SAE Feature Attribution on GPT-2 Small 128K SAE
32.16Input Attribution (%)Query Lens
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Query LensTarget Component=Key2026.05 | 32.16 | 1.44 | |
| Token Changealpha=102026.05 | 22.14 | 11.89 | |
| Token Changealpha=52026.05 | 22.01 | 12.51 | |
| Token Changealpha=12026.05 | 21.19 | 13.17 | |
| Query LensTarget Component=Value2026.05 | 17.73 | 14.43 | |
| Logit LensTarget Component=Value2026.05 | 7.21 | 11.75 | |
| Logit LensTarget Component=Key2026.05 | 6.44 | 3.21 | |
| Token Likelihood2026.05 | 5.36 | 11.32 | |
| Zero-Out2026.05 | 2.44 | 3.21 |