Circuit Discovery on Induction
99ROC-AUCCo-activation
Evaluation Results
| Method | Links | |
|---|---|---|
| Co-activationModel=GPT-2-lg, Signal=co-activation, Calibration=repeated-token2026.07 | 99 | |
| Co-activationModel=GPT-2-md, Signal=co-activation, Calibration=repeated-token2026.07 | 98 | |
| Co-activationModel=GPT-Neo-1.3B, Signal=co-activation, Calibration=repeated-token2026.07 | 97 | |
| Co-activationModel=Pythia-410M, Signal=co-activation, Calibration=repeated-token2026.07 | 95 | |
| Co-activationModel=Pythia-160M, Signal=co-activation, Calibration=repeated-token2026.07 | 94 | |
| Co-activationModel=GPT-2-sm, Signal=co-activation, Calibration=repeated-token2026.07 | 94 | |
| Co-activationModel=Qwen2.5-7B, Signal=co-activation, Calibration=repeated-token2026.07 | 93 | |
| Co-activationModel=Llama-3.1-8B, Signal=co-activation, Calibration=repeated-token2026.07 | 92 | |
| Co-activationModel=Gemma-2-9B, Signal=co-activation, Calibration=repeated-token2026.07 | 89 | |
| Co-activationModel=Gemma-2-2B, Signal=co-activation, Calibration=repeated-token2026.07 | 87 | |
| Co-activationModel=OLMo-2-7B, Signal=co-activation, Calibration=repeated-token2026.07 | 86 | |
| Co-activationModel=Pythia-1.4B, Signal=co-activation, Calibration=repeated-token2026.07 | 83 | |
| Conditional Co-ablationModel=GPT-2-lg, Signal=co-ablation, Calibration=repeated-token2026.07 | 79 | |
| Conditional Co-ablationModel=Qwen2.5-7B, Signal=co-ablation, Calibration=repeated-token2026.07 | 71 | |
| Conditional Co-ablationModel=Llama-3.1-8B, Signal=co-ablation, Calibration=repeated-token2026.07 | 58 | |
| Conditional Co-ablationModel=Pythia-410M, Signal=co-ablation, Calibration=repeated-token2026.07 | 53 | |
| Conditional Co-ablationModel=GPT-2-md, Signal=co-ablation, Calibration=repeated-token2026.07 | 52 | |
| Conditional Co-ablationModel=GPT-2-sm, Signal=co-ablation, Calibration=repeated-token2026.07 | 48 | |
| Conditional Co-ablationModel=Pythia-160M, Signal=co-ablation, Calibration=repeated-token2026.07 | 43 | |
| Conditional Co-ablationModel=GPT-Neo-1.3B, Signal=co-ablation, Calibration=repeated-token2026.07 | 40 | |
| Conditional Co-ablationModel=OLMo-2-7B, Signal=co-ablation, Calibration=repeated-token2026.07 | 40 | |
| Conditional Co-ablationModel=Pythia-1.4B, Signal=co-ablation, Calibration=repeated-token2026.07 | 39 | |
| Conditional Co-ablationModel=Gemma-2-2B, Signal=co-ablation, Calibration=repeated-token2026.07 | 27 | |
| Conditional Co-ablationModel=Gemma-2-9B, Signal=co-ablation, Calibration=repeated-token2026.07 | 26 |