Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster D in Poster Session D: Wednesday, August 5, 2:00 – 3:45 pm, Kimmel Center, Shorin & Rosenthal Rooms

Neural Alignment Reveals Selective Reorganization in Domain-Adapted Language Models

Miriam Havin¹, Refael Tikochinski², Ariel Goldstein¹; ¹Hebrew University of Jerusalem, ²University College London

Presenter: Miriam Havin

Neural alignment analyses have shown that large language models reliably predict brain activity during naturalistic comprehension and are increasingly used to compare and evaluate model representations. However, alignment scores are often treated as a single measure of “brain-likeness,” making it hard to understand what differences between models actually mean computationally or cognitively. Here, we propose using known functional differences between brain networks to interpret model differences more directly. Focusing on the language-selective and domain-general multiple-demand (MD) networks, which support dissociable aspects of comprehension, we examine how domain-specific pretraining alters neural alignment during technical lecture viewing. Using voxelwise fMRI encoding models, we compare contextual embeddings from a general-purpose model (BERT) and a domain-adapted model (SciBERT) while participants viewed extended and short lectures in computer science and history. Across datasets, SciBERT showed a selective advantage over BERT in predicting activity in the MD network during computer science lectures, while the two models performed similarly in the language network. This dissociation was strongest in deeper transformer layers and did not generalize to non-technical content. Complementary representational analyses revealed systematic reorganization of embedding geometry following domain-adaptive pretraining. Together, these findings suggest that domain-specific pretraining selectively changes representations relevant for effortful, integrative comprehension, while leaving core linguistic tracking relatively stable. More broadly, the results show how network-specific neural alignment can help explain not just whether models differ, but how they differ.

Topic Area: Methods, Tools, Theory & Neural Coding