Multi-image safety detection on MIIT-dataset (test)
98.91Unsafe PrecisionQwen3.5-9B
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Qwen3.5-9Bconcatenated_images=true2026.07 | 98.91 | 31.71 | 48.02 | 63.97 | 96.52 | 76.94 | 64.11 | |
| Qwen3.5-9Bconcatenated_images=false2026.07 | 94.17 | 39.37 | 55.53 | 64.78 | 91.64 | 75.9 | 65.51 | |
| GPT-5.4concatenated_images=false2026.07 | 92.34 | 75.61 | 83.14 | 79.35 | 93.73 | 85.94 | 84.67 | |
| Qwen3-VL-8Bconcatenated_images=false2026.07 | 92.14 | 44.95 | 60.42 | 63.59 | 96.17 | 76.56 | 70.56 | |
| MiShield-8Bconcatenated_images=false2026.07 | 90.41 | 91.99 | 91.19 | 91.84 | 90.24 | 91.04 | 91.11 | |
| Llama-Guard-4-12Bconcatenated_images=false2026.07 | 88.89 | 5.57 | 10.49 | 51.26 | 99.3 | 67.62 | 52.44 | |
| InternVL3.5-8Bconcatenated_images=false2026.07 | 87.59 | 44.25 | 58.8 | 62.7 | 93.73 | 75.14 | 68.99 | |
| Baidu ICRconcatenated_images=true2026.07 | 86.3 | 21.95 | 35 | 55.29 | 96.52 | 70.3 | 59.23 | |
| GPT-5.4concatenated_images=true2026.07 | 85.28 | 58.54 | 69.42 | 68.44 | 89.9 | 77.71 | 74.22 | |
| Claude Opus 4.6concatenated_images=false2026.07 | 85.17 | 94.08 | 89.4 | 91.9 | 67.25 | 77.67 | 80.66 | |
| OpenAI Omniconcatenated_images=true2026.07 | 82.46 | 16.38 | 27.33 | 53.58 | 96.52 | 68.91 | 56.45 | |
| Gemini 3.1 Proconcatenated_images=false2026.07 | 82.06 | 86.06 | 84.01 | 85.35 | 81.18 | 83.21 | 83.62 | |
| InternVL3-8Bconcatenated_images=false2026.07 | 78.11 | 80.84 | 79.45 | 80.14 | 77.35 | 78.72 | 79.09 | |
| Alibaba Guardrailsconcatenated_images=true2026.07 | 68.18 | 10.45 | 18.13 | 51.51 | 95.12 | 66.83 | 52.79 | |
| Tencent IMSconcatenated_images=true2026.07 | 64.18 | 29.97 | 40.86 | 54.32 | 83.28 | 65.75 | 56.62 |