生成式AI在引导性干扰下的答案一致性研究  

Research on answer consistency of generative AI under guided interference

作　　者：王萍[1] 游强华 WANG Ping;YOU Qianghua(College of Basic Medicine and Forensic Medicine,North Sichuan Medical College,Nanchong,Sichuan 637000,China;College of Management,North Sichuan Medical College,Nanchong,Sichuan 637007,China)

机构地区：[1]川北医学院基础医学与法医学院,四川南充637000 [2]川北医学院管理学院,四川南充637007

出　　处：《计算机应用文摘》2025年第4期173-177,共5页

基　　金：南充市社会科学研究“十四五”规划2024年度项目:人工智能时代ChATGPT大型语言模型中风险信息生成的输入模式研究(NC24B219)。

摘　　要：生成式AI在自然语言处理领域展示了强大的文本生成能力,但其在引导性干扰下的答案一致性和稳定性问题仍未得到充分研究。文章以7个主流生成式AI模型为对象,围绕数学、语文、计算机、英语和地理5个学科,设计了涵盖易、中、难3种难度等级的测试题目,并通过多轮引导性干扰来评估模型的答案一致性与稳定性。研究结果表明,各模型在基础知识点的首次回答中表现较好,但在处理复杂推理任务和多轮干扰时,一致性和稳定性显著下降。部分模型在特定学科中表现出优势,但不同模型在面对引导性干扰时的表现差异明显,部分模型在推理能力和上下文记忆方面存在不足。Generative AI has demonstrated powerful text generation capabilities in the field of natural language processing,but its answer consistency and stability under guiding interference have not been fully studied.The article focuses on seven mainstream generative AI models and designs test questions covering easy,medium,and difficult levels around five disciplines:mathematics,Chinese,computer science,English,and geography.Multiple rounds of guided interference are used to evaluate the consistency and stability of the models̓answers.The research results indicate that each model performs well in the initial response to basic knowledge points,but its consistency and stability significantly decrease when dealing with complex reasoning tasks and multiple rounds of interference.Some models have shown advantages in specific disciplines,but there are significant differences in the performance of different models when facing guiding interference.Some models have shortcomings in reasoning ability and contextual memory.

关 键 词：生成式AI 引导性干扰 答案一致性 多学科测试 模型评估 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]