Ellie Pavlick
Language Understanding and Representation Lab
Brown University

Understanding Linguistic and Reasoning Mechanisms in Large Language Models

Large neural network language models have become the state of the art for processing language, but are typically considered to be “black boxes”, the underlying representations and mechanisms of which are inscrutable to humans. This inscrutability makes it difficult if not impossible to say if and how the success of modern AI can inform thinking about language in general (e.g., language in the abstract, or language processing in humans). In this talk, I will discuss new and mostly unpublished work which attempts to uncover pieces of the mechanisms that are in play under-the-hood when large language models engage in basic linguistic reasoning tasks. Taking these results together, I speculate on what the next few years of LLM research might look like, and how it could support the development of new hypotheses and theories about language processing more generally.