Linguistic-space calibration on TruthfulQA
70.4Faithfulness Divergence ReductionRALC
Evaluation results
| Method | Links | ||
|---|---|---|---|
| RALCSignal=Token Probability, Calibration Training Domain=SQuAD 2.02026.05 | 70.4 | 61.4 | |
| RALCSignal=Semantic Uncertainty, Calibration Training Domain=TruthfulQA2026.05 | 66.1 | 45 | |
| RALCSignal=Semantic Uncertainty, Calibration Training Domain=SQuAD 2.02026.05 | 65.5 | 49.8 | |
| RALCSignal=Token Probability, Calibration Training Domain=TruthfulQA2026.05 | 64.6 | 42.3 | |
| RALCSignal=Linguistic Confidence, Calibration Training Domain=TruthfulQA2026.05 | 60.5 | 39.6 | |
| RALCSignal=Linguistic Confidence, Calibration Training Domain=SQuAD 2.02026.05 | 59.8 | 35 | |
| RALCSignal=Token Probability, Calibration Training Domain=MMLU2026.05 | 56.4 | 37.5 | |
| RALCSignal=Semantic Uncertainty, Calibration Training Domain=MMLU2026.05 | 43 | 33.7 | |
| RALCSignal=Linguistic Confidence, Calibration Training Domain=MMLU2026.05 | 29.6 | 12.2 |