Exploring Latent Associations Between Exercises and Concepts With Large Language Models for Cognitive Diagnosis

Abstract

Cognitive diagnosis (CD) is a fundamental task in intelligent education that aims to assess students’ mastery of fine-grained knowledge concepts. Despite recent progress, most existing CD approaches heavily rely on expert-annotated Q-matrices, which are often incomplete and fail to capture latent exercise–concept associations arising in authentic problem-solving processes. To address this limitation, we propose Exploring Latent Associations between exercises and concepts with large language models (LLMs) for CD (ELACD), a novel framework that formulates latent association discovery as a dual-view cognitive inference problem. ELACD identifies latent exercise–concept associations from two complementary perspectives. From the semantic view, a large language model (LLM) enhances exercise representations by extracting implicit conceptual cues from textual content, facilitating the discovery of latent exercise–concept associations. From the structural view, we construct a student–exercise–concept interaction graph and employ graph neural networks (GNNs) to capture high-order relational dependencies that reflect empirical co-occurrence patterns in learner interactions. To effectively reconcile heterogeneous evidence from these two views, we introduce a consensus-informed fusion mechanism that adaptively integrates semantic and structural latent associations, enabling more faithful modeling of student–exercise interactions. Extensive experiments across multiple benchmark datasets demonstrate that ELACD consistently outperforms state-of-the-art CD methods. Furthermore, case studies reveal that the discovered latent associations align closely with real cognitive mastery patterns, highlighting the interpretability and practical reliability of the proposed framework.