MiSCHiEF (Safety) - Caption-to-Image Matching: leaderboard
Metric: Accuracy (%, 0-100) on MiS: 190 safety minimal pairs, a safe and an unsafe version of one scenario; given one caption and the pair's two images, pick the matching image (chance 50); captions written by Gemini 2.5 Pro and images generated and edited with the GPT-Image API, both manually verified; default Hugging Face generation settings, one seed; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o | 93.15 | #333 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=mischief-safety-caption-to-image-matching · How It Works · Data refreshed daily, snapshot 2026-10-11.