稀疏加速论文趋势 2023–2026

主要结论

• 稀疏加速论文从 2023 年的 54 篇增长到 2026 年的 89 篇，三年间扩大到约 1.6 倍。
• NeurIPS 是发稀疏加速论文最多的会议，2023–2026 年共 80 篇，占四个会议总量的 31%。
• 交叉方向里涨得最快的是多模态大模型：同时属于两个主题的论文从 2023 年的 0 篇增至 2026 年的 13 篇。
• 在「硬件与编译」的 5 个主题里，它的 2026 年论文数排第 1；同组增长最快的端侧推理是 4.0 倍，稀疏加速是 1.6 倍。

代表论文

按年份和会议均匀挑选，方便快速看到每一届的样貌

• NeurIPS 2026AC/DC on a Budget -- Alternating Sparse Phases
• ICML 2026An Evidential Route to Asymptotic Bayes Optimality under Sparsity
• ICLR 2026Hilbert-Guided Sparse Local Attention
• CVPR 2026Improving Sparse Autoencoder with Dynamic Attention
• NeurIPS 2025Bilevel Network Learning via Hierarchically Structured Sparsity
• ICML 2025COGNATE: Acceleration of Sparse Tensor Programs on Emerging Hardware using Transfer Learning
• ICLR 2025Beyond Worst-Case Dimensionality Reduction for Sparse Vectors
• CVPR 2025ICP: Immediate Compensation Pruning for Mid-to-high Sparsity
• NeurIPS 2024ELSA: Exploiting Layer-wise N:M Sparsity for Vision Transformer Acceleration
• ICML 2024Accelerating Transformer Pre-training with 2:4 Sparsity
• ICLR 2024Adaptive Sharpness-Aware Pruning for Robust Sparse Networks
• CVPR 2024Masked Spatial Propagation Network for Sparsity-Adaptive Depth Refinement
• NeurIPS 2023Dynamic Sparsity Is Channel-Level Sparsity Learner
• ICML 2023Conformalization of Sparse Generalized Linear Models
• ICLR 2023Bit-Pruning: A Sparse Multiplication-Less Dot-Product
• CVPR 2023Adaptive Channel Sparsity for Federated Learning Under System Heterogeneity
• NeurIPS 2026Beyond Masked Sparsity: SNACK Enables Truly Sparse Neural Networks on GPU
• ICML 2026FlashSketch: Sketch-Kernel Co-Design for Fast Sparse Sketching on GPUs
• ICLR 2026Identifiability Challenges in Sparse Linear Ordinary Differential Equations
• CVPR 2026Revisiting Pose Sensitivity in Splat-based Computed Tomography under Sparse-view Reconstruction

全部稀疏加速论文

共 75 篇，第 1 / 2 页

1. ICML 2026AdaSplash-2: Faster Differentiable Sparse Attention
2. ICML 2026An Evidential Route to Asymptotic Bayes Optimality under Sparsity
3. ICML 2026AsyncSpade: Efficient Test-Time Scaling with Asynchronous Sparse Decoding
4. ICML 2026Attention Sparsity is Input-Stable: Training-Free Sparse Attention for Video Generation via Offline Sparsity Profiling and Online QK Co-Clustering
5. ICML 2026Column Thresholding for Sparse Spiked Wigner Models: Improved Signal Strength Requirements
6. ICML 2026DC-Leap: Training-Free Acceleration of dLLMs via Draft-Guided Contiguous Leaping Decoding
7. ICML 2026Elastic Attention: Test-time Adaptive Sparsity Ratios for Efficient Transformers
8. ICML 2026Faster Than Flash: Exploiting Attention Sparsity for Efficient Long-Context Decoding
9. ICML 2026FlashSketch: Sketch-Kernel Co-Design for Fast Sparse Sketching on GPUs
10. ICML 2026HASTE: Hardware-Aware Dynamic Sparse Training for Large Output Spaces
11. ICML 2026Incremental Learning of Sparse Attention Patterns in Transformers
12. ICML 2026Lie-Algebraic Acceleration of Neural Koopman Dynamics
13. ICML 2026MACKO: Sparse matrix-vector multiplication for low sparsity
14. ICML 2026Memory-Efficient LLM Training with Dynamic Sparsity: From Stability to Practical Scaling
15. ICML 2026Narrowing the ANN–SNN Gap for Continuous 1D Temporal Signal Classification with Multi-Scale Temporal Encoding and Sparsity-Regularized Transform Encoding
16. ICML 2026Prism: Spectral-Aware Block-Sparse Attention
17. ICML 2026RSTR: Reducing SpatioTemporal Redundancy in Diffusion Transformers
18. ICML 2026RT-Lynx: Putting the GEMM Sparsity In a Right Way for Diffusion Models
19. ICML 2026Size Transferability of Graph Convolutional Networks across Sparsity: A Generalized Graphon Perspective
20. ICML 2026SlideSparse: Fast and Flexible (2N-2):2N Structured Sparsity
21. ICML 2026SparseOpt: Addressing Normalization-induced Gradient Skew in Sparse Training
22. ICML 2026Sparser Block-Sparse Attention via Token Permutation
23. ICML 2026Weight-sparse transformers have interpretable circuits
24. ICML 2026When Does Sparsity Mitigate the Curse of Depth in LLMs
25. ICML 2025Accelerating Linear Recurrent Neural Networks for the Edge with Unstructured Sparsity
26. ICML 2025AdaSplash: Adaptive Sparse Flash Attention
27. ICML 2025BAME: Block-Aware Mask Evolution for Efficient N:M Sparse Training
28. ICML 2025Be a Goldfish: Forgetting Bad Conditioning in Sparse Linear Regression via Variational Autoencoders
29. ICML 2025COGNATE: Acceleration of Sparse Tensor Programs on Emerging Hardware using Transfer Learning
30. ICML 2025CoreMatching: A Co-adaptive Sparse Inference Framework with Token and Neuron Pruning for Comprehensive Acceleration of Vision-Language Models
31. ICML 2025Determining Layer-wise Sparsity for Large Language Models Through a Theoretical Perspective
32. ICML 2025Dynamic Sparse Training of Diagonally Sparse Networks
33. ICML 2025Fast Inference with Kronecker-Sparse Matrices
34. ICML 2025FlashTP: Fused, Sparsity-Aware Tensor Product for Machine Learning Interatomic Potentials
35. ICML 2025Grokking Beyond the Euclidean Norm of Model Parameters
36. ICML 2025Lexico: Extreme KV Cache Compression via Sparse Coding over Universal Dictionaries
37. ICML 2025Metadata Conditioning Accelerates Language Model Pre-training
38. ICML 2025Network Sparsity Unlocks the Scaling Potential of Deep Reinforcement Learning
39. ICML 2025Pivoting Factorization: A Compact Meta Low-Rank Representation of Sparsity for Efficient Inference in Large Language Models
40. ICML 2025Position: A Theory of Deep Learning Must Include Compositional Sparsity
41. ICML 2025PROXSPARSE: REGULARIZED LEARNING OF SEMI-STRUCTURED SPARSITY MASKS FOR PRETRAINED LLMS
42. ICML 2025SpargeAttention: Accurate and Training-free Sparse Attention Accelerating Any Model Inference
43. ICML 2025Sparse Video-Gen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity
44. ICML 2025Sparsing Law: Towards Large Language Models with Greater Activation Sparsity
45. ICML 2025When Dynamic Data Selection Meets Data Augmentation: Achieving Enhanced Training Acceleration
46. ICML 2025XAttention: Block Sparse Attention with Antidiagonal Scoring
47. ICML 2024Accelerating Transformer Pre-training with 2:4 Sparsity
48. ICML 2024Better Locally Private Sparse Estimation Given Multiple Samples Per User
49. ICML 2024Block Acceleration Without Momentum: On Optimal Stepsizes of Block Gradient Descent for Least-Squares
50. ICML 2024Distributional Bellman Operators over Mean Embeddings

常见问题

2023–2026 年四大会议一共有多少篇稀疏加速论文？

共 260 篇，按年份分别是 2023 年 54 篇、2024 年 53 篇、2025 年 64 篇、2026 年 89 篇。

NeurIPS 2026 有多少篇稀疏加速论文？

NeurIPS 2026 有 29 篇稀疏加速论文，上一届是 21 篇。

ICLR 2026 有多少篇稀疏加速论文？

ICLR 2026 有 24 篇稀疏加速论文，上一届是 16 篇。

ICML 2026 有多少篇稀疏加速论文？

ICML 2026 有 24 篇稀疏加速论文，上一届是 22 篇。

CVPR 2026 有多少篇稀疏加速论文？

CVPR 2026 有 12 篇稀疏加速论文，上一届是 5 篇。

哪个会议的稀疏加速论文最多？

NeurIPS，2023–2026 年共 80 篇。