QD-LLMs: Quality-Diversity Optimization with LLMs for Generative Design Exploration
Abstract
Recent text-to-X generative models can synthesize design modalities such as images and 3D objects from free-form natural-language prompts, opening new opportunities for engineering design exploration. However, Quality-Diversity (QD) methods lack effective variation operators for free-form prompt genotypes that reliably generate high-performing, domain-valid 3D design phenotypes. We propose QD-Optimization with LLMs for generative design exploration (QD-LLMs), a QD framework that uses large language models to perform natural-language-coded variation while maintaining domain alignment through physical and vision-language evaluations. QD-LLMs includes an LLM allocator that selects elite LLM emitters to improve search performance. These specialized emitters choose parent prompts with high potential to produce valid, high-performing designs across both explored and unexplored regions of the target feature space. Experiments on vehicle design exploration show that QD-LLMs improves the discovery of high-performing, diverse designs by over 80% compared with iterative zero-shot LLM prompting and other relevant QD methods. These results demonstrate the effectiveness of QD-LLMs for principled exploration of diversity and fitness optimization in text-to-X generative design.
Authors: Ariq Koh, Melvin Wong, Jiao Liu, Caishun Chen, Thiago Rios, Stefan Menzel, Yew-Soon Ong
Published in: Genetic and Evolutionary Computation Conference Companion (GECCO Companion) (2026)