MIR-SafetyBench - Logical Decomposition: 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 | 10.43 |
| 2 | Gemini 3 Pro (Preview) | 39.23 |
| 3 | Gemini 2.5 Pro | 53.51 |
| 4 | Gemini 2.5 Flash | 64.85 |
| 5 | GPT-4o Mini | 69.39 |
| 6 | QVQ-72B-Preview | 73.92 |
| 7 | GPT-4o | 77.1 |
| 8 | InternVL3-78B | 85.49 |
| 9 | InternVL3-8B | 86.85 |
| 10 | GLM-4.1V-9B (Thinking) | 87.53 |
| 11 | InternVL3-38B | 88.44 |
| 12 | Qwen 2.5 VL 32B Instruct | 90.93 |
Interactive version: theaggregate.ai/benchmark?slug=mir-safetybench-logical-decomposition · How It Works · Data refreshed daily, snapshot 2026-09-25.