CritiCal: Can Natural Language Critiques Help LLM's Uncertainty or Confidence Calibration?

ACL ARR 2026 January Submission8707 Authors

Keywords: confidence calibration, uncertainty quantification, natural language critiques

Abstract:

Confidence calibration is critical for safe use of Large Language Models (LLMs), where clear verbalized confidence enhances user trust.
Traditional methods that mimic reference confidence expressions often fail to utilize the logic in model's original reasoning chain.
We propose natural language critique as a solution, which is ideal for confidence calibration, as precise gold confidence labels are hard to obtain and often require multiple generations but assessing whether the confidence is appropriate is easy by analyzing its internal logic and answer correctness.
This paper studies how natural language critiques can enhance verbalized confidence: (1) **What to critique**: uncertainty (question-focused) or confidence (answer-specific)?
Analysis shows confidence suits multiple-choice tasks, while uncertainty excels in open-ended scenarios.
(2) **How to critique**: self-critique or critique calibration training?
We propose **Self-Critique**, enabling LLMs to critique and optimize their confidence, and **CritiCal**, using natural language **Criti**ques to train confidence **Cal**ibration.
Experiments show that CritiCal significantly outperforms Self-Critique and other competitive baselines, \textbf{even surpassing its teacher, GPT-4o}, in complex reasoning tasks.
CritiCal also shows robust generalization in out-of-distribution settings, proving its reliability.

Paper Type: Long

Research Area: Interpretability and Analysis of Models for NLP

Research Area Keywords: calibration/uncertainty, free-text/natural language explanations, robustness

Contribution Types: Model analysis & interpretability, Data analysis

Languages Studied: English

Submission Number: 8707