SAE Feature Attribution on Gemma-3-270M 16K SAE
17.3Input Success Rate (%)Query Lens
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Query LensTarget Component=Key2026.05 | 17.3 | 0.86 | |
| Query LensTarget Component=Value2026.05 | 8.54 | 10.26 | |
| Token Changealpha=102026.05 | 7.66 | 9.36 | |
| Token Changealpha=12026.05 | 7.55 | 7.11 | |
| Token Changealpha=52026.05 | 7.54 | 8.99 | |
| Zero-Out2026.05 | 5.25 | 4.44 | |
| Logit LensTarget Component=Key2026.05 | 2.76 | 1.55 | |
| Token Likelihood2026.05 | 1.76 | 4.96 | |
| Logit LensTarget Component=Value2026.05 | 1.65 | 5.46 |