Confidence Introspection: A Self-Reflection Method for Reliable and Helpful Large Language Models

Large Language Models (LLMs) suffer from factual hallucinations, meaning the LLMs confidently provide responses that are inconsistent with reality. Previous studies explored fine-tuning-based verbalized confidence calibration to mitigate hallucinations, yet these approaches often resulted in overly conservative models, compromising their ability to provide relevant knowledge. Inspired by human introspection processes, we propose Confidence Introspection Tuning, a novel confidence calibration framework that enables LLMs to accurately express their confidence while maintaining helpfulness. This is a two-stage framework: first, it estimates the confidence through question paraphrasing and sampling. Subsequently, through self-generated training data, the model develops the ability to classify questions asknown,uncertain, orunknownwhile providing appropriate responses or relevant knowledge for the questions. Extensive comparisons and analysis on in-distribution and out-of-distribution datasets, four model series, and diverse metrics demonstrate that our method achieves superior performance over previous state-of-the-arts by enhancing reliability while preserving the model's helpfulness.