Generate-Then-Validate: Question Generation for Education

Synthesis: A novel generate-then-validate pipeline for educational question generation that reduces LLM hallucination by 62% compared to direct generation, validated on STEM datasets with 89% accuracy and a 23% improvement over baseline LLMs on relevance metrics. The two-stage approach first generates candidate questions, then validates them against domain constraints and pedagogical criteria.

Approach

The paper introduces a two-stage pipeline for automated educational question generation:

1. Generate — an LLM produces candidate questions based on source material and specified learning objectives

2. Validate — a separate validation module checks each candidate against domain constraints, factual accuracy, and pedagogical quality criteria

This architecture addresses a core limitation of direct generation: LLMs produce plausible-sounding but factually incorrect or pedagogically inappropriate questions at high rates. The validation stage acts as a quality filter, discarding or flagging candidates that fail domain-specific checks.

Key Findings

62% reduction in hallucination compared to direct LLM generation

89% accuracy on STEM datasets (physics, chemistry, biology)

23% improvement over baseline LLMs on relevance and pedagogical alignment metrics

The validate stage catches factual errors, inappropriate difficulty levels, and misaligned learning objectives

Significance

Automated question generation reduces manual authoring burden for educators and enables adaptive assessment at scale. The generate-then-validate approach is particularly relevant for STEM domains where factual precision is critical and hallucinated content can mislead learners. This work connects to the broader Automated Question Generation and Automated Assessment literature.

Connected Concepts

Automated Question Generation

Automated Assessment

Connected Articles

Code Gen — CODE-GEN: A Human-in-the-Loop RAG-Based Agentic AI System for Multiple-Choice Question Generation

Kt4eqg Personalized Question Generation — KT4EQG: Personalized Exercise Question Generation via Knowledge Tracing

Nsmq Riddles Science Math Benchmark — NSMQ Riddles: A Benchmark of Scientific and Mathematical Riddles for Quizzing Large Language Models

Deeptutor — DeepTutor: Towards Agentic Personalized Tutoring

Multimodal Item Parameter Estimation 2026 — Multimodal Item Parameter Estimation using Simulated Response Probabilities

Slidesqaqa Pedagogical Question Generation — Slide Deck Q&A Quality Assurance App: A Multi-Stage Pipeline for Pedagogical Question Generation

Citation

F, A.W.Y.S.J.C.P. (2026). Generate-Then-Validate: Question Generation for Education. (LAK 2026), April 27-May 01 Wei, Y., Stamper, J., et al. (2025). Generate-Then-Validate: A Novel Question Generation Approach. arXiv preprint arXiv:2512.10110.