The scale of the investigation alone sets it apart. Hayes and Henderson tracked the eye movements of 100 participants as each viewed 100 real-world scenes, yielding a massive corpus of fixation behavior over natural images. But the real innovation lay in how the scenes were decomposed. At every location in every image, the researchers modeled four distinct semantic features. The first was similarity to the scene category itself, capturing how much a given patch resembled the gist of the whole image, such as a countertop resembling a typical kitchen. The second was similarity to other objects in the scene, measuring how semantically related a location was to the broader cast of objects present. The third was similarity to the most recently fixated object, a dynamic, scanpath-dependent quantity that changed with every eye movement. The fourth was local semantic intensity, the density of meaningful content in a region, in the tradition of the meaning maps that Henderson and Hayes have developed over the past decade.

To connect these features to behavior, the authors fitted complementary Bayesian models of fixation likelihood and fixation onset time, using the Bambi interface for Bayesian linear modeling in Python. This dual approach matters because a feature could attract fixations strongly but only late in viewing, or weakly but immediately, and the two patterns carry very different theoretical implications. By modeling both whether a location was fixated and when it was fixated, the researchers could trace the temporal trajectory of each semantic signal with a precision that earlier aggregate analyses could not achieve. The Bayesian framework also allowed the team to quantify uncertainty in the estimated effects rather than relying solely on point estimates, giving a more honest picture of how strong each temporal pattern really is.

The results split cleanly along the four feature dimensions. Scene-category similarity and local semantic intensity were powerful predictors of early fixations, but their influence weakened steadily as viewing progressed. This pattern is exactly what one would expect if viewers extract the gist of a scene within the first few hundred milliseconds, as work by Oliva, Potter, and others has long suggested, and then use that rapidly assembled representation to guide initial saccades toward the most category-relevant and semantically dense regions. Once those informative regions have been sampled, the marginal value of visiting yet another category-similar location declines, and the guidance signal correspondingly fades. In information-theoretic terms, the early viewing period behaves like an effort to maximize the acquisition of novel, non-redundant semantic content.

The second major finding concerned the preference for semantically distinctive objects, operationalized as low similarity to other objects in the scene. Unlike the gist-derived signals, this preference for semantic distinctiveness remained stable throughout the entire viewing period. An object that stood apart from its neighbors conceptually, a fire extinguisher in a living room, say, or a piece of farm equipment in a domestic scene, attracted attention early and continued to attract attention late. This constancy suggests that semantic distinctiveness is not a transient product of initial gist extraction but a durable property of the scene representation that keeps exerting its pull for as long as the image remains on screen. It also aligns with the broader literature on semantic inconsistency effects, going back to Biederman’s classic demonstrations that viewers detect and prioritize objects that violate scene expectations.

Perhaps the most striking result was a temporal reversal in the third feature. Similarity to the previously fixated object, relative to each observer’s own scanpath baseline, actually predicted under-selection of semantically related regions early in viewing and over-selection of those same regions later on. In other words, immediately after fixating an object, viewers tended to avoid looking at conceptually related locations, but as time went on they increasingly sought them out. The researchers interpret this reversal as evidence for a shift from information maximization to conceptual integration. Early on, visiting a location similar to what you just looked at would be redundant, so the visual system steers away to gather new content. Later, once broad coverage has been achieved, attention turns toward weaving the scene’s elements into a coherent semantic structure, and related locations become valuable precisely because they connect to what has already been processed.

Critical to the credibility of these conclusions is the set of control analyses the authors conducted. Two mundane alternatives could, in principle, explain the temporal reversal without invoking semantics at all. The first is the well-documented central fixation bias, the tendency of viewers to begin exploring images near the center, which Tatler and others have shown is partly independent of image content. The second is simple oculomotor geometry: regions close to the previous fixation are easier and faster to reach, so any apparent relationship between semantic relatedness and fixation could be a byproduct of distance. Hayes and Henderson controlled for both factors, and the temporal dynamics of all four semantic features survived intact. The patterns, they conclude, cannot be reduced to purely spatial or motor explanations; they reflect genuine semantic computation guiding the eyes.

The theoretical upshot is a departure from the traditional stage-based picture of scene perception, in which an early salience-driven phase hands off to a later semantics-driven phase, or vice versa. Instead, the findings suggest that multiple semantic processes operate in parallel from the very beginning, with their relative weights shifting continuously over the course of viewing. Gist-based guidance and semantic intensity dominate the opening moments, distinctiveness exerts a constant pull throughout, and relational similarity to the recent fixational history flips from repulsive to attractive. This parallel-and-shifting-weights architecture resonates with predictive processing frameworks, such as those articulated by Clark and Friston, in which the brain continuously generates and updates a generative model of the world, and with Henderson’s own proposal that gaze control is fundamentally an act of prediction. Each fixation, on this view, is a query posed to the scene, and the queries change character as the internal model matures.

The methodological machinery behind the study is as noteworthy as the findings. The semantic feature maps were derived using modern vector-space and vision-language tools, building on the DeepMeaning approach the authors introduced previously, which estimates scene meaning with a vision-language transformer, alongside knowledge resources such as WordNet and ConceptNet and embedding techniques descended from word2vec and GloVe. Combining computational semantics with eye-tracking at this scale, and making the full pipeline reproducible, represents a template for the next generation of attention research. Indeed, the authors have placed the scene stimuli, fixation data, semantic feature maps, fitted model outputs, and all analysis code on the Open Science Framework, allowing any laboratory to verify, extend, or challenge the results.

Why should a general reader care about the choreography of eye movements across pictures? Because the same machinery operates every waking moment. When you drive, cook, read a room at a party, or scan a hospital X-ray, your visual system is deciding where to look next based on what the scene means, not just how it looks. The new findings imply that these decisions are temporally smart: your brain first harvests the most informative, novel semantic content, then pivots to integrating what it has gathered into a coherent understanding. For applied domains, the implications are concrete. Radiologists searching for anomalies, security screeners inspecting luggage, and designers of driver-assistance systems all depend on models of visual attention, and those models have often treated semantic priority as static. This study shows that any faithful model must be dynamic, tracking how redundancy, novelty, and conceptual relatedness trade off over seconds of viewing. Funded by the National Eye Institute of the National Institutes of Health under award R01EY027792, and approved by the UC Davis Institutional Review Board with informed consent from all participants, the work opens a window on a process so fast and so automatic that we never notice it, yet so sophisticated that it quietly reorganizes our priorities hundreds of times per minute. The eyes, it turns out, are not just windows on the world; they are instruments of an ongoing, self-correcting conversation between perception and knowledge.

Subject of Research: Temporal dynamics of semantic guidance of attention and eye movements during real-world scene viewing

Article Title: The temporal dynamics of semantic guidance in scenes

Article References: The temporal dynamics of semantic guidance in scenes. (n.d.). https://doi.org/10.3758/s13414-026-03347-z

Image Credits: AI Generated

DOI: 10.3758/s13414-026-03347-z

Keywords: scene perception, eye movements, attention, semantic guidance, temporal dynamics, gist, meaning maps, visual attention, semantic memory, Bayesian modeling, fixation, psychophysics