I’m a CS PhD and a philosophy MA advised by Ellie Pavlick at Brown University. I study how developmental psychology and comparative cognition inform what types of behavioral and representational evaluations we should have for language models in order to measure to what extent they really understand language.

My current research focuses on evaluating language models’ understanding of prompts. I am a co-first-author of T0 (https://arxiv.org/abs/2110.08207), one of the first LLMs trained on hundreds of manually curated prompts that enable zero-shot generalization to novel tasks. Meanwhile, despite the substantial performance improvement of T0-style instruction tuning, I am also a critic of my own work, presenting my paper at this conference that even T0 and GPT-3 still fall far short of human understanding (https://arxiv.org/abs/2109.01247).