Metric-Bench: leaderboard
Metric-Bench (Xi et al., 2026) tests in-context spatial metric reasoning of vision-language models on 1,340 questions. Each question gives the true metric values of several reference objects in a ScanNetV2 indoor image and asks for a target object quantity in metres, answered zero-shot; mean relative accuracy in percent over tolerance thresholds from 0.50 to 0.95; higher is better.
Metric: Mean relative accuracy (%; all questions; 1,340 template questions over ScanNetV2 indoor test scenes, each giving the real metric values of reference objects and asking for a target object's metric in camera coordinates; zero-shot; mean relative accuracy over thresholds 0.50-0.95; GPT-5-mini and Gemini-2.5-flash on a random 400-question subset). Source: arxiv.org. Saturation forecast: Estimated already saturated. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 Mini | 37.84 | #349 |
| 2 | Gemini 2.5 Flash | 33.95 | #452 |
| 3 | Qwen 2.5 VL 32B Instruct | 32.71 | #1055 |
| 4 | Qwen 2.5 VL 7B Instruct | 27.58 | #1842 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=metric-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.