Introduction

Successfully managing even the most mundane situations requires a multitude of mental processes. For example, when shopping in a supermarket, you first need to retrieve the item you wish to purchase from your memory. Next, to identify the item, you need to perceptually process the visual information around you. If you cannot locate the desired item on the shelf right away, you will need to compare the appearance of the items on the shelf with the representations of the item you are looking for in your memory. Often, it can help to remember the last time you purchased the item in this store: instead of a time-consuming visual search, you can simply focus your attention on the most likely location of the item. Finally, once you have located the item and put it into your shopping cart, you need to maintain and update your memory of other items you still want to buy. To efficiently perform in such situations, you need to flexibly adapt to the demands of changing contexts and dynamic environments. If these demands exceed your range of functional cognitive flexibility over a prolonged period of time, cognitive plasticity can be triggered1. Plasticity is the brain’s capacity to implement lasting changes that alter its functional and behavioural repertoire2,3. One way to advance our understanding of plasticity is to study cognitive training, which refers to behavioural interventions designed to improve cognitive performance, and measure its effects in both the laboratory and everyday life.

Cognitive training interventions typically target cognitive abilities that are central to human learning, problem-solving and innovation throughout the lifespan, including the basic abilities required in the situation described above. Many of these cognitive abilities develop rapidly until young adulthood, stabilize in adult maturity and, then, begin to decline with age4. In addition, these abilities can be affected by developmental neurocognitive disorders, such as attention deficit hyperactivity disorder5 or autism spectrum disorder6; age-related disorders, such as dementia7; and impairments after acquired brain injury8. Thus, affordable, easy-to-administer behavioural interventions that improve cognitive abilities are highly desirable.

Owing to the flexibility and potential of cognitive training, research exploring its effectiveness has seen a surge in popularity over the past 20 years. Studies have consistently demonstrated training effects (also referred to as training gains): performance improves in the training tasks from the first training session to the last training session. However, the ultimate goal of cognitive training is to establish transfer of training to contexts or outcomes that differ from the trained tasks. Inconsistent evidence for such transfer effects, pervasive methodological concerns and a shift towards more refined theoretical accounts of the mechanisms9 underpinning training-induced cognitive change have led to heated debates in this field.

In this Review, we emphasize theories of training and transfer as well as the current state of evidence of the malleability of the most frequently targeted cognitive abilities: perception and attentional control, working memory, episodic memory and multitasking. In contrast to previous reviews of the cognitive training literature10,11,12, we highlight the importance of identifying the mechanisms that underlie training-induced cognitive performance improvements, and how a deeper theoretical understanding of these mechanisms can be harnessed to develop more robust, reliable and powerful cognitive training interventions.

Design of cognitive training studies

The gold standard for testing whether training generates transfer is to use a pretest–post-test study design (Fig. 1). Training-induced improvements are measured by assessing cognitive performance before and after cognitive training, with some studies additionally including a follow-up assessment to evaluate the durability of effects over a longer period (often 3–6 months after the end of training). Training interventions vary in content: some involve the repetitive practice of cognitive laboratory tasks (Table 1), whereas others focus on metacognitive instructions, for example teaching of strategies. Many training interventions comprise 10–20 sessions, each between 20 and 60 min long; however, the number of sessions can vary markedly13,14.

Fig. 1: A typical cognitive training study design.
figure 1

Cognitive training is evaluated by comparing changes in measures of interest (training task, near transfer task and far transfer task) between pretest (before training), post-test (immediately after training) and follow-up (typically 3–6 months after the end of training) assessments. Change is compared between the training group or groups that perform the training task and the control group or groups that complete alternative interventions (active control groups) and/or no intervention (passive control groups). Note that not all training studies include control groups and a follow-up assessment.

Table 1 Sample cognitive training tasks by domain

Assessments of training-induced improvements vary in the degree to which the processes they measure overlap with the processes targeted by the training tasks. Gains in tasks that assess the same processes as the training task but with different stimulus materials and/or task structure are considered near transfer effects. Gains in tasks thought to rely on the trained processes to a lesser degree (for example, gains in related but different cognitive processes or in everyday life functioning) are interpreted as far transfer effects15,16.

To distinguish training-induced effects from changes in performance that arise simply from repeatedly completing the same set of tasks (test–retest effects), the group of participants undergoing the training intervention (training group) is compared with passive control groups or active control groups. Passive control groups do not undergo an intervention. Active control groups undergo an intervention that does not rely on the cognitive ability targeted by the training intervention. The advantage of active control groups is that they additionally control for placebo and expectancy effects17,18 as well as other non-specific changes that occur due to taking part in an intervention, for example, benefits to everyday time management from learning to adhere to a regular training schedule19. In addition, more elaborate active control group activities enable identification of the specific mechanisms underlying any training effects20.

The design of control group activities is one of the biggest challenges in this field. The activities must be as plausible and believable as the training group’s activities to elicit similar expectations regarding training success. Importantly, however, they must not rely on the same cognitive processes as the training activities; otherwise, no conclusions can be inferred about the effects of training. What constitutes an ideal control group is still debated and varies between training approaches for different processes19,20,21. Regardless of the type of control group, it is critical to demonstrate that all groups perform similarly in the transfer tasks before training to avoid confounding pretraining differences and regression to the mean with post-training between-group differences that reveal the effects of training.

Other methodological considerations for cognitive training studies include the validity and statistical power of transfer assessment. Transfer effects can manifest in behavioural, neural and biopsychosocial outcomes. Past studies primarily assessed behavioural outcomes using laboratory tasks, with some also considering their neural correlates using techniques such as neuroimaging22 (Box 1) and electroencephalography23. Only a few studies have also evaluated gains in everyday cognition outside the laboratory24 or biopsychosocial benefits such as increased quality of life, well-being and physical and mental health25.

The validity of transfer assessment has been questioned because cognitive abilities and other outcomes are often assessed by only a single measure26,27. Single measures are not process-pure because they generate task-specific variance in addition to individual differences in the outcome of interest28. For example, the ability to update contents in working memory can be assessed with a letter keep-track task, in which participants are presented with a continuous series of letters and must recall the most recent three letters once the sequence has ended. Gains detected in this task might reflect better performance with the particular stimuli (that is, letters) or the specific paradigm (that is, keep track) in addition to the ability to update information in working memory. To assess changes in cognitive abilities unconfounded with task-specific gains, performance needs to be assessed using multiple measures.

A criticism of many training studies is inclusion of sample sizes that are too small to detect any transfer effects, which are expected to be small (Cohen’s d = 0.20) or medium (Cohen’s d = 0.50) at best29. For example, training groups in the literature typically comprise no more than 30 participants30,31. For a medium effect size, 30 participants translate into theoretical statistical power of 48%. Thus, even medium transfer effects would be detected only in about every second study. Thus, low statistical power is problematic because it can lead to false negative results32. Ironically, low statistical power also increases the likelihood of false positive results33. Finally, low statistical power can substantially inflate estimates of effect sizes34. For example, a simulation study demonstrated that, for a true medium effect (Cohen’s d = 0.50) that is tested with 30 participants per group, about 98% of effect sizes are inflated35. One way to address these problems is to more adequately power cognitive training studies and to evaluate the strength of evidence with Bayesian inference29 (Box 2).

Despite these methodological concerns, past studies should not be indiscriminately dismissed if they did not include active control groups, broad assessment of transfer or adequate sample sizes. Instead, the strengths and weaknesses of the methodologies used must be taken into account when evaluating the overall evidence for the effectiveness of cognitive training, in particular when interpreting findings from meta-analyses. Because meta-analyses average across effect size estimates reported in the primary literature, they directly rely on the methodological quality of the empirical studies included36. For example, when averaging across overestimated effect sizes from studies with small sample sizes, the overall effect size estimate will be equally inflated. Similarly, if meta-analyses do not distinguish between studies with active and passive control groups, non-specific training effects may contribute to the average effect size estimates.

Theories of training and transfer

The theoretical questions in cognitive training research are whether training generalizes and which cognitive processes change during training. These two questions are inherently intertwined. To develop maximally effective interventions, we need to know not only when transfer occurs but also which cognitive processes are prone to change so interventions can target them directly. However, if we can predict how cognitive performance changes during training but lack any theoretical idea of how this change transfers to other contexts, we cannot develop training interventions that effectively improve human cognition. Theoretical frameworks that speak to both questions can help guide cognitive training research. One such framework is the capacity-efficiency model of cognitive training and transfer19.

The capacity-efficiency model

The capacity-efficiency model proposes that training can induce transfer through two pathways (Fig. 2): training can expand cognitive capacity, that is, the overall cognitive resource available to an individual, or training can increase the efficiency with which the existing capacity is used. To better understand the distinction between training-induced enhancements in capacity and efficiency, a physical training analogy might be helpful. When training on weightlifting with the goal to lift heavy objects, trainees who enhance their capacity increase their muscle mass. By contrast, trainees who improve efficiency learn to use leverage to lift heavy objects without increasing their muscle mass. Critically, and in contrast to the assumptions of other theoretical frameworks1,37, according to the capacity-efficiency account, enhanced cognitive efficiency is not limited to the acquisition of strategies or general task knowledge. Other possible mechanisms that may underlie training-induced enhancements of efficiency include an increased level of automatization or increased speed of information processing that frees up cognitive resources for other concurrent tasks. Gains in capacity and efficiency are not necessarily mutually exclusive; training may yield broad benefits through enhanced efficiency as well as capacity.

Fig. 2: The capacity-efficiency model of cognitive training and transfer.
figure 2

Cognitive training can lead to near or far transfer through enhanced cognitive capacity or enhanced cognitive efficiency. For example, an increase in the amount of information that can be held in memory at one time is an increase in capacity, whereas using a strategy to remember more items more easily is an increase in efficiency. Transfer can manifest in behavioural, neural and/or biopsychosocial outcomes that overlap with the trained processes to varying degrees, resulting in near or far transfer. Moderating variables can modulate performance gains during training but also the extent of transfer by affecting either or both mechanisms of transfer.

Moderators at the level of training38,39,40 or transfer41,42 can amplify or attenuate training benefits (Fig. 2). Moderators can include factors related to the intervention (for example, training tasks43 or conditions44), between-person differences (for example, age, initial cognitive ability, biological and neural predispositions or personality), within-person fluctuations (for example, affect, motivation, well-being, physical and mental health or everyday leisure activities) or the environment (for example, external events or the environment of the intervention).

Gains in capacity and efficiency do not map simply onto single behavioural indicators of near and far transfer effects, nor is a change in any particular neurobiological metric conclusively indicative of either enhanced capacity or efficiency45. Moreover, training can both increase or decrease brain structure and function46. For example, gains can be accompanied by a reduction in overall energy required to complete the task at hand47, or even an increase in functional neural activity16. Thus, relying on single behavioural tasks and neurobiological markers is too simplistic. Instead, conclusively distinguishing between training-induced improvements in capacity and efficiency requires theoretical identification of potential mechanisms of cognitive efficiency and selection of transfer outcomes that allow for systematically pitting them against changes in capacity.

Transfer of capacity and efficiency gains

Most past cognitive training studies aimed to enhance capacity based on the idea that it will generalize to untrained outcomes that draw on the same capacity limit, maximizing far transfer. This basic rationale loosely builds on common-elements theory48, which hypothesizes that transfer occurs if knowledge components are identical across tasks. In later variations and extensions of this theory, identical knowledge has been replaced by sharing a cognitive process49 or functional overlap22. Specifically, if two tasks overlap in the cognitive processes they demand, any gains in these underlying processes should transfer from training on one task to performing the other task. However, this basic rationale of functional overlap comes with two challenges. First, training will involve more than just practising what is considered the underlying process. For example, training on the letter keep-track task described above involves the processes of updating information in working memory but also visual processing, encoding and recalling the stimuli. Second, tasks may seem to be functionally overlapping because of similar surface structures or because they correlate well, but our understanding of the underlying processes might be wrong.

Under the functional overlap rationale, one approach for deriving testable predictions of transfer effects is to consider the structure of individual differences in cognitive abilities50. Put simply, if individual differences in two cognitive abilities are strongly correlated, it is assumed that they have processes in common and, therefore, transfer should be more likely. For example, one suggestion16 is to define transfer distance based on the three-stratum factor model50, which differentiates between 69 narrow abilities that are grouped into 8 broader abilities, with general intelligence on top of the hierarchy. Transfer from one task to another within the same narrow ability would constitute nearest transfer, whereas transfer within one broad ability would be intermediate transfer, and transfer to a different broad ability would reflect far transfer through the change in general intelligence. For example, for an object-location memory training task51, in which participants have to remember and recall the associations of objects and their locations, nearest transfer within the same narrow ability (visual memory) would be gains in a spatial memory task, intermediate transfer within the same broad ability (meaningful memory) would be improvements in a verbal episodic memory task and far transfer to a different broad ability (reasoning) would be gains in a matrix reasoning task. One criticism of using the factor-analytic structure of cognitive abilities to determine functional overlap is the implication that correlation reflects causation, as changes in one ability are assumed to causally lead to changes in a correlated ability. However, correlations between abilities do not necessarily reflect common processes but could be due to another common factor, such as a common biological basis that contributes to their development52. Indeed, evidence from recent working memory training studies has shown that transfer can be absent even between strongly correlated tasks53, with statistical modelling suggesting that correlations do not necessarily reflect common cognitive processes54. Independent of the approach used to determine transfer abilities, enhanced capacity is assumed to manifest in transfer to a wide range of cognitive tasks19.

Similar to changes in capacity, enhanced efficiency of overlapping processes or acquisition of strategies or routines that are useful across different contexts could translate into patterns of relatively broad transfer effects19. The triarchic theory of learning55 proposes that, initially, novice learners of a new cognitive task will rely mainly on their metacognitive system to generate and establish new behavioural routines. These routines may also involve strategies such as grouping of information or mental imagery. Once these routines are formed, the role of the metacognitive system will diminish, and learners will mostly engage their cognitive control network to execute these new routines. Finally, with sufficient practice, learners will move from controlled towards automatic task execution. For example, students of arithmetic might first rely on explicit multiplication but, later, can automatically retrieve the previously stored answer (for example, when answering ‘what is two times two?’). Using the acquired cognitive routines flexibly in different contexts with reduced involvement of the cognitive control network may establish transfer without necessarily increasing the capacity of the cognitive system.

In a similar vein, the cognitive routine framework37 suggests that training on a task involves learning a new skill by developing new cognitive routines. In the beginning of training, when the training task is still novel and unfamiliar, general cognitive resources are needed to identify and execute the routine, which becomes automated by the end of the training regimen. Transfer will be observed if the newly acquired automatic cognitive routine can also be applied in the transfer measure. Depending on the particular mechanism affected by training, transfer through cognitive efficiency can be narrower (for example, if efficiency is enhanced through acquisition of stimulus-specific strategies) or broader (for example, if the speed of visual processing has increased). Detailed task analyses and formalized models of cognition56 (Box 3) are useful for generating testable predictions of when transfer can be expected across tasks through either capacity or efficiency gains, but these models are rare.

Training effects and mechanisms

Cognitive training studies typically target one or more specific cognitive abilities. The abilities most commonly targeted include perception and attentional control, working memory, episodic memory and multitasking. Below, we will discuss the empirical evidence for transfer effects and the mechanisms assumed to underpin training and transfer effects for each of these abilities.

Perception and attentional control

Perception is the organization, identification and interpretation of sensory information to form mental representations. Attentional control is the ability to regulate information processing during goal-directed behaviour. Both skills are critical for understanding the environment and executing intended behaviours. Training interventions that target early basic perceptual and attentional control range from laboratory, tasks such as speed of processing training57, to interventions that build more directly on real-life activities, such as mindfulness meditation58,59,60 (not discussed here) and action video gaming61,62.

Speed of processing training is a process-based training approach in which participants are trained on variations of the useful field of view test63,64 to improve their visual search and divided attention abilities. Typical speed of processing tasks include identifying, locating and/or comparing stimuli (for example letters, pictures or digits) as quickly and as accurately as possible. Often, to keep the task challenging over the course of training, task difficulty is adaptively adjusted to individual performance by displaying the stimuli for less time or further apart from each other, or by increasing the amount of visual or auditory distraction or the number of concurrent tasks. A meta-analysis reported, on average, small training-induced improvements in speed of processing (d = 0.22) and spatial and sustained attention (d = 0.14)57. Thus, speed of processing training transfers to other laboratory tasks that measure the trained cognitive constructs. Moreover, the same meta-analysis reported evidence for small far transfer effects to real-world and biopsychosocial outcomes such as activities of daily living (d = 0.27), well-being (d = 0.21) and driving (d = 0.36)57.

Another family of cognitive training interventions involves training inhibitory control, which requires voluntary, goal-oriented attention to particular features or objects, in tasks such as the go/no-go task65,66,67 or stop-signal task68,69. In these types of tasks, participants have to respond (‘go’) to certain stimuli or stimulus characteristics and inhibit their response (‘no-go’) to others. The stimuli in the training task can be artificial (for example, abstract visual patterns) or real-world (for example, photographs of food). Although inhibitory control training rarely transfers to other laboratory tasks70, a recent meta-analysis71 across 19 studies revealed a small to medium overall benefit (d = 0.38) of this type of training intervention on health behaviours such as reducing the consumption of alcohol72 or high-calorie food73. The effect was stronger for interventions that used stimuli specific to the health behaviour of interest, such as food stimuli for interventions aiming at reducing high-calorie food intake, than for interventions that used unrelated stimuli. Hence, learning seems to be stimuli-specific but generalizes to other contexts with similar stimuli.

Different from typical laboratory cognitive training tasks, visual environments in action video games are complex and dynamic. The fast-paced and constantly changing task conditions of such games require players to perform multiple tasks simultaneously and to continuously update and adapt their task goals and actions74. Moreover, action video games are typically highly engaging and immersive, unlike standard laboratory training tasks. Thus, a growing body of research has explored their potential as interventions for enhancing perceptual and attentional abilities. However, evidence for the effectiveness of action video game training is mixed, with meta-analyses reporting non-significant (g = –0.12 to g = 0.10)75 or small near and intermediate transfer (g = 0.34)62 effects on cognitive performance averaged across domains. Estimates vary between domains but results overall indicate only small training benefits, if any. For example, there is some evidence for small but significant effects of video game training on top-down attention (g = 0.31)62, mixed evidence for improvements in spatial cognition (ranging from g = –0.04 (ref.75) to g = 0.45 (ref.62)) and no significant evidence for effects on perception (g = 0.26)62 or visual attention and processing (from g = –0.01 to g = 0.22)75.

In contrast to the use of action video games as a cognitive training regimen, habitually playing action video games was consistently significantly associated with better general cognitive performance across domains (g = 0.40 (ref.75) to g = 0.55 (ref.62)) and on measures of perception (g = 0.79)62, visual attention and processing (g = 0.45)75, spatial ability (from g = 0.47 (ref.75) to g = 0.75 (ref.62)), attentional control (from g = 0.27 (ref.75) to g = 0.31 (ref.62)), multitasking (g = 0.55)62 and verbal cognition (g = 0.33)62 relative to control participants who played no75 or less than 1 h62 of action video games. However, cross-sectional studies investigating habitual video game players cannot rule out self-selection effects: improved cognitive performance relative to non-players may exist independently of any impact of playing video games. Nonetheless, these findings might indicate that playing action video games for at least 3–5 h per week for at least 6 months rather than the 10–50 h of a typical training intervention may yield cognitive benefits. Overall, the impact of action video games on cognitive performance remains controversial. Differences in study inclusion criteria, meta-analytic methods, corrections for publication bias, and study design and participant characteristics (for example, different control groups and participant age) likely contribute to the variations in effect sizes; the discussion continues on best practices for meta-analysis and primary research methods20,76.

Transfer effects of training interventions that target perception and attention have been attributed predominantly to enhanced efficiency. Specifically, transfer effects have been attributed to the acquisition of knowledge and skills that, although not directly applicable to new tasks, enable new tasks to be learned more efficiently owing to improved probabilistic inference77,78,79. Specifically, the learning to learn account77 suggests that training increases effectiveness in extracting and accumulating evidence from the task environment, thereby optimizing decision-making and resource allocation. Because learning how to better use evidence from repeated presentations is thought to be a general mechanism of cognitive efficiency, it can yield performance improvements in both trained and untrained tasks.

Working memory

Working memory is the ability to temporarily access mental representations needed for complex cognition in the present moment. Working memory capacity is strongly correlated with many other abilities, such as fluid intelligence80. The hypothesis81 that increasing working memory capacity through training could lead to neuroplastic changes that benefit other related cognitive abilities led to extensive study of working memory training approaches.

Initial excitement about the promise of working memory training arose after early studies found large improvements in working memory capacity and far transfer to fluid intelligence82,83. However, subsequent studies that addressed methodological concerns with the early studies, such as the lack of active control groups, failed to replicate most of these far transfer effects when training was conducted using either tasks similar to those used in the original studies84 or other working memory training tasks53,85,86,87. To date, the findings from working memory training studies are inconsistent; some meta-analyses have shown evidence of small yet significant far transfer effects (estimates ranging from g = 0.18 to g = 0.20)88,89,90,91, whereas effects were non-significant in others (estimates ranging from g = 0.01 to g = 0.20)31,92,93. Furthermore, there is also inconsistent support for near transfer within working memory; transfer has been observed across working memory tasks using the same stimuli or paradigm but not across dissimilar working memory tasks37,91.

Given the narrow and variable effects of working memory training at the behavioural level, the claim that training increases capacity81 is not well supported. However, neuroplastic changes have been observed, including greater functional interconnectivity94 and improved white matter integrity95 of brain networks known to support working memory. These findings raise the question of how these apparent neuroplastic changes can be reconciled with the lack of evidence for behavioural transfer. These neural changes possibly reflect an increase in processing efficiency and connectivity between existing structures rather than an increase in neural capacity per se. Indeed, no training-induced changes were found for potential biomarkers of increased capacity, such as in grey matter volume96.

Some evidence suggests that these changes in processing efficiency following working memory training are associated with the acquisition of strategies that are difficult to apply to new contexts97. The strategy mediation hypothesis posits that such task-specific strategies are developed during training to compensate for the challenges imposed by the training task98 and, therefore, will only benefit performance in highly similar tasks. For example, working memory training has been shown to increase the use of strategies such as grouping, visualization and forming semantic associations, but only small training-related gains were observed in untrained tasks98. The context-specificity of strategies is well exemplified in the case study of participant S.F.99, who trained in a working memory task with digits for 2 years. S.F. was a runner who recoded digit sequences into running times from their long-term memory. With this strategy, they were able to expand their memory span for digits from a typical size of 7 to 79 items. However, their memory span for letter sequences remained unchanged because this strategy of utilizing running times was not applicable to letter stimuli. However, strategy acquisition can yield benefits in untrained tasks if they afford the same strategies. For example, improvements in untrained working memory tasks are moderated by the degree to which participants use various encoding strategies on similar untrained tasks97.

Some studies have investigated other potential mechanisms underpinning working memory training effects, including interference resolution53,100, removal of no longer relevant information53 and switching attention between representations53,101, or more global strategies, such as relying more on familiarity-based processing than recollection. Finally, a further change in efficiency may result in improvements in probabilistic learning77, enabling participants to better take advantage of task regularities. For example, over the course of completing just a single working memory task, performance increased as participants learned to take advantage of statistical regularities in the task. These regularities enabled participants to compress the presented information and use their available working memory capacity more efficiently102. These statistical regularities can be stimuli-specific and task-specific, and therefore may not be evident in dissimilar transfer tasks, which explains why near transfer is observed only in tasks with structures similar to those of the training task.

Episodic memory

Episodic memory is the ability to encode and retrieve information with its appropriate context. Episodic memory gives individuals a sense of continuity and identity, and is critical to maintain independence across the lifespan. Episodic memory training often uses strategy-based training and involves teaching mnemonics to support depth and specificity of encoding103,104,105,106 based on mental imagery (for example, method of loci) or elaboration (for example, semantic associations). Mnemonics are powerful training devices and typically produce near transfer effects103,104,105,106,107. However, mnemonics must be used with material that lends itself to the strategy. Hence, far transfer is usually not assessed with entirely different material but with self-report measures of metacognition. Although these interventions seem to improve knowledge and use of mnemonics in everyday life, little evidence supports benefits for complex daily activities24,104,106,107. Nonetheless, a meta-analysis of 30 studies showed that training of episodic memory strategies can, on average, induce small benefits to activities of daily living (d = 0.32), mood (d = 0.16) and metacognition (d = 0.37)108. Hence, whereas gains in cognitive efficiency from strategy-based episodic memory have narrow cognitive benefits, they can positively impact everyday well-being.

To facilitate transfer, some studies have combined strategy training with explicit metacognitive approaches. For example, the Méthode d’Entraînement pour Mémoire Optimale (MEMO) programme105,106 teaches a range of mnemonics and includes explicit instructions regarding when and for which type of material the mnemonics are appropriate. The programme also teaches about age-related memory changes and how to improve metacognition and self-efficacy. MEMO was found to improve memory and metacognition in healthy older adults and in those with mild cognitive impairment105,106. Similarly, strategy-adaptation training, which encourages participants to test memory strategies in different contexts, was found to yield transfer to other domains and memory processes109. These findings show that instructing participants in how to adapt trained strategies is critical to meet the demands of unpractised tasks and contexts to produce training transfer.

Process-based approaches to improving episodic memory include interventions that manipulate memory load, retrieval intervals or interference during retrieval. Studies that target memory load administer episodic associative memory tasks in which participants have to remember an adaptively increasing number of associations, for example, objects and their locations. Findings of these studies are mixed. In an object-location memory training study, near transfer to spatial episodic memory and far transfer to reasoning were observed 4 months after training51. However, another associative memory training study found no transfer effects to untrained tasks involving episodic associative memory (near transfer) or reasoning (far transfer)110.

Studies that manipulate the spacing between repeated retrieval attempts build on the robust finding that memory is better after spaced learning than massed learning111. One explanation for the benefits of spaced learning is that the intervals between study episodes introduce variability at encoding, which increases the spectrum of retrieval cues and improves performance. Another possible explanation is that spaced learning ensures an optimal balance between retrieval effort and retrieval success112. Generally, practising with equal intervals between study and retrieval attempts can be equally effective in terms of long-term retention of learned materials as practising with intervals that gradually increase over the training phase113,114. However, positive feedback is more frequent during the latter paradigm, which might increase interest in the training task and reduce frustration113.

Another approach to episodic memory training is to manipulate the level of interference at retrieval by interspersing recall with other materials, referred to as repetition-lag training. The rationale of repetition-lag training is to increase reliance on consciously controlled, recollection-based memory processes rather than automatic, familiarity-based memory processes115,116,117,118. Participants first study lists of items and, then, memory is probed in recognition trials that include a gradually increasing number of lures that are repeated at increasingly longer intervals. This procedure has been shown to benefit recollection in adults115,116,117,118,119,120,121 and in people with dementia118. These positive effects persisted over a 3-month delay period117. However, transfer of those gains to novel materials or tasks is weak115,116,117,119,120. Hence, benefits from repetition-lag training are limited to the trained tasks and materials.

Taken together, both strategy-based and process-based episodic memory training approaches seem to predominantly affect cognitive efficiency, with training benefits typically limited to highly similar tasks and materials. Approaches triggering metacognitive processing, such as strategy-adaptation interventions, are more promising avenues to producing far transfer.

Multitasking

Multitasking entails flexible task switching or performing two tasks concurrently (dual tasking)122. Task switching, such as performing a colour identification task on some trials and a shape identification task on other trials, requires switching between tasks on different trials. Training studies have demonstrated substantial reductions in switch costs in the trained tasks across the lifespan89,123, including in clinical groups such as children with attention deficit hyperactivity disorder124. Most studies also reported near transfer effects in untrained switching tasks125, suggesting improvements in the ability to flexibly switch tasks on a trial by trial basis. However, the extent of near transfer varies; for example, one study found that near transfer was much greater in children (aged 8–10 years) and older adults (aged 62–76 years) than in younger adults (aged 18–26 years)126. Findings of far transfer following task switching training are inconsistent, with some studies reporting transfer to other executive functions and fluid intelligence126,127 and others observing no far transfer at all128,129. One hypothesis is that only training interventions that require proactive control strategies lead to transfer. Indeed, studies that reported broad transfer typically used alternating-run task switching training paradigms, in which participants have to monitor the task sequence to switch tasks at the appropriate time but can also prepare for upcoming task switches. Thus, participants must use a proactive cognitive control strategy56,130.

The Primitive Information Processing Elements (PRIMs; Box 3) model explains why training a proactive control strategy can lead to transfer effects49. The PRIMs model conceptualizes a proactive strategy in task switching as two operators acting in conjunction: the first operator initiates task preparation and adjusts tasks goals, and the second operator executes the task goal. Based on this model, training task switching is thought to be effective because the proactive operators are trained and, therefore, become more efficient to use. According to this account, transfer will be demonstrated in tasks that can reuse the trained operators.

Similar to improvements in task switching, dual-task costs can be extremely reduced and, under some conditions, even eliminated after training131,132. Training studies that aim to reduce dual-task costs often compare fixed priority conditions with variable priority conditions. Under fixed priority conditions, participants are asked to emphasize both tasks equally throughout training. In variable priority conditions, participants are instructed to flexibly change their task response priorities, thereby varying how attentional resources are split between the two tasks. Findings from variable priority training studies suggest that participants acquire skills during dual-task training that optimize allocation of limited attentional resources when processing multiple competing tasks133,134,135,136,137,138,139. Larger training-related improvements have also been observed under the variable priority conditions than in fixed priority conditions for untrained dual tasks133,138,139. Thus, to some extent, this optimization of resource allocation is generalizable to other contexts. Alternatively, dual-task training could induce acquisition of skills that improve coordination of multiple tasks. To test this hypothesis, effects of fixed priority dual-task training were compared with pure single-task training, in which participants trained on two tasks separately131,140. Dual-task training led to larger improvements in dual-task performance than single-task training, both in trained and untrained dual tasks132,141,142, demonstrating transferable gains in task coordination skills. However, whether these skills are accessible in other task contexts is disputed, with some studies showing short-term far transfer effects (for example, in measures of sustained attention and working memory23,143) whereas others do not144,145.

In sum, multitasking training that involves task monitoring or flexibly changing priorities may enhance efficiency through optimizing engagement of proactive control strategies, attentional resource allocation and task coordination. Near and far transfer is therefore most likely to occur in contexts where these mechanisms operate as well.

Summary and future directions

Cognitive training interventions usually lead to large improvements in the trained task, regardless of the cognitive ability targeted and the training approach used. However, broad transfer to cognitive, neural or biopsychosocial measures other than the training task is much more elusive. To develop cognitive training interventions that are successful in generating transfer, it is critical to identify the cognitive mechanisms underpinning training and transfer effects. There are several candidate mechanisms (Table 2).

Table 2 Observed patterns of transfer and mechanisms with empirical support

Returning to the capacity-efficiency model of cognitive training and transfer described above (Fig. 2), these findings suggest that when transfer occurs it is primarily driven by improvements in cognitive efficiency, with little convincing evidence for gains in cognitive capacity. Thus, future research should focus on attempts to enhance cognitive efficiency rather than increase cognitive capacity. Notably, assumptions of a fixed cognitive capacity were the predominant view146 until findings from training studies that challenged this view were published in 2002 (ref.147) and 2008 (ref.82). The notion of a fixed cognitive capacity might seem to contradict to the idea of cognitive and neural plasticity. However, a fixed capacity does not imply that cognitive performance is immutable. Instead, some of the processes prone to training-induced enhancements of cognitive efficiency could be potentially useful in a wider range of cognitive tasks as well as in real-world contexts. To advance our understanding of cognitive plasticity, future research should investigate these mechanisms of cognitive efficiency further and identify how they can be harnessed for interventions that induce generalizable cognitive improvements. For example, explicit metacognitive instruction that supports strategy application across contexts105,106,148 and adaptive process-based metacognitive training, in which participants practise to increase accuracy of their performance estimates149, are promising avenues for far transfer. Metacognitive instruction is also a potential avenue for maximizing transfer of efficiency gains in tasks that require probabilistic inference and attention allocation.

Cognitive training can be a powerful experimental tool for studying cognition. Patterns of transfer can yield insights into how cognitive abilities are related and can advance our understanding of the causality underlying these relationships54. Moreover, identifying the mechanisms of cognitive training can offer insights about the nature of individual differences in cognition. Especially when combined with neuroimaging (Box 1) and computational modelling approaches (Box 3), cognitive training can help to delineate the contribution of individual differences in cognitive capacity and cognitive efficiency to overall cognitive performance, on both a conceptual and an analytical level. Critically, to fully exploit cognitive training as an experimental tool to advance our understanding of human cognition and cognitive plasticity, we must move past vague notions of common elements and develop theories that can generate falsifiable, testable hypotheses.