Fact-checking on Causal and Downstream Robustness Ablation Suite Averaged over 4 models
3.7Fact EMΔHETA
Evaluation Results
| Method | Links | |
|---|---|---|
| HETAMethod Variant=Full2026.04 | 3.7 | |
| HETAMethod Variant=LR+WIN2026.04 | 3.3 | |
| HETAMethod Variant=w/o Hessian2026.04 | 2.5 | |
| HETAMethod Variant=w/o KL2026.04 | 2 | |
| ReAGentMethod Variant=Standard2026.04 | 2 | |
| SEA-CoTMethod Variant=Standard2026.04 | 1.7 | |
| HETAMethod Variant=w/o Transition2026.04 | 1.6 | |
| Progressive InferenceMethod Variant=Standard2026.04 | 1.5 | |
| fAMLMethod Variant=Standard2026.04 | 1.3 | |
| ContextCiteMethod Variant=Standard2026.04 | 1.2 | |
| TDD-backwardMethod Variant=Standard2026.04 | 1.1 | |
| Peering (PML)Method Variant=Standard2026.04 | 1 | |
| Integrated GradientsMethod Variant=Standard2026.04 | 0.9 | |
| Attention RolloutMethod Variant=Standard2026.04 | 0.5 |