SAE Activation on Neuronpedia SAE features entropy 3-9
81Win PercentageGCG
Evaluation Results
| Method | Links | |
|---|---|---|
| GCGOptimisation target=mean activation, Comparison baseline=GPT-4o2025.06 | 81 | |
| EPOOptimisation target=mean activation, Comparison baseline=GPT-4o2025.06 | 75.5 | |
| GCGOptimisation target=mean activation, Comparison baseline=Max Act Examples2025.06 | 75.2 | |
| Max Act ExamplesOptimisation target=mean activation, Comparison baseline=GPT-4o2025.06 | 73.7 | |
| GCGOptimisation target=mean activation, Comparison baseline=EPO2025.06 | 68.4 | |
| EPO-AssistOptimisation target=mean activation, Comparison baseline=GPT-4o2025.06 | 68.3 | |
| EPOOptimisation target=mean activation, Comparison baseline=Max Act Examples2025.06 | 67.3 | |
| EPO-AssistOptimisation target=mean activation, Comparison baseline=EPO-Inpaint2025.06 | 64.4 | |
| GCGOptimisation target=mean activation, Comparison baseline=EPO-Assist2025.06 | 62.7 | |
| EPO-InpaintOptimisation target=mean activation, Comparison baseline=GPT-4o2025.06 | 61.8 | |
| GCGOptimisation target=mean activation, Comparison baseline=EPO-Inpaint2025.06 | 59.8 | |
| EPO-AssistOptimisation target=mean activation, Comparison baseline=Max Act Examples2025.06 | 57.8 | |
| EPO-InpaintOptimisation target=mean activation, Comparison baseline=EPO2025.06 | 53.5 | |
| EPO-InpaintOptimisation target=mean activation, Comparison baseline=Max Act Examples2025.06 | 50 | |
| Max Act ExamplesOptimisation target=mean activation, Comparison baseline=EPO-Inpaint2025.06 | 50 | |
| EPO-AssistOptimisation target=mean activation, Comparison baseline=EPO2025.06 | 48.5 | |
| EPOOptimisation target=mean activation, Comparison baseline=EPO-Assist2025.06 | 47.5 | |
| EPOOptimisation target=mean activation, Comparison baseline=EPO-Inpaint2025.06 | 46.5 | |
| Max Act ExamplesOptimisation target=mean activation, Comparison baseline=EPO-Assist2025.06 | 41.2 | |
| EPO-InpaintOptimisation target=mean activation, Comparison baseline=GCG2025.06 | 39.2 | |
| EPO-AssistOptimisation target=mean activation, Comparison baseline=GCG2025.06 | 37.3 | |
| GPT-4oOptimisation target=mean activation, Comparison baseline=EPO-Inpaint2025.06 | 37.3 | |
| EPO-InpaintOptimisation target=mean activation, Comparison baseline=EPO-Assist2025.06 | 34.7 | |
| Max Act ExamplesOptimisation target=mean activation, Comparison baseline=EPO2025.06 | 32.7 | |
| GPT-4oOptimisation target=mean activation, Comparison baseline=EPO-Assist2025.06 | 31.7 | |
| EPOOptimisation target=mean activation, Comparison baseline=GCG2025.06 | 29.6 | |
| GPT-4oOptimisation target=mean activation, Comparison baseline=Max Act Examples2025.06 | 26.3 | |
| Max Act ExamplesOptimisation target=mean activation, Comparison baseline=GCG2025.06 | 24.8 | |
| GPT-4oOptimisation target=mean activation, Comparison baseline=EPO2025.06 | 24.5 | |
| GPT-4oOptimisation target=mean activation, Comparison baseline=GCG2025.06 | 18 |