Steering effectiveness correlation analysis on Gemma-2-2B concept families (across 26 layers)
0.871Correlation (rho) Alin vs delta PDifference-of-means steering
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Difference-of-means steeringFamily=Word transform, Model=Gemma-2-2B, Number of Layers=262026.04 | 0.871 | -0.744 | 0.809 | 10 | |
| Difference-of-means steeringFamily=Analogy, Model=Gemma-2-2B, Number of Layers=262026.04 | 0.866 | -0.923 | 0.647 | 0.001 | |
| Difference-of-means steeringFamily=Sequence, Model=Gemma-2-2B, Number of Layers=262026.04 | 0.814 | -0.83 | 0.869 | 10 | |
| Difference-of-means steeringFamily=Pooled (n = 130), Model=Gemma-2-2B, Number of Layers=262026.04 | 0.777 | -0.706 | 0.507 | 10 | |
| Difference-of-means steeringFamily=Geography, Model=Gemma-2-2B, Number of Layers=262026.04 | 0.758 | -0.729 | 0.372 | 0.067 | |
| Difference-of-means steeringFamily=Arithmetic, Model=Gemma-2-2B, Number of Layers=262026.04 | 0.719 | -0.99 | 0.688 | 10 |