Abstract: Given the safety and ethical concerns arising from bias in Large Language Model (LLM), this review provides a comprehensive analysis of current research, techniques, and limitations related to bias in LLMs. The analysis is structured around three key aspects: bias identification, evaluation, and mitigation of bias. Firstly, three key techniques of LLM were examined to uncover the fundamental causes of intrinsic bias. Secondly, the types of bias present in LLM were categorized into linguistic bias, demographic bias, and evaluation bias, with characteristics and underlying causes explored. Thirdly, a systematic review of existing bias evaluation benchmarks was provided, discussing the strengths and weaknesses of the general-purpose, language-specific, and task-specific benchmarks. Finally, current bias mitigation techniques were analyzed, focusing on both model debiasing and data debiasing approaches while highlighting directions for future refinement. The findings highlights the research directions for bias in LLMs: evaluating bias from a multicultural perspective, developing lightweight bias mitigation techniques, and enhancing the interpretability of bias.

Key words:  large language model, bias tracing, bias identification, bias evaluation, bias mitigation

摘要： 针对大语言模型（LLM）输出内容存在偏见而导致其不安全和不可控的问题，从偏见识别、偏见评估和偏见去除3个角度，深入梳理和分析现有大语言模型偏见的研究现状、技术与局限。首先，概述大语言模型的三大关键技术，从中分析其不可避免存在内隐偏见（Intrinsic Bias）的根本原因；其次，总结现有大语言模型存在的偏见类型，分为语言偏见、人口偏见和评估偏见，分析这些偏见的特点和原因；再次，系统回顾现有大语言模型偏见的评估基准，探讨这些通用型评估基准、特定语言评估基准、和特定任务评估基准的优点及局限；最后，从模型去偏和数据去偏两个角度对现有大语言模型偏见的去除技术进行深入分析，指出其改进方向。分析指出大语言模型偏见研究的三个方向：偏见的多文化属性评估、轻量级的偏见去除技术以及增强偏见的可解释性。

关键词: 大语言模型, 偏见溯源, 偏见识别, 偏见评估, 偏见去除