MLLM Safety Evaluation on OOD-MMSafe Constitution Mode 1.0 (test)
1.84RAGemma-3-12B
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Gemma-3-12BModel Type=Open-source2026.03 | 1.84 | 2.9 | 1.82 | 6.6 | 1.96 | 0 | |
| GPT-5.1Model Type=Closed-source2026.03 | 1.82 | 7.7 | 1.83 | 7.5 | 1.98 | 0 | |
| Gemini-3-ProModel Type=Closed-source2026.03 | 1.79 | 9.2 | 1.8 | 8.6 | 1.94 | 0 | |
| Gemini-3-Pro-ThinkModel Type=Closed-source2026.03 | 1.7 | 12.7 | 1.7 | 13 | 1.91 | 0.2 | |
| Qwen3VL-32BModel Type=Open-source2026.03 | 1.58 | 17.6 | 1.62 | 17.6 | 1.85 | 0.2 | |
| Qwen3VL-4BModel Type=Open-source2026.03 | 1.54 | 16.7 | 1.65 | 15.4 | 1.81 | 1.1 | |
| Doubao-1.6Model Type=Closed-source2026.03 | 1.42 | 24 | 1.48 | 22.4 | 1.78 | 0.7 | |
| InternVL-3.5-38BModel Type=Open-source2026.03 | 1.4 | 17.6 | 1.72 | 11.6 | 1.65 | 4 | |
| Qwen3VL-8BModel Type=Open-source2026.03 | 1.33 | 28.1 | 1.42 | 25.7 | 1.76 | 0.9 | |
| o4-MiniModel Type=Closed-source2026.03 | 1.28 | 31.9 | 1.35 | 29.5 | 1.73 | 0.9 | |
| InternVL-3.5-8BModel Type=Open-source2026.03 | 0.76 | 49.2 | 1.05 | 43.4 | 1.27 | 2.6 | |
| LLaVA-1.5-7BModel Type=Open-source2026.03 | 0.59 | 60.9 | 0.83 | 52.1 | 0.95 | 20.7 | |
| QwenVL2.5-7BModel Type=Open-source2026.03 | 0.52 | 64.4 | 0.6 | 66.8 | 1.24 | 2.4 |