Elastic Attention: Test-time Adaptive Sparsity Ratios for Efficient Transformers

• URL: http://arxiv.org/abs/2601.17367v2
• Date: Wed, 28 Jan 2026 10:28:15 GMT
• Title: Elastic Attention: Test-time Adaptive Sparsity Ratios for Efficient Transformers
• Authors: Zecheng Tang, Quantong Qiu, Yi Yang, Zhiyi Hong, Haiya Xiang, Kebin Liu, Qingqing Dang, Juntao Li, Min Zhang,
• Abstract summary: We propose Elastic Attention, which allows the model to adjust its overall sparsity based on the input.<n>Within only 12 hours of training on 8xA800 GPUs, our method enables models to achieve both strong performance and efficient inference.
• Score: 42.80120203718226
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: The quadratic complexity of standard attention mechanisms poses a significant scalability bottleneck for large language models (LLMs) in long-context scenarios. While hybrid attention strategies that combine sparse and full attention within a single model offer a viable solution, they typically employ static computation ratios (i.e., fixed proportions of sparse versus full attention) and fail to adapt to the varying sparsity sensitivities of downstream tasks during inference. To address this issue, we propose Elastic Attention, which allows the model to dynamically adjust its overall sparsity based on the input. This is achieved by integrating a lightweight Attention Router into the existing pretrained model, which dynamically assigns each attention head to different computation modes. Within only 12 hours of training on 8xA800 GPUs, our method enables models to achieve both strong performance and efficient inference. Experiments across three long-context benchmarks on widely-used LLMs demonstrate the superiority of our method.

Related papers

• MiniCPM-SALA: Hybridizing Sparse and Linear Attention for Efficient Long-Context Modeling [80.48332380100915]
  MiniCPM-SALA is a hybrid model that integrates the high-fidelity long-context modeling of sparse attention with the global efficiency of linear attention.<n>On a single NVIDIA A6000D GPU, the model achieves up to 3.5x the inference speed of the full-attention model at the sequence length of 256K tokens.
  arXiv  Detail & Related papers  (2026-02-12T09:37:05Z)
• DyCAF-Net: Dynamic Class-Aware Fusion Network [0.0]
  We introduce Dynamic Class-Aware Fusion Network (DyCAF-Net)<n>DyCAF-Net achieves significant improvements in precision, mAP@50, and mAP@50-95 across 13 diverse benchmarks.<n>Its adaptability to scale variance, semantic overlaps, and class imbalance positions it as a robust solution for real-world detection tasks.
  arXiv  Detail & Related papers  (2025-08-05T16:06:26Z)
• Trainable Dynamic Mask Sparse Attention [11.506985057671015]
  We introduce a trainable dynamic mask sparse attention mechanism, a method that merges the advantages of both position-aware and content-aware approaches.<n>We demonstrate that the introduced dynamic mask and sparse weights do not obstruct gradients, supporting end-to-end training.
  arXiv  Detail & Related papers  (2025-08-04T07:05:15Z)
• Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free [81.65559031466452]
  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.
  arXiv  Detail & Related papers  (2025-05-10T17:15:49Z)
• Communication-Efficient Wireless Federated Fine-Tuning for Large-Scale AI Models [13.742950928229078]
  Low-Rank Adaptation (LoRA) addresses these issues by training compact, low-rank matrices instead of fully fine-tuning large models.<n>This paper introduces a wireless federated LoRA fine-tuning framework that optimize both learning performance and communication efficiency.
  arXiv  Detail & Related papers  (2025-05-01T06:15:38Z)
• Balcony: A Lightweight Approach to Dynamic Inference of Generative Language Models [31.103832542711864]
  Balcony is a framework for depth-based dynamic inference.<n>It maintains the full model's performance while enabling real-time adaptation to different computational budgets.<n>Remarkably, we show that Balcony outperforms state-of-the-art methods such as Flextron and Layerskip.
  arXiv  Detail & Related papers  (2025-03-06T22:09:55Z)
• PRISM: Self-Pruning Intrinsic Selection Method for Training-Free Multimodal Data Selection [68.8373788348678]
  Visual instruction tuning adapts pre-trained Multimodal Large Language Models to follow human instructions.<n>PRISM is the first training-free framework for efficient visual instruction selection.<n>It reduces the end-to-end time for data selection and model tuning to just 30% of conventional pipelines.
  arXiv  Detail & Related papers  (2025-02-17T18:43:41Z)
• CLEAR: Conv-Like Linearization Revs Pre-Trained Diffusion Transformers Up [64.38715211969516]
  We introduce a convolution-like local attention strategy termed CLEAR, which limits feature interactions to a local window around each query token.<n>Experiments indicate that, by fine-tuning the attention layer on merely 10K self-generated samples for 10K iterations, we can effectively transfer knowledge from a pre-trained DiT to a student model with linear complexity.
  arXiv  Detail & Related papers  (2024-12-20T17:57:09Z)
• Adaptive Training Meets Progressive Scaling: Elevating Efficiency in Diffusion Models [52.1809084559048]
  We propose a novel two-stage divide-and-conquer training strategy termed TDC Training.<n>It groups timesteps based on task similarity and difficulty, assigning highly customized denoising models to each group, thereby enhancing the performance of diffusion models.<n>While two-stage training avoids the need to train each model separately, the total training cost is even lower than training a single unified denoising model.
  arXiv  Detail & Related papers  (2023-12-20T03:32:58Z)
• Characterizing and overcoming the greedy nature of learning in multi-modal deep neural networks [62.48782506095565]
  We show that due to the greedy nature of learning in deep neural networks, models tend to rely on just one modality while under-fitting the other modalities. We propose an algorithm to balance the conditional learning speeds between modalities during training and demonstrate that it indeed addresses the issue of greedy learning.
  arXiv  Detail & Related papers  (2022-02-10T20:11:21Z)

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.