VisReason (Vision-Centric) - 3D-Spatial Reasoning: leaderboard
Metric: Accuracy (%) on the 3D-Spatial Reasoning category (structural reasoning: reason about implicit 3D geometric relationships to infer the correct answer) of VisReason, answers scored by format (regular-expression match, GPT-5-mini judge for open-ended answers, IoU above 0.5 for boxes), zero-shot with chain-of-thought prompt templates, one per answer format; rows not shaded gray in the paper use explicit reasoning (GPT-5 and Gemini 3 Pro at low reasoning effort); higher is better. Source: arxiv.org. Saturation forecast: Around 2031. 22 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3 Pro (Preview) (Low) | 43.9 |
| 2 | Qwen 3 VL 235B A22B (Thinking) | 36.2 |
| 3 | Qwen 3 VL 8B (Thinking) | 33.1 |
| 4 | Gemini 2.5 Flash | 31.5 |
| 5 | InternVL3-8B | 31.5 |
| 6 | Qwen 3 VL 30B A3B (Thinking) | 30.8 |
| 7 | Qwen 3 VL 32B (Thinking) | 30.8 |
| 8 | GPT-5.2 (Thinking) | 30.8 |
| 9 | GPT-5 Mini (Low) | 30.8 |
| 10 | GPT-5 Mini (Medium) | 28.5 |
| 11 | O4 Mini | 27.7 |
| 12 | Qwen 2.5 VL 32B Instruct | 27.7 |
| 13 | GPT-5 Mini (High) | 26.9 |
| 14 | GPT-5 (Low) | 26.2 |
| 15 | GPT-4o | 25.4 |
Interactive version: theaggregate.ai/benchmark?slug=visreason-vision-centric-3d-spatial-reasoning · How It Works · Data refreshed daily, snapshot 2026-10-07.