Pedagogical LLM Agents
Adaptive Teacher and Student Multi-Agent Simulation: Genetic Policy Adaptation Meets Persona-based Retrieval Augmented Generation (Accepted at EMNLP 2025).
Investigating Pedagogical Teacher and Student LLM Agents: Genetic Adaptation Meets Retrieval Augmented Generation Across Learning Style
Accepted at EMNLP 2025 Multi-Agent Simulation Persona-RAG
Effective teaching requires dynamically adapting instructional strategies to accommodate the diverse cognitive and behavioral profiles of students. In this project, we introduce an adaptive simulation framework pairing LLM-based heterogeneous student agents with a self-optimizing teacher agent whose pedagogical policy dynamically evolves using genetic algorithms.
Direct Links & Resources
- 📄 arXiv Paper: https://arxiv.org/abs/2505.19173
- 📑 PDF: https://arxiv.org/pdf/2505.19173.pdf
Authors & Collaborators
Key Innovations
- Heterogeneous Student Agents: Simulates students with varying learning styles (visual, verbal, reflective, sequential) based on validated educational psychology taxonomies.
- Persona-RAG (Retrieval Augmented Generation): Empowers student agents to query and retrieve knowledge representations specifically tailored to their individual learning styles.
- Genetic Teacher Policy Optimization: The teacher agent generates lesson plans and explanations, observes student performance, and uses crossover, mutation, and fitness evaluation to evolve pedagogical strategies across generations.
- Empirical Findings: The adaptive teacher policy achieves statistically significant learning gains compared to static or non-personalized prompt baselines across diverse student cohorts.