Large Language Model for Multiobjective Evolutionary Optimization

Multiobjective evolutionary algorithms (MOEAs) are major solutions for solving multiobjective optimization problems (MOPs). However, to design efficient and powerful operators in MOEAs, a tedious process of trial-and-error with domain experts is usually required. This paper investigates a novel approach that leverages large language models (LLM) to design MOEA operators. Firstly, with proper prompt engineering, we directly employ an LLM to serve as a black-box search operator for decomposition-based MOEA in a zero-shot manner. In this way, we can get rid of the time-consuming manual operator design process. In addition, we further design a white-box operator to interpret and approximate the behaviour of LLM and propose a new version of decomposition-based MOEA, termed MOEA/D-LMO. This white-box operator improves generalization and reduces runtime by eliminating costly LLM interaction. Experimental studies on different test benchmarks show that our proposed method can achieve competitive performance with widely used MOEAs. Furthermore, the operators only learned from LLMs on a few instances have good generalization performances on unseen problems with different patterns and settings. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2025.