This study presents a novel model-agnostic framework that integrates large language models (LLMs) with cognitive diagnosis models (CDMs) to enhance the accuracy and robustness of cognitive assessments in intelligent education systems. The core innovation lies in bridging the semantic space of LLMs with the behavioral space of CDMs through a two-stage approach: LLM diagnosis and cognitive level alignment.

The framework addresses two critical limitations of conventional CDMs. First, it overcomes the cold-start problem by leveraging LLMs’ rich domain knowledge and reasoning capabilities to generate comprehensive textual diagnoses of students' cognitive states and exercise attributes. Second, it resolves the representational gap between LLMs’ semantic understanding and CDMs’ behavioral modeling through advanced alignment techniques. The LLM diagnosis module employs SOLO taxonomy (Structure of Observed Learning Outcomes) to systematically analyze learning processes across five cognitive levels, from surface understanding to deep integrative comprehension. This enables fine-grained assessment of both students and exercises, particularly valuable for data-sparse scenarios.

The cognitive level alignment module employs two complementary strategies: mixed contrastive alignment and interactive reconstruction alignment. These techniques effectively project LLM-generated semantic representations into CDMs’ behavioral space while preserving mutual information through contrastive learning and mask-reconstruction mechanisms. The framework’s model-agnostic design allows integration with various CDM architectures, demonstrating consistent performance improvements across multiple real-world datasets.

Empirical results show significant enhancements in diagnostic accuracy (measured by AUC, ACC, and RMSE), particularly in cold-start scenarios where traditional CDMs typically struggle. The framework’s effectiveness is further validated through ablation studies, visualization of embedding spaces, and comparisons across different LLM implementations. While introducing additional computational overhead, the benefits of enriched semantic knowledge integration justify the costs, especially for large-scale educational applications.

This research contributes to intelligent education systems by: (1) establishing a principled approach for combining LLMs’ conceptual understanding with CDMs’ interaction modeling, (2) introducing SOLO taxonomy for structured cognitive assessment, (3) developing novel alignment techniques between semantic and behavioral spaces, and (4) demonstrating practical improvements across diverse educational domains. The work opens new directions for developing more adaptive and explainable cognitive diagnosis systems that better emulate human educators’ reasoning processes.
DOI:10.1007/s44366-025-0057-8