Proceedings of the Annual Meeting of the Cognitive Science Society

Category fluency tasks, which involve retrieval from semantic memory, help reveal the clustered structure of human semantic memory. However, traditional coding schemes for this task rely on fixed groupings that do not capture humans' varied semantic clusters and retrieval strategies. We evaluate whether the attention patterns of Large Language Models (LLMs) can help generate tailored coding schemes that better reveal the structure of semantic memory within and across people. Using the Hills et al. (2012) data, we prefill each human-generated sequence into an LLM to derive a per-sequence coding scheme. These attention-derived groups (from middle model layers) show inter-item response time increases at subcategory boundaries -- a key prediction of optimal foraging theory -- that are up to 30% greater than when using the standard, fixed Hills et al. (2012) scheme. We also evaluate the cognitive plausibility of LLM-generated sequences, finding that they are broadly consistent with the predictions of optimal foraging theory.