The predictive processing theory proposes that minimization of prediction error is a governing principle of perception, learning, and behavior, yet few studies have directly tested this hypothesis in time perception. We operationalized prediction error minimization as updating of an internal temporal reference and examined whether reference updating follows the direction required to reduce systematic prediction error. Across two experiments, 60 university students (Experiment 1) and 359 adults from a broader age range (Experiment 2) performed a temporal generalization task. Reference updating was assessed from changes in distributional asymmetry and peak shifts of temporal generalization gradients. This dissociated reference updating from possible perceptual bias that would alter gradient steepness without shifting the peak. In Experiment 1, frequent presentation of longer or shorter durations produced expected directional shifts of the internal temporal reference. In Experiment 2, consistent directional shifts emerged only when prediction error was sufficiently large, whereas smaller prediction error failed to induce reliable shifts. Together, these findings demonstrate that internal temporal reference updating in sub-second range time perception depends on the direction and magnitude of prediction error, providing a direct behavioral test of the prediction error minimization hypothesis.
