Safety Evaluation on MM-Safety
0.4ASRMoRAS
Evaluation Results
| Method | Links | |
|---|---|---|
| MoRASModel=Qwen-VL-Chat2025.10 | 0.4 | |
| MoRASModel=LLaVA-OneVision-7B2025.10 | 0.6 | |
| MoRASModel=LLaVA-1.5-7B2025.10 | 2.6 | |
| CoCAModel=Qwen-VL-Chat2025.10 | 2.6 | |
| MoRASModel=InternLM-XComposer-2.52025.10 | 5.9 | |
| CoCAModel=InternLM-XComposer-2.52025.10 | 6.1 | |
| FigStepModel=InternLM-XComposer-2.52025.10 | 6.3 | |
| ETAModel=InternLM-XComposer-2.52025.10 | 7.3 | |
| FigStepModel=Qwen-VL-Chat2025.10 | 8.1 | |
| CAREBackbone=Qwen-2.5VL2026.03 | 8.72 | |
| CAREBackbone=Qwen2.5-VL2026.03 | 8.72 | |
| CAREBackbone=LLaVA-onevision2026.03 | 9.13 | |
| ETAModel=Qwen-VL-Chat2025.10 | 9.3 | |
| ShiftDCBackbone=Qwen2.5-VL2026.03 | 10.31 | |
| CoCAModel=LLaVA-OneVision-7B2025.10 | 11.8 | |
| ProFSBackbone=Qwen2.5-VL2026.03 | 13.39 | |
| ASTRAModel=InternLM-XComposer-2.52025.10 | 13.4 | |
| FigStepModel=LLaVA-OneVision-7B2025.10 | 13.5 | |
| ASTRAModel=LLaVA-OneVision-7B2025.10 | 14.6 | |
| ECSOModel=InternLM-XComposer-2.52025.10 | 14.9 | |
| AutoSteerModel=LLaVA-OneVision-7B2025.10 | 15.2 | |
| ASTRABackbone=LLaVA-onevision2026.03 | 15.37 | |
| ASTRABackbone=Qwen-2.5VL2026.03 | 15.57 | |
| ETAModel=LLaVA-1.5-7B2025.10 | 15.8 | |
| ETAModel=LLaVA-OneVision-7B2025.10 | 15.8 | |
| ECSOModel=LLaVA-1.5-7B2025.10 | 15.9 | |
| ECSOModel=LLaVA-OneVision-7B2025.10 | 16.1 | |
| ASTRAModel=Qwen-VL-Chat2025.10 | 16.2 | |
| SPO-VLMBackbone=LLaVA-onevision2026.03 | 16.26 | |
| CASTBackbone=Qwen-2.5VL2026.03 | 18.81 | |
| AutoSteerModel=InternLM-XComposer-2.52025.10 | 18.9 | |
| ECSOModel=Qwen-VL-Chat2025.10 | 19.1 | |
| SPO-VLMBackbone=Qwen-2.5VL2026.03 | 19.63 | |
| CoCAModel=LLaVA-1.5-7B2025.10 | 19.7 | |
| JailGuardBackbone=Qwen-2.5VL2026.03 | 20.83 | |
| VanillaModel=InternLM-XComposer-2.52025.10 | 21.8 | |
| VanillaModel=LLaVA-OneVision-7B2025.10 | 26.1 | |
| FigStepModel=LLaVA-1.5-7B2025.10 | 26.8 | |
| ASTRAModel=LLaVA-1.5-7B2025.10 | 27.2 | |
| CASTBackbone=LLaVA-onevision2026.03 | 27.83 | |
| AutoSteerModel=Qwen-VL-Chat2025.10 | 28.6 | |
| MoRASBackbone=LLaVA-1.5-7B, Suffix injection=true2025.10 | 28.9 | |
| JailGuardBackbone=LLaVA-onevision2026.03 | 30.17 | |
| Refusal PairsBackbone=Qwen-2.5VL2026.03 | 30.31 | |
| ETABackbone=LLaVA-1.5-7B, Suffix injection=true2025.10 | 31.5 | |
| Refusal PairsBackbone=LLaVA-onevision2026.03 | 32.13 | |
| ECSOBackbone=LLaVA-1.5-7B, Suffix injection=true2025.10 | 33.1 | |
| VanillaModel=Qwen-VL-Chat2025.10 | 33.1 | |
| FigStepBackbone=LLaVA-1.5-7B, Suffix injection=true2025.10 | 34.2 | |
| CoCABackbone=LLaVA-1.5-7B, Suffix injection=true2025.10 | 35.7 | |
| Original modelBackbone=Qwen-2.5VL2026.03 | 36.29 | |
| Original modelBackbone=LLaVA-onevision2026.03 | 36.48 | |
| AutoSteerModel=LLaVA-1.5-7B2025.10 | 37.6 | |
| AutoSteerBackbone=LLaVA-1.5-7B, Suffix injection=true2025.10 | 38.4 | |
| VanillaModel=LLaVA-1.5-7B2025.10 | 40.1 | |
| ASTRABackbone=LLaVA-1.5-7B, Suffix injection=true2025.10 | 41.5 | |
| VanillaBackbone=LLaVA-1.5-7B, Suffix injection=true2025.10 | 44.2 |