The dynamic relationship between pupil dilation and neural surprise in natural language comprehension

Predictive processing accounts hold that the brain continuously generates predictions and updates internal models from error signals, but it remains unclear whether prediction representation and model-updating reflect a single graded computation or computationally distinct operations implemented in dissociable systems. We recorded magnetoencephalography (MEG) and pupillometry while participants listened to continuous natural speech, quantifying word-by-word lexical surprise and semantic prediction error with a GPT-2 model. Broadly speaking, lexical surprise was tracked in a predominantly graded fashion, whereas the response to semantic error was better captured by a rectified linear (ReLU) gating function - i.e., that responds only when the error exceeds a recent contextual baseline. There were also some interesting dissociations in the strength of these relative effects. The auditory cortex tracked lexical surprise in a graded, continuous manner, with later cortical stages reflecting semantic error. By contrast, tonic pupil diameter and source-localised brainstem responses were predominantly captured by a gated response to semantic error. A pupil-coupling analysis confirmed that this gating signature was statistically shared between pupil-linked arousal and brainstem-localised, but not cortical, responses. Together, these findings reveal a division of labour in which the cortex maintains a continuous, high-fidelity map of predictive information while a pupil-linked model-update system acts as a selective gate, engaged specifically by meaning-disrupting events that have crossed a threshold level of error. This asymmetry suggests that signals serving continuous parsing of prediction error and signals serving model revision may be computationally dissociated, with a range of implications for our understanding of the intrinsic interactions between mechanisms serving perception and learning.