Fig. 2 |. fMRI results.

Visualizations for four (S1–S4) of the nine participants, with the remainder available in Supplementary Figs. 3–5. Voxel-level significance is FDR corrected at level 0.01. a, Voxels significantly (sig.) predicted by full-context embeddings (blue), residual-context embeddings (red) or both (white), visualized in Montreal Neurological Institute space. Most of the temporal cortex and inferior frontal gyri are predicted by full context embeddings, with residual context embeddings mostly predicting a subset of those areas. b, ROI-level results. Top: language system ROIs²⁹ and two semantic ROIs³¹. Bottom: proportion of ROI voxels significantly predicted by full context (left) and residual context (right) embeddings. The investigated ROI correspond to the bilateral dorsomedial prefrontal cortex (DMPFC), posterior cingulate cortex (PC), inferior frontal gyrus pars orbitalis (IFGorb), middle frontal gyrus (MFG), inferior frontal gyrus (IFG), angular gyrus (AG), anterior temporal lobe (ATL) and posterior temporal lobe (PTL). Displayed are the median proportions across all nine participants and the 95% confidence interval on the median. Full context predicts all ROIs (ROI-level Holm–Bonferroni correction, P < 0.05), while residual context predicts only bilateral ATL and PTL (significant predictions indicated by an asterisk). c, Spatial generalization matrices. Models trained to predict PTL voxels are used to predict PTL and ATL voxels, within participant (left) and across participants (right). PTL cross-voxel correlations form two clusters: models that predict activity for voxels in one cluster can also predict activities of other voxels in the same cluster, but not activities for voxels in the other cluster. Across participants, only one of these clusters has voxels that predict ATL voxels. d, Performance of models trained on ATL and PTL voxels at predicting other participants’ ATL. All participants show a cluster of predictive voxels in the pSTS.