This study investigates the hypothesis that the human brain utilizes a hierarchical predictive coding system for language processing, unlike current language models that primarily predict nearby words. Using fMRI data from 304 participants listening to short stories, the researchers demonstrate that enhancing language models with multi-timescale predictions improves their alignment with brain responses. Furthermore, they reveal a hierarchical organization of predictions, with frontoparietal cortices predicting higher-level, longer-range representations than temporal cortices. This supports the role of hierarchical predictive coding in human language processing.

Charlotte Caucheteux, Alexandre Gramfort, Jean-Rémi King