This study investigates methods to enhance large language models (LLMs) with robust capabilities for visual reasoning in the context of education. While LLMs demonstrate strong performance on symbolic reasoning tasks, their inability to reliably interpret diagrams in domains such as geometry and calculus often results in erroneous explanations, thereby undermining their effectiveness as tutoring systems. Typical errors include misidentifying geometric relationships—for instance, inferring intersections between disjoint circles or misclassifying the location of a labelled point relative to a region. To address these limitations, we aim to propose a framework for augmenting LLMs to integrate visual understanding with formal reasoning, with the objective to establish an AI tutoring system capable of delivering precise, visually grounded feedback that supports learning while minimizing instructional error.

I’m excited to pursue this SuperUROP in order to develop my research skills, and to apply what I’ve learned to further the field of artificial intelligence and education.