Large Language Models Must Be Taught to Know What They Don't Know

TL;DR

This paper demonstrates that fine-tuning large language models with a small, graded dataset significantly improves their ability to estimate uncertainty, which is crucial for high-stakes applications and human-AI collaboration.

Contribution

It introduces a fine-tuning approach using a small dataset to enhance LLM uncertainty calibration, outperforming baseline methods with minimal computational overhead.

Findings

Fine-tuning with 1,000 graded examples improves uncertainty calibration.

Model features are essential for effective uncertainty estimation.

Many models can serve as general-purpose uncertainty estimators.

Abstract

When using large language models (LLMs) in high-stakes applications, we need to know when we can trust their predictions. Some works argue that prompting high-performance LLMs is sufficient to produce calibrated uncertainties, while others introduce sampling methods that can be prohibitively expensive. In this work, we first argue that prompting on its own is insufficient to achieve good calibration and then show that fine-tuning on a small dataset of correct and incorrect answers can create an uncertainty estimate with good generalization and small computational overhead. We show 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 and tractable for large open-source models when using LoRA. We also investigate the mechanisms that enable reliable LLM uncertainty estimation, finding that many models can be used as general-purpose uncertainty estimators, applicable not just to their own uncertainties but also the uncertainty of other models. Lastly, we show that uncertainty estimates inform human use of LLMs in human-AI collaborative settings through a user study.