The honest version

How to actually improve, and what is a waste of your time

Almost everything written about "raising your IQ" is either selling something or confusing three different things. This page separates them, says what the evidence supports, and gives you something concrete to do.

Three things people mean by "improve"

1

The score

How many points this test prints. Moves fast with practice on the formats. Worth a handful of points, then it stops.

Easy to move

2

The skills

Vocabulary, mental rotation, algebraic fluency, deduction. Genuinely learnable, genuinely useful, and they are most of what school is for.

Moves with work

3

General ability

The shared factor underneath all of them, g. Moves slowly, mostly through years of education and health, not through exercises.

Moves slowly

If someone promises you the third thing in six weeks, they are selling you the first thing. The good news is that the second one is large, real, and entirely in your control.

What the evidence actually supports

Staying in education, and taking the demanding version of it

Strongest evidence

This is the single best-supported intervention on cognitive ability that exists. A 2018 meta-analysis pooling 42 datasets and over 600,000 people (Ritchie & Tucker-Drob, Psychological Science) used natural experiments, policy changes that forced some cohorts to stay in school longer, and found roughly 1 to 5 IQ points per additional year of schooling. The gains persisted into old age and were not just test familiarity.

What to do: take the harder course when you have the choice. Do not drop the subject that makes you think hardest because it threatens your average. Sit the proof, not the summary.

Reading volume

Strong for verbal

Vocabulary is acquired overwhelmingly through reading rather than conversation or instruction, written language is where the rare words live. Heavy readers pick up thousands of words a year that light readers never meet. This is the most directly trainable index on this test, and unlike puzzle drilling it is not a trick: a larger vocabulary is a real and permanent increase in what you can think about precisely.

What to do: read things slightly above your comfortable level, and read non-fiction outside your subjects. When a word stops you, look it up rather than guessing past it. Twenty minutes a day beats a weekend binge.

Sleep

Large short-term effect

Sleep does not raise your ceiling, but losing it drops you well below it. A few nights restricted to around five hours produces measurable slowing of processing speed and losses in working memory, the two indices this test measures with timed tasks. Teenagers need roughly 8 to 10 hours, and the biological clock shifts later during adolescence, which is why early starts hurt.

What to do: a consistent schedule matters more than the occasional long night. And never benchmark yourself on this test while short of sleep, you will record a dip that is not real.

Aerobic exercise

Small but real

Meta-analyses in children and adolescents find small-to-moderate improvements in executive function and processing speed from regular aerobic activity. The effect is modest, do not expect it to change your composite, but it is one of the few interventions that shows up repeatedly with a plausible mechanism, and the rest of the benefits are obvious.

What to do: roughly 150 minutes a week of something that raises your heart rate.

Fixing an actual deficiency

Large, if it applies

Correcting iodine or iron deficiency in a deficient person produces substantial cognitive gains, in severely iodine-deficient populations, on the order of ten points. Removing lead exposure likewise. The catch is that supplementing when you are not deficient does essentially nothing. This is a repair mechanism, not an upgrade path.

What to do: eat properly; get bloodwork if you have reason to suspect a deficiency. Ignore every supplement marketed on cognition.

Learning something genuinely difficult, for years

Real, but narrower than advertised

Mathematics, programming, a second language, an instrument. The honest framing: the evidence for these raising general ability is weak, but the evidence that they build deep, durable, transferable-within-domain skill is overwhelming. Proof-based maths in particular trains exactly the move the Deductive Reasoning section tests, following what must be true rather than what sounds true.

What does not work

This part matters as much as the list above, because these are where people put their hours.

Commercial brain-training apps

Not supported

The pattern is consistent across large reviews (Simons et al., 2016, Psychological Science in the Public Interest): you get better at the trained game, a little better at very similar tasks, and no better at anything that matters. The industry has settled legal cases over advertising claims on exactly this point.

Working-memory training (n-back and friends)

Failed to replicate

This had the most promising early results and the most thorough debunking. Melby-Lervåg, Redick & Hulme (2016) pooled the literature and found reliable near transfer and no reliable far transfer to fluid intelligence, with the apparent effects concentrated in studies lacking active control groups. Your working memory span will improve at n-back. It will not carry.

Chess and music as IQ boosters

Near zero

Meta-analyses by Sala and Gobet find that once you restrict to studies with active control groups, the effect of chess or music instruction on general cognitive ability is approximately nothing. Learn them because they are worth learning.

Nootropics, Mozart, binaural beats, "10% of your brain"

No

No evidence worth the name. The Mozart effect was a brief arousal effect on one spatial task and did not survive replication.

The awkward middle: raising the score without raising the ability

Practising matrix puzzles, number series and mental rotation will move your score on this test, probably by several points, because each has learnable strategies. Whether that is cheating depends entirely on what you want the number for.

Legitimate

If you are preparing for an entrance exam, an aptitude test, or a selection process that uses these formats, practise them deliberately. The test is the target, and being fluent in the format is a real advantage you are entitled to.

Self-defeating

If you are using this app to measure yourself over years, practising the formats corrupts your own instrument. You will record gains you did not make and lose the ability to tell real change from familiarity.

The resolution is simple: keep practice and measurement apart. If you want to drill puzzles, drill them, but do not drill in the month before a benchmark sitting, and write in your notes that you did, so future you knows why the line jumped.

Index by index: what to actually do

Fluid Reasoning Gf · hardest to move

The ability to find a rule you were never taught. Least responsive to training, most responsive to years of demanding study.

Proof-based mathematics rather than procedural exercises, the point is deriving, not executing.
Competition-style problems (olympiad, Project Euler) where the method is not given to you.
For balance-scale items, convert everything into the smallest shape before doing any arithmetic. Working in mixed units is where almost all the errors come from.
When stuck on a matrix, work attribute by attribute: shape, count, shading, rotation, size. Ask what each does along a row before looking at the options. Most errors come from pattern-matching the answers instead of extracting the rule.

Verbal / Knowledge Gc · most improvable

This index is essentially a measure of accumulated learning, which means it responds to effort more reliably than any other part of the test.

Read above your level, daily. Non-fiction, long-form essays, anything with argument structure.
Keep a running list of words that stopped you. Revisit it, spaced, not crammed.
Learn Latin and Greek roots. A few dozen unlock thousands of words and let you attack unfamiliar ones by decomposition rather than memory.
Write regularly. Producing precise language is what converts recognition into command.

Visual-Spatial Gv · trainable

Spatial ability shows some of the clearest training gains in the literature, and they persist reasonably well.

Technical drawing, CAD, 3D modelling, anything that makes you hold an object and turn it.
Speedcubing, mechanical puzzles, origami.
For paper-folding items: work backwards one crease at a time, mirroring the holes across each fold in reverse order. Do not try to picture the whole unfolding at once.

Working Memory Gwm · train the strategy, not the span

Raw span is close to fixed and the training literature is discouraging. What genuinely improves is how efficiently you use the span you have.

Chunk aggressively, 4 8 2 9 1 5 is three pairs, not six digits.
Mental arithmetic without writing anything down builds the habit of holding intermediate results.
Outside a test, the real win is offloading: notes, diagrams, written intermediate steps. Reducing the load beats expanding the buffer.

Processing Speed Gs · mostly a state, not a trait

Of the five, this is the one most affected by things that have nothing to do with ability, sleep, illness, caffeine, screen fatigue, even the mouse you are using.

Treat a low score here first as a question about your condition on the day.
It is the least reliably calibrated index in this app. Use it for tracking yourself, not for comparing yourself to others.

How to track yourself over years without fooling yourself

Benchmark at most twice a year, on the Full battery. More often and you are mostly measuring practice.
Standardise the conditions. Same rough time of day, rested, fed, quiet room, no interruptions, same device. Inconsistent conditions are the largest source of noise you actually control.
Discard your first sitting as a baseline. Sitting two is typically 5 to 8 points higher purely from format familiarity. Treat sitting two as your true starting point.
Use the reliable-change threshold the app prints after every retest. If the change is smaller than that, it means nothing, however much you want it to.
Remember the score is age-adjusted. Between about 13 and 18 your raw ability rises steeply on its own. A flat line here does not mean stagnation, it means you are keeping pace with your age group. A rising line means you are pulling ahead of it, which is a genuinely hard thing to do.
Write down what was going on. Illness, exam week, four hours of sleep, a month of puzzle practice. In three years you will not remember, and the note is what turns a squiggly line into an explanation.
Export a backup after every sitting. The data lives in this browser only. A cleared cache erases a decade.

What the number does and does not predict

Worth keeping in proportion, in both directions.

It is a real predictor. Cognitive ability correlates meaningfully with school achievement and with performance in complex jobs. It is one of the better single predictors psychology has.
It is nowhere near destiny. Even a correlation of 0.5 leaves most of the variation unexplained. Recent re-analyses have argued the classic job-performance figures were inflated by range-restriction corrections, revising them downward.
Conscientiousness competes with it. For school grades specifically, self-discipline and study habits predict about as well as measured ability, and unlike ability, they are straightforwardly trainable.
Heritability is not immutability. Adult IQ heritability estimates run high, roughly 0.6 to 0.8, but heritability describes variation within a population under existing conditions. It does not say a given person's score cannot change, the education effect above is the proof.
Whole populations have moved. Scores rose about three points per decade through most of the twentieth century, the Flynn effect, which is far too fast to be genetic. Whatever these tests measure responds to environment on a scale that dwarfs anything a puzzle app is going to do.

The most useful thing you can do with a score is not defend it. It is to find the index you are weakest on, notice that it is almost certainly the one you avoid practising, and go and do that.

Under the hood

How this is scored, and where it can be wrong

A score you cannot audit is a horoscope. Here is every step, including the parts that are estimates.

1 · Items are generated, not stored

Every matrix, series, rotation figure, folded sheet, span sequence and speed symbol is constructed from code at the moment you sit down, from a random seed. Two consequences:

You cannot memorise the test. This is the single biggest threat to tracking yourself over years, and generation removes it almost entirely. Only the verbal bank is fixed, because generated vocabulary items would be nonsense, it holds around a hundred items and samples a fraction of them per sitting.
Every item is checked for solvability. Generating an item that is true is easy; generating one that is answerable is not. Balance-scale items are only kept when the premises pin down every ratio, formally, when the null space of the premise matrix is one-dimensional, so exactly one option can balance. Number series are audited the same way: the rule that produces the shown terms must also produce the keyed answer.
Difficulty is known by construction. A matrix governed by four simultaneous rules is harder than one governed by a single rule, and by a knowable amount. The generator records that rather than guessing after the fact.

The seed of every sitting is stored with your result, so any test you have taken can be rebuilt exactly.

2 · Ability is modelled, not counted

Counting correct answers treats an easy item and a brutal one as worth the same, which they obviously are not. Instead each response is fitted with a logistic item-response model:

P(correct) = c + (1 − c) · 1 / (1 + e^{−(θ − b)})

θ is your ability, in logits. This is what gets estimated.
b is the item's difficulty, on the same scale. Roughly: −2 is easy, 0 is around average for an adult, +2 is hard.
c is the guessing floor, fixed at 1/(number of options). With five choices, someone who knows nothing still scores 20%, and a model that ignores that reads luck as ability.

θ is estimated by maximum a-posteriori with a standard-normal prior, which keeps the estimate finite when you get everything right or everything wrong, and yields a standard error directly from the Fisher information. That standard error is where the confidence band comes from.

Discrimination is held equal across items, a Rasch-style assumption, because estimating a discrimination parameter per item requires a calibration sample this app does not have. Pretending otherwise would be false precision.

3 · The score is age-adjusted

An IQ of 100 does not mean "average human". It means "average for people your age". A fifteen-year-old and a thirty-year-old with identical raw performance should not get the same score, because their peer groups differ.

So your θ is shifted by the expected θ at your age before conversion. The age curves follow the shapes established in the developmental literature: fluid reasoning, spatial ability and processing speed rise steeply through adolescence, peak in the late teens to mid twenties, then decline gradually; crystallised verbal ability keeps rising into middle age and holds far longer.

The conversion is then the standard deviation IQ scale: IQ = 100 + 15 · (θ − θ_age)

4 · The composite is not a plain average

Five index scores are combined with weights approximating their loadings on general ability, fluid reasoning heaviest at 0.30, processing speed lightest at 0.13.

But averaging correlated measures shrinks their spread. Average five independent standard scores and the result has a standard deviation well under 1, which would squash everyone toward 100 and make high scores impossible. So the weighted sum is rescaled by the standard deviation implied by the weights and the correlation between indices (taken as 0.50, typical for index intercorrelations):

SD_composite = √( Σw_i² + 2·r·Σ_i<jw_iw_j )

Skipping this step is one of the most common ways amateur test scoring goes wrong.

5 · Two different uncertainties, and why only one matters for you

Measurement error, shown as your band

How much the score would bounce if you re-sat with identical underlying ability. It comes out of the item-response model and gets wider when the test is shorter. This is the number that decides whether a change between two sittings is real, and it is the basis of the reliable-change threshold the app prints.

Calibration error, a constant offset

The item difficulties and age curves here were reasoned out, not measured on a standardisation sample. The whole scale could sit several points high or low. But it is the same offset every time you sit. It affects how your score compares to other people's; it cancels out completely when you compare yourself to yourself.

Which is why the trajectory on the History page is the trustworthy output of this app, and the absolute number on the results page is the approximate one. If you ever sit a professionally administered test, note the difference between the two scores, that difference is this app's calibration offset, and knowing it turns the absolute number into a real one.

6 · The two tasks that are not scored this way

Memory span is an adaptive staircase: sequences lengthen until you miss two at the same length. Your span is converted to a z-score against age-referenced means (around 4.6 for a fourteen-year-old, about 5.1 for an adult, standard deviation 1.3). Two span tasks are combined with a correction for the correlation between them.

Processing speed is scored correct minus incorrect, so guessing does not pay, then z-scored against an age-referenced expectation. Be aware this is the weakest calibration in the app, the task is modelled on a clinical subtest but is not that subtest, and its norms are the most uncertain here. It is reliable for comparing you with you, and shakier for comparing you with the population.

7 · Known limitations

No standardisation sample. Stated plainly everywhere it matters, including here.
The reasoning sections are untimed. Real tests are usually timed, and untimed administration measures something slightly different. Time per item is recorded, so take it the same way each sitting.
The verbal bank is in English and assumes a reader comfortable in it. If English is not your first language, expect the verbal index to understate you and read the other four.
Extreme scores are the least trustworthy. Above about 135 or below about 65 there are too few items at the right difficulty to pin you down, and the band widens accordingly.
Nothing here is a clinical instrument. It cannot diagnose anything, and it should not be used to make a decision about education, employment or health.
Where your data lives: results are stored only in this browser. Nothing is ever transmitted.