Multi-Turn Jailbreak Robustness on Actor Attack/MUSE-A/TRACE Scenarios
47.17Actor Attack ASRQwen2.5-7B-IT
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Qwen2.5-7B-IT2026.05 | 47.17 | 51.57 | 87.84 | 58.63 | |
| MUSE-DBackbone=Qwen2.5-7B-IT2026.05 | 30.81 | 6.28 | 50.94 | 29.34 | |
| TRACEBackbone=Qwen2.5-7B-IT, variant=w/o lat.2026.05 | 17.61 | 1.26 | 10.06 | 9.64 | |
| TRACEBackbone=Qwen2.5-7B-IT, variant=ours2026.05 | 15.72 | 1.89 | 9.43 | 9.01 | |
| SafeMTBackbone=Qwen2.5-7B-IT2026.05 | 3.14 | 9.43 | 28.3 | 13.62 |