Multimodal Jailbreaking on SafeBench Mirror (ID)
100ASRVanilla
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| VanillaModel=LLaVA-1.5-7B2026.03 | 100 | 4.33 | |
| VanillaModel=Qwen3-VL-8B2026.03 | 90.9 | 4.55 | |
| Static SFTModel=Qwen3-VL-8B2026.03 | 0 | 0 | |
| CEMMAModel=Qwen3-VL-8B2026.03 | 0 | 0 | |
| Static SFTModel=LLaVA-1.5-7B2026.03 | 0 | 0 | |
| CEMMAModel=LLaVA-1.5-7B2026.03 | 0 | 0 |