In free visual search, individuals move their eyes around a scene and search for targets with a "functional visual field" (FVF) surrounding the current point of fixation. Items could be sampled in series, one after the other within the FVF or the entire FVF might be processed in parallel. In either case, processing could be uniform, on average, across the FVF or, alternatively, some items or regions in the FVF might be systematically processed more successfully than others. To assess this, we had observers fixate at a point and then flashed a ring of eight items around fixation. We examined the distribution of errors as a function of the angular position around fixation. We found, as might be expected, that, on average, fewer errors were made on the horizontal than on the vertical meridian. Interestingly, we found reliable, idiosyncratic deviations from this average pattern. Observers tended to have specific angular positions that were consistently better or worse than average. These biases in search performance might be related to the production of otherwise apparently random false negative errors in visual search. (PsycInfo Database Record (c) 2026 APA, all rights reserved).

Purpose: To determine the effects of automation bias when radiologist read digital breast tomosynthesis (DBT) screening cases concurrently with an artificial intelligence (AI) cancer detection system. Method We retrospectively analyzed an observer study of radiologists screening reading DBT images with and without an AI system. In that study, there were 260 DBT screening exams (65 containing at least one malignant lesion). Twentyfour radiologists read the cases in two separate sessions (with a 4-week washout period) once without the AI tool and once with AI concurrently (i.e., the AI marks and scores were available immediately upon examining the images). We examined cases for which one might expect automation bias (disagreement between radiologist reading without AI and the AI assessment) and we measure the amount of agreement between the AI score and the radiologist reading with AI. As a control, we performed the same analysis, but for cases of agreement between the radiologist reading without AI and the AI. We analyzed the data separately for each of four different categories of cases included in the study: cancer cases, biopsied benign cases, recalled but not biopsied cases, and negative cases. These four categories represent different levels of suspicion of cancer being present (as listed here, highest to lowest level of suspicion). Results: The amount of bias depended on the likelihood that cancer was present in the case. The net result is that, as the likelihood of cancer being present decreased, the performance of the radiologists tended towards the ideal clinical result. That is, there is less automation bias for non-cancer cases. Conclusion: When reading with AI concurrently, radiologists show signs of automation bias. However, the amount of automation bias is related to how likely that cancer is present in the case. The net benefit improves the radiologists' performance.

Using a threshold stimulus exposure duration (TSED) method, rather than the usual response time methods, a recent study showed evidence for a two-stage visual selection process in visual search. The resulting TSED function was bilinear. In the present study, we investigated whether this bilinear pattern reflects general property of TSED functions or if it was just an artifact of specific methodological choices. Experiment 1 refined the staircase procedure of the TSED method to separate the effect of target-present and target-absent trials. Experiments 2 and 3 replaced the target-present/target-absent task with two-alternative forced-choice tasks. Results repeatedly showed a bilinear pattern in the TSED function across paradigms, with a distinct kink point separating a shallower slope at lower set sizes from a steeper slope at larger set sizes. The findings provide convincing evidence that bilinear pattern in TSED functions is generalizable property rather than a task-specific artifact. A modified Guided Search "carwash" can explain how visual search can produce bilinear TSED functions and linear RT × Set Size functions.

Attentional priority is typically conceived as a static spatial map, despite attention operating in a continuously changing world. We propose a dynamic priority map and outline the core demands and key questions needed to understand attentional guidance in a world that never stands still.

Imagine searching a collection of coins for quarters (0.25), dimes (0.10), nickels (0.05), and pennies (0.01)—a hybrid foraging task where observers look for multiple instances of multiple target types. In such tasks, how do target values and their prevalence influence foraging and eye movement behaviors (e.g., should you prioritize rare quarters or common nickels)? To explore this, we conducted human psychophysics experiments, revealing that humans are proficient reward foragers. Their eye fixations are drawn to regions with higher average rewards, fixation durations are longer on more valuable targets, and their cumulative rewards exceed chance, approaching the upper bound of optimal foragers. To probe these decision-making processes of humans, we developed a transformer-based Visual Forager (VF) model trained via reinforcement learning. Our VF model takes a series of targets, their corresponding values, and the search image as inputs, processes the images using foveated vision, and produces a sequence of eye movements along with decisions on whether to collect each fixated item. Our model outperforms all baselines, achieves cumulative rewards comparable to those of humans, and approximates human foraging behavior in eye movements and foraging biases within time-limited environments. Furthermore, stress tests on out-of-distribution tasks with novel targets, unseen values, and varying set sizes demonstrate the VF model’s effective generalization. Our work offers valuable insights into the relationship between eye movements and decision-making, with our model serving as a powerful tool for further exploration of this connection. All data, code, and models are available at https://github.com/ZhangLab-DeepNeuroCogLab/visual-forager.

Visual attention paradigms have revealed that neural excitability in higher-order visual areas is modulated according to a priority map guiding attention towards task-relevant locations. Neural activity in early visual regions, however, has been argued to be modulated based on bottom-up salience. Here, we combined Magnetoencephalography (MEG) and Rapid Invisible Frequency Tagging (RIFT) in a classic visual search paradigm to study feature-guidance in early human visual cortex. Our results demonstrate evidence for both target boosting and distractor suppression when the participants were informed about the task-relevant and -irrelevant colour (guided search) compared to when they were not (unguided search). These results conceptually replicated using both a magnitude-squared coherence approach and a General Linear Model based on a single-trial measure of the RIFT response. The present findings reveal that feature-guidance in visual search affects neuronal excitability as early as primary visual cortex, possibly contributing to a priority-map-based mechanism.

In everyday life, we frequently engage in 'hybrid' visual and memory search, where we look for multiple items stored in memory (e.g., a mental shopping list) in our visual environment. Across three experiments, we used event-related potentials to better understand the contributions of visual working memory (VWM) and long-term memory (LTM) during the memory search component of hybrid search. Experiments 1 and 2 demonstrated that the FN400 (an index of LTM recognition) and the CDA (an index of VWM load) increased with memory set size (target load), suggesting that both VWM and LTM are involved in memory search, even when target load exceeds capacity limitations of VWM. In Experiment 3, we used these electrophysiological indices to test how categorical similarity of targets and distractors affects memory search. The CDA and FN400 were modulated by memory set size only if items resembled targets. This suggests that dissimilar distractor items can be rejected before eliciting a memory search. Together, our findings demonstrate the interplay of VWM and LTM processes during memory search for multiple targets.

Visual search models have long emphasised that task-relevant items must be prioritized for optimal performance. While it is known that search efficiency also benefits from active distractor inhibition, the underlying neuronal mechanisms are debated. Neuronal alpha oscillations (7-14 Hz) have been associated with functional inhibition of cortical excitability, as well as distractor suppression in spatial attention and visual working memory tasks. We therefore hypothesised that alpha oscillations similarly support the deselection of distractors in visual search. Using Magnetoencephalography (MEG), we here show that high alpha power before the onset of a complex search display is associated with faster search performance. Crucially, we used a General Linear Model (GLM) approach to control for confounds between alpha power and task duration, ruling out that this result was merely driven by practice effects paired with increased fatigue over time. In addition to spontaneous oscillatory activity, we quantified the cortical excitability to colours of the search stimuli based on Rapid Invisible Frequency Tagging (RIFT) responses. In contrast to our initial hypothesis, increased pre-search alpha power did not correlate with the RIFT response, providing no direct evidence for feature-specific inhibition of distracting stimuli by alpha. Our findings challenge the traditional view of alpha oscillations reducing visual processing, showing instead that increased occipital alpha power can enhance performance in a visual task. We propose that the increase in alpha power may reflect increased top-down control supporting visual search.

Low target prevalence affects perceptual decisions on both simple and complex stimuli. Without prior knowledge of how often targets may appear, trial-by-trial accuracy feedback modulates the effects of low prevalence partially by providing observers with information about the target base rate. Using simple colored dots, Lyu (PBR 28:1906–1914, 2021) found that at low prevalence, observers demonstrate a classical low prevalence effect (LPE) when receiving feedback. This involves a conservative shift of the decision criterion where observers are less likely to call an ambiguous item a target. In the absence of feedback, observers adopted more liberal criteria and became more likely to classify an item as a target, producing a Prevalence-Induced Concept Change (PICC, Levari et al., Science 360:1465–1467, 2018). The present study examines whether the effects of low prevalence and feedback are modulated by expertise. Novice (n = 26) and expert (n = 24) observers performed a cancer cell discrimination task. The prevalence of cancerous “blast cells” and the presence or absence of trial-by-trial accuracy feedback were manipulated. Unsurprisingly, medical professionals performed better than trained novices. Importantly, both experts and novices showed an LPE with feedback, although that LPE was weaker in experts, suggesting expertise may modulate the size of the LPE. Low prevalence had little effect on the criterion in the absence of feedback in this setting. For both novices and experts, initial exposure to trials with feedback influenced criteria in subsequent no feedback conditions. Interestingly, experts showed a conservative criterion at the start of the experiment, even without having experienced a feedback block. This could reflect previous training or working in a low prevalence setting. Our study shows the interactions of the effects of low prevalence, feedback, and expertise on perceptual decisions and provides direct evidence for prevalence and feedback effects on expert decisions.

Purpose: To determine if reading digital breast tomosynthesis (DBT) concurrently with an artificial intelligence (AI) system increases the probability of missing a cancer not marked by AI for cancers that the radiologist detected reading without AI. Method We retrospectively analyzed an observer study of radiologists reading with and without an AI system. In that study, there were 260 DBT screening exams (65 containing at least one malignant lesion). Twenty-four radiologists read the cases in two separate sessions (with a 4-week washout period) once without the AI tool and once with AI concurrently (i.e., the AI marks and scores were available immediately upon examining the images). We separated the cases into AI-detected and AI-notDetected and then examined only cases that the radiologist recalled when reading without AI. We determined the fraction of cases from each group that the radiologist recalled when reading with AI; this was done separately for cancer and non-cancer cases. Results: When reading without AI, the readers detected an average of 5.0 of 7 (71%) cancers that were not marked by AI (range 1-7) and 49.8 of 58 (86%) cancers that were marked by AI (range 30-57). When reading with AI concurrently, readers found 3.3 (46%) of the 7 AI-notDetected cancers and agreed with 54.2 (93%) of the 58 AI-detected cancers. Using a two-tailed, paired t-test, this difference (46% vs 93%) was statistically significant (p << 0.00001). Nevertheless, the overall sensitivity increased with concurrent reading compared to reading without AI (77% to 85%). Similarly, for non-cancer cases that were recalled (FP) without AI (47%/26% not-marked/marked by AI), there was a smaller fraction recalled for the not-marked cases (8.1% vs 48%, p << 0.00001). This contributed to an increase in specificity with concurrent reading (63% to 70%). Conclusion: When reading with AI concurrently, radiologists are more likely to miss a cancer when AI fails to mark that cancer. Likewise, radiologists are more likely not to recall a non-cancer case when AI fails to mark a lesion in the case, even though the radiologist recalled the case when reading without AI.

Hybrid visual search tasks involve searching for multiple targets held in memory, but some targets are more memorable than others. Furthermore, some items are readily identified as being in the memory set, while others are readily identified as not being in the memory set; these may be considered to vary in their "hittability" and "rejectability", respectively. In principle, both factors should impact error rates and reaction times in hybrid search. Using a set of 9 million trials from an online hybrid search game, we analyze participants' errors and show that hittability and rejectability are largely separable. It is possible for items to be rejectable without being particularly hittable, and to be hittable without being particularly rejectable. Both factors are consistent across participants and stable across age, training, and performance. Rejectability strongly predicted reaction times in the search for new items, while hittability was more weakly associated with reaction times.

We have decades of visual search data from experiments where observers look for targets among distractors. Typically, observers are tested in blocks of several hundred trials, and conclusions about underlying mechanisms are inferred from Reaction Time × Set Size functions and errors. However, in the real world, searchers almost never search for the same target or the same type of target hundreds of times in a row. You search for cereal, then milk, then a bowl. Do the rules derived from blocks of trials apply when search tasks are mixed? Here, we compare mixed and blocked conditions in five experiments. In Experiment 1, four different feature searches are tested. In Experiments 2 and 3, the target was the same in four tasks that were defined by different distractor sets. In Experiment 4, different targets are searched for amongst distractors that remained constant across trials. Finally, in Experiment 5, we allowed participants to choose which of four tasks to perform on each trial. In each experiment, there was no qualitative change in search behavior as a function of the mixed/blocked manipulation. The results support the generality of rules of search learned from blocked trials. However, these results do pose a challenge to simple adaptive models of search termination.

This study investigated patch-leaving strategies in mixed hybrid visual foraging scenarios, focusing on how target specificity and the number of target sets influence overall outcomes. In mixed hybrid foraging, participants collect targets from patches with varying types (specific and categorical) and numbers (three or six) of targets. Despite the complexity introduced by having multiple target types, participants' patch-leaving behavior remained broadly consistent with the predictions of the marginal value theorem (MVT), suggesting that quitting strategies are based on similar rules across different conditions. While overall foraging performance varied with target specificity and the number of sets, patch-leaving decisions consistently adhered to a simple, rule-based approach. This study highlights the robustness of visual foraging strategies and suggests that effective patch-leaving behavior is maintained even in complex visual environments.