Nora Dunphy Rose Hills
Uncovering Structured Intent Representations in Language Models
Large Language Models (LLMs) like ChatGPT excel at generating human-like responses by predicting the next word in a sequence based on statistical patterns in their training data. While effective, this approach does not fully align with how humans use language. Human communication is not merely predictive—it is goal-oriented and intent-driven. When we speak, we do so with purpose, whether giving a command, asking a question, or engaging in small talk to build rapport. This project aims to explore how LLMs represent this concept of speaker intent without relying on human-labeled data. By using techniques like Task Vectors and variational modeling, I will investigate whether intent representations are already encoded within pretrained language models and how these structures can be extracted and evaluated. This project will examine whether these latent representations can be used to improve model interpretability, controllability, and large-scale text analysis. The broader goal is to align language models more closely with human communication by uncovering and leveraging the intent-driven structures behind everyday language use.
Message To Sponsor
Thank you for your generous support of my summer research. This opportunity allows me to explore how large language models represent speaker intent, combining my interests in linguistics and machine learning. I’m truly grateful for the opportunity to deepen my research experience and grow as an interdisciplinary thinker.