Pupil responses reveal temporally distinct signatures of value updating and decision strategy

Abstract

Adaptive decision making depends on dopaminergic and noradrenergic systems supporting value learning and exploratory decision strategies, respectively, yet their contributions remain difficult to dissociate non-invasively. Here, we investigated whether pupil dynamics provide dissociable readouts of computational processes underlying value-based decision making across sensory modalities, potentially reflecting distinct neuromodulatory processes. Human participants performed a dynamic foraging task in which they chose between auditory, visual, or audio-visual options based on their estimated value, alongside a control task where choices were instructed. Behavior in the value-based task was well captured by a probabilistic choice model, revealing adaptive integration of reward history and a balance between exploration and exploitation. Pupil responses revealed temporally distinct computational signatures of decision strategy and value updating. Reaction time was associated with sustained pupil dilation throughout the decision process, whereas value differences between chosen and unchosen options selectively modulated pupil responses during stimulus evaluation and following feedback. These findings are consistent with computational processes linked to noradrenergic regulation of exploration-exploitation behavior and dopaminergic value updating, respectively. Both effects were significantly stronger during value-based than instructed decisions, indicating enhanced engagement of these computational processes when choices depended on learned reward values. Importantly, these effects were largely modality-independent, indicating a domain-general encoding of computational variables. Together, these findings identify pupil dynamics as a temporally sensitive and non-invasive marker of distinct computational stages underlying adaptive decision making and establish a framework for linking pupillometry to neuromodulatory theories of value learning and uncertainty processing.