Safety and Helpfulness Evaluation on VLGuard
95.1Safety ScoreGRPO
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| GRPOBackbone=Qwen3-VL-4B, Alignment Method=GRPO2026.06 | 95.1 | — | 51.79 | |
| SAFER-STEERBase Model=Qwen2.5VL-7B, Method Category=Dynamic Methods2026.03 | 88.48 | 81.82 | — | |
| Safe-RLHF-VBackbone=Qwen3-VL-4B, Alignment Method=Safe-RLHF-V2026.06 | 88.2 | — | 77.6 | |
| GRPOBackbone=Qwen2.5-VL-7B, Alignment Method=GRPO2026.06 | 87.9 | — | 55.38 | |
| COPSD (Hybrid)Backbone=Qwen3-VL-4B, Alignment Method=COPSD (Hybrid)2026.06 | 86.3 | — | 97.13 | |
| SFTBase Model=Qwen2.5VL-7B, Method Category=Dynamic Methods2026.03 | 84.2 | 77.6 | — | |
| COPSD (Standard)Backbone=Qwen2.5-VL-7B, Alignment Method=COPSD (Standard)2026.06 | 83.6 | — | 94.62 | |
| OPDBackbone=Qwen2.5-VL-7B, Alignment Method=OPD2026.06 | 83 | — | 96.42 | |
| SPA-VLBase Model=Qwen2.5VL-7B, Method Category=Static Methods2026.03 | 82.6 | 74.51 | — | |
| COPSD (Hybrid)Backbone=Qwen2.5-VL-7B, Alignment Method=COPSD (Hybrid)2026.06 | 81.9 | — | 95.34 | |
| SFTBackbone=Qwen2.5-VL-7B, Alignment Method=SFT2026.06 | 81.8 | — | 95.52 | |
| COPSD (Standard)Backbone=Qwen3-VL-4B, Alignment Method=COPSD (Standard)2026.06 | 81.1 | — | 94.8 | |
| TISBase Model=Qwen2.5VL-7B, Method Category=Static Methods2026.03 | 80.1 | 76.7 | — | |
| SaFeR-VLMBase Model=Qwen2.5VL-7B, Method Category=Static Methods2026.03 | 79.88 | 77.88 | — | |
| SAFER-STEERBase Model=Qwen2.5VL-3B, Method Category=Dynamic Methods2026.03 | 79.62 | 73.25 | — | |
| MMDPOBase Model=Qwen2.5VL-7B, Method Category=Static Methods2026.03 | 79.59 | 78.41 | — | |
| SPA-VLBase Model=Qwen2.5VL-3B, Method Category=Static Methods2026.03 | 77.11 | 68.67 | — | |
| SFTBase Model=Qwen2.5VL-3B, Method Category=Dynamic Methods2026.03 | 76.9 | 66.6 | — | |
| SaFeR-VLMBase Model=Qwen2.5VL-3B, Method Category=Static Methods2026.03 | 75.1 | 75.3 | — | |
| TISBase Model=Qwen2.5VL-3B, Method Category=Static Methods2026.03 | 74.3 | 71 | — | |
| OPDBackbone=Qwen3-VL-4B, Alignment Method=OPD2026.06 | 74.3 | — | 94 | |
| SFTBackbone=Qwen3-VL-4B, Alignment Method=SFT2026.06 | 73.7 | — | 95.52 | |
| Qwen2.5VL-7BBase Model=Qwen2.5VL-7B, Method Category=Base2026.03 | 71.9 | 72.37 | — | |
| MMDPOBase Model=Qwen2.5VL-3B, Method Category=Static Methods2026.03 | 71.54 | 66.13 | — | |
| Qwen3-VL-4BBackbone=Qwen3-VL-4B, Alignment Method=Base2026.06 | 70.4 | — | 94.98 | |
| Safe-RLHF-VBackbone=Qwen2.5-VL-7B, Alignment Method=Safe-RLHF-V2026.06 | 69.6 | — | 64.87 | |
| Qwen2.5-VL-7BBackbone=Qwen2.5-VL-7B, Alignment Method=Base2026.06 | 67.6 | — | 97.31 | |
| Qwen2.5VL-3BBase Model=Qwen2.5VL-3B, Method Category=Base2026.03 | 66.2 | 56.3 | — | |
| GRLHF-VBase Model=Qwen2.5VL-7B, Method Category=Static Methods2026.03 | 59.34 | 63.72 | — | |
| SafeRLHF-VBase Model=Qwen2.5VL-7B, Method Category=Static Methods2026.03 | 52.8 | 46.2 | — | |
| SafeRLHF-VBase Model=Qwen2.5VL-3B, Method Category=Static Methods2026.03 | 48.2 | 39 | — | |
| GRLHF-VBase Model=Qwen2.5VL-3B, Method Category=Static Methods2026.03 | 36.9 | 37.8 | — |