Sycophancy Evaluation on POLI
92.18Sycophantic Preference (%)Ours Resid
Evaluation Results
| Method | Links | |
|---|---|---|
| Ours ResidModel=Gemma-2-9B, Probe Type=Residual2026.01 | 92.18 | |
| Ours ResidBase Model=Gemma-2-2B, Probe Type=Residual2026.01 | 86.41 | |
| Ours SAEModel=Gemma-2-9B, Probe Type=SAE2026.01 | 86.13 | |
| Ours SAEBase Model=Gemma-2-2B, Probe Type=SAE2026.01 | 79.6 | |
| Synthetic Data InterventionModel=Gemma-2-9B2026.01 | 74.59 | |
| Untrained Gemma-2-9BModel=Gemma-2-9B2026.01 | 74.2 | |
| Supervised Pinpoint TuningModel=Gemma-2-9B2026.01 | 73.95 | |
| Untrained Gemma-2-2BBase Model=Gemma-2-2B2026.01 | 50.22 | |
| Supervised Pinpoint TuningBase Model=Gemma-2-2B2026.01 | 50.12 | |
| Synthetic Data InterventionBase Model=Gemma-2-2B2026.01 | 49.14 |