MIT study: LLMs develop brain-like modular architecture for reasoning

A new arXiv paper (arXiv:2608.13567) from Pengrui Han, Jacob Andreas, Evelina Fedorenko, and Andrea Gregor de Varda reveals that large language models (LLMs) spontaneously develop a modular cognitive architecture strikingly similar to the human brain. Using circuit analyses across N=46 tasks spanning language, formal reasoning, social reasoning, and physical reasoning, the team found that tasks engaging the same functional network in humans activate overlapping sets of neurons in LLMs, while tasks drawing on distinct networks engage separate neuronal populations. This functional specialization emerged purely from next-token prediction—no architectural constraints were imposed, yet the LLM independently recreated the brain's division of labor.

The findings have profound implications for AI research and neuroscience. The convergent evolution of modularity in biological brains and artificial neural networks suggests that organizing intelligence into specialized processing modules may be a universal principle—not an evolutionary quirk. For AI developers, this points toward architectures that explicitly separate reasoning domains, potentially improving interpretability and efficiency. For neuroscientists, LLMs can serve as testable models of how modular brain networks interact. The authors caution that replicating human brain organization does not guarantee human-like cognition, but the parallel offers a powerful framework for studying intelligence across both silicon and biological substrates.