Large language models appear to distribute certain tasks among distinct groups of artificial neurons, in a way that recalls the organization of the human brain.
To understand this finding, we need to look under the hood of artificial intelligence systems. A large language model contains very many computational elements called artificial neurons. They gradually transform the information they receive to produce a response. Researchers wanted to know whether all of them participated equally in the same tasks.
The team analyzed 46 tasks divided into four families. Some involved language, while others concerned formal reasoning, social relationships, or understanding the physical world. The scientists then identified which artificial neurons were most involved in each activity.
A clear pattern emerged. Two tasks belonging to the same family activated more shared neurons. Conversely, activities belonging to different domains more often involved distinct groups. This separation forms what researchers call a modular organization.
The connection with the brain comes from this point. In humans, several brain networks are more heavily recruited depending on the activity being performed. Language processing therefore does not rely on exactly the same networks as certain forms of social reasoning. The study finds a comparable separation in the models tested, despite their very different operating mechanisms.
The authors also find that tasks associated with the same network in humans tend to share more neurons in the models.
This finding interests researchers for another reason. The models studied were not built by deliberately assigning one area to language and another to physics. This specialization emerges during training. It could therefore be an efficient way to organize the processing of different tasks.
However, the study remains an analysis of computer models, and the comparison with the brain concerns their functional organization. The authors now propose examining more closely why this separation emerges and whether it can be found in other artificial intelligence architectures.