title: "Rethinking Confidence in Reasoning Models | Cohere"

A method named CALIBER is proposed to better evaluate the confidence of AI reasoning models. It works by reporting both an initial and a final confidence level for the model's outputs. This approach has been shown to significantly reduce the Expected Calibration Error on benchmarks like BigMathDigits. Additionally, the CALIBER method improves the AUROC, indicating a better ability to distinguish between correct and incorrect reasoning.