Proceedings of the Annual Meeting of the Cognitive Science Society

Semantic fluency tasks reveal the temporal structure of memory retrieval through "clustering and switching'' behavior. Existing models rely on static similarity (e.g., cosine similarity in Word2Vec) or heuristic switching rules, failing to capture this sequential decision-making process. Recently, fine-tuned semantic vectors to better fit transitions, but assumed a single context-independent weighting throughout the task. We propose Deep Inverse Reinforcement Learning for Semantic Fluency (DIRL-SF), modeling the participant as an agent navigating a semantic graph to maximize an unknown reward function. A Transformer encoder captures retrieval history, while a dynamic attention mechanism re-weights semantic dimensions per step, accounting for similarity, novelty, and effort in a state-dependent manner. On a 27-category dataset, DIRL-SF significantly outperforms static and LSTM baselines, providing a theoretically grounded tool for comparing search strategies across populations.