Year: 2026
Type: dissertation
Abstract: Large Language Models exhibit increasingly sophisticated behaviors, yet we lack systematic methods to evaluate their true reasoning capabilities.Current approaches-such as collecting human-generated data-cannot reliably predict model failure modes, distinguish genuine reasoning from complex memorization patterns, nor guarantee synthetic data quality.This thesis explores the idea that imbuing tasks with explicit structure provides the control and verification mechanisms necessary for more principled evaluation and reliable data generation.Structure unlocks optimization and search algorithms that can systematically discover model flaws while supporting both higher-quality training data generation and I am deeply grateful to my advisors Yulia Tsvetkov and Yejin Choi for guiding me throughout my PhD, for taking a chance on me when I had extremely limited research experience, and for their amazing and always complementary research insights.To Yulia, thank you for encouraging me to believe in myself and for giving me the freedom to pursue the directions I believed in most.You offered guidance along the way while trusting me with the space to think critically and grow into an independent researcher-the best result of that is what became Part I of this thesis.Your warmth and care made me feel like more than just a student in your lab, and you taught me that academic excellence and personal well-being can go hand in hand.To Yejin, thank you for teaching me by example to keep dreaming bigger, to find new horizons, and to rapidly adapt to changing times.Thank you for teaching me to be stronger and communicate my research more clearly and compellingly.I have often found your advice in my head when I write any kind of text and when I present my work.As we will cover in Part III, having a robust theory of mind is important, and you taught me to sharpen mine.To my committee members Luke Zettlemoyer and Max Kleiman-Weiner, thank you for your thoughtful questions and support throughout the defense process.
Author: Sclar, Melanie
Institution: University of Washington
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7213483608