A research team has introduced a graph-augmented cognitive diagnosis model named HCGCDM in Engineering, which leverages large language models to automatically map multi-type links between educational knowledge concepts and deliver more precise assessments of students’ learning proficiency.

Cognitive diagnosis (CD) acts as the foundational technical support for online adaptive learning systems, relying on student answer logs and expert-built Q-matrices to quantify how well learners grasp discrete knowledge concepts. Conventional CD frameworks, ranging from psychometric models like multidimensional item response theory to deep learning-based neural cognitive diagnosis variants, mostly focus on student-exercise interaction signals. Prior graph-enhanced CD methods only model prerequisite dependencies between concepts, ignoring parallel and synergistic connections that exist widely in educational knowledge systems. For example, matrix row operation and matrix row echelon transformation lack prerequisite ties yet mutually reinforce comprehension, a nuance unaccounted for by single-relation graph structures.

The core innovation of HCGCDM lies in its two sequential functional modules built around LLM reasoning and retrieval-augmented generation (RAG). The heterogeneous concept graph construction module automates graph building without heavy manual annotation or large training datasets. It first embeds all knowledge concepts via BERT-base-nli-mean-tokens to filter high-similarity concept pairs above a predefined cosine similarity threshold, narrowing down candidate pairs for relation detection. The framework then deploys RAG to pull domain-specific external knowledge bases, concatenating retrieved context with structured prompts to guide LLMs in generating relational triples covering three defined edge categories: prerequisite, parallel, and synergistic relations. A second RAG-powered LLM validation step screens unreliable triples to reduce graph noise, and validated triples form a unified heterogeneous graph split into three independent relation-specific subgraphs.

The second core module, heterogeneous concept graph modeling and aggregation, applies tailored graph neural network architectures to each subgraph to match relation characteristics. Two-layer graph attention networks process directed prerequisite subgraphs to capture asymmetric contribution weights of child concepts to parent nodes; two-layer graph convolutional networks encode parallel subgraphs to preserve hierarchical consistency among peer concepts; CoAttention-GNN handles bidirectional information flow within synergistic subgraphs to model mutual knowledge reinforcement. An adaptive selector computes dynamic attention weights for three subgraph embeddings, adjusting the contribution of each knowledge relation type based on exercise features such as difficulty and discrimination. A graph information aggregator then fuses weighted graph embeddings into base representations of knowledge concepts and exercises, feeding integrated features into a multilayer perceptron prediction head to output student response probabilities, with cross-entropy loss used for model parameter optimization.

The research team ran comparative tests against mainstream baseline CD models across multiple real-world educational datasets. Ablation experiments confirmed each core module delivers consistent performance gains, with the LLM-driven heterogeneous graph construction component producing the largest diagnostic metric shifts when removed. Additional controlled trials swapped the base LLM between GPT-4o and Llama3-8B-Instruct, confirming that LLM capability correlates with triple construction accuracy and downstream diagnostic performance. The paper notes several existing limitations, including limited support for datasets lacking textual concept descriptions and reliance on predefined relation categories, and outlines future directions to expand automated relation extraction and incorporate multi-agent LLM pipelines for graph refinement.

The paper “Enhancing Cognitive Diagnosis via LLM-Driven Heterogeneous Concept Graph Construction,” is authored by Yaqing Sheng, Jiuyang Tang, Weixin Zeng, Xiang Zhao, Yuejin Tan. Full text of the open access paper: https://doi.org/10.1016/j.eng.2026.02.017. For more information about Engineering, visit the website at https://www.sciencedirect.com/journal/engineering.