Abstract
Personalized learning represents a promising educational strategy withinintelligent educational systems, aiming to enhance learners' practiceefficiency. However, the discrepancy between offline metrics and onlineperformance significantly impedes their progress. To address this challenge, weintroduce Agent4Edu, a novel personalized learning simulator leveraging recentadvancements in human intelligence through large language models (LLMs).Agent4Edu features LLM-powered generative agents equipped with learner profile,memory, and action modules tailored to personalized learning algorithms. Thelearner profiles are initialized using real-world response data, capturingpractice styles and cognitive factors. Inspired by human psychology theory, thememory module records practice facts and high-level summaries, integratingreflection mechanisms. The action module supports various behaviors, includingexercise understanding, analysis, and response generation. Each agent caninteract with personalized learning algorithms, such as computerized adaptivetesting, enabling a multifaceted evaluation and enhancement of customizedservices. Through a comprehensive assessment, we explore the strengths andweaknesses of Agent4Edu, emphasizing the consistency and discrepancies inresponses between agents and human learners. The code, data, and appendix arepublicly available at https://github.com/bigdata-ustc/Agent4Edu.