What question did this study set out to answer?

This research aims to enhance the alignment between verbalized and internal confidence in large language models.

December 19, 2025Open Access

Direct Confidence Alignment: Aligning Verbalized Confidence with Internal Confidence In Large Language Models

View Full Paper

Ask AI

Bookmark

Authors

Proposed Direct Confidence Alignment improves reliability in model responses while highlighting calibration challenges.

This research aims to enhance the alignment between verbalized and internal confidence in large language models.
Developed Direct Confidence Alignment using Direct Preference Optimization.
Evaluated DCA across multiple open-weight LLMs on various datasets.
Introduced new calibration error-based metrics for alignment assessment.
DCA improved alignment metrics in some model architectures.
Reduced inconsistencies in confidence expression within models.
Highlighted ineffectiveness of DCA in certain architectures.

Zhang et al. (2025) studied this question.

View Full Paper

Ask AI

Bookmark