SAE Pareto Dominance Analysis on Pythia-1B activations
100Dominance FractionSA-half (SA-GSAE)
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| SA-half (SA-GSAE)Hookpoint=mlp_out, Overlap L0=[15.3, 144.4]2026.05 | 100 | 155 | 145 | 671 | 98.5 | |
| SA-half (SA-GSAE)Hookpoint=resid, Overlap L0=[17.9, 129.3]2026.05 | 100 | 191 | 492 | 1.49 | 129.3 | |
| SA-half (SA-GSAE)Hookpoint=attn, Overlap L0=[13.7, 97.8]2026.05 | 26 | 1.5 | 1.5 | 1.6 | 67.7 |