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

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

A Low-Rank GPT-2 XL Subspace Captures Linear Alignment with Intracranial Responses during Language Listening

Zachary Paris¹, Emily S Finn¹; ¹Dartmouth College

Presenter: Zachary Paris

Large language models (LLMs) are increasingly used as computational models of human language processing, yet what drives their alignment with neural activity remains unclear. Here, we use ridge reduced-rank regression on intracranial ECoG recordings from nine participants listening to a naturalistic podcast to identify the subspace of GPT-2 XL embeddings responsible for neural prediction. We find that approximately 20 dimensions out of 1,600 are sufficient to capture the linearly encodable neural variance of GPT2-XL layer 24, and that dimensions outside this subspace do not improve encoding performance. These results suggest that, in this setting, linear LLM–brain alignment is concentrated in a compact representational subspace.

Topic Area: Auditory, Speech & Language Processing