SAE Feature Activation on GemmaScope 102 features
98.7Win PercentageEPO-Ast.
Evaluation results
| Method | Links | |
|---|---|---|
| EPO-Ast.Opponent=GPT-4o, Target type=normalised max activation, Cross-entropy range=3–92025.06 | 98.7 | |
| EPO-Inp.Opponent=GPT-4o, Target type=normalised max activation2025.06 | 98.6 | |
| EPOOpponent=GPT-4o, Target type=normalised max activation, Cross-entropy range=3–9, Bootstrapped 95% confidence intervals=n=100002025.06 | 97.3 | |
| EPO-Inp.Opponent=Max Act Ex., Target type=normalised max activation2025.06 | 96.9 | |
| EPOOpponent=Max Act Ex., Target type=normalised max activation, Cross-entropy range=3–9, Bootstrapped 95% confidence intervals=n=100002025.06 | 95.9 | |
| EPO-Ast.Opponent=Max Act Ex., Target type=normalised max activation, Cross-entropy range=3–92025.06 | 94.9 | |
| EPO-Ast.Opponent=GCG, Target type=normalised max activation, Cross-entropy range=3–92025.06 | 93.7 | |
| EPOOpponent=GCG, Target type=normalised max activation, Cross-entropy range=3–9, Bootstrapped 95% confidence intervals=n=100002025.06 | 92.4 | |
| EPO-Inp.Opponent=GCG, Target type=normalised max activation2025.06 | 92.4 | |
| GCGOpponent=GPT-4o, Target type=normalised max activation2025.06 | 82.1 | |
| Max Act Ex.Opponent=GPT-4o, Target type=normalised max activation2025.06 | 81.3 | |
| GCGOpponent=Max Act Ex., Target type=normalised max activation2025.06 | 68.8 | |
| EPO-Inp.Opponent=EPO, Target type=normalised max activation2025.06 | 60 | |
| EPO-Ast.Opponent=EPO, Target type=normalised max activation, Cross-entropy range=3–92025.06 | 57 | |
| EPO-Inp.Opponent=EPO-Ast., Target type=normalised max activation2025.06 | 56 | |
| EPO-Ast.Opponent=EPO-Inp., Target type=normalised max activation, Cross-entropy range=3–92025.06 | 42 | |
| EPOOpponent=EPO-Ast., Target type=normalised max activation, Cross-entropy range=3–9, Bootstrapped 95% confidence intervals=n=100002025.06 | 38 | |
| EPOOpponent=EPO-Inp., Target type=normalised max activation, Cross-entropy range=3–9, Bootstrapped 95% confidence intervals=n=100002025.06 | 37 | |
| Max Act Ex.Opponent=GCG, Target type=normalised max activation2025.06 | 31.2 | |
| GPT-4oOpponent=GCG, Target type=normalised max activation2025.06 | 17.9 | |
| GPT-4oOpponent=Max Act Ex., Target type=normalised max activation2025.06 | 17.3 | |
| GCGOpponent=EPO-Inp., Target type=normalised max activation2025.06 | 7.6 | |
| GCGOpponent=EPO, Target type=normalised max activation2025.06 | 6.3 | |
| GCGOpponent=EPO-Ast., Target type=normalised max activation2025.06 | 5.1 | |
| Max Act Ex.Opponent=EPO-Ast., Target type=normalised max activation2025.06 | 5.1 | |
| Max Act Ex.Opponent=EPO, Target type=normalised max activation2025.06 | 4.1 | |
| Max Act Ex.Opponent=EPO-Inp., Target type=normalised max activation2025.06 | 3.1 | |
| GPT-4oOpponent=EPO, Target type=normalised max activation2025.06 | 2.7 | |
| GPT-4oOpponent=EPO-Inp., Target type=normalised max activation2025.06 | 1.4 | |
| GPT-4oOpponent=EPO-Ast., Target type=normalised max activation2025.06 | 1.3 |