Advances in deep learning have led to the development of predictive deep language models (DLMs). This study investigates whether the human brain and autoregressive DLMs share computational principles when processing natural narratives. Nine participants underwent electrocorticography (ECOG) while listening to a podcast. The findings show that both the human brain and DLMs engage in continuous next-word prediction, match pre-onset predictions to incoming words to calculate surprise, and use contextual embeddings to represent words. This suggests that autoregressive DLMs offer a biologically plausible framework for understanding language neural processing.

Ariel Goldstein, Zaid Zada, Eliav Buchnik, Mariano Schain, Amy Price, Bobbi Aubrey, Samuel A. Nastase, Amir Feder, Dotan Emanuel, Alon Cohen, Aren Jansen, Harshvardhan Gazula, Gina Choe, Aditi Rao, Catherine Kim, Colton Casto, Lora Devore, Adeen Flinker, Liat Hasenfratz, Omer Levy, Avinatan Hassidim, Sasha Fanda, Werner Doyle, Daniel Friedman, Patricia Dugan, Lucia Melloni, Roi Reichart, Michael Brenner, Yossi Matias, Kenneth A. Norman, Orrin Devinsky, Uri Hasson