pith. sign in

arxiv: 2606.12487 · v1 · pith:DV4JHGU2new · submitted 2026-06-10 · 💻 cs.LG

DynamicPTQ: Mitigating Activation Quantization Collapse via Residual-Stream Dynamics

Zimo Zhao , Maolin Wang , Bowen Yu , Bowen Liu , Xiao Han , Xiangyu Zhao This is my paper

Pith reviewed 2026-06-27 10:34 UTC · model grok-4.3

classification 💻 cs.LG
keywords post-training quantizationactivation quantizationresidual streammassive activationsmixed-precisionlarge language modelsW4A4KV4
0
0 comments X

The pith

DynamicPTQ raises activation precision to 8 bits only in layers where residual-stream phase changes dominate 4-bit scales.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that massive activations in large language models appear and vanish in distinct phases across network depth, producing abrupt residual-stream shifts that static smoothing transformations cannot stabilize. These shifts let new layer updates overwhelm the quantization scale and erode prior information, collapsing performance when weights, activations, and KV caches are all forced to 4 bits. The authors introduce Jump Ratio and Historical Feature SNR to locate the unstable layers and apply 8-bit activations solely there while retaining 4-bit precision everywhere else. When combined with existing PTQ baselines, the policy improves perplexity and zero-shot QA scores on LLaMA-2 and LLaMA-3 under W4A4KV4 settings and yields modest throughput gains. A reader would care because the approach keeps most of the model in low precision yet recovers accuracy without retraining.

Core claim

Massive activations emerge and disappear in a phase-wise pattern across network depth, triggering large residual changes. These changes cause newly injected layer-wise updates to dominate the 4-bit quantization scale and weaken historical residual information. Static transformation-based smoothing cannot fully resolve the resulting dynamic instability. DynamicPTQ therefore identifies quantization-sensitive layers from residual-stream dynamics and assigns 8-bit activation precision only to those layers, keeping weights, KV caches, and remaining activations at 4 bits.

What carries the argument

Jump Ratio and Historical Feature SNR, which measure the sudden appearance of massive activations and the dominance of new updates over retained historical features in the residual stream.

If this is right

  • Integration with QuaRot, SpinQuant, or FlatQuant yields consistent perplexity reductions under W4A4KV4.
  • Zero-shot QA accuracy rises on both LLaMA-2 and LLaMA-3 models.
  • Throughput improves by a factor of 1.05 to 1.07 with only modest added memory.
  • The policy supplies a direct route to robust low-bit inference without full retraining.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same residual metrics could be used to decide bit widths during training rather than only after training.
  • Hardware schedulers might switch activation precision on the fly by tracking Jump Ratio across successive batches.
  • Similar phase detection might apply to other dynamic-range problems such as KV-cache eviction or activation sparsity.

Load-bearing premise

Phase-wise residual-stream changes, rather than static per-layer statistics, are the main driver of activation quantization collapse and can be corrected by raising precision only in the affected layers.

What would settle it

Measuring whether layers flagged by high Jump Ratio or low Historical Feature SNR produce the largest activation quantization error, and whether restricting 8-bit precision to exactly those layers recovers the reported perplexity gains while uniform 4-bit or uniform 8-bit baselines do not.

Figures

Figures reproduced from arXiv: 2606.12487 by Bowen Liu, Bowen Yu, Maolin Wang, Xiangyu Zhao, Xiao Han, Zimo Zhao.

Figure 1
Figure 1. Figure 1: Massive activations and residual-stream dynam [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of residual-stream dynamics under [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Average zero-shot QA accuracy of QuaRot with and without DynamicPTQ under W4A4KV4 quantization. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Residual-stream dynamics of DeepSeek-V2-Lite [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Layer-wise residual-stream dynamics across LLaMA-2 models. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Layer-wise residual-stream dynamics across different 7B-scale decoder-only model families. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
read the original abstract

Post-training quantization (PTQ) is essential for efficient large language model inference, but reliably quantizing activations remains challenging when weights, activations, and KV caches are all quantized to 4-bit precision. A key difficulty lies in massive activations, whose extreme values dominate the activation range and amplify quantization errors. State-of-the-art methods mainly mitigate massive activations through transformation-based smoothing, such as orthogonal rotations and affine scaling, but overlook the cross-layer dynamics of the residual stream. In this paper, we show that massive activations emerge and disappear in a phase-wise pattern across network depth, triggering large residual changes. These changes cause newly injected layer-wise updates to dominate the 4-bit quantization scale and weaken historical residual information. To characterize this behavior, we introduce Jump Ratio and Historical Feature SNR. This suggests that static transformation-based smoothing cannot fully resolve dynamic quantization instability caused by cross-layer residual changes. Based on this analysis, we propose DynamicPTQ, a Dynamic Post-Training Quantization policy for phase-aware mixed-precision activation quantization. DynamicPTQ identifies quantization-sensitive layers from residual-stream dynamics and assigns 8-bit activation precision only to these layers, while keeping weights, KV caches, and other activations in 4-bit precision. It can be directly integrated with strong PTQ baselines such as QuaRot, SpinQuant, and FlatQuant. Experiments on LLaMA-2 and LLaMA-3 show that DynamicPTQ consistently improves perplexity and zero-shot QA performance under W4A4KV4 quantization, while achieving 1.05 to 1.07 times throughput improvement with modest memory overhead. These results demonstrate a practical path toward robust low-bit LLM inference.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper claims that massive activations emerge and disappear in a phase-wise pattern across LLM layers, causing dynamic residual-stream changes that lead to quantization instability under W4A4KV4 settings. Static smoothing methods (e.g., rotations and scaling in QuaRot/SpinQuant/FlatQuant) are insufficient because they overlook these cross-layer dynamics. The authors introduce two new metrics—Jump Ratio and Historical Feature SNR—to characterize the behavior, and propose DynamicPTQ, a mixed-precision policy that identifies sensitive layers via these metrics and elevates only their activations to 8 bits while keeping weights, KV caches, and other activations at 4 bits. Experiments on LLaMA-2 and LLaMA-3 report consistent gains in perplexity and zero-shot QA when DynamicPTQ is integrated with the baselines, plus 1.05–1.07× throughput with modest memory cost.

Significance. If the central claim holds after proper controls, the work offers a targeted, low-overhead way to mitigate activation quantization collapse by exploiting residual-stream phase dynamics rather than uniform or purely magnitude-based fixes. The explicit compatibility with multiple strong PTQ baselines and the reported throughput numbers are concrete strengths that could influence practical low-bit inference pipelines.

major comments (2)
  1. [Experiments] Experiments section: the manuscript reports perplexity and QA gains from DynamicPTQ but contains no ablation that assigns the extra 8-bit activations to randomly chosen layers or to layers selected solely by per-layer activation magnitude (the statistic already used by smoothing baselines). Without this control it is impossible to isolate whether the residual-stream metrics (Jump Ratio, Historical Feature SNR) add predictive power beyond simply giving extra bits to some layers. This directly bears on the claim that phase-wise dynamics are the primary driver missed by static methods.
  2. [Method] Method / metric definitions: Jump Ratio and Historical Feature SNR are introduced to capture phase-wise residual changes, yet the text provides neither explicit equations for their computation nor quantitative evidence (e.g., correlation plots or regression against observed quantization error) that they predict instability better than existing activation-range statistics. This absence undermines verification that the metrics are load-bearing for the proposed policy.
minor comments (1)
  1. [Abstract] Abstract and experimental description: the claimed 1.05–1.07× throughput improvement is stated without specifying the hardware platform, batch size, or exact baseline configuration, which is needed for reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight important aspects of experimental validation and metric clarity. We address each major comment below and will revise the manuscript to strengthen the presentation of our contributions.

read point-by-point responses

  1. Referee: [Experiments] Experiments section: the manuscript reports perplexity and QA gains from DynamicPTQ but contains no ablation that assigns the extra 8-bit activations to randomly chosen layers or to layers selected solely by per-layer activation magnitude (the statistic already used by smoothing baselines). Without this control it is impossible to isolate whether the residual-stream metrics (Jump Ratio, Historical Feature SNR) add predictive power beyond simply giving extra bits to some layers. This directly bears on the claim that phase-wise dynamics are the primary driver missed by static methods.

    Authors: We agree that the requested ablations are necessary to isolate the contribution of the residual-stream metrics. In the revision we will add experiments that compare DynamicPTQ's metric-driven layer selection against (i) random layer selection and (ii) selection based solely on per-layer activation magnitude (the statistic already employed by the smoothing baselines). These controls will be reported on the same LLaMA-2 and LLaMA-3 models and W4A4KV4 setting, allowing direct assessment of whether the proposed metrics provide predictive value beyond magnitude-based or random allocation. revision: yes

  2. Referee: [Method] Method / metric definitions: Jump Ratio and Historical Feature SNR are introduced to capture phase-wise residual changes, yet the text provides neither explicit equations for their computation nor quantitative evidence (e.g., correlation plots or regression against observed quantization error) that they predict instability better than existing activation-range statistics. This absence undermines verification that the metrics are load-bearing for the proposed policy.

    Authors: We will add the explicit mathematical definitions of both Jump Ratio and Historical Feature SNR to the revised manuscript. We will also include quantitative supporting evidence in the form of correlation plots and regression analyses that relate these metrics to observed per-layer quantization error, demonstrating their relationship to instability beyond standard activation-range statistics used by prior smoothing methods. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper's chain consists of empirical observation of residual-stream patterns, definition of Jump Ratio and Historical Feature SNR from those observations, a heuristic policy selecting layers for 8-bit activations, and experimental validation on LLaMA models when combined with external baselines (QuaRot, SpinQuant, FlatQuant). No equation or claim reduces by construction to a fitted parameter, self-referential definition, or self-citation; the metrics and policy are presented as derived from data rather than presupposing the performance outcome. This is the normal self-contained case.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

The claim rests on the validity of two newly defined metrics (Jump Ratio, Historical Feature SNR) whose definitions and layer-selection thresholds are not supplied, plus the domain assumption that residual-stream phase changes dominate quantization error in W4A4KV4 settings.

axioms (1)
  • domain assumption Massive activations emerge and disappear in a phase-wise pattern across network depth and trigger large residual changes that dominate 4-bit quantization scales.
    Stated directly in the abstract as the key difficulty overlooked by prior smoothing methods.
invented entities (2)
  • Jump Ratio no independent evidence
    purpose: Characterize the phase-wise pattern of massive activations and residual changes
    New metric introduced to quantify the observed cross-layer behavior.
  • Historical Feature SNR no independent evidence
    purpose: Measure weakening of historical residual information due to new layer-wise updates
    New metric introduced to quantify the observed cross-layer behavior.

pith-pipeline@v0.9.1-grok · 5846 in / 1504 out tokens · 35284 ms · 2026-06-27T10:34:24.547984+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 19 canonical work pages · 15 internal anchors

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Alt- man, Shyamal Anadkat, et al. 2023. Gpt-4 technical report.arXiv preprint arXiv:2303.08774(2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  2. [2]

    Saleh Ashkboos, Amirkeivan Mohtashami, Maximilian L Croci, Bo Li, Pash- mina Cameron, Martin Jaggi, Dan Alistarh, Torsten Hoefler, and James Hens- man. 2024. Quarot: Outlier-free 4-bit inference in rotated llms.Advances in Neural Information Processing Systems37 (2024), 100213–100240

    2024

  3. [3]

    Yonatan Bisk, Rowan Zellers, Jianfeng Gao, Yejin Choi, et al . 2020. Piqa: Reasoning about physical commonsense in natural language. InProceedings of the AAAI conference on artificial intelligence, Vol. 34. 7432–7439

    2020

  4. [4]

    Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. 2018. Think you have solved question answering? try arc, the ai2 reasoning challenge.arXiv preprint arXiv:1803.05457(2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  5. [5]

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. 2021. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168(2021)

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  6. [6]

    Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. 2022. Gpt3. int8 (): 8-bit matrix multiplication for transformers at scale.Advances in neural information processing systems35 (2022), 30318–30332

    2022

  7. [7]

    Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. 2022. Gptq: Accurate post-training quantization for generative pre-trained transformers. arXiv preprint arXiv:2210.17323(2022)

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  8. [8]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al . 2024. The llama 3 herd of models.arXiv preprint arXiv:2407.21783(2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  9. [9]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948(2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  10. [10]

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. Measuring mathematical problem solving with the math dataset.arXiv preprint arXiv:2103.03874 (2021)

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  11. [11]

    Coleman Hooper, Sehoon Kim, Hiva Mohammadzadeh, Michael W Mahoney, Yakun S Shao, Kurt Keutzer, and Amir Gholami. 2024. Kvquant: Towards 10 million context length llm inference with kv cache quantization.Advances in Neural Information Processing Systems37 (2024), 1270–1303

    2024

  12. [12]

    Naman Jain, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. 2025. Livecodebench: Holistic and contamination free evaluation of large language models for code. InInternational Conference on Learning Representations, Vol. 2025. 58791– 58831

    2025

  13. [13]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, De- vendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timo- thée Lacroix, and William El Sayed. 2023. Mistral 7B. arXiv:2310.0...

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  14. [14]

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. 2020. Scaling laws for neural language models.arXiv preprint arXiv:2001.08361(2020)

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  15. [15]

    Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Wei-Ming Chen, Wei-Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, and Song Han. 2024. Awq: Activation-aware weight quantization for on-device llm compression and acceleration.Proceedings of machine learning and systems6 (2024), 87–100

    2024

  16. [16]

    Ruikang Liu, Yuxuan Sun, Manyi Zhang, Haoli Bai, Xianzhi Yu, Tiezheng Yu, Chun Yuan, and Lu Hou. 2025. Quantization hurts reasoning? an empirical study on quantized reasoning models.arXiv preprint arXiv:2504.04823(2025)

    work page arXiv 2025

  17. [17]

    Zirui Liu, Jiayi Yuan, Hongye Jin, Shaochen Zhong, Zhaozhuo Xu, Vladimir Braverman, Beidi Chen, and Xia Hu. 2024. Kivi: A tuning-free asymmetric 2bit quantization for kv cache.arXiv preprint arXiv:2402.02750(2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  18. [18]

    Zechun Liu, Changsheng Zhao, Igor Fedorov, Bilge Soran, Dhruv Choudhary, Raghuraman Krishnamoorthi, Vikas Chandra, Yuandong Tian, and Tijmen Blankevoort. 2025. Spinquant: Llm quantization with learned rotations. In International Conference on Learning Representations, Vol. 2025. 92009–92032

    2025

  19. [19]

    Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. 2016. Pointer sentinel mixture models.arXiv preprint arXiv:1609.07843(2016)

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  20. [20]

    Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort. 2020. Up or down? adaptive rounding for post-training quantization. InInternational conference on machine learning. PMLR, 7197– 7206

    2020

  21. [21]

    Denis Paperno, Germán Kruszewski, Angeliki Lazaridou, Ngoc-Quan Pham, Raffaella Bernardi, Sandro Pezzelle, Marco Baroni, Gemma Boleda, and Raquel Fernández. 2016. The LAMBADA dataset: Word prediction requiring a broad discourse context. InProceedings of the 54th annual meeting of the association for computational linguistics (volume 1: Long papers). 1525–1534

    2016

  22. [22]

    Enrique Queipo-de Llano, Álvaro Arroyo, Federico Barbero, Xiaowen Dong, Michael Bronstein, Yann LeCun, and Ravid Shwartz-Ziv. 2025. Attention sinks and compression valleys in llms are two sides of the same coin.arXiv preprint arXiv:2510.06477(2025)

    work page arXiv 2025

  23. [23]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer.Journal of machine learning research21, 140 (2020), 1–67

    2020

  24. [24]

    David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R Bowman

  25. [25]

    Gpqa: A graduate-level google-proof q&a benchmark.arXiv preprint arXiv:2311.12022(2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  26. [26]

    Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi

  27. [27]

    Winogrande: An adversarial winograd schema challenge at scale. Commun. ACM64, 9 (2021), 99–106

    2021

  28. [28]

    Wenqi Shao, Mengzhao Chen, Zhaoyang Zhang, Peng Xu, Lirui Zhao, Zhiqian Li, Kaipeng Zhang, Gao Peng, Yu Qiao, and Ping Luo. 2024. Omniquant: Omnidirectionally calibrated quantization for large language models. In International Conference on Learning Representations, Vol. 2024. 45472–45496

    2024

  29. [29]

    Mingjie Sun, Xinlei Chen, J Zico Kolter, and Zhuang Liu. 2024. Massive activations in large language models.arXiv preprint arXiv:2402.17762(2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  30. [30]

    Yuxuan Sun, Ruikang Liu, Haoli Bai, Han Bao, Kang Zhao, Yuening Li, Jiaxin Hu, Xianzhi Yu, Lu Hou, Chun Yuan, et al. 2024. Flatquant: Flatness matters for llm quantization.arXiv preprint arXiv:2410.09426(2024)

    work page arXiv 2024

  31. [31]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288(2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  32. [32]

    Guangxuan Xiao, Ji Lin, Mickael Seznec, Hao Wu, Julien Demouth, and Song Han. 2023. Smoothquant: Accurate and efficient post-training quantization for large language models. InInternational conference on machine learning. PMLR, 38087–38099

    2023

  33. [33]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al . 2025. Qwen3 technical report.arXiv preprint arXiv:2505.09388(2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  34. [34]

    Zhewei Yao, Reza Yazdani Aminabadi, Minjia Zhang, Xiaoxia Wu, Conglong Li, and Yuxiong He. 2022. Zeroquant: Efficient and affordable post-training quantization for large-scale transformers.Advances in neural information processing systems35 (2022), 27168–27183

    2022

  35. [35]

    Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi

  36. [36]

    InProceedings of the 57th annual meeting of the association for computational linguistics

    Hellaswag: Can a machine really finish your sentence?. InProceedings of the 57th annual meeting of the association for computational linguistics. 4791–4800

  37. [37]

    Yilong Zhao, Chien-Yu Lin, Kan Zhu, Zihao Ye, Lequn Chen, Size Zheng, Luis Ceze, Arvind Krishnamurthy, Tianqi Chen, and Baris Kasikci. 2024. Atom: Low-bit quantization for efficient and accurate llm serving.Proceedings of Machine Learning and Systems6 (2024), 196–209

    2024

  38. [38]

    Yue Zheng, Yuhao Chen, Bin Qian, Xiufang Shi, Yuanchao Shu, and Jiming Chen. 2025. A Review on Edge Large Language Models: Design, Execution, and Applications. arXiv:2410.11845 [cs.DC] https://arxiv.org/abs/2410.11845 Received 20 February 2007; revised 12 March 2009; accepted 5 June 2009

    work page arXiv 2025