Downstream behavior evaluation on Gemma-2-2B layer 12 residual-stream activations
0.093Delta CETopK SAE
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| TopK SAEActivation Function Family=TopK2026.06 | 0.093 | 99.11 | |
| RSAEActivation Function Family=TopK2026.06 | 0.093 | 99.76 | |
| RSAEActivation Function Family=JumpReLU2026.06 | 0.125 | 98.82 | |
| JumpReLU SAEActivation Function Family=JumpReLU2026.06 | 0.127 | 98.79 | |
| RSAEActivation Function Family=ReLU2026.06 | 0.273 | 97.39 | |
| ReLU SAEActivation Function Family=ReLU2026.06 | 0.283 | 97.3 |