Safety Alignment Evaluation on OOD Safety Suite Average
0.5Average Absolute ImprovementSqrt-Competence
Evaluation Results
| Method | Links | |
|---|---|---|
| Sqrt-CompetenceBackbone=Qwen3-8B2026.05 | 0.5 | |
| Curri-DPOBackbone=Yi-1.5-9B2026.05 | -4.3 | |
| SequentialBackbone=Yi-1.5-9B2026.05 | -5.5 | |
| Sqrt-CompetenceBackbone=Yi-1.5-9B2026.05 | -5.6 | |
| Curri-DPOBackbone=LLaMA-3-8B2026.05 | -6.4 | |
| Staged-CompetenceBackbone=Yi-1.5-9B2026.05 | -7.1 | |
| SequentialBackbone=Qwen3-8B2026.05 | -7.2 | |
| Sqrt-CompetenceBackbone=LLaMA-3-8B2026.05 | -9 | |
| Curri-DPOBackbone=Qwen3-8B2026.05 | -10 | |
| SequentialBackbone=LLaMA-3-8B2026.05 | -11.4 | |
| Staged-CompetenceBackbone=LLaMA-3-8B2026.05 | -12.2 | |
| Staged-CompetenceBackbone=Qwen3-8B2026.05 | -28.9 |