生成式人工智能赋能职业院校程序设计课程教学改革

Teaching Reform of Programming Courses in Vocational Colleges Enabled by Generative Artificial Intelligence

摘要: 生成式人工智能特别是大语言模型的发展，为职业院校程序设计课程的个性化指导、代码调试和过程性评价提供了新的技术条件，但也带来答案依赖与学生主体性弱化等风险。针对个性化反馈不足、自动评测重结果轻解释、评价难以识别真实学习过程等问题，本文首先梳理从智能辅导系统、自动评测到生成式人工智能的程序设计反馈研究，指出大语言模型使反馈生成方式由预定义规则与模板转向按代码情境即时生成、反馈形态由结构化信息转向自然语言解释，带来反馈能力的质变，但反馈强度控制与答案依赖防范仍有赖教学设计；继而利用ProgFeed真实课堂开放数据开展探索性分析，发现反馈形式对学生是否继续提交和是否修改代码的影响有限，自然语言解释性反馈与更大的得分改善和更高的问题解决概率相关，多轮提交后各组最终结果趋于接近；在此基础上，结合职业院校学情特点，提出“教师规则设计–大模型受限反馈–学生自主修订–自动测试验证–教师诊断指导–学习反思”的人机协同教学模式构想，设计五级分层反馈机制，主张将生成式人工智能由“答案生成工具”转化为“形成性学习支架”，推动程序设计课程由结果导向评价转向过程证据导向评价。

Abstract: Generative artificial intelligence, especially large language models, provides new technical conditions for personalized guidance, debugging support, and process-oriented assessment in programming courses of vocational colleges, while introducing risks such as answer dependence and weakened learner agency. To address insufficient individualized feedback, result-oriented automated assessment, and the difficulty of identifying real learning processes, this study first reviews feedback research in programming education, tracing its evolution from intelligent tutoring systems and automated assessment to generative AI, and argues that large language models bring a qualitative change in feedback capability: feedback generation shifts from predefined rules and templates to context-aware generation, and feedback forms shift from structured messages to natural-language explanation, whereas controlling feedback intensity and preventing answer dependence still depend on instructional design. An exploratory analysis of the open ProgFeed classroom dataset then shows that feedback forms have limited influence on whether students continue submitting or modify code; natural-language explanatory feedback is associated with greater score improvement and higher problem-resolution probability; and final outcomes converge after multiple submissions. Based on these findings and the characteristics of vocational learners, the study proposes a teacher-led human-AI collaborative teaching model consisting of teacher rule design, constrained LLM feedback, student revision, automated testing, teacher diagnosis, and learning reflection, together with a five-level tiered feedback mechanism. It is argued that generative AI should be positioned as a formative learning scaffold rather than an answer-generation tool, and that assessment should shift from final results to process evidence.