How to Check Whether a Visual Reasoning Rule Fits the Whole Grid

What does a whole-grid explanation have to do?

A candidate rule explains a visual matrix when it describes the relevant evidence already shown and predicts what belongs in the blank. A resemblance is only a starting observation: name the visible feature, state how it changes or relates across panels, and say what that relation requires in the missing panel. In “What One Intelligence Test Measures: A Theoretical Account of the Processing in the Raven Progressive Matrices Test,” the authors analyze Raven problems as matrices governed by relations whose constraints converge on a missing entry. That makes the proposed rule answerable to the displayed item rather than to one eye-catching pair. A rule that fits the examples may still not be the only complete description; fit and uniqueness are separate questions. This is a way to test an explanation, not a universal algorithm: items can differ, and the paper’s analysis concerns Raven materials. Its prediction can then be compared with the proposed answer, so the rule has a consequence that could fail against the item.

Sources: What One Intelligence Test Measures: A Theoretical Account of the Processing in the Raven Progressive Matrices Test

How can one striking match hide an uncovered feature?

A useful distinction is between a true observation and a complete account. An observation can be accurate while covering only one part of what changes. To notice the gap, describe features separately before deciding that one proposed relation explains the whole pattern. This is an informal reasoning aid, not a validated scoring method.

Consider this invented verbal illustration, not a test item. Imagine a sequence of imaginary postcards in which a small blue circle moves from the left side of each card to the right, while the number of circles also changes: one, then two, then three. The observation “the circle moves right” may capture a real positional relation. But if the proposed rule says only that position shifts right, it says nothing about why the count changes. Conversely, “add one circle each time” accounts for quantity but leaves the movement open. The distinction is useful even if the two features happen together in the example: the position claim concerns where marks are, while the count claim concerns how many there are.

Write down the observation in ordinary terms, then state the relation being proposed. Ask which feature that relation covers and what visible remainder it leaves unexplained. In this illustration, the position relation covers direction of movement; the increasing-count relation remains. A fuller description would need to account for both, or give a reason that one is irrelevant to the item’s structure. Keeping the partial observation is better than discarding it: it may be one component of a successful account. The error is silently promoting that component into a complete explanation without checking the rest of the evidence.

This separation is a way to make reasoning inspectable, not proof that every feature can be isolated cleanly. In an actual visual design, attributes may interact: a rotation can change apparent position, or a transformation can affect both shape and count. The record captures what a proposed account says about the item; it does not certify that the observer’s feature labels are exhaustive or objectively independent. “What One Intelligence Test Measures: A Theoretical Account of the Processing in the Raven Progressive Matrices Test” describes Raven examples involving multiple relations and distractors that satisfy incomplete subsets. That supports the distinction between a partial match and broader coverage in those materials, while leaving the postcard example and this accounting exercise as original illustrations rather than empirical findings.

In practice, this distinction also makes a proposed correction easier to evaluate. If a later panel breaks the movement pattern but preserves the count pattern, the two claims do not fail together; each can be reconsidered on its own terms. That does not prove the puzzle intended separate rules. It simply prevents one successful observation from shielding an unexplained change from scrutiny.

Sources: What One Intelligence Test Measures: A Theoretical Account of the Processing in the Raven Progressive Matrices Test

When can rows and columns contribute different constraints?

No single relation has to govern every direction in every matrix. A row can constrain one attribute while a column constrains another, provided the proposed conditions can both hold in the same missing panel. In “A Bayesian Model of Rule Induction in Raven's Progressive Matrices,” the model represents row and column rule sets separately for encoded Raven items and evaluates the feature values implied for the missing object. That makes direction-specific relations a model-supported possibility, not a universal instruction for solving every visual puzzle.

Consider an invented verbal illustration, not a puzzle: descriptions of a set of display cards say that moving across each row changes the material named for an object, while moving down each column changes the number of labels printed beside it. If those statements concern different attributes, they can impose distinct conditions on one card. The unseen card would need a material consistent with its row and a label count consistent with its column. No particular answer or complete sequence is needed to see the logic: each direction supplies a condition, and compatibility asks whether both conditions can be true of the same missing panel.

This is different from listing several visible features within one directional relation. The question here is whether evidence organized across one direction and evidence organized across another can coexist at the blank. A solver can state the row-derived condition in words, state the column-derived condition separately, and compare their demands. If both can be satisfied together, the two-rule account remains possible. If they demand mutually exclusive descriptions of the same attribute—for example, one requires a single label while the other requires two—the account as stated is incompatible. It cannot explain that item without changing a rule or revising the assumed attribute description.

It also helps to notice whether the two directions constrain separate attributes or make claims about the same one. Separate attributes can often coexist without affecting each other, as in the card illustration. When both directions constrain the same attribute, the account needs a single description that satisfies both sets of demands; matching one direction while violating the other is not joint compatibility. This is a logical check on the proposed descriptions, not a demand that every item expose two cleanly separable rules. The item itself must support treating those observations as directional constraints.

An incompatibility is a result about the candidate account, not proof that the puzzle itself is defective. The solver may have misread a feature, selected an overly narrow relation, or assumed that each direction carries a rule when the item is organized differently. Conversely, a jointly satisfiable pair is not by itself proof that the account is intended or unique; it only passes this compatibility check. Little and colleagues' model shows that separate row and column rule sets can be represented for its Raven materials, while other items may use a shared transformation or a different organization. Treat direction-specific rules as a possibility to test against the item, not a template to impose in advance.

Sources: A Bayesian Model of Rule Induction in Raven's Progressive Matrices

How should competing explanations be compared?

When two accounts fit what is visible, compare what each explains and what each still leaves open; a neat story is not enough. An editorial comparison aid—not a validated score—can keep the judgment concrete: coverage asks which displayed relations the account explains; exceptions asks whether it needs a special adjustment for one panel; prediction asks what it requires at the blank; and unresolved alternatives asks whether another account also meets those demands. These questions do not mechanically identify an author's intended rule. They show where one explanation is better supported by the item and where uncertainty remains.

For an abstract illustration, imagine that a set of labels is arranged in a repeating sequence. Account A says the order follows a short cycle. Account B says the same labels follow a longer route but pauses at one point to explain an apparent irregularity. This example supplies no particular sequence to solve; it illustrates what to compare. Does each account cover every observed transition? Does one need a special exception that applies nowhere else? What does each predict about the unshown continuation? Does the evidence distinguish their predictions, or could both still fit? A short account may be preferable if it covers the same evidence and predicts the same blank without a patch, but shortness alone cannot settle the case.

The paper “A Bayesian Model of Rule Induction in Raven's Progressive Matrices” formalizes a related comparison computationally. It represents candidate rules, assigns priors, and evaluates how well candidate descriptions fit features encoded from Raven items; the resulting predictions can then be considered alongside response-option similarity. In that framework, rule preference is not simply a reader saying that one explanation feels simpler. The model's posterior depends on the candidate representations and prior probabilities as well as the fit to the encoded item features. A practical reader's preference for fewer exceptions is an editorial reasoning aid, not a calculation of that posterior and not a substitute for the model's specified assumptions.

The model also has boundaries that matter when borrowing its vocabulary. Its candidate rules, priors, and item features are choices made in constructing the model, and the implementation hand-codes features for its selected Raven materials. If a feature or plausible rule is absent from that representation, the calculation cannot compare it. The paper therefore demonstrates one formal way to compare encoded alternatives; it does not establish a validated procedure for human solvers, a universal ranking of all possible accounts, or the intended authorship of a particular puzzle. A model can be internally explicit while still depending on what researchers chose to encode.

For your own comparison, treat the four questions as prompts for inspection, not points to total. If one account needs a one-off exception, ask whether that exception is independently supported by the displayed structure or was introduced only to rescue the account. If its blank prediction differs from a rival's, identify what evidence would discriminate between them; if the item offers none, report the ambiguity instead of inventing a tie-breaker. Parsimony helps resist adding a new clause every time evidence resists a preferred explanation. It is a disciplined preference among accounts with comparable coverage, not proof that the shortest account is correct or that a puzzle has only one defensible reading.

Sources: A Bayesian Model of Rule Induction in Raven's Progressive Matrices

What can answer choices confirm, and what can they not?

A plausible answer choice can test a rule you have already stated, but choosing it does not by itself show that the rule explains the grid. In “What One Intelligence Test Measures: A Theoretical Account of the Processing in the Raven Progressive Matrices Test,” Carpenter, Just, and Shell describe a response process in which a solver forms a candidate completion from relations in the matrix and then matches that prediction to an available alternative. Their analysis also explains why a distractor can look convincing: it may satisfy only part of the set of constraints. That account concerns Raven material and a theoretical task analysis; it is not evidence that every multiple-choice puzzle was built or solved in exactly this way.

Consider an abstract illustration, not a test item. Suppose your proposed account predicts that a missing panel should contain three marks arranged in a rising line, while also requiring the marks to alternate between filled and open. One option has three marks in a rising line, but all are filled. It matches the count and arrangement yet leaves the alternation condition unsatisfied. Its resemblance is informative because it shows which part of your prediction it captures; it does not confirm the whole account. A different option might preserve alternation but have the wrong count. Comparing choices against each stated feature can reveal these partial matches and expose which prediction remains unmet.

This is why response selection and rule evaluation are related but distinct steps. First make the candidate account specific enough to predict relevant properties of the blank. Then inspect whether an option meets those predictions, whether it contradicts any, and whether the choices expose an assumption you had not made explicit. If the only reason an option seems right is that it resembles one visible panel, that is a cue to revisit the explanation. If it fits the full prediction, the match supports that candidate as an answer within the available set, though it cannot establish that the candidate rule is uniquely intended when another account also fits.

Elimination remains useful: a choice that violates a required feature can be ruled out, and comparing alternatives may clarify a prediction. The caution is against letting the list generate a story after the fact, where each attractive resemblance gets a new exception. The Raven analysis gives a concrete reason to check coverage because some distractors correspond to incomplete rule sets; it does not justify a blanket instruction to avoid looking at options. Across tasks, choices can play different roles, so use them as evidence about a stated prediction rather than as a substitute for one.

Sources: What One Intelligence Test Measures: A Theoretical Account of the Processing in the Raven Progressive Matrices Test

What does eye-tracking evidence say about checking a grid?

Eye-tracking research offers evidence about where groups of solvers looked while completing matrices, but it does not turn a gaze pattern into a personal recipe. The indexed abstract for “The Development of Relational Reasoning: An Eyetracking Analysis of Strategy Use and Adaptation in Children and Adults Performing Matrix Completion” reports a study of 6-year-olds, 9-year-olds, and adults. Across those groups, indicators of scanning across rows and columns were associated with better overall performance. This pattern is consistent with the practical value of attending to relations across the whole matrix: a solver who only inspects one appealing pair may fail to notice a constraint expressed elsewhere. The result is an association in this task, however, not proof that scanning caused higher scores or that the same movement will help every person on every item.

The abstract also reports that quicker and more frequent consultation of answer options was associated with poorer performance. That finding should be read alongside the scanning result, not as a rule that options are harmful. Eye movements record visual attention, not a full account of thought: a look toward an option cannot tell us whether someone is checking a carefully stated prediction, searching for a cue, or simply moving their eyes between regions. Nor does a group-level association explain why one participant answered a particular item incorrectly. The study supports treating systematic attention to rows and columns as relevant behavior; it cannot diagnose an individual's reasoning from the direction or duration of a glance.

The study's reported strategy patterns also changed with item difficulty. As tasks became harder, the relation between observed strategy indicators and performance was not simply fixed across all conditions; the abstract describes adaptation in strategy use. This matters because a useful approach may depend on what an item demands. An easy matrix may permit a quick check of a salient relation, while a harder one may call for more deliberate comparison across directions or closer evaluation of candidate responses. The evidence does not specify a universal threshold for switching methods, and it does not show that every solver should adopt the same sequence. It does caution against treating one observed behavior as uniformly good or bad regardless of difficulty.

There is also a limit on how specifically these findings can be summarized: the study's full text was inaccessible during research, so the evidence available here is its indexed abstract. Claims should stay at that level—participant groups, the reported group associations, and the stated change with difficulty—without adding unverified measures, effect sizes, or explanations. For an individual reviewing a missed matrix, the findings can prompt a modest question: did the attempt include attention to both rows and columns, and did option checking follow a prediction or replace one? Those questions are reflection prompts, not an inference that a particular gaze pattern caused the miss. The study describes tendencies in its sample and task; it cannot tell an individual what strategy will work best on a different matrix.

Taken together, the observations support a restrained conclusion: whole-grid scanning is associated with stronger performance in the studied groups, while frequent, fast option consultation is associated with weaker performance, and strategy indicators vary with difficulty. Association does not establish that changing gaze will change accuracy. Option use also has a legitimate checking role once a solver has a candidate prediction, so the abstract cannot support a categorical ban. The useful application is to keep the behavior tied to the reasoning question: inspect whether the proposed relation accounts for the matrix, then use choices to check its consequences. That distinction makes the study relevant without claiming that eye tracking has settled an individual's best method.

Sources: The Development of Relational Reasoning: An Eyetracking Analysis of Strategy Use and Adaptation in Children and Adults Performing Matrix Completion

Why does a wrong answer fail to reveal its own cause?

A missed item records an outcome: the selected response did not match the scored answer. By itself, it does not reveal the mental step that produced it. A solver might have overlooked a feature, applied a relation incorrectly, predicted the wrong missing panel, or made a selection error after forming a sound prediction. Those routes can end at the same observed answer. To infer an individual's process, one would need evidence about the intermediate reasoning, not just the final response.

The primary study “Using Cognitive Diagnostic Models to Evaluate the Two-Process Theory of Matrix Reasoning” compared two formal accounts using a 26-item figural-matrix test. Its sample comprised 633 high-ability volunteers aged 13 to 58, who completed paper-and-pencil or online administrations. The researchers compared an additive linear logistic test model (LLTM) with a multiplicative linear logistic sequential design model (LSDM). In this test and sample, the additive LLTM reconstructed item parameters better. These details describe the study's task, participants, administration modes, model comparison, and reported result; the available indexed record does not support adding unreported effect sizes or further procedural detail.

The distinction between a model-level result and a personal explanation matters. The models describe how specified components account for item difficulty across the test. Better reconstruction by one account therefore favors that account for the studied material under the comparison made. It does not show that the winning model has read a particular person's thoughts, nor does it connect one wrong response to one omitted relation. An additive account can fit aggregate item parameters better while individual solvers still arrive at errors through different routes. The response alone underdetermines which route occurred. Item parameters summarize how items behave in the model's analysis; they are not labels assigned to a person's particular mistake. Moving between levels would require additional process evidence.

Nor does this comparison establish that components never interact. It tests specified additive and multiplicative accounts on one test and a selected high-ability volunteer sample; other tasks, populations, or model specifications could produce a different comparison. The study does not universally disprove interaction or working-memory explanations. The indexed record supports a bounded conclusion about which model reconstructed item parameters better there, not an unrestricted theory of matrix reasoning.

For a person reviewing a miss, the defensible statement is consequently narrow: the answer was wrong, while its cause remains unknown unless separate evidence identifies the process. A wrong answer cannot tell you whether you failed to notice a feature, misapplied a relation, made an inaccurate prediction, or selected the wrong option despite a good prediction. The model result explains why aggregate difficulty accounts should not be mistaken for individual error diagnoses; it does not supply a diagnostic key for any one response.

Sources: Using Cognitive Diagnostic Models to Evaluate the Two-Process Theory of Matrix Reasoning

What can practice on reasoning rules transfer to?

Improvement after practice should be named at the level actually tested. It may mean learning the trained material; it may also extend to new combinations or rules within a related task. Neither result alone establishes a lasting gain across unrelated reasoning tasks or a rise in general intelligence. For that stronger conclusion, evidence would need to test outcomes farther from the training task and establish what persists over time. The distinction is about transfer distance: success on a practiced operation is closer to training than success on a different task or a broad ability measure.

The primary study “Knowledge Representations: Individual Differences in Novel Problem Solving” examined training on figural analogies. Participants learned individual rules and then completed multi-rule analogy problems, including a condition with new rules. The final analytic sample was 220. The indexed study record reports that training improved performance on the study task and that working-memory measures contributed to transfer findings. Thus the result goes beyond repeating only the individually trained rules: performance was examined on analogies that combined rules and included novel-rule material. The available record does not warrant a more precise effect estimate or a claim that every participant transferred equally.

That is evidence of learning and transfer within a related figural-analogy task, not proof of broad transfer. A multi-rule analogy still shares task family and representational demands with the training. The finding does not test the row-and-column verification approach discussed in this article, and it cannot establish that using such a check sequence improves matrix scores. Nor does it show persistence beyond the study's reported testing, improvement on unrelated cognitive tasks, or an increase in general intelligence. The working-memory finding is part of the study's account of transfer; it is not evidence that training raised working-memory capacity or general ability.

The practical implication is to describe progress precisely. If someone practices a particular relation and later handles it more accurately, that supports learning on that operation. If they also solve fresh examples that combine familiar relations in new ways, that is a useful sign of near transfer within the related task. To claim more, one would need evidence from more distant outcomes and later follow-up; neither is established by the indexed study details. This boundary preserves the useful result without turning task improvement into a claim about a person's underlying general intelligence.

As an informal, unvalidated way to make practice interpretable, keep a short note after a session: name the operation practiced, then mark whether a later example used a familiar arrangement or a new combination of rules. This is a record-keeping suggestion, not a tested intervention or a method shown to increase transfer. It can help distinguish success on familiar material from success on a related variation when reflecting on one's own work. The analogy study makes that distinction worth tracking, but it did not test this note-taking suggestion or the article's checking sequence.

The counterpoint is central: an analogy task is not matrix verification. Even the novel-rule condition remains inside a related task, and indexed access limits what can safely be said about the study's details. The result cannot establish long-term persistence, transfer across all reasoning formats, or an increase in general intelligence. Practice can be worthwhile for understanding and performing the operations practiced; broader claims require broader evidence.

Sources: Knowledge Representations: Individual Differences in Novel Problem Solving

What should the reader do with this reasoning practice?

Choose the next step by what you need the result to do. If you are curious or want low-stakes practice with patterns, number relationships, and spatial reasoning, take the free 50-question test at /test, then use its worked answers to examine how a proposed rule accounts for the information shown. Its provisional estimated IQ range comes from an unvalidated conversion with no validated adult reference population; it is not a normed adult comparison or a confidence interval. The product page, “Test IQ Free: Free IQ Assessment, 50 Adaptive Questions, Estimated IQ Range,” describes the 50-question assessment, while the saved product specification sets this measurement boundary.

The optional $9 report may be useful if you want a fuller explanation of the same result, examples, and a structured practice plan. It uses the same provisional conversion, so payment adds explanation, not validity. Treat any practice as work on particular reasoning operations; this result and the report do not establish a change in general intelligence. If an educational, clinical, or employment decision depends on an ability estimate, consult a qualified professional who can select and interpret an appropriate assessment in context.

Sources: Test IQ Free: Free IQ Assessment, 50 Adaptive Questions, Estimated IQ Range

Questions readers ask