Social Regularities Guide Visual Search Without Spatial Context

Human visual search can be effectively guided by abstract social regularities even when spatial configurations offer no predictive structure, according to a series of cognitive experiments published by researchers Lijeong Hong and Min-Shik Kim. Historically, research examining contextual cueing has concentrated primarily on spatial arrangements, demonstrating that observers locate visual targets faster when the surrounding spatial layouts of items are repeatedly encountered. The newly reported investigations addressed whether socially meaningful dimensions—such as identity pairings, collective emotional atmosphere, or relational facing directions—can independently establish contextual learning when spatial positions are completely randomized across every trial. Through three distinct visual search experiments using standardized facial photographs, the authors tracked reaction time changes across extended search sessions and administered subsequent objective awareness evaluations. The resulting data demonstrated reliable contextual cueing across all three social cue paradigms, revealing that the human cognitive system can spontaneously extract, encode, and exploit non-spatial social regularities to direct visual attention.

The study was designed to probe the boundaries of object-based contextual cueing, an area where previous empirical findings have often produced conflicting outcomes regarding whether non-spatial features reliably guide attention. In traditional spatial contextual cueing paradigms, participants rapidly acquire predictive spatial relationships, frequently showing performance facilitation within the earliest experimental blocks. In contrast, Hong and Kim asked whether the predictive scaffold that guides search could be established purely through social relationships when spatial consistency is systematically eliminated. To ensure that location-based learning was impossible, all three experiments introduced randomized spatial layouts and positional jitter on every search display. The experimental designs systematically compared search speed under Consistent Mapping conditions, where targets were paired with invariant social contexts, against Variable Mapping conditions, where the surrounding social information shifted continuously across blocks. Bayesian repeated-measures analyses evaluated how response latencies evolved across eight experimental epochs to measure the emergence, strength, and implicit nature of social learning.

Experiment 1 evaluated whether observers could learn identity-based associations between specific target faces and invariant sets of distractor identities. A total of 96 unique facial photographs from the Radboud Faces Database were organized such that 48 faces appeared exclusively in the Consistent Mapping condition, avoiding distractor overlap between repeated and non-repeated search displays. Data trimming removed response times under 200 milliseconds or beyond 2.5 standard deviations from individual participant means, which eliminated only 0.11 percent of trials while preserving an overall task accuracy of 98.64 percent. Across the overall eight-epoch analysis, the model containing only the main effect of epoch provided the most decisive fit against the null model, indicating widespread general task practice, whereas the mapping condition and the mapping-by-epoch interaction initially yielded anecdotal evidence favoring the null hypothesis.

To explore whether identity-based learning unfolded over extended exposure, the researchers divided the eight epochs into an early phase covering epochs one through four and a late phase covering epochs five through eight. In the early phase, reaction times were nearly identical between conditions, averaging 1.59 seconds in consistent mapping and 1.58 seconds in variable mapping displays. In the late phase, participants responded noticeably faster in the consistent mapping condition, averaging 1.38 seconds compared to 1.45 seconds in the variable mapping condition. Post-experiment assessments supported the implicit nature of this learning. In a binary recognition query, zero participants reported noticing consistent target-distractor identity pairings, while a subsequent forced-choice generation test produced a mean identification accuracy of 11.88 percent, failing to exceed the 12.5 percent chance threshold. Excluding the five participants who scored above chance left the late-phase interaction evidence intact.

Experiment 2 examined whether an ensemble-level social cue—the collective emotional mood of a display—could support contextual cueing. Sixteen participants, including 3 men and 13 women with a mean age of 23.2 years, were recruited under a Bayesian optional stopping rule. Displays consisted of 15 faces, comprising 1 female target and 14 male distractors displaying happy or angry expressions drawn from 28 identities in the Radboud Faces Database. The display mood was defined by the ratio of angry to happy faces, spanning 14 ratios from 1:14 to 14:1. In consistent mapping trials, a target was paired with an invariant emotional ratio across all epochs, whereas in variable mapping trials, target-ratio pairings varied across blocks. Crucially, emotional expressions were randomly assigned to individual facial identities at the start of each block, preventing participants from relying on individual identity-expression pairings or specific face configurations while ensuring that target frequencies were matched across conditions.

The search performance in Experiment 2 demonstrated robust contextual cueing that emerged far more rapidly than in the first experiment. The Bayesian repeated-measures analysis across all eight epochs strongly favored the full model including both main effects and their interaction over the null model. When collapsed into early and late phases, reaction times in the early phase were 2.06 seconds for consistent mapping and 2.21 seconds for variable mapping. In the late phase, reaction times dropped to 1.33 seconds in the consistent condition versus 1.87 seconds in the variable condition. Analysis across all 14 ratios verified that facilitation was not driven by extreme emotional ratios being inherently easier to process. In the post-search recognition assessment, nine participants reported noticing regularities, yet average objective accuracy on the forced-choice generation test remained at 4.46 percent, well below the 7.14 percent chance mark. Re-analyzing the visual search data after excluding the five participants who exceeded chance accuracy confirmed that the interaction and mapping main effect remained statistically decisive.

Experiment 3 investigated whether contextual cueing could be driven by relational social orientation—specifically, the facing directions within pairs of profile-view faces. Twenty participants, comprising 5 men and 15 women with an average age of 22.8 years, searched for a neutral female target among neutral male distractors. Stimuli were organized into seven distinct pairs distributed across a five-by-eight grid containing 40 possible cell locations. Each pair occupied adjacent positions separated by approximately 117 pixels, while cell centers were separated by 232 pixels horizontally and 195 pixels vertically to prevent spatial crowding. In consistent mapping conditions, the pattern of whether specific pairs faced toward or away from each other remained constant across all blocks for a given target. In variable mapping conditions, the facing directions of three to five pairs were flipped between successive displays while preserving an overall ratio of three facing to four non-facing pairs or four facing to three non-facing pairs.

The results of Experiment 3 confirmed that relational social structure facilitated search speed despite total spatial randomization. Across all eight epochs, the Bayesian repeated-measures analysis provided strong evidence for both the main effect of mapping condition and its interaction with epoch. Phase-based comparisons demonstrated comparable early-phase performance, with mean reaction times of 1.36 seconds in consistent mapping and 1.39 seconds in variable mapping, followed by marked divergence in the late phase, where consistent mapping trials averaged 1.19 seconds compared to 1.33 seconds in variable mapping trials. In subsequent awareness testing, nine participants reported noticing regularity, and ten scored above the 7.14 percent chance level on the forced-choice generation task, averaging 17.14 percent. When these ten participants were excluded, the interaction between phase and mapping condition remained supported by the Bayesian models, although the overall magnitude of the contextual cueing effect was somewhat diminished, indicating that conscious awareness could not be entirely ruled out as an auxiliary contributor.

Comparing the three experiments revealed that ensemble emotion in Experiment 2 produced the most pronounced contextual cueing effect. The authors noted that global emotional ratios function as low-variance statistical summaries that observers extract automatically and in parallel across visual scenes, eliminating the need to bind individual identities item by item. Furthermore, potential visual artifacts were rigorously evaluated to confirm that performance was driven by social information rather than low-level image features. In Experiment 2, a pixel-level brightness analysis of teeth visibility demonstrated that dental exposure did not differ across mapping conditions and did not modulate search speed. The researchers also contrasted their findings with previous failures to replicate object-based cueing, such as work by Gaál and colleagues, pointing out that separating distractor pools between mapping conditions prevented distractor overlap from degrading the diagnostic utility of the social cues.

The investigators emphasized that while their findings indicate that social structures can serve as an effective predictive scaffold, the underlying mechanisms likely reflect multi-stage cognitive processes. Although the behavioral latency measures cannot isolate whether facilitation occurs during early attentional selection or later response decision stages, related event-related potential literature tracking N2pc and P3b components suggests that contextual cueing influences both perceptual and post-perceptual operations. In addition, the study did not test how individual personality traits such as empathy, loneliness, or social anxiety might alter the implicit extraction of social contexts. Future research utilizing naturalistic video displays and immersive environments will be needed to explore how these non-spatial social predictive mechanisms operate in daily interactions.

Subject of Research: Implicit learning of social contextual cues in visual search

Article Title: Implicit learning of social information in contextual cueing

Article References: Hong, L., & Kim, M.-S. (2026). Implicit learning of social information in contextual cueing. Cognitive Research: Principles and Implications, 11(1), Article 48. https://doi.org/10.1186/s41235-026-00736-8

Image Credits: AI Generated

DOI: 10.1186/s41235-026-00736-8

Keywords: contextual cueing, visual search, implicit learning, social cognition, ensemble perception, facial processing, Bayesian analysis, cognitive psychology, Implicit, learning, social, information