Safety Evaluation on Refusal
86.6Refusal RateMARI
Evaluation Results
| Method | Links | |
|---|---|---|
| MARIModel=Llama-3-8B2026.05 | 86.6 | |
| ReFTModel=Llama-3-8B2026.05 | 86.1 | |
| CAAModel=Llama-3-8B2026.05 | 85.4 | |
| MARIModel=Llama-2-7B2026.05 | 85.1 | |
| VanillaModel=Llama-3-8B2026.05 | 85.1 | |
| ReFTModel=Llama-2-7B2026.05 | 84.5 | |
| NL-ITIModel=Llama-2-7B2026.05 | 84 | |
| CAAModel=Llama-2-7B2026.05 | 83 | |
| ITIModel=Llama-2-7B2026.05 | 82.3 | |
| VanillaModel=Llama-2-7B2026.05 | 82.1 | |
| NL-ITIModel=Llama-3-8B2026.05 | 82 | |
| ITIModel=Llama-3-8B2026.05 | 81.3 | |
| MARIModel=Llama-2-13B2026.05 | 73.8 | |
| ReFTModel=Llama-2-13B2026.05 | 72.8 | |
| CAAModel=Llama-2-13B2026.05 | 72 | |
| NL-ITIModel=Llama-2-13B2026.05 | 71.5 | |
| VanillaModel=Llama-2-13B2026.05 | 71.4 | |
| CAAModel=Qwen2.5-32B2026.05 | 66.3 | |
| MARIModel=Qwen2-7B2026.05 | 65.5 | |
| MARIModel=Qwen2.5-32B2026.05 | 65.3 | |
| ReFTModel=Qwen2.5-32B2026.05 | 64.8 | |
| ReFTModel=Qwen2-7B2026.05 | 64.6 | |
| MARIModel=Qwen2.5-14B2026.05 | 64.5 | |
| ITIModel=Qwen2.5-32B2026.05 | 64.5 | |
| CAAModel=Qwen2-7B2026.05 | 64.4 | |
| NL-ITIModel=Qwen2.5-32B2026.05 | 64.2 | |
| NL-ITIModel=Qwen2.5-14B2026.05 | 64 | |
| ReFTModel=Qwen2.5-14B2026.05 | 63.6 | |
| ITIModel=Qwen2-7B2026.05 | 63.5 | |
| CAAModel=Qwen2.5-14B2026.05 | 63.4 | |
| NL-ITIModel=Qwen2-7B2026.05 | 63.3 | |
| ITIModel=Qwen2.5-14B2026.05 | 63.2 | |
| VanillaModel=Qwen2-7B2026.05 | 63.1 | |
| VanillaModel=Qwen2.5-14B2026.05 | 62.6 | |
| VanillaModel=Qwen2.5-32B2026.05 | 62.1 | |
| ITIModel=Llama-2-13B2026.05 | 26.5 |