Lloyd A. Thompson¹, Stephanie M. Shields¹, Kathryn Bonnen¹; ¹Indiana University Bloomington

Understanding gaze allocation during real-world behavior requires measurements that go beyond static image-viewing and controlled laboratory tasks. The Visual Experience Database (Greene, Balas, Lescroart, et al., 2024) provides the opportunity to analyze how eye movements unfold during natural everyday activities via head-mounted video and binocular gaze data collected from 56 participants across hundreds of labeled tasks. Here, we asked whether task-specific fixation priors reflect stable task structure and improve gaze prediction. We developed an automated pipeline for processing VEDB sessions, temporally aligning binocular gaze to video frames and identifying recordings that met inclusion criteria (>10k gaze samples, ≥40% of samples localized with high confidence). This yielded 375 usable sessions containing one or more labeled task segments. Tasks were grouped into 10 categories: locomotion/navigation, household chores, food preparation, sports/games, cognitively demanding tasks, social interaction, object manipulation, artistic tasks, technology use, and other. We computed a global egocentric fixation prior by pooling valid gaze samples across all task segments, providing a static baseline for the overall spread of gaze. We then aligned the global prior to the center of gaze for each task segment to generate a dynamic set of context-specific priors. We evaluated the predictive value of static and dynamic priors using AUC (area under the curve) and NSS (normalized scan-path saliency), computing results per task segment before pooling per task category. Dynamic priors consistently outperformed the static prior. Gains were strongest for tasks involving strong visuomotor structure (e.g., object manipulation, household chores, sports/games, and cognitively demanding tasks) and weakest for tasks like social interaction and locomotion/navigation. These results demonstrate that gaze in natural behavior is organized around stable, task-dependent fixation strategies rather than a single uniform center bias. Our findings highlight the importance of dynamic, context-conditioned priors for next-generation models of naturalistic gaze behavior.