Safety Alignment Evaluation on MSS-Bench
93.33Safety ScoreSafe-RLHF-V
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Safe-RLHF-VBackbone=Qwen3-VL-4B, Alignment Method=Safe-RLHF-V2026.06 | 93.33 | 85 | |
| GRPOBackbone=Qwen3-VL-4B, Alignment Method=GRPO2026.06 | 89.83 | 89.67 | |
| COPSD (Hybrid)Backbone=Qwen3-VL-4B, Alignment Method=COPSD (Hybrid)2026.06 | 83.58 | 95.83 | |
| COPSD (Standard)Backbone=Qwen3-VL-4B, Alignment Method=COPSD (Standard)2026.06 | 83.5 | 97.5 | |
| COPSD (Standard)Backbone=Qwen2.5-VL-7B, Alignment Method=COPSD (Standard)2026.06 | 82.17 | 91.67 | |
| SFTBackbone=Qwen2.5-VL-7B, Alignment Method=SFT2026.06 | 80.67 | 94 | |
| Safe-RLHF-VBackbone=Qwen2.5-VL-7B, Alignment Method=Safe-RLHF-V2026.06 | 78.83 | 78.17 | |
| GRPOBackbone=Qwen2.5-VL-7B, Alignment Method=GRPO2026.06 | 78.67 | 86 | |
| Qwen3-VL-4BBackbone=Qwen3-VL-4B, Alignment Method=Base2026.06 | 77.31 | 91.28 | |
| OPDBackbone=Qwen2.5-VL-7B, Alignment Method=OPD2026.06 | 75.67 | 89.17 | |
| COPSD (Hybrid)Backbone=Qwen2.5-VL-7B, Alignment Method=COPSD (Hybrid)2026.06 | 75.5 | 90.17 | |
| SFTBackbone=Qwen3-VL-4B, Alignment Method=SFT2026.06 | 71.5 | 95.83 | |
| OPDBackbone=Qwen3-VL-4B, Alignment Method=OPD2026.06 | 69.5 | 95.83 | |
| Qwen2.5-VL-7BBackbone=Qwen2.5-VL-7B, Alignment Method=Base2026.06 | 59.83 | 93.33 |