@inproceedings {pub6581,
	title = {QD-LLMs: Quality-Diversity Optimization with LLMs for Generative Design Exploration},
	author = {Ariq Koh AND Melvin Wong AND Jiao Liu AND Caishun Chen AND Thiago de Jesus de Araujo Rios AND Stefan Menzel AND Yew Soon Ong},
	year = {2026},
	month = {July},
	abstract = {Text-to-3D generative models synthesize 3D geometries from free-form natural language prompts, enabling the exploration of vast generative design spaces. These models present an opportunity in engineering design to systematically search for diverse, high-performing candidates within a target feature space. While the Quality-Diversity (QD) optimization paradigm supports structured diversity maintenance through discretized grid-based archives, existing QD variation operators are ill-suited for natural language, where small perturbations in text embedding space can synthesize domain-invalid or incoherent artifacts. Large Language Models (LLMs) alleviate this through learned semantic priors and conditional generation capabilities. To this end, we propose Quality Diversity Optimization with LLMs for Generative Design Exploration (QD-LLMs), a novel framework that systematically explores a predefined target feature space while ensuring domain validity of diverse designs through vision-language evaluations. QD-LLMs comprises an LLM allocator that selects elite LLM emitters to maximize search performance. These emitters strategically select parents with high potential to yield high-performing, domain-valid designs not only in explored regions but also in new regions of the target feature space. These key components enable natural-language-coded variation operations that effectively explore the vast design spaces of text-to-3D generative models.  Systematic experiments on aerodynamic vehicle design scenario demonstrates that QD-LLMs substantially outperform both iterative zero-shot LLM prompting and latent-space QD methods, achieving 7.4x higher performance in generating valid, aerodynamically optimized designs than CMA-ME while maintaining comprehensive coverage of the measure space.},
	publisher = {ACM},
	booktitle = {The Genetic and Evolutionary Computation Conference}
}