How Task Representations Integrate Information from Multiple Sources

Adaptive behavior requires integrating multiple sources of information, yet neural and computational mechanisms underlying this process are poorly understood. We introduce a novel Bayesian computational framework elucidating how internally maintained and externally cued information are integrated to guide goal-directed actions. Our model infers trial-by-trial beliefs for inputs and integrates them into a joint probability distribution, where entropy quantifies overall uncertainty. This integrated representation predicts a probabilistic task belief and generates continuous measures of prediction errors and updating. Decoding analyses revealed distinct neural substrates encoded external and internal inputs, with convergence in frontoparietal hub regions where activity fluctuations scaled with entropy. Increasing integrated uncertainty modulated connectivity patterns, biasing the generation of integrated task outputs and prediction errors. Together, these findings capture the full integration process—encoding inputs, forming an integrated representation jointly encoding uncertainty, and producing a unified task output. This work provides novel insights into neural mechanisms underlying cognitive integration.