LessonBench-V1: First Standardized Benchmark for AI Lesson Generation in STEM

As large language models increasingly power AI educational content generation, researchers have lacked a standardized way to benchmark their lesson-planning capabilities. To fill this gap, Ravidu Suien Rammuni Silva and colleagues from Nottingham Trent University and other institutions created LessonBench-V1. This dataset comprises 647 human-written lessons reverse-engineered into lesson plans using LLMs, covering 240 topics across four STEM disciplines: mathematics, physics, chemistry, and computer science. The lessons are drawn from 97 trusted open sources including LibreTexts, arXiv, and GeeksForGeeks. Each lesson plan was human-reviewed and built on a pedagogically grounded methodology that synthesizes Bloom's Taxonomy, Gagné's Nine Events of Instruction, Merrill's First Principles of Instruction, and the 5E Instructional Model (Engage, Explore, Explain, Elaborate, Evaluate).

The dataset captures 3,620 learning objectives, each annotated with pedagogical metadata such as cognitive level, instructional event, and prerequisite relationships. This metadata allows for granular, reproducible evaluation of AI agents that generate lessons. The study also proposes a three-dimensional evaluation pipeline: (1) content accuracy and coverage, (2) pedagogical alignment with established frameworks, and (3) coherence and flow of the generated lesson. By providing a common benchmark, LessonBench-V1 enables researchers and edtech companies to compare different AI lesson generators systematically, moving beyond anecdotal or task-specific evaluations. The dataset and evaluation methodology are publicly available on arXiv, aiming to catalyze further research in AI-assisted education.