Distinct hierarchical effects of goal specificity and reward uncertainty on decision-making: An active inference...

Adaptive behaviour depends on the interplay between bottom-up processing of environmental inputs and top-down adjustments of internal representations. However, how reward uncertainty (as a constraint on sensory evidence) and goal setting (as a top-down constraint on policy precision) jointly shape this hierarchical organisation remains unclear. To address this question, we combined EEG and fNIRS with a partially observable Markov decision process model within an active inference framework during a probabilistic reversal learning task performed under concurrent auditory interference. Participants completed the task under two levels of reward uncertainty and were assigned either a nonspecific ("do your best") or a specific performance goal. EEG captured trial-by-trial dynamics using linear mixed-effects models informed by computational metrics, whereas fNIRS characterised block-level cortical entropy and fronto-parietal network topology using graph theory. Results revealed a dissociated pattern across hierarchical levels of decision-making. Specific goal setting improved accuracy and was associated with more stable top-down policy selection, reflected in choice-locked P3 amplitude modulations, reduced centro-parietal beta desynchronisation during choice preparation, and increased HbO entropy within the DLPFC. Goal specificity also modulated feedback-related computational quantities, including outcome surprise and belief entropy, although these effects were not accompanied by corresponding goal-related modulations of feedback-locked EEG signals. Reward uncertainty was primarily associated with bottom-up belief updating, as evidenced by belief-entropy effects on the feedback-locked P3b and greater fronto-parietal global efficiency with shorter path length under high uncertainty. Together, these findings support a relative dissociation, with goals primarily stabilising action policies and uncertainty predominantly modulating feedback-driven belief updating.