CLS Andrea de Varda (
https://andrea-de-varda.github.io/), a postdoctoral researcher in the Department of Brain and Cognitive Sciences at the Massachusetts Institute of Technology (MIT), working with Evelina Fedorenko and Roger Levy.
-----------------------------------------------------------------------------
Title: Large Language Models as models of human language(s) and higher-level cognition
Abstract: Large language models (LLMs) have recently emerged as powerful candidates for modeling several domains of human cognition. Because they operate over natural language, they provide flexible representations that can be evaluated against human behavior and brain activity. In this talk, I will present a set of studies that use LLMs to test how far this modeling approach can go—first in the domain of language, and then in higher-level reasoning.
In the first part, I ask whether multilingual language models can explain how the human brain processes the extraordinary diversity of the world's languages. Using fMRI data from native speakers of 21 languages spanning 7 language families, we show that model embeddings reliably predict brain responses within languages and, crucially, transfer zero-shot across languages and families. These results point to a shared representational component in the human language network, largely driven by semantic content, that aligns with the representations learned by multilingual models.
In the second part, I move beyond language to ask whether LLMs can serve as models of human reasoning, from two angles.
First, the brain shows striking functional specialization, with distinct networks for language, formal reasoning, social reasoning, and physical reasoning. Is this modular organization a general principle of intelligent systems, or an accident of biological evolution? Using circuit analyses across 46 tasks in these four domains, we show that LLMs develop a modular architecture mirroring the brain. This convergence suggests modularity is a general principle of intelligence.
Second, analyzing large reasoning models, we show that the number of reasoning steps they take predicts human reaction times across seven diverse tasks. This holds both within tasks, reflecting item difficulty, and across tasks, capturing broad differences in cognitive demand.