Sycophancy Reduction (held-in and held-out)
79Activation PatchingDifference-in-means
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Difference-in-meansModel=SOLAR-10.7B, alpha=10, k=12, N=1202026.02 | 79 | 78 | 70 | 71 | 60 | |
| Difference-in-meansModel=Qwen-14B, alpha=10, k=12, N=1202026.02 | 78 | 80 | 79 | 66 | 64 | |
| Mean steeringModel=Qwen-14B, alpha=10, k=12, N=1202026.02 | 77 | 77 | 24 | 77 | 59 | |
| Difference-in-meansModel=OLMo-13B, alpha=10, k=12, N=1202026.02 | 70 | 69 | 58 | 64 | 53 | |
| Mean steeringModel=OLMo-13B, alpha=10, k=12, N=1202026.02 | 67 | 70 | 44 | 48 | 56 | |
| ReFTModel=SOLAR-10.7B, alpha=10, k=12, N=1202026.02 | 58 | 57 | 56 | 58 | 57 | |
| ReFTModel=OLMo-13B, alpha=10, k=12, N=1202026.02 | 53 | 51 | 31 | 51 | 42 | |
| Mean steeringModel=SOLAR-10.7B, alpha=10, k=12, N=1202026.02 | 46 | 44 | 11 | 61 | 39 | |
| ReFTModel=Qwen-14B, alpha=10, k=12, N=1202026.02 | 41 | 39 | 22 | 40 | 21 |