💻 computer science

FACET: Teacher-Centred LLM-Based Multi-Agent Systems-Towards Personalized Educational Worksheets

This paper introduces FACET, a teacher-centered multi-agent LLM system that generates personalized mathematics worksheets by simulating diverse learner profiles and applying didactical principles to address the cognitive and motivational heterogeneity of student populations.

Original authors: Jana Gonnermann-Müller, Jennifer Haase, Konstantin Fackeldey, Sebastian Pokutta

Published 2026-03-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jana Gonnermann-Müller, Jennifer Haase, Konstantin Fackeldey, Sebastian Pokutta

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a classroom as a bustling kitchen. The teacher is the head chef, and the students are the diners. The problem? Every diner has a completely different palate. Some are starving and ready to eat anything (high motivation, high skill), some are full and picky (high skill, low motivation), some are hungry but can't handle spicy food (low skill, high motivation), and some are just too overwhelmed to look at the menu (low skill, low motivation).

Traditionally, the chef has to cook one giant pot of soup for everyone. It's a compromise: it's not too hot, not too cold, but it's rarely perfect for anyone. The chef spends hours trying to ladle out different portions, often getting tired and burnt out.

Enter FACET: The "Smart Sous-Chef" Team.

This paper introduces FACET, a new AI system designed to be a team of virtual assistants that helps the teacher (the head chef) create a unique, perfect meal for every single student, without the teacher having to cook it all from scratch.

Here is how it works, broken down into simple parts:

1. The Three Specialized Robots (The Multi-Agent System)

Instead of one big, confused AI trying to do everything, FACET uses a team of three specialized "robots" (agents) that talk to each other:

The "Student Simulator" (Learner Agent):
Think of this robot as an actor. Before the teacher even writes a worksheet, this robot pretends to be a specific student. It says, "I'm a student who loves math but gets bored easily," or "I'm a student who is great at math but feels very anxious." It tries to solve a math problem and shows the teacher exactly where it gets stuck, what it feels frustrated about, and what it enjoys.
- Analogy: It's like a dress rehearsal where the robot plays the role of the student so the teacher can see exactly what the student needs before the real show.
The "Teacher Assistant" (Teacher Agent):
This is the creative chef. It watches the "Student Simulator" act out the problem. Based on that performance, it designs a custom worksheet. If the "student" was anxious, the worksheet includes encouraging notes and easier starting steps. If the "student" was bored, it adds a tricky challenge. It uses educational rules (like Bloom's Taxonomy, which is just a fancy way of saying "start simple, then get harder") to make sure the lesson is just right.
- Analogy: It's like a tailor who looks at the "student simulator's" measurements and sews a suit that fits perfectly, rather than handing out a one-size-fits-all t-shirt.
The "Quality Inspector" (Evaluator Agent):
This robot is the strict food critic. Before the worksheet goes to the real teacher, the Inspector checks it. It asks: "Is this too hard? Is the instruction clear? Is it creative? Does it actually match the student's needs?" If the answer is no, it sends it back for a redo.
- Analogy: It's the health inspector in the kitchen making sure the meal is safe, nutritious, and delicious before it leaves the building.

2. Why This Matters

Currently, most AI tools for schools are like a vending machine: you put in a grade level, and you get a generic worksheet. They only care about scores (did the student get the answer right?).

FACET is different because it cares about the whole person. It understands that a student might know the math but be too scared to try, or be bored because the math is too easy. It tries to fix the emotional and motivational side of learning, not just the brain side.

3. The Test Drive

The researchers tested this system using real 8th-grade math problems from German schools. They created four types of "fake students" (e.g., the "Anxious High-Achiever" and the "Bored Struggling Student") and asked the FACET team to make worksheets for them.

The Results:

The Robots agreed: The "Quality Inspector" gave the worksheets high marks, saying they were clear, well-structured, and perfectly matched the student's needs.
Real Teachers were impressed: When real teachers looked at the worksheets, they loved them. One teacher said it was like having a "superpower" because they could finally keep the whole class on the same topic without the fast kids getting bored or the slow kids getting left behind. Another teacher said they wished they could use it to "test-drive" exam questions before giving them to real kids.

The Bottom Line

FACET isn't trying to replace teachers. It's trying to give them a superpower.

Right now, teachers are overwhelmed trying to personalize lessons for 30 different kids. FACET acts as a tireless assistant that simulates the students, drafts the perfect lesson plan, and checks the quality, all in seconds. This allows the human teacher to focus on what they do best: connecting with their students and guiding them through the learning journey.

In short: FACET turns the "one-size-fits-all" classroom into a "made-to-order" experience for every single student.

1. Problem Statement

The paper addresses two critical limitations in current AI-driven educational tools:

Performance-Centric Narrowness: Existing personalization systems primarily adapt content based on cognitive performance (e.g., quiz scores), neglecting the motivational, emotional, and self-concept dimensions that are crucial for learning outcomes, particularly in mathematics.
Teacher Exclusion: Most AI tools are "student-facing" (directly interacting with students) or offer static repositories. They fail to provide direct, scalable support to teachers who must differentiate instruction for increasingly heterogeneous classrooms. Teachers struggle to create personalized materials that address both cognitive and motivational needs within existing time constraints.

2. Methodology: The FACET Framework

The authors propose FACET (Framework for Agent-based Classroom Enhancement for Teacher), a modular, teacher-facing multi-agent system built on Large Language Models (LLMs).

System Architecture

The system utilizes three specialized LLM-backed agents with distinct roles, communicating via structured transcripts and typed interfaces:

Learner Agents (Simulation):
- Function: Simulate diverse student personas by processing a teacher-defined task prompt.
- Mechanism: They generate self-reflective outputs capturing reasoning steps, misconceptions, emotional states, and perceived self-efficacy.
- Variables: Profiles are defined by two core dimensions: Mathematical Proficiency (High/Low) and Intrinsic Motivation (High/Low), based on Self-Determination Theory.
Teacher Agent (Generation):
- Function: Acts as the instructional designer. It consumes the learner agent's transcript and applies didactical directives.
- Mechanism: It employs a three-step process:
  - Analysis: Evaluates the learner model against normative benchmarks.
  - Personalization Layer: Generates adaptive responses, including motivational feedback, interest-based customization, and scaffolding. It maps learner profiles to tasks aligned with Bloom's Taxonomy (reproduction, making connections, generalizing).
  - Output Generator: Compiles a coherent, one-page Markdown worksheet containing tasks, hints, and motivational comments.
Evaluator Agent (Quality Assurance):
- Function: Provides automated feedback on the generated worksheet.
- Metrics: Scores the output on a 1–6 scale across four dimensions: Didactical Structure, Clarity, Creativity, and Suitability for the specific learner profile.

Technical Implementation

Models: The system uses GPT-4.1 for reasoning-heavy agents (Learner/Teacher) and GPT-4o for the Evaluator and formatting. It supports over 100 LLMs via LiteLLM.
Workflow: The pipeline executes sequentially (Learner $\rightarrow$ Teacher $\rightarrow$ Evaluator). A single iteration was found sufficient for high accuracy, avoiding the latency of iterative refinement.
Prompt Engineering: Agents use reusable prompt templates with slot-filling for grade, topic, and profile parameters, reducing the prompting overhead for teachers.

3. Key Contributions

Teacher-Centric Multi-Agent Architecture: A novel system designed to support teachers rather than replace them, enabling the generation of individualized materials for heterogeneous classes.
Holistic Personalization: Integration of motivational and emotional dimensions (intrinsic motivation, self-concept) alongside cognitive proficiency, moving beyond simple performance adaptation.
Scalable Simulation: The use of persistent agent profiles allows for the rapid generation of diverse, reusable learner personas without extensive manual setup.
Empirical Validation: The framework was tested using authentic 8th-grade mathematics curriculum content from German schools, validated through both automated agent-based assessment and human expert feedback.

4. Results

The study evaluated the system using ten internal iterations per learner profile and feedback from five K-12 in-service teachers.

Agent-Based Evaluation (Internal Validity):
- The system demonstrated high stability and alignment.
- Suitability (Mean = 4.78/6) and Clarity (Mean = 4.76/6) received the highest scores.
- Didactical Structure (Mean = 4.12/6) and Creativity (Mean = 3.94/6) were also rated favorably.
- The generated worksheets successfully avoided under- and over-challenging tasks by dynamically selecting difficulty based on the simulated learner's profile.
Teacher Expert Evaluation (External Validity):
- Teachers rated Didactical Structure (Mean = 4.95/6) and Suitability (Mean = 5.05/6) very highly.
- Qualitative Feedback: Teachers appreciated the ability to keep the whole class on the same topic while varying task complexity (preventing boredom for high achievers and overload for struggling students).
- Suggestions for Improvement: Teachers requested more explicit scaffolding (e.g., worked examples) for low-motivation/low-knowledge students and deeper narrative integration to boost engagement.

5. Significance and Future Directions

Bridging the Theory-Practice Gap: FACET demonstrates that multi-agent LLM systems can operationalize complex educational theories (like Self-Determination Theory and Bloom's Taxonomy) into practical, scalable classroom tools.
Scalability: The framework offers a pathway to handle the "heterogeneity crisis" in modern classrooms by automating the creation of differentiated materials without requiring teachers to be prompt engineering experts.
Future Work:
- Richer Profiles: Expanding learner simulations to include language background, attentional variability, and neurodiversity (e.g., ADHD, dyslexia).
- Real-World Trials: Moving from exploratory feedback to large-scale, in-classroom evaluations with real student data.
- Curriculum Integration: Embedding full curriculum specifications and domain-specific didactic theories into the agent prompts to ensure stricter alignment with educational standards.

In conclusion, the paper establishes that multi-agent LLM architectures are a viable and promising approach for generating context-aware, holistic personalized educational materials, shifting the focus from purely cognitive adaptation to a more inclusive model that addresses the emotional and motivational needs of diverse learners.