TeachMateGPT automates science test creation with AI agents

A team of researchers from multiple institutions has developed **TeachMateGPT**, a multi-agent AI framework designed to automate the generation of science assessment items from curriculum materials. Published on arXiv (arXiv:2608.13708), this system addresses key limitations in traditional retrieval-augmented generation (RAG) approaches by introducing a hierarchical knowledge base called COPE, which organizes textbook content into a graph-based structure aligned with syllabus progression.

The framework replaces one-shot RAG pipelines with a staged agent system that includes routing gates, evidence fusion, and specialized agents for generating both objective and constructed-response questions. A source-attributed verification protocol (SAVER) further ensures question quality by scoring faithfulness, relevance, and hallucination risk. In evaluations using the NCTB-SciGen8 dataset (198 questions spanning 14 textbook chapters), TeachMateGPT achieved 96% faithfulness and 89% answer relevancy compared to 68% and 60% for vanilla RAG baselines.