Who wins the first glance: your goal or the loudest object?
In an eye-tracking study that varied display density, whether the target was pre-cued, and the recent history of trials, early selection was dominated by top-down knowledge and trial history. When attention was biased toward the target feature—either by a valid cue or by automatic priming from previous trials—the target was localized immediately, even in dense displays [2]. Bottom-up feature contrast only modulated selection when the target was unknown and attention was biased toward the non-targets [2].
A separate study in a pseudo-realistic scene with a highly salient singleton found something similar but with a twist. In a free-view baseline, people's first fixation stayed locked onto the high-salience item across a range of saccade latencies. Yet once given a search instruction to find a red target, that salient item captured very few saccades, and it took time for saccades to be correctly aimed at the target [1]. The authors suggest attention may be "in limbo" early in a search trial—caught between the pull of salience and the push of the goal [1].
Salience may help you reject distractors rather than grab your eyes
The classic idea is that a high-contrast pop-out target is grabbed immediately in dense displays but not in sparse ones. The eye-tracking work challenges this: reliable feature-contrast effects showed up in average response times, but those effects traced to later, identification-related processes such as how long the eyes dwelled on the target—not to the initial selection [2]. The authors propose that bottom-up contrast in dense displays doesn't directly steer attention but instead helps you reject non-targets, possibly by making them easier to group together [2].
This reframes salience as something that can smooth the search rather than hijack it. It also fits the finding that salient singletons captured few saccades once a search goal was in place [1].
When goals can't override the pull of salience
Top-down control is not unlimited. In a task where items came in two colors and the target was more likely to appear in the minority-colored group, people's attention was naturally captured by that minority group. They did use the environmental regularities to adapt their strategy and search more efficiently, but they could only be biased away from the salient group when the task encouraged serial search [3]. That asymmetry suggests top-down expectations modulate bottom-up tendencies unevenly depending on what kind of search the task demands [3].
Top-down control also appears to work jointly with learned regularities rather than simply overriding them. In an EEG study combining valid versus neutral cues with locations where targets appeared more or less often, both cueing and location frequency improved performance, but the two interacted: the statistical-learning benefit emerged only when top-down guidance was absent [5]. The neural marker of attentional selection (the N2pc) showed a cueing effect for targets in low-frequency locations but not high-frequency ones, which the authors read as top-down control and statistical learning working in a conjoint, integrated manner—with learning reduced or absent when a fully reliable top-down signal is at play [5].
How the two forces combine over the course of a search
A neural dynamic model of triple-conjunction search—where targets are defined by three features at once—attempts to explain how these interactions arise without positing a separate pre-attentive binding step. It shows that search is not guided only top-down; its efficiency is also affected by bottom-up salience, and the model produces these complex interactions as an emergent property of the underlying neural dynamics [4]. That is a modeling proposal rather than a direct measurement, but it offers a concrete picture of how goal signals and salience signals could combine continuously rather than one simply beating the other.
Taken together, the evidence points to a division of labor that shifts with time and certainty. Early in a trial, before the goal is fully engaged, salience can hold attention [1]. Once a reliable goal or recent priming is in place, it can direct the first saccade regardless of clutter [2]. Salience then tends to matter more for later stages, such as rejecting distractors [2], and top-down control can override a natural pull toward salient items only under some task conditions [3].
Sources used in this answer
Evaluating the contributions of top-down and bottom-up processing on eye movements during parallel visual search
In a pseudo-realistic search task, a high-salience singleton held the first fixation in free viewing, but captured few saccades once participants searched for a red target, and correct targeting of the target took time to develop.
Which processes dominate visual search: Bottom-up feature contrast, top-down tuning or trial history?
Early visual selection was dominated by top-down knowledge and trial history, with bottom-up feature contrast modulating selection only when the target was unknown and attention was biased to non-targets; contrast effects in mean response times reflected later target-identification processes.
Top-down Control and Bottom-up Salience Guide Attention During Visual Search
Attention was naturally captured by the minority-color group, and people adapted their strategy using environmental regularities but could be biased away from salient items only when the task promoted serial search.
A neural dynamic process model of combined bottom-up and top-down guidance in triple conjunction visual search
A neural dynamic model accounts for triple-conjunction search efficiency through emergent interactions of top-down guidance and bottom-up salience without requiring preattentive feature binding.
Integrated effects of top-down attention and statistical learning during visual search: An EEG study
Top-down cueing and statistical learning of target locations interacted, with statistical-learning effects emerging only in the absence of top-down guidance, consistent with the two mechanisms working in a conjoint, integrated manner.