EASE-ing into global optimization with LLMs : A competition entry on LLM-designed evolutionary algorithms at the Genetic and evolutionary computation conference (GECCO) 2025

Abstract

This paper presents an extended abstract describing an entry into the LLM for evolutionary algorithms tailored benchmarking competition using a GNBG-generated test suite. This study explores the generative design of optimization algorithms using large language models (LLMs) within the EASE modular framework, which supports iterative prompting and feedback-driven refinement. Across five successive generations, we observed a progressive transformation in algorithmic structure. The findings suggest opportunities for further research into the role of prompt design, feedback phrasing, and framework architecture in guiding the emergence of more task-adaptive, domain-specialized algorithmic behavior.