Study finds modular cognitive architecture emerging in large language models across 46 tasks

A new arXiv preprint reports that large language models (LLMs) develop modular cognitive architectures that resemble those observed in the human brain. Using circuit analyses across 46 tasks spanning four cognitive domains—language, formal reasoning, social reasoning, and physical reasoning—the authors find that tasks recruiting the same functional networks in humans also recruit overlapping neurons in LLMs. Conversely, tasks drawing on distinct human networks recruit distinct neurons in the models.

The study, titled 'Modular Cognitive Architecture Emerges in Large Language Models,' was authored by Pengrui Han, Jacob Andreas, Evelina Fedorenko, and Andrea Gregor de Varda. It was submitted to arXiv on June 27, 2026, and is currently listed under the Artificial Intelligence (cs.AI), Computation and Language (cs.CL), and Machine Learning (cs.LG) categories.

The authors argue that the convergent emergence of modularity in both biological brains and artificial neural networks suggests modularity may be a fundamental property of intelligent systems, rather than an evolutionary accident specific to biology. Their analysis is based on circuit-level examinations of model activations during task performance, comparing patterns of neuron recruitment across domains.

The paper does not claim that LLMs replicate human cognition or that their internal organization is identical to the brain. Instead, it presents evidence that both systems exhibit functional specialization at the level of task-relevant neuron groups, even though they arise from vastly different optimization processes.