Ying Xu, Ph.D.

Assistant Professor
Graduate School of Education
Harvard University

AI system interacting with a child during reading and math

In this series of projects, we explore ways to improve algorithms and system designs to enable AI models to better support student learning across subject areas such as math, science, and literacy. Our goals are to:

  1. understand how AI can personalize learning experiences
  2. compare the effectiveness of different AI techniques
  3. generate insights that inform both research and practice in AI-driven personalized learning

Highlights

  • AI-personalized story reading: Children, parents, and teachers in our study valued the personalized features of STORYMATE—an interactive tool powered by GPT-4. They described its conversations with children as “interesting,” “intelligent,” and “meaningful.” Drawing on our findings, we present critical reflections and design recommendations to inform future LLM-driven personalized storytelling technologies. [paper in CHI 2025]

  • Knowledge tracing for student learning: We introduce KT²—a tree-structured probabilistic framework for interpretable, online knowledge tracing. Starting from just a few exercise responses, KT² incrementally updates its estimates of a student’s knowledge state as learning progresses. By modeling learning across a structured knowledge concept tree, KT² can infer mastery not only of practiced concepts but also of related, unpracticed ones—providing real-time insights into areas where a student may be struggling and why. Available in arXiv and Submitted to ACL

  • QA dataset for real-world knowledge: We present StorySparkQA—a dataset of 5,868 expert-annotated QA pairs designed to bring real-world knowledge into interactive story reading. By capturing educators’ reasoning and leveraging knowledge graphs, this work supports AI systems that go beyond story content to promote deeper learning. Available in arXiv

Upcoming Studies

  1. A between-subject experiment involving two hundred children to investigate how different AI-led questioning strategies—constrained-response (multiple-choice) vs. open-ended dialogic questions—affect children’s comprehension of science-related informational texts.

  2. A study to examine how teachers evaluate the two AI-powered eBook systems by inviting them to use them firsthand and sharing their perspectives through interviews—focusing on educational value, usability, and classroom integration.

Publications

  • Gao, X., Wu, Q., Zhang, Y., Liu, X., Qian, K., Xu, Y., & Chang, S. (2025). A Hierarchical Probabilistic Framework for Incremental Knowledge Tracing in Classroom Settings. arXiv preprint arXiv:2506.09393.

  • Chen, J., Tang, M., Lu, Y., Yao, B., Fan, E., Ma, X., … & He, L. (2025, April). Characterizing LLM-Empowered Personalized Story Reading and Interaction for Children: Insights From Multi-Stakeholder Perspectives. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (pp. 1-24). [DOI]

  • Chen, J., Lu, Y., Zhang, S., Yao, B., Dong, Y., Xu, Y., … & Sun, Y. (2023). StorySparkQA: Expert-Annotated QA Pairs with Real-World Knowledge for Children’s Story-Based Learning. arXiv preprint arXiv:2311.09756.

Updated: