How The IQ Gym works — and what it honestly can't do

The IQ Gym is built on a simple, well-evidenced claim: practice, familiarity, and strategy reliably improve performance on IQ tests. It deliberately avoids a second claim that the evidence does not support: that training here permanently raises your underlying general intelligence. This page explains both halves, and how our scores are computed.

What the science supports

Meta-analyses of test-retest research covering more than 150,000 participants find that simply retaking a cognitive ability test raises scores by roughly a third of a standard deviation — about five IQ points — with further gains on a third attempt before the effect plateaus. Gains are larger when practice is accompanied by coaching, and taking a nonverbal reasoning test such as matrices a second time is worth roughly eight IQ points by one estimate: exactly the material most IQ tests lead with.¹²³

Solving strategy is trainable, too. Research on matrix reasoning distinguishes 'constructive matching' — predicting the answer before scanning the options — from 'response elimination', and finds that the predictive strategy is what stronger performers do. Every The IQ Gym lesson teaches these strategies explicitly: rows-first scanning, one attribute at a time, decision ladders for number series, bridge sentences for analogies, set diagrams for syllogisms.⁴

What we will not claim

Studies that trained working memory hoping to raise fluid intelligence — the famous dual n-back literature — did not hold up: training improves the trained task and close variants, with little credible transfer to broader ability. We measure working memory because a battery that samples it describes you better — not because drilling it will raise your general ability. Regulators have also acted against brain-training marketing that promised real-world cognitive gains without evidence. So our claim is exact: The IQ Gym trains your test performance — your familiarity, pattern knowledge, strategy, speed, and confidence — and how much it moves varies from person to person. We do not promise a new brain.⁵⁶⁷

What the test samples — and what it doesn't

Modern intelligence research organizes abilities in the Cattell–Horn–Carroll (CHC) model: a general factor at the top, broad abilities beneath it — fluid reasoning, comprehension-knowledge, visual-spatial processing, working memory, processing speed — and narrow abilities under those. Every The IQ Gym family is tagged with the CHC ability it draws on: matrix puzzles and series load on fluid reasoning (induction), vocabulary items on comprehension-knowledge, rotation and folding on visual-spatial processing. We score six domains — abstract, numerical, verbal & logic, spatial, working memory and processing speed — which sample the broad abilities that mainstream IQ tests weight most heavily. The free quick test samples the first four; the full mock exam covers all six, with working memory and processing speed carrying deliberately small weight in the composite.⁸

Sampling is the honest word. A short online battery measures a useful slice of the CHC hierarchy, not the whole of it — a full clinical evaluation samples more abilities, in person, with an examiner. And a domain profile is a study guide, not a diagnosis: scoring lower in one domain here means that domain deserves training time, nothing more.

One honest caveat about the categories themselves: our "Verbal" domain groups two different abilities — vocabulary knowledge (synonyms, antonyms, analogies) and verbal reasoning (letter series, deduction, syllogisms). They correlate, but a sharp gap between them is meaningful, so read a Verbal result as a blend and lean on the lesson-level breakdown when the two diverge. Numerical and Spatial likewise mix pure reasoning with a little learned knowledge. We tag every family with its precise CHC ability; your report already breaks Verbal into its vocabulary (Gc) and reasoning (Gf) strands, and we are working toward the same split for the other mixed domains.

How your score is estimated

Start with the part of this that most online tests leave out: a short test cannot produce a precise number, and ours does not pretend to. Twelve minutes of questions is a small sample of how you reason, and a small sample gives a wide answer. That is not a defect we are working around — it is the arithmetic of measurement, and the honest response to it is to show you the width rather than hide it behind a confident-looking integer. A longer test narrows the range; nothing makes it disappear.

Every question in our bank carries a difficulty rating that updates as real people answer it — the same family of methods (Rasch/Elo and item-response theory) used by serious adaptive testing programs. Today those ratings start from each question's designed difficulty: calibration against real answers has only just begun. Anchor questions will be frozen to hold the scale in place once each has 200 answers, and ultra-fast responses are screened out so guessing sprees don't distort the statistics. When you finish an assessment we estimate your ability in each domain from which questions you answered correctly and how hard they were, then combine those domains into one overall estimate.

That overall estimate then has to be put on the familiar IQ scale, and this is the step worth being precise about, because it is the one that gives the number its meaning. An IQ score is norm-referenced: it says where you sit relative to a reference population, not how many questions you got right. So your estimate is ranked against the distribution of estimates this same test produces across that population — 100 is that distribution's middle, and 15 points is one standard step away from it. That is the same scale convention the Wechsler and Stanford–Binet tests use; it is not the same instrument, the same supervision, or the same precision, and this page says so twice more below. The percentile printed beside your score is then simply the position of that number on the same curve, which is why the two can never tell you different things.

We always report a band, never a bare number, because measurement without uncertainty is fiction. Even gold-standard clinical tests carry a margin of several points; a 12-minute online sample carries more. Your band is an honest 90% confidence range — the coverage the assessment standards use for an individual score — and it narrows as a form samples each domain more deeply. The exact width of yours is printed on your report, and it is wider on the short free test than on the full-length exam for exactly the reason above.

One assumption inside that band is worth stating plainly, because it is the one that decides how wide the band is. Your six domain scores are not measurements of six unrelated things — people who reason well abstractly tend to reason well numerically too. Combining scores that move together cannot cancel out their errors the way combining independent ones would. So we carry a provisional design value for the inter-domain correlation — how strongly the underlying abilities move together — of 0.65, the middle of the range published for broad-ability test batteries, and work out from it how much of each domain's uncertainty is shared rather than independent. Averaged over everyone who takes a form, a band we label 90% then contains the true score about 90% of the time; without this, our own simulations put that figure nearer 70%. The correction makes every band we publish wider, and wider is the honest direction.

That 90% is an average across everyone, and an average can hide its tails — so here are ours. In simulation the band contains the true score about 90–93% of the time for takers in the middle of the range, but only about 82–87% of the time for the lowest-scoring and highest-scoring tenths. The reason is the norm-referencing step above: a norm-referenced score deliberately sits between your own standing and the population average, so at the far ends of the scale a symmetric band around it misses on the outside more often than it should. A longer, more reliable test shrinks the effect — it is smallest on our 49-item mock, where the tails hold 86–87% — and nothing else does. We would rather print the range than an average that flatters it.

Getting that step right took us two goes, and both are worth stating. First: because domain scores move together, an average of them sits closer to the middle than any single one does, so converting that average as though it varied as much as one domain squeezed every result toward 100. Second, and larger: an estimate from a short test is itself less spread out than the true ability behind it — measurement pulls estimates toward the average — so ranking it against a table of true scores squeezed every result toward 100 all over again. Both are the same mistake, which is reading a score off a table built for something else. Your estimate is now ranked against the estimates this test really produces, so the spread of reported scores is the spread the scale claims, and the percentile is exactly the percentile of the number you see. Anything scored before this keeps the number it was given: we version our scoring and never quietly rescore an old result.

Testing professionals evaluate a score the way the Standards for Educational and Psychological Testing require: state the intended interpretation, then back every link in the argument. Ours, in one sentence: your The IQ Gym band estimates your current performance level on IQ-test-style reasoning tasks, scored against disclosed provisional design norms — backed by checked items with worked explanations and model-based scoring. It is not certified evidence of clinical IQ, and we never sell it as such.

Why online scores are estimates

Professional IQ tests are administered one-on-one by trained examiners under controlled conditions, and normed on carefully constructed population samples. An online test can control none of that: people take it on phones at midnight, sometimes twice. Our reference distribution is built from our item design, and we will recalibrate it from real response data once there is enough of it — its provisional nature is disclosed here rather than hidden. No online score, ours included, qualifies you for high-IQ societies or constitutes a clinical assessment.

The training method

• Teach the generating rules of each question family — the small set of patterns items are built from.
• A written lesson for each question family, some with a worked example, then drills with instant worked explanations.
• Adaptive difficulty tuned to the learning research: roughly 85% success for your first 20 drill answers in a domain, then about 72% — hard enough that errors surface, easy enough that corrections land.
• Miss an item and you get one unaided retry before the explanation — generating the correction yourself is what makes it stick.
• A one-tap confidence rating after each answer trains calibration: lucky guesses get scheduled for review even when they landed, and confident errors get the correction that memory research shows sticks hardest.
• Spaced review: your misses — and your slow or lucky corrects — return on an expanding schedule until the pattern sticks.
• Timed sets and full mock exams under exam conditions, because pacing is a skill of its own.
• A trajectory chart of banded mock scores; a change smaller than the band can be noise.

One more honesty note: because our assessments never reuse an item you have seen, a rise in your trend line cannot come from remembering answers — though a change smaller than the band can still be noise.

Sources

1. Scharfen, J., Peters, J. M., & Holling, H. (2018). Retest effects in cognitive ability tests: A meta-analysis. Intelligence, 67, 44–66. doi:10.1016/j.intell.2018.01.003 — 122 studies, 153,185 participants; gains stop after the third sitting.
2. Hausknecht, J. P., Halpert, J. A., Di Paolo, N. T., & Moriarty Gerrard, M. O. (2007). Retesting in selection: A meta-analysis of coaching and practice effects for tests of cognitive ability. Journal of Applied Psychology, 92(2), 373–385. doi:10.1037/0021-9010.92.2.373 — larger gains with coaching.
3. Sala, G., & Gobet, F. (2024, July 8). IQ tests: can you improve your score by practising? The Conversation — about eight IQ points on a retaken nonverbal reasoning test, with no effect on unfamiliar tasks.
4. Vigneau, F., Caissie, A. F., & Bors, D. A. (2006). Eye-movement analysis demonstrates strategic influences on intelligence. Intelligence, 34(3), 261–272. doi:10.1016/j.intell.2005.11.003 — constructive matching versus response elimination on matrices.
5. Melby-Lervåg, M., Redick, T. S., & Hulme, C. (2016). Working memory training does not improve performance on measures of intelligence or other measures of “far transfer”. Perspectives on Psychological Science, 11(4), 512–534. doi:10.1177/1745691616635612
6. Sala, G., & Gobet, F. (2019). Cognitive training does not enhance general cognition. Trends in Cognitive Sciences, 23(1), 9–20. doi:10.1016/j.tics.2018.10.004
7. US Federal Trade Commission (2016, January 5). Lumosity to pay $2 million to settle FTC deceptive advertising charges for its “brain training” program.
8. McGrew, K. S. (2009). CHC theory and the human cognitive abilities project: Standing on the shoulders of the giants of psychometric intelligence research. Intelligence, 37(1), 1–10. doi:10.1016/j.intell.2008.08.004