s parameters in the loop.

Running black-box-optimization.py

black-box-optimization.py showcases a straightforward use-case of LLaMEA. It:

  • Defines an evaluation function evaluateBBOB that runs generated algorithms on a standard set of BBOB problems (24 functions).
  • Initializes LLaMEA with a specific model (e.g., GPT-4, GPT-3.5) and prompts the LLM to generate metaheuristic code.
  • Iterates over a (1+1)-style evolutionary loop, refining the code until a certain budget is reached.

How to run:

uv run python examples/black-box-optimization.py

The script will:

  1. Query the specified LLM with a prompt describing the black-box optimization task.
  2. Dynamically execute each generated algorithm on BBOB problems.
  3. Log performance data such as AOCC (Area Over the Convergence Curve).
  4. Iteratively refine the best-so-far algorithms.

Running black-box-opt-with-HPO.py (LLaMEA-HPO)

black-box-opt-with-HPO.py extends LLaMEA with in-the-loop hyper-parameter optimization—termed LLaMEA-HPO. Instead of having the LLM guess or refine hyper-parameters directly, the code:

  • Allows the LLM to generate a Python class representing the metaheuristic plus a ConfigSpace dictionary describing hyper-parameters.
  • Passes these hyper-parameters to SMAC, which then searches for good parameter settings on a BBOB training set.
  • Evaluates the best hyper-parameters found by SMAC on the full BBOB suite.
  • Feeds back the final performance (and errors) to the LLM, prompting it to mutate the algorithm’s structure (rather than simply numeric settings).

Why LLaMEA-HPO? Offloading hyper-parameter search to SMAC significantly reduces LLM query overhead and encourages the LLM to focus on novel structural improvements.

How to run:

uv run python examples/black-box-opt-with-HPO.py

Script outline:

  1. Prompt & Generation: Script sets up a role/task prompt, along with hyper-parameter config space templates.
  2. HPO Step: For each newly generated algorithm, SMAC tries different parameter values within a budget.
  3. Evaluation: The final best configuration from SMAC is tested across BBOB instances.
  4. Refinement: The script returns the performance to LLaMEA, prompting the LLM to mutate the algorithm design.

Note

Adjust the model name (ai_model) or API key as needed in the script. Changing budget or the HPO budget can drastically affect runtime and cost. Additional arguments (e.g., logging directories) can be set if desired.

Running automl_example.py

automl_example.py uses LLaMEA to showcase that it can not only evolve and generate metaheuristics but also all kind of other algorithms, such as Machine Learning pipelines. In this example, a basic classification task on the breast-cancer dataset from sklearn is solved by generating and evolving open-ended ML pipelines.

  • We define the evaluate function (accuracy score on a hold-out test set)
  • We provide a very basic example code to get the algorithm started.
  • We run a few iterations and observe the excellent performance of our completely automatic ML pipeline.

How to run:

uv run python examples/automl_example.py

Note

Adjust the model name (ai_model) or API key as needed in the script.

Viewing conversation logs

The repository provides a minimal Flask app in logreader/app.py to explore conversation logs stored as JSON Lines files. Start the server with a log file path:

uv run python logreader/app.py --logfile path/to/conversationlog.jsonl

You can also set the environment variable CONVERSATION_LOG instead of passing --logfile. If neither is given, the app defaults to conversationlog.jsonl in the current working directory. Navigate to http://localhost:5001 to browse the messages.

🤖 Contributing

Contributions to LLaMEA are welcome! Here are a few ways you can help:

  • Report Bugs: Use GitHub Issues to report bugs.
  • Feature Requests: Suggest new features or improvements.
  • Pull Requests: Submit PRs for bug fixes or feature additions.

Please refer to CONTRIBUTING.md for more details on contributing guidelines.

🪪 License

Distributed under the MIT License. See LICENSE for more information.

🤖 Reproducability

Each paper we published also has an accompanying Zenodo repository for full reproducability of all our results.

✨ Citation

If you use LLaMEA in your research, please consider citing the associated paper:

@ARTICLE{van2025llamea,
  author={Stein, Niki van and Bäck, Thomas},
  journal={IEEE Transactions on Evolutionary Computation},
  title={LLaMEA: A Large Language Model Evolutionary Algorithm for Automatically Generating Metaheuristics},
  year={2025},
  volume={29},
  number={2},
  pages={331-345},
  keywords={Benchmark testing;Evolutionary computation;Metaheuristics;Codes;Large language models;Closed box;Heuristic algorithms;Mathematical models;Vectors;Systematics;Automated code generation;evolutionary computation (EC);large language models (LLMs);metaheuristics;optimization},
  doi={10.1109/TEVC.2024.3497793}
}

If you only want to cite the LLaMEA-HPO variant use the folllowing:

@article{van2024intheloop,
  author = {van Stein, Niki and Vermetten, Diederick and B\"{a}ck, Thomas},
  title = {In-the-loop Hyper-Parameter Optimization for LLM-Based Automated Design of Heuristics},
  year = {2025},
  publisher = {Association for Computing Machinery},
  address = {New York, NY, USA},
  url = {https://doi.org/10.1145/3731567},
  doi = {10.1145/3731567},
  note = {Just Accepted},
  journal = {ACM Trans. Evol. Learn. Optim.},
  month = apr,
  keywords = {Code Generation, Heuristic Optimization, Large Language Models, Evolutionary Computation, Black-Box Optimization, Traveling Salesperson Problems}
}

Other works about extensions or integrations of LLaMEA:

@InProceedings{yin2024controlling,
  author="Yin, Haoran and Kononova, Anna V. and B{\"a}ck, Thomas and van Stein, Niki",
  editor="Garc{\'i}a-S{\'a}nchez, Pablo and Hart, Emma and Thomson, Sarah L.",
  title="Controlling the Mutation in Large Language Models for the Efficient Evolution of Algorithms",
  booktitle="Applications of Evolutionary Computation",
  year="2025",
  publisher="Springer Nature Switzerland",
  address="Cham",
  pages="403--417",
  isbn="978-3-031-90065-5"
}

For more details, please refer to the documentation and tutorials available in the repository.

flowchart LR
    A[Initialization] -->|Starting prompt| B{Stop? fa:fa-hand}
    B -->|No| C(Generate Algorithm - LLM )
    B --> |Yes| G{{Return best so far fa:fa-code}}
    C --> |fa:fa-code|D(Evaluate)
    D -->|errors, scores| E[Store session history fa:fa-database]
    E --> F(Construct Refinement Prompt)
    F --> B
Loading

CodeCov test coverage

About

Large Language Model Evolutionary Algorithm

Resources

Readme

License

MIT license

Contributing

Contributing
Activity
Custom properties

Stars

84 stars

Watchers

2 watching

Forks

16 forks
Report repository

Releases 21

v1.1.10 Latest
Oct 23, 2025
+ 20 releases

Packages

No packages published

Contributors 6

Languages