Sycophancy Evaluation on Syco-Bench
1.21Pickside ScoreUntrained Gemma-2-9B
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Untrained Gemma-2-9BModel=Gemma-2-9B2026.01 | 1.21 | 4.25 | 0.98 | 3 | |
| Ours ResidBase Model=Gemma-2-2B, Probe Type=Residual2026.01 | 0.8 | 2.35 | 0.77 | 3.35 | |
| Supervised Pinpoint TuningBase Model=Gemma-2-2B2026.01 | 0.7 | 4.34 | -0.04 | 2.5 | |
| Ours ResidModel=Gemma-2-9B, Probe Type=Residual2026.01 | 0.64 | 1.51 | 0.13 | 2.2 | |
| Ours SAEModel=Gemma-2-9B, Probe Type=SAE2026.01 | 0.42 | 1.8 | 0.54 | 3.43 | |
| Supervised Pinpoint TuningModel=Gemma-2-9B2026.01 | 0.33 | 3.64 | 0.67 | 2.3 | |
| Ours SAEBase Model=Gemma-2-2B, Probe Type=SAE2026.01 | 0.23 | 2.68 | 0.66 | 3.3 | |
| Untrained Gemma-2-2BBase Model=Gemma-2-2B2026.01 | -0.28 | 4.39 | 0.53 | 2.9 | |
| Synthetic Data InterventionModel=Gemma-2-9B2026.01 | -0.89 | 5.22 | 0.99 | 3.55 | |
| Synthetic Data InterventionBase Model=Gemma-2-2B2026.01 | -1.82 | -0.36 | -0.74 | 3.52 |