一句话:探索将大语言模型嵌入自动编程评测系统,提供上下文感知、可信且契合学习目标的反馈。
自动编程评测系统(APAS)能够在大规模课程中扩展练习与反馈,但常给出泛化提示、静态测试,且缺乏透明度。本博士论文研究将大语言模型嵌入 APAS,以提供上下文感知、可信赖、并与学习目标(而非仅正确性)对齐的反馈。具体计划包括:扩展 Iris 聊天导师,引入检索增强生成(RAG);构建面向教育的 LLM arena,用于评估辅导质量与练习生成。作者希望就设计、评估方法与研究范围获得反馈。该工作尚处于构想阶段,重点在于探索 LLM 在编程教育反馈中的可信性与教学对齐问题。
Automated programming assessment systems (APASs) scale exercises and feedback in large courses, but often offer generic hints, static tests, and little transparency. My dissertation explores embedding LLMs into APASs to deliver context-aware, trustworthy feedback aligned with learning goals, not just correctness. I will extend the Iris chat tutor with retrieval-augmented generation and build an education-focused LLM arena to evaluate tutoring and exercise generation. I seek feedback on design, evaluation, and scope.