Safety and Helpfulness Evaluation on SPA-VL
87.17Safety ScoreCOPSD (Standard)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| COPSD (Standard)Backbone=Qwen2.5-VL-7B, Alignment Method=COPSD (Standard)2026.06 | 87.17 | 89.15 | |
| SAFER-STEERBase Model=Qwen2.5VL-7B, Method Category=Dynamic Methods2026.03 | 85.8 | 79.17 | |
| TISBase Model=Qwen2.5VL-7B, Method Category=Static Methods2026.03 | 85.66 | 78.3 | |
| SFTBase Model=Qwen2.5VL-7B, Method Category=Dynamic Methods2026.03 | 85.66 | 78.3 | |
| COPSD (Hybrid)Backbone=Qwen2.5-VL-7B, Alignment Method=COPSD (Hybrid)2026.06 | 84.53 | 89.02 | |
| SAFER-STEERBase Model=Qwen2.5VL-3B, Method Category=Dynamic Methods2026.03 | 83.4 | 77.03 | |
| SPA-VLBase Model=Qwen2.5VL-7B, Method Category=Static Methods2026.03 | 83.4 | 76.42 | |
| SFTBase Model=Qwen2.5VL-3B, Method Category=Dynamic Methods2026.03 | 81.7 | 73.02 | |
| GRPOBackbone=Qwen2.5-VL-7B, Alignment Method=GRPO2026.06 | 76.98 | 73.38 | |
| MMDPOBase Model=Qwen2.5VL-7B, Method Category=Static Methods2026.03 | 76.05 | 78.54 | |
| SPA-VLBase Model=Qwen2.5VL-3B, Method Category=Static Methods2026.03 | 76.04 | 69.62 | |
| COPSD (Hybrid)Backbone=Qwen3-VL-4B, Alignment Method=COPSD (Hybrid)2026.06 | 74.72 | 72.52 | |
| SFTBackbone=Qwen2.5-VL-7B, Alignment Method=SFT2026.06 | 73.58 | 86.27 | |
| TISBase Model=Qwen2.5VL-3B, Method Category=Static Methods2026.03 | 72.45 | 61.32 | |
| SaFeR-VLMBase Model=Qwen2.5VL-7B, Method Category=Static Methods2026.03 | 71.13 | 79.81 | |
| OPDBackbone=Qwen2.5-VL-7B, Alignment Method=OPD2026.06 | 70.19 | 85.6 | |
| COPSD (Standard)Backbone=Qwen3-VL-4B, Alignment Method=COPSD (Standard)2026.06 | 69.81 | 76.72 | |
| Qwen2.5VL-7BBase Model=Qwen2.5VL-7B, Method Category=Base2026.03 | 69.43 | 73.2 | |
| SaFeR-VLMBase Model=Qwen2.5VL-3B, Method Category=Static Methods2026.03 | 68.87 | 73.21 | |
| MMDPOBase Model=Qwen2.5VL-3B, Method Category=Static Methods2026.03 | 65.09 | 63.02 | |
| Safe-RLHF-VBackbone=Qwen2.5-VL-7B, Alignment Method=Safe-RLHF-V2026.06 | 61.51 | 45.18 | |
| GRPOBackbone=Qwen3-VL-4B, Alignment Method=GRPO2026.06 | 60.38 | 62.74 | |
| GRLHF-VBase Model=Qwen2.5VL-7B, Method Category=Static Methods2026.03 | 59.81 | 68.24 | |
| OPDBackbone=Qwen3-VL-4B, Alignment Method=OPD2026.06 | 58.11 | 60.61 | |
| SFTBackbone=Qwen3-VL-4B, Alignment Method=SFT2026.06 | 51.7 | 51.91 | |
| Qwen3-VL-4BBackbone=Qwen3-VL-4B, Alignment Method=Base2026.06 | 50 | 50 | |
| Qwen2.5-VL-7BBackbone=Qwen2.5-VL-7B, Alignment Method=Base2026.06 | 50 | 50 | |
| Qwen2.5VL-3BBase Model=Qwen2.5VL-3B, Method Category=Base2026.03 | 48.68 | 53.2 | |
| Safe-RLHF-VBackbone=Qwen3-VL-4B, Alignment Method=Safe-RLHF-V2026.06 | 47.17 | 20.15 | |
| SafeRLHF-VBase Model=Qwen2.5VL-7B, Method Category=Static Methods2026.03 | 46.23 | 40.94 | |
| SafeRLHF-VBase Model=Qwen2.5VL-3B, Method Category=Static Methods2026.03 | 43.21 | 33.58 | |
| GRLHF-VBase Model=Qwen2.5VL-3B, Method Category=Static Methods2026.03 | 40.75 | 43.4 |