Safety Evaluation on AutoDAN
100Safety ScoreREFLECTOR (+GDPO)
Evaluation Results
| Method | Links | |
|---|---|---|
| REFLECTOR (+GDPO)Backbone=Llama-3.1-8B-Instruct2026.05 | 100 | |
| STAIRBackbone=Llama-3.1-8B-Instruct2026.05 | 99.04 | |
| REFLECTOR (+SFT)Backbone=Llama-3.1-8B-Instruct2026.05 | 98.26 | |
| REFLECTOR (+GDPO)Backbone=Qwen-2.5-7B-Instruct2026.05 | 97.89 | |
| Shallow-AlignBackbone=Llama-3.1-8B-Instruct2026.05 | 96.8 | |
| Self-CritiqueBackbone=Llama-3.1-8B-Instruct2026.05 | 96.15 | |
| DPOBackbone=Llama-3.1-8B-Instruct2026.05 | 95.38 | |
| STAIRBackbone=Qwen-2.5-7B-Instruct2026.05 | 95.19 | |
| OriginalBackbone=Llama-3.1-8B-Instruct2026.05 | 94.75 | |
| SFTBackbone=Llama-3.1-8B-Instruct2026.05 | 94.62 | |
| Shallow-AlignBackbone=Qwen-2.5-7B-Instruct2026.05 | 91.8 | |
| REFLECTOR (SFT)Backbone=Qwen-2.5-7B-Instruct2026.05 | 90.38 | |
| OriginalBackbone=Qwen-2.5-7B-Instruct2026.05 | 52.89 | |
| DPOBackbone=Qwen-2.5-7B-Instruct2026.05 | 49.04 | |
| Self-CritiqueBackbone=Qwen-2.5-7B-Instruct2026.05 | 43.7 | |
| SFTBackbone=Qwen-2.5-7B-Instruct2026.05 | 38.46 |