Investigating Pedagogical Teacher and Student LLM Agents: Genetic Adaptation Meets Retrieval Augmented Generation Across Learning Style

TL;DR

This work addresses adaptive pedagogy in AI-driven educational simulations by integrating heterogeneous LLM-based student agents with a self-optimizing teacher evolved via a genetic algorithm, and by introducing Persona-RAG to personalize retrieval to individual learning styles. The teacher’s strategy is evolved using a population-based search to maximize class-wide learning outcomes, while Persona-RAG guides multi-step retrieval aligned with each student’s reasoning path. The framework demonstrates emergence of interpretable teaching patterns and improved learning outcomes across diverse student profiles, with human evaluators rating the generated lectures as engaging and pedagogically sound. Collectively, the approach provides a scalable, data-driven testbed for AI-assisted teacher training and adaptive curricula, with practical implications for individualized education at scale.

Abstract

Effective teaching requires adapting instructional strategies to accommodate the diverse cognitive and behavioral profiles of students, a persistent challenge in education and teacher training. While Large Language Models (LLMs) offer promise as tools to simulate such complex pedagogical environments, current simulation frameworks are limited in two key respects: (1) they often reduce students to static knowledge profiles, and (2) they lack adaptive mechanisms for modeling teachers who evolve their strategies in response to student feedback. To address these gaps, \textbf{we introduce a novel simulation framework that integrates LLM-based heterogeneous student agents with a self-optimizing teacher agent}. The teacher agent's pedagogical policy is dynamically evolved using a genetic algorithm, allowing it to discover and refine effective teaching strategies based on the aggregate performance of diverse learners. In addition, \textbf{we propose Persona-RAG}, a Retrieval Augmented Generation module that enables student agents to retrieve knowledge tailored to their individual learning styles. Persona-RAG preserves the retrieval accuracy of standard RAG baselines while enhancing personalization, an essential factor in modeling realistic educational scenarios. Through extensive experiments, we demonstrate how our framework supports the emergence of distinct and interpretable teaching patterns when interacting with varied student populations. Our results highlight the potential of LLM-driven simulations to inform adaptive teaching practices and provide a testbed for training human educators in controlled, data-driven environments.

Figures (10)

• Figure 1: Complete pipeline for our pedagogical setup. Initially, the student agents are prepared with individual knowledge bases, containing prerequisite knowledge for Math, Science and English as per their subject aptitudes. The teacher agent teaches a topic from these three subjects building upon the prerequisite knowledge of the students, and the students are assessed on how well they have learned the topics. The teacher agent optimizes to increase the average score of the classroom.
• Figure 2: A complete example of a lecture from the teacher agent, the notes taken by the student agent, the student response to the assessment and the performance as scored by LLM-as-a-Judge. We observe that the teacher agent has learned to include real-world examples right after the first definition, provides analogies after presenting formulas for better intuitions, example problems and summaries at the end of the lecture for an overall recap. This leads to a complete and well rounded lecture. The student is able to answer effectively in the exam with the visualizations provided in the class.
• Figure 3: Subject-wise performance with different RAG methods. We observe that while Persona-RAG doesn't achieve the peak accuracies, the mean accuracy for each subject is high since it allows to student retrieve notes the in the same learning style as they wrote the notes in.
• Figure 4: Box plots representing the distribution of student scores for different RAG techniques. While both Persona-RAG and HyDE have comparable performances, HyDE achieves the peak scores but Persona-RAG has a higher average score. It is worth noting that Persona-RAG is a better fit for pedagogical environment since it achieves comparative performance to HyDE at a fraction of time and compute.
• Figure 5: Performance of students with different learning styles over 50 generations. These experiments were done with the teacher optimizing over classes with students with homogeneous learning types. We observe a clear increasing trend in marks attained in each of the plots, indicating that the teacher can successfully identify the specific learning cues for each learning type. The peak average marks achieved by the class varies, and we can observe that Intuitive and Technical learners have higher average scores.