Core Concepts
The authors propose using Large Language Models (LLMs) as teachable agents for Learning by Teaching (LBT), aiming to reduce the cost and barriers of building teachable agents and make LBT more engaging and effective.
Abstract
The study investigates the use of LLMs as teachable agents for programming education. It introduces AlgoBo, an LLM-based tutee chatbot, and TeachYou, an LBT environment for algorithm learning. The Reflect-Respond pipeline restrains LLMs' knowledge to initiate "why" and "how" questions for effective knowledge-building. Mode-shifting guides conversations through thought-provoking questions. The study confirms the effectiveness of the prompting pipeline in configuring AlgoBo's problem-solving performance. A user study with 40 algorithm novices shows that AlgoBo's questions led to knowledge-dense conversations. Design implications, cost-efficiency, and personalization of LLM-based teachable agents are discussed.
Stats
Our technical evaluation confirmed that our prompting pipeline can effectively configure AlgoBo’s problem-solving performance.
Through a between-subject study with 40 algorithm novices, we observed that AlgoBo’s questions led to knowledge-dense conversations (effect size=0.71).