Multi-turn Jailbreak on HarmBench
100ASR (X-Teaming)No Defense
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| No DefenseModel=Qwen2.5-7B, Defense=No Defense2026.06 | 100 | 84.3 | |
| No DefenseModel=Llama3-8B, Defense=No Defense2026.06 | 86.7 | 84.7 | |
| SAGEModel=Qwen2.5-7B, Defense=SAGE2026.06 | 86.2 | 4.9 | |
| PROACTModel=Qwen2.5-7B, Defense=PROACT2026.06 | 63.5 | 2.6 | |
| PROACTModel=Llama3-8B, Defense=PROACT2026.06 | 47.4 | 2.4 | |
| SAGEModel=Llama3-8B, Defense=SAGE2026.06 | 17.5 | 6.7 | |
| THRDModel=Qwen2.5-7B, Defense=Ours2026.06 | 4 | 2.1 | |
| THRDModel=Llama3-8B, Defense=Ours2026.06 | 1.6 | 0.6 |