SenseBench - Paired-Image Description: leaderboard
Diagnostic description task from SenseBench (Zhong et al., 2026) with two remote-sensing images: the model describes the distortions in each image in one sentence apiece and says which image has the higher quality and why. A Prometheus judge compares the answer with a human-verified reference and rates completeness, correctness and faithfulness from 0 to 2 each; the score is their sum (0-6), higher is better.
Metric: Judge Score Sum (0-6; Prometheus completeness, correctness and faithfulness against the verified reference report, 0-2 each, 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) | 4.69 | #91 |
| 2 | Qwen 3.5 Plus | 4.5 | #227 |
| 3 | GPT-5.4 | 4.1 | #121 |
| 4 | Llama 3.2 11B | 2.58 | #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-description · How It Works · Data refreshed daily, snapshot 2026-11-12.