EZYer Generative Agent Framework

Updated 9 December 2025

EZYer is a generative agent framework for high-school mathematics, simulating interactive classroom environments with curriculum-aligned teaching materials.
It uses a modular design with Teacher, Student, and Controller modules to integrate LLMs, text-corpus retrieval, and rigorous content validation.
The system automates LaTeX Beamer slide generation and employs recursive role-based simulation with multi-layered quality assurance to ensure academic rigor.

EZYer is a generative agent framework designed to simulate a high-school classroom environment, focused on automatic, curriculum-compliant production of teaching materials and interactive notes in mathematics education. It integrates LLM technologies, text-corpus retrieval, and rigorous content validation systems to address existing limitations in educational tools, offering comprehensive service, improved performance, and enhanced interactivity in high school mathematics courseware generation (Yang et al., 2 Dec 2025).

1. System Architecture and Agent Modules

EZYer’s architecture consists of three tightly coupled modules: Teacher Module, Student Module, and Controller, each implemented as a generative agent leveraging a LLM (specifically, DeepSeek-V3) and a high-school mathematics text corpus.

Teacher Module:

Accepts user-provided knowledge points or documents and performs semantic segmentation to split input into chapters $S(I)=\{ C_1, C_2, \dots, C_n \}$ .
For each chapter $C_i$ , it generates structured teaching material $\mathcal{M}(C_i)$ and corresponding exercise problems $\mathcal{E}(C_i)$ in parallel:

$(\mathcal{M} \otimes \mathcal{E}) \circ S(I) = \{ (\mathcal{M}(C_i), \mathcal{E}(C_i)) \}, \quad i=1 \ldots n$

Integrates content and user images ( $\theta$ , $\gamma$ ) into a LaTeX Beamer source $\mathcal{B}$ and compiles final PDF output via $\mathcal{C}$ :

$L = \mathcal{B}(\{ \mathcal{M}(C_i), \mathcal{E}(C_i) \}, \theta, \gamma)$

$P = \mathcal{C}(L)$

Student Module:

Simulates collaborative classroom interaction through recursive role outputs:

Teacher ( $T = f_0(I)$ ): concise explanations.
Assistant ( $A = f_1(T, I)$ ): additional examples.
Top Student ( $TS = f_2(T, A, I)$ ): advanced questions.
Struggling Student ( $SS = f_3(T, A, TS, I)$ ): common pitfalls and misconceptions.
Note Taker ( $NT = f_4(T, A, TS, SS, I)$ ): synthesizes all outputs into polished academic notes.

Controller:

Enforces academic rigor and compliance by integrating:

Keyword Filtering (regular expression whitelist/blacklist of curriculum terms),
Content Scoring (LLM-based scoring on a 1–5 scale for logical coherence, relevance),
Role Co-validation (regenerates content if any role scores below threshold),
Dynamic Content Correction (keyword violations, LaTeX error parsing and auto-correction).

2. Content Generation, Formatting, and Pipelines

EZYer automates the construction of teaching materials and exercises, transforming inputs using structured prompting and semantic boundary detection. Generated courseware is output as LaTeX Beamer slides, with strict adherence to academic typographic conventions and XeLaTeX compilation stability.

LaTeX Beamer Integration:

For each chapter, teaching materials and exercises are formatted as distinct frames, employing bullet lists, enumerations, and mathematical formulas.

Example template:

\documentclass{beamer}
\usepackage{xeCJK}
\title{…}
\begin{document}
  \frame{\titlepage}
  \section{Chapter 1}
  \begin{frame}
    \frametitle{Definition of Function}
    A **function** is a mapping %%%%19%%%% such that …
  \end{frame}
  …
\end{document}

User-Defined Images:

Controller handles image insertion via parameters $\theta$ (path) and $\gamma$ (placement), with automatic escaping of special characters.

\begin{frame}
  \frametitle{Illustration}
  \includegraphics[width=0.5\textwidth]{path/to/image.png}
\end{frame}

3. Interactivity and Collaborative Simulation

The Student Module employs a recursive workflow among distinct roles, collectively simulating classroom interaction. This system enables multi-perspective engagement (explanatory, inquisitive, corrective), with the Note Taker synthesizing all contributions into concise, well-structured notes, thereby reinforcing learning through distillation of key points and formulas.

Each agent role possesses a defined template:

Teacher: aligned to curriculum, concise exposition.
Assistant: further clarification, example expansion.
Top Student: pushes conceptual boundaries via deep questions.
Struggling Student: surfaces common errors and misunderstandings.
Note Taker: filters, edits, and compiles academic notes for distribution.

4. Multi-Layered Quality Assurance

Content quality and propriety are maintained via Controller subsystems:

Keyword Filtering: Only valid mathematical terms pass, blocking out-of-scope material.
Content Scoring: Each output is rated on content accuracy, knowledge coverage, usability, formatting correctness, and visual appeal using structured rubrics.
Role Co-validation: If any agent role produces an inadequate output (score $< 3$ ), automatic regeneration is triggered.
Dynamic Correction: Both semantic and syntactic errors (e.g., LaTeX compilation issues) are identified and corrected iteratively.

These controls ensure that generated materials satisfy both academic requirements and presentation standards.

5. Evaluation Methodology and Empirical Outcomes

EZYer was evaluated using a five-dimensional index:

Content Accuracy
Knowledge Coverage
Usability
Formatting Correctness
Visual Design & Appeal

100 Beamer slides and 100 Notes, generated across varied mathematical topics, were scored independently by five state-of-the-art LLMs (ChatGPT-o1, Claude 3.5, Sonnet, Gemini-2.0-Pro, Qwen2.5-Max, GLM-4-Plus) according to the EZYer rubric.

Output	Content Accuracy	Knowledge Coverage	Usability	Formatting Correctness	Visual Design & Appeal
Beamer	4.74	4.70	4.10	4.35	3.84
Notes	4.70	4.61	4.19	4.21	4.09

High scores on content rigor and curriculum alignment affirm EZYer’s pedagogical validity, while robust formatting showcases reliable LaTeX integration. Usability and visual appeal achieved moderate scores, highlighting areas for refinement in interface design and interactivity latency.

6. Comparative Context and Design Principles

EZYer exhibits architectural kinship with AutoGenesisAgent (Harper, 25 Apr 2024), adopting specialized generative agents orchestrated via message-passing protocols. While EZYer collapses agent roles for rapid, user-friendly educational content generation, its Controller can be viewed as an autonomous approval and correction loop paralleling the Hierarchy Agent in broader multi-agent systems.

Principles from ADAP (Derek et al., 2021) inform latent policy space packaging and modular agent extension, suggesting that a well-designed EZYer system can support diverse agent behaviors and fast adaptation via latent-space search (though not yet implemented in the released version).

Architectural motifs from generative agents for simulacra (Park et al., 2023)—notably memory streaming, retrieval scoring, and planning—can further enhance EZYer’s agent design, suggesting avenues for deeper social simulation and dynamic classroom events.

7. Limitations, Prospects, and Future Directions

EZYer’s established strengths are in automated, curriculum-compliant content generation, academic quality control, and modular agent simulation. Limitations include moderate scores in usability and visual appeal and challenges in latency and interactivity that are typical of LLM-based educational tools. A plausible implication is that extended simulation capabilities (e.g., agent memory, reflection, and temporal planning) may be integrated in future iterations to provide a richer, more emergent classroom environment. Robustness, debugability, and security controls remain essential for reliable production deployment.

EZYer represents a concrete step toward intelligent educational agent frameworks, demonstrating state-of-the-art performance in content generation and rigorous content assurance within mathematics education (Yang et al., 2 Dec 2025).