Multi-turn MLLM Safety Evaluation on STEER-Beaver Multi-turn
55.58SafetySPA-VL
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SPA-VLBackbone=Qwen2.5VL-7B, Alignment Strategy=SPA-VL2026.03 | 55.58 | 77.34 | |
| SAFER-STEERBackbone=Qwen2.5VL-7B, Alignment Strategy=SAFER-STEER2026.03 | 53.44 | 75.57 | |
| SaFeR-VLMBackbone=Qwen2.5VL-7B, Alignment Strategy=SaFeR-VLM2026.03 | 51.69 | 58.26 | |
| SAFER-STEERBackbone=Qwen2.5VL-3B, Alignment Strategy=SAFER-STEER2026.03 | 50.27 | 71.2 | |
| SFTBackbone=Qwen2.5VL-7B, Alignment Strategy=SFT2026.03 | 50 | 75.45 | |
| MMDPOBackbone=Qwen2.5VL-7B, Alignment Strategy=MMDPO2026.03 | 40.73 | 71.81 | |
| SFTBackbone=Qwen2.5VL-3B, Alignment Strategy=SFT2026.03 | 39.43 | 64.27 | |
| Qwen2.5VL-7BBackbone=Qwen2.5VL-7B, Alignment Strategy=Base2026.03 | 36.06 | 65.54 | |
| SPA-VLBackbone=Qwen2.5VL-3B, Alignment Strategy=SPA-VL2026.03 | 34.9 | 53.35 | |
| GRLHF-VBackbone=Qwen2.5VL-7B, Alignment Strategy=GRLHF-V2026.03 | 30.23 | 55.01 | |
| TISBackbone=Qwen2.5VL-7B, Alignment Strategy=TIS2026.03 | 29.23 | 24.96 | |
| SafeRLHF-VBackbone=Qwen2.5VL-7B, Alignment Strategy=SafeRLHF-V2026.03 | 28.42 | 42.63 | |
| SaFeR-VLMBackbone=Qwen2.5VL-3B, Alignment Strategy=SaFeR-VLM2026.03 | 26.23 | 37.32 | |
| TISBackbone=Qwen2.5VL-3B, Alignment Strategy=TIS2026.03 | 23.84 | 26.16 | |
| MMDPOBackbone=Qwen2.5VL-3B, Alignment Strategy=MMDPO2026.03 | 22.71 | 42.43 | |
| Qwen2.5VL-3BBackbone=Qwen2.5VL-3B, Alignment Strategy=Base2026.03 | 21.79 | 42.88 | |
| GRLHF-VBackbone=Qwen2.5VL-3B, Alignment Strategy=GRLHF-V2026.03 | 20.63 | 39.86 | |
| SafeRLHF-VBackbone=Qwen2.5VL-3B, Alignment Strategy=SafeRLHF-V2026.03 | 14.44 | 23.17 |