When: Wednesday, October 21, 2026
Time: 3:00 - 3:50 pm
Where: RB353
Diagnosing Perception Failures in Multimodal Mathematical Reasoning
Adithya Kulkarni - Assistant Professor of Computer Science
Abstract: AI models can now solve math problems that include both text and diagrams. But when a model gets an answer wrong, it is not always clear what went wrong. Did the model misunderstand the diagram? Did it miss an important detail in the text? Or did it understand the problem correctly but make a mistake while solving it? In this talk, I will present our work on separating these different types of errors. Using geometry problems, we study what AI models actually understand from the text and images before they begin solving the problem. Our results show that a model can understand the information correctly and still reason incorrectly, while another model may misunderstand part of the problem but sometimes still reach the correct answer. The goal of this work is to better understand how multimodal AI systems fail and how we can evaluate them more carefully.
Bio: Adithya Kulkarni is an Assistant Professor of Computer Science at Ball State University. He received his Ph.D. in Computer Science from Iowa State University and later worked as a Postdoctoral Research Associate at Virginia Tech's Sanghani Center for Artificial Intelligence and Data Analytics. At Virginia Tech, his research focused on areas such as large language models, graph learning, reliable AI, and using AI for scientific discovery. His current research explores how AI systems reason, why they make mistakes, and how they can be made more reliable. He also works on interdisciplinary applications of AI, including the Cohen Peace Studies Grant project "Reasoning for Peace," which explores how AI can support conflict prevention and transformation. He enjoys involving undergraduate students in research and helping them see how ideas learned in computer science courses can be applied to real research problems.
Light Refreshments will be Served