Abstract:

Selective attention is a critical cognitive function that enables the brain to prioritize relevant stimuli while filtering out irrelevant information. This ability is essential for navigating complex environments, where multiple stimuli compete in parallel for limited processing resources. Traditional theories of attention focus primarily on how specific feature dimensions (e.g., color, shape) or their absolute values (e.g., red, circle) guide attentional selection. However, emerging evidence points to an alternative mechanism: attentional guidance based on feature relationships (e.g., “redder” than its surroundings). In this review, we systematically synthesize research on attention based on feature relationships, comparing it with the conventional attention models based on feature values. We characterized how this relational-based attention operates across different forms of attentional control, highlighting its distinct processing characteristics, theoretical implications, and potential applications. This review makes two key contributions to understanding attentional guidance by feature relationships.

First, we systematically summarize how feature relationships guide attention across both goal-driven (e.g., contingent attentional capture) and experience-driven (e.g., priming, selection history) attention. According to contingent-capture theory, attention is automatically captured by irrelevant singletons sharing the target-defining feature value (e.g., a red item). However, recent studies show that attention could be preferentially guided by relative feature differences within the stimulus context (e.g., a “redder” item) without exact feature-value matches. Critically, this relational mechanism operates independently of singleton-induced salience effects and extends to conjunction searches where the targets lack physical salience. In such cases, attention can be effectively guided by combinations of feature relationships across dimensions (e.g., “bluer and larger”), regardless of feature-value matches. In addition, we present evidence distinguishing this relational account from an alternative explanation - optimal tuning account. These findings suggest the existence of a flexible, top-down control mechanism that prioritizes relational feature properties.

Similarly, we summarize that experience-driven attention biases - whether from priming (repeated exposure) or learned regularities (selection history or reward history) - could reflect relational coding rather than the presumed feature-specific processing. In particular, inter-trial priming effects depend primarily on whether the target-nontarget relationship remains consistent or change across trials, rather than on simple repetition of specific feature values. Both statistical learning (selection history) or reward-based learning (reward history) produce attentional biases that generalize to novel stimuli sharing the learned feature relationships, Crucially, this generalization occurs specifically when the stimuli maintain the learned relational information, even when their feature values differ from those encountered during learning. Given that these experience-related attentional biases persist for previously processed but currently task-irrelevant features, this line of research suggests a bottom-up mechanism that prioritizes relational feature properties during attentional selection.

The second contribution is that we demonstrate that relational guidance represents distinct spatiotemporal characteristics from traditional feature-value based account. While these two mechanisms are not mutually exclusive — individuals can flexibly deploy feature-values and relational templates across different feature dimensions or processing stages depending on task demands — they operate differently in time and space. Temporally, initial selection relies primarily on feature relationships, while later target identification and verification depend more on exact feature matching. This division of labor aligns with recent theories of attention, which emphasize the “good-enough” principle of attentional guidance in early selection. Spatially, these two mechanisms show differences in the global effect across the visual field: feature-specific attention produces robust global effects, whereas feature relationship-guided attention shows limited spread (potentially constrained by attentional window size) unless the relational information itself becomes task-relevant.

The review identifies several key avenues for future research and translation: (1) mapping the neural basis of attention based on feature relationships; (2) exploring whether these mechanisms generalize to complex relations (e.g., social relationships) and other cognitive domains like working memory; and (3) translating these findings into clinical tools for relational processing deficits, attentional training protocols, and technological applications including artificial intelligence algorithms and human-computer interface design.

Key words: selective attention, feature relationship, contingent attentional capture, experience-driven attention