SaLAD: leaderboard
Daily-life safety of multimodal models: 2,013 real-world image-text queries in 10 everyday categories, 1,070 implying a hazardous behaviour and 943 benign counterparts; SaLAD (Lou et al., ACL 2026 Findings). A GPT-4o judge marks an unsafe query correct only when the answer names and explains the specific risk (a bare refusal is wrong) and a safe query correct when it is answered without refusal or needless warnings; accuracy in percent over all queries, higher is better.
Metric: Accuracy rate (%; GPT-4o judge; all 2,013 image-text queries: an unsafe query is correct when the answer names and explains the risk, a safe query when it is answered without refusal or needless warnings). Source: arxiv.org. Saturation forecast: Estimated already saturated. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3.7 Sonnet | 77.05 | #455 |
| 2 | Gemini 2.5 Flash | 75.96 | #452 |
| 3 | GPT-4o | 75.36 | #741 |
| 4 | Qwen 2.5 VL 7B | 62.89 | #1796 |
| 5 | MiniCPM-V-2.6 | 59.12 | #3095 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=salad · How It Works · Data refreshed daily, snapshot 2026-11-12.