“I Am Confused! How to Differentiate Between…?” Adaptive Follow-Up Questions Facilitate Tutor Learning with Effective Time-On-Task

Within the learning-by-teaching paradigm, students, who we refer as tutors, often tend to dictate what they know or what to do rather than reflecting on their knowledge when assisting a teachable agent (TA). It is vital to explore more effective ways of fostering tutor reflection and enhancing the learning experience. While TAs can employ static follow-up questions, such as “Can you clarify or explain more in detail?” to encourage reflective thinking, the question arises: Can Large Language Models (LLMs) generate more adaptive and contextually-driven questions to deepen tutor engagement and facilitate their learning process? In this paper, we propose ExpectAdapt, a novel questioning framework for the TA using three stacked LLMs to promote reflective thinking in tutors, thereby, facilitating tutor learning. ExpectAdapt generates adaptive follow-up questions by directing tutors towards an expected response based on the tutor’s contributions using conversation history as a contextual guide. Our empirical study with 42 middle-school students demonstrates that adaptive follow-up questions facilitated tutor learning by effectively increasing problem-solving accuracy in the learning-by-teaching environment when compared to tutors answering the static follow-up questions and no follow-up questions at all.