UltraVR: leaderboard
UltraVR (2026) tests evidence-grounded reasoning over ultra-resolution images, where the answer depends on tiny, distant or scattered details, in CCTV, remote sensing, whole-slide pathology and industrial anomaly grids. The score is the domain-balanced macro accuracy. Models answer directly from the original ultra-resolution image under their own preprocessing; accuracy in percent, higher is better.
Metric: Domain-balanced macro accuracy (%): mean of the CCTV, remote sensing, whole-slide imaging and anomaly detection accuracies, direct answering (S0) with the original ultra-resolution image under each model default preprocessing (V2); four-choice questions except anomaly detection, which asks for the grid coordinate of the one anomalous cell out of 100; higher is better. Source: arxiv.org. Saturation forecast: Around April 2028. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 44.9 | #49 |
| 2 | Gemini 3.1 Pro (Preview) | 41.4 | #94 |
| 3 | Qwen 3.5 27B | 34.2 | #359 |
| 4 | GPT-5.4 | 32.9 | #123 |
| 5 | Gemma 4 26B | 30.7 | #612 |
| 6 | GLM-4.5V | 28.2 | #744 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=ultravr · How It Works · Data refreshed daily, snapshot 2026-11-13.