教学场景中大模型偏见的存在性和顽固性特点研究

A Study on the Existence and Stubbornness Characteristics of Large Language Model Bias in Educational Scenarios

摘要: 大模型偏见表现为模型输出存在刻板印象或对特定实体代表性异常，是一类难以完全避免的大模型安全问题。当前，大模型在教育领域的应用日益广泛，其偏见可能引发的问题后果严重，但相关研究却十分匮乏。鉴于此，文章参考现有偏见研究成果，聚焦教学场景中的性别、从众和学科三类偏见展开系统性研究。首先，采用“大模型合成 + 人工检测”的方式，构建了三个数据集，分别用于探测职业判断时性别偏见、多学科问题回答时从众偏见以及模拟招生场景下学科偏见。其次，基于上述数据集，选取我国教育领域常用的智谱清言、通义千问和DeepSeek大模型，对其表现出的性别、从众和学科偏见进行量化评估，结果显示所有模型均存在不同程度的偏见。最后，通过探究反思和偏见教育等提示工程手段对大模型偏见的纠正效果来分析偏见的顽固性。实验发现，通过思维链实现的反思对部分偏见有一定改善作用，但存在过度纠正或受模型特性限制的问题。同时，通过无偏见提示语干预实施的偏见教育能在一定程度上缓解偏见，不过由于模型异质性，效果差异显著。综上所述，本研究认为大模型在教学场景中的偏见具有多样性和个性化特征。为防范大模型偏见影响学生认知，造成不良后果，教育工作者应强化偏见风险意识，结合具体场景设计针对性的偏见纠正策略，并密切关注学生使用大模型的情况。

Abstract: Large language model (LLM) bias manifests as outputs containing stereotypes or abnormal representations of specific entities, constituting a type of LLM security issue that is difficult to completely avoid. Currently, the application of LLMs in the educational field is increasingly widespread, and the potential problems caused by their bias can have serious consequences; however, related research remains scarce. In view of this, drawing on existing bias research, this paper focuses on a systematic study of three types of bias in educational scenarios: gender bias, conformity bias, and discipline bias. Firstly, this paper adopts a “LLM synthesis + manual detection” approach to construct three datasets for detecting gender bias in career judgments, conformity bias in multi-disciplinary question answering, and discipline bias in simulated admission scenarios, respectively. Secondly, based on the aforementioned datasets, this paper selects LLMs commonly used in China’s educational field—Zhipu Qingyan, Tongyi Qianwen, and DeepSeek—to quantitatively evaluate their exhibited gender, conformity, and discipline biases. The results show that all models exhibit biases to varying degrees. Finally, this paper analyzes the stubbornness of LLM bias by exploring the corrective effects of prompt engineering techniques, such as triggering reflection and bias education, on this bias. Experiments found that reflection, achieved through chain-of-thought, can partially mitigate certain biases, but issues like over-correction or limitations due to model characteristics exist. Simultaneously, bias education implemented through intervention with unbiased prompts can alleviate bias to some extent; however, due to model heterogeneity, the effectiveness varies significantly. In conclusion, this paper argues that LLM bias in educational scenarios is characterized by diversity and individuality. To prevent LLM bias from affecting students’ cognition and causing adverse consequences, educators should strengthen their awareness of bias risks, design targeted bias correction strategies based on specific scenarios, and closely monitor students’ use of LLMs.