LessonAgent: A Multimodal Pipeline for Automated Generation of Lesson Plans, Presentations, and Podcasts

Lesson preparation plays a crucial role in structuring and organizing the teaching process. However, traditional lesson design and presentation creation require teachers to spend a considerable amount of time reviewing the literature and organizing materials. Therefore, developing an intelligent and multimodal technology capable of automatically generating lesson materials holds great significance. Such technology can potentially reduce teachers’ workloads and improve the efficiency and quality of lesson preparation, as indicated by teacher satisfaction and preference judgments. In this paper, we introduce LessonAgent, a multimodal and interactive pipeline that leverages large language models (LLMs) to generate lesson plans, presentations, and podcasts. Our system enhances the quality of generated materials through diverse input modalities, refined generation mechanisms, and interactive feedback with teachers. Specifically, we present the Plan10k dataset—a high-quality bilingual collection of lesson plans—and employ it to train and evaluate our framework. The pipeline consists of three main modules: a query rewriting module that handles multimodal teacher inputs (e.g., textual concepts, images, or textbook excerpts), a lesson plan generation module that produces structured content, and a chapter correction module that integrates retrieval-based tools to improve factual accuracy and contextual relevance. Furthermore, teachers can interact with intermediate results, allowing adaptive refinement throughout the generation process. Based on the generated lesson plans, the framework further produces corresponding visual presentations and podcasts, forming a comprehensive multimodal teaching assistant system. Extensive experiments and teacher evaluations demonstrate the superior performance and satisfaction of our approach.