Feature Visualization on Gemma Scope Gemma 2 2B 16k-wide residual stream 486-latents SAE (train/test)
84Win RateADAPT
Evaluation Results
| Method | Links | |
|---|---|---|
| ADAPTComparison baseline=BEAST2026.02 | 84 | |
| ADAPTComparison baseline=EPO2026.02 | 78.6 | |
| ADAPTComparison baseline=GCG2026.02 | 70.4 | |
| GCGComparison baseline=EPO2026.02 | 63.2 | |
| GCGComparison baseline=BEAST2026.02 | 56.4 | |
| BEASTComparison baseline=EPO2026.02 | 50.2 | |
| EPOComparison baseline=BEAST2026.02 | 49.8 | |
| BEASTComparison baseline=GCG2026.02 | 43.6 | |
| EPOComparison baseline=GCG2026.02 | 36.8 | |
| GCGComparison baseline=ADAPT2026.02 | 29.6 | |
| EPOComparison baseline=ADAPT2026.02 | 21.4 | |
| BEASTComparison baseline=ADAPT2026.02 | 12.8 |