---

Reasoning Models Better Express Their Confidence

Dongkeun Yoon, Seungone Kim, Sohee Yang, Sunkyoung Kim, Soyeon Kim, Yongil Kim, Eunbi Choi, Yireun Kim, Minjoon Seo

Abstract: We demonstrate that reasoning models with extended chain-of-thought (CoT) achieve superior confidence calibration compared to non-reasoning models. Across six datasets, reasoning models showed better calibration in 33 of 36 settings. We found that slow thinking behaviors – such as exploring alternatives and backtracking – enable these models to dynamically adjust confidence throughout their reasoning process. Notably, calibration improves as the CoT unfolds, and removing slow thinking behaviors significantly reduces calibration quality. We also reveal that non-reasoning models can enhance calibration through in-context learning guidance to slow think, isolating slow thinking as the key source of calibration improvements.

[Paper]