LLaMEA: Automatically Generating Metaheuristics with Large Language Models

Genetic and Evolutionary Computation Conference Companion (GECCO Companion), pp. 81–82

Abstract

This paper summarizes our recent work published in IEEE Transactions on Evolutionary Computation (2025), entitled "LLaMEA: A Large Language Model Evolutionary Algorithm for Automatically Generating Metaheuristics". In this work we propose a novel evolutionary algorithm framework that leverages large language models (LLMs) like GPT-4 to automatically generate, evaluate, and refine metaheuristic optimization algorithms. Evaluations on the BBOB benchmark suite demonstrate that algorithms automatically discovered by LLaMEA can outperform established state-of-the-art methods, such as CMA-ES and DE, especially on lower-dimensional optimization problems. Our results illustrate the significant potential of using LLMs for automated algorithm design, setting a new direction for evolutionary computation research.

Authors 2

1. Bas van Stein Aachen

   Leiden University

   Affiliation as printed

   Leiden University, Leiden, Netherlands

2. Thomas Bäck Aachen

   Leiden University

   Affiliation as printed

   Leiden University, Leiden, Netherlands