Schedule of Events | Symposia

Tracking the temporal dynamics of conceptual learning during a STEM lecture

Poster Session B - Sunday, March 30, 2025, 8:00 – 10:00 am EDT, Back Bay Ballroom/Republic Ballroom
Also presenting in Data Blitz Session 4 - Saturday, March 29, 2025, 10:30 am – 12:00 pm EDT, Constitution B.

Yeongji Lee¹ (yeongji.lee.gr@dartmouth.edu), David Kraemer¹; ¹Dartmouth College

During a science lecture, successful understanding emerges as individual pieces of information are revisited, interconnected, and integrated into a unified network during the course of the lecture. In this fMRI study, participants watched a video lecture on Newtonian physics concepts and then verbally recalled what they remembered and learned from the lesson while still inside the scanner. Using the embedding space of a large language model (LLM), we first used representational similarity analysis (RSA) to identify brain regions where patterns of neural activity reflected the semantic network structure of the lecture, supporting successful comprehension as measured by verbal recall performance. In a separate analysis, we then used a voxelwise forward encoding model to predict the degree to which participants understand each underlying concept as they are built up over time throughout the lecture. We fitted a linear mapping to predict neural responses from LLM-derived semantic features, and then this mapping was aligned again with human-labeled individual concepts. Whereas the RSA approach reveals where in the brain neural patterns reflect the overall semantic organization of concepts discussed during the whole lecture, the voxelwise timecourse analysis enabled us to quantify understanding of specific concepts at given time points during the lecture, and then to compare these estimates with post-lecture quiz scores and verbal recall performance. The findings demonstrate that mapping the timecourse of neural representations with a large language model provides a powerful tool to characterize individual differences in conceptual understanding by revealing the temporal dynamics of abstract knowledge construction during learning.

Topic Area: THINKING: Reasoning