Can AI Tutors Know When to Help and When to Hold Back?
Imagine you're a student struggling with a math problem. A good tutor wouldn't immediately give you the answer, but instead, ask questions to help you understand the concept and guide you to find the solution yourself. But what about AI tutors? Can they balance the delicate trade-off between helping and holding back?
Researchers at AllenAI have developed a framework called TutorMoments to evaluate whether cutting-edge language models can make this crucial decision. The results show that, while AI tutors tend to over-help by providing too much support, spelling out the trade-off in the prompt can improve performance. However, they still differ widely in how reliably they make this call, and human tutors consistently outperform them.
The key to effective tutoring is not just about providing answers, but about adapting to each student's needs. As AI tutors become more prevalent, it's essential to develop models that can make pedagogical decisions that matter most. By releasing a dataset of de-identified tutoring transcripts, code, and model tutor replays, the researchers hope to encourage the development of AI tutors that can truly support student learning.