Attention Sink Forges Native MoE in Attention Layers: Sink-Aware Training to Address Head Collapse

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Fu, Zizhuo, Zeng, Wenxuan, Wang, Runsheng, Li, Meng
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913164993495040
author Fu, Zizhuo
Zeng, Wenxuan
Wang, Runsheng
Li, Meng
author_facet Fu, Zizhuo
Zeng, Wenxuan
Wang, Runsheng
Li, Meng
contents Large Language Models (LLMs) often assign disproportionate attention to the first token, a phenomenon known as the attention sink. Several recent approaches aim to address this issue, including Sink Attention in GPT-OSS and Gated Attention in Qwen3-Next. However, a comprehensive analysis of the relationship among these attention mechanisms is lacking. In this work, we provide both theoretical and empirical evidence demonstrating that the sink in Vanilla Attention and Sink Attention naturally construct a Mixture-of-Experts (MoE) mechanism within attention layers. This insight explains the head collapse phenomenon observed in prior work, where only a fixed subset of attention heads contributes to generation. To mitigate head collapse, we propose a sink-aware training algorithm with an auxiliary load balancing loss designed for attention layers. Extensive experiments show that our method achieves effective head load balancing and improves model performance across Vanilla Attention, Sink Attention, and Gated Attention. We hope this study offers a new perspective on attention mechanisms and encourages further exploration of the inherent MoE structure within attention layers.
format Preprint
id arxiv_https___arxiv_org_abs_2602_01203
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Attention Sink Forges Native MoE in Attention Layers: Sink-Aware Training to Address Head Collapse
Fu, Zizhuo
Zeng, Wenxuan
Wang, Runsheng
Li, Meng
Computation and Language
Machine Learning
Large Language Models (LLMs) often assign disproportionate attention to the first token, a phenomenon known as the attention sink. Several recent approaches aim to address this issue, including Sink Attention in GPT-OSS and Gated Attention in Qwen3-Next. However, a comprehensive analysis of the relationship among these attention mechanisms is lacking. In this work, we provide both theoretical and empirical evidence demonstrating that the sink in Vanilla Attention and Sink Attention naturally construct a Mixture-of-Experts (MoE) mechanism within attention layers. This insight explains the head collapse phenomenon observed in prior work, where only a fixed subset of attention heads contributes to generation. To mitigate head collapse, we propose a sink-aware training algorithm with an auxiliary load balancing loss designed for attention layers. Extensive experiments show that our method achieves effective head load balancing and improves model performance across Vanilla Attention, Sink Attention, and Gated Attention. We hope this study offers a new perspective on attention mechanisms and encourages further exploration of the inherent MoE structure within attention layers.
title Attention Sink Forges Native MoE in Attention Layers: Sink-Aware Training to Address Head Collapse
topic Computation and Language
Machine Learning
url https://arxiv.org/abs/2602.01203