Two forms of uncertainty hierarchically shape lexical predictions during natural speech processing

Prediction under uncertainty is a hallmark of human cognition(Friston, 2010), yet how the brain represents uncertainty remains unresolved(Bach and Dolan, 2012). Speech provides a powerful test case: lexical predictions involve thou-sands of alternatives, making classical Shannon entropy an ambiguous measure of uncertainty. Here we explored a generalization of Shannon entropy, the Renyi entropy family (Renyi, 1961), combined with information-theoretic tools and whole-brain intracranial EEG during naturalistic listening. Word-related neural activity was best explained by two complementary forms of uncertainty: dis-persion, reflecting the number of plausible alternatives, and strength, reflecting the probability of the most likely word. These uncertainties were consistently encoded in each participant by distinct neural populations, with opposing com-putational goals. Dispersion shapes anticipatory sensory activity and dampens responses to unexpected inputs, consistent with Bayesian optimization(Pouget et al., 2013). Strength instead emerges only for surprising words in lexical circuits, amplifying prediction errors and promoting information seeking. Despite bidirec-tional exchange between these two populations, their dynamics are dominated by local computations(Heilbron et al., 2022), challenging purely hierarchical models of predictive processing. Our findings reveal that the brain approximates large lexical probability distributions using multiple summary statistics, providing a flexible, low-dimensional representation of uncertainty in complex cognitive tasks and advancing principles of predictive processing across cognition.