Multi-Agent Graph-Enhanced Knowledge Tracing Shows Performance Gains

Global: Multi-Agent Graph-Enhanced Knowledge Tracing Advances Student Performance Prediction

In January 2026, a team of researchers released a new preprint on arXiv introducing Multi-Agent Graph-Enhanced Knowledge Tracing (MAGE-KT), a framework designed to improve the modeling of student learning trajectories and next‑question prediction. The work aims to overcome shortcomings in existing graph‑based knowledge tracing methods by more accurately representing relationships among students, questions, and knowledge concepts while reducing computational overhead.

Background on Knowledge Tracing

Knowledge Tracing (KT) is a subfield of educational data mining that seeks to infer a learner’s mastery of underlying concepts from their interaction history and to forecast future performance. Accurate KT models support personalized learning pathways and adaptive assessment systems.

Limitations of Existing Graph Approaches

Recent graph‑based KT paradigms have demonstrated promise, yet many rely solely on interaction sequences to infer inter‑concept relations, potentially missing richer semantic connections. Moreover, encoding entire heterogeneous KT graphs can be computationally intensive and prone to noise, causing attention mechanisms to allocate resources to irrelevant portions of the graph.

Proposed Multi-Agent Graph-Enhanced Framework

MAGE‑KT addresses these issues by constructing a multi‑view heterogeneous graph that merges two components: a multi‑agent knowledge‑concept (KC) relation extractor that captures semantic links, and a student‑question interaction graph that reflects behavioral data. This dual‑view representation provides complementary signals for downstream prediction.

Subgraph Retrieval and Fusion Mechanism

Conditioned on a target student’s historical interactions, MAGE‑KT retrieves compact, high‑value subgraphs rather than processing the full graph. An Asymmetric Cross‑attention Fusion Module then integrates the extracted subgraphs, focusing attention on relevant regions while mitigating diffusion into student‑irrelevant areas.

Experimental Validation

The authors evaluated MAGE‑KT on three widely used KT datasets. Results indicated notable improvements in KC‑relation accuracy and measurable gains in next‑question prediction metrics compared with previously reported methods.

Potential Impact and Future Directions

By enhancing the fidelity of inter‑concept modeling and streamlining computation, MAGE‑KT could enable more responsive and scalable adaptive learning platforms. The authors suggest further exploration of additional data modalities and real‑time deployment scenarios.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.