Automatic Lesson Plan Generation via Large Language Models with Self-critique Prompting

In this paper, we utilize the understanding and generative abilities of large language models (LLMs) to automatically produce customized lesson plans. This addresses the common challenge where conventional plans may not sufficiently meet the distinct requirements of various teaching contexts and student populations. We propose a novel three-stage process, that encompasses the gradual generation of each key component of the lesson plan using Retrieval-Augmented Generation (RAG), self-critique by the LLMs, and subsequent refinement. We generate math lesson plans for grades 2 to 5 at the elementary school levels, covering over 80 topics using this method. Three experienced educators were invited to develop comprehensive lesson plan evaluation criteria, which are then used to benchmark our LLM-generated lesson plans against actual lesson plans on the same topics. Three evaluators assess the quality, relevance, and applicability of the plans. The results of the evaluation indicate that our approach can generate high-quality lesson plans. This innovative approach can significantly streamline the process of lesson planning and reduce the burden on educators.