Educational vision-language material

Generating reliable visual material for childhood education.

As a graduate research assistant at the Center for Unified Biometrics and Sensors, I have been building a proof of concept for vision-language educational material with the AI4ExceptionalEd Institute.

The work uses large language models, vision-language models, and text-to-image models, together with prompting strategies aimed at material that is factually accurate, reliable, and appropriate for children. Two research questions in that setting are how to generate complex scenes that contain many compositional elements, and how to evaluate hallucination in text-to-image generation systematically.

This project does not yet have a public paper.