Modular Cognitive Architecture Emerges in Large Language Models
06:00 · August 17, 2026 · arXiv cs.AI RSS
The human brain exhibits a striking degree of functional specialization, with distinct networks supporting language, formal reasoning, reasoning about other minds, and reasoning about the physical world. Is this modular organization a fundamental principle of how intelligent systems must be built, or an evolutionary accident specific to biological brains? Here, we test whether a similar organization emerges in Large Language Models--another class of intelligent systems created through a very different optimization process. Using circuit analyses across N=46 tasks spanning four cognitive domains (language, formal reasoning, social reasoning, physical reasoning), we find that LLMs develop a modular architecture that mirrors the human brain: tasks drawing on the same network in humans recruit overlapping neurons in LLMs, whereas tasks drawing on different networks recruit distinct neurons. The convergent emergence of modularity in brains and neural networks suggests that it may be a fundamental property of intelligent systems.
Summary
The human brain displays clear functional specialization, with separate networks handling language, formal reasoning, social reasoning about other minds, and reasoning about the physical world. Researchers have long wondered whether this modular layout reflects a core requirement for any intelligent system or simply an outcome of biological evolution. To address the question, the study examined whether large language models develop a comparable organization despite being trained through an entirely different process.
The authors performed circuit-level analyses on models across 46 tasks drawn from the same four cognitive domains used to map human brain networks. Tasks that engage the same network in humans were found to activate overlapping sets of neurons in the models, while tasks from different domains consistently recruited largely distinct neuronal populations. This pattern held across the tested models and produced a modular architecture that closely parallels the specialization observed in human neuroimaging studies.
The results indicate that modularity is not limited to biological tissue but can arise in artificial systems optimized for next-token prediction. The convergent appearance of domain-specific neuron groups in both brains and language models suggests that such organization may be a general property of systems that must handle diverse cognitive demands efficiently.
Why it matters
This paper provides deep insights into the mechanistic interpretability of LLMs, a key area for Dutch AI researchers focused on transparent and ethical AI. Understanding the modular nature of LLMs can help local research institutions and advanced practitioners design more efficient, explainable, and aligned AI systems.
