In a new preprint, researchers investigate whether the modular organization of the human brain—with distinct networks for language, formal reasoning, social reasoning, and physical reasoning—is a fundamental principle of intelligent systems or an evolutionary accident. They test this by analyzing large language models (LLMs), which are built through a very different optimization process: gradient descent rather than biological evolution.
Using circuit analyses across 46 tasks spanning four cognitive domains, the team finds that LLMs develop a modular architecture that mirrors the human brain. Tasks that recruit the same network in humans tend to activate overlapping neurons in LLMs, while tasks drawing on different networks activate distinct neurons.
The convergent emergence of modularity in both brains and neural networks suggests that modularity may be a fundamental property of intelligent systems, regardless of how they are created. The paper is available on Hugging Face as preprint 2608.13567.