Visual Analogy Puzzles

• Visual analogy puzzles are abstract visual reasoning tasks that require identifying mapping functions and hidden relational rules across image sets.
• They underpin benchmarks like Raven’s matrices and Bongard problems, offering rigorous tests for AI model generalization and cognitive evaluation.
• Recent advances leverage unified architectures and contrastive learning to enhance rule inference and puzzle generation, addressing compositional and domain-shift challenges.

Visual analogy puzzles constitute a core class of abstract visual reasoning (AVR) challenges, where the solver must identify, extend, or map symbolic or relational structures between sets of images. These puzzles are foundational both in cognitive science and as rigorous benchmarks for artificial intelligence, notably spanning Raven’s Progressive Matrices (RPMs), Bongard problems, Visual Analogy Problems (VAP), and diverse variants in the diagnostics of large vision-LLMs (LVLMs).

1. Formalization and Problem Classes

A visual analogy puzzle is succinctly formalized as a triple $X = \langle I, R, S \rangle$ (Lymperaiou et al., 20 Jan 2026), where:

• $I$ is the set of input images (and possibly text context),
• $R$ is a (possibly implicit) rule set encoding the relations or transformations among elements of $I$,
• $S$ is the finite solution space (e.g., candidate answer images or labels).

The central task is to infer a mapping function $f: A \to B$—typically unknown—that translates relational structure from a known (source) part of the puzzle to an unknown (target) part. Solving the puzzle amounts to evaluating candidates from $S$ and selecting the answer $s_*$ that optimally preserves the underlying analogy, often formalized as minimizing a dissimilarity measure $\Delta( f(A), B )$ with respect to hypotheses $f$ (Lymperaiou et al., 20 Jan 2026). Such puzzles range from panel arrangement tasks (Raven’s matrices), set distinction (Bongard), to mapping high-level relations across domains (REBUS, MARVEL).

2. Representation and Benchmark Construction

Puzzles are typically constructed by algorithmically arranging object instances according to abstract rules, with strict control over compositional structure and perceptual properties (Chia et al., 2024, Basioti et al., 30 Mar 2025). Each instance may involve:

• Demonstration images establishing a relational pattern,
• One or more query items with masked or missing attributes,
• Multiple candidate answers, from which the correct completion must be selected.

Benchmark datasets implement this framework at varying levels of abstraction:

• PuzzleVQA offers 2,000 puzzles across four atomic concepts (color, number, shape, size) and their pairwise combinations, formalizing each instance as $I$0, with robust procedural generation and ground-truth reasoning chains (Chia et al., 2024).
• RAVEN and PGM instantiate RPM-style puzzles as $I$1 grids with diverse, systematically parameterized rule sets (Kim et al., 2020, Basioti et al., 30 Mar 2025).
• Bongard-style datasets organize inputs as positive/negative sets, targeting rule abstraction at a meta-level (Lymperaiou et al., 20 Jan 2026).
• VAP comprises over 700,000 analogy instances requiring domain transfer of Boolean or structural operations (Małkiński et al., 2024).

3. Model Architectures and Learning Paradigms

Unified Relational Architectures

Recent advances emphasize the importance of unified, role-agnostic architectures. The Unified Model for Abstract Visual Reasoning (UMAVR) processes each puzzle as a single image, with spatial relations discovered endogenously (Małkiński et al., 2024). Its core modules include:

• Convolutional encoder: Extracts local features from an arbitrary-sized input image.
• MetaFormer-style global relational blocks: Stackable layers combine token-mixing operations (spatial and channel permutations) and MLP-based mixing for relational abstraction.
• Auxiliary rule-prediction head: Outputs one-hot predictions of abstract rules, crucial for tasks like RPM.

Universal models, such as UMAVR, enable multitask, transfer, and curriculum learning across distinct puzzle formats, challenging the conventional panel-centric designs.

Embedding-Based and Contrastive Methods

The VISALOGY approach leverages a quadruple-Siamese CNN to learn difference embeddings between image pairs, with a double-margin contrastive loss pulling analogous transformations together and pushing apart non-analogous pairs (Sadeghi et al., 2015). The objective is to ensure that the embedding $I$2 for the context and candidate analogies are proximal if the analogy holds, and distant otherwise.

Few-shot and meta-analogical contrastive learning approaches further infuse robustness to data scarcity and domain shift (Kim et al., 2020). Here, contrastive objectives align relational representations across task episodes, and meta-level alignment against different attribute domains yields pronounced improvements, especially on cross-domain and low-shot splits.

Generative and Rule-Informed Models

Generative approaches such as GenVP model not only the puzzle-solving process but also the generation of new puzzles consistent with user-specified rules (Basioti et al., 30 Mar 2025). Hierarchical variational autoencoder architectures disentangle rule-relevant and nuisance latents, explicitly inject rule matrices, and enforce contrastive consistency at both puzzle and rule levels.

4. Evaluation Protocols and Empirical Findings

Standard metrics include (i) accuracy—the fraction of puzzles solved correctly; (ii) consistency—requiring correct answers on all equivalent reformulations per instance; (iii) explanation faithfulness—alignment of model-generated rationales with ground truth (Lymperaiou et al., 20 Jan 2026).

A cross-section of empirical results:

Benchmark	Model	In-Distribution Accuracy	OOD Shift Impact
VAP (nₐ=4)	UMAVR	97.3%	TL to 98.0%
VAP (nₐ=4)	Slot Transformer Scoring	98.5%	—
PuzzleVQA (single-conc)	GPT-4V	46.4%	Human: 91.6%
PuzzleVQA	Claude 3 Opus	39.4%	—
Bongard in Wonderland	GPT-4V	28%	−25–40pp OOD
RAVEN (low-shot, 161)	CoPINet+gen-analog	40.1%	—
GenVP (RAVEN, OOD)	GenVP	94.7% (ID); +20–50pp vs SOTA (OOD)	—

Even for state-of-the-art LVLMs, accuracy under OOD and compositional shifts deteriorates rapidly: e.g., GPT-4V drops from 35% to 12% accuracy on VisualPuzzles with out-of-distribution shape stylings (Lymperaiou et al., 20 Jan 2026), and consistency scores remain below 12% across comprehensive benchmarks.

Critical bottlenecks for neural models include:

• Over-reliance on superficial cues or low-level features,
• Insufficient abstraction in relational mapping,
• Brittle generalization under composition or domain shift,
• A pronounced explanation-execution gap (rationales often plausible but not actually followed in answer selection).

5. Generation and Synthesis of Analogy Puzzles

GenVP inaugurates the capability to synthesize entire RPM-style puzzles and candidate sets under arbitrary user-specified abstract rules (Basioti et al., 30 Mar 2025). The hierarchical generative framework models the composition:

1. Sample rule matrix $I$3,
2. Generate puzzle-level relationship latents,
3. Compose image-level rule-relevant and nuisance latents,
4. Render object arrays consistent with the chosen abstract structure.

This pipeline supports the generation of multiple logically-valid answers per puzzle, OOD configuration sampling, and robust manipulation of puzzle structure. Empirical results demonstrate 75–85% coherence in rule adherence on generated puzzles and 93% logical consistency when generating sets of valid alternatives.

6. Limitations, Bottlenecks, and Future Directions

Contemporary models, including top-performing LVLMs, reveal a persistent gap to human-level abstraction (Chia et al., 2024, Lymperaiou et al., 20 Jan 2026):

• Visual perception errors frequently confound pattern induction,
• Inductive reasoning—inferring the correct rule from minimal demonstrations—is the principal weak link,
• Deductive reasoning is a secondary bottleneck for weaker models,
• Curriculum learning and transfer across puzzle families offer gains but are not universally sufficient for full generalization.

Recommendations for benchmark and model design (Lymperaiou et al., 20 Jan 2026):

• Embed explicit compositionality controls and systematically varied rules,
• Require verifiable intermediate mappings and explanation consistency,
• Explore debate or multi-agent modeling frameworks for candidate verification,
• Develop evaluation protocols emphasizing OOD and compositional robustness,
• Incorporate geometry-aware and abductive puzzle variants with incomplete or ambiguous evidence.

A plausible implication is that bridging the performance chasm will require both stronger visual encoding modules and reasoning architectures with explicit relational abstraction, compositional inductive bias, and context-sensitive transfer mechanisms.

7. Synthesis and Outlook

Visual analogy puzzles serve as a compact, diagnosis-rich substrate for probing abstract reasoning in both humans and artificial systems. Unified views that treat each puzzle instance holistically—as an image to be relationally parsed, rather than a set of discrete, role-labeled panels—enable the development of universal solvers that rival, or even surpass, specialist systems on a wide spectrum of tasks (Małkiński et al., 2024).

Nevertheless, existing multimodal and vision-centric models remain constrained by superficial pattern-matching, limited generalization, and failures of true abstraction. Integrating generative, contrastive, and explicit rule-inference frameworks, alongside rigorous compositional benchmarks, delineates a promising path for both research and diagnostic evaluation in abstract visual reasoning. This domain continues to illuminate essential gaps between current artificial intelligence and the general, adaptable fluid intelligence that characterizes human problem solving.