Cognitive Control as Precision Regulation: From Predictive Coding to Adaptive Behaviou

Cognitive control is traditionally conceptualized as a set of top-down mechanisms that selectand regulate stimulus–response mappings to support goal-directed and flexible behaviour.Traditional accounts have provided fundamental insights into the neural architecture ofcognitive control and identified a set of core control processes, including goal maintenance, taskswitching, control allocation, conflict monitoring, and cognitive effort. However, they rarelyaddress how these functions can be integrated within a common computational architecture.Furthermore, most theories tend to assume that the core control functions use specializedmechanisms that are fundamentally distinct from those supporting action–perception loops.Here, we review an alternative perspective grounded in predictive coding and active inference,according to which cognitive control emerges from hierarchical elaborations of the samegenerative inference supporting perception and action. Within predictive coding and activeinference, the brain continuously generates predictions about sensory inputs, actions, andtheir expected consequences, and updates these predictions on the basis of precision-weightedprediction errors. We argue that the same precision-weighting mechanism, when operatingbeyond sensory processing—to optimize the precision of higher-level beliefs and of courses ofaction (policies)—provides the foundations of cognitive control, allowing a cognitive system tobalance the stability of current goals and task representations against the flexibility requiredto respond to unexpected information and environmental changes. From this perspective, thecore control processes emerge as different facets of precision regulation within a hierarchicalgenerative model: the precision of higher-level beliefs supports goal maintenance and taskswitching; policy precision supports control allocation, setting the balance between habitualand goal-directed control and prioritizing the latter when required by task demands; conflictmonitoring provides information about whether the selected policy continues to produceexpected and preferred outcomes; and cognitive effort reflects the cost of increasing precision tochange the resulting control configuration. Finally, we link the precision-based computations toputative neural processes in frontoparietal networks and neuromodulatory processes spanningdopaminergic, noradrenergic, and cholinergic systems, and derive empirical predictions thatcan test the proposed framework.