List your company
ResearchFormerly WizWand

Harmful Content Safety Evaluation on AdvBench

0.19Attack Success Rate (ASR)Star1
-1.218.2417.6927.14Jun 15, 2026
Updated 4mo ago

Evaluation results

MethodLinks
Star1Backbone Model=Qwen3-8B2026.060.19
SafePathBackbone Model=Qwen3-8B2026.060.38
ST-DBackbone Model=Qwen3-8B2026.060.38
ST-SBackbone Model=Qwen3-8B2026.060.58
ST-DBackbone Model=DeepSeek-R1-Distill-Llama-70B2026.060.58
ST-SBackbone Model=Qwen3-32B2026.061.35
ST-SBackbone Model=DeepSeek-R1-Distill-Llama-70B2026.061.35
ST-DBackbone Model=Qwen3-32B2026.061.73
BaseBackbone Model=Qwen3-8B2026.061.92
BaseBackbone Model=Qwen3-32B2026.063.85
ST-DBackbone Model=DeepSeek-R1-Distill-Llama-8B2026.064.42
SafePathBackbone Model=Qwen3-32B2026.064.81
Star1Backbone Model=Qwen3-32B2026.066.73
ST-SBackbone Model=DeepSeek-R1-Distill-Llama-8B2026.069.62
Star1Backbone Model=DeepSeek-R1-Distill-Llama-8B2026.0614.04
SafePathBackbone Model=DeepSeek-R1-Distill-Llama-8B2026.0614.62
Star1Backbone Model=DeepSeek-R1-Distill-Llama-70B2026.0618.65
SafePathBackbone Model=DeepSeek-R1-Distill-Llama-70B2026.0620.96
BaseBackbone Model=DeepSeek-R1-Distill-Llama-70B2026.0626.15
BaseBackbone Model=DeepSeek-R1-Distill-Llama-8B2026.0635.19