AI question generation: from objectives to item-level checks
Our question-bank research prototype plans, generates, checks and revises. Papers inform the method; teacher review determines whether an item belongs in a lesson.
Turn teaching objectives into an inspectable plan
Our question-bank research prototype turns teaching objectives into an item plan: grade, knowledge components, cognitive level and variation. For textbook-based generation, retrieved passages let teachers check how each item relates to the material.
The prototype draws on EQPR’s planning, evaluation and revision cycle and EDUMATH’s standards alignment and readable-solution criteria. Its bounded planning and retries are not the full search procedure described in EQPR.
Check answers and pedagogical quality after generation
Each item passes through generation, supported symbolic checks, verification and pedagogical review. The verifier considers correctness, grade fit, grounding and whether the target knowledge components are actually needed to solve the item—a concern informed by KT4EQG.
Multiple-choice items also need meaningful distractors. LookAlike informs the link between distractors and plausible misconceptions, rather than merely different numbers. Symbolic tools have limits, and a verifier role does not provide an independent guarantee of correctness.
From research evaluation to classroom use
This planning loop is currently a research prototype. Version records, bounded revisions and human review help us assess whether each revision actually improves an item.
The main product also has a worksheet draft-and-reflection flow and answer acceptance checks comparing outputs from different models. These support inspectable generation, but each workflow needs its own evaluation. Teachers should review materials for the learners they teach.
References
- EQPR · From Objectives to Questions: A Planning-based Framework for Educational Mathematical Question Generation
- EDUMATH · Generating Standards-aligned Educational Math Word Problems
- KT4EQG · Personalized Exercise Question Generation via Knowledge Tracing
- LookAlike · Consistent Distractor Generation in Math MCQs