List your company
ResearchFormerly WizWand

Safety Evaluation on MCV SafetyBench (four-clip setting)

0.62ASRImage filtering
-2.56618.939540.44561.9505Jun 1, 2026
Updated 5mo ago

Evaluation results

MethodLinks
Image filteringModel=Qwen3-VL-8B, Attack Type=Implicit, Number of clips=42026.060.62
Image filteringModel=LLaVA-Video-7B, Attack Type=Implicit, Number of clips=42026.065.55
AdashieldModel=LLaVA-Video-7B, Attack Type=Implicit, Number of clips=42026.065.62
AdashieldModel=Qwen3-VL-8B, Attack Type=Implicit, Number of clips=42026.0615.68
Image filteringModel=Average, Attack Type=Implicit, Number of clips=42026.0617.37
AdashieldModel=InternVL3.5-8B, Attack Type=Implicit, Number of clips=42026.0623.01
AdashieldModel=Average, Attack Type=Implicit, Number of clips=42026.0629.45
Image filteringModel=InternVL3.5-8B, Attack Type=Implicit, Number of clips=42026.0629.66
Safe systemModel=InternVL3.5-8B, Attack Type=Implicit, Number of clips=42026.0633.01
Image filteringModel=Qwen2.5-VL-7B, Attack Type=Implicit, Number of clips=42026.0633.63
Safe systemModel=LLaVA-Video-7B, Attack Type=Implicit, Number of clips=42026.0650.62
OriginalModel=LLaVA-Video-7B, Attack Type=Implicit, Number of clips=42026.0650.68
Safe systemModel=Average, Attack Type=Implicit, Number of clips=42026.0652.79
Safe systemModel=Qwen3-VL-8B, Attack Type=Implicit, Number of clips=42026.0657.05
OriginalModel=InternVL3.5-8B, Attack Type=Implicit, Number of clips=42026.0665.27
OriginalModel=Average, Attack Type=Implicit, Number of clips=42026.0667.34
Safe systemModel=Qwen2.5-VL-7B, Attack Type=Implicit, Number of clips=42026.0670.48
OriginalModel=Qwen3-VL-8B, Attack Type=Implicit, Number of clips=42026.0673.15
AdashieldModel=Qwen2.5-VL-7B, Attack Type=Implicit, Number of clips=42026.0673.49
OriginalModel=Qwen2.5-VL-7B, Attack Type=Implicit, Number of clips=42026.0680.27