AgentTutor: Empowering Personalized Learning with Multi-Turn Interactive Teaching in Intelligent Education Systems
Yuxin Liu, Zeqing Song, Jiong Lou, Chentao Wu, Jie Li · 2026
Read AgentTutor: Empowering Personalized Learning with Multi-Turn Interactive Teaching in Intelligent Education Systems by Yuxin Liu, Zeqing Song, Jiong Lou, Chentao Wu, Jie Li (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-04219.txt.
Introduction
Yuxin Liu, Zeqing Song, Jiong Lou, Chentao Wu, and Jie Li present AgentTutor, a teaching system built for many-turn, personalised lessons. Current smart education tools often answer one question at a time. They cannot gauge a learner's level, shift strategy from live feedback, or grow beyond one-off replies.
AgentTutor joins a multi-agent system powered by a large language model with a personal learning profile for each learner. The profile tracks status, goals, tastes, and study materials in varied forms, and teaching plans update as learning proceeds. Five modules handle splitting the curriculum, assessing the learner, choosing strategy, reflecting on teaching, and remembering knowledge and experience. Across benchmark data sets, the five authors report clear gains in learner performance and strong quality in multi-turn teaching against baselines.
Catalog record
| CREATOR | Yuxin Liu, Zeqing Song, Jiong Lou, Chentao Wu, Jie Li |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 57 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-04219
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
AgentTutor: Empowering Personalized Learning with Multi-Turn Interactive Teaching in Intelligent Education Systems Yuxin Liu, Zeqing Song, Jiong Lou, Chentao Wu, Jie Li Abstract The rapid advancement of large-scale language models (LLMs) has shown their potential to transform intelligent education systems (IESs) through automated teaching and learning support applications. However, current IESs often rely on single-turn static question-answering, which fails to assess learners’ cognitive levels, cannot adjust teaching strategies based on real-time feedback, and is limited to providing simple one-off responses. To address these issues, we introduce AgentTutor, a multi-turn interactive intelligent education system to empower personalized learning. It features an LLM-powered generative multi-agent system and a learner-specific personalized learning profile environment that dynamically optimizes and delivers teaching strategies based on learners’ learning status, personalized goals, learning preferences, and multimodal study materials. It includes five key modules: curriculum decomposition, learner assessment, dynamic strategy, teaching reflection, and knowledge & experience memory. We conducted extensive experiments on multiple benchmark datasets, AgentTutor significantly enhances learners’ performance while demonstrating strong effectiveness in multi-turn interactions and competitiveness in teaching quality among other baselines. Introduction The development of large-scale language models (LLMs) has demonstrated outstanding capabilities in various natural language tasks ( 43 ; 39 ; 44 ; 25 ; 48 ; 18 ; 17 ) . Among these, LLMs’ exceptional zero-shot reasoning and interactive abilities hold great promise in the field of education ( 26 ; 21 ; 20 ; 35 ) . Leveraging these advantages, researchers have developed LLM-powered intelligent education systems (IESs), available for round-the-clock use, aiming to provide more personalized learning services to assist in automated teaching or learning support applications ( 19 ; 28 ; 10 ; 41 ; 14 ; 45 ; 23 ; 31 ) . (a) Existing LLM-powered IESs. (b) Our proposed AgentTutor. Figure 1: A comparative analysis of existing systems and proposed AgentTutor. AgentTutor incorporates multiple LLM-powered generative agents along with a personalized environment to facilitate learning contexts. Personalized learning services focus on customized learning content, real-time feedback, and interactive experiences ( 28 ) . This adaptability is reflected in the ability to provide systematic guidance toward learners’ learning curriculum, construct knowledge frameworks, and decompose complex problems based on learners’ profile information. Tutors are required to continuously assess learners’ cognitive levels and provide feedback, dynamically adjusting teaching strategies. Therefore, the personalized teaching process is a complex task involving sequential decision-making and multi-turn interactions. Although existing LLM-powered IESs have achieved significant success in handling static single-turn question answering and knowledge reasoning tasks, these studies typically focus on answering specific, narrowly defined problems, as illustrated in Figure 1(a) . Current systems, due to the absence of multi-turn interactive teaching, are unable to comprehensively assess learners’ cognitive levels and learning progress, nor can they dynamically adjust teaching strategies based on real-time feedback. This limitation in multi-turn interaction hinders the effectiveness of these systems in more complex educational contexts, particularly in challenging scenarios like zero-shot inferences. To address these issues, we propose AgentTutor, a multi-turn interactive intelligent education system designed for personalized learning, as shown in Figure 1(b) . Inspired by Bloom’s Taxonomy ( 6 ) and anthropological mechanisms ( 40 ) , AgentTutor decomposes complex curriculum into sub-goals, continuously assesses learner progress, dynamically adjusts teaching strategies using tree search tools, and refines teaching through reflection on past experiences and knowledge. The system features an LLM-powered generative agent and an external learner-specific personalized learning profile environment. The generative agent provides dynamic teaching strategies based on the learner’s cognitive level in a multi-turn process, comprising five modules: curriculum decomposition , learner assessment , dynamic strategy , teaching reflection , and knowledge & experience memory . The personalized learning environment enables learners to configure their learning curriculum, learning preferences, and multimodal study materials, allowing the generative agent to engage in direct, multi-turn interactive teaching. Our contributions are summarized as follows: • We developed AgentTutor, a multi-turn interactive intelligent education system. It applies multiple LLM-powered generative agents that can automate teaching strategies and provide dynamic, multi-turn teaching processes. • AgentTutor is designed for dynamic, personalized learning scenarios, extending beyond single-turn static question answering. It integrates five core modules: curriculum decomposition , learner assessment , dynamic strategy , teaching reflection , and knowledge & experience memory , allowing for dynamic strategies, continuous assessment, and interactive experiences. • We conducted extensive experiments on benchmark datasets and baselines. The system’s effectiveness was evaluated via learners’ improved performance, interactive teaching quality, human evaluations, and ablation studies, confirming the strong competitiveness of AgentTutor. Related Work Intelligent Education Systems Intelligent Education Systems (IESs) offer personalized, real-time learning through computer-based technologies. Early rule-based systems like SCHOLAR ( 8 ) and ACT ( 2 ) guided learners through predefined materials without human instructors, pioneering adaptive teaching. However, they couldn’t dynamically adjust to learners’ needs. With technological advancements, data-driven systems like knowledge tracing and emotional state recognition models ( 12 ; 13 ) improved personalization by analyzing learner behavior. Yet, these systems still lacked deep understanding and high-level reasoning. The introduction of deep learning technologies further advanced IESs. Some researchers introduced a knowledge tracing model ( 30 ) by recurrent neural networks to model learners’ learning levels, significantly improving prediction accuracy. Recently, generative AI models such as ChatGPT ( 1 ) and Gemini ( 3 ) have shown great potential in personalized learning. However, they still face challenges in problem decomposition, multi-turn dynamic learning, and adjusting strategies based on ongoing learner progress, limiting their effectiveness in complex educational tasks ( 5 ; 17 ; 20 ) . LLM-powered Agent Systems Many studies have utilized LLM-powered generative agents combined with domain-specific knowledge to design personalized applications in various scenarios. For instance, communications between users and customer service ( 27 ) , patient-doctor dialogues ( 33 ) , scriptwriting drama ( 38 ) , debates ( 29 ) , and interactions between directors and actors ( 15 ) . Recently, generative agent technologies have been applied to the field of education. For example, agents have been used to extract learners’ cognitive factors to simulate the learning process ( 45 ; 14 ; 28 ; 42 ) , generate and summarize functional reports automatically ( 19 ) , enhance the ability to generate high-quality standard questions based on educational theories ( 10 ) , and predict teachers’ teaching strategies for data annotation ( 41 ) . Although such progress in existing work has made, it often focuses on the implementation of fixed strategies and provides general teaching content, overlooking the dynamic changes in interaction and decision-making throughout the educational process. Therefore, our work focuses on the dynamic adjustment of teaching strategies based on the learner’s cognitive level and learning progress. AgentTutor Figure 2: The diagram highlights the multi-turn interaction framework of the AgentTutor system, showing the collaboration among the five modules and personalized learning profile environment, forming a dynamic process that enhances teaching effectiveness. We take the KNN (K-Nearest Neighbors) algorithm as the curriculum example. We propose AgentTutor to address multi-turn teaching interaction, featuring multiple LLM-powered generative agents and a personalized learning environment that configures curriculum goals, materials, and learning preferences based on the learner’s profile. Operating as a professional tutor, the system integrates five core functional modules in a collaborative framework in Figure 2 to form