# From Entropy to Calibrated Uncertainty: Training Language Models to Reason About Uncertainty

Source: [From Entropy to Calibrated Uncertainty: Training Language Models to Reason About Uncertainty](<https://www.sota2.com/research/paper/69ae53e82d603e6de3e42997>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-03-10T05:00:05.425Z

Authors: Azza Jenane, Nassim Walha, Lukas Kuhn, Florian Buettner

Published: 2026-03-06T14:21:42.000Z

## Abstract

Large Language Models (LLMs) that can express interpretable and calibrated uncertainty are crucial in high-stakes domains. While methods to compute uncertainty post-hoc exist, they are often sampling-based and therefore computationally expensive or lack calibration. We propose a three-stage pipeline to post-train LLMs to efficiently infer calibrated uncertainty estimates for their responses. First, we compute fine-grained entropy-based uncertainty scores on the training data, capturing the distributional variability of model outputs in embedding space. Second, these scores are calibrated via Platt scaling, producing reliable and human-interpretable uncertainty signals. Finally, the target LLM is post-trained via reinforcement learning to align its policy with these calibrated signals through a verifiable reward function. Unlike post-hoc uncertainty estimation methods, our approach provides interpretable and computationally efficient uncertainty estimates at test time. Experiments show that models trained with our pipeline achieve better calibration than baselines and generalize to unseen tasks without further processing, suggesting that they learn a robust uncertainty reasoning behavior.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Uncertainty Calibration on TriviaQA + Natural Questions ID held-out (test)](<https://www.sota2.com/research/sota/uncertainty-calibration-on-triviaqa-plus-natural-questions-id-held-out-test.md>): Uncertainty Calibration; TriviaQA + Natural Questions ID held-out (test); ECE: 7.2.

- [Uncertainty Calibration on GSM8k OOD (test)](<https://www.sota2.com/research/sota/uncertainty-calibration-on-gsm8k-ood-test.md>): Uncertainty Calibration; GSM8K OOD (test); ECE: 3.15.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2603.06317>)

[PDF](<https://arxiv.org/pdf/2603.06317.pdf>)
