Abstract:

Research on Intelligent Tutoring Systems (ITS) has consistently shown that effective tutoring depends less on correctness of answers and more on pedagogical structure, scaffolded guidance, and formative feedback that support learner understanding. However, recent developments in artificial intelligence tutors based on large language models (LLMs) have primarily focused on improving language generation capabilities, while the pedagogical structure of the tutoring interaction often remains underdeveloped. As a result, many AI tutors provide fluent explanations but lack instructional coherence and meaningful formative feedback.

This study investigates how explicitly embedding Instructional Design (ID) principles into AI-based mathematics tutoring systems affects pedagogical quality and tutoring behavior. Instructional design refers to structured pedagogical frameworks that guide the organization of explanations, feedback, and learning activities to support understanding. Grounded in contemporary approaches to scaffolding and feedback design, this study explores how AI tutors can operationalize these pedagogical principles.

We propose two complementary strategies for integrating instructional design into tutoring systems powered by large language models. The first approach applies instructional design–guided prompting at inference time, structuring how the model generates tutoring responses. The second approach embeds instructional design principles directly during model training through parameter-efficient fine-tuning of an open-source language model. Both strategies aim to promote core tutoring behaviors commonly associated with effective learning support, including activating prior knowledge, step-by-step scaffolding, formative micro-checks during problem-solving, and coherent instructional progression.

To evaluate tutoring effectiveness beyond simple answer correctness, we introduce a Bayesian evaluation framework that estimates the probability that responses generated by instructional design–enhanced tutors provide stronger instructional support than those from baseline tutors. The evaluation focuses on mathematics problem-solving tasks and simulated personalized tutoring dialogues.

Results show that instructional design–guided prompting produces modest but consistent improvements in tutoring quality. In contrast, embedding instructional design principles during model fine-tuning leads to more stable and pedagogically consistent tutoring behavior, reducing variability in explanation quality and improving the structure of the learning interaction. These findings suggest that integrating instructional design principles into AI tutors is a promising strategy for improving the pedagogical reliability of large-language-model–based tutoring systems.

Keywords:

Intelligent Tutoring Systems, Instructional Design, AI Tutors, Large Language Models, Personalized Learning, Mathematics Education.