TeachMateGPT: Multi-Agent Framework for Generating Science Assessments
TeachMateGPT is an innovative multi-agent framework designed to enhance the automated creation of assessment items based on textbooks for science education. This system, outlined in a paper on arXiv (2608.13708), tackles the shortcomings of current retrieval-augmented generation (RAG) systems. It introduces four significant improvements: COPE, a structured knowledge base; a staged, fail-closed agent pipeline; along with features for generating multiple questions and aligning with curricula. By automating the development of assessment items, it alleviates the burden on science educators. The paper is cross-listed, suggesting it has been submitted to various arXiv categories, and is pertinent to both artificial intelligence and education, particularly in low-resource environments with board-exam frameworks.
Key facts
- TeachMateGPT is a multi-agent framework for generating science assessment items.
- It addresses limitations in existing RAG systems: flat retrieval, single-question generation, lack of evidence safeguards, and unsuitability for low-resource curricula.
- COPE is a hierarchical knowledge base that segments documents along syllabus structure and links them at three granularities via a graph-based lineage.
- The system uses a staged, fail-closed agent pipeline with routing gates and a coverage gate that withholds generation when evidence is weak.
- The paper is available on arXiv with ID 2608.13708.
- The framework aims to reduce science teachers' workload.
- It is designed for board-exam-structured curricula in low-resource settings.
- The announcement type is 'cross', indicating multiple arXiv categories.
Entities
Institutions
- arXiv