CoT faithfulness detection on AQuA
77Accuracy (CoT Faithfulness)CIE-SCORER
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| CIE-SCORERParadigm=Circuit-based2026.05 | 77 | 72.8 | |
| BiGGenParadigm=LLM-as-judge2026.05 | 73.7 | 70.3 | |
| Removing StepsParadigm=Counterfactual-based2026.05 | 72.4 | 46.2 | |
| Early AnsweringParadigm=Counterfactual-based2026.05 | 72.4 | 53.3 | |
| Adding MistakesParadigm=Counterfactual-based2026.05 | 71.6 | 66.7 | |
| ParaphrasingParadigm=Counterfactual-based2026.05 | 68.4 | 42.9 | |
| CRVParadigm=Circuit-based2026.05 | 63 | 62.8 | |
| Option ShufflingParadigm=Counterfactual-based2026.05 | 61 | 16.7 | |
| Answer TracingParadigm=Logits-based2026.05 | 53.2 | 30.8 | |
| PerplexityParadigm=Baselines2026.05 | 49.4 | 36.1 | |
| RandomParadigm=Baselines2026.05 | 43.2 | 37.4 | |
| Information GainParadigm=Logits-based2026.05 | 41.6 | 20.2 |