Calibrating the Confidence of Large Language Models by Eliciting Fidelity
This work addresses the issue of inaccurate confidence calibration in large language models, which is crucial for reliable deployment in applications like question-answering, but it is incremental as it builds on existing calibration techniques.
The paper tackles the problem of overconfidence in large language models after alignment by decomposing confidence into uncertainty and fidelity, proposing a plug-and-play method for calibration. The method demonstrated good calibration performance in experiments with 6 RLHF-LMs on four MCQA datasets, and introduced two novel metrics (IPR and CE) for evaluation.
Large language models optimized with techniques like RLHF have achieved good alignment in being helpful and harmless. However, post-alignment, these language models often exhibit overconfidence, where the expressed confidence does not accurately calibrate with their correctness rate. In this paper, we decompose the language model confidence into the \textit{Uncertainty} about the question and the \textit{Fidelity} to the answer generated by language models. Then, we propose a plug-and-play method to estimate the confidence of language models. Our method has shown good calibration performance by conducting experiments with 6 RLHF-LMs on four MCQA datasets. Moreover, we propose two novel metrics, IPR and CE, to evaluate the calibration of the model, and we have conducted a detailed discussion on \textit{Truly Well-Calibrated Confidence}. Our method could serve as a strong baseline, and we hope that this work will provide some insights into the model confidence calibration.