A new framework treats large language models like brains, mapping how their internal skills actually cluster together.
Researchers built a network-based framework that links cognitive skills, model architecture, and training data, adapting analytical tools from biology to study how capabilities organize inside LLMs. Applied to existing models, the approach found "module communities" - clusters of components sharing related skills - whose patterns partially resemble the distributed, interconnected organization seen in avian and small-mammal brains, rather than the tightly localized specialization found in some biological systems. The researchers also flagged a divergence: LLMs appear to pick up new skills mainly through dynamic, cross-module interaction rather than the neuroplasticity and regional specialization biological brains rely on.
That distinction matters for anyone trying to fine-tune these models. The paper argues that because skill acquisition depends on interaction across the network rather than isolated modules, tuning strategies that surgically edit single components are likely to underperform broader, distributed approaches. It's a data point in the wider effort to make LLMs less of a black box, alongside other interpretability techniques like probing classifiers and sparse autoencoders.
Still, "partially mirrors" bird and small-mammal brains is a hedge, not a breakthrough - this is a mapping exercise built and tested by its own authors, not proof that a language model reasons anything like an animal does.