Can Large Language Models Recognize and Respond to Student Misconceptions?

Expert human tutors can observe learner mistakes to understand their misconceptions and procedural errors. Highly capable, but opaque large language models have shown remarkable abilities across numerous domains, and may be useful for adaptive instruction in a variety of ways. Working with publicly available data from the National Assessment of Educational Progress, (388 questions selected from 4th, 8th and 12th grade math and science) we examined these three questions: Discussion focuses on how these capabilities can be used for test design and adaptive instruction.