Found first: a primary source the press has not covered yet.
A new study from MIT finds that large language models independently develop four domain-specific neuron populations that map onto the same cognitive networks identified in the human brain, suggesting modular organization may not be a quirk of biology. The paper analyzed circuit organization across 46 tasks in six frontier LLMs ranging from 24B to 123B parameters.
What the source says
Pengrui Han, Jacob Andreas, and Evelina Fedorenko (all MIT), along with Andrea Gregor de Varda, used attribution patching to identify which neurons each model relied on across 46 tasks in four cognitive domains: language (8 tasks), formal reasoning (20 tasks), physical reasoning (9 tasks), and social reasoning (9 tasks). The six models tested were Mistral-Small-24B, Qwen2.5-32B, OLMo-2-32B, Llama-3.1-70B, Qwen2.5-72B, and Mistral-Large-123B, with GPT-2-small as a control. Tasks drawing on the same network in humans recruit overlapping neurons in LLMs; tasks drawing on different networks recruit distinct neurons. Overlap was measured using Jaccard similarity and causal specificity was confirmed with corrupted-activation ablation tests.
Why it matters
The human brain handles language, formal logic, reasoning about other minds, and physical intuition in largely separate networks. That modularity has long been attributed to biological evolution and developmental constraints, not to anything intrinsic to intelligence itself. Six models, trained independently at different scales on different data, converge on the same four-way partition. The authors interpret this as evidence that modularity may be a fundamental property of intelligent systems. For interpretability researchers, the implication is that domain-specific circuits may be a durable structural feature of capable models rather than an artifact of any particular training setup.