It has previously been established that pupil dilation tracks reward prediction errors (RPEs) and surprise during reinforcement learning and that this relationship is scaled by confidence. Whether and how these interactions change across learning however, is unclear. This study aims to replicate that confidence modulates post-feedback pupil responses to RPEs, with the post-feedback pupil dilation being hypothesized to be larger for high confidence decisions, and investigates whether this modulation changes across learning, expecting the modulation to diminish at a high learning level. 

A sample of 26 young adults completed a four-choice probabilistic reinforcement learning task requiring high- or low-confidence choices while undergoing monocular eye-tracking. Post-feedback pupil dilation was analysed using linear mixed-effects models incorporates scaled Q-difference scores from a reinforcement learning model to track learning. Results confirmed a main effect of unsigned RPE on post-feedback pupil dilation, which was significantly modulated by confidence, where RPE predicted post-feedback pupil dilation for high-confidence choices but not for low-confidence ones. 

Further, a significant three-way interaction was found between RPE, confidence, and scaled Q-difference scores, demonstrating learning alters this relationship. Contrary to the initial hypothesis that post-feedback pupil dilation, scaling surprise, would desensitize as learning increases, pupil sensitivity to RPEs during low confidence decisions increased at high learning, reaching levels comparable to those during high confidence ones. These findings demonstrate that post-feedback pupil dilation is modulated by confidence and changes during learning, potentially reflecting a shift from resolving estimation uncertainty during early task stages to evaluating unexpected outcomes once value representations are established.