Item

Abstract

Large language models appear quite creative, often per- forming on par with the average human on creative tasks. However, research on LLM creativity has fo- cused solely on products, with little attention on the creative process. Process analyses of human creativ- ity often require hand-coded categories or exploit re- sponse times, which do not apply to LLMs. We pro- vide an automated method to characterise how humans and LLMs explore semantic spaces on the Alternate Uses Task, and contrast with behaviour in a Verbal Flu- ency Task. We use sentence embeddings to identify response categories and compute semantic similarities, which we use to generate jump profiles. Our results cor- roborate earlier work in humans reporting both persis- tent (deep search in few semantic spaces) and flexible (broad search across multiple semantic spaces) path- ways to creativity, where both pathways lead to sim- ilar creativity scores. LLMs were found to be biased towards either persistent or flexible paths, that varied across tasks. Though LLMs as a population match hu- man profiles, their relationship with creativity is differ- ent, where the more flexible models score higher on cre- ativity. Our dataset and scripts are available on GitHub.