Interpretable Difficulty-Aware Knowledge Tracing in Tutor-Student Dialogues

Analysis

This paper bridges LLM-based dialogue tutoring and interpretable student modeling. By mapping opaque LLM representations to Item Response Theory parameters — student ability (θ) and question difficulty (b) — the framework makes turn-by-turn predictions both accurate and cognitively meaningful. This connects directly to Knowledge Tracing IRT by extending IRT beyond static assessment into live dialogue.

The framework was validated across two tutor-student dialogue datasets and outperformed existing KT baselines. The approach also operationalizes Intelligent Tutoring by enabling tutors to calibrate scaffolds based on explicit difficulty-aware readiness estimates, and supports LLM Student Modeling Memory by providing a principled way to convert LLM outputs into structured student state representations.