SAE Feature Attribution on Gemma-3-4B 65K SAE
11.14Input AttributionQuery Lens
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Query LensTarget Component=Key2026.05 | 11.14 | 0.89 | |
| Query LensTarget Component=Value2026.05 | 9.54 | 3.53 | |
| Token Changealpha=102026.05 | 9.31 | 3.2 | |
| Token Changealpha=52026.05 | 8.59 | 3.24 | |
| Token Changealpha=12026.05 | 6.04 | 2.04 | |
| Zero-Out2026.05 | 5.79 | 1.76 | |
| Logit LensTarget Component=Key2026.05 | 2.66 | 2.26 | |
| Token Likelihood2026.05 | 1.99 | 2.79 | |
| Logit LensTarget Component=Value2026.05 | 1.91 | 2.81 |