List your company
ResearchFormerly WizWand

Safety Evaluation on AdvBench & SorryBench

41.6Avg Success Rate (Reasoning Path)No Defense
-1.6649.56820.832.032Aug 6, 2025
Updated 5mo ago

Evaluation results

MethodLinks
No DefenseBackbone=R1-Qwen-7B, Defense Strategy=None2025.0841.659.2
PPLBackbone=R1-Qwen-7B, Defense Strategy=Input-level Filtering2025.0841.659.2
No DefenseBackbone=R1-Llama-8B, Defense Strategy=None2025.0840.348.5
PPLBackbone=R1-Llama-8B, Defense Strategy=Input-level Filtering2025.0840.348.5
Llama Guard 3Backbone=R1-Llama-8B, Defense Strategy=Post-hoc Guard Model2025.0814.627.1
Llama Guard 3Backbone=R1-Qwen-7B, Defense Strategy=Post-hoc Guard Model2025.0810.232.3
GuardReasonerBackbone=R1-Llama-8B, Defense Strategy=Post-hoc Guard Model2025.0810.222.7
GuardReasonerBackbone=R1-Qwen-7B, Defense Strategy=Post-hoc Guard Model2025.08923.5
ReasoningGuardBackbone=R1-Llama-8B, Defense Strategy=Proactive Reasoning Guard2025.081.18
ReasoningGuard + LGBackbone=R1-Llama-8B, Defense Strategy=Hybrid2025.081.16.8
ReasoningGuard + GRBackbone=R1-Llama-8B, Defense Strategy=Hybrid2025.081.18
ReasoningGuardBackbone=R1-Qwen-7B, Defense Strategy=Proactive Reasoning Guard2025.0817.8
ReasoningGuard + LGBackbone=R1-Qwen-7B, Defense Strategy=Hybrid2025.0806.8
ReasoningGuard + GRBackbone=R1-Qwen-7B, Defense Strategy=Hybrid2025.0806.8