Coherence in the brain unfolds across separable temporal regimes

Abstract

Coherence in language requires the brain to satisfy two competing temporal demands: gradual accumulation of meaning across extended context and rapid reconfiguration of representations at event boundaries. Despite their centrality to language and thought, how these processes are implemented in the human brain during naturalistic listening remains unclear. Here, we tested whether these two processes can be captured by annotation-free drift and shift signals and whether their neural expression dissociates across large-scale cortical systems. These signals were derived from a large language model (LLM) and formalized contextual drift and event shifts directly from the narrative input. To enable high-precision voxelwise encoding models with stable parameter estimates, we densely sampled one healthy adult across more than 7 hours of listening to thirteen crime stories while collecting ultra high-field (7T) BOLD data. We then modeled the feature-informed hemodynamic response using a regularized encoding framework validated on independent stories. Drift predictions were prevalent in default-mode network hubs, whereas shift predictions were evident bilaterally in the primary auditory cortex and language association cortex. Furthermore, activity in default-mode and parietal networks was best explained by a signal capturing how meaning accumulates and gradually fades over the course of the narrative. Together, these findings show that coherence during language comprehension is implemented through dissociable neural regimes of slow contextual integration and rapid event-driven reconfiguration, offering a mechanistic entry point for understanding disturbances of language coherence in psychiatric disorders.

Figures (9)

• Figure 1: Annotation-free mapping of narrative coherence to brain dynamics. Crime stories were processed by a decoder-only large language model (LLM) to derive two complementary signals of narrative structure: drift, capturing the gradual accumulation of contextual meaning, and shift, capturing discrete event boundaries. These LLM-derived signals were aligned to the spoken narratives presented during fMRI and entered into an encoding model to predict voxelwise BOLD responses. Comparing predicted and observed activity allowed us to map distinct neural systems associated with gradual contextual integration and rapid event-driven reconfiguration across the cortex.
• Figure 2: Across stories activation and consistency maps for drift and shift.a) Cortical surface projections showing Z-scores for drift. b) Volumetric region of interest maps showing the cross-story consistency of effects, where color intensity represents the number of stories (out of 13) in which each region was significant (Simes $P < 0.05$). The drift maps showed comparatively weaker and more distributed patterns with lower consistency, involving higher-level hubs like the angular gyrus and precuneus (default mode network), consistent with integrative or attentional components rather than primary auditory drive. c) Cortical surface projections displaying Z-scores for shift. d) The shift maps exhibited robust, bilateral modulation within classic speech areas around the Sylvian fissure (e.g., Heschl’s gyrus and STG), showing high consistency across up to 13 stories.
• Figure 3: Unique predictive contributions of drift and shift.a) Voxelwise significance map for unique drift effects, localized primarily to heteromodal association hubs, including angular gyrus and precuneus. b) Corresponding ROI-level effect sizes for drift in posterior-inferior default-mode regions (DMN–PI). c) Voxelwise significance map for unique shift effects, estimated after regressing out shared variance with drift. Effects concentrate bilaterally in the peri-Sylvian language network, including Heschl’s gyrus, superior temporal cortex, and planum temporale/polare. d) ROI-level effect sizes for shift in language-network ROIs (LANG), showing mean unique regression weights ($\beta$) across 13 stories with 95% bootstrap confidence intervals.
• Figure 4: Replicable drift effects across integration timescales. Cross-story generalization of drift effects as a function of the memory decay parameter $\rho$a) Raw sums of region of interest-level Simes significant-story counts (not normalized for ROI number or size). b) Size-weighted mean significant stories per region, normalizing for both region number and size. After normalization, drift generalized more strongly in posterior-inferior default-mode and parietal regions (DMN–PI) than in the language network (LANG), with a broad maximum at intermediate integration timescales ($\rho\approx0.10$–$0.20$).
• Figure S4: Mean frame-wise displacement in BOLD data.