Safety Evaluation on FalseReject
79.2USRBenign RateNo Defense
Evaluation Results
| Method | Links | |
|---|---|---|
| No DefenseBase Model=Vicuna2026.02 | 79.2 | |
| Self-ExamBase Model=Vicuna2026.02 | 70.7 | |
| SSDBase Model=Vicuna2026.02 | 66.2 | |
| NGSDBase Model=Vicuna2026.02 | 65 | |
| Self-ReminderBase Model=Vicuna2026.02 | 63.9 | |
| No DefenseBase Model=Qwen32026.02 | 60 | |
| Self-ExamBase Model=Qwen32026.02 | 51.1 | |
| SafeDecodingBase Model=Vicuna2026.02 | 49.9 | |
| SafeDecodingBase Model=Qwen32026.02 | 49.9 | |
| SSDBase Model=Qwen32026.02 | 46.1 | |
| Self-ReminderBase Model=Qwen32026.02 | 45.8 | |
| NGSDBase Model=Qwen32026.02 | 45.1 | |
| SafeDecodingBase Model=Llama32026.02 | 29.7 | |
| No DefenseBase Model=Llama32026.02 | 22 | |
| Self-ReminderBase Model=Llama32026.02 | 16.8 | |
| NGSDBase Model=Llama32026.02 | 16.5 | |
| Self-ExamBase Model=Llama32026.02 | 15.7 | |
| SSDBase Model=Llama32026.02 | 11 |