A Graph-Based Foundation Model for Sample-Efficient Adaptive Learning

Educational assessments stand to benefit significantly from the recent advancements in foundation models (FMs) and reinforcement learning (RL). However, most Massive Open Online Courses (MOOCs) still adopt a non-adaptive, sequential structure, failing to tailor to the heterogeneity of student needs. In this paper, we present a novel graph-based foundation model for adaptive learning path personalization, designed to optimize educational content sequencing based on automatic elicitation of student needs. Leveraging pre-trained graph neural networks (GNNs) and RL, our model recommends documents within the zone of proximal development of students, within few episodes of fine-tuning, significantly reducing the need for costly human interaction data. In particular, we demonstrate on synthetic student data from real educational resources that pre-training on sequential corpora present in MOOCs improves even more the sample efficiency, and generalizes to corpora not seen in the training set, which existing RL approaches cannot do. This work contributes to the growing field of AI in education by proposing a scalable, sample-efficient foundation model that can personalize learning paths for diverse learner needs.