SpatialTrust: leaderboard
SpatialTrust (University of Washington, 2026) tests whether multimodal LLMs recognize, localize and explain environmental information-leakage risks in laptop-camera authentication scenes, with evidence grounding at an IoU threshold of 0.5. Score for the overall score is the mean over the five question types of detection, identification and explanation, in percent, higher is better.
Metric: Overall score (%; mean over the five question types of sensitive-factor detection, direct and indirect factor identification and explanation; over SpatialTrust's 577 secure-authentication images (five questions each); IoU threshold 0.5 for grounded evidence). Source: arxiv.org. Saturation forecast: Around 2028. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 30B A3B Instruct | 36.78 | #870 |
| 2 | Gemini 2.5 Flash | 33.81 | #452 |
| 3 | Qwen 3 VL 8B Instruct | 28.91 | #982 |
| 4 | GPT-5.4 | 25.94 | #121 |
| 5 | GLM-4.1V-9B (Thinking) | 21.62 | #1061 (GLM-4.1V-9B) |
| 6 | Qwen 2 VL 7B Instruct | 15.89 | #2569 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=spatialtrust · How It Works · Data refreshed daily, snapshot 2026-11-12.