Scalable Human–AI Workflow for Middle School Math Question Creation

• In Event: AERA Poster Session Wednesday 11:45am
  In Poster Session: ATL Poster Session 1: AI Integration and Student Outcomes in K–12 Settings

Abstract

This study presents a human–AI pipeline including Generate, Filter for scalable creation of middle school mathematics multiple‑choice questions. GPT‑4 generated 5,634 items guided by curricular benchmarks and real‑world contexts. Expert educators labeled items as deployable, revision‑needed, or unsalvageable. A gradient‑boosted decision‑tree model trained on 376 text and pedagogical features achieved 92 % accuracy in removing unsalvageable items and 82 % accuracy in distinguishing deployable from revision‑needed questions (Agarwal & Mannem, 2011; Iqbal et al., 2023). Key metrics map to cognitive load management and concept coherence, guiding explainable authoring tools.

Authors

• Hai Li, University of Florida

• Rui Guo, University of Florida

• Chenglu Li, University of Utah

• Wanli Xing, University of Florida