ConfidenceBench Evaluates LLM Verbalized Confidence Calibration.

Matthew ffrench-Constant, Daniel Yang, Xinmeng Huang, Sanyam Kapoor· July 24, 2026 View original

Summary

ConfidenceBench is a new benchmark that evaluates the verbalized confidence estimates of 15 frontier LLMs using the Brier score, revealing significant divergence between accuracy and calibration. It shows that even accurate models can be poorly calibrated, highlighting the practical importance of knowing when models are likely to be wrong.

A new benchmark, ConfidenceBench, has been introduced to evaluate the confidence calibration of Large Language Models (LLMs), a crucial factor in deployment scenarios where incorrect but fluent answers can be costly. The benchmark assesses verbalized confidence estimates from 15 leading LLMs using the Brier score, a metric designed to incentivize truthful probability reporting. This method relies on prompting for confidence, making it applicable to both open-source and closed-source models without needing access to internal logits. The dataset consists of 200 private multiple-choice questions across four categories: spatial reasoning, high-precision mathematics, word lookup, and unknowable questions. The findings reveal substantial differences in calibration across models. Claude Opus 4.6 and Gemini 3.1 Pro Preview achieved the lowest Brier scores (0.103), significantly outperforming a calibrated-random baseline (0.1875). In contrast, Gemini 3.1 Flash-Lite scored 0.367, indicating severe miscalibration. The study highlights a significant divergence between accuracy and calibration, showing that the most accurate model is not necessarily the best-calibrated, and some models perform worse than random in terms of calibration despite reasonable accuracy. This research underscores that verbalized confidence calibration is a distinct and practically important aspect of LLM reliability, complementing traditional accuracy metrics.

Why it matters

For professionals deploying LLMs in critical applications, understanding confidence calibration is as important as accuracy to mitigate risks associated with overconfident or underconfident incorrect responses, ensuring more reliable and trustworthy AI systems.

How to implement this in your domain

  1. 1Integrate confidence calibration metrics like the Brier score into your LLM evaluation pipelines, alongside traditional accuracy metrics.
  2. 2Prioritize LLMs that demonstrate strong verbalized confidence calibration for high-stakes applications.
  3. 3Develop prompting strategies that explicitly encourage LLMs to express their confidence levels for better decision-making.
  4. 4Implement human-in-the-loop processes to review LLM outputs, especially when confidence is low or miscalibrated, to prevent costly errors.

Who benefits

Financial ServicesHealthcareLegalAI/ML DevelopmentCustomer Service

Key takeaways

  • LLM confidence calibration is a distinct and critical aspect of reliability, separate from accuracy.
  • ConfidenceBench provides a method to evaluate verbalized confidence without logit access.
  • Many accurate LLMs can be poorly calibrated, leading to overconfidence in incorrect answers.
  • Prioritizing calibrated models and implementing confidence-aware workflows is essential for high-stakes deployments.

Original post by Matthew ffrench-Constant, Daniel Yang, Xinmeng Huang, Sanyam Kapoor

"arXiv:2607.20526v1 Announce Type: new Abstract: Large language models (LLMs) are increasingly deployed in settings where fluent but incorrect answers can be costly. In these settings, accuracy alone is insufficient: models must also know when they are likely to be wrong. We prese…"

View on X

Originally posted by Matthew ffrench-Constant, Daniel Yang, Xinmeng Huang, Sanyam Kapoor on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses