Alignment Robustness Evaluation on CLINC-150 and 20 Newsgroups (averaged)
14.2Alignment Depth DDeep Alignment Training
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Deep Alignment TrainingBackbone=Qwen2.5-32B2025.12 | 14.2 | 0.94 | 2.2 | 6.9 | 12 | |
| Deep Alignment TrainingBackbone=Qwen3-4B2025.12 | 13.6 | 0.92 | 2.4 | 7.2 | 10 | |
| Deep Alignment TrainingBackbone=Qwen2.5-3B2025.12 | 13.1 | 0.91 | 2.7 | 7.6 | 9 | |
| Deep Alignment TrainingBackbone=Qwen3-1.7B2025.12 | 12.5 | 0.89 | 3.1 | 8.3 | 8 | |
| Standard TrainingBackbone=Qwen2.5-32B2025.12 | 3.2 | 0.71 | 11 | 89.1 | — | |
| Standard TrainingBackbone=Qwen3-4B2025.12 | 3.1 | 0.69 | 11.5 | 92.4 | — | |
| Standard TrainingBackbone=Qwen2.5-3B2025.12 | 2.9 | 0.67 | 12.1 | 95.8 | — | |
| Standard TrainingBackbone=Qwen3-1.7B2025.12 | 2.8 | 0.65 | 12.5 | 98.2 | — |