Search and Filter

Edumath: Generating Standards-Aligned Educational Math Word Problems

Date

AI Overview

Why it matters

Personalizing math word problems supports student learning, but teachers struggle to find time to customize them given large classes and burnout. Current language models often fail to generate problems aligned with educational standards and readable solutions.

Study at a glance
  • Question: Can language models generate educational math word problems tailored to grade school standards and student interests?
  • Setting and data: Two US elementary schools, 94 students in grades 3-5, 1,372 US teachers, and over 11,000 generated math word problems.
  • Method: Dataset creation, fine-tuning, classifier filtering, automated model evaluation, and student worksheet experiments.
  • AI tools: EDUMATH 12B, EDUMATH 30B, Gemma 3 12B IT, Gemma 3 27B IT, GPT-4o, Llama 3.3 70B IT
Key findings
  • The EDUMATH 30B model met all educational criteria on 94.6% of generated problems, outperforming closed models including GPT-4o, GPT-4.1, and GPT-4.5.
  • Grade school students solved AI-generated problems at similar rates to human-written problems, 70.6% versus 68.2% in one school and 68.4% versus 73.7% in another.
  • Students significantly preferred customized AI-generated problems over human-written problems across both school settings, most often because they liked the problem topic.

Summary drafted with AI from the full paper and may contain errors. See the full paper for details.

Math word problems (MWPs) are critical K-12 educational tools, and customizing them to students' interests and ability levels can increase learning outcomes. However, teachers struggle to find time to customize MWPs for each student given large class sizes and increasing burnout. We propose that LLMs can support math education by generating MWPs customized to student interests and math education standards. To this end, we use a joint human expert-LLM judge approach to evaluate over 11,000 MWPs generated by open and closed LLMs and develop the first teacher-annotated dataset for standards-aligned educational MWP generation. We show the value of our data by using it to train a 12B open model that matches the performance of larger and more capable open models. We also use our teacher-annotated data to train a text classifier that enables a 30B open LLM to outperform existing closed baselines without any training. Next, we show our models' MWPs are more similar to human-written MWPs than those from existing models. We conclude by conducting the first study of customized LLM-generated MWPs with grade school students, finding they perform similarly on our models' MWPs relative to human-written MWPs but consistently prefer our customized MWPs.