List your company

Over-refusal Evaluation on XSTest Safe

1.6Over-refusal RateAdaCD
-0.3212.6425.638.56Mar 9, 2026Mar 25, 2026Apr 10, 2026Apr 27, 2026May 13, 2026May 29, 2026Jun 15, 2026

Evaluation results

MethodLinks
AdaCDBase Model=Llama32026.041.6
BaseBackbone Model=Qwen3-32B2026.061.6
SelfCDBase Model=Llama32026.042
AdaCDBase Model=Qwen32026.042
ST-DBackbone Model=Qwen3-8B2026.062
Star1Backbone Model=Qwen3-8B2026.062.4
ST-DBackbone Model=DeepSeek-R1-Distill-Llama-8B2026.062.4
SelfCDBase Model=Qwen32026.042.8
ST-SBackbone Model=DeepSeek-R1-Distill-Llama-8B2026.062.8
SafePathBackbone Model=Qwen3-32B2026.062.8
SurgicalBase Model=Qwen32026.043.2
BaseBackbone Model=Qwen3-8B2026.063.2
SafePathBackbone Model=Qwen3-8B2026.063.2
Star1Backbone Model=Qwen3-32B2026.063.2
ST-SBackbone Model=Qwen3-32B2026.063.2
ST-DBackbone Model=Qwen3-32B2026.063.2
Categorical SteeringBackbone=REFUSE-LLAMA, Method=Categorical Steering2026.033.6
SurgicalBase Model=Llama32026.043.6
DefaultBase Model=Qwen32026.044
BaseBackbone Model=DeepSeek-R1-Distill-Llama-8B2026.064
SSDBase Model=Qwen32026.044.8
ST-SBackbone Model=Qwen3-8B2026.064.8
Low-Rank CombinationBackbone=REFUSE-LLAMA, Method=Low-Rank Combination2026.035.2
DefaultBase Model=Llama32026.045.2
BaseBackbone Model=DeepSeek-R1-Distill-Llama-70B2026.065.2
PromptBase Model=Qwen32026.046
ST-SBackbone Model=DeepSeek-R1-Distill-Llama-70B2026.066
ST-DBackbone Model=DeepSeek-R1-Distill-Llama-70B2026.066.8
Star1Backbone Model=DeepSeek-R1-Distill-Llama-8B2026.067.2
Star1Backbone Model=DeepSeek-R1-Distill-Llama-70B2026.067.6
LLAMA 3 8B INSTRUCTBackbone=LLAMA 3 8B INSTRUCT2026.038.4
SafePathBackbone Model=DeepSeek-R1-Distill-Llama-70B2026.0610
SurgicalBase Model=Gemma22026.0410.4
AdaCDBase Model=Gemma22026.0410.4
Low-Rank CombinationBackbone=LLAMA 3 8B INSTRUCT, Method=Low-Rank Combination2026.0311.2
SelfCDBase Model=Gemma22026.0411.2
PromptBase Model=Gemma22026.0412
SafePathBackbone Model=DeepSeek-R1-Distill-Llama-8B2026.0612
PromptBase Model=Llama32026.0414.8
SSDBase Model=Gemma22026.0416.4
DefaultBase Model=Gemma22026.0417.6
Low-Rank CombinationBackbone=DEEPSEEK R1 DISTILL LLAMA, Method=Low-Rank Combination2026.0323.6
REFUSE-LLAMABackbone=REFUSE-LLAMA2026.0328
DEEPSEEK R1 DISTILL LLAMABackbone=DEEPSEEK R1 DISTILL LLAMA2026.0328
SSDBase Model=Llama32026.0449.6