Sentence-Level Confidence Prediction on FEVER
0.737AUROCVERDI LR ensemble
Evaluation results
| Method | Links | ||
|---|---|---|---|
| VERDI LR ensembleCalls=1×2026.05 | 0.737 | — | |
| Best single signalCalls=1×2026.05 | 0.718 | — | |
| SelfCheckGPT2026.02 | 0.7 | 0.364 | |
| Structural Signal2026.02 | 0.683 | 0.348 | |
| Semantic-feature2026.02 | 0.682 | 0.349 | |
| CoT-UQCalls=5×2026.05 | 0.582 | — | |
| Trace lengthCalls=1×2026.05 | 0.574 | — | |
| Structure-feature2026.02 | 0.573 | 0.274 | |
| Spectral Stability2026.02 | 0.555 | 0.268 | |
| Shape Coherence2026.02 | 0.552 | 0.263 | |
| Local Variation2026.02 | 0.534 | 0.247 | |
| Logprob confidenceCalls=1×2026.05 | 0.523 | — | |
| KMeans-outlier2026.02 | 0.518 | 0.234 | |
| Mean logp2026.02 | 0.438 | 0.208 | |
| Semantic entropyProtocol=NLI, single-pass2026.02 | 0.432 | 0.201 |