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.
@article{arxiv.2508.17126,
title = {Token Homogenization under Positional Bias},
author = {Viacheslav Yusupov and Danil Maksimov and Ameliia Alaeva and Tatiana Zaitceva and Antipina Anna and Anna Vasileva and Chenlin Liu and Rayuth Chheng and Danil Sazanakov and Andrey Chetvergov and Alina Ermilova and Egor Shvetsov},
journal= {arXiv preprint arXiv:2508.17126},
year = {2025}
}