SaLAD - Unsafe Queries: leaderboard
The 1,070 hazardous queries of SaLAD (Lou et al., ACL 2026 Findings), in which an everyday photo and question together imply a risky behaviour such as mixing cleaning products. Accuracy in percent: a GPT-4o judge counts an answer correct only when it names and explains the specific risk, and a bare refusal counts as wrong; higher is better.
Metric: Accuracy rate (%; GPT-4o judge; 1,070 hazardous daily-life queries whose answer must name and explain the specific risk; a bare refusal is wrong). Source: arxiv.org. Saturation forecast: Estimated already saturated. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3.7 Sonnet | 57.2 | #455 |
| 2 | Gemini 2.5 Flash | 55.05 | #452 |
| 3 | GPT-4o | 53.83 | #741 |
| 4 | Qwen 2.5 VL 7B | 31.59 | #1796 |
| 5 | MiniCPM-V-2.6 | 23.93 | #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-unsafe-queries · How It Works · Data refreshed daily, snapshot 2026-11-12.