Designing an LLM-Based Dialogue Tutoring System for Novice Programming

Abstract

In this paper, we present the design of a dialogue-based tutoring system called DT4-Coding to help novice programmers understand their misconceptions about programming concepts and work towards correcting them. The error or misconception detection and correction are based on Ohlsson's Theory of Learning from Performance Error, which suggests that a mentor (in this case, a software tutor) is necessary to help learners detect these errors and provide feedback to repair faulty knowledge. We utilized the constraint-based model (CBM) for domain and knowledge modeling, which allows human experts to represent knowledge as a constraint. These constraints were used to assess student answers and to generate LLM-based feedback via a question- answer form of dialogue. By incorporating human-written domain knowledge, the system could potentially reduce the problem of LLMs in generating solutions that may be factually incorrect.