Abstract: As the capabilities of LLMs have grown, so has interest in using them to understand how language works in the human brain. Some have even suggested that LLMs do or should mimic how language is divided in the human brain: Mahowald et al. (2024) propose that, just as the brain has separate areas for formal linguistic competence (morphology, syntax, lexical semantics) as opposed to functional linguistic competence (world knowledge, reasoning, pragmatics), LLMs should learn a similar division—and may have even already done so. How can we translate such hypotheses from brains to LLMs, and test them in the latter? To answer this question, I'll first introduce circuits, a causal framework that allows us to localize the abilities of LLMs to a small set of components. Next, I'll apply these techniques to the question of formal and functional linguistic competence, showing that LLMs may in part learn distinct mechanisms for the two. Finally, I'll discuss alternative patterns that might underlie the mechanistic organization of LLMs.
Bio: Michael Hanna is a third-year PhD student at the University of Amsterdam, advised by Sandro Pezzelle and Yonatan Belinkov. His research lies at the intersection of mechanistic interpretability and cognitive science, using causal techniques to uncover the low-level mechanisms that support LLMs' strong linguistic abilities. To this end, he has worked on not only developing new methods for interpreting LLMs, but also applying these techniques to linguistic topics such as incremental sentence processing. Michael is an ELLIS PhD student and is supported by an OpenAI Superalignment Fellowship.