Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free

• URL: http://arxiv.org/abs/2505.06708v1
• Date: Sat, 10 May 2025 17:15:49 GMT
• Title: Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free
• Authors: Zihan Qiu, Zekun Wang, Bo Zheng, Zeyu Huang, Kaiyue Wen, Songlin Yang, Rui Men, Le Yu, Fei Huang, Suozhi Huang, Dayiheng Liu, Jingren Zhou, Junyang Lin,
• Abstract summary: We conduct experiments to investigate gating-augmented softmax attention variants.<n>We find that a simple modification-applying a head-specific sigmoid gate after the Scaled Dot-Product Attention (SDPA)-consistently improves performance.
• Score: 81.65559031466452
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Gating mechanisms have been widely utilized, from early models like LSTMs and Highway Networks to recent state space models, linear attention, and also softmax attention. Yet, existing literature rarely examines the specific effects of gating. In this work, we conduct comprehensive experiments to systematically investigate gating-augmented softmax attention variants. Specifically, we perform a comprehensive comparison over 30 variants of 15B Mixture-of-Experts (MoE) models and 1.7B dense models trained on a 3.5 trillion token dataset. Our central finding is that a simple modification-applying a head-specific sigmoid gate after the Scaled Dot-Product Attention (SDPA)-consistently improves performance. This modification also enhances training stability, tolerates larger learning rates, and improves scaling properties. By comparing various gating positions and computational variants, we attribute this effectiveness to two key factors: (1) introducing non-linearity upon the low-rank mapping in the softmax attention, and (2) applying query-dependent sparse gating scores to modulate the SDPA output. Notably, we find this sparse gating mechanism mitigates 'attention sink' and enhances long-context extrapolation performance, and we also release related $\href{https://github.com/qiuzh20/gated_attention}{codes}$ and $\href{https://huggingface.co/QwQZh/gated_attention}{models}$ to facilitate future research.

Related papers

• S*: Test Time Scaling for Code Generation [55.11863577956177]
  We propose S*, the first hybrid test-time scaling framework for code generation.<n>S* substantially improves the coverage and selection accuracy of generated code.
  arXiv  Detail & Related papers  (2025-02-20T09:18:53Z)
• FuXi-$α$: Scaling Recommendation Model with Feature Interaction Enhanced Transformer [81.12174905444229]
  Recent advancements have shown that expanding sequential recommendation models to large-scale recommendation models can be an effective strategy.<n>We propose a new model called FuXi-$alpha$ to address these issues.<n>Our model outperforms existing models, with its performance continuously improving as the model size increases.
  arXiv  Detail & Related papers  (2025-02-05T09:46:54Z)
• Towards Scalable and Deep Graph Neural Networks via Noise Masking [59.058558158296265]
  Graph Neural Networks (GNNs) have achieved remarkable success in many graph mining tasks.<n> scaling them to large graphs is challenging due to the high computational and storage costs.<n>We present random walk with noise masking (RMask), a plug-and-play module compatible with the existing model-simplification works.
  arXiv  Detail & Related papers  (2024-12-19T07:48:14Z)
• DiTMoS: Delving into Diverse Tiny-Model Selection on Microcontrollers [34.282971510732736]
  We introduce DiTMoS, a novel DNN training and inference framework with a selector-classifiers architecture. A composition of weak models can exhibit high diversity and the union of them can significantly boost the accuracy upper bound. We deploy DiTMoS on the Neucleo STM32F767ZI board and evaluate it based on three time-series datasets for human activity recognition, keywords spotting, and emotion recognition.
  arXiv  Detail & Related papers  (2024-03-14T02:11:38Z)
• Simple linear attention language models balance the recall-throughput tradeoff [60.06020449520365]
  We propose BASED, a simple architecture combining linear and sliding window attention.<n>We train language models up to 1.3b parameters and show that BASED matches the strongest sub-quadratic models in perplexity and outperforms them on real-world recall-intensive tasks by 6.22 accuracy points.
  arXiv  Detail & Related papers  (2024-02-28T19:28:27Z)
• TransNormerLLM: A Faster and Better Large Language Model with Improved TransNormer [34.790081960470964]
  We present TransNormerLLM, the first linear attention-based Large Language Model (LLM) We make advanced modifications that include positional embedding, linear attention acceleration, gating mechanisms, tensor normalization, and inference acceleration and stabilization. We validate our model design through a series of ablations and train models with sizes of 385M, 1B, and 7B on our self-collected corpus.
  arXiv  Detail & Related papers  (2023-07-27T16:45:33Z)
• Anomaly Detection via Multi-Scale Contrasted Memory [3.0170109896527086]
  We introduce a new two-stage anomaly detector which memorizes during training multi-scale normal prototypes to compute an anomaly deviation score. Our model highly improves the state-of-the-art performance on a wide range of object, style and local anomalies with up to 35% error relative improvement on CIFAR-10.
  arXiv  Detail & Related papers  (2022-11-16T16:58:04Z)
• Sparse Attention with Linear Units [60.399814410157425]
  We introduce a novel, simple method for achieving sparsity in attention: we replace the softmax activation with a ReLU. Our model, which we call Rectified Linear Attention (ReLA), is easy to implement and more efficient than previously proposed sparse attention mechanisms. Our analysis shows that ReLA delivers high sparsity rate and head diversity, and the induced cross attention achieves better accuracy with respect to source-target word alignment.
  arXiv  Detail & Related papers  (2021-04-14T17:52:38Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.