Tangram Puzzle: Geometry & Applications

• Tangram is a classic dissection puzzle composed of seven polygonal pieces that form target silhouettes without overlap through rotation and reflection.
• It has been rigorously studied in combinatorial geometry, with classification results such as 13 convex figures and 53 simple pentagonal configurations.
• Tangram serves as a versatile testbed for multimodal spatial reasoning, cognitive science experiments, and robotic assembly benchmarks.

Tangram is a classic dissection puzzle in which seven polygonal pieces are arranged to exactly cover a target silhouette without overlap. In formal geometric terms, it is a problem of deciding whether a fixed set of polygons can be placed in the plane so that their union is a given target polygon, with rotation and reflection allowed but scaling disallowed (Fox-Epstein et al., 2014). The standard set consists of five isosceles right triangles, one square, and one parallelogram, and it can be modeled as a partition of sixteen identical right isosceles triangles with side lengths $1,1,\sqrt{2}$ (Pohl et al., 2020). Beyond recreation, tangram has become a technical object in combinatorial geometry, cognitive science, multimodal reasoning, and robotics (Zong et al., 5 Feb 2026).

1. Classical form and geometric representation

The standard tangram set contains two small triangles with legs of length $1$, one medium triangle with legs of length $\sqrt{2}$, two large triangles with legs of length $2$, one unit square, and one parallelogram with side lengths $1$ and $\sqrt{2}$ and acute angle $\frac{\pi}{4}$ (Pohl et al., 2020). A tangram figure is any polygon that can be dissected into isometric copies of these seven tans, with the interiors of the pieces pairwise disjoint. In the usual geometric model, every tangram has area $8$, and when the resulting figure is a simple polygon its interior angles are integer multiples of $\frac{\pi}{4}$ (Pohl et al., 2020).

A standard mathematical reduction treats the seven tans as unions of sixteen identical right isosceles triangles. This reduction is important because it places tangram in a common combinatorial universe with related dissection puzzles such as Sei Shōnagon Chie no Ita, and it explains why tangram figures inherit a restricted set of admissible edge directions and angle patterns (Fox-Epstein et al., 2014). A plausible implication is that much of tangram theory is best understood as a constrained tiling problem on a rotated square lattice rather than as a free-form polygon assembly problem.

Historically, the puzzle is of anonymous origin; one account notes that the first known reference in literature is from 1813 in China (Fox-Epstein et al., 2014). In modern work, the same seven-piece construction is also the canonical substrate for target silhouettes representing animals, people, letters, and objects, which is why tangram remains useful as both a mathematical and cognitive testbed (Zhao et al., 17 May 2025).

2. Enumerative geometry and classification results

The most classical finite result is the convex classification. Sixteen identical isosceles right triangles can form exactly $20$ convex polygons, but the standard seven-piece tangram can realize only $1$0 of them (Fox-Epstein et al., 2014). Those $1$1 convex tangram figures consist of one triangle, six quadrangles, two pentagons, and four hexagons (Pohl et al., 2020). This places the standard tangram below the maximum expressive power of seven-piece dissections of the same sixteen triangles: Sei Shōnagon Chie no Ita realizes $1$2 convex polygons, a constructed seven-piece puzzle realizes $1$3, no seven-piece puzzle realizes all $1$4, and $1$5 pieces are necessary and sufficient to realize all $1$6 (Fox-Epstein et al., 2014).

The pentagonal case is the last major finite classification problem for simple tangram polygons. It has been solved completely: there are exactly $1$7 simple pentagonal tangrams up to isometry, comprising $1$8 convex pentagons, $1$9 non-convex lattice pentagons, and $\sqrt{2}$0 non-lattice pentagons (Pohl et al., 2020). In the lattice case, all vertices lie in one lattice, whereas in the non-lattice case the pentagon can be decomposed into convex pieces whose lattices differ by a rotation of $\sqrt{2}$1 (Pohl et al., 2020). This shows that tangram pentagons are not merely a subset of convex classifications; they expose a deeper distinction between single-lattice and two-lattice constructions.

For $\sqrt{2}$2, the situation changes qualitatively. The families of all $\sqrt{2}$3-gonal tangram figures are either infinite or empty (Pohl et al., 2020). This suggests a sharp structural transition: low-sided polygons admit exhaustive classification, whereas higher-sided polygon families typically admit continuous deformation parameters.

Mathematical class	Count
Convex polygons realizable by 16 identical isosceles right triangles	20
Convex tangram figures	13
Simple pentagonal tangrams	53

3. Tangram in cognition, naming, and conceptualization

Tangram silhouettes are intrinsically ambiguous, and that ambiguity has made them a long-standing instrument in psycholinguistics and cognitive science. In tangram naming settings, different observers can conceptualize the same shape as a “crab,” “sink,” “space ship,” “turtle,” or “person,” and dialogue-based tasks study how partners converge on shared references for such abstract figures (Morita et al., 2023). In one model-model version of the Tangram Naming Task, two agents see the same six tangram shapes under different placements and rotations and must agree on names through language alone; the reported communication accuracy was above the chance level of $\sqrt{2}$4, with an initial value of $\sqrt{2}$5 and a statistically significant increase to $\sqrt{2}$6 after incremental learning on successful episodes $\sqrt{2}$7 (Morita et al., 2023).

Large-scale annotation work has turned this ambiguity into a quantitative resource. KiloGram contains $\sqrt{2}$8 distinct tangram silhouettes, including $\sqrt{2}$9 scanned and vectorized shapes and $2$0 canonical tangrams, together with $2$1 annotations, whole-shape descriptions, part segmentations, and part labels (Patil et al., 2022). The dataset makes it possible to measure shape naming divergence, part naming divergence, and part segmentation agreement, and it shows that tangram interpretation is both semantically rich and structurally variable. This suggests that tangram is not merely a whole-shape recognition problem; it is also a part–whole abstraction problem.

SceneGram adds explicit scene context to this paradigm. It pairs $2$2 tangrams from KiloGram with $2$3 scene conditions, yielding $2$4 items and $2$5 human annotations, and shows that scene context systematically shifts conceptualization toward scene-typical labels while preserving substantial variability (Junker et al., 13 Jun 2025). A bathroom context increases labels such as “sink,” a beach context favors “crab,” and a mountain context favors “mountain.” Multimodal LLMs in the same setting exhibit some context sensitivity, but they do not reproduce the richness and variability of human conceptualizations and often label the scene instead of the tangram (Junker et al., 13 Jun 2025).

4. Tangram as a benchmark for multimodal spatial and geometric reasoning

Several recent AI benchmarks use tangram or tangram-like constructions to isolate geometric competence. One benchmark named “Tangram” evaluates geometric element recognition in large multimodal models using $2$6 diagrams and $2$7 visual-question-answer pairs about letters, triangles, circles, and line segments. The best closed-source configuration, GPT-4o with zero-shot chain-of-thought, achieved $2$8 overall accuracy, whereas human-student and human-expert baselines were $2$9 and $1$0, respectively (Zhang et al., 2024). Here, “Tangram” functions as a low-level perception benchmark rather than a seven-piece assembly task, but its motivation is directly analogous: geometric reasoning fails if diagram perception fails.

TangramPuzzle turns the classical puzzle itself into a geometry-grounded benchmark for compositional spatial reasoning. It contains $1$1 unique tangram configurations and $1$2 problem instances, formalized in a symbolic Tangram Construction Expression that specifies target outlines, initial and final piece states, and adjacency graphs exactly (Liu et al., 23 Jan 2026). In Outline Prediction, Gemini 3-Pro reached $1$3 accuracy with $1$4 invalid outputs; in End-to-End Code Generation, Gemini 3-Pro achieved a Validation Pass Rate of $1$5 and a Success rate of $1$6, while almost all other models had Validation Pass Rate $1$7 and Success $1$8 (Liu et al., 23 Jan 2026). The reported failure mode is consistent: models often match the target silhouette while violating rigid-body constraints, overlapping pieces, or distorting piece geometry.

TangramSR pushes the benchmark from symbolic composition to continuous geometry. Each piece is represented by a pose vector $1$9, where $\sqrt{2}$0 is position, $\sqrt{2}$1 is angle, and $\sqrt{2}$2 is scale, and correctness is measured by polygon IoU (Zong et al., 5 Feb 2026). Across five representative VLMs, the average IoU is about $\sqrt{2}$3 on single-piece tasks and about $\sqrt{2}$4 on two-piece composition; a training-free verifier–refiner loop using in-context learning and reward-guided feedback raises medium-triangle IoU from $\sqrt{2}$5–$\sqrt{2}$6 to $\sqrt{2}$7 without retraining (Zong et al., 5 Feb 2026). The paper interprets this as evidence that current VLMs recognize geometric patterns semantically but do not natively compute with geometry in a continuous metric sense.

A broader spatial benchmark, STARE, places tangram puzzles beside cube-net folding as integrated spatial reasoning tasks. Its tangram instances ask whether a set of pieces can exactly fill a $\sqrt{2}$8 or $\sqrt{2}$9 board, optionally with textual assembly steps and intermediate visual simulations. Models perform close to random chance, whereas humans achieve $\frac{\pi}{4}$0 without visual simulation and $\frac{\pi}{4}$1 with visual simulation; tangrams are also the task on which humans show the largest reduction in response time when intermediate visual states are provided (Li et al., 5 Jun 2025). A plausible implication is that tangram remains one of the clearest operational probes of multi-step visual simulation.

5. Robotic assembly and transfer learning from tangram structure

Tangram assembly has also become a benchmark for robotic reasoning, planning, and manipulation. In MRChaos, the robot starts from known, separated locations of the seven standard pieces, receives only a silhouette prompt $\frac{\pi}{4}$2, and must place pieces sequentially in a fixed order from largest to smallest (Zhao et al., 17 May 2025). The task is modeled as an MDP with observation $\frac{\pi}{4}$3, action $\frac{\pi}{4}$4, and a vision-based reward

$\frac{\pi}{4}$5

where $\frac{\pi}{4}$6 is the placed piece’s occupied pixels and $\frac{\pi}{4}$7 is its target region (Zhao et al., 17 May 2025). Training uses PPO in PyBullet with a two-stage curriculum; in simulation, MRChaos achieves $\frac{\pi}{4}$8 final coverage on random composites and $\frac{\pi}{4}$9 on unseen H-Fiendish tangram objects, and on the real robot it reaches $8$0 on Random and $8$1 on H-Hard (Zhao et al., 17 May 2025). The same framework also transfers to cutlery and soda arrangements.

A different line of work uses tangram trajectories as a source of visual pre-training. The Tangram dataset in this setting contains $8$2 unique puzzles, $8$3 human solution trajectories, and more than $8$4 binary snapshots, from which the authors define a completeness contrast loss over intermediate states and a puzzle meaning loss mapping final states to $8$5-dimensional GloVe embeddings (Zhao et al., 2021). The resulting $8$6k-parameter convolutional feature extractor improves state ranking in folding clothes and room-layout evaluation and facilitates convergence in few-shot handwriting tasks such as Omniglot and Multi-digit MNIST (Zhao et al., 2021). This suggests that the tangram solution process provides supervision not only for end-state recognition but also for progress estimation and aesthetic ordering.

Taken together, these results position tangram as a robotics-friendly abstraction of object assembly: the pieces are rigid, the target is specified by silhouette, and success depends on precise geometric placement rather than semantic recognition alone (Zhao et al., 17 May 2025). The same structure makes tangram useful as a source domain for low-resolution configuration learning (Zhao et al., 2021).

6. Extended meanings, metaphors, and homonyms

The term “tangram” has acquired several specialized meanings beyond the classical dissection puzzle. In combinatorics on words, a tangram is a finite word in which every letter occurs an even number of times; such a word can be cut into pieces and rearranged into two identical words, and the paper “4-tangrams are 4-avoidable” proves $8$7 for the minimum alphabet size needed to avoid tangrams with cut number at most $8$8 or $8$9 in an infinite word (Ochem et al., 28 Feb 2025). This is a terminological extension rather than a geometric one, but it preserves the underlying notion of decomposing and reassembling structured pieces.

In mathematical physics, “The 5d Tangram” describes a tessellation framework for 5d SCFTs in which rigid brane-web building blocks correspond to primitive T-cones and locked superpositions, and different tessellations of a Generalized Toric Polygon correspond to different cones of the extended Coulomb branch (Bolla et al., 2024). In measurement-based quantum computation, a tangram-like puzzle maps Clifford circuits to placements of colored polyomino blocks on a square-grid cluster state, so that correct placements implement valid MBQC patterns while more compact placements correspond to lower-overhead embeddings (Patil et al., 2022). In both cases, the word denotes compositional assembly under strict local constraints.

The name has also been adopted for systems in which heterogeneous components are stitched together efficiently. One “Tangram” is an LLM serving system for non-uniform KV cache compression that uses Deterministic Budget Allocation, Head Group Page, and Ahead-of-Time Load Balancing to improve throughput by up to $\frac{\pi}{4}$0 while preserving model accuracy (Kim et al., 4 Jun 2026). Another is a cloud–edge video analytics system that stitches variable-size patches into fixed-size canvases and uses online SLO-aware batching, reducing bandwidth consumption by up to $\frac{\pi}{4}$1 and computation cost by up to $\frac{\pi}{4}$2 while keeping SLO violations within $\frac{\pi}{4}$3 (Peng et al., 2024). These homonymous uses do not refer to the seven-piece puzzle directly, but they retain the idea of constrained composition from smaller parts.

Across these domains, the classical tangram persists as more than a recreational artifact. It is a rigorous dissection problem, a finite classification problem in polygonal geometry, an instrument for studying naming and common ground, a benchmark for continuous and compositional spatial reasoning, and a template for assembly-oriented abstractions in robotics and other computational settings (Pohl et al., 2020).