Aligning generative AI with hierarchical K-12 curricula: a RAG and multi-agent framework for EFL content generation

While large language models (LLMs) like GPTs show strong problem-solving capabilities, their application in K-12 education is limited by the hierarchical nature of school curricula, which pre-trained models often fail to capture. This study proposes a Contextualized Content Generation Framework (CCGF) that integrates Retrieval-Augmented Generation (RAG) and a Multi-Agent System (MAS) to enhance contextual relevance and reliability in English as a Foreign Language (EFL) content creation. The framework assigns three agents’ distinct roles: knowledge localization, question generation, and result verification. A CCGF-based prototype was used to generate middle school EFL exercises and evaluated by 18 teachers through a blind evaluation of 60 questions from GPTs, manual sources, and CCGF. Results indicated that CCGF outperformed GPTs and was comparable to teacher-crafted items. Further validation with 139 students using Item Response Theory showed no significant differences between CCGF-generated and human-made assessments. These findings support the framework’s validity and usability in real classroom settings and suggest its potential to enhance generative AI applications in interactive K-12 learning environments.