ExpertAgent: Enhancing Personalized Education through Dynamic Planning and Retrieval-Augmented Long-Chain Reasoning

AI summary

Overview

Research area: Artificial intelligence in education (AIED), specifically LLM-based pedagogical agents that combine retrieval-augmented generation (RAG), chain-of-thought (CoT) reasoning, and dynamic planning with a continuously updated student model.

Technical level: Intermediate. The paper is a system and framework description rather than a mathematically dense methods paper; it presumes familiarity with LLM concepts such as RAG and CoT but presents them in accessible terms.

Scope: The paper proposes and describes ExpertAgent, an integrated learning-agent framework that grounds instructional content in a validated curriculum repository and adapts to a learner via a dynamic student model, then reports an internal user-acceptance test with volunteers scored on four dimensions.

What This Paper Is About

Current AI-in-education systems struggle with real-time adaptability, shallow personalization, limited long-term memory across learning dialogues, and unreliable output such as hallucinations. ExpertAgent addresses these problems by pairing proactive teaching with a student model that is updated after every interaction, and by grounding every generated explanation in retrieved material from a validated curriculum repository. The goal is a learning agent that behaves less like an information provider and more like an adaptive teacher that adjusts difficulty, recommends targeted exercises, and explains its reasoning and sources.

Key Contributions

1. Design and implementation of ExpertAgent, a framework that delivers an interactive, personalized learning experience by integrating dynamic planning with RAG for reliable, adaptive educational content.
2. A dynamic student model that tracks user progress, records knowledge states and performance trajectories, and provides feedback that drives subsequent instructional planning.
3. An integration argument for three normally separate techniques — RAG for factual reliability, CoT for reasoning transparency, and dynamic student modeling — unified into a single teaching–learning feedback loop.
4. A working multi-module system comprising an Interactive Learning module, a "Track My Progress" module, a Document Management module, a Feedback Mechanism, a Data Storage system (CSV and JSON), and a Modular Architecture that ties the interface, content generation, retrieval, and storage together.

Main Findings

• User acceptance results: In an internal test with volunteers using the User Acceptance Model across four categories rated on a Likert scale, performance expectancy scored 4.33, effort expectancy 4.22, and facilitation conditions 4.22. Social influence was noticeably lower at 2.78, indicating that external drivers such as peer recommendation or organizational support remain limited for the system.
• Perceived strengths: The high scores on performance expectancy, effort expectancy, and facilitation conditions indicate users generally believed the system could improve learning efficiency, was easy to use, and offered adequate environmental and resource support.
• Perceived weakness: The low social influence score points to a need for stronger external adoption and peer-driven participation.
• Interface-level claims: The teaching interface (Brief Summary, Content Type Selection, Knowledge Details, AI Q&A, Retrieved Snippets, Learning Feedback) and the practice interface (Practice Session, Answer Review, Learning Advice, Knowledge Map, Feedback Area, AI Learning Tips) are described as forming complete learning loops with "positive effects" on learning efficiency, knowledge mastery, and learner satisfaction.
• Knowledge map coding: Learner mastery status is visualized with four colors — blue for "Untouched," yellow for "Learning," green for "Mastered," and red for "Weak."
• Literature gap identified: Prior work typically applies RAG, CoT, and dynamic planning in isolation — RAG mostly in open-domain QA with little curricular grounding, CoT narrowly on reasoning tasks, and dynamic planning concentrated in robotics and gaming with minimal adaptation to student modeling.
• Not reported: The paper does not report the number of volunteer participants, any accuracy or benchmark metrics, dataset sizes, baseline comparisons, or statistical significance tests. There is no quantitative evidence beyond the four User Acceptance Model averages.

Methodology in Plain English

The authors built a working educational agent and evaluated it with people rather than on a benchmark.

On the teaching side, instructional materials such as textbooks and documents are first segmented and embedded into a vector database. When a student submits a query, the system retrieves relevant context through RAG, passes that retrieved knowledge to the ExpertAgent module to enhance the prompt, and produces a structured response using chain-of-thought reasoning. This step is designed to do more than give a correct answer: it explains why a solution or concept applies and indicates the source of the reference.

On the student side, the system maintains a structured student model recording progress across exercises, knowledge states, and performance trajectories, including conceptual dependencies. The model is updated after each interaction and captures both correct reasoning and misconceptions. Those updates then feed back into instructional planning — recommending targeted exercises, adjusting task difficulty, and highlighting weak or emerging concepts.

The two sides are connected by a bidirectional loop: the student submits a query or exercise response; ExpertAgent retrieves context and generates a pedagogically informed answer; the student model updates; and that updated state guides the next round of planning.

For evaluation, the authors ran an internal test with volunteers and analyzed responses with the User Acceptance Model, covering performance expectancy, effort expectancy, social influence, and facilitation conditions, each rated on a Likert scale and averaged.

Why This Matters

Impact on research: The paper argues for integrating three techniques that prior AIED work typically treats separately. It also frames the specific failure modes each technique carries — RAG's rigid pipeline and error propagation, CoT's narrow applicability to mathematical or logical reasoning, and reward hacking in dynamic planning when educational rewards are oversimplified (for example, test scores) — which sharpens the case for combined frameworks rather than isolated ones.

Real-world applications:

• Personalized tutoring that recommends targeted exercises and adjusts difficulty based on an evolving model of the learner rather than static rules.
• Teacher support through document management that processes multi-format materials, extracts key topics, and answers student questions grounded in validated curriculum content.
• Learning assessment via interactive quizzes, knowledge maps, and AI-generated review recommendations tied to specific errors and weaknesses.
• Source-traceable explanations in study tools, where retrieved snippets let learners verify where an answer came from.

Industry relevance: The framework addresses the trust barrier that most directly slows adoption of generative AI in classrooms — hallucination, biased output, and overreliance. Grounding instructional content in a validated curriculum repository and surfacing references is a concrete mechanism for reducing those risks. Its modular architecture and structured storage of feedback and learning trajectories (CSV and JSON) also point toward a deployable, scalable product shape rather than a research prototype.

Future Directions

• Scaling the evaluation across diverse subjects, different learning content, and different user groups, since the current test is internal and small.
• Longitudinal study to investigate the long-term impact of ExpertAgent on both educators and students.
• Strengthening mathematical and symbolic reasoning through reinforcement learning, so the system can assist with technical calculations, process analysis, and quantitative decision support.
• Raising social influence, the one User Acceptance Model dimension that scored noticeably low, which the authors attribute to limited external drivers such as peer recommendation and organizational support.

Target Audience

Researchers and practitioners in AI in education, learning-science, and educational technology who want a concrete example of how RAG, chain-of-thought reasoning, and dynamic student modeling can be combined into one agent. It is also useful for developers building LLM-based tutoring or assessment products who need to understand grounding and traceability as design choices. Instructors, curriculum designers, and institutional decision-makers evaluating AI teaching assistants will find the modular system breakdown and the User Acceptance Model results accessible. Readers looking for rigorous quantitative benchmarking, baseline comparisons, or large-scale learner studies will not find them here; the evidence is a small internal acceptance test.

Authors’ abstract

The application of advanced generative artificial intelligence in education is often constrained by the lack of real-time adaptability, personalization, and reliability of the content. To address these challenges, we propose ExpertAgent - an intelligent agent framework designed for personalized education that provides reliable knowledge and enables highly adaptive learning experiences. Therefore, we developed ExpertAgent, an innovative learning agent that provides users with a proactive and personalized learning experience. ExpertAgent dynamic planning of the learning content and strategy based on a continuously updated student model. Therefore, overcoming the limitations of traditional static learning content to provide optimized teaching strategies and learning experience in real time. All instructional content is grounded in a validated curriculum repository, effectively reducing hallucination risks in large language models and improving reliability and trustworthiness.