SAE Feature Attribution on Gemma-3-1B 16K SAE
14.71Input AttributionQuery Lens
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Query LensTarget Component=Key2026.05 | 14.71 | 2.32 | |
| Query LensTarget Component=Value2026.05 | 13.39 | 10.87 | |
| Token Changealpha=102026.05 | 13.28 | 9.6 | |
| Token Changealpha=52026.05 | 13.08 | 9.26 | |
| Token Changealpha=12026.05 | 11.75 | 7.22 | |
| Zero-Out2026.05 | 10.76 | 6.81 | |
| Logit LensTarget Component=Key2026.05 | 3.45 | 5.67 | |
| Logit LensTarget Component=Value2026.05 | 1.82 | 9.61 | |
| Token Likelihood2026.05 | 1.82 | 9.58 |