Disentangling the temporal signatures of conditioned pupil dilation: Distinct valence- and prediction-error-related components revealed by mega-analysis

Abstract

Pupil dilation offers a sensitive, non-invasive window into the neurocognitive processes supporting human associative learning. Yet, the field lacks consensus on how to quantify conditioned pupil responses, particularly which time intervals best serve as readouts. Moreover, it remains unresolved whether pupil dilation during learning reflects a unitary arousal signal or multiple (overlapping yet distinct) processes that unfold sequentially (e.g., cue valuation, uncertainty-driven attention, anticipation, prediction-error [PE] signaling). To address this, we conducted a large-scale individual-participant-data reanalysis (mega-analysis) of harmonized primary studies from our lab (K = 7 independent samples; total N = 385; 562 sessions), reprocessed through an identical pipeline, with systematic variation in outcome valence and modality. Temporal PCA of epochs locked either to the conditioned (CS) or unconditioned stimulus (US) yielded highly stable component structures across trials and studies, which carried dissociable information: An early (CS-locked) component selectively tracked appetitive value, distinguishing reward cues before aversive differentiation emerged. Later phases reflected differential conditioning across both valences, increasingly coupled to arousal toward US onset. Aversive learning was most specifically expressed in responses peaking around expected US delivery, which also predicted negative valence ratings. Two subsequent components captured unconditioned responding and its modulation by unexpected outcomes and omissions, consistent with unsigned PE signaling (amplified after aversive cues) and, later, positive PEs. Generalized additive mixed models largely converged with the PCA structure. Pupil dilation is therefore not a unitary marker of conditioning, but rather a high-resolution readout of sequential processes such as valuation, arousal/attention, and outcome updating, offering a promising framework for future computational models of learning.