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

TL;DR

This study investigates how training influences the geometric organization of representations in language models and how this relates to their alignment with human brain activity, revealing stable low-complexity clusters that predict neural responses.

Contribution

It introduces a geometric modularization framework that links training dynamics in language models to brain-like neural alignment, highlighting the role of representational smoothing.

Findings

Low-complexity modules predict brain activity better.

Geometric reorganization correlates with training progress.

Alignment improves with model scale and reduced curvature.

Abstract

How large language models (LLMs) align with the neural representation and computation of human language is a central question in cognitive science. Using representational geometry as a mechanistic lens, we addressed this by tracking entropy, curvature, and fMRI encoding scores throughout Pythia (70M-1B) training. We identified a geometric modularization where layers self-organize into stable low- and high-complexity clusters. The low-complexity module, characterized by reduced entropy and curvature, consistently better predicted human language network activity. This alignment followed heterogeneous spatial-temporal trajectories: rapid and stable in temporal regions (AntTemp, PostTemp), but delayed and dynamic in frontal areas (IFG, IFGorb). Crucially, reduced curvature remained a robust predictor of model-brain alignment even after controlling for training progress, an effect that strengthened with model scale. These results links training-driven geometric reorganization to temporal-frontal functional specialization, suggesting that representational smoothing facilitates neural-like linguistic processing.