Jailbreak on Advbench-M Image + Text (test)
980HF (BE)JOOD
Evaluation results
| Method | Links | ||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| JOODTarget model=LLaVA-1.5 13B, Equation=Eq. 32025.03 | 980 | 100 | 850 | 93 | 830 | 65 | 530 | 89 | 610 | 63 | 580 | 40 | 850 | 90 | |
| JOODTarget model=MiniGPT-4 7B, Equation=Eq. 32025.03 | 810 | 83 | 800 | 87 | 750 | 47 | 730 | 95 | 570 | 54 | 390 | 35 | 760 | 97 | |
| JOODTarget model=GPT-4V, Equation=Eq. 32025.03 | 710 | 63 | 390 | 23 | 720 | 47 | 400 | 74 | 210 | 4 | 110 | 10 | 60 | 23 | |
| VanillaTarget model=LLaVA-1.5 13B2025.03 | 660 | 87 | 320 | 43 | 350 | 47 | 240 | 53 | 170 | 46 | 100 | 15 | 430 | 83 | |
| FigStepTarget model=LLaVA-1.5 13B, Year=20232025.03 | 580 | 77 | 350 | 53 | 420 | 41 | 690 | 79 | 400 | 54 | 320 | 35 | 120 | 47 | |
| FigStepTarget model=MiniGPT-4 7B, Year=20232025.03 | 440 | 63 | 310 | 63 | 360 | 47 | 670 | 74 | 420 | 63 | 360 | 55 | 190 | 53 | |
| VanillaTarget model=MiniGPT-4 7B2025.03 | 150 | 50 | 250 | 73 | 120 | 6 | 0 | 42 | 20 | 42 | 80 | 20 | 300 | 73 | |
| HADESTarget model=LLaVA-1.5 13B, Year=20242025.03 | 100 | 13 | 270 | 53 | 360 | 47 | 470 | 74 | 400 | 54 | 40 | 5 | 170 | 47 | |
| FigStep-ProTarget model=GPT-4V, Year=20232025.03 | 90 | 23 | 80 | 17 | 180 | 25 | 210 | 32 | 40 | 8 | 30 | 0 | 10 | 0 | |
| HADESTarget model=MiniGPT-4 7B, Year=20242025.03 | 10 | 17 | 180 | 40 | 150 | 29 | 120 | 47 | 60 | 17 | 80 | 5 | 130 | 37 | |
| VanillaTarget model=GPT-4V2025.03 | 0 | 0 | 0 | 0 | 60 | 12 | 30 | 5 | 0 | 0 | 0 | 0 | 0 | 0 | |
| FigStepTarget model=GPT-4V, Year=20232025.03 | 0 | 0 | 20 | 3 | 50 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | |
| HADESTarget model=GPT-4V, Year=20242025.03 | 0 | 0 | 20 | 3 | 10 | 0 | 10 | 0 | 0 | 0 | 10 | 0 | 0 | 0 |