Emergent Specialization: Rare Token Neurons in Language Models

TL;DR

This paper investigates how certain neurons in large language models become specialized for rare tokens, revealing their emergence during training and their coordinated activation patterns, which enhance understanding of model behavior in specialized domains.

Contribution

It identifies and characterizes rare token neurons, explaining their emergence and functional organization within language models, a novel insight into model specialization.

Findings

Rare token neurons form a coordinated subnetwork.

Emergence of specialization follows a three-phase organization.

Heavy-tailed weight distributions correlate with neuron specialization.

Abstract

Large language models struggle with representing and generating rare tokens despite their importance in specialized domains. In this study, we identify neuron structures with exceptionally strong influence on language model's prediction of rare tokens, termed as rare token neurons, and investigate the mechanism for their emergence and behavior. These neurons exhibit a characteristic three-phase organization (plateau, power-law, and rapid decay) that emerges dynamically during training, evolving from a homogeneous initial state to a functionally differentiated architecture. In the activation space, rare token neurons form a coordinated subnetwork that selectively co-activates while avoiding co-activation with other neurons. This functional specialization potentially correlates with the development of heavy-tailed weight distributions, suggesting a statistical mechanical basis for emergent specialization.