Scaling Attention via Feature Sparsity

• URL: http://arxiv.org/abs/2603.22300v1
• Date: Tue, 17 Mar 2026 08:41:50 GMT
• Title: Scaling Attention via Feature Sparsity
• Abstract summary: Scaling Transformers to ultra-long contexts is bottlenecked by the $O(n2 d)$ cost of self-attention.<n>We propose Sparse Feature Attention, where queries and keys are represented as $k$-sparse codes that preserve high-dimensional expressivity.<n>Across GPT-2 and Qwen3 pretraining, SFA matches dense baselines while improving speed by up to $2.5times$ and reducing FLOPs and KV-cache by nearly 50%.
• Score: 50.64995497733461
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Scaling Transformers to ultra-long contexts is bottlenecked by the $O(n^2 d)$ cost of self-attention. Existing methods reduce this cost along the sequence axis through local windows, kernel approximations, or token-level sparsity, but these approaches consistently degrade accuracy. In this paper, we instead explore an orthogonal axis: feature sparsity. We propose Sparse Feature Attention (SFA), where queries and keys are represented as $k$-sparse codes that preserve high-dimensional expressivity while reducing the cost of attention from $Θ(n^2 d)$ to $Θ(n^2 k^2/d)$. To make this efficient at scale, we introduce FlashSFA, an IO-aware kernel that extends FlashAttention to operate directly on sparse overlaps without materializing dense score matrices. Across GPT-2 and Qwen3 pretraining, SFA matches dense baselines while improving speed by up to $2.5\times$ and reducing FLOPs and KV-cache by nearly 50\%. On synthetic and downstream benchmarks, SFA preserves retrieval accuracy and robustness at long contexts, outperforming short-embedding baselines that collapse feature diversity. These results establish feature-level sparsity as a complementary and underexplored axis for efficient attention, enabling Transformers to scale to orders-of-magnitude longer contexts with minimal quality loss. Code is available at https://github.com/YannX1e/Sparse-Feature-Attention.

Related papers

• LimiX-2M: Mitigating Low-Rank Collapse and Attention Bottlenecks in Tabular Foundation Models [56.999481798138625]
  LimiX-2M is a 2M- parameter model that outperforms larger TabPFN-v2 and TabICL baselines on widely used benchmarks.<n>We present a unified tokenize-and-route framework for strong Tabular foundation models (TFMs)<n>Results highlight value-aware tokenization and readout-aligned routing as key levers for improving the accuracy--efficiency trade-off in TFMs.
  arXiv  Detail & Related papers  (2026-06-03T06:07:33Z)
• AdaSplash-2: Faster Differentiable Sparse Attention [14.338708749838881]
  We introduce AdaSplash-2, which reduces the number of iterations needed to compute $$ to typically 1--2.<n>AdaSplash-2 matches or improves per-step training time relative to FlashAttention-2 when block sparsity is moderate-to-high.<n>On downstream tasks, models trained with our efficient $$-entmax attention match softmax baselines at short-context lengths and achieve substantial gains in long-context settings.
  arXiv  Detail & Related papers  (2026-04-16T16:03:13Z)
• FASA: Frequency-aware Sparse Attention [56.26881872333624]
  We propose FASA, a novel framework that achieves query-aware token eviction by dynamically predicting token importance.<n>Our key finding is that a small, identifiable subset of "dominant" FCs consistently exhibits high contextual agreement with the full attention head.<n>Across a spectrum of long-context tasks, FASA consistently outperforms all token-eviction baselines and achieves near-oracle accuracy.
  arXiv  Detail & Related papers  (2026-02-03T06:09:06Z)
• SAGA: Selective Adaptive Gating for Efficient and Expressive Linear Attention [10.607730369798551]
  We introduce input-adaptive learnable gates to selectively modulate information aggregation into the $KV$ feature map.<n>SAGA achieves a 1.76$times$ improvement in throughput and a 2.69$times$ reduction in peak GPU memory compared to PVT-T.<n>It improves top-1 accuracy by up to 4.4% on the ImageNet dataset, demonstrating both computational efficiency and model effectiveness.
  arXiv  Detail & Related papers  (2025-09-16T08:36:05Z)
• Spark Transformer: Reactivating Sparsity in FFN and Attention [53.221448818147024]
  We introduce Spark Transformer, a novel architecture that achieves a high level of activation sparsity in both FFN and the attention mechanism.<n>This sparsity translates to a 2.5x reduction in FLOPs, leading to decoding wall-time speedups of up to 1.79x on CPU and 1.40x on GPU.
  arXiv  Detail & Related papers  (2025-06-07T03:51:13Z)
• VAMO: Efficient Zeroth-Order Variance Reduction for SGD with Faster Convergence [6.574641780732972]
  Large-scale non problems are common in deep learning.<n>First-order (FO)s serve as today's baselines.<n>ZO algorithms reduce the size of computation and memory costs.<n>VAMO achieves these gains with smaller dynamic memory requirements.
  arXiv  Detail & Related papers  (2025-05-20T05:31:15Z)
• Second-order Optimization of Gaussian Splats with Importance Sampling [51.95046424364725]
  3D Gaussian Splatting (3DGS) is widely used for novel view rendering due to its high quality and fast inference time.<n>We propose a novel second-order optimization strategy based on Levenberg-Marquardt (LM) and Conjugate Gradient (CG)<n>Our method achieves a $3times$ speedup over standard LM and outperforms Adam by $6times$ when the Gaussian count is low.
  arXiv  Detail & Related papers  (2025-04-17T12:52:08Z)
• Hamming Attention Distillation: Binarizing Keys and Queries for Efficient Long-Context Transformers [18.469378618426294]
  We introduce Hamming Attention Distillation (HAD), a framework that binarizes keys and queries in the attention mechanism to achieve significant efficiency gains.<n>We implement HAD in custom hardware simulations, demonstrating superior performance characteristics compared to a custom hardware implementation of standard attention.
  arXiv  Detail & Related papers  (2025-02-03T19:24:01Z)
• Squeezed Attention: Accelerating Long Context Length LLM Inference [61.787865959140994]
  We propose Squeezed Attention to accelerate applications where a large portion of the input context is fixed.<n>During inference, we compare query tokens from the user input with the centroids to predict which keys from the fixed context are semantically relevant.<n>We also present a hierarchical version of our algorithm which can reduce the complexity of attention from linear to logarithmic with respect to the fixed context length.
  arXiv  Detail & Related papers  (2024-11-14T18:54:19Z)
• A Training-free Sub-quadratic Cost Transformer Model Serving Framework With Hierarchically Pruned Attention [43.211427581302715]
  We propose Hierarchically Pruned Attention (HiP) to increase context length in large language models.<n>HiP reduces the time complexity of the attention mechanism to $O(T log T)$ and the space complexity to $O(T)$, where $T$ is the sequence length.<n>We show that HiP significantly reduces both prefill and decoding latencies, as well as memory usage, while maintaining high-quality generation with minimal degradation.
  arXiv  Detail & Related papers  (2024-06-14T08:32:45Z)
• p-Laplacian Transformer [7.2541371193810384]
  $p$-Laplacian regularization, rooted in graph and image signal processing, introduces a parameter $p$ to control the regularization effect on these data. We first show that the self-attention mechanism obtains the minimal Laplacian regularization. We then propose a novel class of transformers, namely the $p$-Laplacian Transformer (p-LaT)
  arXiv  Detail & Related papers  (2023-11-06T16:25:56Z)

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.