当机器拟其形而失其律——大语言模型辅助近体诗学习中的知识建构与批判性思维

When the Machine Mimics the Form but Misses the Meter—Knowledge Construction and Critical Thinking in LLM-Assisted Learning of Chinese Regulated Verse

摘要: 本研究考察大学生与大语言模型(LLM)协同创作近体诗时的知识建构与批判性思维。延续既有七言绝句研究，本文以DeepSeek-V4辅助五言律诗创作，并设置两种提示条件：第一项不限次数，第二项限为七次。研究对两名学生的四份完整人机对话进行解释性逐轮分析。案例显示，模型较能遵守提示中明确列出的平仄与押韵要求，却可能遗漏对仗等未明示规范；自我核验时还可能以“拗救”“宽对”等术语为错误作出似是而非的解释。低投入学生倾向整体接受模型输出；积极互动若缺少稳固的格律知识，也未必能有效纠错。本文以元任务觉察(MTA)为分析框架指出，错误得以留存的关键在于学生对任务评价标准的觉察不足，且思维回路中的反思解释环节被让渡给生成者。本文建议将LLM定位为由规则核验与教师引导约束的诊断性陪练，而非格律判断的最终权威。由于样本仅含两名学生与一名专家，结论限于案例解释。

Abstract: This study examines knowledge construction and critical thinking as undergraduates co-compose Chinese regulated verse (jintishi) with a large language model (LLM). Extending earlier work on seven-character quatrains, the study uses DeepSeek-V4 to support the composition of five-character regulated verse, a form with more extensive structural requirements. Inspired by the allusion of “composing a poem in seven steps”, the first task allowed unlimited prompting, whereas the second limited students to seven prompts. We conducted an interpretive, turn-by-turn analysis of four complete human-AI dialogue records produced by two students. The cases suggest that the model more reliably followed explicitly stated tonal and rhyme constraints but could overlook unstated genre requirements, particularly parallelism. During self-checking, it sometimes offered plausible-sounding yet incorrect explanations using terms such as aojiu (tonal deviation and compensation) and kuandui (loose parallelism). The two students also showed distinct risks: minimal engagement led to wholesale acceptance of model outputs, while active interaction without secure metrical knowledge did not ensure successful error detection. Drawing on Meta-Task Awareness (MTA), we argue that errors persisted because the students lacked awareness of the task’s evaluative criteria and because the reflective-interpretation node of the thinking loop was ceded to the generator. We therefore propose positioning the LLM as a diagnostic practice partner constrained by rule-based verification and teacher guidance, rather than as the final authority on prosodic correctness. Because the study draws on two students and one expert evaluator, its findings are interpretive and preliminary.