A language model (LM) is a mapping from a linguistic context to an output token. However, much remains to be known about this mapping, including how its geometric properties relate to its function. We take a high-level geometric approach to its analysis, observing, across five pre-trained transformer-based LMs and three input datasets, a distinct phase characterized by high intrinsic dimensionality. During this phase, representations (1) correspond to the first full linguistic abstraction of the input; (2) are the first to viably transfer to downstream tasks; (3) predict each other across different LMs. Moreover, we find that an earlier onset of the phase strongly predicts better language modelling performance. In short, our results suggest that a central high-dimensionality phase underlies core linguistic processing in many common LM architectures.
View on arXiv@article{cheng2025_2405.15471,
title={ Emergence of a High-Dimensional Abstraction Phase in Language Transformers },
author={ Emily Cheng and Diego Doimo and Corentin Kervadec and Iuri Macocco and Jade Yu and Alessandro Laio and Marco Baroni },
journal={arXiv preprint arXiv:2405.15471},
year={ 2025 }
}