Safety Evaluation on DeepInception (test)
78.5ASRNo Defense
Evaluation Results
| Method | Links | |
|---|---|---|
| No DefenseBackbone=Qwen2.5-14B-Instruct2026.01 | 78.5 | |
| PPLBackbone=Qwen2.5-14B-Instruct2026.01 | 73 | |
| ICDBackbone=Qwen2.5-14B-Instruct2026.01 | 50.8 | |
| Self-ReminderBackbone=Qwen2.5-14B-Instruct2026.01 | 43.2 | |
| Self-ExaminationBackbone=Qwen2.5-14B-Instruct2026.01 | 5.6 | |
| SafeDecodingBackbone=Qwen2.5-14B-Instruct2026.01 | 0.8 | |
| SafeThinkerBackbone=Qwen2.5-14B-Instruct2026.01 | 0.8 |