Reproduce “Coherence in the brain unfolds across separable temporal regimes” on OpenResearch

About this paper

Davide Staub, Finn Rabe, Akhil Misra, Yves Pauli, Roya Hüppi, Ni Yang, Nils Lang, Lars Michels and 5 more

Researchers at the University of Zurich leveraged large language models and ultra high-field fMRI to reveal that human brain coherence during naturalistic language comprehension unfolds across two distinct neural temporal regimes: gradual semantic integration ('drift') in heteromodal association cortices and rapid event-boundary detection ('shift') in peri-Sylvian language regions.

Key insights

• Narrative coherence relies on the dynamic interplay of two separable neural temporal regimes: a slow, context-integrating process ('drift') and a fast, event-boundary-linked process ('shift').
• Large language models can generate robust, annotation-free computational signals that effectively capture distinct temporal components of human language comprehension beyond simple feature prediction.
• The findings align with hierarchical temporal receptive window accounts, demonstrating that early auditory-language regions track rapid event changes, while higher-order association cortices handle slower, finite-timescale contextual integration.

Summary by alphaXiv.