华东师范大学学报(教育科学版)

华东师范大学学报(教育科学版) ›› 2025, Vol. 43 ›› Issue (5): 44-56.doi: 10.16382/j.cnki.1000-5560.2025.05.004

• 教育数字化转型：学习与多智能体（特约主持人：顾小清） • 上一篇    下一篇

从单智能体到多智能体：大模型智能体支持下的激励型学习活动设计与实证研究

黄昌勤1,2,3, 钟益华1, 王希哲3, 韩中美3, 魏同权1   

  1. 1. 华东师范大学上海智能教育研究院，上海 200062
    2. 浙江大学教育学院，杭州 310058
    3. 浙江全省智能教育技术与应用重点实验室，金华 321004
  • 出版日期:2025-05-01 发布日期:2025-04-21
  • 基金资助:
    教育部哲学社会科学研究重大课题攻关项目“人工智能和教育深度融合研究”(24JZD011)。

From Single-agent to Multi-agent: Motivational Learning Activities Design and Empirical Study Supported by LLM-based Agents

Changqin Huang1,2,3, Yihua Zhong1, Xizhe Wang3, Zhongmei Han3, Tongquan Wei1   

  1. 1. Shanghai Institute of Artificial Intelligence for Education, East China Normal University, Shanghai 200062, China
    2. College of Education, Zhejiang University, Hangzhou 310058, China
    3. Zhejiang Key Laboratory of Intelligent Education Technology and Application, Zhejiang Normal University, Jinhua 321004, China
  • Online:2025-05-01 Published:2025-04-21

RichHTML

18

PDF

1244

摘要：

大模型与智能代理技术的不断进步，使得大模型智能体成为教育领域中实现教与学提质增效的重要新工具。基于大模型智能体的功能定位差异，单一智能体虽然已能够针对各类教与学任务提供诸如内容生成、智能反馈与评估等支持，但单智能体的交互特点、功能属性具有较高同质性，在促进深层次认知发展方面存在一定局限。相比之下，多智能体能够通过模拟多种教育主体角色，提升学习互动的多样性和深度，进而实现更为个性化和深度的学习体验。鉴于智能体在学习过程中的应用主要依靠学习者自发性，为了保障学习活动的有效开展，本研究基于ARCS动机模型分别设计基于单智能体与多智能体的激励型学习活动方案，并面向英语阅读场景开展了准实验研究。实验结果发现：基于多智能体的激励型学习活动相较单智能体能够显著提升学生在推理、评价与应用方面的学习成绩，具有更强的学习动机，且有效促进了其深层次认知发展，尤其是抽象与概括能力。研究证明了多智能体在支持学生深度学习中的价值，为未来进一步探讨多智能体在教育中的应用提供了借鉴。

关键词: 大语言模型, 多智能体, ARCS动机模型, 激励型学习活动, 深度学习

Abstract:

The continuous advancements in large language models (LLMs) and agent technologies have made LLM-based agents a promising new tool for enhancing the quality and efficiency of teaching and learning in the educational domain. Based on the functional positioning of LLM-based agents, single-agent systems are capable of supporting various teaching and learning tasks, such as content generation, intelligent feedback, and assessment. However, due to the homogeneous nature of their interaction characteristics and functional attributes, single-agent systems face limitations in promoting deeper cognitive development. In contrast, multi-agent systems can simulate various educational roles, enhancing the diversity and depth of learning interactions and thereby providing more personalized and profound learning experiences. Considering that the application of agents in the learning process primarily relies on learners’ initiative, this study based on the ARCS model designed motivational learning activities for both single-agent and multi-agent systems to ensure the effective implementation of learning activities and conducted a quasi-experimental study in the context of English reading. The experimental results reveal that motivational learning activities supported by multi-agent systems significantly outperform single-agent systems in improving students’ performance in reasoning, evaluation, and application tasks. Furthermore, multi-agent systems exhibit stronger motivational effects and effectively promote deeper cognitive development, particularly in abstraction and generalization skills. The study underscores the value of multi-agent systems in supporting deeper learning and provides insights for further exploration of their applications in education.

Key words: large language model, multi-agent, ARCS model, motivational learning activities, deeper learning