SpatialViz-Bench: leaderboard

Each multimodal model answers all 1,180 four-option spatial visualization problems (mental rotation, paper and cube folding, cross-sections and cube counting, arrow, block and gear motion) zero-shot with step-by-step reasoning; the score is accuracy, read by rules. Unlike spatial-perception boards that ask where objects are in a photo, every item asks the model to transform a generated diagram in its head.

Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 26 models tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1Gemini 2.5 Pro (Preview 03-25)44.66#239
2O1 (2024-12-17)41.36#314
3Gemini 2.5 Flash (Preview 04-17)36.86#455
4Llama 4 Scout Instruct34.24#1575
5Claude 3.7 Sonnet (20250219)33.9#421
6Qwen 2.5 VL 72B Instruct33.31#849
7doubao-1-5-vision-pro-32k-25011533.31
8Claude 3.5 Sonnet (20240620)32.54#693
9Qwen 2.5 VL 32B Instruct32.12#1074
10Qwen-VL-Max-2025-04-0832.03
11Llama 4 Maverick Instruct31.78#1151
12GPT-4o (2024-08-06)31.1#746
13InternVL3-38B30.34#927
14InternVL3-8B30.08#1756
15InternVL3-78B29.75#755

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=spatialviz-bench · How It Works · Data refreshed daily, snapshot 2026-11-17.