LLM-assisted interactive evolutionary exploration for aesthetic preference clarification: an exploratory field case

Interactive evolutionary exploration elicits choices between alternatives, while perceptual and semantic approaches articulate the qualities underlying those choices. We examine how a multimodal large language model (LLM), mediated by a designer, connects these activities and allows exploration rules to change. In an exploratory field case, the author helped a subject clarify an initially vague Art Deco ring preference. Eight populations of nine existing rings were presented over eight days. Each daily session allowed 15–20 min for reflection, selections and optional semantic comments. Visual selections, human comments and LLM hypotheses progressively revised the preference frame and retrieval instructions. We reconstruct who initiated consequential changes. Retrospective visual descriptors were coded by the LLM; dissimilarities were then computed using explicit formulas. Mean distance to a subsequently purchased ring decreased by 28.0%, while internal diversity increased by 48.8%. A designer-initiated reopening temporarily increased both measures. The subject reported finding her dream ring in less than two hours on a street containing four jewellery stores. The principal contribution is a framework coupling population, preference frame and exploration instructions through dialogic operator adaptation. Semantic conditioning and exploration reopening support this process. The outcome-aware single case supports feasibility and situated usefulness.