In the deployment of large language models (LLMs), fluent yet incorrect answers can incur significant costs; thus, accuracy alone is insufficient. Models must also discern when they are likely to be wrong. We introduce ConfidenceBench, a calibration benchmark that evaluates verbalized confidence estimates in 15 frontier LLMs using the Brier score, a proper scoring rule that incentivizes truthful probability reporting. Confidence is elicited via prompting, requiring no access to model logits, making the framework applicable to both closed-source and open-source systems.
The benchmark comprises 200 private multiple-choice questions across four categories: spatial reasoning, high-precision mathematics, word lookup, and unknowable questions. Across three independent runs, Claude Opus 4.6 and Gemini 3.1 Pro Preview achieved the lowest Brier scores, both reported as 0.103. Both significantly outperform the calibrated-random baseline of 0.1875, while Gemini 3.1 Flash-Lite scored 0.367, indicating severe miscalibration. Accuracy and calibration diverge substantially across model families: the most accurate model is not necessarily the best-calibrated, and several models perform worse than the calibrated-random Brier baseline despite reasonable accuracy. These results highlight that verbalized confidence calibration is a distinct and practically important axis of LLM reliability, complementing standard accuracy-based evaluation.
Blogger's Review: The increasing complexity of LLM applications makes the introduction of ConfidenceBench a timely development, offering a fresh perspective on the assessment of model confidence calibration. By emphasizing the accuracy of confidence, researchers can better understand the limitations of model performance, leading to more informed decisions in practical applications.