Causal Faithfulness Evaluation on Edgar Domain
0.988nAOPCGAE
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| GAEModel=Pythia-1.4B2026.05 | 0.988 | 1.652 | 0.023 | |
| GAEModel=GPT-2 Small2026.05 | 0.981 | 1.618 | 0.0009 | |
| SAEBoostModel=GPT-2 Small2026.05 | 0.979 | 1.542 | 0.0072 | |
| FaithfulSAEModel=GPT-2 Small2026.05 | 0.976 | 1.134 | 0.0197 | |
| FixedModel=GPT-2 Small2026.05 | 0.975 | 1.025 | 0.0201 | |
| FinetuneModel=GPT-2 Small2026.05 | 0.971 | 1.117 | 0.0047 | |
| RetrainModel=Pythia-1.4B2026.05 | 0.97 | 1.848 | 0.0261 | |
| GAEModel=Pythia-1.4B2026.05 | 0.968 | 1.946 | 0.0098 | |
| TERMModel=GPT-2 Small2026.05 | 0.964 | 0.999 | 0.0198 | |
| TERMModel=Pythia-1.4B2026.05 | 0.959 | 1.787 | 0.0167 | |
| FinetuneModel=Pythia-1.4B2026.05 | 0.959 | 1.832 | 0.0131 | |
| SAEBoostModel=Pythia-1.4B2026.05 | 0.955 | 1.598 | 0.0152 | |
| FixedModel=Pythia-1.4B2026.05 | 0.953 | 1.69 | 0.017 | |
| FaithfulSAEModel=Pythia-1.4B2026.05 | 0.953 | 1.769 | 0.0205 | |
| RetrainModel=GPT-2 Small2026.05 | 0.936 | 1.034 | 0.0015 | |
| SAEBoostModel=Pythia-1.4B2026.05 | 0.903 | 1.297 | 0.0296 | |
| TERMModel=Pythia-1.4B2026.05 | 0.895 | 0.981 | 0.0298 | |
| RetrainModel=Pythia-1.4B2026.05 | 0.746 | 0.821 | 0.0329 | |
| FixedModel=Pythia-1.4B2026.05 | 0.725 | 0.828 | 0.03 | |
| FaithfulSAEModel=Pythia-1.4B2026.05 | 0.724 | 0.822 | 0.0323 | |
| GAEModel=GPT-2 Small2026.05 | 0.723 | 0.809 | 0.0037 | |
| RetrainModel=GPT-2 Small2026.05 | 0.715 | 0.797 | 0.0065 | |
| FinetuneModel=Pythia-1.4B2026.05 | 0.684 | 0.86 | 0.0282 | |
| FinetuneModel=GPT-2 Small2026.05 | 0.658 | 0.687 | 0.0068 | |
| FaithfulSAEModel=GPT-2 Small2026.05 | 0.657 | 0.683 | 0.0074 | |
| TERMModel=GPT-2 Small2026.05 | 0.655 | 0.682 | 0.0145 | |
| FixedModel=GPT-2 Small2026.05 | 0.65 | 0.667 | 0.0089 | |
| SAEBoostModel=GPT-2 Small2026.05 | 0.604 | 0.65 | 0.0098 |