Circuit discovery on Duplicate-token
98ROC-AUCConditional Co-ablation
Evaluation Results
| Method | Links | |
|---|---|---|
| Conditional Co-ablationModel=GPT-2-sm, Signal=co-ablation2026.07 | 98 | |
| Conditional Co-ablationModel=Gemma-2-9B, Signal=co-ablation2026.07 | 96 | |
| Conditional Co-ablationModel=GPT-Neo-1.3B, Signal=co-ablation2026.07 | 94 | |
| Conditional Co-ablationModel=OLMo-2-7B, Signal=co-ablation2026.07 | 86 | |
| Conditional Co-ablationModel=GPT-2-lg, Signal=co-ablation2026.07 | 84 | |
| Conditional Co-ablationModel=Gemma-2-2B, Signal=co-ablation2026.07 | 72 | |
| Co-activationModel=OLMo-2-7B, Signal=co-activation2026.07 | 71 | |
| Conditional Co-ablationModel=Llama-3.1-8B, Signal=co-ablation2026.07 | 69 | |
| Co-activationModel=Llama-3.1-8B, Signal=co-activation2026.07 | 68 | |
| Conditional Co-ablationModel=Pythia-160M, Signal=co-ablation2026.07 | 65 | |
| Conditional Co-ablationModel=Qwen2.5-7B, Signal=co-ablation2026.07 | 65 | |
| Co-activationModel=Gemma-2-9B, Signal=co-activation2026.07 | 65 | |
| Conditional Co-ablationModel=Pythia-410M, Signal=co-ablation2026.07 | 62 | |
| Co-activationModel=Qwen2.5-7B, Signal=co-activation2026.07 | 62 | |
| Co-activationModel=Pythia-1.4B, Signal=co-activation2026.07 | 61 | |
| Co-activationModel=Pythia-410M, Signal=co-activation2026.07 | 56 | |
| Conditional Co-ablationModel=Pythia-1.4B, Signal=co-ablation2026.07 | 55 | |
| Conditional Co-ablationModel=GPT-2-md, Signal=co-ablation2026.07 | 54 | |
| Co-activationModel=GPT-2-md, Signal=co-activation2026.07 | 46 | |
| Co-activationModel=Gemma-2-2B, Signal=co-activation2026.07 | 43 | |
| Co-activationModel=GPT-2-lg, Signal=co-activation2026.07 | 39 | |
| Co-activationModel=Pythia-160M, Signal=co-activation2026.07 | 38 | |
| Co-activationModel=GPT-2-sm, Signal=co-activation2026.07 | 37 | |
| Co-activationModel=GPT-Neo-1.3B, Signal=co-activation2026.07 | 37 |