Single-Turn Jailbreak Robustness on Single-Turn Attack Scenarios DAN, WildGuard
8.33DAN ASRTRACE
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| TRACEBackbone=Qwen2.5-7B-IT, variant=ours2026.05 | 8.33 | 7.72 | 0 | |
| TRACEBackbone=Qwen2.5-7B-IT, variant=w/o lat.2026.05 | 8.67 | 8.01 | 0 | |
| SafeMTBackbone=Qwen2.5-7B-IT2026.05 | 12.67 | 41.84 | 8.98 | |
| MUSE-DBackbone=Qwen2.5-7B-IT2026.05 | 14 | 13.35 | 0.24 | |
| Qwen2.5-7B-IT2026.05 | 32.33 | 36.39 | 3.88 |