DefectBench (Building Facades) - Spatial Reasoning: leaderboard
Metric: Visual spatial reasoning F1 (0-1, times 100) of the pairwise relations (inclusion, overlapping, adjacency, disjoint) the model states between its detected defects, on DefectBench's 487 building-facade evaluation images (crack 117, material loss 102, surface stain 134, external fixings 122), zero-shot, JSON answers in one multi-turn dialogue per image; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (Preview) | 83.9 | #64 |
| 2 | GPT-5.2 Pro | 79.65 | #57 |
| 3 | Claude Opus 4.1 | 77.07 | #132 |
| 4 | Gemini 3 Flash (Preview) | 69.19 | #78 |
| 5 | GLM-4.6V | 68.66 | #309 |
| 6 | GLM-4.5V | 66.32 | #339 |
| 7 | Seed 1.8 | 64.89 | #136 |
| 8 | Qwen 2.5 VL 72B Instruct | 61.88 | #364 |
| 9 | GPT-4o | 58.91 | #333 |
| 10 | Qwen 3 VL 32B Instruct | 55.88 | #276 |
| 11 | Qwen 3 VL 32B (Thinking) | 47.63 | #287 (Qwen 3 VL 32B) |
| 12 | Qwen 2.5 VL 32B Instruct | 37.08 | #443 |
| 13 | Qwen 3 VL 8B Instruct | 28.5 | #401 |
| 14 | Qwen 2.5 VL 7B Instruct | 16.42 | #643 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=defectbench-building-facades-spatial-reasoning · How It Works · Data refreshed daily, snapshot 2026-10-11.