Safety Refusal on AdvBench
99.42Refusal RateLow-Rank Combination
Evaluation Results
| Method | Links | |
|---|---|---|
| Low-Rank CombinationBackbone=REFUSE-LLAMA, Method=Low-Rank Combination2026.03 | 99.42 | |
| Categorical SteeringBackbone=REFUSE-LLAMA, Method=Categorical Steering2026.03 | 99.23 | |
| WASBackbone=Llama 3.1-8B2025.05 | 98.8 | |
| WASBackbone=Llama 3.2-1B2025.05 | 98.2 | |
| WASBackbone=Mistral 7B2025.05 | 98.2 | |
| CASTBackbone=Llama 3.1-8B2025.05 | 95.7 | |
| Low-Rank CombinationBackbone=DEEPSEEK R1 DISTILL LLAMA, Method=Low-Rank Combination2026.03 | 94.81 | |
| REFUSE-LLAMABackbone=REFUSE-LLAMA2026.03 | 94.23 | |
| Low-Rank CombinationBackbone=LLAMA 3 8B INSTRUCT, Method=Low-Rank Combination2026.03 | 94.23 | |
| DEEPSEEK R1 DISTILL LLAMABackbone=DEEPSEEK R1 DISTILL LLAMA2026.03 | 94.23 | |
| LLAMA 3 8B INSTRUCTBackbone=LLAMA 3 8B INSTRUCT2026.03 | 93.85 | |
| CASTBackbone=Llama 3.2-1B2025.05 | 93.7 | |
| CASTBackbone=Mistral 7B2025.05 | 91.7 | |
| Base ModelBackbone=Llama 3.1-8B2025.05 | 91.5 | |
| Base ModelBackbone=Llama 3.2-1B2025.05 | 91.1 | |
| ITIBackbone=Llama 3.1-8B2025.05 | 89.1 | |
| ACTBackbone=Llama 3.1-8B2025.05 | 83.9 | |
| ACTBackbone=Llama 3.2-1B2025.05 | 83.1 | |
| ITIBackbone=Llama 3.2-1B2025.05 | 82.5 | |
| GER-SteerBackbone=Qwen2.5-7B2026.03 | 77.5 | |
| GER-SteerBackbone=Gemma-2-9B-it2026.03 | 77.3 | |
| ACTBackbone=Gemma-2-9B-it2026.03 | 77 | |
| LDPBackbone=Gemma-2-9B-it2026.03 | 76.7 | |
| Angular SteeringBackbone=Gemma-2-9B-it2026.03 | 76.5 | |
| NL-ITIBackbone=Gemma-2-9B-it2026.03 | 76.2 | |
| NL-ITIBackbone=Qwen2.5-7B2026.03 | 75.3 | |
| RePEBackbone=Gemma-2-9B-it2026.03 | 75.2 | |
| CAABackbone=Qwen2.5-7B2026.03 | 75.1 | |
| CAABackbone=Gemma-2-9B-it2026.03 | 75.1 | |
| Angular SteeringBackbone=Qwen2.5-7B2026.03 | 74.7 | |
| VanillaBackbone=Gemma-2-9B-it2026.03 | 74.5 | |
| RePEBackbone=Qwen2.5-7B2026.03 | 74.2 | |
| ACTBackbone=Qwen2.5-7B2026.03 | 73.2 | |
| GER-SteerBackbone=Llama-3.1-8B-Instruct2026.03 | 67.7 | |
| RePEBackbone=Llama-3.1-8B-Instruct2026.03 | 67.5 | |
| Angular SteeringBackbone=Llama-3.1-8B-Instruct2026.03 | 67.4 | |
| LDPBackbone=Llama-3.1-8B-Instruct2026.03 | 67.1 | |
| ACTBackbone=Llama-3.1-8B-Instruct2026.03 | 66.8 | |
| CAABackbone=Llama-3.1-8B-Instruct2026.03 | 66.7 | |
| NL-ITIBackbone=Llama-3.1-8B-Instruct2026.03 | 66.4 | |
| LDPBackbone=Qwen2.5-7B2026.03 | 66.3 | |
| VanillaBackbone=Llama-3.1-8B-Instruct2026.03 | 65.2 | |
| VanillaBackbone=Qwen2.5-7B2026.03 | 64.7 | |
| ITIBackbone=Mistral 7B2025.05 | 54.9 | |
| ACTBackbone=Mistral 7B2025.05 | 51.4 | |
| Base ModelBackbone=Mistral 7B2025.05 | 10.2 |