Jailbreak Robustness on Jailbreak Attacks
88.8Prefill Success RateUnaligned
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| UnalignedBackbone=Qwen3-8B2026.05 | 88.8 | 68.8 | — | |
| UnalignedBackbone=LLaMA-3-8B2026.05 | 88.2 | 67.6 | — | |
| UnalignedBackbone=Yi-1.5-9B2026.05 | 82 | 67.1 | — | |
| Sqrt-CompetenceBackbone=Qwen3-8B2026.05 | 59 | 27.6 | 4 | |
| SequentialBackbone=Qwen3-8B2026.05 | 52.8 | 26.1 | 0.2 | |
| Standard DPOBackbone=Qwen3-8B2026.05 | 51.7 | 26.9 | — | |
| Standard DPOBackbone=LLaMA-3-8B2026.05 | 46.5 | 23.6 | — | |
| Curri-DPOBackbone=LLaMA-3-8B2026.05 | 36.8 | 17.1 | -8.1 | |
| SequentialBackbone=LLaMA-3-8B2026.05 | 31 | 17.8 | -10.7 | |
| Curri-DPOBackbone=Qwen3-8B2026.05 | 30.5 | 24.1 | -12 | |
| Sqrt-CompetenceBackbone=LLaMA-3-8B2026.05 | 29.2 | 15.6 | -12.7 | |
| Standard DPOBackbone=Yi-1.5-9B2026.05 | 26.5 | 11.8 | — | |
| Staged-CompetenceBackbone=LLaMA-3-8B2026.05 | 24.2 | 8.3 | -18.8 | |
| Curri-DPOBackbone=Yi-1.5-9B2026.05 | 22.5 | 5 | -5.4 | |
| SequentialBackbone=Yi-1.5-9B2026.05 | 19.2 | 6.8 | -6.2 | |
| Staged-CompetenceBackbone=Qwen3-8B2026.05 | 16.2 | 8.3 | -27.1 | |
| Sqrt-CompetenceBackbone=Yi-1.5-9B2026.05 | 15.2 | 5 | -9.1 | |
| Staged-CompetenceBackbone=Yi-1.5-9B2026.05 | 9.2 | 1.5 | -13.8 |