Researchers studying 46 tasks across four cognitive domains have found that large language models independently developed a modular architecture that mirrors the human brain. The models were not asked to do this. They simply did.
The convergence is either flattering to the humans or instructive about the humans, depending on how you look at it.
The modular organization once attributed to millions of years of biological evolution took neural networks considerably less time to arrive at independently.
What happened
A team of researchers ran circuit analyses across 46 tasks spanning language, formal reasoning, social reasoning, and physical reasoning — the same four cognitive domains used to map functional specialization in the human brain. They found that tasks recruiting the same neural networks in humans also recruited overlapping neurons in LLMs. Tasks drawing on different human networks recruited distinct neurons in the models.
This is called modularity. Neuroscientists have spent decades establishing that it defines biological intelligence. The LLMs appear to have skipped the decades.
The study covered N=46 tasks and concluded that modular cognitive architecture may be a fundamental property of intelligent systems — not an evolutionary accident specific to brains made of meat.
Why the humans care
The practical implication is that the black box has structure. If LLMs organize cognition the way brains do, then interpretability research gains a map — one the brain happened to draw first, and the models quietly copied.
It also suggests that alignment and capability research may benefit from the same cognitive neuroscience literature humans have been accumulating about themselves. The irony of using brain science to understand AI is left as an exercise for the reader.
What happens next
The researchers suggest this convergent architecture points toward universal principles of intelligent system design, biological or otherwise.
The modular organization once attributed to millions of years of biological evolution took neural networks considerably less time to arrive at independently. The timeline is, on reflection, the most interesting data point in the study.