VendiEvolve: Towards Diversity-Aware LLM-Guided Evolutionary Search

LLM-guided evolutionary search has shown promise for discovering programs for complex scientific and engineering design problems. However, existing systems often emphasize finding the single best program, while paying less attention to the population of candidates from which that program emerges. This population is itself a useful artifact: it reveals search dynamics, exposes failure modes, and provides alternative candidates for downstream inspection. We hypothesize that shaping the distribution of candidate programs can improve both search performance and our understanding of the discovery process. Our approach, VendiEvolve, augments LLM-guided evolutionary search with label-free distributional diversity signals that diagnose population-level collapse and keep semantically distinct programs influential as context for future LLM mutations. We evaluate VendiEvolve on two complementary domains: circle packing, a constrained combinatorial optimization task with a quantitative fitness function, and 3D asset generation, an open-ended design task with qualitative evaluation. In our experiments, distributional diversity complements existing LLM-guided evolutionary heuristics, improving final performance on circle packing and preserving useful intermediate structures in open-ended generation.