Reinforcement Learning Measurement Model for Interactive Assessment

Key Findings

Interactive assessments generate sequential process data that conventional item response models (IRT) cannot adequately handle. This paper proposes a reinforcement learning measurement model that links action choices to state-action values, extending beyond existing MDP-based measurement approaches.

The model addresses the gap between traditional static assessment models and the dynamic, interactive nature of modern computer-based assessments. It builds on prior work (LaMar, 2018) but improves reliability of estimates for interactive assessments where students' action sequences carry diagnostic information.

Connections to AIED

This work directly extends Knowledge Tracing IRT by replacing static IRT with a dynamic RL-based approach. It connects to Assessment by providing a new framework for scoring interactive assessments that capture student problem-solving processes rather than just final answers.

The approach also relates to Knowledge Tracing and Learning Analytics as it extracts diagnostic signals from interaction data that go beyond traditional performance metrics.

Connected Concepts

Connected Articles

Citation

Xu, W., & Ji, F. (2026). Reinforcement learning measurement model. arXiv:2605.09305.