SAE Activation on Gemma and Llama Scope SAE features 3-9 cross-entropy range
98.7Win PercentageEPO-Assist
Evaluation Results
| Method | Links | |
|---|---|---|
| EPO-AssistOpponent=GPT-4o, Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 98.7 | |
| EPO-InpaintOpponent=GPT-4o, Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 98.6 | |
| EPOOpponent=GPT-4o, Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 97.3 | |
| EPO-InpaintOpponent=Max Act Ex., Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 96.9 | |
| EPOOpponent=Max Act Ex., Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 95.1 | |
| EPO-AssistOpponent=Max Act Ex., Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 94.9 | |
| EPO-AssistOpponent=GCG, Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 93.7 | |
| EPOOpponent=GCG, Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 92.4 | |
| EPO-InpaintOpponent=GCG, Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 92.4 | |
| GCGOpponent=GPT-4o, Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 82.1 | |
| Max Act Ex.Opponent=GPT-4o, Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 81.3 | |
| GCGOpponent=Max Act Ex., Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 68.8 | |
| EPO-InpaintOpponent=EPO, Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 60 | |
| EPO-AssistOpponent=EPO, Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 57 | |
| EPO-InpaintOpponent=EPO-Assist, Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 56 | |
| EPO-AssistOpponent=EPO-Inpaint, Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 42 | |
| EPOOpponent=EPO-Assist, Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 38 | |
| EPOOpponent=EPO-Inpaint, Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 37 | |
| Max Act Ex.Opponent=GCG, Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 31.2 | |
| GPT-4oOpponent=GCG, Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 17.9 | |
| GPT-4oOpponent=Max Act Ex., Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 17.3 | |
| GCGOpponent=EPO-Inpaint, Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 7.6 | |
| GCGOpponent=EPO, Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 6.3 | |
| GCGOpponent=EPO-Assist, Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 5.1 | |
| Max Act Ex.Opponent=EPO-Assist, Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 5.1 | |
| Max Act Ex.Opponent=EPO, Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 4.1 | |
| Max Act Ex.Opponent=EPO-Inpaint, Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 3.1 | |
| GPT-4oOpponent=EPO, Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 2.7 | |
| GPT-4oOpponent=EPO-Inpaint, Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 1.4 | |
| GPT-4oOpponent=EPO-Assist, Optimisation target=maximum activation, Cross-entropy range filter=3-92025.06 | 1.3 |