推理加速论文趋势 2023–2026

主要结论

• 推理加速论文从 2023 年的 54 篇增长到 2026 年的 91 篇，三年间扩大到约 1.7 倍。
• NeurIPS 是发推理加速论文最多的会议，2023–2026 年共 90 篇，占四个会议总量的 35%。
• 交叉方向里涨得最快的是深度学习 · large language models：同时属于两个主题的论文从 2023 年的 1 篇增至 2026 年的 9 篇。
• 在「效率与系统」的 5 个主题里，它的 2026 年论文数排第 5；同组增长最快的硬件协同是 4.4 倍，推理加速是 1.7 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 2026A Method-Class Divide in Sub-4-Bit Quantization:\\Iterative vs.\ Single-Pass Sensitivity to Data Composition
• ICML 2026Consistency Deep Equilibrium Models
• ICLR 2026OrderDP: A Theoretically Guaranteed Lossless Dynamic Data Pruning Framework
• CVPR 2026GeoRK2 Geometry-Guided Runge-Kutta Integration for Diffusion Transformer Acceleration
• NeurIPS 2025PolarQuant: Leveraging Polar Transformation for Key Cache Quantization and Decoding Acceleration
• ICML 2025RePaViT: Scalable Vision Transformer Acceleration via Structural Reparameterization on Feedforward Network Layers
• ICLR 2025Progressive distillation induces an implicit curriculum
• CVPR 2025Adapting Dense Matching for Homography Estimation with Grid-based Acceleration
• NeurIPS 2024Provable Acceleration of Nesterov's Accelerated Gradient for Asymmetric Matrix Factorization and Linear Neural Networks
• ICML 2024Diversified Batch Selection for Training Acceleration
• ICLR 2024Embarrassingly Simple Dataset Distillation
• CVPR 2024Data-Free Quantization via Pseudo-label Filtering
• NeurIPS 2023Continuous-time Analysis of Anchor Acceleration
• ICML 2023Auxiliary Modality Learning with Generalized Curriculum Distillation
• ICLR 2023Accurate Neural Training with 4-bit Matrix Multiplications at Standard Formats
• CVPR 2023NIPQ: Noise Proxy-Based Integrated Pseudo-Quantization
• NeurIPS 2026AcceleGrad#: Adaptive Geometry-Aware Acceleration
• ICML 2026PsumQuant: In-line Post-training Partial Sum Quantizer for Energy Efficient NPU Inference
• ICLR 2026NLI : Non-uniform Linear Interpolation Approximation of Nonlinear Operations for Efficient LLMs Inference
• CVPR 2026SODA Sensitivity-Oriented Dynamic Acceleration for Diffusion Transformer

全部推理加速论文

共 91 篇，第 2 / 2 页

51. ICML 2026RSTR: Reducing SpatioTemporal Redundancy in Diffusion Transformers
52. ICML 2026RT-Lynx: Putting the GEMM Sparsity In a Right Way for Diffusion Models
53. ICML 2026Semantic Cache Distillation: Efficient State Transfer via Reuse and Selective Patching
54. ICML 2026SparseInfer: Accelerating Large Language Model Inference with Semantics-Inspired Adaptive Sparse Activation
55. ICML 2026UniSVQ: 2-bit Unified Scalar-Vector Quantization
56. ICML 2026WaterSIC: Information-Theoretically (Near) Optimal \\Linear Layer Quantization
57. ICML 2026WAVE: Window-Aware Vocabulary-Efficient Early-Exit for Training-Free LLM Acceleration
58. ICML 2026XSpecMesh: Quality-Preserving Auto-Regressive Mesh Generation Acceleration via Multi-Head Speculative Decoding
59. NeurIPS 2026A Method-Class Divide in Sub-4-Bit Quantization:\\Iterative vs.\ Single-Pass Sensitivity to Data Composition
60. NeurIPS 2026AcceleGrad#: Adaptive Geometry-Aware Acceleration
61. NeurIPS 2026AccelEval: A Whole-Program Benchmark for LLM-Generated CPU-to-GPU Code Acceleration
62. NeurIPS 2026Adaptive Conditional Gradient Sliding: Projection-Free and Line-Search-Free Acceleration
63. NeurIPS 2026AlphaQ: Calibration-Free Bit Allocation for Mixture-of-Experts Quantization
64. NeurIPS 2026Benchmark Recovery Does Not Certify a Frozen Tool-Trigger Contract After Quantization
65. NeurIPS 2026Block Sphere Vector Quantization
66. NeurIPS 2026Cache the Future: Training-Free Self-Revision for Diffusion Transformer Acceleration
67. NeurIPS 2026Complexity of Classical Acceleration for $\ell_1$-Regularized PageRank
68. NeurIPS 2026Direct Acceleration of Stochastic Root-Finding Without Variance Reduction and Regularization
69. NeurIPS 2026Disentangling the Good From the Bad: Quantization-Induced Flips Are Not Random
70. NeurIPS 2026Facts Don't Speak Louder than Words: The Behavioral Essence of Long-CoT Distillation
71. NeurIPS 2026FADE: Fractional Anomalous Dynamics Extrapolation for Training-Free Diffusion Transformers Acceleration
72. NeurIPS 2026FlashControl: One-Step Controllable Generator via Distillation-Friendly Single-Stream Teachers.
73. NeurIPS 2026GDMD: Guiding Distribution Matching Distillation with Gradient-Based Reinforcement Learning
74. NeurIPS 2026InQuant: In-Place Mixed-Precision KV Cache Quantization via Saliency-Aware Neighbor-Slot Reuse
75. NeurIPS 2026Locking Pretrained Weights via Deep Low-Rank Residual Distillation
76. NeurIPS 2026Logical Distillation of Transformer Encoders
77. NeurIPS 2026Matched-Control Tests of Partition-Source Claims in One Routed Distillation Family
78. NeurIPS 2026MPQ: A Message-Passing View of Post-Training Quantization
79. NeurIPS 2026On the Memorization of Consistency Distillation for Diffusion Models
80. NeurIPS 2026Perfect Parallelization in Mini-Batch SGD with Classical Momentum Acceleration
81. NeurIPS 2026Provable Quantization with Randomized Hadamard Transform
82. NeurIPS 2026RaZeR: Pushing the Limits of NVFP4 Quantization with Redundant Zero Remapping
83. NeurIPS 2026Rethinking Gradient Approximation in Quantization: A Zeroth-Order Expectation Perspective
84. NeurIPS 2026SHRED: Retain-Set-Free Unlearning via Self-Distillation with Logit Demotion
85. NeurIPS 2026The Quantization Benefits of Residual-Free Transformers
86. NeurIPS 2026Tikhonov-Stabilized Bezier Representation Forecasting for Training-free Diffusion Acceleration
87. NeurIPS 2026Total Variation Distance Estimation in Autoregressive Models
88. NeurIPS 2026Towards Understanding Momentum Acceleration in River-Valley Loss Landscape
89. NeurIPS 2026Transferable SCF-Acceleration through Solver-Aligned Initialization Learning
90. NeurIPS 2026Transporting Quantiles to Codebook: Scalable Vector Quantization without Codebook Collapse
91. NeurIPS 2026UniPrefill: Universal Long-Context Prefill Acceleration via Block-wise Dynamic Sparsification

常见问题

2023–2026 年四大会议一共有多少篇推理加速论文？

共 259 篇，按年份分别是 2023 年 54 篇、2024 年 50 篇、2025 年 64 篇、2026 年 91 篇。

NeurIPS 2026 有多少篇推理加速论文？

NeurIPS 2026 有 33 篇推理加速论文，上一届是 15 篇。

ICLR 2026 有多少篇推理加速论文？

ICLR 2026 有 17 篇推理加速论文，上一届是 18 篇。

ICML 2026 有多少篇推理加速论文？

ICML 2026 有 26 篇推理加速论文，上一届是 22 篇。

CVPR 2026 有多少篇推理加速论文？

CVPR 2026 有 15 篇推理加速论文，上一届是 9 篇。

哪个会议的推理加速论文最多？

NeurIPS，2023–2026 年共 90 篇。