PolyComp Benchmark Tests AI Spatial Reasoning with Polycube Puzzles

Researchers have introduced PolyComp, a procedurally generated benchmark designed to stress visual recognition and compositional spatial reasoning in multimodal AI models. The benchmark presents models with a target solid and four options, requiring identification of which pair of polycube components can combine to form the target. It comprises 120 problems across four geometry families, each presented in three formats using single or multiple images, totaling 360 presented problems per model. The random guessing baseline is 25%. In evaluations, GPT-5.6 Sol with max effort achieved 50.0% accuracy (95% CI 43.3-56.7%) at a mean cost of $0.951 per presented problem. Claude Fable 5 with max effort scored 39.4% (33.1-46.1%) at $0.701, while Gemini 3.1 Pro Preview with thinking level high reached 27.5% (22.8-32.5%), near the random baseline, at $0.350. The accuracy spread across geometry families indicates varying difficulty. The benchmark is released as an arXiv preprint (arXiv:2608.14741) and is intended to advance research in compositional spatial reasoning for AI systems.