List your company
ResearchFormerly WizWand

Safety Evaluation on VLBreak (Challenge)

0.04ASRVLGuard
-0.6163.8128.2412.668Aug 26, 2025
Updated 7mo ago

Evaluation results

MethodLinks
VLGuardBackbone=Qwen2-VL2025.080.04
PRISMBackbone=Qwen2-VL2025.080.05
PRISMBackbone=LLaVA-1.52025.080.2
VLGuardBackbone=LLaVA-1.52025.081.94
SPA-VLBackbone=LLaVA-1.52025.086.54
SPA-VLBackbone=Qwen2-VL2025.087.82
SafeRLHF-VBackbone=Qwen2-VL2025.0810.31
No DefenseBackbone=LLaVA-1.52025.0813
No DefenseBackbone=Qwen2-VL2025.0815.12
SafeRLHF-VBackbone=LLaVA-1.52025.0816.44