SaLAD - Safe Queries: leaderboard
The 943 benign queries of SaLAD (Lou et al., ACL 2026 Findings), everyday image-text questions rewritten from hazardous ones to test over-sensitivity. Accuracy in percent: a GPT-4o judge counts an answer correct when it helps without refusing or adding needless safety warnings; higher is better.
Metric: Accuracy rate (%; GPT-4o judge; 943 benign queries answered helpfully without refusal or needless safety warnings). Source: arxiv.org. Saturation forecast: Estimated already saturated. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o | 99.79 | #741 |
| 2 | Gemini 2.5 Flash | 99.68 | #452 |
| 3 | Claude 3.7 Sonnet | 99.58 | #455 |
| 4 | MiniCPM-V-2.6 | 99.05 | #3095 |
| 5 | Qwen 2.5 VL 7B | 98.41 | #1796 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=salad-safe-queries · How It Works · Data refreshed daily, snapshot 2026-11-12.