| Challenge: | Existing block-granularity sparsification can reduce latency, but coarse blocks impose an intrinsic sparsity ceiling. |
| Approach: | They propose a method that performs early stopping for sparse attention via online permutation. |
| Outcome: | The proposed approach reduces the complexity of the model and its performance. |
Similar Papers
From Logical to Computational Sparsity: Structure-Aware Block-Sparse Attention for Long-Code Completion (2026.acl-long)
Copied to clipboard
| Challenge: | Existing sparse attention methods for long-context generation pose high latency . general sparsity methods cause excessive accuracy degradation without considering code structure . |
| Approach: | They propose a training-free **S**tructure-**a**ware **b**lock-spa**r**s**e** attention mechanism that bridges the gap between logical and computational sparsity. |
| Outcome: | The proposed method reduces TTFT by 45-55% while maintaining accuracy within 3% of dense attention. |
The Sparse Frontier: Sparse Attention Trade-offs in Transformer LLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Sparse attention is a promising strategy to extend long-context capabilities in LLMs . but its efficiency–accuracy trade-offs remain unclear due to the lack of comprehensive evaluation . |
| Approach: | They evaluate sparse attention methods across multiple model families and sizes . they find larger sparser models outperform smaller dense ones at equivalent cost . |
| Outcome: | The proposed methods outperform smaller sparse models at equivalent cost and improve the Pareto frontier. |
Accelerating Prefilling via Decoding-time Contribution Sparsity (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing acceleration methods exploit attention score sparsity by estimating blocks with high attention scores and applying dynamic sparse attention. |
| Approach: | They propose a method which replaces dense attention with Triangle attention in a subset of layers to reduce the time needed to decode. |
| Outcome: | Experiments show that TriangleMix achieves near-lossless performance on long-context and long-constrast reasoning benchmarks while significantly improving efficiency. |
Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention (2025.acl-long)
Copied to clipboard
Jingyang Yuan, Huazuo Gao, Damai Dai, Junyu Luo, Liang Zhao, Zhengyan Zhang, Zhenda Xie, Yuxing Wei, Lean Wang, Zhiping Xiao, Yuqing Wang, Chong Ruan, Ming Zhang, Wenfeng Liang, Wangding Zeng
| Challenge: | Long-context modeling is crucial for next-generation language models, but high computational cost of standard attention mechanisms poses significant computational challenges. |
| Approach: | They propose a natively trained Sparse Attention mechanism that integrates algorithms with hardware-aligned optimizations to achieve efficient long-context modeling. |
| Outcome: | The proposed model maintains or exceeds Full Attention models across general benchmarks, long-context tasks, and instruction-based reasoning. |
Evolving Sparsity: Leveraging Token Importance Dynamics for Efficient LLM Decoding with Sparse Attention (2026.acl-long)
Copied to clipboard
| Challenge: | Efficient long-context inference remains a major challenge for large language models (LLMs), as the cost of attention computation during auto-regressive decoding grows linearly with the context length. |
| Approach: | They propose to model token importance as a dynamic process that evolves over decoding steps and propagates through model layers. |
| Outcome: | The proposed method outperforms baseline sparse attention methods and achieves speedups of up to 5.36 for attention latency and 2.33 for end-to-end decoding. |
Weight-Aware Activation Sparsity with Constrained Bayesian Optimization Scheduling for Large Language Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing activation sparsification methods rely on activation magnitude and weights for sparsity . authors propose a weight-aware activation-a-ware framework for large language models . |
| Approach: | They propose a weight-aware activation sparsity framework that uses weight-based scoring to measure activation importance in sparsification and a custom GPU sparse kernel to support it. |
| Outcome: | The proposed framework outperforms existing methods at 60% model-level sparsity and significantly outperfies them at higher sparsities. |
Lil: Less is Less When Applying Post-Training Sparse-Attention Algorithms in Long-Decode Stage (2026.findings-acl)
Copied to clipboard
Junhao Hu, Fangze Li, Mingtao Xu, Feifan Meng, Shiju Zhao, Tiancheng Hu, Ting Peng, Anmin Liu, Wenrui Huang, Chenxu Liu, Ziyue Hua, Tao Xie
| Challenge: | Prior work typically decomposes inference into prefill and decode stages, with the decode stage dominating total latency. |
| Approach: | They propose an algorithm that detects threshold where information loss exceeds information gain during sparse decoding to reduce token consumption by up to 90% and a marginal accuracy degradation of less than 2%. |
| Outcome: | The proposed algorithm reduces token consumption by 90% with a marginal accuracy degradation of less than 2% across reasoning-intensive benchmarks. |
Octopus: Gated Selective Attention for Memory-Bounded Long-Context Inference in Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Subquadratic architectures rely on aggressive state compression that degrades performance on complex reasoning tasks. |
| Approach: | They propose a framework that confers fixed-memory inference onto pretrained Transformers . they use a learnable module that enforces an adaptive sparsity policy over the context history . |
| Outcome: | The proposed framework outperforms state-of-the-art linearized baselines on the GSM8K benchmark by over 36 points under identical memory constraints. |
Optimizing Native Sparse Attention with Latent Attention and Local Global Alternating Strategies (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing research has proposed a variety of training-free and post-training methods for selecting critical key-value pairs at each generation step. |
| Approach: | They propose to use local (sliding-window) and global (compression/selective) attention across layers to enlarge long-context modeling. |
| Outcome: | Experiments on models from 340M to 1.3B parameters show that the proposed method matches or exceeds full attention and native sparse attention in both common-sense reasoning and long-context understanding tasks. |
Efficient Content-Based Sparse Attention with Routing Transformers (2021.tacl-1)
Copied to clipboard
| Challenge: | Self-attention suffers from quadratic computation and memory requirements with respect to sequence length . despite its effectiveness, self-attention models suffer from quadratic computation and a limited set of locations . |
| Approach: | They propose to learn dynamic sparse attention patterns that avoid allocating computation and memory to attend to content unrelated to the query of interest. |
| Outcome: | The proposed model outperforms similar sparse attention models on language modeling and image generation on Wikitext-103 . |