A team at MIT has identified a striking parallel between artificial intelligence and biology: six large language models exhibit the same modular organization of four cognitive domains that neuroscientists have described in the human brain. The study, authored by Pengrui Han, Jacob Andreas, Evelina Fedorenko, and Andrea Gregor de Varda, was submitted to arXiv on June 27, 2026 (reference 2608.13567).
Fedorenko is one of MIT’s most cited neurolinguists, with decades of research on the brain’s language network. Her involvement matters: she understands what modularity means in biological systems, and now she is searching for it in silicon.
- Four cognitive domains: MIT analyzed 46 tasks across language, formal reasoning, physical reasoning, and social reasoning—mirroring systems in the human cortex.
- Six models confirmed: The pattern held from Mistral-Small-24B (24B parameters) to Mistral-Large-123B (123B), plus Qwen2.5-32B, OLMo-2-32B, Llama-3.1-70B, and Qwen2.5-72B. GPT-2-small served as a control.
- Causal evidence: Using attribution patching, the Jaccard index, and causal ablation, researchers showed the modular overlap is functionally relevant—not merely correlational.
What the MIT team found
The researchers examined 46 cognitive tasks divided into four domains: language (8 tasks), formal reasoning (20 tasks), physical reasoning (9 tasks), and social reasoning (9 tasks). That taxonomy maps onto four cognitive systems identified in the human cortex.
The hypothesis was straightforward: if language models use internal representations similar to those of the brain, tasks that activate the same neural systems in humans should activate overlapping groups of artificial neurons in the models. Tasks that engage different systems in humans should activate distinct groups.
Results confirmed the hypothesis across six models, ranging from Mistral-Small-24B with 24 billion parameters to Mistral-Large-123B with 123 billion parameters. The line-up also included Qwen2.5-32B, OLMo-2-32B, Llama-3.1-70B, and Qwen2.5-72B. GPT-2-small served as a control.
The technical approach had three components: attribution patching to identify which parameters influence a response, the Jaccard index to measure overlap between activated neuron sets, and causal ablation by activation to verify that the overlap is functionally meaningful rather than just correlational.
Separate research on astronaut brains under microgravity also points to plasticity and cognitive reorganization: cognitive systems are more modular and robust than previously assumed, both in biology and apparently in silicon.
From correlation to causation
The distinction between “AI models work like brains” as a convenient metaphor and as an experimental finding is important. Causal ablation is the key here: the team did not just measure correlations. They deactivated specific groups of artificial neurons and verified that doing so selectively affected tasks in the corresponding domain without affecting the others.
This is a notable departure from much interpretability research. Most papers on large language model interpretability identify patterns but do not prove those patterns have functional relevance. This study does. That methodological leap—from correlational observation to causal verification—is what makes the result worth attention.
The implications extend to multimodal systems. Multimodal hallucinations in AI vision models are one of the problems interpretability research aims to address. If physical reasoning modules and language modules are functionally distinct, errors in one type should not contaminate the other. That has practical consequences for improving the reliability of multimodal AI.
Why modularity matters for AI design
The most immediate implication is philosophical, but it carries practical weight. If cognitive modularity emerges independently in systems as different as the biological brain and artificial neural networks, that suggests the organization is not an evolutionary accident. It may be a structural solution to cognitive efficiency.
The analyzed models were not designed with explicit cognitive modules. No one told them during training to use one set of parameters for language and another for physical reasoning. Modularity emerged from the optimization process over data.
The Socratic learning mode introduced by Claude suggests AI systems can adapt to different processing modes based on context. MIT’s research provides a theoretical framework for why that is possible: different reasoning types are organized into modules that activate relatively independently. It also suggests that future AI systems could benefit from explicitly reinforcing those separations. Fault diagnosis, in turn, could become far more precise.
The study’s most relevant contribution is not confirming that LLMs are “like brains”—that analogy has limits. It is the possibility that modularity is an inevitable consequence of optimizing any system for general information processing, regardless of substrate. If true, that insight could shape both the architecture and the debugging of the next generation of AI.