Does mental abacus training actually work? What the research says — nulls included

Two questions sit behind almost every enquiry we get about mental abacus training. Does it actually make a child better at maths, and could it do any harm? Both have real answers in the published literature, and neither is quite the answer this category usually markets. The short version: the arithmetic gains are supported, but only over years — and the claims about intelligence are not supported at all. This page walks through the four papers that carry the weight, including the trial that found nothing.

What the evidence supports: arithmetic, measured in years

The strongest single piece of evidence is a randomised controlled trial that ran for three years. Barner, Alvarez, Sullivan, Brooks, Srinivasan and Frank published it in Child Development, volume 87, issue 4, 2016, pages 1146 to 1158. In Vadodara, India, 204 children were enrolled and 183 completed, receiving three hours of instruction a week. The mental-abacus group outperformed the control group on arithmetic, with an effect size of Cohen’s d = .60 and a 95% confidence interval from .30 to .89 on the study’s own measure. That is a genuine, moderate advantage on exactly the skill the training is about.

The same paper is also where the honest caveats live. The authors report that the training "did not alter basic cognitive abilities". And the gain was not evenly spread: baseline spatial working memory moderated how much each child learned, and children below the median on that measure did not differ from controls. Moderated, not mediated — the paper does not claim the training built the working memory. It reports that the working memory a child already had predicted how much of the training landed. That finding constrains our marketing as much as anyone else’s.

What it does not support: intelligence

The most useful synthesis available is Chunjie Wang’s 2020 review in Frontiers in Neuroscience. It concludes that abacus-based mental calculation training "has the potential to enhance various cognitive skills including mathematics, working memory and numerical magnitude processing" — and that hedge is the review’s own wording, not ours. On fluid intelligence it is plainly negative: a five-year longitudinal study with active controls found "no significant group difference ... on the Raven’s intelligence scores", and the review states that the effect on fluid intelligence "appears to not go beyond a placebo effect". It describes the wider pattern as benefits "only in tasks tightly related to the trained tasks (near transfer)", and flags small samples, cross-sectional designs and passive control groups as weaknesses that "may artificially inflate the overall effect".

Two consequences follow for anyone reading a brochure. First, the framing those brochures use — "right-brain training", "whole-brain development", "activating both hemispheres" — is not what the imaging literature actually describes; the network reported is bilateral, frontal and parietal, on both sides. We do not use that language about our own product and we would not try to defend it. Second, the most circulated brain-imaging citation in this category, Ku and colleagues in PLOS ONE 7(5):e36410 from 2012, is explicitly a case study of one expert. It tells you something interesting about one extraordinary calculator. It is not evidence of brain change in learners generally, and it should not be quoted as though it were.

The one-year trial that found nothing

The counterweight comes from the same lab. Barner, Athanasopoulou, Chu, Lewis, Marchand, Schneider and Frank, in Journal of Numerical Cognition, volume 3, issue 3, 2017, pages 540 to 558, randomised classes of 180 US first- and second-graders to either standard maths instruction or standard instruction plus mental abacus, for one school year. They "did not see evidence of differential change in performance for either the in-house arithmetic or standardized WJ-III measures". The only signal was a marginal place-value trend in second graders, at p = .052. And one mechanical detail explains a great deal: only 21% of first graders could accurately decode multi-digit abacus representations by the end of the year. The authors conclude: "Overall, our results suggest caution in the adoption of MA as a short-term educational intervention."

Set the two trials side by side and the pattern is not a contradiction — it is a timescale. Three years at three hours a week produced a moderate arithmetic advantage. One year produced no measurable one. Wang’s review reaches the same place from the other direction, noting that most studies detecting a benefit did so after around three years. There is a plausible mechanism. Frank and Barner, in Journal of Experimental Psychology: General, volume 141, issue 1, 2012, pages 134 to 149, are the standard reference for the finding that mental-abacus users appear limited to holding three or four abacus columns at a time, as though each column were a separate object in visuospatial working memory — a laboratory result about how the representation is structured, not a trial of teaching outcomes. Building that image and then extending it is slow work. A single term of classes does not do it.

Who gains the most, and where a beginner should start

• Children who already had stronger spatial working memory got more out of the three-year trial. Children below the median on that baseline measure did not differ from controls. This is the finding programs quote least and parents should ask about most.
• Duration beats intensity. The evidence for a measurable arithmetic advantage sits at multi-year exposure, not at a single term. If a program cannot describe what year two and year three look like, the research does not cover what it is selling.
• Near transfer is the honest promise. Expect gains on arithmetic and on tasks close to arithmetic. Do not budget for a change in general ability.
• The starting rung matters more than the starting age. A child who already knows the complements is bored by a bead-introduction level and lost in a multi-row one. That is a placement problem, not a birthday problem.
• Small samples are the norm in this literature. Most positive findings outside the two trials above come from studies small enough that one unusual classroom moves the result.

The concerns parents raise, taken seriously

Search in Arabic and one of the most common queries is not whether mental arithmetic works, but whether it harms. An Arabic guide on an Egyptian schools directory, dated 30 June 2026, is headlined as a warning about the harms and risks of mental arithmetic for children, and lists pressure from speed-focused drilling, memorising shortcuts instead of understanding, anxiety about making mistakes, and training time displacing play and reading. That page is search-oriented directory content by an unnamed editorial team, not research, so read it as a list of what parents worry about rather than a set of findings. The worries are still legitimate, and they apply to speed-drilled mental arithmetic in general — ours included.

• Pressure from speed drilling: keep the timer off until the method is secure. Speed is the last thing to train, not the first.
• Shortcuts without understanding: a child who can move beads for 7 + 5 but cannot say why the five-bead comes down has learned a gesture. Ask for the reason, not only the answer.
• Anxiety about mistakes: private practice, no audience and no leaderboard in the early weeks. Errors should be cheap.
• Time displacing play and reading: cap the session. One abacus teacher, writing on a teaching blog in July 2024, observed that classes commonly run about 75 minutes while five- and six-year-olds hold focus for about 10 to 15 minutes, and that a few of her own young students left after a single day while others dropped out after a couple of months. That is one teacher describing her own classes rather than retention data — but it matches what every parent of a five-year-old already suspects.

How we hold ourselves to this

• Arithmetic is the promise. Faster, more accurate mental calculation, and comfort with numbers. Not intelligence, not IQ, not hemispheres.
• Ten minutes a day, not a 75-minute sitting. The session length is set by what a young child can actually attend to, and the research does not reward intensity over duration anyway.
• Level is an explicit placement, never derived from age, because the trial that worked measured what children could already do before it measured what changed.
• We state sample sizes and funding, and that goes double when a study flatters the format we sell. The 2024 randomised trial of the tablet app SoroTouch, published in PLOS ONE on 12 March 2024, had 20 participants, 10 per arm, and its funding statement discloses that the app’s developer funded the work and had a role in data collection. It is a legitimate study and a thin reed, and we will not lean on it. No operator in this category publishes audited outcome figures — Kani included.

Last reviewed 30 July 2026, against Child Development 87(4) 2016, Journal of Numerical Cognition 3(3) 2017, Frontiers in Neuroscience 2020 and Journal of Experimental Psychology: General 141(1) 2012. Brand names mentioned here are trademarks of their respective owners; Kani is not affiliated with, endorsed by, or licensed by any of them.