Calibration on DeepSeek models Aggregated datasets
0.067ECE2Test-time self-consistency (TT SC)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Test-time self-consistency (TT SC)2026.04 | 0.067 | 0.116 | |
| Unsupervised Confidence CalibrationEmbeddings=Internal2026.04 | 0.076 | 0.14 | |
| Unsupervised Confidence CalibrationEmbeddings=External (Black-Box)2026.04 | 0.077 | 0.139 | |
| Supervised Platt scaling (Sup.)2026.04 | 0.083 | 0.17 | |
| Verbalized Confidence (VC)2026.04 | 0.243 | 0.222 | |
| Token Probs. (TP)2026.04 | 0.25 | 0.234 | |
| Answer Probs. (Ans. TP)2026.04 | 0.336 | 0.259 |