一句话：系统综述 LLM 不确定性量化，分类偶然与认知不确定性，并讨论校准、案例与可靠性挑战。
本文针对大语言模型（LLM）在可靠部署中面临的固有不确定性问题，对 LLM 不确定性量化（UQ）进行了系统性综述。作者分析 ASE、NeurIPS、ICML、Nature 等顶会/期刊的 45 余篇论文，梳理 UQ 的演进脉络，将不确定性划分为偶然不确定性与认知不确定性，并总结概率建模、置信度估计与校准等技术。通过医学诊断和代码生成等高风险评估案例，论文展示 UQ 对提升可靠性的关键作用，并讨论现有局限、伦理考量与未来方向，强调细粒度可解释性及人机协作对安全、可信和负责任地集成 LLM 的重要性。
Large Language Models (LLMs) have revolutionized AI, yet their inherent uncertainties pose significant challenges to reliable deployment. This paper presents a comprehensive systematic review of uncertainty in LLMs, bridging theoretical foundations and cutting-edge methodologies. We analyze over 45 papers from top venues—including ASE, NeurIPS, ICML, and Nature—to trace the evolution of uncertainty quantification (UQ). We categorize uncertainty into aleatoric and epistemic types, detailing probabilistic modeling, confidence estimation, and calibration techniques. Through illustrative case studies in high-stakes domains such as medical diagnosis and code generation, we demonstrate UQ’s pivotal role in enhancing reliability. We further discuss limitations, ethical considerations, and future directions, emphasizing the need for granular interpretability and human-AI collaboration. This work advances the understanding of LLM uncertainty to enable safer, trustworthy, and responsible real-world integration.
DOI 原文 ·
@inproceedings{paperbot9431,
title = {Understanding Uncertainty In LLMs},
author = {Chandan Kumar Sah},
booktitle = {ASE},
year = {2025},
doi = {10.1109/ase63991.2025.00392}
}