Scaling High-Leverage Curriculum Scaffolding in Middle-School Mathematics

Despite well-designed curriculum materials, teachers often face significant challenges in their implementation due to the diverse learning needs present in classrooms. This paper examines whether and how Large Language Models (LLMs) can be leveraged to enhance K-12 math education by facilitating the creation of high-quality curriculum scaffolds that reflect expert teachers' strategies. Through an in-depth qualitative analysis with experienced middle-school math teachers, we identified crucial instructional supplements such as warm-up tasks and example-problem pairs that are essential for engaging students and supporting diverse learner needs. Building on these insights, we developed ScaffGen, an LLM-powered tool designed to generate curriculum-aligned educational materials. While LLMs alone may fall short in educational contexts, when enhanced with expert teacher insights, they can effectively emulate the cognitive processes required for pedagogically robust material creation. We plan to assess the effectiveness of these AI-generated materials through rigorous evaluations involving comparisons with expert-written benchmarks and field tests in classroom settings. This research highlights the potential of LLMs to mimic expert decision-making in educational material creation, offering significant implications for scalable instructional support.