LLMs may separate language, logic and the physical world into different neurons

If an LLM really thinks with one universal "neural network", different kinds of reasoning ought to engage roughly the same internal mechanism. But an analysis of 46 tasks shows a different picture: language, formal logic, understanding others' intentions and the physical world more often rely on different groups of the model's neurons.

The researchers compared four cognitive domains with the way they are organised in the human brain. Tasks that in people use a single network activated overlapping groups of neurons in the LLM. Tasks from different domains were already more separate.

That is more interesting than another claim about models' "human thinking". It turns out specialisation can arise not only in a biological brain but in systems trained in a completely different way. Perhaps modularity is not an accident of evolution but a convenient way to organise complex behaviour.

The practical lesson for studying LLMs is simple: looking for one universal "reasoning centre" may be less useful than taking the model apart by task type. An error in logic and an error in understanding the physical world are potentially tied to different internal circuits — which means fixing them in future may also have to be done differently.

But there is an important caveat. The work shows the character of activations and the overlap of neuron groups; it does not prove the modules found are independent causes of the abilities. Nor can the brief description tell us which models exactly were analysed or how robust the result is to other task sets. So this is a strong signal about the internal structure of an LLM, but not yet a map of its "brain regions".

If models really do specialise by type of thinking, which matters more for reliability: finding such modules, or learning to test how they work on specific tasks? Source: cs.AI updates on arXiv.org