PolyComp: leaderboard
PolyComp (Patel, 2026) tests compositional 3D spatial reasoning: given images of a target polycube solid, the model picks which of four options shows two components that combine to form it. Accuracy in percent over 120 procedurally generated and verified problems, each shown in three single- or multi-image formats; higher is better and random guessing scores 25 percent.
Metric: Accuracy (%; pass@1 over 360 presented problems per model: 120 procedurally generated and exactly verified polycube problems, each asking which of four options shows two components that combine into the target solid, in three single- or multi-image presentation formats; chance 25%; a refusal counts as wrong). Source: arxiv.org. Saturation forecast: Around December 2027. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (Max) | 50 | #30 (GPT-5.6 Sol) |
| 2 | Claude Fable 5 (Max) | 39.4 | #27 (Claude Fable 5) |
| 3 | Gemini 3.1 Pro (Preview) | 27.5 | #94 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=polycomp · How It Works · Data refreshed daily, snapshot 2026-11-13.