Raven's Progressive Matrices: what the test is, how it is scored, and how to practise the format

What the test is

Lionel Penrose and John C. Raven described a new series of perceptual tests in the British Journal of Medical Psychology in 1936, and the standard form was published in 1938. That standard form, the Standard Progressive Matrices, has five sets of twelve puzzles presented in black and white, 60 in all. Each puzzle is a pattern with one piece missing, and the test-taker chooses which of several pieces completes it. The puzzles become progressively more difficult. Two companion forms exist: the Advanced Progressive Matrices, a set of 36 more difficult items, and the Coloured Progressive Matrices.

The test depends very little on language. Researchers describe the standard form as a test of abstract reasoning and a nonverbal estimate of fluid intelligence, and it often appears in clinical assessment batteries and research studies. A well-known analysis of how college students solve the advanced form, by Carpenter, Just and Shell (1990), found that everyone used the same incremental strategy, working out the regularities in each puzzle a piece at a time. What separated higher from lower scorers was mainly the ability to induce abstract relations and the ability to manage a large set of problem-solving goals in working memory.

How it is given and scored today

The current edition, Raven's 2, was published in 2018. Its publisher's product page gives an age range of 4 to 90; a 2019 flyer for the same edition said 4 to 85. It is sold at qualification level B, meaning to buyers with the professional credentials the publisher requires. It can be given on paper or on screen, to one person or a small group. The administrator receives a raw score and a percentile rank based on norms collected from the general population.

The test can be given with or without a time limit. Bilker and colleagues (2012), who built two shortened forms of the standard test, note that the standard and advanced forms can be given as speeded tests, and cite an average of 17 minutes for a computerised version of the 60-item standard form.

Why a free copy online is not the test

The puzzles, the norms and the manual are all part of what is sold, at qualification level B. A site that offers a free version under the name is doing one of two things. It may be using original puzzles in the same format, which is what we do. Or it may be copying published items, which needs the publisher's written permission.

Our guide to which psychological tests may be reproduced quotes the publisher's own wording on that.

Either way, what makes a score on the real test mean something is missing. A percentile rank comes from comparing your raw score with the norm sample the publisher collected, under the conditions the manual sets out. A copied puzzle taken in a browser, with no norm sample behind it, can tell you how many puzzles you solved. It cannot tell you where that places you.

How a matrix puzzle works

A matrix puzzle is a grid of figures with one cell missing. Something changes in a regular way along each row and down each column: a shape turns, a count grows, a shading darkens, or lines are added or cancelled. You work out the rule, apply it to the empty cell, and only then look for the option that matches.

Our guide how to solve matrix reasoning problems walks through three examples step by step, with every rule named.

The flaw that lets a matrix test answer itself

A matrix test is only fair if you cannot find the answer without looking at the grid. That sounds obvious. It is easy to get wrong.

What researchers found. RAVEN is a research dataset of 70,000 matrix puzzles used to test machine reasoning; it is not the published test. In 2021, Hu and colleagues showed that its answer options gave the answer away. Every wrong option had been made by changing one feature of the right one. So the right answer was simply the option with the most common value of each feature, and it could be found without looking at the puzzle at all. A model shown only the eight options, never the grid, picked the right one 94.2% of the time. When the authors rebuilt the options so that every feature value was shared evenly, the same model's score fell to 14.2%, close to the 12.5% that guessing among eight gives.

What we found in our own puzzles. We ran the same check on ours. On 25 September 2026 the rule "pick the option that shares the most with the others" found the answer in 91–99% of the five-option matrix and series puzzles we built the old way, against a chance rate of 20%. Across the 205 answer sets on the eight tests of ours that use figure puzzles, it scored 62.5% where guessing would score 21.0%; on our 20-grid matrix reasoning test the single most typical option was the right answer on 18 of the 20 grids.

What we changed. The answer options are now built as a small balanced grid around the right answer. Two or three features each take the right answer's value plus one to three others. Where the number of options allows, every value appears equally often; where it does not, the options are a random draw from that grid that never looks at which one is right. On all but 3 of the 205 answer sets the right answer is no longer the single most typical option, and none of the three option-only rules we test picks it out more often than guessing would. On the same 205 answer sets the shortcut now scores 21.0%, and on the matrix test no grid has a single most typical option that is the right answer. A sampled version of the check runs before any change to the puzzles is published, so a change that let the options give the answer away again would be caught first.

What is still open. Balanced options do not remove every clue. A simple statistical model trained on our puzzles still scored 30.1% on 123 of the live answer sets, where guessing scores 20.8%. It did so because some kinds of puzzle fix where the answer sits: in a progression, for example, the missing cell is usually the last step. Removing that means changing the puzzles themselves, not just their options, and that work is not done yet.

How to check any matrix test yourself

Cover the grid and look only at the options. If one option shares a feature with nearly every other one, while the others differ from each other in only one way, be suspicious: you may be able to pick it without reasoning at all. A fair test survives this check. Practising on one that fails it teaches you to scan options, which is the opposite of what the real test rewards.

Practise the format

Our matrix reasoning test: 20 untimed grids covers five kinds of rule: fill, count, rotation, lines that combine and lines that cancel. The answer review names the rule behind each one.

For the same kind of grid against a clock, the timed 35-puzzle matrix practice runs 35 in 25 minutes.

None of these puzzles comes from Raven's Progressive Matrices. They are original, they report how many you solved rather than a percentile, and they are not a substitute for the real test.

MyTestAtlas is not affiliated with or endorsed by Pearson, which publishes Raven's Progressive Matrices and owns its trademark.