Your Pre-trained LLM is Secretly an Unsupervised Confidence Calibrator

In this work, we propose a noval unsupervised calibration method to mitigate the over-confidence problem of LLMs introduced by post-training techniques. With the observation of the inherent well-calibrated nature of Pre-trained LMs (PLMs), we propose to leverage the output of PLMs on unlabeled data for post-hoc calibration.