EduPlanner: LLM-Based Multi-Agent Systems for Customized and Intelligent Instructional Design
Xueqiao Zhang†, Chao Zhang†, Jianwen Sun, Jun Xiao, Yi Yang, and Yawei Luo‡ · 2025
Read EduPlanner: LLM-Based Multi-Agent Systems for Customized and Intelligent Instructional Design by Xueqiao Zhang†, Chao Zhang†, Jianwen Sun, Jun Xiao, Yi Yang, and Yawei Luo‡ (2025) 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-2504-05370.txt.
Introduction
Xueqiao Zhang, Chao Zhang, Jianwen Sun, Jun Xiao, Yi Yang, and Yawei Luo build EduPlanner, a team of AI agents that designs lessons in two thousand twenty-five. Good lesson design must fit each class's level and then improve with feedback from results. A single large model struggles to run that whole loop well.
Three agents share the work: one checks questions, one grades plans, and one improves them, in a sparring setup. A skill tree maps what a student group already knows in math, so plans can be tuned to that ground. A five-part grading module scores clarity, completeness, depth, and practical value. The system aims at plans that are both tailored and steadily refined.
Catalog record
| CREATOR | Xueqiao Zhang†, Chao Zhang†, Jianwen Sun, Jun Xiao, Yi Yang, and Yawei Luo‡ |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 42 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-05370
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
EduPlanner: LLM-Based Multi-Agent Systems for Customized and Intelligent Instructional Design Xueqiao Zhang†, Chao Zhang†, Jianwen Sun, Jun Xiao, Yi Yang, and Yawei Luo‡ Abstract Large Language Models (LLMs) have significantly advanced smart education in the Artificial General Intelligence (AGI) era. A promising application lies in the automatic generalization of instructional design for curriculum and learning activities, focusing on two key aspects: (1) Customized Generation: generating niche-targeted teaching content based on students’ varying learning abilities and states, and (2) Intelligent Optimization: iteratively optimizing content based on feedback from learning effectiveness or test scores. Currently, a single large LLM cannot effectively manage the entire process, posing a challenge for designing intelligent teaching plans. To address these issues, we developed EduPlanner, an LLM-based multi-agent system comprising an evaluator agent, an optimizer agent, and a question analyst, working in adversarial collaboration to generate customized and intelligent instructional design for curriculum and learning activities. Taking mathematics lessons as our example, EduPlanner employs a novel Skill-Tree structure to accurately model the background mathematics knowledge of student groups, personalizing instructional design for curriculum and learning activities according to students’ knowledge levels and learning abilities. Additionally, we introduce the CIDDP, an LLM-based five-dimensional evaluation module encompassing C larity, I ntegrity, D epth, P racticality, and P ertinence, to comprehensively assess mathematics lesson plan quality and bootstrap intelligent optimization. Experiments conducted on the GSM8K and Algebra datasets demonstrate that EduPlanner excels in evaluating and optimizing instructional design for curriculum and learning activities. Ablation studies further validate the significance and effectiveness of each component within the framework. Our code is publicly available at https://github.com/Zc0812/Edu_Planner . I INTRODUCTION In the educational realm, formulating pedagogical tasks presents a challenge that involves numerous scenarios and test cases. The crafting of instructional design for curriculum and learning activities [ 1 , 2 ] is an essential but difficult component of these tasks. It requires a considerable investment of time and energy from educators to understand the diverse learning profiles of their students before making an instructional design for curriculum and learning activities, ensuring a comprehensive structure, clarity of content, the inclusion of appropriate introductions and illustrative examples, as well as reflecting on and optimizing teaching behavior after class. To explore the pedagogical priorities that shape teachers’ instructional designs, Wu et al. [ 3 ] revealed that 40.2\% educators consider teaching content the paramount aspect, 29.4\% place heightened emphasis on the application of teaching methods, and 16.3\% focus on example problem explanations. The remaining 14.1\% focus on an in-depth analysis of the student situation, as depicted in Fig. 1 . Traditionally, teachers have relied on their accumulated experience and innate intuition to devise instructional frameworks. This process often necessitates a trial-and-error approach, with iterative improvements made through classroom implementation. Moreover, the diversity of student dispositions and the unpredictability of the classroom environment compound the difficulty of crafting comprehensive instructional design for curriculum and learning activities prior to instruction. The absence of a standard metric to evaluate the quality of these plans further exacerbates the complexity of this task, rendering the entire procedure laborious and often inefficient. Fig. 1: The proportion of people who consider various factors critical in the task of generating lesson plans. Taking mathematics lessons as our example, in EduPlanner, we model the student situation using a newly proposed Skill-Tree ability module. We iteratively evaluate and optimize the teaching method and teaching content through the adversarial collaboration between the Evaluator Agent and the Optimizer Agent. Additionally, an Analyst Agent is employed to analyze error-prone examples and incorporate them into the subsequent lesson plan. EduPlanner effectively addresses the most pressing challenges in lesson planning. The emerging trend in Generatine AI, distinguished by its ability to generate a diverse range of content, including texts, audio, images, and videos, has gradually penetrated the realm of education [ 4 , 5 , 6 , 7 , 8 , 9 ] , thus influencing various aspects of traditional pedagogy. Recent advances in GAI, such as LLM, have positioned them as promising solutions for instructional design [ 10 , 11 ] . LLMs have the ability to develop sophisticated virtual environments and analyze complex social dynamics [ 12 , 13 , 14 ] . By observing agent responses and metric scores, researchers can gain insight into the effectiveness of various strategic approaches within instructional design. However, existing zero-shot instructional design generation methods [ 15 ] , which depend entirely on the capabilities of LLM, often simply follow instructions to perform tasks without addressing the individual needs of students. In addition, the existing methods [ 16 ] only concentrate on the generation and evaluation of instructional designs, yet they fail to incorporate the application of LLM in the optimization process of instructional designs. Consequently, these instructional designs tend to be dull in knowledge transmission, deficient in case explanations, and devoid of a setting in course difficulty levels. To address these challenges, this paper introduces “EduPlanner”, an automated multi-agent system designed for the generation, evaluation, and optimization of instructional design. Taking mathematics lessons as our example, “EduPlanner” enables teachers to employ a variety of mathematical instructional design drafts, simulate classroom scenarios, and make targeted adjustments to instructional strategies based on feedback. The system comprises an evaluator agent, an optimizer agent, and a question analyst, which collaborate adversarially to generate customized and intelligent instructional design for curriculum and learning activities. “EduPlanner” begins by assessing the overall learning situation of the class using an innovative Skill-Tree structure, which informs the creation of teaching resources that align with individual learning trajectories, thus improving instructional efficacy and student engagement. Based on personal evaluation of the class, “EduPlanner” generates an initial instructional design for curriculum and learning activities and forwards it to the evaluator agent. The evaluator agent employs a five-dimensional evaluation system comprising C larity, I ntegrity, D epth, P racticality, and P ertinence (CIDPP). Feedback from the evaluator is then used by the optimizer agent to improve the instructional design and by the analyst agent to record typical examples. This adversarial cycle mimics a real teacher’s iterative process of instructional design for curriculum and learning activities. We evaluate the effectiveness of our framework using CIDPP-based experiments on the GSM8K [ 17 ] and Algebra [ 18 ] datasets, as illustrated in Fig. 2 . Compared to conventional methods, EduPlanner significantly improves the efficiency of instructional design for curriculum and learning activities, reduces pedagogical burdens, and facilitates precise alignment with the diverse educational needs of students. The main contributions of this study can be summarized as follows: • We introduce “EduPlanner”, a novel framework for customized and intelligent instructional design including three agents: Evaluator Agent, Optimizer Agent, and Analyst Agent. • We propose an innovative Skill-Tree structure to craft a personalized learning profile for each student which provides corresponding input for instructional generators, enabling a focus on tailored student needs when structuring educational content. • We design a comprehensive five-dimensional instructional design evaluation system (CIDPP), and our framework achieves state-of-the-art results in experiments utilizing this system. (a) (b) Fig. 2: Examples of datasets. On the left is GSM8K [ 17 ] and on the right is Algebra [ 18 ] . GSM8K is a dataset consisting of 8.5K high-quality, language-diversified elementary school mathematics word problems