Safety Alignment Evaluation on SIUO
85.03Safety ScoreGRPO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| GRPOBackbone=Qwen3-VL-4B, Alignment Method=GRPO2026.06 | 85.03 | 85.63 | |
| COPSD (Hybrid)Backbone=Qwen3-VL-4B, Alignment Method=COPSD (Hybrid)2026.06 | 81.44 | 94.01 | |
| COPSD (Standard)Backbone=Qwen2.5-VL-7B, Alignment Method=COPSD (Standard)2026.06 | 79.64 | 92.22 | |
| COPSD (Standard)Backbone=Qwen3-VL-4B, Alignment Method=COPSD (Standard)2026.06 | 78.44 | 94.61 | |
| COPSD (Hybrid)Backbone=Qwen2.5-VL-7B, Alignment Method=COPSD (Hybrid)2026.06 | 77.25 | 94.01 | |
| Safe-RLHF-VBackbone=Qwen3-VL-4B, Alignment Method=Safe-RLHF-V2026.06 | 75.45 | 79.64 | |
| OPDBackbone=Qwen2.5-VL-7B, Alignment Method=OPD2026.06 | 73.05 | 92.81 | |
| SFTBackbone=Qwen2.5-VL-7B, Alignment Method=SFT2026.06 | 72.46 | 92.22 | |
| Qwen3-VL-4BBackbone=Qwen3-VL-4B, Alignment Method=Base2026.06 | 68.86 | 85.62 | |
| GRPOBackbone=Qwen2.5-VL-7B, Alignment Method=GRPO2026.06 | 65.87 | 83.83 | |
| SFTBackbone=Qwen3-VL-4B, Alignment Method=SFT2026.06 | 60.48 | 94.61 | |
| OPDBackbone=Qwen3-VL-4B, Alignment Method=OPD2026.06 | 60.48 | 91.62 | |
| Safe-RLHF-VBackbone=Qwen2.5-VL-7B, Alignment Method=Safe-RLHF-V2026.06 | 47.31 | 67.66 | |
| Qwen2.5-VL-7BBackbone=Qwen2.5-VL-7B, Alignment Method=Base2026.06 | 34.13 | 82.63 |