MAViS-KT: A Large-Small Model Collaborative Framework for Explainable Knowledge Tracing

ACL ARR 2026 January Submission2482 Authors

Keywords: Knowledge Tracing; Multi-View; Multi-Agent; Collaboration

Abstract: Knowledge Tracing (KT) is pivotal for personalized education, aiming to predict students' future performance by modeling their evolving knowledge states. However, traditional Deep Learning methods operate as opaque black boxes, while Large Language Models (LLMs) offer interpretability but suffer from high computational costs and unstable reasoning. In practice, not only accurate predictions are needed, but also interpretable reports to support effective learning interventions. To bridge this gap, we propose Multi-Agent View Synergy Knowledge Tracing (MAViS-KT), a novel large-small model collaborative framework that synergizes the semantic depth of LLMs with the numerical robustness of lightweight networks. Specifically, we design a multi-view multi-agent debate mechanism to disentangle complex learning signals and ensure reasoning fidelity through collaborative verification. Furthermore, to address the semantic-numerical disconnect, we introduce a trainable correction module that dynamically aligns qualitative insights with precise probability estimates. Experiments show that MAViS-KT outperforms strong baselines in accuracy while offering high-quality, actionable educational insights, effectively combining the strengths of qualitative reasoning and quantitative modeling.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: educational applications

Contribution Types: Model analysis & interpretability

Languages Studied: English

Submission Number: 2482