Haein Jeon , Bo-Yeong Kang
Expert Systems with Applications, 2026
Intelligent tutoring systems (ITSs) have gained increasing attention for their potential to support individualized instruction. While human tutors naturally adapt their teaching strategies by considering both cognitive and non-cognitive traits such as mastery level and personality, most existing ITSs lack mechanisms to model these individual differences effectively. This study proposes a framework for pedagogically adaptive tutoring} that uses a large language model to infer students’ mastery level and Big Five-related behavioral tendencies from dialogue history. The inferred traits guide teaching strategy selection through a rule-based pedagogical engine integrated into response generation. We evaluate the framework across six generative models including open-weight and proprietary systems using two public language-learning dialogue datasets (TSCC and CIMA). Performance is assessed through automatic metrics capturing lexical overlap (ROUGE-L), semantic similarity (BERTScore), and human-like engagement (DialogRPT), as well as a human evaluation study.Experimental results indicate consistent gains across models and datasets, with up to 6.7% relative improvement on automatic metrics and improved human-rated instructional quality.These findings suggest that integrating dialogue-based student modeling into response generation enhances the response-level instructional quality of automated tutoring.