Does Working-Memory Training Transfer to Intelligence Tests?

The short answer: transfer is possible, but a reliable broad IQ gain is not established

Working-memory training can improve performance on the exercises people practice, and gains sometimes carry over to other memory tasks. Whether that improvement reaches intelligence-test performance is a narrower and less settled question. Reviews support a cautious answer: broad transfer is not established, though focused analyses report small positive effects. This makes “training can never transfer” too strong, but does not establish a durable, broad IQ increase. The distinction matters because “working memory improved” and “intelligence-test performance improved” describe different results. In “Working Memory Training Does Not Improve Performance on Measures of Intelligence or Other Measures of ‘Far Transfer,’” a 2016 meta-analysis considered 87 publications and 145 experimental comparisons. It found reliable immediate gains on related verbal and visuospatial working-memory measures, but no convincing evidence of reliable gains on more distant outcomes, including nonverbal and verbal ability, when training was compared with a treated control condition. This is a broad baseline, not proof that every method must produce no transfer; the review covered varied programs and outcomes. A newer analysis, “The Asymmetry of Working Memory Training Transfer: A Systematic Review and Meta-Analysis,” reports that adaptive updating training produced a small positive average on non-working-memory outcomes; its span-training estimate was near zero, and the direct difference between those estimates was only marginal. It examines intelligence-related outcomes separately. Such group averages do not forecast individual IQ change or show that effects persist. The apparent disagreement is therefore partly about which training family and outcome a review examines. Improvement on a practiced exercise, transfer to a similar memory task, and change on a more distant intelligence measure should not be collapsed into one claim. The next sections examine those differences, which are central to interpreting claims about transfer. For now, the responsible conclusion is neither that transfer is impossible nor that memory practice reliably raises intelligence-test scores. Practice a memory task for the skill it measures, and look for any broader benefit on new, relevant tasks. Do not choose a program on a promise of an IQ gain.

Sources: Working Memory Training Does Not Improve Performance on Measures of Intelligence or Other Measures of ‘Far Transfer’: Evidence From a Meta-Analytic Review; The Asymmetry of Working Memory Training Transfer: A Systematic Review and Meta-Analysis

What counts as transfer to an intelligence test?

Transfer is improvement on an untrained task attributable to practice, compared with a control group. Better performance on the trained exercise shows learning there; it does not show that learning carried elsewhere. “Working memory” means temporarily holding information while using or changing it. A span task emphasizes retaining items for recall; an updating task requires monitoring changing information and revising what is held in mind. An n-back task is one updating exercise: the participant judges whether the current item matches one shown a set number of steps earlier. The 2026 review “The Asymmetry of Working Memory Training Transfer” examines whether gains travel between these task families. Near and far transfer describe overlap between practiced and measured tasks. Remembering one short list, then recalling a new list, is relatively near transfer because both involve retaining items. Practicing monitoring changing items, then taking a novel matrix reasoning task, is more distant because it calls for different problem solving. “Far” does not mean more valuable, and “near” does not mean trivial: improvement on a related memory task supports a narrower conclusion than improvement on untrained reasoning. A new measure can still resemble training; a different screen does not make it far transfer if the operation is much the same. An intelligence-related test may share demands with working-memory tasks without being the same task. The 2016 meta-analysis “Working Memory Training Does Not Improve Performance on Measures of Intelligence or Other Measures of ‘Far Transfer’” treated verbal and visuospatial memory outcomes as intermediate transfer and examined nonverbal ability separately. A raw score is performance under test rules. An IQ interpretation asks how it compares with an appropriate reference group under a defined scoring system. A research meta-analysis pools group differences on selected measures. Hedges’ g, used in the 2026 review, expresses the training-versus-control difference in change in standardized units; it is not IQ points or a personal conversion. A positive pooled effect indicates greater average improvement, not one reader’s normed intelligence score. A treated, or active, control gives participants an activity intended to match some of the experience, such as time or engagement, while withholding target training. It helps separate training-specific change from retesting or attention, though no control matches every expectation. The 2016 review reports intermediate gains but no convincing far-transfer improvement against treated controls. “Improved after practice” and “improved more than a comparable control on untrained reasoning” are different claims.

Sources: The Asymmetry of Working Memory Training Transfer: A Systematic Review and Meta-Analysis; Working Memory Training Does Not Improve Performance on Measures of Intelligence or Other Measures of ‘Far Transfer’: Evidence From a Meta-Analytic Review

Why the control group changes what a gain means

A score rising after practice does not, by itself, show that the practice caused the rise. Retesting, attention, and ordinary variation can also change scores. A comparison group helps estimate what changed beyond those influences. The useful question is whether trainees improved more than a comparison group tested on the same schedule. Matching conditions reduces alternative explanations, though it cannot remove every difference between participants or activities. That difference estimates the added benefit of training. In “Working Memory Training Does Not Improve Performance on Measures of Intelligence or Other Measures of ‘Far Transfer’: Evidence From a Meta-Analytic Review,” the authors compared working-memory training with both treated and untreated controls across studies of far-transfer outcomes. A treated control does an activity intended to match some features of training, such as time or engagement, while omitting the target practice. Untreated controls may leave contact or expectations unmatched. Neither label guarantees that groups were matched perfectly; the actual activities and conditions matter. For immediate nonverbal-ability outcomes, the 2016 review reported a pooled standardized difference of Hedges’ g = 0.05 for training compared with treated controls (95% confidence interval, −0.02 to 0.13; 67 studies). Against untreated controls, the corresponding estimate was g = 0.20 (95% confidence interval, 0.11 to 0.28; 53 studies). The review found that this difference by control type was statistically significant for nonverbal ability. Here, Hedges’ g expresses the average between-group difference on a standardized scale; it is not an IQ-point change. The larger untreated-control estimate shows that the apparent effect depends on the comparison. It does not establish that all gains reflect retesting or expectations, or that every finding vanishes with an active control. A good control is judged by what it accounts for: were participants tested at comparable times, given similar contact and task demands, and offered activities likely to produce comparable engagement and expectations? An active comparison is not automatically ideal: the control activity itself could affect the outcome, and the pooled result cannot diagnose each study. So when a report says an intelligence-related score rose after working-memory practice, check the comparison behind that statement. Was the outcome an intelligence measure, and did the trainees improve more than a suitably engaged control group? Without that contrast, a before-and-after rise cannot isolate the training effect. Even with it, the result is evidence about that outcome and study design, not a personal forecast.

Sources: Working Memory Training Does Not Improve Performance on Measures of Intelligence or Other Measures of ‘Far Transfer’: Evidence From a Meta-Analytic Review

What do broad reviews say about far transfer?

Broad reviews support a cautious baseline: working-memory training can improve performance on related memory measures, but dependable improvement on more distant ability tests has not been established. In a 2016 meta-analysis, “Working Memory Training Does Not Improve Performance on Measures of Intelligence or Other Measures of ‘Far Transfer’,” it synthesized 87 publications and 145 experimental comparisons. It found immediate gains on intermediate verbal and visuospatial memory measures, but no convincing improvement over treated controls on farther outcomes: nonverbal and verbal ability, reading, decoding, or arithmetic. This is a conclusion about the reviewed evidence, not proof that transfer is impossible. The distinction between the two result groups matters. Improvement on a related memory measure shows that performance changed on a task sharing relevant demands with training. It does not, by itself, establish a broader change in reasoning. Their mediation analyses did not find that relationship. Thus the evidence did not support a simple chain in which more improvement on the trained memory capacity reliably predicts more improvement on distant outcomes. A person may get better at memory exercises without that gain serving as a dependable proxy for performance on an intelligence test. The review's quantitative results make the cautious conclusion more specific. Against treated controls, its pooled immediate effect for nonverbal ability was Hedges' g = 0.05, with a 95% confidence interval from −0.02 to 0.13; for verbal ability, g = 0.05, with an interval from −0.07 to 0.17. Both estimates are small and their intervals include zero. These group comparisons are not IQ points or individual forecasts. Evidence for delayed far transfer was limited, so immediate post-training performance should not be called durable improvement. A separate perspective comes from “Near and Far Transfer in Cognitive Training: A Second-Order Meta-Analysis,” published in 2019. Rather than combining individual trials directly, this study synthesized results from earlier meta-analyses. It distinguished memory from farther outcomes such as reasoning, processing speed, and language; near transfer appeared, while far transfer was small or null. It reinforces the outcome distinction, but is not a fresh analysis of every participant: its estimates inherit the coverage and study-selection decisions of earlier reviews. Together, the reviews make a reliable general intelligence benefit a poor expectation, while leaving room for targeted exceptions. They do not establish that no training design can ever affect an intelligence-related task; the 2016 review explicitly cautioned against that universal claim. Nor should a later positive result in a narrower training family be dismissed simply because broad averages were near zero. It should instead be compared with this baseline: Is the outcome genuinely distant from practice? Was the improvement measured against a suitable comparison group? Does the finding concern an average study effect, or does it show durable change on an appropriate intelligence measure? Until those questions are answered, evidence of memory-task learning remains evidence of that learning, not a dependable forecast of higher intelligence-test performance.

Sources: Working Memory Training Does Not Improve Performance on Measures of Intelligence or Other Measures of ‘Far Transfer’: Evidence From a Meta-Analytic Review; Near and Far Transfer in Cognitive Training: A Second-Order Meta-Analysis

What the n-back counterevidence does—and does not—show

N-back findings make a categorical claim that working-memory training never transfers too strong. In “Improving fluid intelligence with training on working memory: a meta-analysis,” a synthesis of 20 controlled studies of healthy adults aged 18–50 found a small, statistically significant average effect on fluid intelligence after n-back training. This is counterevidence: the outcome was a reasoning ability measure beyond the trained exercise, and the studies included control groups. The result is specific to n-back programs, that age range, and studies available to the review. It is a pooled group result, not evidence that every participant improved, that any particular person should expect a gain, or that the change would last. Nor does a positive result on selected fluid-intelligence measures establish a general increase across intelligence tests.

A later synthesis asked a more finely divided question. “Working memory training revisited: A multi-level meta-analysis of n-back training studies” analyzed 203 effects from 33 published randomized controlled trials in healthy adults. The authors separated untrained n-back tasks from other working-memory tasks, fluid intelligence, and cognitive-control tasks, accounting for multiple outcomes per study. The clearest transfer was to untrained n-back tasks; estimates for other working-memory measures, fluid intelligence, and cognitive control were very small. This pattern does not directly replicate or cancel the earlier meta-analysis: the reviews used different inclusion and analytic choices, and the later review explicitly distinguished outcomes by their relation to the practiced task. It does show why “a different test improved” needs a second question: how different was that test? A new n-back task retains much of the monitoring and updating demand. Fluid reasoning is a more distant outcome. A result in the first category cannot simply be counted as a result in the second.

The estimate drew a methodological challenge. “Reliable gains? Evidence for substantially underpowered designs in studies of working memory training transfer to fluid intelligence” reanalyzed the early n-back evidence and found indications that the underlying studies had low statistical power. Its bias-adjusted estimates were smaller than the original pooled estimate and their uncertainty included zero. They found no clear evidence of publication selection, where positive findings appear more often than negative ones. Small studies may still yield unstable estimates: some effects look large by chance, while genuine small effects are hard to detect. The authors called these estimates tentative; their reanalysis cannot settle later research or prove the true effect is zero. Together, these reviews support a narrower conclusion: n-back practice can transfer beyond the exact trained task, and one synthesis reported a small positive fluid-intelligence signal, but the breadth and reliability of that signal remain uncertain. Improvement on another n-back exercise is meaningful task-family transfer; it is not, by itself, evidence of a broad change in general intelligence or a predictable rise in an individual’s IQ score.

Sources: Improving fluid intelligence with training on working memory: a meta-analysis; Working memory training revisited: A multi-level meta-analysis of n-back training studies; Reliable gains? Evidence for substantially underpowered designs in studies of working memory training transfer to fluid intelligence

What does the 2026 updating-versus-span comparison establish?

The 2026 review adds a positive result to the working-memory transfer debate: adaptive updating practice transferred to untrained span tasks, and its estimate for outcomes outside working memory was small and positive. This does not establish a dependable rise in intelligence-test performance. The direct difference between updating and span training on those non-memory outcomes was marginal; the separate intelligence-related analysis showed a similar pattern, not a validated estimate of IQ gain. “The Asymmetry of Working Memory Training Transfer” combined 55 independent studies, 208 effect sizes, and 3,492 healthy participants in its main analysis. The three-level method accounted for related outcomes from the same study. Searches covered studies through September 30, 2024; eligible studies used adaptive training and included a control group. Span tasks asked participants to retain or recognize items after a delay. Updating tasks required ongoing monitoring and manipulation of changing information, using formats such as n-back or running memory. These are task families, not interchangeable versions of one exercise. For transfer between these families, updating training improved span outcomes on average (Hedges’ g = 0.176, p < .001); span training’s estimate for updating outcomes was smaller and not statistically distinguishable from zero (g = 0.048, p = .453). However, the direct test comparing the two directions was marginal (p = .073). The pattern looks asymmetric, but separate significance tests do not establish that one direction reliably exceeds the other. The review examined outcomes beyond working memory separately. Updating training had a small positive pooled estimate (g = 0.120, p = .001), while span training’s estimate was near zero (g = 0.008, p = .885). The direct contrast was marginal (p = .076). “Significant after updating, not significant after span” does not by itself show that updating is better; the comparison between estimates is the relevant test. A supplementary subset of intelligence-related measures showed a similar pattern, but the main article gives no numerical estimate for it. Why might updating carry over to span? The authors propose a resource-demand account: updating continually tracks, replaces, and manipulates information, potentially recruiting operations that also support span performance. This is a theoretical explanation, not direct evidence that training expanded a general mental resource. The compared task families differ in both demand and breadth, so the analysis cannot isolate which feature produced the pattern. The review therefore sharpens, rather than overturns, a cautious conclusion. It supports possible limited transfer beyond the practiced exercise, including selected intelligence-related outcomes in this synthesis. It does not show how much an individual’s intelligence-test score would change, whether a change would last, or whether it would appear on another intelligence measure. These healthy-sample group effects are not personal score conversions. Better trials would specify intelligence outcomes, report direct contrasts and follow-up, and separate them from gains on related memory tasks.

Sources: The Asymmetry of Working Memory Training Transfer: A Systematic Review and Meta-Analysis; The Asymmetry of Working Memory Training Transfer

Why task overlap is only part of the explanation

Shared features offer a sensible first explanation for why practice sometimes carries over: when two tasks use similar materials or mental operations, learning one may help with the other. But similarity alone does not predict why transfer might run more strongly in one direction than the reverse. The 2026 review, “The Asymmetry of Working Memory Training Transfer: A Systematic Review and Meta-Analysis,” sets out two additional possibilities. A resource-based account predicts that practice on a more demanding task could support a less demanding one because the former calls on a broader set of resources. An ability-based account instead predicts that training an operation useful across many tasks could help a narrower task that also relies on it. The review uses span and updating tasks to put those predictions in tension. In a span task, a person briefly holds information and then recalls or recognizes it. An updating task requires continued monitoring and manipulation of changing information. Under the review’s framing, updating is generally more demanding, while maintenance in span tasks is treated as a broader component on which updating also depends. The resource account therefore predicts stronger transfer from updating practice to span performance. The ability-based account makes the opposite directional prediction: if span practice strengthens a broadly useful maintenance component, it might help performance on updating tasks that also require maintenance. A similarity account, without further assumptions, predicts no consistent directional advantage when shared features are the main explanation. The synthesis favored the resource-based prediction in its span-versus-updating comparison, but it cannot show that demand alone caused the asymmetry. Its database covered studies published through September 30, 2024, and its included samples were healthy participants. The result therefore does not establish the same pattern in clinical groups or in every form of memory practice. The paradigms differ in both demand and the breadth assigned to the operations they recruit; those dimensions were not independently varied. That is an identification problem: a pattern matching one account does not isolate its proposed mechanism when another feature changes at the same time. For a reader asking whether memory practice raises intelligence-test performance, this distinction limits the conclusion. Task overlap remains relevant, and directional predictions may help researchers design stronger comparisons, but one span/updating contrast cannot supply a universal training recipe. A more decisive study would vary task demand and cross-task breadth separately, use credible comparison conditions, and specify an untrained intelligence-related outcome in advance. Until such evidence distinguishes the explanations, treat the review’s mechanism as a useful hypothesis about why transfer patterns may differ, not proof that a particular program will improve an individual’s intelligence-test result.

Sources: The Asymmetry of Working Memory Training Transfer: A Systematic Review and Meta-Analysis

Why a group effect cannot be read as your IQ gain

A group-level effect is not an IQ-point conversion. Hedges’ g summarizes a standardized difference in change between training and comparison groups; it does not say how many IQ points any participant gained. An individual claim would require an appropriate measure, a defensible score interpretation, and evidence that training rather than other causes produced the observed difference. Even then, an average describes the study group, not a guaranteed outcome for each person. Keep four steps separate: observed performance on a task; a norm-referenced score compared with a defined reference group; an interpretation about broader ability; and causal attribution to training. Evidence at one step does not establish the next. A study can report improvement on a reasoning measure without showing that general intelligence changed for each participant or that any change lasted. The 2016 meta-analysis, “Working Memory Training Does Not Improve Performance on Measures of Intelligence or Other Measures of ‘Far Transfer’,” is relevant to the proposed mechanism. Across studies, its mediation analyses found that the size of working-memory gains was not related to gains on far-transfer measures, including nonverbal and verbal ability. This cross-study analysis cannot settle every causal pathway or show transfer is impossible. It does undercut a simple inference: because working memory and intelligence scores are associated before training, improving a practiced memory-task score must raise intelligence-test performance. A baseline association between people and a causal effect of training are different questions. The newer “The asymmetry of working memory training transfer” meta-analysis offers a small signal on selected outcomes, but its pooled estimates do not translate into personal IQ gains. It combined 55 studies of healthy participants, with 3,492 people, and found a small average transfer from adaptive updating training to span tasks. It separately analyzed outcomes beyond working memory, including intelligence-related measures. These standardized group comparisons concern particular measures and conditions; they are not a conversion scale, a forecast for one trainee, or proof of persistence. Repeated testing adds another problem: familiarity can improve performance, and ordinary variation can move scores either way. A before-and-after result alone cannot separate these from a training effect. Describe a retest as performance on that occasion, not proof of changed general ability. The online result needs the same restraint. Test IQ Free’s assessment presents a provisional estimated IQ range from an unvalidated conversion with no validated adult reference population; its page says the displayed band is not a confidence interval. It cannot serve as a validated measure of training-caused IQ change. The optional report adds explanation and a practice plan, not accuracy or validation. Do not convert a pooled effect into points, diagnose ability, or claim a personal intelligence gain from a browser quiz. Report a retest as task performance under those conditions.

Sources: The Asymmetry of Working Memory Training Transfer: A Systematic Review and Meta-Analysis; Working Memory Training Does Not Improve Performance on Measures of Intelligence or Other Measures of ‘Far Transfer’: Evidence From a Meta-Analytic Review; Test IQ Free 50-question reasoning test

What can you responsibly do next?

Choose practice by naming the outcome first. If you want to get better at a working-memory exercise, practice that operation and judge progress on the exercise or a closely related task. If your goal is broader reasoning, include unfamiliar problems that require the relevant reasoning operation, then report performance on those problems directly. Improvement on a practiced task is useful evidence of learning that task; it is not, by itself, evidence that general intelligence changed. The distinction matters because the research reviewed above measures group differences on particular outcomes, not a guaranteed result for an individual.

For a personal check, use comparable conditions each time and prefer fresh examples over repeating items whose answers you have learned. Record the task and the observed result, rather than converting a change into IQ points. A retest can reflect familiarity as well as learning, and a browser score cannot settle that interpretation. Test IQ Free offers a free, adaptive 50-question exercise in pattern, number, and spatial reasoning. Its estimated IQ range is provisional: the conversion is unvalidated and has no validated adult reference population. Treat it as a low-stakes reasoning exercise, not a normed measure or evidence that training raised IQ. You can take the [free test](https://testiqfree.org/test) for practice and review; the optional paid report adds explanation, not validity.

Sources: Test IQ Free 50-question reasoning test