In recent years, large language models have demonstrated great potential for personalized question answering, leveraging their vast parametric knowledge and powerful natural language processing capabilities. However, existing approaches still face several limitations, including insufficient exploitation of student- and course-specific information, limited answer relevance, and constrained response accuracy due to the lack of reliable background knowledge. To address these challenges, this paper proposes a personalized question-answering method for LLMs based on query expansion and reinforcement learning. Specifically, original queries are expanded using a pedagogical knowledge graph to incorporate personalized information related to students and courses; a reinforcement learning algorithm based on a student background identification task is designed to guide the model in aligning more closely with student characteristics via a feedback mechanism; and a retrieval-augmented generation framework is employed to dynamically extract supporting materials from a teaching knowledge base, thereby enhancing the accuracy and reliability of responses. Experimental results on the MOOCCubeX dataset show that, compared with the best method among the baseline methods, the proposed method improves the matching degree across four personalization dimensions—knowledge background, learning style, learning progress, and interest orientation—by 0.210, 0.156, 0.222, and 0.168, respectively, and increases answer accuracy by 0.086 1.
答疑是教学过程中的重要环节,能够促进学生对知识的深入理解,并激发其学习积极性和主动性。然而随着高等教育普及化与在线教育规模化发展,答疑资源短缺问题日益凸显,因而迫切需要引入人工智能技术以提升教学答疑的效率。作为人工智能领域的最新成果之一,大语言模型(Large Language Model, LLM)凭借海量的参数化知识与强大的自然语言理解和生成能力,在教学答疑等场景中展现出巨大的应用潜力,有望实现“师师有助教,生生有学伴”这一愿景[1-2]。
LLM通过在大规模数据上进行预训练,能够学习到丰富的自然语言知识与通用领域知识,并且通过指令微调(Instruction Tuning)[5]以及基于人类反馈的强化学习(Reinforcement Learning from Human Feedback,RLHF)[6]等技术拥有面向多任务的通用求解能力。其中,指令微调利用格式化的训练数据加强LLM的通用任务泛化能力,而基于人类反馈的强化学习通过训练与人类偏好对齐的奖励模型,使得LLM能够更好地遵循用户意图。除此之外,还可以使用检索增强生成(Retrieval-Augmented Generation,RAG)[7]和思维链(Chain of Thought, CoT)[8]等技术增强LLM回答效果。
随着LLM在各类通用任务中大放异彩,个性化大语言模型(Personalized Large Language Model, PLLM)也逐渐受到关注,它旨在根据用户的特定数据、历史交互和上下文信息,生成满足用户独特需求和偏好的响应,对于提高人机交互体验具有重要意义[9]。PLLM的实现路径有显式和隐式两种,前者主要依靠模型强大的上下文学习(In-context Learning)能力,在推理时将用户偏好信息直接注入指令,以引导模型生成个性化内容。例如,Wang等[10]提出的Cue-CoT方法使用思维链技术提取用户情绪和个性等信息,Liu等[11]提出的ONCE方法使用提示调优技术引导闭源LLM总结用户浏览历史,以获得用户的兴趣偏好;相比之下,后者通过调整LLM内部参数或训练目标函数,将用户偏好隐式地嵌入模型中,代表性技术包括参数高效微调(Parameter-Efficient Fine-Tuning, PEFT)[12]和个性化RLHF(P-RLHF)[13]等。
DAIL, ZHAOX W, ZHUZ T. A new inquiry learning: Conversational learning with ChatGPT[J]. Open Education Research, 2023, 29(6): 42-51, 111. DOI:10.13966/j.cnki.kfjyyj.2023.06.005(Ch ).
WUL A, YANH B, HUANGF L, et al. The application analysis and realistic challenge of large language models in higher education[J]. Modern Educational Technology, 2023, 33(8): 29-37. DOI:10.3969/j.issn.1009-8097.2023.08.004(Ch ).
[5]
HUANGL, YUW J, MAW T, et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions[J]. ACM Transactions on Information Systems, 2025, 43(2): 1-55. DOI:10.1145/3703155 .
LIUZ Y, WANGP J, SONGX B, et al. Survey on hallucinations in large language models[J]. Journal of Software, 2025, 36(3): 1152-1185. DOI: 10.13328/j.cnki.jos.007242(Ch ).
[8]
WEIJ, BOSMAM, ZHAOV Y, et al. Finetuned language models are zero-shot learners[EB/OL]. 2021: 2109.01652.
[9]
OUYANGL, WUJ, JIANGX, et al. Training language models to follow instructions with human feedback[J]. Advances in Neural Information Processing Systems, 2022, 35: 27730-27744. DOI: 10.5555/3600270.3602281 .
[10]
LEWISP, PEREZE, PIKTUSA, et al. Retrieval-augmented generation for knowledge-intensive NLP tasks[J]. Advances in Neural Information Processing Systems, 2020, 33: 9459-9474. DOI: 10.5555/3495724.3496517 .
[11]
KOAK J L, MAY S, NG R, et al. Learning to generate explainable stock predictions using self-reflective large language models[C]//Proceedings of the ACM Web Conference 2024. Singapore:ACM, 2024: 4304-4315. DOI:10.1145/3589334.3645611 .
[12]
CHENJ, LIUZ, HUANGX, et al. When large language models meet personalization: Perspectives of challenges and opportunities[J]. World Wide Web, 2024, 27(4): 42. DOI:10.1007/s11280-024-01276-1 .
[13]
WANGH R, WANGR, MIF, et al. Cue-CoT: Chain-of-thought prompting for responding to In-depth dialogue questions with LLMs[EB/OL].[2025-04-11]. DOI: 10.18653/v1/2023.findings-emnlp.806 .
[14]
LIUQ J, CHENN, SAKAIT, et al. ONCE: Boosting content-based recommendation with both open- and closed-source large language models[C]//Proceedings of the 17th ACM International Conference on Web Search and Data Mining. New York: ACM, 2024: 10.1145/3616855.3635845. DOI:10.1145/3616855.3635845 .
[15]
LIUX, JIK X, FUY C, et al. P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks[C]//Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). Stroudsburg: Association for Computational Linguistics, 2022: 61-68. DOI:10.18653/v1/2022.acl-short.8 .
[16]
LIX Y, ZHOUR Y, LIPTONZ C, et al. Personalized language modeling from personalized human feedback[EB/OL]. 2024: 2402.05133.
[17]
ABDELGHANIR, WANGY H, YUANX D, et al. GPT-3-driven pedagogical agents to train children’s curious question-asking skills[J]. International Journal of Artificial Intelligence in Education, 2024, 34(2): 483-518. DOI:10.1007/s40593-023-00340-7 .
[18]
TSAIM L, ONGC W, CHENC L. Exploring the use of large language models(LLMs) in chemical engineering education: Building core course problem models with Chat-GPT[J]. Education for Chemical Engineers, 2023, 44: 71-95. DOI:10.1016/j.ece.2023.05.001 .
[19]
KASNECIE, SESSLERK, KÜCHEMANNS, et al. ChatGPT for good? On opportunities and challenges of large language models for education[J]. Learning and Individual Differences, 2023, 103: 102274. DOI:10.1016/j.lindif.2023.102274 .
LIUJ, SUNX, ZHANGY Q. Educational resource content review method based on knowledge graph and large language model collaboration[J]. Journal of East China Normal University (Natural Science), 2024(5): 57-69. DOI: 10.3969/j.issn.1000-5641.2024.05.006(Ch ).
[22]
TABARSIB. Developing LLM-powered trustworthy agents for personalized learning support[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2025, 39(28): 29301-29302. DOI:10.1609/aaai.v39i28.35228 .
[23]
GANW S, QIZ L, WUJ Y, et al. Large language models in education: Vision and opportunities[C]//2023 IEEE International Conference on Big Data (BigData). New York: IEEE Press, 2023: 4776-4785. DOI:10.1109/BigData59044.2023.10386291 .
[24]
PARKM, KIMS, LEES, et al. Empowering personalized learning through a conversation-based tutoring system with student modeling[EB/OL]. [2024-12-10]. DOI: 10.1145/3613905.3651122 .
[25]
DANY H, LEIZ K, GUY Y, et al. EduChat: A large-scale language model-based chatbot system for intelligent education[EB/OL]. 2023: 2308.02773. DOI: 10.1007/978-981-96-1809-5_22 .
ZHANGX F, ZHANGL P, YANS, et al. Personalized learning recommendation in collaboration of knowledge graph and large language model[J]. Journal of Computer Applications, 2025,45(3):773-784. DOI:10.11772/j.issn.1001-9081.2024070971(Ch ).
[28]
POSCENTEK. Book review-The three dimensions of learning: Contemporary learning theory in the tension field between the cognitive, the emotional and the social [J]. The International Review of Research in Open and Distributed Learning, 2006, 7(1): 1-3. DOI:10.19173/irrodl.v7i1.305 .
[29]
WALKINGTONC, BERNACKIM L. Motivating students by “personalizing” learning around individual interests: A consideration of theory, design, and implementation issues[EB/OL].[2025-05-04]. DOI:10.1108/S0749-742320140000018004 .
HONEYP, MUMFORDA. The Manual of Learning Styles[M]. 3rd edition. Maidenhead: Peter Honey Publications, 1992.
[32]
ZHENGL M, CHIANGW L, SHENGY, et al. Judging LLM-as-a-judge with MT-Bench and Chatbot Arena[EB/OL]. [2024-12-21].
[33]
KAHNGM, TENNEYI, PUSHKARNAM, et al. LLM comparator: Interactive analysis of side-by-side evaluation of large language models[J]. IEEE Transactions on Visualization and Computer Graphics, 2025, 31(1): 503-513. DOI:10.1109/TVCG.2024.3456354 .