SenseBench - Paired-Image Perception: leaderboard
Remote-sensing low-level perception from SenseBench (Zhong et al., 2026) with two images: a pristine view against its degraded copy, or two acquisitions of the same place at different times. The model answers zero-shot multiple-choice questions on whether, what and how severe the degradation is; exact-match accuracy in percent, averaged over the three question types, higher is better.
Metric: Exact-Match Accuracy (%; mean of the whether, what and how question types, image pairs). Source: arxiv.org. Saturation forecast: Around February 2027. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 81.28 | #91 |
| 2 | Qwen 3.5 Plus | 76.88 | #227 |
| 3 | GPT-5.4 | 76.15 | #121 |
| 4 | Llama 3.2 11B | 45.63 | #6434 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=sensebench-paired-image-perception · How It Works · Data refreshed daily, snapshot 2026-11-12.