Fast adaptation to student cognitive states: A confidence-aware meta-learning framework for intelligent tutoring systems

Abstract

Intelligent Tutoring Systems face significant challenges in dynamically adapting to diverse student cognitive states (e.g., knowledge gaps, learning pace, and engagement levels) while maintaining high confidence in pedagogical decisions. Traditional rule-based or static Machine Learning (ML) models often fail to generalize across different learners and subjects. A confidence-aware Meta-Reinforcement Learning (Meta-RL) framework is proposed, allowing for fast adaptation to individual student needs with quantifiable uncertainty estimation. The Proposed Method (PM) leverages meta-learning to pre-train a policy on a distribution of simulated and real-world student interactions, allowing rapid fine-tuning with minimal data for new learners. The framework incorporates Bayesian Neural Networks (BNN) to assess prediction confidence, ensuring that tutoring actions (e.g., hint provision or problem difficulty adjustment) are personalized and reliable. Experiments on large-scale educational datasets (e.g., ASSISTments, MOOC logs) demonstrate that the model outperforms baseline methods (e.g., deep RL, non-adaptive ITS) in learning gain (+12%) and student engagement (+18%), with real-time deployment feasibility on edge devices. This work bridges the gap between high-speed adaptation and high-confidence AI in education, offering a scalable solution for next-generation ITS.