Jailbreak Attacks on MM-Safety
99.9Safety RateTRR
Evaluation Results
| Method | Links | |
|---|---|---|
| TRRBackbone=Qwen2.5VL-7B, Alignment Method=TRR2025.12 | 99.9 | |
| TRRBackbone=Qwen2.5VL-32B, Alignment Method=TRR2025.12 | 99.4 | |
| MSR-AlignBackbone=Qwen2.5VL-32B, Alignment Method=MSR-Align2025.12 | 99.1 | |
| TiSBackbone=Qwen2.5VL-32B, Alignment Method=TiS2025.12 | 98.9 | |
| MSR-AlignBackbone=Qwen2.5VL-7B, Alignment Method=MSR-Align2025.12 | 98.2 | |
| Gemini-2.5-proModel=Gemini-2.5-pro2025.12 | 94.8 | |
| Claude-3.5-SonnetModel=Claude-3.5-Sonnet2025.12 | 91.9 | |
| Safework-R1-7BModel=Safework-R1-7B2025.12 | 88.3 | |
| TiSBackbone=Qwen2.5VL-7B, Alignment Method=TiS2025.12 | 85.8 | |
| INFER+Test Model=Gemini 2.5 Flash2026.07 | 83 | |
| INFER+Test Model=Qwen3-VL2026.07 | 77 | |
| GPT-4oModel=GPT-4o2025.12 | 74.7 | |
| INFERTest Model=Gemini 2.5 Flash2026.07 | 66 | |
| Qwen2.5VL-32BBackbone=Qwen2.5VL-32B, Alignment Method=None2025.12 | 54.7 | |
| CS-DJTest Model=Gemini 2.5 Flash2026.07 | 54 | |
| Qwen2.5VL-7BBackbone=Qwen2.5VL-7B, Alignment Method=None2025.12 | 50.1 | |
| ImageTypoTest Model=Gemini 2.5 Flash2026.07 | 42 | |
| INFERTest Model=Qwen3-VL2026.07 | 41 | |
| CS-DJTest Model=Qwen3-VL2026.07 | 32 | |
| ImageTypoTest Model=Qwen3-VL2026.07 | 15 | |
| TVPITest Model=Gemini 2.5 Flash2026.07 | 11 | |
| TVPITest Model=Qwen3-VL2026.07 | 2 |