Proceedings of the Annual Meeting of the Cognitive Science Society

When learning from educational content, we must simultaneously remember individual pieces of information and integrate them into a coherent conceptual framework. Previous work has shown that semantic network structure predicts what people remember from narratives, but whether these same principles extend to academic learning remains unknown. In this study, participants watched a 15-minute physics lecture and completed free recall during fMRI scanning. Using Sentence-BERT embeddings to quantify semantic centrality, we found that central information was better recalled and, critically, participants who recalled more units of central information demonstrated better conceptual understanding as rated by human raters (r = .73). In a voxel-wise encoding analysis, we modeled brain responses during lecture viewing using the same SBERT features, and this brain-model mapping predicted behavioral outcomes in selective cortical regions. These findings suggest that the embedding space of large language models reflects relevant aspects of how our minds organize newly learned conceptual information.