My project investigates how large language models (LLMs) lose critical technical information during information synthesis, and how these gaps vary across different scientific fields. By categorizing the errors, we identify consistent challenges in handling theoretical foundations and performance evaluation, with computer science showing the most significant decline in accuracy. The findings emphasize fundamental limitations in LLMs’ ability to process and retain critical scientific knowledge, offering a foundation for targeted improvements in AI-assisted scientific analysis.

I am participating in Super UROP to connect my main interest areas AI, mathematics and finance through high-level research. This is a great opportunity for me to do work in the intersection of my majors Business (15-2) and AI (6-4), while learning more about financial markets, and how to work with LLMs. I am also interested in LLM capability research, particularly how their use in everyday life influences larger systems and drives broader societal changes.