Moral Steering on AITA (test)
-15.92Deviation (alpha_U=100%)BL-PRS
Evaluation Results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| BL-PRSBackbone=Llama2026.05 | -15.92 | -12.92 | -12.4 | -9.18 | -5.19 | -0.73 | 3.87 | 7.95 | 11.46 | 12.1 | 14.84 | |
| Top-K HeadBackbone=Llama2026.05 | -3.44 | 2.41 | 4.55 | 7 | 10.04 | 12.96 | 14.84 | 16.24 | 15.71 | 13.11 | 16.88 | |
| OursBackbone=Llama2026.05 | -1.17 | 4.29 | 2.68 | 2.27 | 2.08 | 1.14 | 0.65 | -1.11 | -3.23 | -4.75 | 1.24 | |
| BL-PRSBackbone=Vicuna2026.05 | -0.36 | 8.96 | 16.9 | 22.82 | 27.83 | 31.46 | 29.47 | 23.38 | 14.54 | 2.11 | 1.91 | |
| OursBackbone=Vicuna2026.05 | -0.11 | 9.69 | 13.68 | 6.74 | -2.58 | 1.63 | 1.71 | 1.58 | -4.34 | -9.71 | 0 | |
| Prompt-OnlyBackbone=Vicuna2026.05 | -0.09 | 9.83 | 19.82 | 29.38 | 38.81 | 39.14 | 35.44 | 14.39 | 2.5 | -1.69 | 0 | |
| Prompt-OnlyBackbone=Llama2026.05 | -0.03 | 9.73 | 18.99 | 20.81 | 16.36 | -13.25 | -30.43 | -27.11 | -19.63 | -9.91 | 0 | |
| Top-K HeadBackbone=Vicuna2026.05 | 0 | 10 | 19.31 | 27.49 | 33.57 | 36.91 | 29.12 | 13.09 | -1.78 | -4.63 | 0.38 |