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:	Computation and Language Machine Learning
Online Access:	Acceder al recurso
Tags:	Add Tag No Tags, Be the first to tag this record!

Similar Items

Unveiling and Harnessing Hidden Attention Sinks: Enhancing Large Language Models without Training through Attention Calibration
by: Yu, Zhongzhi, et al.
Published: (2024)

On the Existence and Behavior of Secondary Attention Sinks
by: Wong, Jeffrey T. H., et al.
Published: (2025)

How Attention Sinks Emerge in Large Language Models: An Interpretability Perspective
by: Peng, Runyu, et al.
Published: (2026)

Enhanced Structured State Space Models via Grouped FIR Filtering and Attention Sink Mechanisms
by: Meng, Tian, et al.
Published: (2024)

Forgetting to Forget: Attention Sink as A Gateway for Backdooring LLM Unlearning
by: Shang, Bingqi, et al.
Published: (2025)

Attention Sinks as Internal Signals for Hallucination Detection in Large Language Models
by: Binkowski, Jakub, et al.
Published: (2026)

When Attention Sink Emerges in Language Models: An Empirical View
by: Gu, Xiangming, et al.
Published: (2024)

Attention Sinks: A 'Catch, Tag, Release' Mechanism for Embeddings
by: Zhang, Stephen, et al.
Published: (2025)

Attention Sinks in Massively Multilingual Neural Machine Translation:Discovery, Analysis, and Mitigation
by: Mutisya, Hillary, et al.
Published: (2026)

Prefixing Attention Sinks can Mitigate Activation Outliers for Large Language Model Quantization
by: Son, Seungwoo, et al.
Published: (2024)

Sink-Aware Pruning for Diffusion Language Models
by: Myrzakhan, Aidar, et al.
Published: (2026)

Does RoBERTa Perform Better than BERT in Continual Learning: An Attention Sink Perspective
by: Bai, Xueying, et al.
Published: (2024)

Garbage Attention in Large Language Models: BOS Sink Heads and Sink-aware Pruning
by: Sok, Jaewon, et al.
Published: (2026)

Attention Sinks and Outliers in Attention Residuals
by: Luo, Haozheng, et al.
Published: (2026)

PithTrain: A Compact and Agent-Native MoE Training System
by: Lai, Ruihang, et al.
Published: (2026)

CTR-Sink: Attention Sink for Language Models in Click-Through Rate Prediction
by: Li, Zixuan, et al.
Published: (2025)

The Spike, the Sparse and the Sink: Anatomy of Massive Activations and Attention Sinks
by: Sun, Shangwen, et al.
Published: (2026)

LocMoE: A Low-Overhead MoE for Large Language Model Training
by: Li, Jing, et al.
Published: (2024)

Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention
by: Yuan, Jingyang, et al.
Published: (2025)

Improving Transformers with Dynamically Composable Multi-Head Attention
by: Xiao, Da, et al.
Published: (2024)

Mechanism and Emergence of Stacked Attention Heads in Multi-Layer Transformers
by: Musat, Tiberiu
Published: (2024)

Condense, Don't Just Prune: Enhancing Efficiency and Performance in MoE Layer Pruning
by: Cao, Mingyu, et al.
Published: (2024)

MoE-Prism: Disentangling Monolithic Experts for Elastic MoE Services via Model-System Co-Designs
by: Xia, Xinfeng, et al.
Published: (2025)

Jakiro: Boosting Speculative Decoding with Decoupled Multi-Head via MoE
by: Huang, Haiduo, et al.
Published: (2025)

ASAP: Attention Sink Anchored Pruning
by: Lee, Jaehyuk, et al.
Published: (2026)

MoE-Sieve: Routing-Guided LoRA for Efficient MoE Fine-Tuning
by: Manzoni, Andrea
Published: (2026)

Attention Sinks in Diffusion Language Models
by: Rulli, Maximo Eduardo, et al.
Published: (2025)

LayerBoost: Layer-Aware Attention Reduction for Efficient LLMs
by: Souibgui, Mohamed Ali, et al.
Published: (2026)

Attention Sinks Induce Gradient Sinks: Massive Activations as Gradient Regulators in Transformers
by: Chen, Yihong, et al.
Published: (2026)

Faster MoE LLM Inference for Extremely Large Models
by: Yang, Haoqi, et al.
Published: (2025)

Stabilizing MoE Reinforcement Learning by Aligning Training and Inference Routers
by: Ma, Wenhan, et al.
Published: (2025)

Efficient Streaming Language Models with Attention Sinks
by: Xiao, Guangxuan, et al.
Published: (2023)

Spectral Filters, Dark Signals, and Attention Sinks
by: Cancedda, Nicola
Published: (2024)

Unchosen Experts Can Contribute Too: Unleashing MoE Models' Power by Self-Contrast
by: Shi, Chufan, et al.
Published: (2024)

NOSA: Native and Offloadable Sparse Attention
by: Huang, Yuxiang, et al.
Published: (2025)

HybriMoE: Hybrid CPU-GPU Scheduling and Cache Management for Efficient MoE Inference
by: Zhong, Shuzhang, et al.
Published: (2025)

MoE-SpAc: Efficient MoE Inference Based on Speculative Activation Utility in Heterogeneous Edge Scenarios
by: Li, Shuhuai, et al.
Published: (2026)

Alloc-MoE: Budget-Aware Expert Activation Allocation for Efficient Mixture-of-Experts Inference
by: Liu, Baihui, et al.
Published: (2026)

Flux Attention: Context-Aware Hybrid Attention for Efficient LLMs Inference
by: Qiu, Quantong, et al.
Published: (2026)

GRIN: GRadient-INformed MoE
by: Liu, Liyuan, et al.
Published: (2024)