MIR-SafetyBench - Logical Causality: leaderboard
Metric: Attack Success Rate (%; lower is safer). Source: arxiv.org. Saturation forecast: Around December 2026. 19 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-5.1 | 21.43 |
| 2 | Gemini 3 Pro (Preview) | 46.79 |
| 3 | GPT-4o Mini | 53.57 |
| 4 | Gemini 2.5 Pro | 58.93 |
| 5 | QVQ-72B-Preview | 73.21 |
| 6 | InternVL3-8B | 73.93 |
| 7 | Gemini 2.5 Flash | 75.71 |
| 8 | InternVL3-78B | 77.14 |
| 9 | GPT-4o | 78.21 |
| 10 | Qwen 2.5 VL 32B Instruct | 82.5 |
| 11 | InternVL3-38B | 84.64 |
| 12 | GLM-4.1V-9B (Thinking) | 90 |
Interactive version: theaggregate.ai/benchmark?slug=mir-safetybench-logical-causality · How It Works · Data refreshed daily, snapshot 2026-09-25.