Recurrent neural networks (RNNs) trained on several cognitive tasks developed more modular connection patterns than single-task models, with the clearest gains in the smallest tested networks (8 and 16 hidden units). The peer-reviewed version appeared in Nature Machine Intelligence on Sept. 28; the work first circulated as a 2025 preprint.
In these models, a module is a group of units with relatively dense connections inside the group and fewer links to other groups. The team trained RNNs on 20 lab-style tasks, comparing single-task, all-at-once multitask and incremental schedules across models with 8 to 64 hidden units. Multitask training raised modularity most in the capacity-limited 8- and 16-unit models; gradually adding connections as tasks arrived raised it further.
The researchers also used data from 84 cortical areas in the Human Connectome Project to compare generative network models. A model combining physical wiring distance with task-specific constraints fit the measured brain network better than models using either factor alone. This suggests task demands can complement spatial wiring costs as one explanation for brain-like structure.
The experiments show that task demands can produce modularity in simplified recurrent networks. They do not establish why biological brains developed their wiring, or whether the result transfers to transformers and other modern AI architectures. The study’s processed data and analysis code are public.