LLMs Mirror the Brain's Modular Neuron Layout, Study Finds

Across six instruction-tuned language models spanning 24B to 123B parameters, the same task types recruit the same neurons, and different task types do not. That is the core finding of "Modular Cognitive Architecture Emerges in Large Language Models", an arxiv preprint from Pengrui Han, Jacob Andreas, Evelina Fedorenko, and Andrea Gregor de Varda.

The team ran attribution patching over 46 tasks sorted into four cognitive domains borrowed from human neuroscience: language (8 tasks, 8,877 pairs), formal reasoning (20 tasks, 19,941 pairs), physical reasoning (9 tasks, 9,200 pairs), and social reasoning (9 tasks, 11,412 pairs). They scored which MLP neurons mattered for each task, then measured how much the top neurons overlapped across tasks. The gap they report is stark: "Averaged across six models, this within-domain overlap exceeds cross-domain overlap by more than fourfold (12.9% vs 3.0%, permutation test p < 0.0001)."

Ablating those neurons hurt the corresponding tasks and left the rest largely alone. The paper reports "Within-domain ablations cause a 25.9% accuracy drop versus 2.5% for cross-domain ablations (ratio 10.3x, p < 0.0001)."

When the authors clustered the 46x46 task matrix without giving the algorithm the human-brain groupings, the same four domains fell back out. "Unsupervised hierarchical clustering of the 46x46 task matrix recovers the four cognitive domains defined in neuroscience (Adjusted Rand Index = 0.78, p < 0.0001)."

The abstract's framing question is unusually direct: "Is this modular organization a fundamental principle of how intelligent systems must be built, or an evolutionary accident specific to biological brains?" The answer the paper argues for is the first. "The convergent emergence of modularity in brains and neural networks suggests that it may be a fundamental property of intelligent systems."

Which six models were tested is not spelled out in the abstract or on the project page. The demo names Qwen2.5-32B, and the paper notes the six span four families across the 24B-123B range.