Bongard Problem¶

A concept-learning puzzle that presents positive and negative diagram sets and asks for a rule true of every positive case and no negative case.

Version: v1 · 2026-09-28 · History
Domain-specific #: 8242
Domain group: Applied Sciences & Engineering
Origin domain: Computer Science & Software Engineering
Subdomains: Pattern Recognition, Concept Learning → Computer Science & Software Engineering

Core Idea¶

A Bongard problem is a small contrastive concept-learning puzzle. Two panels contain relatively simple diagrams: every diagram on the positive side shares an intended property, and every diagram on the negative side lacks it. The solver must formulate that distinction as a convincing general rule.

The challenge is underdetermination. Many accidental features can separate a finite sample, so the positive set must show variation and the negative set must contain near misses. Solving requires inventing a visual vocabulary, testing hypotheses against both classes, and preferring a coherent rule that explains why each example was included.

The format originated in pattern-recognition research and now serves as a benchmark for few-shot concept learning and visual reasoning. Success is not mere image classification: the rule should be expressible, compositional where necessary, and capable of classifying new diagrams drawn from the same concept family.

How would you explain it like I'm…

Secret Rule Picture Puzzle

A Bongard problem is a picture puzzle with two sides. All the drawings on one side share a secret rule, and none of the drawings on the other side follow it. Your job is to figure out the secret rule, like "all these shapes are closed" or "all these have three corners."

Spot the Rule Between Two Sides

A Bongard problem is a puzzle made of two groups of simple drawings. Every drawing on the left has some shared feature, and every drawing on the right is missing it. You have to figure out the rule that separates the two sides and say it clearly. It's tricky because lots of small details might happen to differ between the sides, so a good puzzle includes drawings that are very different on the same side and drawings on the other side that are almost the same but just miss the rule. A real solution should work for new drawings too, not just the ones shown.

Contrastive Concept-Learning Puzzle

A Bongard problem is a small puzzle for learning a concept by contrast. It shows two panels of simple diagrams: every diagram on the positive side has some intended property, and every diagram on the negative side lacks it. The solver must state that property as a general rule. The difficulty is underdetermination: with only a few examples, many accidental features could separate the two sides, so well-designed problems vary the positive examples and include near misses on the negative side. Solving one means inventing a vocabulary for describing the pictures, testing guesses against both sides, and choosing the rule that best explains every example. It began in pattern-recognition research and is now used as a benchmark for few-shot concept learning and visual reasoning, where the goal is a rule that can also classify new diagrams, not just sort the given ones.

A Bongard problem is a contrastive concept-learning puzzle consisting of two panels of relatively simple diagrams: all diagrams on the positive side share an intended property, and all on the negative side lack it. The solver must formulate the distinguishing property as an explicit general rule. The core difficulty is underdetermination: with a finite sample, many accidental features could separate the sets, so good problems vary the positive examples and include near misses among the negatives to rule out spurious hypotheses. Solving requires inventing a suitable visual vocabulary, generating and testing hypotheses against both classes, and preferring a coherent rule that explains why each example was included. The format originated in pattern-recognition research and is now a benchmark for few-shot concept learning and visual reasoning. Success is not mere image classification: the rule should be expressible, compositional where needed, and able to classify new diagrams from the same concept family.

Structural Signature¶

Sig role-phrases:

positive set. Provides varied diagrams that all instantiate the target concept. Constitutive evidence class. If altered: One positive example cannot establish which feature is intended rather than incidental.
negative set. Supplies contrast cases that exclude overbroad hypotheses. Constitutive counterevidence. If altered: Without negatives, many incompatible descriptions fit the positives.
visual vocabulary. Determines available objects, relations, transformations, and compositional features. Necessary representation frame. If altered: A solver unable to represent the intended relation cannot express the concept.
candidate rule. States a compact property that classifies every shown diagram. Identity-bearing inference. If altered: A rule with exceptions or index-based memorization fails the task.
contrastive validation. Checks the rule against all positives, all negatives, and plausible alternative explanations. Constitutive acceptance test. If altered: A coincidental rule may fit the finite panels without capturing the intended concept.

What It Is Not¶

Not spot-the-difference. The task seeks one class rule across several diagrams, not local changes between two pictures.
Not odd-one-out. There are two example sets rather than one anomalous member.
Not memorization. A position- or item-specific lookup does not formulate the shared concept.
Not any binary dataset. The small visual panels are designed for explicit concept induction and explanation.

Scope of Application¶

The format applies to pattern-recognition research, cognitive studies, education, and benchmarks for explainable few-shot reasoning.

Human concept learning. Studies hypothesis formation from sparse contrasts.
Artificial intelligence. Benchmarks visual relational reasoning.
Cognitive modeling. Tests representation and analogy mechanisms.
Education. Practices classification and rule articulation.
Puzzle design. Constructs informative positives and near-miss negatives.

Clarity¶

The two-panel structure separates resemblance from definition. A valid rule must classify every example, expose the relation doing the work, and survive negative near misses; merely naming a salient object or aesthetic impression is insufficient.

Manages Complexity¶

A few simple shapes can support a combinatorial number of object, count, topology, orientation, and relation hypotheses. Positive variation and negative contrast prune that space, while explicit rule formulation makes the remaining inductive leap inspectable.

Abstract Reasoning¶

Inventory objects, attributes, counts, and relations without committing to the first salient feature.
Find invariants across all positive diagrams and violations across every negative.
Use near-miss pairs to isolate which relation rather than object identity controls the split.
Prefer the simplest rule that accounts for all panels without panel-specific exceptions.
Test the rule on imagined or held-out diagrams to expose accidental correlations.

Knowledge Transfer¶

The task format transfers literally to new visual domains when positive and negative panels support an all-versus-none rule. Calling any difficult classification a ‘Bongard problem’ is loose analogy unless sparse contrastive examples and explicit concept formulation are central.

Examples¶

Canonical¶

The left panel varies shape, size, and orientation but every large figure contains a small figure; the right panel contains the same object types without containment. ‘A figure contains another figure’ classifies all and only the left cases.

Mapped back: positive set → varied containment cases; negative set → matched non-containment cases; visual vocabulary → figures and inside relation; candidate rule → one figure contains another; contrastive validation → all panels checked.

Applied / In Practice¶

A vision benchmark gives six positive human–object interactions and six negatives differing in role direction. A model must infer ‘person supports object’ rather than memorize object category, then state and apply the relation to held-out scenes.

Mapped back: positive set → support interactions; negative set → role-reversed near misses; visual vocabulary → person, object, and directed support; candidate rule → person supports object; contrastive validation → held-out classification.

Structural Tensions¶

T1: finite fit vs. intended concept. Many rules separate a small sample, but only some capture the designer's generalization. Diagnostic: What new panel would distinguish the leading hypotheses?

T2: perceptual salience vs. relational depth. Obvious shapes can distract from topology, nesting, or role relations. Diagnostic: Does the rule survive changes in object appearance?

T3: human interpretability vs. benchmark difficulty. Harder tasks may demand richer composition while becoming ambiguous or culturally dependent. Diagnostic: Can a competent solver state one defensible all-versus-none rule?

Structural–Framed Character¶

Bongard problem is mixed. The all-versus-none classification relation is formal, while intended visual vocabulary and simplicity judgments are human-framed. It is evaluative as a reasoning benchmark and institutionally rooted in AI and cognitive research. Its character: sparse contrastive evidence designed to elicit an explicit visual concept.

Structural Core vs. Domain Accent¶

Skeletal core. Positive and negative examples constrain a hypothesis that must cover one class and exclude the other.

Domain-bound accent. Diagram panels, visual relations, puzzle design, concept articulation, and benchmark scoring define the format.

Why not prime. Contrastive induction travels, but the Bongard problem is a particular visual task and research artifact.

This entry is a kind of Logic Puzzle.

Contrastive learning. Negative cases make the decision boundary informative.
Induction. The solver generalizes a rule from sparse examples.
No canonical parent edge is asserted in the current DAG.

Relationships to Other Abstractions¶

Current abstraction Bongard Problem Domain-specific

Parents (1) — more general patterns this builds on

Bongard Problem is a kind of Logic Puzzle Domain-specific

Bongard Problem satisfies the defining boundary of Logic Puzzle: A logic puzzle is a deliberately constructed problem that presents entities, states, clues, and explicit or inferable constraints and asks a solver to derive a required configuration, classification, quantity, or explanation primarily through valid deduction and exhaustive consistency rather than hidden factual knowledge.

Hierarchy path (1) — routes to 1 parentless root

Bongard Problem → Logic Puzzle

Neighborhood in Abstraction Space¶

Bongard Problem sits in a crowded region of the domain-specific corpus (39^th percentile for distinctiveness): several abstractions share nearly its structure, so a description that fits it tends to fit its neighbors too.

Family — Visual & Cinematic Composition Techniques (24 abstractions)

Nearest neighbors

Semiorder — 0.88
Logical graph — 0.87
Constructional System — 0.87
Suppressed Correlative — 0.87
Writing system — 0.87

Computed from structural-signature embeddings · 2026-10-08

Not to Be Confused With¶

Raven's Progressive Matrices. Tell: Is the task class separation or completion of a relational matrix?
Odd-one-out puzzle. Tell: Are there two labeled sets or one anomalous item?
Spot the difference. Tell: Is the answer a general concept across cases or a local image change?
Binary image classification. Tell: Must the solver formulate the learned rule from sparse designed contrasts?

References¶

Frozen Wikipedia discovery revision: https://en.wikipedia.org/wiki/Bongard_problem (revision 1352890140).
Preserved source candidate: http://app.ebape.fgv.br/comum/arq/Linhares2.pdf
Preserved source candidate: https://web.archive.org/web/20110811145352/http://app.ebape.fgv.br/comum/arq/Linhares2.pdf
Preserved source candidate: http://www.foundalis.com/soc/why_no_more_Bongard.html
Preserved source candidate: http://aace.org/conf/edmedia
Preserved source candidate: https://papers.nips.cc/paper/2020/file/bf15e9bbff22c7719020f9df4badc20a-Paper.pdf
Preserved source candidate: https://openaccess.thecvf.com/content/CVPR2022/papers/Jiang_Bongard-HOI_Benchmarking_Few-Shot_Visual_Reasoning_for_Human-Object_Interactions_CVPR_2022_paper.pdf
Preserved source candidate: https://openaccess.thecvf.com/content/CVPR2023/papers/Spratley_Unicode_Analogies_An_Anti-Objectivist_Visual_Reasoning_Challenge_CVPR_2023_paper.pdf
Preserved source candidate: http://www.oebp.org

The frozen Wikipedia revision is discovery provenance. The retained source set was reviewed for identity, formal or operational relation, and scope. The encyclopedia's structural synthesis is bounded to those claims; a thin authority surface is recorded as a nonblocking source-strengthening repair rather than concealed.