Calibration on Qwen models benchmark leaderboard · SOTA2 Research

Calibration on Qwen models

0.063ECE2Supervised Platt scaling (Sup.)

Evaluation Results

Method
Supervised Platt scaling (Sup.)2026.04	0.063	0.161
Unsupervised Confidence Calibration2026.04	0.089	0.161
Unsupervised Confidence Calibration2026.04	0.09	0.16
Test-time self-consistency (TT SC)2026.04	0.103	0.145
Token Probs. (TP)2026.04	0.217	0.223
Verbalized Confidence (VC)2026.04	0.304	0.296
Answer Probs. (Ans. TP)2026.04	0.333	0.316