Pipeline for Training Simulated Student Models with Reduced Human Data

Team: Tammy Kwan and Brenden Lake

Overview: This project develops a training pipeline for neural-network-based simulated student models. The approach combines synthetic reasoning traces generated by rule-based student models with human student reasoning data to substantially reduce the amount of human data required.

Subject Focus: Fraction arithmetic within grades 6-8, with exploratory transfer to decimals, percentages, and early algebra.

Targeted Universalism Focus: This project responds to the reality that high-quality, individualized diagnostic attention in mathematics is unevenly distributed, frequently leaving students in under-resourced classrooms without consistent formative feedback due to large class sizes, limited staffing, or high learner variability.

Public Goods & Deliverables:

  • An open-source codebase (released via GitHub under the Apache 2.0 license) detailing workflows for dual synthetic-human optimization.
  • Fine-tuned neural weights that replicate empirically observed human accuracy, error distributions, and individual student response patterns in fraction arithmetic.
  • Implementation Guidance: Technical reports, metadata schemas, and adaptation guidelines to help learning engineers extend the pipeline to adjacent mathematics domains.

Return to see all grantees.