MOSAIC Improves Knowledge Tracing with LLM Semantic Alignment

Knowledge Tracing (KT), a crucial component of personalized education, traditionally faces limitations due to its reliance on shallow, ID-based representations and single-granularity mastery estimation. These issues neglect the semantic depth of knowledge and hierarchical dependencies between concepts. This research introduces MOSAIC (Multi-granularity Online Semantic AI for Collaborative Knowledge), a novel framework designed to overcome these challenges. MOSAIC orchestrates large language model (LLM)-driven semantic alignment with sequential modeling. Unlike methods that use LLMs solely for prediction, MOSAIC leverages a frozen LLM to generate dynamic, context-aware embeddings and hierarchical prediction prompts. This approach explicitly captures collaborative signals and peer interactions, enriching the understanding of student learning. Furthermore, the framework incorporates a cross-granularity consistency objective, which jointly regularizes mastery estimation across different levels: individual concepts, topic clusters, and overall proficiency. Extensive experiments on datasets like ASSISTments, EdNet, and a new large-scale MOOC dataset demonstrate MOSAIC's superior performance, achieving AUC improvements of up to 3.4% and Accuracy gains of up to 2.5%. MOSAIC also shows enhanced robustness in collaborative and long-sequence learning scenarios, offering both high predictive precision and semantically grounded interpretability.