Fine-grained Moral Steering on AITA
0.966RhoCDR
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| CDRModel=Llama2026.05 | 0.966 | 0.6 | |
| BL-PRSModel=Llama2026.05 | 0.952 | 1.6 | |
| Top-K HeadModel=Llama2026.05 | 0.936 | 1.7 | |
| CDRModel=Vicuna2026.05 | 0.928 | 3.5 | |
| Prompt-OnlyModel=Llama2026.05 | 0.927 | 0 | |
| Prompt-OnlyModel=Vicuna2026.05 | 0.871 | 0.2 | |
| BL-PRSModel=Vicuna2026.05 | 0.866 | 0.4 | |
| Top-K HeadModel=Vicuna2026.05 | 0.86 | 1.1 |