The paper discusses the importance of teaching large language models (LLMs) to understand and express uncertainty, particularly in high-stakes applications. It highlights the need for calibrated uncertainties to ensure reliable decision-making. The authors argue that prompting alone is insufficient for achieving good calibration and propose a method involving fine-tuning on a small dataset of correct and incorrect answers to create more accurate uncertainty estimates. They demonstrate that a thousand graded examples are sufficient to outperform baseline methods and that training through the features of a model is necessary for good performance. The study also explores the generalization of uncertainty estimates across distribution shifts and the potential of using LLMs to assist human decision-making. A user study supports the effectiveness of calibrated uncertainty estimates in collaborative settings. The findings suggest that fine-tuning can significantly improve the reliability of LLMs' uncertainty estimates, making them more useful in practical applications.The paper discusses the importance of teaching large language models (LLMs) to understand and express uncertainty, particularly in high-stakes applications. It highlights the need for calibrated uncertainties to ensure reliable decision-making. The authors argue that prompting alone is insufficient for achieving good calibration and propose a method involving fine-tuning on a small dataset of correct and incorrect answers to create more accurate uncertainty estimates. They demonstrate that a thousand graded examples are sufficient to outperform baseline methods and that training through the features of a model is necessary for good performance. The study also explores the generalization of uncertainty estimates across distribution shifts and the potential of using LLMs to assist human decision-making. A user study supports the effectiveness of calibrated uncertainty estimates in collaborative settings. The findings suggest that fine-tuning can significantly improve the reliability of LLMs' uncertainty estimates, making them more useful in practical applications.