Proceedings of the Annual Meeting of the Cognitive Science Society

Modeling cognitive dynamics in web media exploration can contribute to healthier information environments, for example by enabling inference of users' false beliefs and estimating how media designs may affect belief formation. We propose Free-Energy-driven Web Explorer (F-Explorer), a cognitive LLM agent that instantiates active inference for web exploration. In contrast to data-driven LLM approaches that learn behavior end-to-end, F-Explorer takes a theory-driven approach: it embeds an LLM as the belief module within a cognitive model grounded in the Free-Energy Principle (FEP), extending prior work that models human web exploration under FEP. We evaluate F-Explorer on a virtual SNS exploration dataset from a previous study and compare it with a reproduction of WEB-FEP with aligned initial-belief distributions. Using maximum likelihood estimation over 840 candidate settings per model (initial belief x learning rate), F-Explorer achieves higher action-sequence log-likelihood and improved top-k choice accuracy. Moreover, a proxy belief measure shows larger correspondence with participants' self-reported initial opinion scores, while correspondence with final opinion scores remains comparable. These results suggest that integrating LLM-based belief representations can improve FE–based models of web exploration while preserving interpretability through explicit latent-state and parameter inference.