Feature absorption on first-letter benchmark
37.2Mean Fractional AbsorptionStandard SAE
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Standard SAEBackbone=GPT-2, Architecture=Standard2026.06 | 37.2 | 49.6 | |
| Standard SAEBackbone=Mistral-7B, Architecture=Standard2026.06 | 24 | 23.3 | |
| SASABackbone=Mistral-7B, Architecture=SASA2026.06 | 18.3 | 11.9 | |
| SASABackbone=GPT-2, Architecture=SASA2026.06 | 6.6 | 4.7 |