Is IQ Just Pattern Recognition? What the Test Items Show

No. Pattern recognition is the best single route into g, the general factor behind IQ, but it is a minority of what an IQ test asks. On the WAIS-5, the current adult Wechsler test, the Full Scale IQ is built from seven subtests and only two, Matrix Reasoning and Figure Weights, are fluid reasoning, the family of tasks where pattern-finding lives. In an independent 2026 analysis of the WAIS-5 standardisation sample (2,020 people) those two do load highest on g, at .74 and .70, but the other five Full Scale subtests, vocabulary, verbal reasoning, spatial construction, working memory and speed, load .56 to .68, so g runs through all seven. Some tests, Raven's matrices and the NNAT3 among them, really are made only of patterns, and even Raven's shares only about half of its variance with g. This guide goes through the items, the numbers and the best case for "yes".

What is on a modern IQ test, and how much of it is pattern recognition?

The WAIS-5, published by Pearson in 2024 for ages 16 to 90, has a primary battery of ten subtests grouped into five indexes. Seven of the ten feed the Full Scale IQ, which Pearson says takes about 45 minutes; the full ten take about 60. Pearson's brochure does not name the seven; an independent journal analysis of the test (Canivez and colleagues, 2026) lists them as Similarities, Vocabulary, Block Design, Matrix Reasoning, Figure Weights, Digit Sequencing and Coding.

• Verbal Comprehension: Similarities (say how two things are alike) and Vocabulary (define words), both in the Full Scale
• Visual Spatial: Block Design (rebuild a pattern with blocks), in the Full Scale, and Visual Puzzles
• Fluid Reasoning: Matrix Reasoning (find the rule that completes a grid) and Figure Weights (quantitative reasoning and induction), both in the Full Scale
• Working Memory: Digit Sequencing (hold and reorder digits in mind), in the Full Scale, and Running Digits
• Processing Speed: Coding (match symbols to digits against the clock), in the Full Scale, and Symbol Search

Only Matrix Reasoning is a pure pattern task. Pearson describes the WISC-V Fluid Reasoning Index, Matrix Reasoning plus Figure Weights, as measuring the ability to detect the underlying conceptual relationship among visual objects and to use reasoning to identify and apply rules, and describes Figure Weights on its own as measuring quantitative reasoning and induction. Block Design and Visual Puzzles are visual-spatial tasks. Vocabulary, Similarities, Digit Sequencing and Coding are not pattern tasks in any useful sense. Two of the seven Full Scale subtests, in other words, are fluid reasoning. For how the indexes combine, see our explainer on full-scale IQ versus index scores and the piece on CHC theory.

Do the pattern items carry the most weight in the WAIS-5 score?

They carry the most, but not by much. A loading is the correlation between a subtest and the general factor; squared, it is the share of that subtest's variance that g explains. Canivez and colleagues (2026, Assessment) ran a Schmid-Leiman factor analysis on the WAIS-5 standardisation sample (N = 2,020, ages 16 to 90), which the test's own manual had not done. The loadings for the seven Full Scale subtests:

• Figure Weights .742 (g explains 55% of its variance, the highest of all 20 subtests)
• Matrix Reasoning .700 (49%)
• Block Design .681
• Vocabulary .676 (46%)
• Similarities .663
• Digit Sequencing .632
• Coding .564

The two fluid reasoning subtests top the list, and Matrix Reasoning is above Vocabulary. But the spread is small: the seven loadings average .67, and Vocabulary sits .02 below Matrix Reasoning. Across all 20 WAIS-5 subtests g accounted for 40.4% of total variance and 72.0% of the common variance, and the authors found that the exploratory analysis did not support separate Visual Spatial and Fluid Reasoning factors; the two merged into one perceptual reasoning factor, with measurement dominated by g. They conclude that the Full Scale IQ is best read as an estimate of g. That also means "pattern recognition" cannot be cleanly separated from spatial reasoning in this test.

Did the WAIS-IV give a different answer?

Yes, and that is the useful part. Canivez and Watkins (2010) analysed the WAIS-IV normative sample (2,200 people aged 16 to 90) the same way, and there Vocabulary loaded .70 on g against .66 for Matrix Reasoning:

• Arithmetic .74, Vocabulary .70, Digit Span .68, Similarities .67
• Block Design .66, Matrix Reasoning .66, Information .65, Visual Puzzles .64
• Coding .58, Symbol Search .54

Squared, Vocabulary explained 49% of its variance through g against 44% for Matrix Reasoning. Across the ten core WAIS-IV subtests, g accounted for 42.9% of total variance and 67.0% of the common variance. The ranking of subtests changed between editions, Matrix Reasoning moving from below Vocabulary to above it, which tells you that a loading is a property of a whole battery and a sample, not of an item type. Arithmetic, the top WAIS-IV loader, is not among the WAIS-5's seven Full Scale subtests, although it loads .72 in the WAIS-5 data.

Is the Mensa IQ test just pattern recognition?

Mensa's own page says only that its admission test takes one to two hours and that a qualifying score means testing at or above the 98th percentile of the general population; it does not describe the items, so we cannot say they are patterns. What Mensa does say is that nearly two-thirds of members join by submitting scores from prior testing, and that it accepts about 150 standardised tests on one page and about 200 on another. The accepted list includes the matrix tests (Raven's, the NNAT, Cattell) but also the Wechsler scales, the Stanford-Binet, the Woodcock-Johnson cognitive battery, the GRE and LSAT, and tests with verbal and quantitative content. A pattern-only test is enough to qualify, but it is not required. Our guides to Mensa's score requirement and the Mensa requirements by test give the cut-offs.

Which IQ tests really are all patterns?

Raven's Progressive Matrices is built from grids with a missing piece, and so is the Naglieri Nonverbal Ability Test (NNAT3), whose Levels E to G have 48 items per form. The NNAT3 manual describes the task as looking at a pattern that has a missing section, understanding the relationships among the parts and choosing which of five options fills the gap, and says the thinking is essentially the same as that needed for a verbal analogy such as girl is to woman as boy is to what. Many culture-fair tests lean on such items because they need little language; see our guides to the Cattell Culture Fair test and culture-fair tests. These tests are the closest thing to "just patterns" that exists, which makes what they do and do not measure the real test of the claim.

What about Raven's, the closest thing to a pure pattern test?

Gignac (2015, Intelligence) tested the idea directly with a bifactor model across several large samples. Raven's shared about 50% of its variance with g and about 10% with a fluid reasoning factor beyond g, and about 25% was reliable variance specific to Raven's itself. His conclusion was that Raven's is "not a particularly remarkable test with respect to g," and that researchers should not rely on it alone if they want a valid estimate of g; he suggested a four-subtest Wechsler combination with a g validity of .93 and a 14-minute administration time. Matrix Reasoning's .70 loading in the WAIS-5 data, which squares to 49%, is in line with that 50% for Raven's, though the samples and models differ. What separates people on the Advanced Progressive Matrices, according to Carpenter, Just and Shell (1990, Psychological Review), is the ability to induce abstract relations and to manage a large set of problem-solving goals in working memory, so even this task is more than spotting a pattern; our note on working memory and reasoning covers the overlap. Our explainer on Raven's Progressive Matrices scores covers what its raw score means.

What is the best case for "yes"?

At the level of factors, fluid reasoning and g are very nearly the same thing. Kvist and Gustafsson (2008, Intelligence) found the relationship between g and fluid reasoning was 0.83 when all subjects were treated as one group and close to unity within each of three subgroups. In the WAIS-IV technical-manual data, a higher-order model put g to Fluid Reasoning at 1.0, and after adjustments .99, a sign of empirical redundancy, according to the WAIS-5 analysis that reviews those results; the same kind of model for the WAIS-5 produced an out-of-bounds path of 1.004 from g to Fluid Reasoning. So when researchers say matrices are the cleanest measure of g, they have evidence. The reply is about the test, not the factor: a Full Scale IQ samples seven kinds of task, every one of which loads .56 or more on g, and a measure of the factor is not the same as the whole of what the factor contains.

Where does the pattern-recognition idea come from?

It has a serious pedigree. Charles Spearman identified g in 1904 and, in his 1923 book The Nature of "Intelligence" and the Principles of Cognition, set out principles of cognition including the eduction of relations, seeing the relation between two things, and the eduction of correlates, producing the related item for a given item and relation. His own example of the second is verbal: given the ideas "good" and "opposite to", you obtain "bad". So the root idea is general reasoning, not figural puzzles, even though matrix and analogy items are built to isolate that operation. Our blog post on fluid versus crystallized intelligence explains why a test needs both kinds of task, and is IQ debunked sets this reduction beside the other popular ones.

Is this the same question people ask about AI?

Largely, yes. The ARC-AGI-2 benchmark (Chollet, Knoop, Kamradt, Landers and Pinkard, arXiv 2505.11831, submitted May 2025 and revised in January 2026) presents unique, novel grid tasks that need minimal prior knowledge, are accessible to people and are hard for AI systems. Its authors frame it as a test of general fluid intelligence. Our reports on whether AI can pass an IQ test on ARC-AGI-3 and on a g factor for AI models follow how models perform, and why a benchmark score is not an IQ. On the human side, whether lacking mental imagery changes performance on such items is covered in does aphantasia affect IQ.

What does this mean for practising with puzzles?

Practice raises scores on the same kind of item, for pattern puzzles as for others. Scharfen, Peters and Holling (2018, Intelligence) combined 122 studies and found a retest gain of 0.33 of a standard deviation from first to second try across cognitive tests (95% interval 0.28 to 0.38), 0.32 for figural content such as matrices and 0.33 for verbal content, with the gain 0.37 on an identical form and 0.23 on an alternate form, and no further gains after the third administration. Hausknecht and colleagues (2007) found .26 across 107 samples and 134,436 people. A fluid test is therefore not immune to practice. Our practice effect piece and the review of brain training apps describe the limits, and the matrix puzzle with one rule you can say out loud shows the method. Because a Full Scale IQ mixes verbal, spatial, memory and speed tasks, a pattern-only habit rehearses a fraction of what a complete test samples. To see how a mixed test is put together, read how IQ tests work; to see where a score falls, use the IQ percentile calculator; and for a baseline of your own, take the IQ Metrics IQ test.

The bottom line: pattern recognition is a real and central ingredient of fluid reasoning, and fluid reasoning is the closest thing to g that a single task type measures. It is still not what an IQ score is made of. Two of seven Full Scale subtests, five more that load .56 to .68 on g, a Raven's test that shares about half its variance with g, and a Mensa route that accepts verbal and mixed tests all point the same way. Our guide to how many words the average person knows looks at the vocabulary side in detail, and the Cognitive Reflection Test, which checks a first answer instead of a pattern, is covered in our guide to the CRT.

Common questions