生成式人工智能融入STEM课堂教学的国外实证研究综述

A Review of Empirical Studies in the International Journals on Integrating Generative AI into STEM Instruction

摘要: 随着生成式人工智能的快速发展，大模型在课堂教学中的应用备受关注。本文以近三年国际期刊发表的29篇相关实证研究为分析对象，系统梳理大模型在STEM课堂教学中的应用研究现状。文献统计结果显示，当前该领域研究呈现明显的教育阶段聚焦特征：以高等教育场景为主，相关文献占比达62%；基础教育场景的研究相对薄弱，占比为38%。值得注意的是，截至目前，尚未检索到涉及学前教育或特殊教育领域的相关实证研究成果。从学科分布来看，基础教育的文献主要关注数学和物理，鲜少关注化学、生物、工程等学科。此外，在STEM课堂教学中，常用的3种大模型增强策略包括提示语工程，模型微调和检索增强生成。其中，提示语工程最为常用。最后，大模型在STEM课堂教学中的应用主要涵盖至教师教学(教)、学生学习(学)、学业评价(评)三个维度中。其中，学习评价相关的文献数量较多。总体而言，现有文献表明，大模型能够提升课堂互动性、促进深度学习和增强学生学习动机。未来研究应进一步关注大模型在具体学科中的优化实践，以推动STEM课堂教学向智能化、个性化与高质量方向发展。

Abstract: With the rapid development of generative artificial intelligence, the integration of large language models (LLMs) into STEM education has become an emerging research trend. This review synthesizes 29 empirical studies published in international journals for the past three years. The results of this review paper indicate that LLMs are mainly applied in higher education (62%), with comparatively fewer studies in K-12 contexts (38%), and no empirical studies were found in the context of preschool or special education. At the disciplinary level, K-12 studies tend to focus on mathematics and physics, whereas chemistry, biology, and engineering remain underexplored. Across these reviewed articles, three optimization methods in LLMs are often used: prompt engineering, model fine-tuning, and retrieval-augmented generation (RAG). The most often used method is prompt engineering. The integrations of LLMs into STEM education align with teaching, learning, and assessment in instructional practice. The assessment gains the greatest research attention. Overall, current empirical studies reveal that LLMs can increase instructional effectiveness, promote deeper cognitive engagement, and strengthen students’ motivation. Future research should prioritize discipline-specific instructional design to advance more customized and higher-quality STEM instruction.