Token Homogenization under Positional Bias

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Yusupov, Viacheslav, Maksimov, Danil, Alaeva, Ameliia, Zaitceva, Tatiana, Anna, Antipina, Vasileva, Anna, Liu, Chenlin, Chheng, Rayuth, Sazanakov, Danil, Chetvergov, Andrey, Ermilova, Alina, Shvetsov, Egor
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866914001819009024
author Yusupov, Viacheslav
Maksimov, Danil
Alaeva, Ameliia
Zaitceva, Tatiana
Anna, Antipina
Vasileva, Anna
Liu, Chenlin
Chheng, Rayuth
Sazanakov, Danil
Chetvergov, Andrey
Ermilova, Alina
Shvetsov, Egor
author_facet Yusupov, Viacheslav
Maksimov, Danil
Alaeva, Ameliia
Zaitceva, Tatiana
Anna, Antipina
Vasileva, Anna
Liu, Chenlin
Chheng, Rayuth
Sazanakov, Danil
Chetvergov, Andrey
Ermilova, Alina
Shvetsov, Egor
contents This paper investigates token homogenization - the convergence of token representations toward uniformity across transformer layers and its relationship to positional bias in large language models. We empirically examine whether homogenization occurs and how positional bias amplifies this effect. Through layer-wise similarity analysis and controlled experiments, we demonstrate that tokens systematically lose distinctiveness during processing, particularly when biased toward extremal positions. Our findings confirm both the existence of homogenization and its dependence on positional attention mechanisms.
format Preprint
id arxiv_https___arxiv_org_abs_2508_17126
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Token Homogenization under Positional Bias
Yusupov, Viacheslav
Maksimov, Danil
Alaeva, Ameliia
Zaitceva, Tatiana
Anna, Antipina
Vasileva, Anna
Liu, Chenlin
Chheng, Rayuth
Sazanakov, Danil
Chetvergov, Andrey
Ermilova, Alina
Shvetsov, Egor
Computation and Language
Artificial Intelligence
Machine Learning
This paper investigates token homogenization - the convergence of token representations toward uniformity across transformer layers and its relationship to positional bias in large language models. We empirically examine whether homogenization occurs and how positional bias amplifies this effect. Through layer-wise similarity analysis and controlled experiments, we demonstrate that tokens systematically lose distinctiveness during processing, particularly when biased toward extremal positions. Our findings confirm both the existence of homogenization and its dependence on positional attention mechanisms.
title Token Homogenization under Positional Bias
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2508.17126