Constructing Differential Evolution via LLM Prompt Chaining

This paper presents an extended abstract, designed as a competition entry on LLM-designed Evolutionary Algorithms. We present DEuLLM, a differential evolution (DE) constructed entirely through prompt engineering with the help of a large language model (LLM). DEuLLM integrates key advances in DE, such as success-history-based parameter adaptations, distance-based feedback, linear population size reduction, and careful boundary control, without any code design outside the LLM-driven dialogue. The development process demonstrates how iterative prompting can guide the design of a competitive optimizer from scratch through natural language interaction. The results show that our proposed DEuLLM is able to locate the global optimum for 15 out of 24 test problems.