MAGE-KT: Multi-Agent Graph-Enhanced Knowledge Tracing with Subgraph Retrieval and Asymmetric Fusion

Chi Yu, Hongyu Yuan, Zhiyi DuanarXiv 2026 · arXiv · OpenReview

TL;DR

This paper introduces MAGE-KT, a novel framework for knowledge tracing that significantly improves accuracy by building richer knowledge concept (KC) relationships using a multi-agent system and focusing computation on relevant subgraphs. It achieves state-of-the-art results by intelligently fusing student interactions and KC structures, overcoming limitations of existing methods that struggle with noisy relations and inefficient full-graph processing.

What Problem Does This Paper Address?

Knowledge Tracing (KT) is crucial for personalizing education by predicting a student's future performance based on their learning history. A key challenge is accurately modeling the complex relationships between students, questions, and the underlying knowledge concepts (KCs). While graph-based approaches have shown promise, they often suffer from two main issues:

  1. Incomplete and Noisy KC Relationships: Existing methods infer KC-KC relationships primarily from student interaction sequences, which can lead to sparse and inaccurate connections. Prompting large language models (LLMs) can also introduce generative uncertainty. These errors propagate and distort the student's learned knowledge state.
  2. Computational Inefficiency and Attention Diffusion: Encoding entire knowledge graphs for every prediction is computationally expensive and prone to "attention bleed," where the model focuses on irrelevant parts of the graph, degrading performance and increasing noise.

MAGE-KT aims to address these limitations by creating more accurate KC relationships and enabling efficient, focused computation.

Methods

MAGE-KT employs a two-stage approach: Heterogeneous Graph Construction and Multi-view Fusion Prediction.

Stage 1: Heterogeneous Graph Construction

  • Multi-relational KC Graph: This is built using a novel multi-agent pipeline involving five steps:
  • Knowledge Concept Completion: A "Semantic Agent" standardizes KC names and generates detailed definitions.
  • Preliminary Relationship Judgment: The Semantic Agent uses definitions and interaction history to propose a relationship type (e.g., association, containment, equivalence, sibling, predecessor-successor) for KC pairs.
  • Relationship Score: A "Scoring Agent" evaluates the proposed relationships based on specific criteria.
  • Relationship Review: An "Arbitration Agent" cross-checks the agents' outputs against axioms and graph topology.
  • Circular Relationship Cross-Correction: Experts (teaching, structure/semantics, behavior/cognition) review and vote on uncertain relationships. This multi-agent, multi-expert approach aims for higher-fidelity KC relations.
  • Student-Question (S-Q) Interaction Graph: This graph captures student abilities and question difficulties (estimated via IRT) and models relationships between students and questions, as well as between questions and between students.

Stage 2: Multi-view Fusion Prediction

  • Subgraph Retrieval: For a given student and target question, MAGE-KT retrieves compact, high-value subgraphs from both the KC graph and the S-Q graph. This is conditioned on the student's history, ensuring that only the most relevant information is considered, thus avoiding attention diffusion.
  • Asymmetric Cross-Attention Fusion Module: This module fuses the information from the retrieved subgraphs. It uses two directional pathways (KC-to-Question and KC-to-Student) to inject student and question signals into KC embeddings, and then propagates these enhanced KC embeddings back to student and question embeddings. This asymmetric fusion allows for a more nuanced understanding of how KCs relate to both students and questions.
  • Prediction and Loss Function: A GRU maintains the student's evolving state, and the fused vector is used to predict the next-question performance. The model is trained using a standard Cross-Entropy Loss.

Datasets

The paper evaluates MAGE-KT on three widely used public KT datasets:

  • ASSIST09
  • Junyi
  • Statics2011

The authors preprocessed these datasets by removing records with missing fields and filtering out students with too few interactions or questions answered too few times. They used an 8:1:1 train/validation/test split.

Key Insights and Ideas

  • Multi-Agent for KC Relations is Key: The core innovation is using a collaborative multi-agent system (with expert review) to generate more accurate and nuanced KC-KC relationships than traditional methods. This moves beyond simple co-occurrence or temporal order.
  • Subgraph Retrieval is Crucial for Efficiency and Accuracy: Instead of processing the entire knowledge graph, focusing on student-conditioned subgraphs significantly reduces computational cost and prevents noise from irrelevant parts of the graph from impacting predictions.
  • Asymmetric Fusion Captures Deeper Interactions: The directed, asymmetric cross-attention mechanism allows for a more sophisticated fusion of information from different views (KC, student, question), leading to better performance.
  • Complementary Graphs are Better: Combining a rich KC graph with an S-Q interaction graph provides a more holistic view of the learning process, capturing both conceptual dependencies and individual learning dynamics.

Key Results

  • State-of-the-Art Performance: MAGE-KT achieves the best AUC and ACC scores across all three datasets (ASSIST09, Junyi, Statics2011), demonstrating its superior predictive accuracy. For example, on ASSIST09, MAGE-KT achieved 83.06% ACC and 87.89% AUC.
  • Improved KC-Relation Accuracy: The multi-agent approach significantly enhances the accuracy of KC-relation extraction, as validated by ablation studies.
  • Ablation Studies Confirm Design Choices: Removing any of the core components (Asymmetric Attention, KC Graph, S-Q Graph, Subgraph Retrieval) leads to a consistent drop in performance, highlighting the necessity and complementarity of each design element. The removal of the S-Q graph showed the largest performance degradation.
  • Prerequisite-Successor Extraction Validation: Experiments on the Junyi dataset showed that the full multi-agent pipeline significantly outperforms single-agent or ablated versions in extracting prerequisite-successor relationships.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while the multi-agent system improves KC relations, there's still potential for generative uncertainty from LLMs.
  • The current subgraph retrieval is based on a fixed hop number, which might not always be optimal.
  • Future Work:
    • Extend MAGE-KT to continual learning settings for real-time adaptation.
  • Incorporate richer textual signals to further refine KC relations.
  • Investigate uncertainty calibration and human-in-the-loop correction for improved reliability.
  • Explore more sophisticated subgraph retrieval mechanisms.

Summary updated

Abstract

Knowledge Tracing (KT) aims to model a student's learning trajectory and predict performance on the next question. A key challenge is how to better represent the relationships among students, questions, and knowledge concepts (KCs). Recently, graph-based KT paradigms have shown promise for this problem. However, existing methods have not sufficiently explored inter-concept relations, often inferred solely from interaction sequences. In addition, the scale and heterogeneity of KT graphs make full-graph encoding both computationally both costly and noise-prone, causing attention to bleed into student-irrelevant regions and degrading the fidelity of inter-KC relations. To address these issues, we propose a novel framework: Multi-Agent Graph-Enhanced Knowledge Tracing (MAGE-KT). It constructs a multi-view heterogeneous graph by combining a multi-agent KC relation extractor and a student-question interaction graph, capturing complementary semantic and behavioral signals. Conditioned on the target student's history, it retrieves compact, high-value subgraphs and integrates them using an Asymmetric Cross-attention Fusion Module to enhance prediction while avoiding attention diffusion and irrelevant computation. Experiments on three widely used KT datasets show substantial improvements in KC-relation accuracy and clear gains in next-question prediction over existing methods.
BibTeX
@misc{yu2026mage,
  author = {Chi Yu and Hongyu Yuan and Zhiyi Duan},
  title = {{MAGE-KT: Multi-Agent Graph-Enhanced Knowledge Tracing with Subgraph Retrieval and Asymmetric Fusion}},
  year = {2026},
  eprint = {2601.16886},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2601.16886},
}
Figure