Thesis Defense: Determination of output composition in reaction-advection-diffusion systems and a new method to improve language model performance

Abstract: We present 2 different subjects. The first is on modeling chemical reactions. We formulate the problem of determining how much product is formed from a reactor and model this problem using metric graphs. We develop an efficient method to explicitly solve the problem on metric graphs. We work through examples by hand and with code to solve the problem. The second subject is a novel method to improve language model performance on compositional tasks. Our method teaches a language model to break a given problem into different subproblems, create prompts for these, and then solve the subproblems in separate contexts. The model uses the solutions of the subproblems to return the solution to the original problem given to the model. This method improves performance on 3 common tasks in the length generalization literature.

Advisor: Renato Feres