Multi-Agent Intelligent Tutoring System

As the landscape of Generative AI rapidly expands, there have been several developments in the space of Intelligent tutoring systems, promising the convenience of AI as a supplementary educational tool. But unfortunately, many of these systems fall short by treating pedagogy as a static, one-size-fits-all process. In this study, we propose a multi-agent intelligent tutoring system that dynamically adapts to each student’s emotional affect, cognitive state, and demonstrated mastery on every conversational turn. The brain of our system orchestrates five specialized agents - an emotional affect monitor, a Bayesian Knowledge tracer, a metacognitive agent, a modality generator agent, and a tutoring agent. This unified pipeline produces responses tuned to both what a student knows and how they are feeling in the moment. Knowledge tracing is implemented using Personalized Bayesian Knowledge Tracing (pBKT), a mathematically grounded approach that maintains per-student, per-topic mastery. Retrieval Augment Generation (RAG) grounds all responses in verified material, reducing hallucinations and improving factual accuracy. We evaluate this system using an LLM-as-judge approach with cross-model scoring to help avoid self-preference bias. Our results suggest that structured multi-agent coordination, when backed by established learning principles, offers a scalable path towards truly personalized education.

Speaker: Setareh Rafatirad, UC Davis

Attend in person or click here to watch online