Title: Measuring the Accuracy of Confidence Claims in Large Reasoning Models

Abstract

While trustworthy large language models (LLMs) depend heavily on their ability to communicate uncertainty reliably, they frequently fail at "faithful calibration" (FC)—the task of aligning a model’s internal certainty with its verbalized confidence. This issue is especially pressing for Large Reasoning Models (LRMs), as users often interpret their extended reasoning steps as proof of careful deliberation, skill, and assuredness. However, despite the widespread adoption of LRMs and the significance of FC, there is limited understanding of how well these systems actually convey their confidence accurately.

Current methods for assessing FC are ill-suited for LRMs, which generate lengthy chain-of-thought outputs. These outputs typically lack distinct step boundaries, exhibit inconsistent structural patterns, and embed intricate conditional dependencies throughout the reasoning process, all of which make estimating intrinsic confidence difficult. To overcome these obstacles, this study presents a new framework designed to systematically measure the FC of LRMs. This approach evaluates linguistic decisiveness against three distinct metrics of internal uncertainty: token probabilities, hidden states, and the consistency of sampled responses. Additionally, we introduce a prefix-conditioned sampling technique to account for variations in structure and conditionality across different reasoning traces.

Our application of this framework across various leading models, datasets, and prompts reveals that expressing confidence faithfully remains a major hurdle for LRMs. We observed that engaging in reasoning does not inherently enhance FC, and prompt strategies effective for non-reasoning models fail to improve faithfulness within reasoning contexts. Furthermore, the use of different confidence estimators yields conflicting evaluations of identical traces, highlighting vulnerabilities in existing assessment methods. Ultimately, our findings position FC as a unique and critical objective for reliability and alignment in LRMs, a necessity as these systems are increasingly utilized in high-stakes environments.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC