accuracy-controlled

This paper demonstrates that global calibration metrics like Expected Calibration Error are confounded by model accuracy, and proposes ACE, an accuracy-controlled evaluation framework for fair comparison of large language models.