Puzzles are challenging, multimodal tasks that require interpreting ambiguous hints, such as text, visuals, and tabular data, without any explicit instructions. This makes them ideal for training large-language models to tackle complex, open-ended tasks. This SuperUROP project aims to develop a system of multimodal reasoning agents to collaboratively solve these types of puzzles. Ultimately, the hope is to contribute towards models that not only excel at puzzle-solving, but also demonstrate improved performance on downstream tasks involving logical and mathematical reasoning.

I am participating in SuperUROP because I am excited to explore cutting-edge research on multimodal AI and on enhancing LLM reasoning capabilities. Through this program, I can deepen my understanding of these topics through an extended research experience, applying technical skills from my ML coursework and industry experience.