Visual working memory is the mental scratchpad that keeps a handful of visual items active after they disappear from view. Its capacity is famously small, often estimated at around three or four items, and researchers have debated for decades whether storing the conjunction of features—say, that a particular shape was a particular color—costs more than storing the features themselves. A growing body of evidence suggested that it does, and that the extra cost is paid in a currency called object-based attention. When people must divide that resource across multiple objects during memory maintenance, conjunction memory suffers while memory for individual colors and shapes remains largely spared. The interpretation seemed straightforward: bindings are fragile, and keeping them glued together requires ongoing attention directed at objects.

But there was a catch, and it is the kind of catch that can quietly undermine an entire literature. Nearly all previous demonstrations used long encoding windows of roughly a thousand milliseconds, during which participants could study the memory array at leisure. Within such a generous window, two very different processes could unfold. The first is rapid and automatic: features are bound into objects almost immediately, a process long associated with reentrant processing in visual cortex. The second is slower and strategic: a controlled elaboration that stabilizes, rehearses, and consolidates the initial representations. If object-based attention only mattered for the second process, then the classic results might reveal nothing about bindings as such—only about the products of deliberate mental housekeeping performed during a full second of viewing time.

The Zhejiang University team designed their study to sever these two possibilities. Across three experiments, participants performed a change-detection task on color-shape conjunctions, or alternatively on the constituent colors and shapes presented separately, while a concurrent secondary task manipulated the demand on object-based attention. The secondary task was a variant of the feature-report paradigm introduced by John Duncan in 1984, a foundational demonstration that attention operates on objects rather than purely on spatial locations. In the high-demand condition, participants reported two features belonging to two different objects, forcing object-based attention to split and thereby consuming the resource. In the low-demand condition, they reported two features of the same object, leaving the resource largely available for the memory task.

The critical manipulation was temporal. In Experiment 1, the memory array was shown for only 200 milliseconds before the secondary-task stimulus appeared—far too brief for the slower strategic-control process to run its course, though rapid automatic binding could still occur. In Experiment 2, the window shrank to 100 milliseconds, tightening the constraint further. In Experiment 3, the researchers went all the way: the memory array and the secondary-task stimulus appeared simultaneously, leaving no post-encoding window at all in which object-based attention could act on the memory representations unopposed. If the selective impairment of binding memory vanished as the window closed, that would indicate the effect depended on strategic stabilization. If it persisted, the dependence would have to be intrinsic to the binding representations themselves.

The results were strikingly clean. In all three experiments, consuming object-based attention selectively impaired binding memory relative to feature memory, and the magnitude of that selective cost was statistically indistinguishable across experiments. Whether participants had 200 milliseconds, 100 milliseconds, or effectively zero milliseconds of unopposed processing time, draining the object-based attentional resource hurt their ability to remember which color went with which shape, while leaving their memory for the colors and shapes in isolation comparatively untouched. Bayesian and frequentist analyses alike supported the conclusion that the three experiments measured the same underlying effect, not a shrinking one.

This pattern carries a theoretically loaded message. The dependence of working memory bindings on object-based attention is not a downstream consequence of strategic stabilization during extended encoding; it is a property of binding representations themselves. In other words, even the fastest, most automatic bindings in visual working memory are not self-sustaining. They require a continuous supply of object-based attention to remain coherent during retention, and that requirement is baked into what a binding is, not bolted on by deliberate effort. The finding aligns with the hierarchical binding model advanced by Zaifeng Gao and Mowei Shen, in which object-based attention serves as the glue that maintains integrated object representations over time.

The study also helps reconcile a stubbornly contradictory literature. Some earlier experiments, including work by Allen, Baddeley, and Hitch, found that binding in working memory was relatively undemanding of attention, while others found clear attentional costs. Subsequent resolutions proposed that different forms of attention and different buffer stores were at play. The new results add a temporal dimension to that resolution: studies using long encoding windows may have inadvertently given participants room for strategic control processes that either supplemented or masked the attentional maintenance of bindings. By compressing the timeline, Zhang and colleagues show that the core attentional dependence survives even when strategy is squeezed out, suggesting that genuine disagreements in the literature stem from what happens after encoding rather than from the nature of binding itself.

There are broader implications for how scientists model the architecture of working memory. Computational accounts, such as the neural architecture proposed by Schneegans and Bays, treat binding as arising from conjunctive population codes that can, in principle, persist without extra maintenance. The new evidence pushes back against purely passive conceptions: the conjunction information in working memory behaves as though it is actively sustained by an object-centered attentional resource, one that can be taxed by any concurrent task requiring features to be gathered from multiple objects. That resource is distinct from the general attention needed to encode and retrieve visual memories, which affects features and bindings alike. The selective pattern—bindings hit, features spared—is the signature that makes object-based attention identifiable as the specific mechanism at work.

For everyday cognition, the message is humbling. The vivid, unified objects you hold in mind after a glance are not snapshots; they are ongoing constructions that the brain must keep paying for. Divert object-based attention elsewhere—by asking it to juggle features from several different objects at once—and the glue begins to fail, even if only a tenth of a second has passed since the objects vanished. The study was supported by the National Natural Science Foundation of China and the Zhejiang Key Laboratory of Neurocognitive Development and Mental Health, and the authors have made their data, analysis files, experimental programs, and stimuli publicly available on the Open Science Framework, an openness that should speed up the next round of experiments probing exactly how, and for how long, attention keeps the pieces of our visual world stitched together.

Subject of Research: The role of object-based attention in maintaining feature bindings in visual working memory

Article Title: Object-based attention sustains feature binding in working memory without requiring strategic control

Article References: Zhang, X., Wu, F., Zhou, Q., Shen, M., & Gao, Z. (2026). Object-based attention sustains feature binding in working memory without requiring strategic control. Attention, Perception, & Psychophysics, 88(7), Article 195. https://doi.org/10.3758/s13414-026-03341-5

Image Credits: AI Generated

DOI: 10.3758/s13414-026-03341-5

Keywords: visual working memory, feature binding, object-based attention, attention, change detection, Duncan task, cognitive psychology, encoding, strategic control, Zhejiang University, memory maintenance, binding problem

The theoretical backdrop for the work is the increasingly popular Bayesian account of serial dependence. According to this framework, the brain behaves like an intuitive statistician, combining noisy sensory evidence with expectations derived from recent experience to produce decisions that are optimal under uncertainty. When the current stimulus is encoded poorly, its internal representation is noisy, and the perceptual system should lean more heavily on the previous stimulus, producing stronger attractive biases. Conversely, when the previous stimulus, often called the inducer, is itself noisy, its influence on the current decision should weaken because a unreliable source of prior information deserves less weight. This elegant logic has generated a large body of findings, mostly documenting attractive biases in tasks such as orientation and face estimation. Yet the framework makes a prediction that has been far less explored: in some tasks, the strategy that actually minimizes error is not attraction but repulsion, and uncertainty should then modulate repulsive biases in predictable ways.

Yildirim and Boduroglu designed an experiment precisely to test that prediction. Participants viewed three tilted gratings in rapid succession, each flashed for roughly 83 milliseconds, and later reported the orientation of each one from memory. The critical manipulation lay in the timing. The first two targets, T1 and T2, were separated by either three or seven intervening stimuli, while the third target, T3, always followed T2 at a lag of seven. At the shorter lag, T2 fell inside the temporal window of the well-known attentional blink, the brief period after detecting one target during which attentional resources for the next are depleted. As a result, participants’ representations of T2 were measurably worse at Lag 3 than at Lag 7, as indexed by larger absolute estimation errors and lower precision. This created a natural experiment in representational noise: the same stimulus, T2, sometimes served as a well-encoded target and sometimes as a degraded one, and in both cases it also acted as the inducer for the subsequent report of T3.

The results overturned a straightforward reading of the Bayesian account. When T2’s representation was noisier, produced by the attentional blink at the short lag, the repulsive serial bias away from T1 actually grew stronger, not weaker. Reduced attentional resources amplified the tendency for participants’ estimates of T2 to be pushed away from the orientation they had just seen on T1. This is the first demonstration, the authors report, that diminished attention increases repulsive serial bias in working memory. In other words, when the current representation is unreliable, the system does not simply lean harder on the past; instead, in this task regime, it actively differentiates the present from the past, as if exaggerating the contrast between the two to keep them from being confused.

Equally striking was what did not change. The noisiness of the inducer, that is, the fidelity of T2, had no measurable effect on the bias observed in estimates of T3. Whether T2 had been encoded under the attentional blink or with full attentional resources, the repulsive pull of T2 on T3 estimates remained statistically indistinguishable. Under a strict Bayesian reading, a degraded inducer should contribute a weaker prior and therefore a smaller bias. The absence of such modulation suggests that the brain may treat the stored memory trace of a stimulus as a fixed reference point for differentiation, regardless of how faithfully that stimulus was originally captured, or that the adaptive mechanism at work here is not the classic Bayesian weighting scheme at all.

Perhaps the most intriguing findings concern the temporal reach of the repulsive effect. Serial dependence is usually studied between adjacent items, and many researchers have assumed that its influence decays rapidly with intervening stimuli. But in this experiment, estimates of T3 were repelled not only by T2 but also by T1, an item separated from T3 by a lag of ten stimuli in the rapid stream. The bias therefore extended well beyond the immediate perceptual history, implying that multiple past representations can coexist in working memory and each exert a distinct repulsive influence on a current report. This long-range repulsion is difficult to reconcile with simple low-level adaptation accounts, in which sensory neurons tuned to the previous orientation become fatigued and shift subsequent percepts away, because such aftereffects are typically short-lived and confined to the immediately preceding stimulus.

The study also uncovered evidence of repulsion operating in the opposite direction of time. Estimates of T2 were pushed away from the orientation of T3, the stimulus that followed it. Since T3 had not yet been presented when T2 appeared, this backward bias could not reflect perception at encoding. Instead, the authors argue, it emerged during the maintenance period, while T2’s representation was being held in working memory and before T3’s report was required. This implies that repulsive differentiation is an active, ongoing computation performed on stored representations rather than a passive residue of earlier sensory processing. When a new item enters working memory, existing representations appear to be reorganized to maximize their separation from the newcomer, a form of online uncertainty minimization that unfolds during the retention interval itself.

Taken together, these findings complicate the popular narrative that serial dependence is a single, attractively inclined mechanism serving perceptual stability. The authors conclude that the pattern they observed reflects adaptive uncertainty minimization in working memory that goes beyond standard Bayesian accounts of serial dependence. When a target representation is noisy, the optimal strategy for distinguishing successive items in this task is to push them apart, and the visual system appears to implement exactly that strategy, amplifying repulsion under attentional deprivation. At the same time, the insensitivity of downstream biases to inducer fidelity, the long-range influence of T1 on T3, and the backward repulsion from T3 onto T2 collectively suggest that the underlying mechanism is more like an active separation process operating across the entire contents of working memory than a passive weighting of previous percepts as priors.

The work, which was supported by the Scientific and Technological Research Council of Turkey through a 1001 program grant and doctoral and international fellowships to the first author, carries implications well beyond the laboratory. Visual working memory underlies everyday activities as varied as navigating a crowded street, comparing two paint swatches, or tracking a moving vehicle through traffic, and all of these depend on keeping successive percepts distinct. The discovery that attentional lapses strengthen the active separation of memories offers a new lens on the errors people make under cognitive load, and it may inform models of perceptual decision making, visual search, and even clinical conditions characterized by attentional dysfunction. For now, the study stands as a vivid reminder that memory is not a passive recording but a dynamic negotiation between past and present, one in which the brain sometimes decides that the best way to remember what it just saw is to remember it as deliberately different from what came before.

Subject of Research: Repulsive serial dependence in visual working memory and how target fidelity shapes serial biases

Article Title: Repulsive serial dependence in visual working memory modulated by target fidelity

Article References: Yildirim, B., & Boduroglu, A. (2026). Repulsive serial dependence in visual working memory modulated by target fidelity. BMC Biology. https://doi.org/10.1186/s12915-026-02739-9

Image Credits: AI Generated

DOI: 10.1186/s12915-026-02739-9

Keywords: visual working memory, serial dependence, serial bias, attentional blink, orientation estimation, Bayesian inference, perceptual decision making, attention, visual adaptation, working memory maintenance, repulsive bias, target fidelity