The temporal dynamics of semantic guidance in scenes

When viewing a complex scene, the visual system must decide where to look among many potential targets. How this selection process unfolds over time-and whether it reflects changing semantic priorities-remains poorly understood. We tracked eye movements of 100 participants viewing 100 real-world scenes and modeled four distinct semantic features at each location: similarity to the scene category, similarity to other objects, similarity to the most recently fixated object, and local semantic intensity. Complementary Bayesian models of fixation likelihood and fixation onset time revealed that each feature follows a distinct temporal trajectory. Scene-category similarity and semantic intensity strongly predicted early fixations but weakened over time, consistent with rapid gist-derived guidance. Preference for semantically distinctive objects remained stable throughout viewing. Similarity to the previously fixated object showed a temporal reversal: relative to each scanpath's own baseline, semantically related regions were under-selected early and over-selected later. These dynamics were robust to controls for central viewing bias and distance to the previous fixation, ruling out purely spatial or oculomotor explanations. Together, these patterns suggest a shift from information maximization early in viewing, when attention prioritizes novel, non-redundant semantic content, to conceptual integration later, when attention is increasingly directed to semantically related elements. Rather than discrete processing stages, multiple semantic processes appear to operate in parallel with shifting weights over time, revealing how stored knowledge is associated with attention in complex scenes. OPEN PRACTICES STATEMENT: The scene stimuli, fixation data, semantic feature maps, fitted model outputs, and analysis code for reproducing all results are available on the Open Science Framework: https://osf.io/7uspw . This study was not preregistered.