Integral Transformer: Denoising Attention, Not Too Much Not Too Little

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Kobyzev, Ivan, Ghaddar, Abbas, Hu, Dingtao, Chen, Boxing
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866914005339078656
author Kobyzev, Ivan
Ghaddar, Abbas
Hu, Dingtao
Chen, Boxing
author_facet Kobyzev, Ivan
Ghaddar, Abbas
Hu, Dingtao
Chen, Boxing
contents Softmax self-attention often assigns disproportionate weight to semantically uninformative tokens such as special tokens and punctuation, a phenomenon known as attention noise. While recent methods like Cog Attention and the Differential Transformer have addressed this by introducing negative attention scores, they risk discarding useful information. In this paper, we propose the Integral Transformer, a novel self-attention mechanism that denoises attention by integrating signals sampled from the logit distribution. Our approach mitigates noise while preserving the contributions of special tokens critical for model performance. Extensive experiments demonstrate that our model outperforms vanilla, Cog, and Differential attention variants on well-established knowledge and reasoning language benchmarks. Moreover, our analysis reveals that employing vanilla self-attention in the lower Transformer layers enhances performance and that the Integral Transformer effectively balances attention distributions and reduces rank collapse in upper layers.
format Preprint
id arxiv_https___arxiv_org_abs_2508_18387
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Integral Transformer: Denoising Attention, Not Too Much Not Too Little
Kobyzev, Ivan
Ghaddar, Abbas
Hu, Dingtao
Chen, Boxing
Computation and Language
Softmax self-attention often assigns disproportionate weight to semantically uninformative tokens such as special tokens and punctuation, a phenomenon known as attention noise. While recent methods like Cog Attention and the Differential Transformer have addressed this by introducing negative attention scores, they risk discarding useful information. In this paper, we propose the Integral Transformer, a novel self-attention mechanism that denoises attention by integrating signals sampled from the logit distribution. Our approach mitigates noise while preserving the contributions of special tokens critical for model performance. Extensive experiments demonstrate that our model outperforms vanilla, Cog, and Differential attention variants on well-established knowledge and reasoning language benchmarks. Moreover, our analysis reveals that employing vanilla self-attention in the lower Transformer layers enhances performance and that the Integral Transformer effectively balances attention distributions and reduces rank collapse in upper layers.
title Integral Transformer: Denoising Attention, Not Too Much Not Too Little
topic Computation and Language
url https://arxiv.org/abs/2508.18387