Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Shared and domain-specific neural dynamics of narrative comprehension and mathematical reasoning
Poster Session E, Friday, October 2, 11:00 am - 1:00 pm, Wangari Maathai
Yike Wang1, Zhengwu Ma1, Chengcheng Wang1, Jixing Li1; 1City University of Hong Kong
Introduction. Understanding how the human brain supports complex reasoning across different cognitive domains remains a central question. Narrative comprehension and mathematical reasoning both require the construction and integration of structured representations over time, yet they differ in their reliance on linguistic and abstract relational processes. Recent advances in large language models (LLMs) provide a new computational framework for characterizing such representational dynamics. In the present study, we combined magnetoencephalography (MEG) source analysis with LLM-derived embeddings to investigate the temporal and regional correspondence between human neural activity and LLM representations during narrative comprehension and mathematical reasoning. Methods. Thirty-one native Chinese speakers (15 females; mean age = 26.48 years, SD = 3.93) participated in the MEG experiment. Participants listened to the complete Mandarin audiobook of The Little Prince, divided into nine sections of approximately 10 minutes each. After each section, they answered four multiple-choice narrative comprehension questions presented on the screen. After the audiobook task, participants also answered 10 multiple-choice mathematics questions. We prompted Qwen3-8B with the same narrative comprehension and mathematics questions. Model responses were divided into five sequential processing steps, capturing the progression from problem understanding to final answer generation. Embeddings were extracted from all model layers at each step. For each trial, MEG responses were epoched from question onset to the participant’s button press. Because reaction times varied across trials, each trial was temporally normalized by resampling the source-level activity into 100 evenly spaced time points, allowing neural responses to be compared across trials on a common time scale. We selected eight bilateral regions of interest from the HCP-MMP1 combined atlas, covering a temporal–frontal–parietal network. Representational dissimilarity matrices (RDMs) were computed separately for model embeddings and MEG source activity. Model RDMs were calculated across questions for each layer and processing step, whereas neural RDMs were calculated across questions for each ROI and time bin. Representational similarity analysis was then performed using Spearman correlation to compare neural representations with LLM-derived embeddings, yielding time-resolved similarity maps for each processing step, ROI, and time point. Group-level statistics were conducted across participants to identify significant temporal clusters associated with narrative comprehension and mathematical reasoning. Results. For the narrative comprehension task, significant effects were observed in the left inferior frontal gyrus (IFG) and bilateral temporal regions. The left IFG showed sustained significant correlations across all processing steps, primarily around time points 20–40. In contrast, the right IFG showed a more restricted effect, reaching significance only at Step 1 around time points 35–45. Bilateral middle temporal regions showed significant effects during Steps 1–4, mainly within time points 20–50. For the mathematics task, significant effects were largely left-lateralized. Across Steps 1–3, significant correlations shifted from temporal regions toward the left IFG within time points 15–30, suggesting a progression from early semantic processing to later integrative computation. Together, these findings suggest that both narrative comprehension and mathematical reasoning engage a shared left frontal–temporal network, while narrative comprehension additionally recruits right temporal regions, potentially reflecting contextual and discourse-level integration.
Topic Areas: Computational Approaches,