Trichotomy revisited: A monolithic theory of attentional control

Abstract

The control of attention was long held to reflect the influence of two competing mechanisms of assigning priority, one goal-directed and the other stimulus-driven. Learning-dependent influences on the control of attention that could not be attributed to either of those two established mechanisms of control gave rise to the concept of selection history and a corresponding third mechanism of attentional control. The trichotomy framework that ensued has come to dominate theories of attentional control over the past decade, replacing the historical dichotomy. In this theoretical review, I readily affirm that distinctions between the influence of goals, salience, and selection history are substantive and meaningful, and that abandoning the dichotomy between goal-directed and stimulus-driven mechanisms of control was appropriate. I do, however, question whether a theoretical trichotomy is the right answer to the problem posed by selection history. If we reframe the influence of goals and selection history as different flavors of memory-dependent modulations of attentional priority and if we characterize the influence of salience as a consequence of insufficient competition from such memory-dependent sources of priority, it is possible to account for a wide range of attention-related phenomena with only one mechanism of control. The monolithic framework for the control of attention that I propose offers several concrete advantages over a trichotomy framework, which I explore here.

1. Attentional control as a theoretical dichotomy

Attention is the mechanism that selectively prioritizes information processing in perceptual systems, allowing organisms to more efficiently manage representational capacity limitations (Corbetta & Shulman, 2002; Desimone & Duncan, 1995). Research on the control of attention seeks to understand what perceptual information is prioritized and why: what are the mechanisms by which attentional priority is allocated? In the context of vision, this issue is typically contextualized with respect to the nature of the features (e.g., Anderson, 2014, 2017b; Wolfe et al., 1989; Folk et al., 1992), objects (e.g., Chen, 2012; Clement et al., 2023; Hickey et al., 2015), and regions of space (e.g., Anderson & Kim 2018a, 2018b; Anderson et al., 2022; Liao, Kim, et al., 2023) that an observer preferentially processes. These aspects of control can be inferred from the target of eye movements (to the most prioritized stimulus at the moment of saccade initiation; Anderson & Yantis, 2012; Theeuwes & Belopolsky, 2012; Theeuwes et al., 1998), performance costs generated by task-irrelevant stimuli in the identification of a search target (attributed to elevated priority allocated to the stimuli responsible for the cost; Anderson et al., 2011a, 2011b; Theeuwes, 1992; Yantis et al., 2012), and neural activity generated by stimuli (prioritized stimuli evoke stronger signals along pathways for perceptual information processing and/or are represented with higher fidelity; Anderson, 2017c; Anderson, Kuwabara, et al., 2016; Anderson, Laurent, et al., 2014; Corbetta & Shulman, 2002; Hickey et al., 2015; Hickey & Peelen, 2015; Serences et al., 2005; Serences & Saproo, 2010).

One of the oldest and most fundamental questions that has been examined with respect to the control of attention is the extent to which the observer has control over what is prioritized in information processing. Can an observer essentially “choose” what to attend to, or is attentional prioritization a product of their brain’s response to stimuli in the environment independently of the observer’s intentions? In investigating this issue, two distinct phenomenology quickly became evident: instructional cues are effective in shaping what is preferentially processed in visual displays (e.g., Jonides, 1981; Posner, 1980) and physically salient stimuli (i.e., stimuli with high local feature contrast; see Duncan & Humphreys, 1989; Itti & Koch, 2001; Nothdurft, 1993) can robustly interfere with the processing of a search target (e.g., Theeuwes, 1992, 2010; Theeuwes et al., 1998). The former came to be referred to as the product of goal-directed attentional control, and the latter as the product of stimulus-driven attentional control (e.g., Egeth & Yantis, 1997; Serences et al., 2005; Yantis & Egeth, 1999).

Stimulus-driven attentional control is distinguishable from goal-directed attentional control in that it can operate in spite of conflicting goals and intentions. When a stimulus produces one of the aforementioned markers of attentional prioritization even if it is known to be irrelevant to the task and even when a task-relevant stimulus competes with it for attentional prioritization, such prioritization is typically referred to as attentional capture (see Anderson, 2013; Folk et al., 1992; Theeuwes, 1992). Attentional capture by a task-irrelevant stimulus implies prioritization that is inconsistent with goal-directed attentional processes. More specifically, a situation in which goal-directed attention predicts that one stimulus should be prioritized (e. g., the target) and the evidence indicates that a different stimulus is prioritized, allows one to reject goal-directed attention as a mechanistic explanation for the observed prioritization; such an outcome demands an explanation that appeals to the properties of the prioritized stimulus. Consistently, physical salience has provided one such explanatory factor in characterizing the control of attention, with physically salient stimuli producing evidence of attentional capture (see Luck et al., 2021; Theeuwes, 2010; for reviews).

A major point of debate arose in the study of attentional control concerning the extent to which goal-directed processes could prevent stimulus-driven attentional capture, and by extension, the extent to which the control of attention was principally goal-contingent or stimulus-driven. The contingent involuntary orienting hypothesis posited that only stimuli that are consistent with the goal state of the observer will be prioritized, with goal-directed attentional processes essentially gating stimulus-driven attention (Folk et al., 1992). This hypothesis was supported by studies demonstrating that physically salient stimuli selectively captured attention when possessing a feature that was diagnostic of target status (e.g., a red distractor when the target is defined as the red stimulus; Anderson & Folk, 2010, 2012; Folk et al., 1992; Eimer & Kiss, 2008, 2010; Folk & Anderson, 2010; Folk & Remington, 1998; Serences et al., 2005). This hypothesis conflicted with findings that the more physically salient of two simultaneously presented stimuli tended to be preferentially processed even when the more salient stimulus was not the target, which implies a failure of goal-contingent attentional control (Anderson & Mrkonja, 2021, 2022; Luck et al., 2021; Theeuwes, 1992, 2010; Theeuwes et al., 1998).

Various attempts have been made to reconcile these two sets of apparently conflicting findings (e.g., Burnham, 2007; Gaspelin et al., 2015, 2017; Gaspelin & Luck, 2018a; Hickey et al., 2009; Leber, 2010; Luck et al., 2021; Serences & Yantis, 2006), further discussion of which is beyond the scope of this review. The interested reader is pointed to a review paper by Luck et al. (2021) that summarizes more recent developments in the debate on how goal-contingent and stimulus-driven processes jointly determine attentional prioritization. A consensus view, however, has long been that goal-directed and stimulus-driven influences play distinguishable roles in the control of attention, reflecting distinct mechanisms underlying the determination of attentional priority (e.g., Corbetta & Shulman, 2002; Connor et al., 2004; Luck et al., 2021; Serences et al., 2005; Serences & Yantis, 2006; Theeuwes, 2010). That is, exactly where one draws the line between the influence of goals and salience in the determination of attentional priority has been a point of longstanding contention, but the idea that a line needs to be drawn at all has never been debated.

2. Attentional control as a theoretical trichotomy

A recent review by Anderson et al. (2021) includes a detailed description of the history of observations that were inconsistent with the dichotomy between goal-directed and stimulus-driven attentional control and the eventual adoption of a theoretical trichotomy emphasizing the influence of goals, salience, and prior experience or selection history (see Awh et al., 2012; Fig. 1A) in the control of attention. For the purposes of this review, I need not provide a thorough recounting of that history here. What I will take the time to emphasize is how history repeats itself and why understanding the cyclical development of theories of attentional control is significant for the current state of theorizing.

Fig. 1.

Trichotomy models of attentional control. (A) Basic trichotomy model (adapted from Awh et al., 2012). The focus is on three sources of input into an integrated priority map. (B) Updated trichotomy model proposed by Anderson et al. (2021). The adapted version depicted here is a simplified version of the Anderson et al. (2021) model that focuses on the different sources of input impinging upon the computation of attentional priority, which are assumed to compete at multiple stages of priority computation (not necessarily all integrated at the same processing stage).

Reasoning from competitive relationships has been central to the development of theoretical distinctions concerning the control of attention. Stimulus-driven attention emerged as a concept when the prioritization of physically salient stimuli conflicted with the predictions that arise from goal-directed prioritization. Given that attention determines which among the different stimuli that compete for representation are prioritized (Desimone & Duncan, 1995), when different theories are represented by different stimuli that compete for attention, prioritization of one stimulus precludes mechanistic explanations associated with prioritization of the other stimuli. Specifically, when one hypothesized mechanism of control predicts that one stimulus will be prioritized and another hypothesized mechanism of control predicts that a different stimulus will be prioritized, observed prioritization of one stimulus allows the researcher to reject the mechanistic influence represented by the other stimulus as an explanation. In the specific case of stimulus-driven attention, any model of attentional control that posits only goal-directed influences becomes difficult to defend when such influences would predict that one particular stimulus would be attended and a different, more physically salient stimulus is shown to effectively compete for attention.

As described in Anderson et al. (2021), influences of prior experience or learning history on the control of attention that did not fit neatly into the goal-directed versus stimulus-driven dichotomy were reported even before the dichotomy was formally articulated. These influences, however, were not directly in conflict with goal-directed and/or stimulus-driven influences, allowing for research into the learning-dependent control of attention to proceed more or less in parallel with the goal-directed versus stimulus-driven debate. The transparent clash between goal-directed and stimulus-driven influences lacked a parallel with respect to any influence of learning history. That changed when a study by Anderson et al. (2011b) revisited the concept of mechanistic competition, creating an experimental condition in which a non-salient and task-irrelevant but previously reward-associated stimulus competed with a more physically salient (and, by definition, goal-relevant) target. Both the stimulus-driven control of attention and the goal-directed control of attention predicted that the target should dominate attentional prioritization under these conditions, as neither of these mechanisms predicted any priority should have been afforded to the previously reward-associated stimulus, yet the previously reward-associated stimulus impaired search for the target (Anderson et al., 2011b). Later findings demonstrated that such experience-driven¹ prioritization can be observed in eye movements (Anderson & Yantis, 2012; Theeuwes & Belopolsky, 2012; see also Anderson & Kim, 2019a, 2019b; Kim & Anderson, 2019a, 2023b) and distractor-evoked brain activation (Anderson, Laurent, et al., 2014; see also Kim & Anderson, 2019a, 2019b, 2020a, 2020b; Kim, Gregoire, et al., 2021; Kim, Nanavaty, et al., 2021; Liao, Kim, et al., 2023), and can persist for months after stimulus-reward associations are initially learned (Anderson & Yantis, 2013).

By demonstrating learning-dependent attentional prioritization that was observable in spite of conflicting influences of goals and salience, it became clear that neither goal-directed nor stimulus-driven attentional control could provide a suitable explanation for what Anderson et al. (2011b) termed value-driven attentional capture. Later evidence extended this phenomenon to attentional capture driven by learned associations between stimuli and aversive outcomes, either controlling for the influence of salience (Anderson & Britton, 2020; Schmidt et al., 2015) or similarly using non-salient distractors (Kim & Anderson, 2021b; Nissens et al., 2017). More recent findings suggest that reward learning and aversive conditioning influence attention via a common underlying mechanism of experience-driven prioritization driven by valence or survival relevance (Gregoire et al., 2022; Kim, Lee, et al., 2021; Kim & Anderson, 2021a, 2021b, 2023b; see Anderson et al., 2021, for a review), suggesting a more general influence of associative learning in the control of attention.

Anderson et al. (2021) describe several other influences of learning history on the control of attention that cannot be adequately explained by an influence of goals and physical salience on priority assignment (see also, Failing & Theeuwes, 2018). They include history as a sought target, statistical dependencies among objects, statistical dependencies in object location, statistics concerning stimulus frequency, and intertrial priming. Specifically, participants develop a habitual tendency to preferentially process stimuli that they have a history of intentionally orienting to, reflecting a residual influence of goal-directed prioritization (e.g., Anderson & Britton, 2019; Kim & Anderson, 2019c; Kyllingsbaek et al., 2001; Qu et al., 2017; Sha & Jiang, 2016; Shiffrin & Schneider, 1977). The spatial location at which a target more frequently appears will become prioritized with experience (e.g., Geng & Behrmann, 2002, 2005; Jiang, 2018; Jiang et al., 2013; Jiang & Swallow, 2012), as will the likely target location when the arrangement of objects within the search display is predictive of that location (e.g., Chun & Jiang, 1998, 2003; Colagiuri & Livesey, 2016; Jiang & Wagner, 2004). On the other hand, task-irrelevant stimuli can be deprioritized when frequently appearing in a predictable location (e.g., Britton & Anderson, 2020; Goschy et al., 2014; Leber et al., 2016; Sauter et al., 2019; Wang & Theeuwes, 2018a, 2018b, 2018c) or when frequently rendered in a predictable color (e.g., Failing et al., 2019; Stilwell et al., 2019; Vatterott & Vecera, 2012) or shape (Kim, Ogden, et al., 2023; see also Ogden et al., 2023). Stimuli that infrequently appear in a particular context or situation have also been shown to be prioritized in what is referred to as novel orienting (e.g., Horstmann & Ansorge, 2006, 2016; Horstmann & Herwig, 2016; Horstmann, 2002; Johnston et al., 1990, 1993; Johnston & Schwarting, 1997), and in the case of intertrial priming, selecting a target or rejecting a distractor facilitates the same manner of prioritization on the following few trials (e.g., Geyer et al., 2006; Kristjansson, 2006; Kristjansson & Campana, 2010; Kristjansson & Driver, 2008; Kristjansson et al., 2002; Lamy et al., 2008; Maljkovic & Nakayama, 1994, 1996; Treisman, 1992; Wang et al., 2005).

In light of these and related findings, Anderson et al. (2021) proposed at least three distinct mechanisms of experience-driven attention that collectively constitute selection history (Fig. 1B). The first reflects the influence of learned associations between stimuli and valent outcomes (reward and punishment), the second reflects stimulus–response habit learning, and the third reflects statistical learning. Intertrial priming was subsumed within each of these three mechanisms of control as an immediate consequence of the teaching signals that give rise to more stable and protracted learning-dependent attentional biases. Although the precise mechanistic distinctions proposed by Anderson et al. (2021) are debatable and other taxonomies of attentional control mechanisms could be considered, it has become clear that the third arm of the theoretical trichotomy that is selection history is complex and multifaceted, encompassing more than one distinct learning-dependent attentional process. Any comprehensive theory of attentional control needs to lend understanding to the nature of these distinct learning-dependent attentional processes, and how they relate to one another and to other attentional control processes.

3. The case for a monolithic framework for attentional control

When considering the reasons why current theoretical models of attentional control are organized around a trichotomy framework, the history of thinking on attentional control remains pertinent. The distinction between goal-directed and stimulus-driven attention predated the concept of selection history, which could be reasonably accounted for by appealing to a single learning-dependent mechanism of attentional control at the time of its inception (Awh et al., 2012; Failing & Theeuwes, 2018). At that point in history, as recounted in Section 2, it was natural to focus on distinctions and posit a novel mechanism of attentional control, namely the concept of selection history, that ushered in the theoretical trichotomy. It would have been difficult to have had a productive conversation about the nature of learning-dependent influences on attentional priority without first establishing that they are not merely different flavors of goal-directed attention, allowing for the conclusion that they are, in fact, fundamentally experience-driven.

As the diversity of experience-driven influences on attention started coming to light (see Anderson et al., 2021), the theoretical gravity of the trichotomy framework remained to anchor the evolution of thinking. It is worth asking, had the historical lines of demarcation not existed and we were to evaluate the current state of the literature afresh, would we still land on a theoretical trichotomy to make sense of this diversity? Is a trichotomy still the most natural and parsimonious framework for organizing our thinking concerning mechanisms of attentional control, or is it more of an echo of highly influential theorizing that may no longer best account for the fullness of the data? That is to ask, in a rather ironic way, is the maintenance of a theoretical trichotomy itself a bit of a product of selection history?

3.1. Where lines are drawn

The theoretical model for the control of attention put forth by Anderson et al. (2021; see Fig. 1B) is considerably more complex than the model forwarded by Awh et al. (2012) that coined the term selection history (see Fig. 1A), introducing several additional mechanistic distinctions. The issue that I am raising in this review is not a question of whether either model is correct or incorrect in the distinctions it draws. It seems reasonable to assert that experience-driven influences on the control of attention that are not reducible to goal-directed and stimulus-driven influences exist, as asserted by Awh et al. (2012; Fig. 1A), and also that these experience-driven influences are multifaceted, as asserted by Anderson et al. (2021; Fig. 1B). In this sense, the theoretical positions offered by Anderson et al. (2021) and Awh et al. (2012) are complementary and each capture meaningful variance in the landscape of attentional control. Where they draw their lines is not wrong, but they are by no means the only ways one could characterize attentional control.

3.2. Two steps back

In reevaluating the state of the literature as it stands at the writing of this review, I find myself landing on what I believe to be a more concise and parsimonious framework for conceptualizing the variety of influences on attentional priority computations. I find my thinking centered on a common theoretical motif that permeates all of these influences and favoring a framework that avoids placing arbitrary categorical distinctions at the forefront of its organization. I find myself similarly rejecting a theoretical dichotomy between goal-directed and stimulus-driven influences, as Awh et al. (2012) did more than a decade ago, but instead taking a step in the other direction, toward a monolithic theory of attentional control. My proposed framework emphasizes what goal-directed attention and selection history have in common while simultaneously acknowledging their distinctiveness, affording no special status to distinctions involving goal-directed and experience-driven influences versus distinctions within the context of selection history. This framework also characterizes salience differently than Awh et al. (2012) and Anderson et al. (2021). Although the influence of physical salience on the control of attention is robust, this influence does not reflect the product of a dedicated control mechanism but rather an outpouring of how the computation of attentional priority unfolds and is therefore represented in a fundamentally different way. Although in some ways a superficial modification, the treatment of salience in the proposed framework helps to clarify the nature of attentional control, distinguishing between the influence of control mechanisms and baseline influences on priority computations.

In the remainder of Section 3.2, I outline some of the arguments against a theoretical trichotomy as the organizing framework for conceptualizing the control of attention. From there, in Section 3.3, I outline an alternative theoretical framework for the control of attention that seeks to harmonize some of the complexities that arise from where a trichotomy framework falls short. The review goes on to summarize some of the implications of the proposed framework in Section 4, articulating some of the advantages of a monolithic framework over a trichotomy framework, and then highlights some unresolved issues in Section 5.

3.2.1. The problem with three

When it comes to characterizing mechanisms of attentional control, three is not enough. As described towards the end of Section 2., mechanistic distinctions have been made with respect to different experience-driven influences on the control of attention. These distinctions are rooted in dissociations between experience-driven influences on the control of attention (e.g., Anderson & Britton, 2019; Anderson et al., 2017; Kim & Anderson, 2019a, 2019c) as well as additive effects of different experience-driven influences on the computation of attentional priority (e.g., Kim & Anderson, 2021a; Ogden et al., 2023; Stankevic & Geng, 2014). Anderson et al. (2021) provide a detailed argument for these distinctions and a theoretical model that accommodates them by characterizing selection history as a multifaceted umbrella concept (Fig. 1B).

A central tenant of the trichotomy framework for attentional control is that there is a categorical distinction between the influence of goals and selection history (Awh et al., 2012; Fig. 1A). The key question then becomes whether the distinction between goal-directed and experience-driven influences is any more consequential than other distinctions that could be drawn among the different experience-driven influences classified under the umbrella of selection history. So consequential, in fact, that the distinction between goal-directed and experience-driven influences continues to warrant entirely separate categories of influence in theoretical models (Fig. 1A & 1B).

There are three ways one can respond to the complexity posed by the fractionating of selection history (Anderson et al., 2021). The first is to simply abandon categorical distinctions and focus on the distinctiveness and independence of each putative control mechanism (see Fig. 2). This rejects the concept of a trichotomy in favor of something of a mechanistic multiverse. Theorical models of this flavor are lacking in organization and likely for this reason have not received realistic consideration in the literature. In the end, though, I actually come to the conclusion that the organization of Fig. 2 is not without virtue in eschewing categorical boundaries, although it yet misses the forest for the trees while simultaneously conflating the influence of salience with a control mechanism.

Fig. 2.

Mechanisms of assigning attentional priority without any organizational framework, eschewing a distinction between the influence of current goals and selection history. Rather than fractionate selection history as in Fig. 1B, each factor influencing attentional priority is represented as functionally distinct and therefore gets its own box. Models like this have not been given realistic consideration in the field but may not be without some virtue that will be better captured by the proposed model.

Alternatively, one can maintain the spirit of a trichotomy while fractionating selection history into multiple distinct underlying mechanisms. This is the approach recently taken by Anderson et al. (2021; Fig.1B), which is broadly consistent with what remains as the dominant perspective of the field. With respect to such a framework, a categorical distinction between influences on attention that are experience-driven (selection history) and subject to momentary considerations (goal-directed attention)—and with it the spirit of a trichotomy—is upheld.

3.2.2. Trichotomy revisited

The alternative response to the complexity posed by the revelation of new mechanistic distinctions in the context of selection history is to look for commonalities among all of the different hypothesized mechanisms of attentional control. The question then becomes whether these commonalities overshadow whatever distinctions there may be, in which case new lines of demarcation should be drawn accordingly. With broad enough commonality, a case can be made for the contraction of the theoretical model of attentional control and a reduction in the number of hypothesized categories of influence from a trichotomy back to a dichotomy or potentially even to a monolithic account. In the context of the specific arguments that I will make here, if (1) distinctions between goal-directed and experience-driven influences on attentional priority are not more substantive than distinctions within the domain of selection history and (2) there is common ground across all of these influences, the common ground may be the more productive organizing framework around which to build a theoretical model.

There are two key components to the argument needed to support the viability of a theoretical model rooted in common ground and a corresponding reduction in the number of distinct categories of attentional control mechanisms. The first, which will be tackled in Section 3.2.2.1, is defining what the common ground is, and the second, which will be tackled in Section 3.2.2.2, is deciding whether it is justified to maintain a categorical distinction between the influence of goals and selection history in light of what they do and do not share in common. Section 3.2.2.3 will address how we characterize salience with respect to mechanisms of attentional control, taking the fullness of the historical trichotomy into account.

3.2.2.1. The common ground.

One thing that goal-directed and experience-driven influences on the control of attention have in common is that they all involve memory-dependent bias signals. In the context of selection history, memory dependence is self-evident. In the context of goal-directed influences, it is widely accepted that target templates held in working memory and activated long-term memory serve as biasing signals to the attention system (e.g., Olivers et al., 2006, 2011; Reinhart & Woodman, 2015; Woodman & Arita, 2011; Woodman et al., 2013). Goals do not have a direct influence on attentional priority, but rather goal-directed influences are realized through adjustments in the contents of volitionally activated memory representations. Particularly with respect to working memory, which is likely especially pertinent when adjusting to new task demands (e.g., Woodman et al., 2013), the memory systems involved are clearly different between goal-directed and experience-driven influences; the specific memory systems involved is not a source of common ground. Indeed, there are likely multiple memory systems that support the variety of experience-driven influences on the control of attention (Hutchinson & Turk-Browne, 2012), for example procedural memory supporting habitual orienting and associative memory subsumed within the affective system of the brain (e.g., Anderson, 2019; Kim & Anderson, 2019a; Kim, Nanavaty, et al., 2021).

As simple as it may seem, the fact that goal-directed and experience-driven mechanisms of attentional control are all fundamentally the product of memory-dependent prioritization cannot be understated. Goal-directed influences ultimately reflect the influence of a particular memory system and experience-driven influences are in no way unique with respect to their dependence on memory. Goal-directed and experience-driven mechanisms of attentional control are certainly distinguishable, differing notably in both their flexibility and time course. Goal-directed influences tend to be more flexible and momentary, whereas experience-driven influences tend to be more rigid and emerge over time. But, as will be argued below, these generalizations exist in the context of what is in fact blurry boundaries between the two categories of control. Are the differences between goal-directed and experience-driven influences so intrinsically meaningful as to warrant a categorical distinction that supersedes any distinctions between memory systems subsumed within the domain of selection history? As explicated in Section 3.2.2.2 below and throughout the remainder of the paper, I have come to the conclusion that maintaining this categorical distinction between the influence of goals and selection history is not the most productive way to think about attentional control.

3.2.2.2. Blurring the distinction: On the relationship between goal-directed attention and selection history.

Identifying a common principle across the influence of goals and selection history in the control of attention is not alone sufficient to justify a framework organized around that commonality. To the degree that distinctions between goal-directed and experience-driven influences overshadow this commonality, a categorical distinction in the context of a theoretical model may be defensible. Recent findings, however, point to a fuzzy boundary between where goal-directed influences stop and the influence of selection history begins, which is antithetical to a categorical boundary between them that is built into trichotomy models of attentional control (Fig. 1A & 1B).

The first dimension of the fuzzy boundary between the influence of goals and selection history is with respect to time course. Priming, which has been categorized under the umbrella of selection history since the dawn of the theoretical trichotomy (Awh et al., 2012; see also Anderson et al., 2021), is similarly rapid and transient, and several other components of selection history can be fairly rapidly instantiated and/or can adapt to changing task conditions (e.g., Anderson et al., in press; Anderson, 2015a, 2015b; Gregoire et al., 2021; Kim, Anderson, et al., 2021; Kim, Gregoire, et al., 2023; Le Pelley et al., 2015; Nissens et al., 2017). On the other hand, goal-directed attention can be supported by long-term memory representations built through experience repeatedly searching for a particular stimulus (Woodman et al., 2013). Indeed, it has been hypothesized that attentional control settings that serve in the interest of prioritizing goal-relevant stimuli are in large measure the product of experience with a task, such that goal-directed attentional control emerges gradually with learning (Vecera et al., 2014). Distinctions with respect to time course and flexibility therefore seem continuous rather than categorical in nature and goal-directed attention is not unique in its ability to be instantiated rapidly and/or flexibly, which raises questions concerning the categorical distinction between the influence of goals and selection history advocated by the trichotomy framework for attentional control.

The second dimension of the fuzzy boundary between the influence of goals and selection history in the control of attention is with respect to function. Multiple studies have provided evidence that, when observers have the choice of how to search, goal-directed attention proceeds in a manner that is shaped by selection history. Reward history (Lee et al., 2022), statistics concerning target frequency (Lee & Anderson, in press), and statistics concerning the distribution of stimuli in the search array (Clement et al., 2023; Kim et al., in press) all shape what observers prioritize when the task places limited constraints of what they must search for. It is also the case that habit-like orienting can quickly develop when a searched-for target repeats over trials (Kim & Anderson, 2019c; Kyllingsbaek et al., 2001; Qu et al., 2017; Sha & Jiang, 2016; Shiffrin & Schneider, 1977). Indeed, the extent to which goal-directed attentional control is in fact a voluntary attentional process has been questioned (Anderson, 2018; see also Anderson, in press), and it is hypothesized that one of the functions of experience-driven influences is to offload the need for the active control of attention via the maintenance of a target template in working memory (Anderson, 2021a; Anderson, in press).

All this is to say that the influence of selection history can support the accomplishment of task goals and that attentional control settings are not the exclusive domain of what has traditionally been characterized as goal-directed attention. This muddies the conceptual distinction surrounding what exactly constitutes goal-directed attention in the first place and why it would warrant its own category of influence. If, in many circumstances, prioritization of a goal-relevant stimulus can proceed via experience-driven influences, it does not seem to be the case that biasing signals from working memory or activated long-term memory have a unique role in supporting goal-contingent information processing or the promotion of goal-directed behavior more broadly. In some (perhaps many) cases, selection history is the mechanism by which goal-contingent information processing is realized (Anderson, 2018). This has been argued to reflect a key part of what makes mechanisms of experience-driven influences on the control of attention adaptive (Anderson, 2021a).

3.2.2.3. On the relationship between salience and control

The argument against a theoretical trichotomy as the chief organizational framework for the control of attention has centered on the distinction between goal-directed and experience-driven mechanisms of control. Models of attentional control have consistently represented physical salience as a distinct contributor to attentional priority (Anderson et al., 2021; Awh et al., 2012; Theeuwes, 2018, 2019), and justifiably so given the well-documented cases of stimulus-driven attention (e.g., Theeuwes, 2010). In revisioning the theoretical framework for attentional control, however, I have come to the conclusion that depicting physical salience alongside memory-dependent influences in a parallel fashion, as in Anderson et al. (2021; Fig. 1B) and Awh et al. (2012; Fig. 1A), is more confusing than helpful. This is because, although physical salience absolutely contributes to the computation of attentional priority, it does not do so via a dedicated control mechanism. That is, although stimulus-driven attentional capture is a robust phenomenon, there is no such thing as stimulus-driven attentional control.

By explicitly depicting physical salience alongside hypothesized mechanisms of assigning attentional priority, the models of Awh et al. (2012; Fig. 1A) and Anderson et al. (2021; Fig. 1B) do not in fact depict mechanisms governing the control of attention, but rather depict influences that impinge upon the computation of attentional priority. The result is that some of the depicted influences (goals and selection history) reflect control mechanisms that actively adjust attentional priority while another (physical salience) reflects a passive influence of the strength of sensory signals on attentional priority. Correspondingly, in the model of Anderson et al. (2021; Fig. 1B), goals and the different components of selection history all have plus and minus signs to indicate that they can both upweight and downweight priority (see also Theeuwes, 2018), while physical salience only has a plus sign; they are transparently not comparable.

This conflation between attentional control mechanisms and mere influences on priority can easily engender confusion. When Awh et al. (2012), Anderson et al. (2021), and others discuss the role of goals, selection history, and salience, the conversation can shift between discussing these concepts as a causal agent of priority adjustments and as a category of effect on priority, often without warning. There can be a surprising amount of ambiguity with respect to whether causes or effects are being discussed in the context of attentional control, a distinction that has been a point of criticism when debating the utility of the very concept of selective attention (Anderson, 2011; Hommel et al., 2019; Krauzlis et al., 2014). Within this context, it can be easy to conflate the influence of physical salience with the consequence of a dedicated control mechanism for priority assignment, which would be mistaken. I would argue that we either need to depict the influence of physical salience in the control of attention differently and more clearly in theoretical models (see Luck et al., 2021), or we need to clarify that the model is not one of attentional control but rather of categories of effects on priority.

A model merely describing effects on attentional priority, however, is unlikely to offer much in the way of explanatory power beyond accounting for extant observations and is therefore a low-impact target for theoretical development. Models of attentional control should focus on the processes that modulate attentional priority and how this modulatory influence is applied, a more thorough understanding of which allows one to generate novel predictions concerning what will be prioritized under what conditions. This is the approach I take here. If one takes this approach, it quickly becomes clear that physical salience is best characterized as the baseline level of input upon which mechanisms of attentional control operate, an idea that is built into the monolithic framework for the control of attention that I propose in Section 3.3.

A potential objection to the rejection of a salience-driven mechanism of attentional control could be raised with respect to the concept of singleton detection mode. Singleton detection mode reflects a hypothesized search strategy by which observers search for a physically salient target on the basis of its salience, strategically prioritizing physically salient stimuli even if it comes at the expense of initially orienting to a more physically salient distractor (e.g., Bacon & Egeth, 1994; Lamy & Egeth, 2003; Leber & Egeth, 2006). As inefficient and effortful as such a strategy may be (Lee et al., in press; see also Anderson & Lee, 2023), singleton detection mode can be straightforwardly accounted for using the model I will propose by characterizing it as the absence of (or limited allowance for) memory-dependent priority adjustments. That is, singleton detection mode can be thought of as what happens when an individual does not volitionally bring activated memory representations to bear in the computation of attentional priority, a concept that will be explicated more fully in the subsequent section.

3.3. Attentional control with a single overarching mechanism for assigning priority

Different stimuli in the environment generate signals in sensory organs that are fed forward along multiple information processing pathways (e.g., Anderson, 2019; Corbetta & Shulman, 2002), guiding orienting responses and other behaviors. At different points along these pathways, these signals undergo modulation such that information pertaining to certain sources of perceptual input is amplified while information pertaining to other sources of perceptual input is attenuated (see Anderson, 2021c). The sum of the strength of sensory input and such modulations of this input reflects attentional priority, and the modulations themselves reflect what we might call the control of attention.

3.3.1. The attentional control state defined

It is generally accepted that the activation and maintenance of information content in memory can encompass sensory areas (e.g., Harrison & Tong, 2009; Serences et al., 2009). Although the extent to which memories are actually represented in sensory areas of the brain is unclear (e.g., Ester et al., 2015; Hallenbeck et al., 2021; Sreenivasan et al., 2014), that the activation of memory could, in a general sense, influence brain activity in sensory areas is uncontroversial. Such memory-dependent activation provides a putative mechanism by which modulations of signals evoked in perceptual systems that constitute the control of attention are realized.

The monolithic theory for the control of attention that I am proposing here will borrow a term long-ago introduced in the attentional control literature (Folk et al., 1992) that recently made a comeback (Luck et al., 2021), which is the idea of an attentional control state (see Table 1 & Fig. 3). Here, attentional control state can be defined more specifically and explicitly as the state of activated memory that impinges upon perceptual systems. Such activated memory constituting the attentional control state can be conceptualized as something akin to a filter and/or amplifier, modulating signals passing through neurons whose resting state and responsivity has been modulated by the control state (Fig. 4).

Table 1.

Determinants of attentional priority hypothesized by the proposed model of attentional control.

Term	Definition
Attentional Control State	Activated memory that impinges upon perceptual systems, such that perceptual input consistent with the contents of memory is either amplified or attenuated.
Tonic Attentional Control Signals	Sustained neural activity that constitutes the current attentional control state. Updating the contents of currently activated memory results in a corresponding update in the tonic attentional control signals associated with that memory.
Phasic Attentional Control Signals	The modulation of signals evoked in perceptual systems that constitute priority adjustment, which are caused by tonic attentional control signals. In this respect, phasic attentional control signals are the actualization of attentional control in perceptual information processing and what is typically measured in studies of attentional control using cognitive neuroscience methods.

Fig. 3.

The proposed model of attentional control. There is only one mechanism for assigning attentional priority, subsumed within the attentional control state. Different memory systems collectively comprise the attentional control state, several of which contain representational overlap. Note that the overlap in the circles depicts representational overlap with respect to the architecture underlying the attentional control state, not overlap in the features or locations to which adjustments in priority are applied (which can covary; e.g., the same feature can be prioritized according to both its positive valence and relevance to current goals). Physical salience is depicted in the weight (thickness) of the arrows reflecting different sources of sensory input, which is modified by the control state. Priority weights for features are depicted by colored circles with a plus or minus sign, and priority weights applied to representations of spatial information are depicted with a heat map; the choice of which to link to a given memory system is simply intended to depict one possible manifestation of the attentional control state and is not intended to reflect the full scope of any given source of priority (e.g., there is strong evidence that attentional priority based on statistical learning tied to historical relevance can be spatial in nature in addition to feature-based; see, e.g., Wang & Theeuwes, 2018a, 2018b, 2018c). An “endpoint representation” is depicted as the target of perceptual input that has been modulated by the attentional control state rather than a priory map specifically, as a priority map reflects only one of many perceptual representations that could be affected by the control state.

Fig. 4.

Example of attentional priority computations unfolding over time as hypothesized by the proposed model. Nodes in a hypothetical neural network are indicated by donut-shaped circles, and the flow of neural signals by arrows, with dotted lines reflecting neural signals that are tonically active in a particular context and solid lines reflecting sensory-evoked activity emanating from the eyeball. The thickness of the lines and donut-shaped circles denotes the strength of corresponding neural activity. (A) Tonic attentional control signals in which memory-related activity influences how cells within nodes of the network for visual information processing will respond to input consistent with the color indicated, reflecting red as a goal-relevant color and blue as a color prioritized via selection history. (B) As visually evoked signals move through the network, signals prioritized by the attentional control state are amplified at the points at which they come into contact with tonic attentional control signals. This amplification reflects phasic attentional control signals, which then feed forward through the network. Although the figure only depicts feedforward signals for the sake of simplicity, phasic attentional control signals also likely involve some degree of back-propagation as well. Note that the nodes depicted in the network are not intended to be inclusive; rather, a select few notes were chosen to illustrate the concept.

According to the proposed theory, each distinct mechanism of attentional control reflects a distinct memory system that can contribute to the control state, although the different memory systems involved will have some degree of representational overlap (Fig. 3). While mechanisms of attentional control can be compared and contrasted in a number of ways, each is fundamentally dissociable from the rest and distinctly memory dependent. That is, each relies on a distinct supporting memory system. In this respect, it would be unnatural to draw a categorical distinction between goal-directed attention and the experience-driven mechanisms of attentional control that collectively constitute selection history; each is simply one instance of a memory system contributing to the overarching control state.

3.3.2. The realization of control

This leads us to a critical distinction that in my opinion has not been clearly articulated within existing organizational frameworks for the control of attention, which is the distinction between tonic and phasic attentional control signals (Table 1 and Fig. 4). Tonic attentional control signals reflect sustained neural activity that constitutes the current attentional control state, which reflects the contents of activated memory impinging upon perceptual systems. Phasic attentional control signals reflect the modulation of signals evoked in perceptual systems that constitute priority adjustment, which are caused by the control state. That is, phasic attentional control signals reflect the actualization of attentional control in perceptual information processing.

Phasic attentional control signals are what is typically measured in studies of attentional control using cognitive neuroscience methods (e. g., Anderson, 2019; Kim, Nanavaty, et al., 2021; Corbetta & Shulman, 2002). This can be thought of as the product of the implementation of attentional control. Tonic attentional control signals are more difficult to measure, as they do not reflect transient modulations in the gross intensity of neural activation in any one brain area; rather, tonic attentional control signals reflect shifts in the complexion of neural activity, or the content of patterns of activation. However, the proposed framework makes the straightforward prediction that the attentional control state should be decodable within the brain in the absence of stimulus input, for which there is some evidence (e.g., de Vries et al., 2019; Serences & Boynton, 2007; Wen et al., 2019).

By the proposed account, the relationship between the contents of memory and the attentional control state is an intrinsic property of the organization of the memory system of the brain and its connections with perceptual systems. That is, certain memory systems contain representational overlap with perceptual systems, which serves in part to shape how incoming perceptual information is processed in these systems. Within this context, simply bringing a particular memory representation into an active state is itself tantamount to updating the attentional control state; there is no translation that needs to take place between memory and attentional control. An organism does not need to allow memory to influence attention; rather, the contents of memory will influence attention regardless of the intention for it to do so, an assertion that is well-precedented by evidence linking the contents of memory to the involuntary control of attention (e.g., Nickel et al., 2020; Olivers et al., 2006, 2011).

3.3.3. Accounting for the when and where of attentional control

The proposed framework provides a straightforward explanation for context-dependent effects on experience-dependent attentional biases in which previously reward-associated (Anderson, 20215a, 2015b) and punishment-associated (Gregoire et al., 2021) stimuli capture attention selectively in the contexts in which they predicted the associated outcome. In this case, the context triggers the context-dependent memories that constitute the attentional control state. This framework also provides a straightforward explanation for how value-driven attentional biases can be evident months after learning (e.g., Anderson & Yantis, 2013), as when the associated attentional control state is tied to a specific context, it can be in large measure shielded from interference from subsequent learning in other contexts. Similarly, the precedence of prior over more recent reward learning in value-driven attention could reflect the time needed to consolidate new learning into the attentional control state (Liao & Anderson, 2020).

By the proposed account, the influence of intertrial priming on the control of attention simply reflects the influence of residual task-evoked activity in memory systems subsumed within the attentional control state. Such an explanation is not incompatible with the account proposed by Anderson et al. (2021), which posited that intertrial priming reflects a residual consequence of the teaching signals responsible for the updating of experience-dependent attentional priority. Such teaching signals would be expected to occur at the level of the memory representations that constitute the attentional control state as hypothesized in the proposed account, but the proposed account opens the door for a broader influence of residual brain activity evoked by the prior trial, as the attentional control state includes attentional control signals beyond those subsumed within the updating of experience-driven attentional control.

The proposed account additionally offers a straightforward explanation for phasic attentional control signals linked to selection history that arise strikingly early in the stream of visual information processing, in some cases seemingly too early to reflect much in the way of feedback processing (e.g., Itthipuripat et al., 2019; Kim & Anderson, 2022; MacLean & Giesbrecht, 2015; Serences, 2008; Serences & Saproo, 2010; van Koningsbruggen et al., 2016; see Anderson, 2019, for a review). By the proposed account, early visual cortical representations are subsumed within the attentional control state, allowing for phasic experience-driven attentional control signals to arise as soon as sensory information reaches an area such as V1 (see Fig. 4). Subcortical memory systems responsible for components of selection history such as value-driven and aversively conditioned attentional biases have rich connections with early-stage sensory processing (e.g., Arsenault et al., 2013; Shuler & Bear, 2006), allowing for attentional control to be realized in this way. This framework fits naturally with the model for the neurobiology of value-driven attention proposed by Anderson (2019), including with respect to the explanation why value-driven attentional biases can emerge prior to goal-directed modulations of attention priority (see Fig. 4), providing a richer and more concrete theoretical explanation for why the phasic attentional control signals exhibit the time course that they do.

3.3.4. Accounting for stimulus-driven influences in the control of attention

According to the proposed account, attentional capture by salient stimuli happens when the strength of sensory energy evoked by a salient stimulus is sufficiently strong that phasic modulations of attentional priority caused by the attentional control state fail to outcompete the signal evoked by the salient stimulus, at least at earlier stages of information processing when attentional priority is typically probed in attentional capture tasks (see Anderson & Mrkonja, 2021, 2022; Donk & van Zoest, 2008; Godijn & Theeuwes, 2002; van Zoest et al., 2004). As described above, by the proposed account, the phenomenon of singleton detection mode simply reflects a weak influence of an attentional control state, particularly with respect to working memory representing features that should be prioritized in relation to task goals. There is no mechanism of attentional control devoted to prioritizing salient stimuli; salience rather constitutes the baseline perceptual signals that are modulated by the current attentional control state (see Fig. 3). For the control state to effectively prioritize a stimulus, phasic attentional control signals need to shift the balance of activity in perceptual systems such that any more salient stimuli are no longer the stimuli with the stronger representation.

3.3.5. On the issue of priority and selection

The proposed framework is principally concerned with the computation of attentional priority, specifically whether and how perceptual input is modulated by the attention system. Multiple memory-dependent mechanisms of assigning priority function in parallel, with the result being a mosaic of representational strength across the different sources of perceptual input, which is reflected in the content of Fig. 3. But this cannot be the whole story, as one of the hallmarks of attention is its potential to produce a striking degree of representational selectivity. Indeed, some of the earliest theories of attentional control likened attention to a mechanism that selects what information passes through an information processing bottleneck (Broadbent, 1958) or to a spotlight unto perception (Posner et al., 1980). Although perspectives on the nature of attention have evolved over the years, the highly selective nature of information processing remains a core issue for which attention is expected to provide an explanation (e.g., Serences & Yantis, 2006), including with respect to the striking phenomenon of inattentional blindness (Mack & Rock, 1998; Rensink et al., 1997).

Without further qualification, the proposed theoretical framework would produce a perceptual experience in which every source of perceptual input is represented along a broad continuum, and it would be easy to envision situations in which several sources of input are simultaneously represented robustly. Such a state affairs is indeed what is reflected in the content of Fig. 3. The content of that figure, however, and the proposed framework more broadly, should be understood as the state of information processing before a high degree of selectivity is achieved. Indeed, the proposed framework for attentional control is hypothesized to serve as the foundation upon which perceptual input is ultimately selected for representation at the most capacity-limited stages of information processing.

Although the proposed framework does not itself provide a mechanism for differentiating one or two sources of perceptual input from others in the sort of winner-take-all fashion that traditional conceptualizations of attention tend to assume, it is not difficult to fit the proposed mechanisms of priority adjustment into models of how such ultimate selectivity is achieved. One potentially appealing account of selective attention with which the proposed framework could link up with quite naturally is the diachronic account of attentional selectivity (Zivony & Eimer, 2022). By this account, when attentional priority reaches a certain threshold (referred to as the “engagement threshold”), an attentional episode is triggered that results in intense amplification of the input that triggered the episode and competitive interactions by which other sources of perceptual input are attenuated. More than just a mechanism to facilitate the targeting of saccades, attentional episodes reflect dynamic modulations of covert attention that function to promote highly selective information processing. A natural place for such selectivity to occur is in the context of attentional priority maps in the brain, with selection occurring over the region of highest priority (provided that it has exceeded the engagement threshold), although the same principle could be applied to feature-selective representations as well (Thayer & Sprague, 2023).

3.3.6. Organizational framework in historical context

The idea that the control of attention can be guided by the influence of different memory systems is not new. Hutchinson and Turk-Browne (2012) argued that priming, associative learning, working memory, episodic memory, and semantic memory can all bear an influence on the information that is attended. They further suggest that these memory systems likely influence attention in parallel, consistent with the notion of an attentional control state advocated here. The proposed account provides a concrete and comprehensive theoretical framework through which to contextualize such memory-dependent influences, further implications of which are discussed in the section to follow. When the emphasis is placed on the role of memory in attentional control, as advocated by Hutchinson and Turk-Browne (2012), the lines of demarcation advocated by a trichotomy framework for the control attention become arbitrary; the proposed monolithic framework reflects this (Fig. 3), providing a more comprehensive and unifying account of how attentional prioritization occurs.

Consistent with traditional models of biased competition (Desimone & Duncan, 1995; Duncan et al., 1997), the proposed framework for the control of attention assumes that exogenous (physical salience) and endogenous factors (the control state) combine to jointly determine the selectivity of information processing. Similar assumptions are made with respect to input into priority maps as hypothesized by Wolfe’s model of Guided Search, the most recent iteration of which explicitly acknowledges the role of history-dependent input (Wolfe, 2021). Hypothesizing a single, multifaceted control state bears some resemblance to the concept of pertinence as represented in Bundesen’s Theory of Visual Attention (Bundesen, 1990).

Although not a defining feature, the hypothesized influence of the control state in the proposed framework allows for robust history-dependent signal suppression, in accordance with recent conceptualizations of attentional control in the context of the signal suppression hypothesis (Luck et al., 2021; see also Anderson & Kim, 2020; Gaspelin & Luck, 2018a, 2019). The signal suppression hypothesis seeks to explain why salient stimuli sometimes are and are not prioritized over a task-relevant stimulus, affirming a robust influence of both goal-contingent and stimulus-driven influences on selection and asserting that the control of attention is principally neither. The proposed monolithic framework for the control of attention is entirely compatible with this assertion and can straightforwardly accommodate the predictions made by the signal suppression hypothesis. The influence of current task-relevance (Fig. 3), as reflected in the content of working memory, is expected to be relatively slow in modulating incoming visual information processing, in contrast to the more rapid influence of the variety of history-dependent mechanisms, as hypothesized by Theeuwes (2018). Also consistent with Theeuwes (2018), the influence of current task-relevance on the control of attention is highly flexible and subject to momentary adjustment, which reflects a hallmark of working memory (Baddeley, 1992), while history-dependent influences tend to exhibit more limited flexibility and adjust in response to the accumulation of experience.

The proposed monolithic framework (Fig. 3) contains similar mechanistic distinctions made by Anderson et al. (2021), but beyond the conspicuous move to a single overarching category of control mechanism represented by the attentional control state, relationships between different components of the attentional control state are also emphasized, reflecting hypothesized overlap in the memory systems involved. The memory linked to learned reward and punishment associations share representational overlap, reflecting unsigned valence or survival relevance, but the totality of the underlying memory that constitutes the associated attentional control state is distinguishable between the two. What was referred to as stimulus–response (S-R) habit learning in Anderson et al. (2021; Fig. 1B) is procedural memory linked to historical relevance in the current model (Fig. 3), emphasizing its shared dependence on statistical information. By this account, stimulus input that historically serves to signal a relevant target is represented in the attentional control state differently than stimulus input that has consistently proven to be irrelevant, as only memory for the former would be expected to contain an associated response (including an orienting response in the context of search). Evidence for both mechanistic overlap and distinct influences on priority in the case of procedural (tied to target selection) and statistical (frequently examined with respect to distractor ignoring) memory abound in the literature (Di Caro & Della Libera, 2021; Ferrante et al., 2018; Li et al., 2023; Theeuwes et al., 2022), which the proposed framework can accommodate. Similarly, memory for relational information may encompass an associated orienting response (e.g., contextual cueing; Brockmole & Henderson, 2006; Peterson & Kramer, 2001; Ramey et al., 2019) or may signal information about what is and is not likely to be contextually relevant, implying a distinct underlying memory system in the control of attention with representational overlap with other memory systems implicated.

In the schematic shown in Fig. 3, current task relevance is depicted without overlap with other components of the attentional control state, as the memory system supporting this component of the attentional control state is distinct. However, it is important to note that the contents of this memory system are entirely flexible and can be configured to prioritize the same stimulus input that is prioritized by any other component (this happens, for example, when the searched-for stimulus is also associated with reward; see Anderson, 2018). Consistent with Hutchinson and Turk-Browne (2012), it is assumed that multiple components of the attentional control state operate in parallel, with the control state reflecting the sum of activated memory and its relationship with perceptual systems. Physical salience is reflected in the nature of the input, as the manner in which it contributes to attentional priority is fundamentally different than the manner in which components of the control state operate; this architecture emphasizes the dynamic nature of the different influences on attentional priority, which tend to be underspecified in models of experience-dependent attentional control (Anderson et al., 2021; Awh et al., 2012).

3.3.7. An important caveat

It is worth reiterating that the proposed theoretical framework for the control of attention affirms mechanistic distinctions that are made under a trichotomy framework (Anderson et al., 2021). The architecture of Fig. 3 reflects this. A core assumption of the proposed framework is that each mechanism of attentional control is supported by a distinct underlying memory system, including working memory, which can be rapidly and flexibly updated in the context of goal-directed attentional control. There are ways in which goal-directed attentional control can be distinguished from the different mechanisms that collectively constitute selection history (e.g., Theeuwes, 2018), and the proposed framework should not be taken to suggest that the influence of task goals on the control of attention is tantamount to an influence of selection history. With that said, the proposed framework rejects the categorical distinction drawn between goal-directed attention and selection history, and conceptualizes goal-directed attentional control simply as one additional control mechanism that is both unique and ultimately memory-dependent. That is, the proposed framework affirms a boundary between goal-directed attention and selection history, just not a categorical one.

4. Implications of the proposed framework

A monolithic framework for the control of attention along the lines proposed in this theoretical review offers a number of concrete advantages over a trichotomy framework beyond the motivating considerations already discussed. These advantages proceed from the parsimony of the proposed framework and its concrete mechanistic specificity, providing further evidence for its utility as a model for conceptualizing the control of attention. In the following sections, these advantages will be expounded upon in turn, which are also summarized in Table 2. How the proposed framework lends new insight into historical controversies in the control of attention that could help resolve points of contention in the literature is also considered.

Table 2.

A contrast of the assumptions of a trichotomy framework and the proposed monolithic framework for the control of attention.

Issue	Assumptions of a Trichotomy Framework	Assumptions of a Monolithic Framework
Dynamics of priority computations	No inherent constraints or guiding principles	Emerges systematically from the nature of the connections between different memory systems and perceptual systems
Mechanistic distinctions	Constrained only by three categories of influence; mostly constructed to explain extant literature	Logically constrained by the number of memory systems contained within the organism and their relationship with perceptual systems; makes specific, falsifiable predictions
Flexibility of attentional control	Not explicitly represented	Naturally reflected in the flexibility with which the contents of activated memories are updated
Relationship among mechanisms of control	One level of distinction; two influences on attentional control either do or do not reflect distinct underlying mechanisms	Two levels of distinction, one with respect to tonic and one with respect to phasic components of control; influences on attentional control can differ with respect to one or both
Attentional control state	An abstract concept that is loosely defined	The contents of memory that result in modulations of activity within perceptual systems
State-dependent effects on attentional priority	Agnostic as to the nature of state-dependent influences	State-dependent influences reflect context effects on memory influencing priority computations
Prioritization of regularities	Unclear link to selection history	The product of memory updating affecting the attentional control state
Working memory-related signals in sensory cortex	Unclear relationship with attentional control	A reflection of the nature of the relationship between memory and attentional control
Involuntary influences of the contents of memory on attentional control	Require more than one ancillary assumption to accommodate	Are straightforwardly accommodated by predictions concerning the nature of the attentional control state
Top-down vs. bottom-up	Selection history is neither top-down nor bottom-up	All forms of attentional control are intrinsically top-down, although they exert their influence at different time scales and levels of representation
Voluntary vs. involuntary	Attentional control can be voluntary (goal-directed) or involuntary (stimulus-driven and experience-driven)	Attentional control is not itself voluntary or involuntary, but rather the memory responsible for attentional bias can be either instantiated volitionally or involuntarily triggered by a stimulus or context
Phylogenetic origins of attentional control	Agnostic as to how different mechanisms of attentional control are phylogenetically related; emphasizes differences between mechanisms; unclear predictions across species	Hypothesizes a single overarching adaptation in the control of attention that cuts across all mechanistic influences on the computation of priority; evolution across species should reflect the evolution of memory systems and their relationship to perceptual systems
Abnormal attentional biases	Agnostic as to the root cause of abnormal attention	Abnormal attentional biases are rooted in how the memories that bias attention are stored and retrieved

4.1. Organizational framework

The proposed monolithic theory for the control of attention offers several advantages over trichotomy theories with respect to the scope and precision of its organizational framework, which are expounded upon below.

4.1.1. The dynamics of priority computations

Theories of attentional control emphasizing a trichotomy framework for the mechanisms underlying the computation of priority offer no guiding principles concerning the dynamics of that computation (e.g., Anderson et al., 2021; Awh et al., 2012; Theeuwes, 2018, 2019). These theories make reference to mechanistic distinctions in the time course of influence on priority computations, for example with reward-dependent influences arising earlier in the visual processing stream than goal-directed influences (e.g., Anderson, 2019; Anderson et al., 2021), but these distinctions are largely descriptive in nature and rooted in empirical observations. That is, these distinctions are drawn in an effort to explain extant data, with limited prescriptive utility. Each hypothetical mechanism of control is essentially unconstrained with respect to the time course of its influence under a trichotomy framework predicated on categories of influence, with any constraints placed upon these dynamics constituting non-essential boundary conditions that can be easily adjusted to accommodate new observations. The nature of the dynamics is not dictated by the model, nor does it otherwise fall out of any specifications of the model.

The monolithic framework proposed here places natural constraints on the dynamics of the computation of attentional priority, which arise from the relationship between the perceptual system in question and the different memory systems implicated. Where the relevant memories are instantiated within the perceptual system, how quickly sensory input reaches those priority-adjusted neurons, and which neurons those priority-adjusted neurons project to will determine where and when memory-dependent biases in priority are realized. By this account, the influence of different memory systems on the computation of priority unfolds progressively, with different influences coming online at different stages reflective of the organization of memory. The proposed account can straightforwardly accommodate learning-dependent shifts in baseline visual cortical activity (e.g., Duncan et al., 2023; Wang et al., 2019), more rapidly emerging influences of selection history supported by subcortical memory systems (e.g., Anderson, 2019; Anderson, Laurent, et al., 2014; Kim, Nanavaty, et al., 2021; Yamamoto et al., 2013), additive influences from different memory systems (e.g., Kim & Anderson, 2021a; Ogden et al., 2023; Stankevic & Geng, 2014), and influences arising from memories instantiated in prefrontal cortex-dependent working memory that are slower to emerge (e.g., Anderson, 2019; Anderson & Kim, 2019a, 2019b; Theeuwes, 2018; Woodman et al., 2013).

4.1.2. A logical constraint on mechanistic distinctions

A significant complexity with regards to trichotomy theories of attentional control concerns the question of how best to characterize dissociations observed within experience-driven influences on priority. One approach, as described in Section 3.2.1, is to designate selection history as an overarching mechanism of attentional control that can be fractionated into multiple distinct underlying mechanisms, maintaining the trichotomy framework (Anderson et al., 2021). The same logic can be applied to distinctions with respect to the putative memory systems supporting goal-directed attentional control (Woodman et al., 2013).

The broader the construct of selection history becomes, however, the less explanatory power it possesses and the less meaningful it is to invoke the concept. And as described in Section 3.2., as the number of distinct mechanisms underlying the computation of attentional priority grows, whether the categorical distinctions delineated by a trichotomy framework constitute the most meaningful places to draw theoretical boundaries can be called into question. If selection history can be so many things, we should strive for a theoretical framework that captures and makes sense of that diversity rather than merely acknowledges it.

The proposed monolithic framework for the control of attention captures an underlying theme that cuts across all mechanisms of attentional control while at the same time conceptualizing and constraining the nature of the diversity of underlying mechanisms. Every memory system that contains representational overlap with a perceptual system will influence priority computations in that perceptual system, reflecting an intrinsic relationship between memory and the computation of attentional priority. Evidence to the contrary would call the proposed monolithic framework into question. At the same time, the mechanistic influences on the computation of attentional priority are bounded by the sum of these memory systems. Every putative mechanism of attentional control needs a corresponding memory system to support it, the absence of which would also call the proposed monolithic framework into question. In these and other ways, the framework for the control of attention that I am proposing is straightforwardly falsifiable, making more specific, more concrete, and more numerous predictions than current trichotomy frameworks.

4.1.3. Conceptualizing attentional flexibility

Although the concept of flexibility is not itself new in the context of selection history, the proposed framework substantively expands our understanding of the underlying mechanisms responsible for the nature of attentional flexibility. “Attentional flexibility” is a term often used to characterize the speed with which attention can shift from prioritizing one stimulus or location to another (e.g., Sali et al., 2015, 2016, 2020). The proposed framework provides a deeper context for such flexibility, with the concept of attention shifts expanding to encompass a shift in the contents of activated memory and the corresponding control state, in addition to any influence of an attentional circuit-breaking mechanism (Corbetta & Shulman, 2002; Corbetta et al., 2008; Dube et al., 2022). Furthermore, in the spirit of attentional flexibility, the proposed framework underscores the breadth with which shifts in the contents of activated memory are consequential and why.

By the proposed framework, the ability to update the contents of memory and the corresponding control state to accommodate changes in context and new learning (and related extinction of old learning) takes on critical importance in understanding attentional control. The flexible control of attention includes rapidly updating the memory representations that bias attention with new learning or changes in context, whereas “stable” attention is exemplified by a persistent influence of prior learning in spite of changing environmental contingencies (Liao & Anderson, 2020) or changes in context (Anderson et al., 2012; Mine & Saiki, 2015, 2018; see also Krebs et al., 2010, 2011; Liao et al., 2020). Either of these states of attentional control may be adaptive or maladaptive based on the demands of the current situation, much as is the case in the context of broader conceptualizations of cognitive and attentional flexibility (Cools, 2008). In this regard, it may be the case that optimal attentional control involves something between these two extremes (see Anderson, 2021a), with prior learning having an effect that persists beyond exposure to the contingencies responsible for it while allowing for the current instantiation of priority to adjust to new learning (Liao & Anderson, 2020) and readily generalize to new contexts (Anderson et al., 2012; Mine & Saiki, 2015, 2018) without compromising sensitivity to context-dependent learning (Anderson, 2015a, 2015b; Gregoire et al., 2021).

Trichotomy frameworks do not explicitly address the concept of attentional flexibility, which is in no way intrinsic to a trichotomy model. The proposed framework not only accommodates the notion of attentional flexibility, but it enriches the theoretical basis for the mechanisms underlying shifts of attention (see Shomstein & Yantis, 2004; Yantis et al., 2002). In this way, the proposed monolithic framework has broader explanatory power.

4.1.4. Contextualizing the relationship among mechanisms of control

The proposed framework draws a distinction between tonic and phasic influences on attentional priority (Table 1 and Fig. 4). Tonic attentional control signals reflect the state of activated memory and phasic attentional control signals reflect the application of bias in information processing. It is possible for two mechanisms of attentional control to differ with respect to tonic attentional control signals while producing the same phasic attentional control signals in a particular situation, given overlap in the complexion of activated memory and its relationship with perceptual systems. The nature of the relationships involved makes it possible to observe dissociations or competitive interactions between mechanisms of attentional control under certain task conditions while observing evidence in favor of representational overlap and a common mechanism of attentional control under different task conditions, depending on the stage of information processing and/or the component of the attentional control state probed in the experiment.

This distinction between tonic and phasic attentional control signals provides an immediate resolution to apparently conflicting findings concerning the relationship between the manner in which reward and punishment influence attention. On the one hand, previously rewarded and aversively conditioned stimuli produce similar behavioral indicators of attentional bias (Gregoire et al., 2022; Kim & Anderson, 2019a, 2021a, 2021b, 2023a; see Anderson et al., 2021, for a review), even when care is taken to match the subjective valance of the associated outcomes (Kim & Anderson, 2023b), and orienting to such stimuli is associated with indistinguishable neural correlates reflective of phasic attentional control (Kim, Nanavaty, et al., 2021). How reward history and statistical learning combine to influence attentional priority is also comparable to how aversive conditioning and statistical learning combine to influence attention (Kim & Anderson, 2021b; Ogden et al., 2023). This would seem to suggest a common underlying mechanism for priority assignment (see Anderson et al., 2021). A common underlying mechanism, however, would not seem to predict a competitive relationship between anxiety and value-driven attention in which value-driven attention is blunted by an anxious state (Kim & Anderson, 2020a, 2020c), as anxiety is well-documented to potentiate attention to aversively conditioned stimuli (Bar-Haim et al., 2007; Gregoire & Anderson, in press). A common underlying mechanism also belies the intuition that the experience of reward and punishment evoke distinct neural circuits supporting approach and avoidance responses, respectively (Anderson, 2017a; Anderson, Folk, et al., 2016; Chen & Bargh, 1999; Guitart-Masip et al., 2012; Kim & Anderson, 2019b; Krieglmeyer et al., 2010, 2013; van Wouwe et al., 2015). By the proposed framework, the phasic influence of reward and punishment history on attention reflects a shared mechanism of modulating visual information processing, which receives its input from overlapping memory systems encoding unsigned valence or survival-relevance. That is, the attentional control state for reward-dependent and aversively conditioned attentional bias contains representational overlap in the context of perceptual systems in which phasic attentional control signals are similarly computed. But these two attentional control states each contain unique components to the underlying memory representation as well, which is reflected in the nature of the overlap depicted in Fig. 3. With this in mind, distinct state-dependent influences on attention reflect reward- or punishment-specific modulation of tonic attentional control signals arising within components of the activated memory representation specific to the associated outcome (the area of non-overlap in Fig. 3), with such modulation having a broader influence on the totality of the corresponding attentional control state (including the area of overlap depicted in Fig. 3) given that any given component of the control state reflects a representation that is bound in memory.

The distinct influence of tonic versus phasic attentional control signals makes it possible for other mechanisms of attentional control to simultaneously exhibit similarities and differences depending on the level of analysis, for example with respect to how statistics concerning the target and distractor influence attention (Di Caro & Della Libera, 2021; Ferrante et al., 2018; Li et al., 2023; Theeuwes et al., 2022). By the proposed framework, the level at which attentional priority computations are being influenced by an experimental manipulation needs to be considered when drawing conclusions. It is not the case that identifying a dissociation between two influences on attentional priority implies that these two influences are supported by categorically distinct mechanisms of control, nor does observing a dependence between two such influences or a lack of apparent distinction in the underlying attentional dynamics imply an entirely shared mechanism of control. In this respect, a single diagram with boxes meant to depict the relationship among mechanisms of attentional control, however complex, is likely to reflect something of an oversimplification, as it only captures one level of relationship.

4.1.5. Attentional control state defined

The concept of an attentional control state is not new and indeed well-precedented in theories of attentional control (Folk et al., 1992), but concrete neurobiological definitions of the control state have been lacking. The proposed framework makes straightforward predictions concerning how the control state is instantiated and how it operates. It even lends itself to straightforward predictions concerning where and when the control state can be observed in the brain. The conceptualization of a control state here offers unique insights into the nature of proactive attentional control (see Geng, 2014; Liesefeld et al., in press) by articulating the concept of tonic influences in the control of attention, which have a naturally proactive component in the computation of attentional priority. With respect to the attentional control state, the proposed framework promotes the generation of predictions that are more concrete and falsifiable than previous theories of attentional control, and some of these predictions make natural connections with existing theoretical distinctions concerning the time course of attentional control signals.

4.2. Explanatory power

The proposed monolithic theory of attentional control offers a straightforward and parsimonious explanation for several phenomena that trichotomy theories struggle to accommodate, which are expounded upon here.

4.2.1. A simple framework for understanding state-dependent effects on priority

State-dependent effects on attentional bias are well-documented. For example, negatively valenced stimuli capture attention more robustly when observers are in an anxious or dysphoric state (e.g., Bar-Haim et al., 2007; Gregoire & Anderson, in press; Koster et al., 2005; Mathews & MacLeod, 2005), positively-valenced stimuli are more robustly attended when in a positive mood (e.g., Wadlinger & Isaacowitz, 2006), and attentional biases for calorie-dense foods are reduced when satiated (e.g., Cunningham & Egeth, 2018; see also Pool et al., 2014). Value-driven attentional capture is blunted in an anxious state (Kim & Anderson, 2020a, 2020c), and state anxiety is associated with improved strategic attentional control under certain task conditions (e.g., Kim, Lee, et al., 2021). Theoretical models of attentional control centered on a mechanistic trichotomy offer no explanation for such influences, but they can be straightforwardly accounted for under the proposed framework.

Context-dependent effects on memory have been noted for decades (e.g., Godden & Baddeley, 1975; Smith & Vela, 2001), and an internal state such as anxiety can serve as a context (Bower et al., 1978; Buchanan, 2007), along with the environmental influences that play a role in evoking that state. To the degree that the strength and complexion of memory activation varies by context, including with respect to internal state, the strength of attentional bias resulting from the memories and corresponding attentional control state in question will be proportionally affected. That is, fear-related images are more strongly attended when anxious because anxiety intensifies fear-related memories. Reward-related stimuli capture attention less robustly when anxious because anxiety shifts the contents of memory from appetitive to aversive or otherwise negative information (which are situationally more pertinent). Anxiety is associated with elevated arousal, which can at certain levels boost cognitive performance (Lambourne & Tomporowski, 2010), thereby improving the strategic control of attention. In this way, state-dependent effects on attentional bias follow quite naturally from the proposed framework, which provides a mechanistic explanation for these effects.

4.2.2. Attentional biases for statistical regularities

Attention has been shown to be biased toward portions of a display that contain statistical regularities, even when those regularities are completely irrelevant to the ongoing task (Zhao et al., 2013). Although such a bias can be straightforwardly characterized as falling within the domain of selection history (Failing & Theeuwes, 2018), it is unclear why attention would be biased in this way. Existing models of selection history emphasize a role for statistical learning (see Fig. 1B), but this is typically characterized with respect to task-relevance and/or features of the task pertinent to search guidance (see Anderson et al., 2021; Theeuwes, 2018, 2019; Theeuwes et al., 2022). That is, attentional biases attributable to statistical learning are generally learned in the context of search presumably with the function of facilitating search, hence the term selection history. Although one can argue that attention to regularities reflects a bias towards information-seeking (see Gottlieb et al., 2013), trichotomy theories of attentional control do not offer a mechanistic explanation for this bias.

According to the proposed account, there is a necessary link between activated memory and the attentional control state. As long as an activated memory shares representational overlap with a perceptual system, some consequence for the computation of attentional priority should be expected. Therefore, as observers learn regularities and the associated memory for those regularities updates, this memory updating will be reflected in the attentional control state. Attention is intrinsically biased to prioritize input consistent with the contents of memory, even if such memory is implicit in nature and unrelated to the performance of the current task. By the proposed account, the attentional control state does not discriminate between memory that is and is not pertinent to current task goals and/or the quality of expected outcomes.

4.2.3. Accounting for the neural correlates of working memory

The contents of working memory can be decoded from neural activity in the primary visual cortex during a delay period (Harrison & Tong, 2009; Serences et al., 2009). Trichotomy theories of attentional control make no predictions about the nature of working memory representations specifically or the nature of memory representations more broadly. The proposed monolithic framework for attentional control, however, predicts this exact type of phenomenology. The kind of activated memory representations that contribute to the attentional control state will be reflected in tonic neural activity in perceptual systems. Such memory-related neural activity should be evident regardless of whether visual search is required in the current task context, as the relationship between memory and the attentional control state is an intrinsic property of the organization of the brain (see Section 3.3.) that does not depend on the intent to allocate attention.

4.2.4. Accounting for involuntary memory-related attentional biases

Visual attention has been shown to be biased by the contents of working memory, such that stimuli possessing a feature maintained in working memory are more robustly attended (e.g., Olivers et al., 2006, 2011). Trichotomy theories of attentional control do not explicitly predict this kind of relationship, although it is not difficult to accommodate by assuming that the influence of task goals and working memory contents on the control of attention are highly related. The finding that the contents of context-dependent long-term memory can bias attention without such memory being related to prior episodes of attentional control (Nickel et al., 2020), or selection history in the strict sense of the term (see Awh et al., 2012), is more difficult for trichotomy theories to accommodate, and it bears no obvious relationship with the aforementioned observations concerning the relationship between the contents of working memory and attentional control. By the proposed monolithic account of attentional control, each of these observations has an intuitive explanation in the relationship between the contents of activated memory and the attentional control state, and no ancillary assumptions need to be made to accommodate such phenomenology.

4.3. Broader implications

The proposed monolithic theory for the control of attention has implications for how we think about broader issues with respect to the nature of attentional control, which are explored here.

4.3.1. Rethinking the distinction between top-down and bottom-up

Awh et al. (2012) suggested that the distinction between bottom-up and top-down attentional control was a failed theoretical dichotomy, arguing that the influence of selection history on attention was neither. In the years that followed, debate and confusion concerning the categorization of selection history ensued, with some likening selection history to top-down influences (e.g., Gaspelin & Luck, 2018b) and others characterizing it as more akin to a bottom-up influence (e.g., Hickey et al., 2010; Hickey & van Zoest, 2012). The top-down argument rested on the idea that the attentional bias in question was dependent upon a representation internal to the observer and thus could not be, strictly speaking, stimulus-driven, while the bottom-up argument was bolstered by findings that effects of selection history could arise in very early stages of perceptual processing, with a similar timescale to effects of physical salience on eye movements (e.g., Bucker et al., 2015; Kim & Anderson, 2022; Mulckhuyse et al., 2013; Pearson et al., 2016; Schmidt et al., 2017). Similarly, experience-driven attentional biases were evident very early in visual cortical information processing (e.g., Itthipuripat et al., 2019; MacLean & Giesbrecht, 2015; Serences, 2008; Serences & Saproo, 2010; van Koningsbruggen et al., 2016; see Anderson, 2019, for a review).

With its emphasis on memory-dependent biasing signals, the proposed framework asserts that all forms of attentional control are fundamentally top-down. That is, all attentional bias signals are ultimately generated by signals coming from outside of the perceptual system in question, coming online in advance of sensory information processing, that then influence this system. In this way, all phasic attentional control signals reflect a form of feedback triggered by the activation of a memory. By this account, bottom-up influences on attentional priority are reserved for physical salience alone, the influence of which does not constitute attentional control (see Section 3.2.2.3). Furthermore, goal-directed attention is no more top-down than selection history-driven attention, they just arise at different stages of information processing via the same principles.

Differences in the time course of the emergence of biased information processing in the attentional system that have been used to imply a more “bottom-up” influence in the case of selection history can be explained by the positioning of the relevant memory system within the brain circuit in play. The same can be said for the ability of experience-driven influences to at times overpower goal-directed control. Goal-directed adjustments in attentional priority can be relatively slow to come online because they involve feedback within a prefrontal cortex-dependent memory system quite late in the stream of information processing, whereas the striatal memory systems responsible for certain experience-driven attentional biases influence visual input more rapidly (see Anderson, 2019). Long-term memory influences, including many cases of experience-driven attention, can involve biasing signals that are applied to sensory areas in advance of stimulus input, allowing for rapid and phenotypically “bottom-up” prioritization to emerge early and quickly.

This latter point comes with a caveat, in that it can be in some ways productive to relate effects of selection history on attention more closely to stimulus-driven than goal-dependent effects, particularly when making predictions concerning the dynamics of their unfolding influence on information processing. The assertion that all forms of attentional control are fundamentally top-down must be understood in the context of the considerable heterogeneity of the control mechanisms at play. In some respects, this reiterates the incisive point made by Gaspelin and Luck (2018b) that a “top-down” effect is not tantamount to a “voluntary” effect, which is reflected in how I am choosing to use these terms in the context of this theoretical review (cf., Theeuwes, 2018).

4.3.2. On the distinction between voluntary and involuntary attentional control

According to the proposed framework, the distinction between voluntary and involuntary attentional control is really just a distinction with respect to the nature of the underlying memory system involved and whether the memory that modulates attentional priority is brought online in an endogenous fashion. Voluntary attentional control reflects the influence of working memory and volitionally activated long-term memory, whereas “involuntary” influences on attentional priority reflect bias signals from memories that are triggered by eliciting stimuli and/or brought online without explicit intention by the current task context. Mechanisms of attentional control are not themselves voluntary or involuntary, as attested to by the apparently involuntary influence of the contents of working memory on concurrent visual information processing (e.g., Olivers et al., 2006, 2011). Although asking whether attentional priority is involuntarily allocated in a particular context is not an unproductive question, as the answer does say something about the memory system implicated and the manner in which it is implicated, the framing of the question misses the mark because volition is a product of the memory system of the brain, not the attention system.

4.3.3. Phylogenetic origins of attentional control

A straightforward hypothesis arising from the proposed framework is that selective attention emerged as a product of connections and representational overlap that developed between the memory system and perceptual systems. By this framework, there is a single overarching motif within the interconnectedness of the brain that explains the entire range of phenomenology subsumed within the concept of attentional control—a single biological solution to a complex problem. That is, goal-directed attention and the different components of experience-driven attention reflect the same core adaptation applied to different memory systems.

Non-human primates (e.g., Gottlieb et al., 1998; Peck et al., 2009; Peck & Salzman, 2014; Yamamoto et al., 2013) and even birds (Liao, Dillard, et al., 2023; Quest et al., 2022; Sridharan et al., 2014; Winkowski & Knudsen, 2007, 2008) exhibit both goal-directed attention and value-driven attention. Although it is likely that different mechanisms of attentional control arose at different times in evolutionary history, a natural prediction arising from the proposed framework is that each underlying mechanism of attentional control should emerge in close proximity to the emergence of the underlying memory system, with a strong covariance between the evolution of memory and the evolution of attention. By the proposed framework, attentional control is phylogenetically old and attentional control subsumed within the neocortex (e.g., Corbetta & Shulman, 2002; Corbetta et al., 2008; Vossel et al., 2014) reflects a biological solution precedented in more primitive memory systems. In this way, attention reflects a core system of brain organization that can be scaled up to encompass new mechanisms of assigning priority in parallel with an increasingly complex memory system.

4.3.4. The underpinnings of abnormal attentional biases

According to the proposed framework, there are three principal ways in which the control of attention could be abnormal. The first is with respect to the organization of a particular memory system and its connections with perceptual systems, and the second is with respect to computations within priority maps and the actual representation of attentional priority. The latter accounts for gross deficits in the control of attention per se, such as unilateral spatial neglect (Behrmann et al., 2004), and the former would likely imply a broader memory impairment that has an attentional consequence. The third principal way in which the control of attention could be abnormal, which likely accounts for the vast majority of abnormal attentional biases observed in the literature, is with respect to how memories are formed and stored.

Anxiety disorders are often characterized by abnormally strong attentional biases towards fear-related stimuli (see Bar-Haim et al., 2007), and depression by heightened attentional priority afforded to negatively-valenced stimuli (see Mathews & MacLeod, 2005) along with blunted attentional priority for positively-valenced stimuli (e.g., Anderson, Leal, et al., 2014; Anderson et al., 2017). Drug dependence is associated with pronounced attentional biases for drug-related stimuli (see Field & Cox, 2008) and reward-related stimuli more generally (e.g., Albertella et al., 2017; Albertella, Le Pelley, et al., 2019; Albertella, Watson, et al., 2019; Anderson, 2016; Anderson et al., 2013; Anderson, Kronemer, et al., 2016). By the proposed framework, these attentional biases can be straightforwardly explained by either the formulation of abnormally strong or weak memories that are then evoked by the eliciting stimuli in question, or abnormal tendencies with respect to how broadly and robustly such memories are evoked by different contexts people find themselves in (e.g., overgeneralization). The bidirectional relationship between learning (memory) and attention has long posed a chicken-and-egg problem with respect to abnormal attention (see Anderson, 2016, 2021b), but the proposed framework suggests that the origin would naturally fall on the side of memory, as the contents of the memories brought to bear in a particular situation determine the complexion of attentional bias via the corresponding attentional control state.

This is not to say that abnormal attentional biases do not influence the encoding of information into memory, as attention is widely understood to play a prominent role in determining the information that is available to be encoded (Awh et al., 2006; although see Chen & Wyble, 2015a, 2015b; Yan & Anderson, in press; Yan et al., 2023). In this respect, there could be a vicious cycle in which, say, abnormally strong memories for fear-related stimuli cause fear-related stimuli to be more strongly prioritized by the attentional system, which in turn results in stronger encoding of fear-related stimuli into memory, accentuating attentional biases. However, every such cycle must start somewhere. It is difficult to envision the cycle starting with biased attention under the proposed framework, while a bias that begins in memory storage and then translates to attentional processing is intuitive.

5. Some unresolved issues

Every theory has its limitations, and the proposed theory of attentional control is no exception. The following section will explore three issues, both with respect to empirical observations and philosophical considerations, that the proposed theoretical framework cannot straightforwardly accommodate, as well as highlight an underspecified element of the framework. Potential means of incorporating relevant considerations into the framework are briefly explored, although future research will be needed to provide a more satisfying reconciliation.

5.1. Novelty-driven orienting

The proposed framework is centered on the role of activated memory representations in providing the biasing signal responsible for modulations of attentional priority. It is well established that novel stimuli evoke elevated attentional priority (e.g., Horstmann & Ansorge, 2006, 2016; Horstmann & Herwig, 2016; Horstmann, 2002; Johnston et al., 1990, 1993; Johnston & Schwarting, 1997), and it is not immediately clear how this phenomenon can be accounted for under the proposed framework, as novelty reflects a discrepancy between memory and perceptual input. Perhaps novelty-driven orienting has something to do with mechanisms for generating and identifying the source of perceptual prediction errors (Rao & Ballard, 1999; see also Bunzeck & Duzel, 2006; Guitart-Masip et al., 2010; Horvitz, 2000; Wittmann et al., 2008; Zink et al., 2003), which are fundamentally memory-dependent, and/or updating the memory representations supporting the influence of statistical learning on attention (Theeuwes et al., 2022). As with the influence of statistical regularities per se on the allocation of attentional priority (see Section 4.2.2), an updating of memory that contains representational overlap with perceptual systems would be expected to bear an influence on the attentional control state. From eye movement studies, effects of novelty on attentional priority computations tend to arise somewhat later in information processing than other experience-dependent influences such as reward associations and target history (e.g., Horstmann & Ansorge, 2006, 2016; Horstmann & Herwig, 2016; Horstmann, 2002), so novelty-driven orienting may involve feedback with respect to the detection of input that deviates from what would be expected in the context of currently activated memory representations. With that said, trichotomy frameworks do not offer a mechanistic explanation for this phenomenon either (Anderson et al., 2021), so as Johnston and Schwarting (1997) asserted more than twenty-five years ago, novelty-driven orienting remains something of an enigma.

5.2. Non-target rejection

The proposed framework allows for priority adjustments that both upweight and downweight stimuli, with the latter facilitating non-target rejection. However, it is a framework for the assignment of attentional priory, and sometimes the greatest priority will be assigned to a stimulus that is not what the individual is searching for. The proposed framework does not explicitly address how individuals reject a currently prioritized stimulus as something not of interest and subsequently de-prioritize that stimulus to better allow other stimuli to compete for representation. Such de-prioritization would likely require a dedicated attentional process, particularly in the situation in which a working memory representation is being used to facilitate processing of a searched-for target; in this case, the memory representation responsible for prioritizing a non-target cannot itself be suppressed, as it is needed to guide continued search. Within the proposed framework, non-target rejection may involve some combination of a circuit-breaking mechanism in the control of attention (Corbetta & Shulman, 2002; Corbetta et al., 2008; Dube et al., 2022), inhibition of return (Klien, 2000; Lupianez et al., 2006), and potentially an update to the contents of activated memory to exclude the rejected stimulus.

5.3. Context-invariant influences of selection history

On the surface, the proposed framework would seem to assume that all experience-driven influences on the control of attention are context-dependent at least in situations in which context is diagnostic of task contingencies and what should be prioritized (see Anderson, 2015a, 2015b; Gregoire et al., 2021). In situations in which this is not the case, we risk maladaptive experience-driven attentional priorities that are unhelpful across many situations, chronically misguiding attention. The breadth of memories brought to bear in the control of attention must also be to some degree situationally constrained, lest attentional prioritization ultimately prove indiscriminate and be assumed to rely on a memory system of unrealistically unbounded representational capacity. In part for these reasons, it was a surprising observation that, in contrast to the influence of reward learning and aversive conditioning (Anderson, 2015a, 2015b; Gregoire et al., 2021), statistical learning can have an apparently context-insensitive influence on the allocation of attentional priority (Anderson & Britton, 2019; Britton & Anderson, 2020; de Waard et al., 2022). This is not a unique challenge for the proposed framework, however, it is arguably a challenge for any grounded theory of attentional control. Potential explanations include the idea that some measure of context-dependence is present in these situations that only stronger manipulations of context tend to reveal (de Waard et al., 2023), and it is possible that effects of statistical learning on attention, while exhibiting limited contextual dependence, also rely on memory representations that specifically reflect recent experience and are relatively short-lived (see Di Caro & Della Libera, 2021).

5.4. Beyond modulations of priority

The proposed framework is principally concerned with the assignment of attentional priority, which is dynamic and unfolds in parallel across multiple objects and locations. Later stages of information processing, such as memory storage, decision making, and action selection, are highly selective (e.g., Serences & Yantis, 2006), which reflects the at times highly selective nature of information processing and the failure of unattended information to reach perceptual awareness (Kamitani & Tong, 2005; Mack & Rock, 1998; Rensink et al., 1997). As described in Section 3.3.5, the proposed framework does not explicitly address how variation in resulting attentional priority, following priority adjustments brought about by the attentional control state, is translated to something of a winner-take-all outcome. The transition from differential prioritization to the selective representing of a single source of perceptual input could be achieved through mechanisms of biased competition (Desimone & Duncan, 1995; Serences & Yantis, 2006) at different stages of representation, as well as the strength of the priority signal triggering the kind of selective enhancement described as an attentional episode by Zivony and Eimer (2022). As Zivony and Eimer’s (2022) recent review paper illustrates, translating priority assignment to eventual highly selective information processing reflects an important area of continued development for theories of attention more broadly.

6. Conclusions

The diversity of mechanisms for computing attentional priority are well accounted for by a theoretical framework that emphasizes what they share in common, which is a core process by which activated memory representations alter the strength of perceptual signals evoked in sensory organs. Such a monolithic framework avoids categorical distinctions between goal-directed and experience-driven influences on priority that are stronger than the complexion of these diverse mechanisms warrants, and it provides greater clarity concerning the relationship between mechanisms of attentional control and physical salience. The proposed framework is considerably more mechanistically precise than trichotomy frameworks, which describe the computation of attentional priority at a more conceptual level. Its parsimony and concrete mechanistic specificity offer multiple distinct advantages over existing trichotomy frameworks while providing new insight into historical controversies and points of contention in the literature (see Table 2).

This is not to suggest that the distinction between goal-directed influences and selection history in the control of attention is not a valid one. In discussing mechanisms of attentional control, this is a useful distinction to make. I fully intend to continue making the distinction myself in certain contexts. I have come to the position, however, that a trichotomy is no longer the most productive place for us to anchor our thinking on attentional control as an organizational framework. It should not be the starting point for a dialogue on how attentional control works or what exactly constitutes attentional control. I think the most progress is going to be made in scientific thinking on the nature of attentional control when we anchor our thinking to the common thread that provides concrete underlying structure for the computation of attentional priority, broadly construed. We need a framework that can be used to generate novel predictions and can explain a wide range of phenomena in the control of attention beyond a description of the different factors that contribute to the computation of priority, and in this respect the proposed monolithic framework provides a productive starting point for the next chapter for research on attentional control.

Data availability

No data was used for the research described in the article.