Sophie Su¹, Yining Ding, Jeffrey Zacks; ¹Washington University in Saint Louis, ²Washington University in Saint Louis, ³Washington University in Saint Louis

Humans segment continuous experience into discrete events, but the cognitive signals that trigger segmentation remain unclear. We tested whether gaze entropy—a measure of spatial dispersion in eye movements—indexes moment-to-moment predictive uncertainty during passive viewing of naturalistic activities. Eighty-one participants viewed everyday activity videos while gaze was recorded. A separate group provided fine and coarse event segmentation judgments. We derived computational measures of prediction uncertainty and prediction error using Structured Event Memory (SEM) model variants (Nguyen et al, 2024）. Mixed-effects models revealed that gaze entropy was positively related to computational prediction uncertainty, but not prediction error. This suggests that gaze entropy reflects internal prediction uncertainty, distinct from prediction error. In addition, gaze entropy robustly predicted coarse event boundary probability but not fine-grained boundaries. Time-series FIR analyses further showed that gaze entropy reliably increased prior to both fine and coarse boundaries and decreased afterward, revealing a proactive uncertainty signal that precedes segmentation. Fixation duration and saccade amplitude showed weaker temporal relationship with event boundaries. These findings support a proactive control account of event segmentation in which rising predictive uncertainty triggers boundary detection and situation model updating during naturalistic perception.