迈向高效AI:边缘侧模型压缩技术综述

doi:10.19678/j.issn.1000-3428.0260431

摘要/Abstract

摘要： 大语言模型(LLM)参数规模的激增与边缘终端受限的物理资源间存在结构性矛盾,制约了其规模化落地。传统的云端集中式推理高度依赖网络传输,面临较高的通信延迟,难以满足自动驾驶、智能医疗等场景对极低延迟与严格数据隐私保护的双重诉求。然而,边缘物理硬件从微控制器到边缘网关,具有极大的异构性,云侧通用压缩方案难以直接平移。为此,本文立足于边缘设备的异构物理约束,系统性地综述了面向边缘侧的大模型高效压缩与软硬件协同部署技术体系。首先,本文剖析了模型量化、参数剪枝与知识蒸馏(KD)三大核心压缩技术在边缘场景下的底层机制。在量化方面,训练后量化(PTQ)虽具备部署敏捷性,但面临LLM长尾激活异常导致的表征坍塌难题;量化感知训练(QAT)虽具有一定鲁棒性,却受制于边缘重训算力匮乏的瓶颈。在剪枝方面,本文论证了结构化剪枝在受限访存带宽硬件上的实际能效优势,指出非结构化剪枝的高理论压缩率易被通用边缘芯片的索引寻址开销所抵消。在蒸馏方面,传统浅层参数对齐在跨越教师与边缘学生模型间的容量鸿沟时,存在特征丢失与偏见放大风险。综合来看,单一压缩技术在极端约束下已呈现出明显的边际收益递减效应。其次,为缓解单一技术的性能瓶颈,本文归纳了受模型架构与物理场景双重驱动的多级混合压缩范式。系统梳理了3类核心优化链路:追求较高物理压缩率的串行流水线策略,适配边缘网关的实时推理;面向能效与精度严格折中的深度耦合联合优化流程,将量化、剪枝与低秩分解置于同一框架内同步更新,适配功耗受限的移动终端;面向大参数模型部署的蒸馏驱动机制,利用教师先验引导结构重塑与量化。该多级范式有效拓展了模型规模、计算功耗与保真度之间的多维权衡空间。进一步地,面对跨度极大的算力与能耗阶梯,本文构建了"系统-模型-算子-指令"4层软硬件协同设计机制。明确指出协同优化重心需依据物理基座约束进行动态转移:系统级侧重云边环境的资源感知调度与任务分发;模型级依赖硬件感知架构搜索实现结构自适应;算子级推进跨层融合与访存局部性重构;指令级则聚焦特定微架构(如RISC-V)的定制扩展指令,以精准控制底层能耗。结合模型转换、编译重构与内存管理(如SwapNet)的全链条部署流程,该机制实现了压缩算法向底层物理执行的有效映射,提升了异构算力的综合利用效率。最后,本文前瞻性地指出了边缘智能轻量化领域的未来研究挑战。强调超低位宽(≤4 bit)的鲁棒性补偿机制、硬件自适应动态半结构化剪枝,以及对大模型深层逻辑推理的有效知识转移,是克服当前轻量化瓶颈的核心方向。同时,亟需依托深度学习编译器构建硬件无感的统一工具链,消除碎片化异构设备的部署壁垒。本文通过体系化的技术梳理,为发展低延迟且强隐私保护的边缘智能生态提供了坚实的理论支撑与参考指南。

关键词: 边缘侧大模型, 量化与剪枝, 知识蒸馏, 混合压缩, 软硬件协同设计

Abstract: A considerable structural contradiction exists between the surge in the parameter scale of Large Language Models (LLMs) and the limited physical resources of edge terminals, which restricts their large-scale deployment. Traditional cloud-centric inference relies heavily on network transmission and suffers from high communication latency, making it difficult to meet the dual demands of ultra-low latency and strict data privacy protection in scenarios such as autonomous driving and intelligent healthcare. Meanwhile, edge physical hardware, ranging from microcontrollers to edge gateways, exhibits significant heterogeneity, making it difficult to transfer of general cloud-side compression schemes directly. To this end, grounded in the heterogeneous physical constraints of edge devices, this study systematically reviews the technical system of efficient compression and software—hardware co-deployment of large models on the edge. First, this study analyzes the underlying mechanisms of three core compression techniques, namely model quantization, parameter pruning, and Knowledge Distillation (KD), specifically in edge scenarios. Although Post-Training Quantization (PTQ) offers deployment agility, it faces the challenge of representation collapse caused by abnormal long-tail activations in large language models. While more robust, Quantization-Aware Training (QAT) is constrained by the scarcity of retrained computations at the edge. This study demonstrates the practical energy efficiency advantages of structured pruning on hardware with constrained memory bandwidth, highlighting that the high theoretical compression ratio of unstructured pruning is easily offset by the indexing and addressing overhead on general-purpose edge chips. Traditional shallow parameter alignment risks feature loss and bias amplification when bridging the capacity gap between teacher and edge student models. In summary, the single-compression techniques exhibit clear diminishing returns under extreme constraints. Second, to alleviate the performance bottlenecks of individual techniques, this study generalizes a multi-stage hybrid compression paradigm driven jointly by model architecture and physical scenarios. The three core optimization pipelines are systematically categorized as follows: 1) a serial pipeline strategy that pursues a high physical compression rate and suits real-time inference on edge gateways; 2) a deeply coupled joint optimization flow that strictly balances energy efficiency and accuracy by integrating quantization, pruning, and low-rank decomposition within a single framework for synchronous updates, which is suitable for power-constrained mobile terminals; and 3) a distillation-driven mechanism for deploying large-parameter models that uses teacher priors to guide structural reshaping and quantization. This multi-stage paradigm effectively expands the multi-dimensional trade-off space among model scale, computational power consumption, and fidelity. Furthermore, facing a broad spectrum of computing power and energy consumption levels, this study constructs a four-layer software—hardware co-design mechanism comprising the system, model, operator, and instruction levels. It clearly states that the focus of co-optimization must shift dynamically according to physical substrate constraints. At the system level, it focuses on resource-aware scheduling and task distribution in cloud—edge environments; at the model level, it relies on hardware-aware neural architecture search to achieve structural adaptation; at the operator level, it promotes cross-layer fusion and memory access locality restructuring; and at the instruction level, it targets custom extended instructions for specific microarchitectures (e.g., RISC-V) to precisely control underlying energy consumption. Together with the full-chain deployment process comprising model conversion, compilation reengineering, and memory management (e.g., SwapNet), this mechanism effectively maps compression algorithms to the underlying physical execution and improves the comprehensive utilization efficiency of heterogeneous computing resources. Finally, future research challenges in the field of edge intelligence lightweighting are prospectively identified. It emphasizes that robust compensation mechanisms for ultra-low bitwidth (≤4 bit), hardware-adaptive dynamic semi-structured pruning, and effective knowledge transfer for deep logical reasoning in large models are core directions for overcoming current lightweighting bottlenecks. Simultaneously, a hardware-agnostic unified toolchain based on deep-learning compilers must be built to eliminate deployment barriers caused by fragmented heterogeneous devices. This systematic technical review provides solid theoretical support and a reference guide for developing a low-latency, strong-privacy protection edge intelligence ecosystem.

Key words: edge-side large model, quantization and pruning, knowledge distillation, hybrid compression, software—hardware co-design

中图分类号:

TP393

范兴刚, 时雪刚, 廖思腾, 赵依依, 梁玉珠, 王田. 迈向高效AI:边缘侧模型压缩技术综述[J]. 计算机工程, 2026, 52(7): 1-21.

FAN Xinggang, SHI Xuegang, LIAO Siteng, ZHAO Yiyi, LIANG Yuzhu, WANG Tian. Towards Efficient AI: A Survey of Model Compression Techniques on the Edge[J]. Computer Engineering, 2026, 52(7): 1-21.

https://www.ecice06.com/CN/Y2026/V52/I7/1

参考文献

[1] DEEPSEEK-AI, GUO D Y, YANG D J, et al. DeepSeek-R1: incentivizing reasoning capability in LLMs via reinforcement learning[EB/OL].[2026-03-07]. https://arxiv.org/abs/2501.12948. HUI B Y, YANG J, CUI Z Y, et al. Qwen2.5-Coder technical report[EB/OL].[2026-03-07]. https://arxiv.org/abs/2409.12186.
[3] LIANG Y Z, YIN M J, WANG W H, et al. Collaborative edge server placement for maximizing QoS with distributed data cleaning[J]. IEEE Transactions on Services Computing, 2025, 18(3): 1321-1335.
[4] YE X. calflops: a FLOPs and params calculate tool for neural networks[EB/OL].[EB/OL].[2026-03-07]. https://github.com/MrYxJ/calculate-flops.pytorch. CHOUKROUN Y, KRAVCHIK E, YANG F, et al. Low-bit quantization of neural networks for efficient inference[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision Workshop (ICCVW). Washington D.C., USA: IEEE Press, 2019: 3009-3018. THIRUNAVUKARASU A J, TING D S J, ELANGOVAN K, et al. Large language models in medicine[J]. Nature Medicine, 2023, 29(8): 1930-1940.
[7] 鲁思迪, 何元恺, 施巍松. 车计算: 自动驾驶时代的新型计算范式[J]. 计算机研究与发展, 2025, 62(1): 2-21. LU S D, HE Y K, SHI W S. Vehicle computing: an emerging computing paradigm for the autonomous driving era[J]. Journal of Computer Research and Development, 2025, 62(1): 2-21. (in Chinese)
[8] XIAO G X, LIN J, SEZNEC M, et al. SmoothQuant: accurate and efficient post-training quantization for large language models[EB/OL].[2026-03-07]. https://arxiv.org/abs/2211.10438.
[9] GUO J L, CHEN X H, TANG Y H, et al. SlimLLM: accurate structured pruning for large language models[EB/OL].[2026-03-07]. https://arxiv.org/abs/2505.22689.
[10] JOO D, HOSSEINI H, HADIDI R, et al. Mustafar: promoting unstructured sparsity for KV cache pruning in LLM inference[J]. Advances in Neural Information Processing Systems, 2026, 38: 76342-76366.
[11] 邵仁荣, 刘宇昂, 张伟, 等. 深度学习中知识蒸馏研究综述[J]. 计算机学报, 2022, 45(8): 1638-1673. SHAO R R, LIU Y A, ZHANG W, et al. A survey of knowledge distillation in deep learning[J]. Chinese Journal of Computers, 2022, 45(8): 1638-1673. (in Chinese)
[12] 王睿, 齐建鹏, 陈亮, 等. 面向边缘智能的协同推理综述[J]. 计算机研究与发展, 2023, 60(2): 398-414. WANG R, QI J P, CHEN L, et al. Survey of collaborative inference for edge intelligence[J]. Journal of Computer Research and Development, 2023, 60(2): 398-414. (in Chinese)
[13] CHENG H R, ZHANG M, SHI J Q. A survey on deep neural network pruning: taxonomy, comparison, analysis, and recommendations[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024, 46(12): 10558-10578.
[14] QU G Q, CHEN Q Y, WEI W, et al. Mobile edge intelligence for large language models: a contemporary survey[J]. IEEE Communications Surveys[WT《Times New Roman》]& Tutorials, 2025, 27(6): 3820-3860.
[15] ZHENG Y, CHEN Y H, QIAN B, et al. A review on edge large language models: design, execution, and applications[J]. ACM Computing Surveys, 2025, 57(8): 1-35.
[16] HENDRYCKS D, BURNS C, BASART S, et al. Measuring massive multitask language understanding[EB/OL].[2026-03-07]. https://arxiv.org/abs/2009.03300.
[17] WARSTADT A, PARRISH A, LIU H K, et al. BLiMP: the benchmark of linguistic minimal pairs for English[J]. Transactions of the Association for Computational Linguistics, 2020, 8: 377-392.
[18] DENG J, DONG W, SOCHER R, et al. ImageNet: a large-scale hierarchical image database[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Washington D.C., USA: IEEE Press, 2009: 248-255.
[19] LIN T Y, MAIRE M, BELONGIE S, et al. Microsoft COCO: common objects in context[C]//Proceedings of ECCV 2014. Berlin, Germany: Springer, 2014: 740-755.
[20] KRIZHEVSKY A. Learning multiple layers of features from tiny images[EB/OL].[2026-03-07]. https://api.semanticscholar.org/CorpusID:18268744.
[21] ANTOL S, AGRAWAL A, LU J S, et al. VQA: visual question answering[C]//Proceedings of the IEEE International Conference on Computer Vision (ICCV). Washington D.C., USA: IEEE Press, 2015: 2425-2433.
[22] ALABBASY F M, ABOHAMAMA A S, ALRAHMAWY M F. Compressing medical deep neural network models for edge devices using knowledge distillation[J]. Journal of King Saud University (Computer and Information Sciences), 2023, 35(7): 101616.
[23] RAHMAN M W U, ABRAR M M, COPENING H G, et al. Quantized transformer language model implementations on edge devices[C]//Proceedings of the International Conference on Machine Learning and Applications (ICMLA). Washington D.C., USA: IEEE Press, 2023: 709-716.
[24] LIN J, TANG J M, TANG H T, et al. AWQ: activation-aware weight quantization for on-device LLM compression and acceleration[J]. GetMobile: Mobile Computing and Communications, 2025, 28(4): 12-17.
[25] MA Y X, JIN T S, ZHENG X W, et al. OMPQ: orthogonal mixed precision quantization[C]//Proceedings of the AAAI Conference on Artificial Intelligence. Palo Alto, USA: AAAI Press, 2023: 9029-9037.
[26] WANG K, LIU Z J, LIN Y J, et al. HAQ: hardware-aware automated quantization with mixed precision[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach, USA: IEEE Press, 2019: 8604-8612.
[27] ZHANG Z Y, SHAO W Q, GU J W, et al. Differentiable dynamic quantization with mixed precision and adaptive resolution[EB/OL].[2026-03-07]. https://arxiv.org/abs/2106.02295.
[28] HE Y X, LI C L, WU X X, et al. ZeroQuant: efficient and affordable post-training quantization for large-scale transformers[C]//Proceedings of the Advances in Neural Information Processing Systems 35. New Orleans, USA: Neural Information Processing Systems Foundation, Inc. (NeurIPS), 2022: 27168-27183.
[29] FRANTAR E, ASHKBOOS S, HOEFLER T, et al. GPTQ: accurate post-training quantization for generative pre-trained Transformers[EB/OL].[2026-03-07]. https://arxiv.org/abs/2210.17323.
[30] NAGEL M, FOURNARAKIS M, BONDARENKO Y, et al. Overcoming oscillations in quantization-aware training[C]//Proceedings of the International Conference on Machine Learning.[S. l.]: PMLR, 2022: 16318-16330.
[31] JACOB B, KLIGYS S, CHEN B, et al. Quantization and training of neural networks for efficient integer-arithmetic-only inference[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Washington D.C., USA: IEEE Press, 2018: 2704-2713.
[32] SHAO W Q, CHEN M Z, ZHANG Z Y, et al. OmniQuant: omnidirectionally calibrated quantization for large language models[EB/OL].[2026-03-07]. https://arxiv.org/abs/2308.13137.
[33] 王睿, 张留洋, 高志涌, 等. 面向边缘智能的大模型研究进展[J]. 计算机研究与发展, 2025, 62(10): 2565-2582. WANG R, ZHANG L Y, GAO Z Y, et al. Research progress on large models for edge intelligence[J]. Journal of Computer Research and Development, 2025, 62(10): 2565-2582. (in Chinese)
[34] HASAN J. Optimizing large language models through quantization: a comparative analysis of PTQ and QAT techniques[EB/OL].[2026-03-07]. https://arxiv.org/abs/2411.06084.
[35] LIU S Y, LIU Z C, HUANG X J, et al. LLM-FP4: 4-bit floating-point quantized transformers[C]//Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. Singapore: Association for Computational Linguistics, 2023: 592-605.
[36] DING X, LIU X Y, TU Z J, et al. CBQ: cross-block quantization for large language models[EB/OL].[2026-03-07]. https://arxiv.org/abs/2312.07950.
[37] LIU Z C, OGUZ B, ZHAO C S, et al. LLM-QAT: data-free quantization aware training for large language models[C]//Proceedings of the Findings of the Association for Computational Linguistics: ACL 2024. Bangkok, Thailand: Association for Computational Linguistics, 2024: 467-484.
[38] PAN J Y, WANG C C, ZHENG K F, et al. SmoothQuant+: accurate and efficient 4-bit post-training weight quantization for LLM[EB/OL].[2026-03-07]. https://arxiv.org/abs/2312.03788.
[39] HAN I, ZHANG Z L, WANG Z Y, et al. CalibQuant: 1-bit KV cache quantization for multimodal LLMs[EB/OL].[2026-03-07]. https://arxiv.org/abs/2502.14882.
[40] YUAN Z H, XUE C H, CHEN Y Q, et al. PTQ4ViT: post-training quantization for vision Transformers with twin uniform quantization[C]//Proceedings of ECCV 2022. Berlin, Germany: Springer, 2022: 191-207.
[41] WANG L, ZENG C, ZHANG M, et al. Domain aware post training quantization for vision Transformers in deployment[J]. Pattern Recognition, 2025, 171: 112182.
[42] TEAM G, KAMATH A, FERRET J, et al. Gemma3 technical report[EB/OL].[2026-03-07]. https://arxiv.org/abs/2503.19786.
[43] CHEN M Z, SHAO W Q, XU P, et al. EfficientQAT: efficient quantization-aware training for large language models[C]//Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Vienna, Austria: Association for Computational Linguistics, 2025: 10081-10100.
[44] SHEN X, KONG Z, YANG C, et al. EdgeQAT: entropy and distribution guided quantization-aware training for the acceleration of lightweight llms on the edge[EB/OL].[2026-03-07]. https://arxiv.org/html/2402.10787v1
[45] SHEN X, DONG P Y, LU L, et al. Agile-Quant: activation-guided quantization for faster inference of LLMs on the edge[C]//Proceedings of the AAAI Conference on Artificial Intelligence. Palo Alto, USA: AAAI Press, 2024: 18944-18951.
[46] SHEN S, DONG Z, YE J Y, et al. Q-BERT: Hessian based ultra low precision quantization of BERT[C]//Proceedings of the AAAI Conference on Artificial Intelligence. Palo Alto, USA: AAAI Press, 2020: 8815-8821.
[47] BELLUCCI V, BIRNSTEINOVA S, SATO T, et al. Development of crystal optics for multi-projection X-ray imaging for synchrotron and XFEL sources[EB/OL].[2026-03-07]. https://arxiv.org/abs/2402.13262.
[48] TANG S K, SIEBERLING O, KURTIC E, et al. DarwinLM: evolutionary structured pruning of large language models[EB/OL].[2026-03-07]. https://arxiv.org/abs/2502.07780.
[49] HU H Y, ZHAO P X, LI P, et al. FASP: fast and accurate structured pruning of large language models[EB/OL].[2026-03-07]. https://arxiv.org/abs/2501.09412.
[50] SHEN B W, LIN Z, ZHA D R, et al. Pruning large language models to intra-module low-rank architecture with transitional activations[C]//Proceedings of the Findings of the Association for Computational Linguistics: ACL 2024. Bangkok, Thailand: Association for Computational Linguistics, 2024: 9781-9793.
[51] LI Y Q, LI K, YIN X, et al. SepPrune: structured pruning for efficient deep speech separation[C]//Proceedings of the AAAI Conference on Artificial Intelligence. Palo Alto, USA: AAAI Press, 2026: 31861-31869.
[52] SANDER J, MOE D, COHEN A, et al. Constrained edge AI deployment: fine-tuning vs. distillation for LLM compression[C]//Proceedings of the 2025 IEEE Military Communications Conference (MILCOM). Washington D.C., USA: IEEE Press, 2025: 1500-1505.
[53] HUANG W Y, HU Y Z, JIAN G H, et al. Pruning large language models with semi-structural adaptive sparse training[C]//Proceedings of the AAAI Conference on Artificial Intelligence. Palo Alto, USA: AAAI Press, 2025: 24167-24175.
[54] FANG G F, HEINRICH G, KAUTZ J, et al. MaskLLM: learnable semi-structured sparsity for large language models[C]//Proceedings of the Advances in Neural Information Processing Systems 37. Vancouver, Canada: Neural Information Processing Systems Foundation, Inc. (NeurIPS), 2024: 7736-7758.
[55] NETH A, KAUR S, KHAN M N H, et al. UnIT: scalable unstructured inference-time pruning for MAC-efficient neural inference on MCUs[EB/OL].[2026-03-07]. https://arxiv.org/abs/2507.07885.
[56] NETH A. UnIT Pruner: unstructured inference-time pruning for battery-free systems[EB/OL].[2026-03-07]. https://api.semanticscholar.org/CorpusID:278825241.
[57] SHIHAB I F, AKTER S, SHARMA A. Efficient unstructured pruning of mamba state-space models for resource-constrained environments[C]//Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. Suzhou, China: Association for Computational Linguistics, 2025: 11098-11126.
[58] LI C H, BASU B, SHANKER S. LogicSparse: enabling engine-free unstructured sparsity for quantised deep-learning accelerators[EB/OL].[2026-03-07]. https://arxiv.org/abs/2511.03079.
[59] BEARD J C. The sparse data reduction engine: chopping sparse data one byte at a time[C]//Proceedings of the International Symposium on Memory Systems. New York, USA: ACM Press, 2017: 34-48.
[60] LIVANOS M, DAVIDSON I, WONG S. Cooperative knowledge distillation: a learner agnostic approach[C]//Proceedings of the AAAI Conference on Artificial Intelligence. New York, USA: ACM Press, 2024: 14124-14131.
[61] MANSOURIAN A M, AHMADI R, GHAFOURI M, et al. A comprehensive survey on knowledge distillation[EB/OL].[2026-03-07]. https://arxiv.org/abs/2503.12067.
[62] FANG L Y, YU X W, CAI J Z, et al. Knowledge distillation and dataset distillation of large language models: emerging trends, challenges, and future directions[EB/OL].[2026-03-07]. https://arxiv.org/abs/2504.14772.
[63] LI Z H, XU P F, CHANG X J, et al. When object detection meets knowledge distillation: a survey[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 45(8): 10555-10579.
[64]TANGHATARI E, KAMAL M, AFZALI-KUSHA A, et al. Federated learning by employing knowledge distillation on edge devices with limited hardware resources[J]. Neurocomputing, 2023, 531: 87-99.
[65] LEE J, DAS D, HAYAT M, et al. CustomKD: customizing large vision foundation for edge model improvement via knowledge distillation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Washington D.C., USA: IEEE Press, 2025: 25176-25186.
[66] JANG J, MA C F, LEE B. VL2Lite: task-specific knowledge distillation from large vision-language models to lightweight networks[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Washington D.C., USA: IEEE Press, 2025: 30073-30083.
[67] SETYANTO A, BAYU SASONGKO T, FIKRI M A, et al. Knowledge distillation in object detection for resource-constrained edge computing[J]. IEEE Access, 2025, 13: 18200-18214.
[68] LIANG Y Z, YIN M J, ZHANG Y L, et al. Grouping reduces energy cost in directionally rechargeable wireless vehicular and sensor networks[J]. IEEE Transactions on Vehicular Technology, 2023, 72(8): 10840-10851.
[69] 乐俊青, 谭州勇, 张迪, 等. 面向车联网数据持续共享的安全高效联邦学习[J]. 计算机研究与发展, 2024, 61(9): 2199-2212. LE J Q, TAN Z Y, ZHANG D, et al. Secure and efficient federated learning for continuous IoV data sharing[J]. Journal of Computer Research and Development, 2024, 61(9): 2199-2212. (in Chinese)
[70] CHO I, KANG U. Pea-KD: parameter-efficient and accurate knowledge distillation on BERT[J]. PLoS One, 2022, 17(2): e0263592.
[71] CHEN G, CHEN J W, FENG F L, et al. Unbiased knowledge distillation for recommendation[C]//Proceedings of the 16th ACM International Conference on Web Search and Data Mining. New York, USA: ACM Press, 2023: 976-984.
[72] GUO J Y, HAN K, HAO Z W, et al. One-for-all: bridge the gap between heterogeneous architectures in knowledge distillation[C]//Proceedings of the Advances in Neural Information Processing Systems 36. New Orleans, USA: Neural Information Processing Systems Foundation, Inc. (NeurIPS), 2023: 79570-79582.
[73] ZHANG Z R, XU R T, WANG C W, et al. DFMC: feature-driven data-free knowledge distillation[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2025, 35(10): 9681-9694.
[74] XIANG Q L, ZHANG M, SHANG Y Z, et al. DKDM: data-free knowledge distillation for diffusion models with any architecture[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Washington D.C., USA: IEEE Press, 2025: 2955-2965.
[75] VERMA C, TIMMARAJU A S, HSIEH C J, et al. Matryoshka model learning for improved elastic student models[C]//Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining. Washington D.C., USA: IEEE Press, 2025: 4935-4944.
[76] HSIEH C Y, LI C L, YEH C K, et al. Distilling step-by-step! Outperforming larger language models with less training data and smaller model sizes[C]//Proceedings of the Findings of the Association for Computational Linguistics: ACL 2023. Toronto, Canada: Association for Computational Linguistics, 2023: 8003-8017.
[77] DETTMERS T, SVIRSCHEVSKI R, EGIAZARIAN V, et al. SpQR: a sparse-quantized representation for near-lossless LLM weight compression[EB/OL].[2026-03-07]. https://arxiv.org/abs/2306.03078.
[78] KULSHRESTHA A, LIU X Y, USHIJIMA-MWESIGWA H, et al. QAdaPrune: adaptive parameter pruning for training variational quantum circuits[C]//Proceedings of the IEEE International Conference on Quantum Computing and Engineering (QCE). Washington D.C., USA: IEEE Press, 2024: 120-125.
[79] HABIB G, SALEEM T J, KALEEM S M, et al. A comprehensive review of knowledge distillation in computer vision[EB/OL].[2026-03-07]. https://arxiv.org/abs/2404.00936.
[80] NIU T, TENG Y L, JIN L, et al. Pruning-and-distillation: One-stage joint compression framework for CNNs via clustering[J]. Image and Vision Computing, 2023, 136: 104743.
[81] 魏晓辉, 周博文, 李洪亮, 等. 结合算子选择的卷积神经网络显存优化算法[J]. 吉林大学学报(理学版), 2024, 62(2): 302-310. WEI X H, ZHOU B W, LI H L, et al. Memory optimization algorithm for convolutional neural networks with operator selection[J]. Journal of Jilin University (Science Edition), 2024, 62(2): 302-310. (in Chinese)
[82] CHEN Z, WANG W Y, CAO Y, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling[EB/OL].[2026-03-07]. https://arxiv.org/abs/2412.05271.
[83] 李致金, 汤佳辉, 闫金凤. 基于边缘计算的轻量化识别方法[J]. 计算机工程, 2024, 50(6): 287-295. LI Z J, TANG J H, YAN J F. Lightweight recognition method based on edge computing[J]. Computer Engineering, 2024, 50(6): 287-295. (in Chinese)
[84] WEN H, TIAN S Z, PAVLOV B, et al. AutoDroid-V2: boosting SLM-based GUI agents via code generation[C]//Proceedings of the 23rd Annual International Conference on Mobile Systems, Applications and Services. New York, USA: ACM Press, 2025: 223-235.
[85] TIAN X Y, GU J R, LI B L, et al. DriveVLM: the convergence of autonomous driving and large vision-language models[EB/OL].[2026-03-07]. https://arxiv.org/abs/2402.12289.
[86] SAS C, CAPILUPPI A. AutoFL: a tool for automatic multi-granular labelling of software repositories[EB/OL].[2026-03-07]. https://arxiv.org/abs/2408.02557.
[87] YU J M, MA H J, KONG J L. Receipt recognition technology driven by multimodal alignment and lightweight sequence modeling[J]. Electronics, 2025, 14(9): 1717.
[88] WANG H Y, MA S M, WEI F R. BitNet a4.8: 4-bit activations for 1-bit LLMs[EB/OL].[2026-03-07]. https://arxiv.org/abs/2411.04965.
[89] MOTETTI B A, RISSO M, BURRELLO A, et al. Joint pruning and channel-wise mixed-precision quantization for efficient deep neural networks[J]. IEEE Transactions on Computers, 2024, 73(11): 2619-2633.
[90] 宁欣, 赵文尧, 宗易昕, 等. 神经网络压缩联合优化方法的研究综述[J]. 智能系统学报, 2024, 19(1): 36-57. NING X, ZHAO W Y, ZONG Y X, et al. An overview of the joint optimization method for neural network compression[J]. CAAI Transactions on Intelligent Systems, 2024, 19(1): 36-57. (in Chinese)
[91] 张晓东, 张朝昆, 赵继军. 边缘智能研究进展[J]. 计算机研究与发展, 2023, 60(12): 2749-2769. ZHANG X D, ZHANG C K, ZHAO J J. State-of-the-art survey on edge intelligence[J]. Journal of Computer Research and Development, 2023, 60(12): 2749-2769. (in Chinese)
[92] LIU Z Y, CHEN X, WU H, et al. Integrated sensing and edge AI: realizing intelligent perception in 6G[J]. IEEE Communications Surveys[WT《Times New Roman》]& Tutorials, 2026, 28: 2725-2770.
[93] Arm Ltd. Armv9-A architecture overview[EB/OL].[2026-03-07]. https://www.arm.com/architecture/cpu/a-profile/armv9.
[94] WANG T, LIANG Y Z, JIA W J, et al. Coupling resource management based on fog computing in smart city systems[J]. Journal of Network and Computer Applications, 2019, 135: 11-19.
[95] KIM Y, OH C, HWANG J, et al. DACAPO: accelerating continuous learning in autonomous systems for video analytics[C]//Proceedings of the ACM/IEEE 51st Annual International Symposium on Computer Architecture (ISCA). Washington D.C., USA: IEEE Press, 2024: 1246-1261.
[96] BHARDWAJ R, XIA Z X, ANANTHANARAYANAN G, et al. Ekya: continuous learning of video analytics models on edge compute servers[EB/OL].[2026-03-07]. https://arxiv.org/abs/2012.10557.
[97] KONG Y X, YANG P, CHENG Y. Edge-assisted on-device model update for video analytics in adverse environments[C]//Proceedings of the 31st ACM International Conference on Multimedia. New York, USA: ACM Press, 2023: 9051-9060.
[98] TAMBE T, HOOPER C, PENTECOST L, et al. EdgeBERT: sentence-level energy optimizations for latency-aware multi-task NLP inference[C]//Proceedings of the 54th Annual IEEE/ACM International Symposium on Microarchitecture. New York, USA: ACM Press, 2021: 830-844.
[99] ZHOU Y H, WANG Z B, LIU G Y, et al. Squeezing operator performance potential for the ascend architecture[C]//Proceedings of the 30th ACM International Conference on Architectural Support for Programming Languages and Operating Systems. New York, USA: ACM Press, 2025: 1156-1171.
[100] SABIH M, KARIM A, WITTMANN J, et al. Hardware/software co-design of RISC-V extensions for accelerating sparse DNNs on FPGAs[C]//Proceedings of the International Conference on Field Programmable Technology (ICFPT). Washington D.C., USA: IEEE Press, 2024: 1-9.
[101] LEE K H, KWAK J Y, PARK J, et al. Neural network model transformation framework for on-devices[C]//Proceedings of the 14th International Conference on Information and Communication Technology Convergence (ICTC). Washington D.C., USA: IEEE Press, 2023: 1316-1318.
[102] ZHENG G F, LI J N, GAO W, et al. Operator fusion scheduling optimization for TVM deep learning compilers[C]//Proceedings of the 3rd International Symposium on Computer Technology and Information Science (ISCTIS). Washington D.C., USA: IEEE Press, 2023: 273-277.
[103] XU C N, KONG W M, LIU M, et al. Hybrid parallel inference for large model on heterogeneous clusters for high throughput[C]//Proceedings of the IEEE 29th International Conference on Parallel and Distributed Systems (ICPADS). Washington D.C., USA: IEEE Press, 2023: 549-554.
[104] WANG K, CAO J N, ZHOU Z M, et al. SwapNet: efficient swapping for DNN inference on edge AI devices beyond the memory budget[J]. IEEE Transactions on Mobile Computing, 2024, 23(9): 8935-8950.
[105] HUANG W, QIN H T, LIU Y D, et al. SliM-LLM: salience-driven mixed-precision quantization for large language models[EB/OL].[2026-03-07]. https://arxiv.org/abs/2405.14917.
[106] OUYANG X, GE T, HARTVIGSEN T, et al. Low-bit quantization favors undertrained LLMs: scaling laws for quantized LLMs with 100T training tokens[EB/OL].[2026-03-07]. https://arxiv.org/abs/2411.17691.
[107] XU H T, LIU S W, XU Y Y, et al. SparseByteNN: a novel mobile inference acceleration framework based on fine-grained group sparsity[EB/OL].[2026-03-07]. https://arxiv.org/abs/2310.19509.
[108] ECCLES B J, RODGERS P, KILPATRICK P, et al. DNNShifter: an efficient DNN pruning system for edge computing[J]. Future Generation Computer Systems, 2024, 152: 43-54.
[109] HE X, ZHAO K Y, CHU X W. AutoML: a survey of the state-of-the-art[J]. Knowledge-Based Systems, 2021, 212: 106622.
[110] HE Y H, LIN J, LIU Z J, et al. AMC: AutoML for model compression and acceleration on mobile devices[C]//Proceedings of ECCV 2018. Berlin, Germany: Springer, 2018: 815-832.
[111] WANG M Z, MENG Y, TANG C, et al. JAQ: joint efficient architecture design and low-bit quantization with hardware[WT6,3.75BZ]—[WT6BZ]software co-exploration[C]//Proceedings of the AAAI Conference on Artificial Intelligence. Palo Alto, USA: AAAI Press, 2025: 21171-21179.
[112] ZHANG K L, LI G Y, LU N, et al. Hardware-aware DNN compression for homogeneous edge devices[EB/OL].[2026-03-07]. https://arxiv.org/abs/2501.15240.
[113] MA X L, YUAN G, LIN S, et al. Tiny but accurate: a pruned, quantized and optimized memristor crossbar framework for ultra efficient DNN implementation[C]//Proceedings of the 25th Asia and South Pacific Design Automation Conference (ASP-DAC). Washington D.C., USA: IEEE Press, 2020: 301-306.

[1]	许晓阳, 魏伟, 高重阳. 基于改进YOLOv7-tiny的红外船舶目标检测[J]. 计算机工程, 2026, 52(2): 209-220.
[2]	李强, 谭兴义, 郑唯, 刘震, 杨文海. 基于对抗训练与对比表示蒸馏的图神经网络推理优化[J]. 计算机工程, 2026, 52(1): 126-135.
[3]	刘杰, 黄晓辉, 郭敬博. 基于YOLOv8的轻量级田间棉花品级检测[J]. 计算机工程, 2026, 52(1): 400-413.
[4]	张玉博, 杨帆, 郭亚, 杨文慧. 基于视觉大模型的垃圾分类轻量化算法研究[J]. 计算机工程, 2025, 51(7): 140-151.
[5]	沈勤丰, 黄璐瑶. 一种基于不确定性校正的可信连续学习方法[J]. 计算机工程, 2025, 51(12): 56-67.
[6]	张政秀, 周淳, 杨萌. 基于实例谱关系的知识蒸馏[J]. 计算机工程, 2025, 51(11): 63-71.
[7]	林烁彬, 蔡捷仪, 方晓城, 张正, 卢光明, 陈炳志. 基于强度相关正则化学习的对抗鲁棒蒸馏方法[J]. 计算机工程, 2025, 51(1): 42-50.
[8]	屠乃威, 焦猛, 阎馨. 复杂环境下输电线路鸟巢目标图像检测模型[J]. 计算机工程, 2024, 50(7): 216-226.
[9]	曹坪, 杨怀志, 薄一军, 尤嘉, 张淳杰, 李丹勇. 面向低质量裂缝图像的多知识蒸馏分类[J]. 计算机工程, 2023, 49(7): 204-213.
[10]	毛亮, 赵林均, 余敦辉, 孙斌. 基于知识蒸馏的企业命名实体识别模型[J]. 计算机工程, 2023, 49(5): 90-96.
[11]	詹健浩, 甘利鹏, 毕永辉, 曾鹏, 李晓潮. 基于知识蒸馏的多模态融合行为识别方法[J]. 计算机工程, 2023, 49(10): 280-288, 297.
[12]	王士浩, 王中卿, 李寿山, 周国栋. 基于知识蒸馏与模型集成的事件论元抽取方法[J]. 计算机工程, 2022, 48(7): 97-103.
[13]	鲁统伟, 徐子昕, 闵锋. 基于生成对抗网络的知识蒸馏数据增强[J]. 计算机工程, 2022, 48(4): 70-80.
[14]	何涛, 俞舒曼, 徐鹤. 基于条件生成对抗网络与知识蒸馏的单幅图像去雾方法[J]. 计算机工程, 2022, 48(4): 165-172.
[15]	廖胜兰, 吉建民, 俞畅, 陈小平. 基于BERT模型与知识蒸馏的意图分类方法[J]. 计算机工程, 2021, 47(5): 73-79.

选择文件类型/文献管理软件名称

选择包含的内容