Large Language Models in Physics Education

Chatbots have made large language models widely accessible to non-technical users. The consequences of this have been felt especially heavily in education. In this talk, I will first summarize the working principles of LLMs on a very basic level and make a case that even a basic understanding of how LLMs work can help physics educators appreciate their strenghts and weaknesses, and help them make decisions about their appropriate uses during the education process. Second, I will present some findings from our own research on the capabilities of vision-capable large multimodal models on tasks that involve the interpretation of commonly used physics representations other than language (e.g. graphs, sketches, diagrams). These findings can inform both researchers and educators in physics about the strengths and limitations of some of the widely used vision-capable models. Third, I will discuss one way in which we can take advantage of LLMs' incorrect outputs for the purpose of physics teacher training through exercises involving a Socratic-style dialogue between a teacher and an LLM-based chatbot.