Spurious Correlation Removal on SAEBench Spurious Correlation Removal Pythia-70M activations
2.1Bias (Professor vs Nurse)Ensembling (NB)
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Ensembling (NB)Ensemble size=8, Approach=Naive Bagging2025.05 | 2.1 | 0.4 | 1.4 | 0.3 | |
| Expanded SAE2025.05 | 4.7 | 0.6 | 3.7 | 2.1 | |
| Ensembling (Boosting)Ensemble size=8, Approach=Boosting2025.05 | 6.6 | 1.3 | 4.5 | 2.9 |