Agentic AI
Multi-agent system learns to generate personalized math questions with controlled variety
Researchers developed a collaborative AI framework that creates tailored mathematics questions while meeting specific educational objectives and maintaining diverse problem types.
A research team has introduced EduAgentQG, a multi-agent framework that automates the creation of personalized mathematics questions for educational assessment and adaptive learning systems. The approach addresses a longstanding challenge in AI-assisted education: balancing alignment with learning objectives against the need for variety in problem types and difficulty levels.
How the system works#
Rather than relying on single-agent generation or unguided randomness, EduAgentQG orchestrates multiple specialized agents in a structured loop. The framework guides question creation through four main stages: planning develops generation strategies and outlines, writing produces candidate questions following different directions, evaluation checks logical soundness and verifies alignment with learning targets, and refinement iterates to improve output quality.
The evaluation step is particularly granular, assessing questions across multiple dimensions—including knowledge concepts, difficulty levels, grade appropriateness, and core competencies—rather than applying generic feedback. This fine-grained approach lets the system verify that generated questions genuinely match the intended educational goals.
Benchmark and evaluation framework#
The team constructed a mathematics benchmark comprising over 10,000 questions spanning Grades 1–9, with coverage across 634 distinct knowledge concepts and 16 core competencies at three difficulty tiers. Two specialized evaluation subsets support different question formats: MathChoice covers 489 objectives for multiple-choice items, and MathBlank addresses 500 objectives for fill-in-the-blank problems.
Results and comparison#
Experimental results demonstrate that EduAgentQG outperforms several competing approaches—including chain-of-thought prompting, multi-shot chain-of-thought, ReAct, and an earlier method called EQPR—across three key metrics: diversity of generated questions, alignment with stated educational objectives, and overall quality as judged by human evaluators.
The work suggests that explicitly orchestrating multiple agents with structured evaluation can improve both the pedagogical relevance and creative variety of machine-generated educational content.
