Abstract
Adaptive behavior requires integrating information from multiple sources. These sources can originate from distinct channels, such as internally maintained latent cognitive representations or externally presented sensory cues. Because these signals are often stochastic and carry inherent uncertainty, integration is challenging. However, the neural and computational mechanisms that support the integration of such stochastic information remain unknown. We introduce a computational neuroimaging framework to elucidate how brain systems integrate internally maintained and externally cued stochastic information to guide behavior. Neuroimaging data were collected from healthy adult human participants (both male and female). Our computational model estimates trial-by-trial beliefs about internally maintained latent states and externally presented perceptual cues and then integrates them into a unified joint probability distribution. The entropy of this joint distribution quantifies overall uncertainty, which enables continuous tracking of probabilistic task beliefs, prediction errors, and updating dynamics. Results showed that latent-state beliefs are encoded in distinct regions from perceptual beliefs. Latent state beliefs were encoded in the anterior middle frontal gyrus, mediodorsal thalamus, and inferior parietal lobule, whereas perceptual beliefs were encoded in spatially distinct regions including lateral temporo-occipital areas, intraparietal sulcus, and precentral sulcus. The integrated joint probability and its entropy converged in frontoparietal hub areas, notably middle frontal gyrus and intraparietal sulcus. These findings suggest that frontoparietal hubs read out and resolve distributed uncertainty to flexibly guide behavior, revealing how frontoparietal systems implement cognitive integration.