Dual N-Back for working memory

What lies behind it

The n-back task requires following a stream of stimuli and reporting each time the current stimulus matches the one n steps earlier. In the dual version, two stimulus streams run simultaneously, usually a position in a grid and a heard letter. Working memory has to update its content continuously: discard old entries, take in new ones, keep the order.

The theoretical assumption behind it was that working memory capacity is a bottleneck for thinking performance overall — so training this capacity should broadly raise thinking performance. Precisely this assumption was tested directly and did not hold: across all studies, the extent of improvement in the working memory measures was not linked to the size of the far-transfer effects. What happens instead, in the assessment of the meta-analysts, is mainly strategy acquisition. You learn how to solve this particular task.

The four numbers

The paper that resolves the dispute for the n-back paradigm introduces a distinction that previous meta-analyses had not made: it separates task-specific transfer to untrained n-back tasks from transfer to other working memory tasks. The basis is 203 effect sizes from 33 published randomized controlled trials with healthy adults, 190 effect sizes after data cleaning.

The result is a clear gradient: untrained n-back tasks g = 0.62 (0.44 to 0.81), other working memory tasks g = 0.24 (0.16 to 0.32), fluid intelligence g = 0.16 (0.08 to 0.24), cognitive control g = 0.16 (0.05 to 0.27). All four values are statistically significant, but the message lies in the gap: what is large is the task itself. The authors sum it up by saying that a considerable part of the transfer after n-back training is task-specific.

What remains against active control groups

The broadest analysis covers 87 publications with 145 experimental comparisons. The decisive methodological step was the separation by control condition. Immediately after training, there were reliable improvements in intermediate transfer, that is, in verbal and visuospatial working memory. For far transfer — nonverbal ability, verbal ability, word decoding, reading comprehension, arithmetic — there was no convincing evidence against treated control conditions. The publication bias analysis found that the studies with treated control groups have no evidential value.

The broader training literature comes to the same result. In typically developing children, far transfer against active control groups was g = 0.001 (SE = 0.055; p = 0.982), across 41 studies with 393 effect sizes and 2,375 participants. In older adults it was g = −0.008, while trained tasks with g = 0.877 and near transfer with g = 0.274 clearly responded, across 43 studies with 2,140 participants.

The study that tested it directly

The methodologically strictest single test worked with dual n-back itself, not with just any working memory training. Young adults received 20 sessions of an adaptive dual n-back program, an active placebo program with adaptive visual search, or no training at all. The design closed the main weaknesses of the earlier studies: an active placebo instead of a waiting list, several measurement instruments per cognitive construct instead of a single one, plus a no-contact control group.

All participants completed pre-, mid- and post-tests with measures of fluid intelligence, multitasking, working memory capacity, crystallized intelligence and processing speed. Both training groups improved in their respective task — but, despite high statistical power, there was no positive transfer to any of the cognitive ability tests. It was not a null finding due to a lack of participants.

What the studies show

The meta-analysis that breaks down n-back training

Multilevel meta-analysis of cognitive transfer effects in healthy adults after working memory training with n-back tasks; 203 effect sizes from 33 published randomized controlled trials, 190 effect sizes after data cleaning. Transfer to untrained n-back tasks g = 0.62 (0.44 to 0.81), to other working memory tasks g = 0.24 (0.16 to 0.32), to fluid intelligence g = 0.16 (0.08 to 0.24), to cognitive control g = 0.16 (0.05 to 0.27). The moderator analyses showed no effects of age, training dose, training type or content of the transfer tasks.

The randomized test with an active placebo

Young adults received 20 sessions of an adaptive dual n-back program, an active placebo program with adaptive visual search, or no training at all. All participants completed pre-, mid- and post-tests with several measures of fluid intelligence, multitasking, working memory capacity, crystallized intelligence and processing speed. Result: both training groups improved in their respective task, but despite high statistical power there was no positive transfer to any of the cognitive ability tests.

The broadest analysis of far transfer

Meta-analysis of working memory training studies with a pre-post design and a control group: 87 publications with 145 experimental comparisons. Immediately after training, there were reliable improvements in measures of intermediate transfer. For measures of far transfer, there was no convincing evidence against treated control conditions. Mediation analyses showed that the extent of improvement in the working memory measures was not related to the size of the far-transfer effects, and the publication bias analysis found that the studies with treated control groups have no evidential value.

The original study and the counter-calculation from the same camp

The 2008 paper with which it all began reported transfer from demanding working memory training to measures of fluid intelligence and described the effect as dose-dependent: the more training, the greater the gain. A later meta-analysis from the same research group included 20 studies with n-back training, a fluid intelligence endpoint, a control group and healthy participants between 18 and 50 years of age and found a small but significant positive effect, weighted g = 0.24. The authors themselves caution that this effect is measured on laboratory tests.

The training literature beyond n-back

For typically developing children: 41 studies, 393 effect sizes, 2,375 participants; small to medium effects on memory tasks, proportional to the similarity between training and test task, and with active control groups a far transfer of g = 0.001 (SE = 0.055; p = 0.982). For older adults: 43 studies, 698 effect sizes, 2,140 participants; trained tasks g = 0.877, near transfer g = 0.274, far transfer g = 0.121 — and with active control groups, far transfer was zero at g = −0.008.

Where the transfer stops

Transfer to fluid intelligence is very small even in the most positive analysis: g = 0.16 (0.08 to 0.24) in the n-back-specific meta-analysis, g = 0.24 in the analysis from the proponents’ camp. The better the control group, the smaller the effect, down to zero: against treated control conditions there was no convincing evidence of far transfer; in the children’s meta-analysis it was g = 0.001, in older adults g = −0.008. The causal chain was tested directly and did not hold — the extent of working memory improvement was not linked to the size of far transfer. And the evidential value of the studies with treated control groups was rated as non-existent.

The dose-response claim of the original study is not reproducible in the meta-analysis either: training dose did not moderate transfer. Near transfer is also narrower than the everyday term working memory suggests: in a randomized controlled trial with adaptive dual n-back, there was transfer to an untrained single n-back task and to an updating composite score, but no transfer effects on the composite scores for active and passive working memory. Two widespread additional claims have no basis at all: a literature search found no study on the combination with alpha-GPC or citicoline, and the claim that noticeable improvements in focus and working memory set in from level N = 3 was never investigated — no level was tested as a threshold for transfer in the papers analyzed.

How to do it

In the methodologically strictest randomized test, it was 20 sessions of an adaptive dual n-back program, compared with 20 sessions of adaptive visual search as an active placebo and with a group without any training; testing took place before, in the middle and after, each time with several measures per construct. Adaptive here means: the level rises and falls with performance. The studies analyzed consistently work with adaptive versions, not with a fixed n.

Two common practical rules do not stand up to scrutiny. Dual is not better than single — training type was not a significant factor in the moderator analysis across 33 randomized studies; choosing the dual format makes the task harder, not more effective. And more training volume does not bring more transfer; training dose also had no moderator effect. A figure such as 20 minutes four times a week over four weeks therefore remains a practical suggestion, not an effective dose. What you actually get better at is clearly defined: n-back tasks with g = 0.62 and, to a lesser extent, other working memory tasks with g = 0.24.

Safety

Dual n-back training has no physical risks. It is a software task, and the worst that can happen is frustration. The only real harm lies in opportunity cost: the time invested is missing elsewhere. Anyone who regularly invests in a task whose transfer to everyday thinking is g = 0.16 and disappears against a proper comparison group could have put the same time into endurance training, sleep or learning something they actually want to be able to do. The authors of the broadest review put it this way: given the available evidence, the likelihood of finding reliable and robust far-transfer effects is low, and research efforts should be redirected toward near transfer or other methods. A second point concerns expectations: anyone who expects measurably higher intelligence from the training and does not get it may draw the wrong conclusions about themselves. The finding is not that too little training was done — more training did not bring more transfer in the meta-analysis.

Sources

• Soveri et al., Psychonomic Bulletin and Review 2017
• Melby-Lervåg, Redick and Hulme, Perspect Psychol Sci 2016
• Redick et al., J Exp Psychol Gen 2013
• Jaeggi et al., PNAS 2008
• Au et al., Psychonomic Bulletin and Review 2015
• Soveri et al., Experimental Psychology 2017
• Sala and Gobet, Psychonomic Bulletin and Review 2020
• Gobet and Sala, Perspect Psychol Sci 2023