Training-Driven Representational Geometry Modularization Predicts Brain Alignment in Language Models

ABSTRACT

The degree to which Large Language Models (LLMs) align with human brain activity during language processing remains a central question in both AI and neuroscience. We investigate the impact of training dynamics on the 'representational geometry' of LLMs and its correlation with fMRI and MEG recordings. Our results reveal that as models undergo training, their internal representations naturally become more modularized, mimicking the functional specialization observed in the human language network. We demonstrate that this modularization is a strong predictor of brain alignment, suggesting that the pressure to compress linguistic information leads to a convergent representational structure between biological and artificial systems.