LLM-Based Framework for Interpretable Knowledge Tracing in Tutoring Dialogues

Researchers propose a conversational knowledge tracing framework that uses large language models to assess student performance in AI tutoring systems. The framework explicitly models student abilities and task difficulty at each dialogue turn, integrating Item Response Theory for interpretable predictions. It addresses limitations of existing methods that ignore difficulty modeling and rely on opaque LLM representations. The approach uses original textual questions and subsequent tutor tasks to estimate knowledge states and upcoming difficulty. This work aims to enable personalized support in interactive tutoring systems.