Insights into control over cognitive flexibility from studies of task-switching

Abstract

Cognitive flexibility denotes the ability to disengage from a current task and shift one’s focus to a different activity. An individual’s level of flexibility is not fixed; rather, people adapt their readiness to switch tasks to changing circumstances. We here review recent studies in the task-switching literature that have produced new insights into the contextual factors that drive this adaptation of flexibility, as well as proposals regarding the underlying cognitive mechanisms and learning processes. A fast-growing literature suggests that there are several different means of learning the need for, and implementing, changes in one’s level of flexibility. These, in turn, have distinct consequences for the degree to which adjustments in cognitive flexibility are transferrable to new stimuli and tasks.

‘Cognitive flexibility’ describes the ability to shift from a current task to a different one deemed to be of higher priority, such as switching from writing an email to answering an incoming phone call. This process can be conceptualized as replacing one task set – a set of rules linking stimuli to actions [1] – with another one in working memory, a temporary mental workspace for information that guides how we attend and respond to our surroundings [2,3]. Switching tasks incurs a cognitive cost (the ‘switch cost’), as reflected in slower and less accurate responses when switching from one task to another, compared to repeating the same task [4,5] and the size of the switch cost is commonly treated as an (inverse) measure of cognitive flexibility, with a smaller switch cost indicating greater flexibility [6,7].

What makes people more or less flexible? At the inter-individual level, a number of clinical conditions, including attention deficit disorder [8], autism [9], and Parkinson’s disease [10], are associated with reduced flexibility (greater switch costs) [11••], and trait differences can also be observed in the general population [12]. Within an individual’s lifetime, cognitive flexibility appears to increase from childhood to early adulthood [13,14], and decrease in older age [4,15, but see 16]. Importantly, flexibility also varies at much shorter time-scales, from moment to moment, as a function of contextual factors [6,7]. The central observation in this emerging field of study is that the healthy brain seems to dynamically adapt its’ level of flexibility to suit changing environmental demands, a form of ‘meta flexibility’. The goal of the current paper is to provide a brief review of key recent findings and future challenges in this field, asking what makes people more or less flexible in different situations. Answering this question is crucial for gaining new insights into ‘meta control’ [17–19] – how the control processes involved in switching from one task to another are strategically regulated – as well as into possible causes of deficits in cognitive flexibility [11••].

We begin by describing the key behavioral phenomena in this literature, including boundary conditions, followed by a discussion of how these findings could be explained at the level of cognitive mechanisms and learning processes. We close by noting important gaps in our current understanding.

Behavioral signatures of adapting cognitive flexibility

Imagine performing the following experiment involving two task sets. On each trial, you are presented with one digit (from 1–9) and one letter, and depending on a task cue, you have to either classify the digit as being odd or even, or classify the letter as being a vowel or a consonant (Fig. 1A). Many studies have documented that it takes longer to respond correctly if the task on the current trial were different from the one on the previous trial than if they were the same [1]. Importantly, several factors have been identified that can increase or decrease this cost of switching task sets, in particular, manipulations of switch rate and reward.

Figure 1.

(A) A typical cued task switching protocol involving a digit task (“Is the digit smaller or larger than 5?”) and a letter task (“Is the letter a vowel or consonant?”), indicated by a preceding task cue. (B) Mean response time (RT ± standard error) for task repetition and task switch trials displayed as a function of the block-wide switch rate. Modified from Siqi-Liu & Egner (2020), Experiment 2.

First, as exemplified in Fig. 1B, the contextual rate or likelihood of having to switch tasks is a potent modulator of switch costs: people incur lower average switch costs in blocks of trials with a higher compared to a lower rate of task switches (Fig. 2A) [20–28]. Notably, people do not only adapt their flexibility to time-varying switch demands but also in relation to specific stimuli or items that are predictive of switch rate. For instance, if one were to present one subset of the digits and letters in the above task more frequently on task-switch trials, and another subset more frequently on task-repeat trials, the former would incur a smaller switch cost than the latter [24,29,30], even when the overall switch rate in the block of trials is neutral (50%) (Fig. 2B). Finally, manipulating the rate of cued task switches, either at the block or item level, also affects people’s willingness to switch task: a higher switch rate on trials where participants are told which task to perform promotes a greater willingness to switch tasks on voluntary task selection trials, where participants can freely choose which task to perform [31,32].

Figure 2.

(A) A typical blocked (or “listwide”) switch rate manipulation involves comparing switch costs between blocks of trials with a high vs. a low switch rate. Notably, each task (Tasks A and B) assumes the biased switch rate of each block type. (B) A typical item-specific switch rate manipulation involves comparing switch costs between specific stimuli that occur more frequently on task switches (Stim A and B) or repetitions (Stim C and D), within the context of a neutral (50/50) block-wide switch rate. (C) A design for probing whether flexibility changes due to block-wide switch rate manipulations (driven by Tasks A and B) transfer to a task that itself is not associated with block-wide biases (Task C).

In addition to these switch rate effects, contextual adjustments in cognitive flexibility can also be observed in relation to manipulations of reward. First, if participants are selectively rewarded for task-switching, either by a gain in money [33] or a reduction in cognitive effort [34] (task difficulty), they become faster [34] and more willing to switch task sets [33,34]. Second, a series of recent studies has shown that trial-to-trial changes, and especially increases, in reward prospect result in smaller switch costs [35,36•] and promote a greater willingness to switch tasks in voluntary task selection trials [36•–39], when compared to conditions of constant low or high reward prospects.

Taken together, these data suggest that people learn about the contextual likelihood of switching, and adjust their level of cognitive flexibility accordingly, and that an increase in reward anticipation triggers a greater readiness and willingness to change tasks. Before considering possible mechanisms mediating these adjustments in flexibility, we consider findings speaking to their scope or generalizability.

Boundary conditions of cognitive flexibility in task-switching

Does greater cognitive flexibility – as induced by the above manipulations – make people more ready to switch in general or are these effects expressed more narrowly? This question of the transferability of switch-readiness has been addressed both at the level of stimuli and task sets.

Two recent studies assessed transfer of flexibility across task sets by varying the overall rate of switching between blocks of trials via two “biased” tasks, which rarely required switching in low switch rate blocks, and frequently required switching in high switch rate blocks. All blocks were interspersed with a third, “unbiased” task that occurred equally often on task repetition and task switches in both block types [22,40•] (Fig. 2C). Thus, the authors could probe whether the expected adaptation in cognitive flexibility – smaller switch costs in blocks with higher switch rates – would transfer to the unbiased task. The answer was ‘no’; while performance on the biased task sets displayed typical adaptation effects, the unbiased task was unaffected by switch-rate context. However, adaptations in switch costs were observed for unbiased [22] and even completely novel stimuli [40•] within the switch-biased tasks sets. Equivalent findings have been obtained in a study of voluntary switching: the increased willingness to switch tasks induced by a higher rate of cued task switches transferred to new stimuli within the practiced task sets, but not to new tasks [41].

By contrast, a recent study of trial-by-trial changes in reward prospect produced effects on flexibility that were independent of the specific upcoming task [36•]. The authors employed three task sets occurring with equal frequency, presented with an orthogonal cue of whether the current trial would lead to a low or high reward. Switch costs were reduced for trials with increasing reward prospect. Given that participants could not predict the forthcoming task, the authors interpreted this finding as a generic increase in switch-readiness in response to an increase in reward prospect [36•]. Support for transfer of cognitive flexibility also comes from another study where participants had to infer from probabilistic feedback which out of several possible task rules was currently valid [42]. During a learning phase, the frequency of task rule changes varied between groups of participants (low vs. high); during a subsequent transfer phase, all participants were exposed to a medium switch-frequency condition. People adapted their learning rates (the likelihood of switching tasks in response to error feedback) to the statistics of the learning environment, with a higher learning rate in the higher switch-rate condition. Crucially, those differences in flexibility were maintained in the transfer phase, even when it employed different stimuli and different task sets than the learning phase [42].

Taken together, these studies suggest that task sets form the boundary of flexibility adaptation when the latter is being induced by manipulating the rate of cued task switches: participants become more efficient and willing to switch to tasks that are associated with high switch rates in the current context, and this effect generalizes to novel stimuli within those task sets, but not to other tasks [22,40•,41]. This apparent limitation in the generalizability of cognitive flexibility is also evident in training studies of task switching [43,44]. However, when cognitive flexibility is modulated by changes in reward prospects [36•] or via trial-and-error learning of task rules [42], transferrable adaptations in cognitive flexibility can be observed. These divergent findings suggest that different mechanisms can lead to learned changes in cognitive flexibility, and we discuss those mechanisms next.

Possible mechanisms of control over cognitive flexibility

One proposed mechanism for adapting flexibility to changing contexts is an ‘updating threshold’ control parameter [19] for gating new information into working memory [45]. Environmental cues that signal a need to maintain the current task increase this threshold (making updating working memory content harder), while cues signaling that other goals might be more rewarding lower the threshold (making switching a new task set into working memory easier). This provides a plausible account for generalizable effects of adaptation in cognitive flexibility, as observed in the above-mentioned studies on changing reward prospects [36•] and trial-and-error learning of task rules [42]: a lowered updating threshold should facilitate switching to any other task. However, this mechanism would not predict the type of task-selective flexibility effects observed in studies manipulating the rate of cued task switches [22,40•,41]. Two alternative mechanisms have been suggested to account for these findings.

First, since flexibility adaptation effects in cued task-switching studies have typically been demonstrated in contexts of two task sets only, it has been suggested that in blocks with high switch rates participants may simply be keeping both task sets in working memory, thus presumably reducing the time it takes to switch between them [7]. This idea naturally accounts for the fact that the reduced switch costs between these two task sets would not generalize to a different task set that is not also currently held in working memory. A second possibility is grounded in the idea that specific cognitive control parameter settings can become integrated with particular stimuli and task sets via associative learning [6,46••]. Here, a frequently switched-to task would become associated with cognitive control settings that ease switching to that particular set, leading to a set-specific reduction in switch cost and no transfer effects to other task sets.

Three key points can be taken from this discussion of mechanism. First, the different proposed means of adapting flexibility are not mutually exclusive [7]. Second, a systematic empirical evaluation of the proposed mechanisms remains an important aim for future studies. Third, a normative account for why reward-based studies have produced transfer effects while switch-rate manipulations have not is lacking. One promising source for such an account is the literature on how reward affects cost-benefit calculations for investing control [47]. For instance, higher average reward rates may push people to maximize the benefits, as opposed to minimizing the costs, of control [48]. Finally, contrasting findings of transferrable versus non-transferrable flexibility settings obtained across different experimental contexts suggest that we need to understand the exact learning processes that guide adjustments in flexibility to predict whether the latter are generalizable or not. These are discussed next.

Diverse forms of learning drive adjustments in cognitive flexibility

The observation that adaptation of cognitive flexibility can be tied to both blocks of trials and to individual stimuli is reminiscent of findings in the conflict-control literature, where complementary learning mechanisms for blockwide and item-level effects have been put forward, one that is proactive and guided by recent experience, and another one that is reactive and selective to specific stimuli [46••,49]. Specifically, blockwide effects could be mediated by an incremental reinforcement learning mechanism that nudges levels of flexibility up or down in line with a running average of recent switch demands [42,46••]. This type of mechanism can account for flexibility effects transferring to unbiased or novel stimuli (and tasks [42]) within a high switch rate context [22,40•], but it cannot explain item-based effects that occur in contexts where task switches are no more likely than task repetitions [24,29].

Explaining item-level flexibility adjustments requires a mechanism that learns about item-specific switch demand. This could either take the form of parallel incremental learning of item-switch associations (one learner per item), or of an episodic memory mechanism that reinstates the cognitive control settings associated with previous encounters with a specific stimulus [46••]. In support of the latter possibility, recent work has demonstrated ‘one-shot learning’ of stimulus-switch associations: a single exposure to a stimulus in the context of a task switch as compared to a task repetition led to reduced switch costs when reencountering that stimulus [50,51]. Since a single prior exposure would be insufficient to produce substantive incremental learning effects, this finding provides support for episodic reinstatement of control processes involving a particular stimulus mediating item-level flexibility learning.

Finally, flexibility adjustments to changes in reward prospect may rely on yet another process, since flexibility in these studies is not modulated by learning about recent or item-specific demands on task-switching, but rather by detecting (task-orthogonal) changes in reward prospect from the most recent trial [36•–39]. This could be accomplished either by a trial-by-trial learning mechanism that records and compares the most recently obtained reward to the reward prospect on the present trial, or by a standard reinforcement learner that tracks a running average of recent reward, calculates the mismatch (prediction error) between that estimate and the current prospect, and increases flexibility in response to a positive reward prediction error [36•–39].

Conclusions and Outlook

A fast-growing literature on how people adjust cognitive flexibility to varying contexts has revealed several behavioral phenomena and candidate mechanisms. However, many important questions remain unanswered, representing important targets of future research, especially given the long-term goal of formulating interventions to ameliorate failures of flexibility [11••]. For instance, the precise nature of, and relationship between, the different stipulated mechanisms for adapting flexibility are yet to be determined. Accordingly, it is currently unknown whether an individual who displays high meta-flexibility in cued task-switching would also do so in the context of voluntary task selection or trial-and-error driven rule switching. It is also unclear whether people engage in different strategies for adapting meta-flexibility, including perhaps compensatory ones in populations with reduced switch readiness. Thus, additional individual difference studies of meta-flexibility are needed. Similarly, how the lab-based task-switching protocols discussed here relate to real-life attempts at dual- or multi-tasking is subject to some uncertainty [52] and, more broadly, naturalistic studies of cognitive flexibility, and contextual modulation thereof, are sorely missing. Finally, there is an emerging debate about the relationship between cognitive flexibility and its’ conceptual counterpart, cognitive stability, the ability to focus on one’s current task [7,18,25,38,46••,53,54]. While it is intuitive to assume that one’s degree of task focus and readiness to switch tasks should be inversely related [7,17–19], some recent data suggest instead that the two might be independent [25,54]; this possibility and its’ implications remain to be fully evaluated [46••].