pith. sign in

arxiv: 2606.04238 · v1 · pith:R355RZMMnew · submitted 2026-06-02 · 💻 cs.LG · cs.AI

Recover-LoRA for Aggressive Quantization: Reclaiming Accuracy in 2-Bit Language Models via Low-Rank Adaptation with Knowledge Distillation on Synthetic Data

Devleena Das , Rajeev Patwari , Elliott Delaye , Ashish Sirasao This is my paper

Pith reviewed 2026-06-28 10:38 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords Recover-LoRAquantizationlow-rank adaptationknowledge distillationsynthetic datalanguage modelsmixed precisionaccuracy recovery
0
0 comments X

The pith

Recover-LoRA restores 80-95% accuracy in 2-bit quantized LLMs using logit distillation on synthetic data alone.

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

The paper extends Recover-LoRA to recover accuracy after aggressive 2-bit quantization of selected layers in large language models. It introduces a mixed-precision GateUp setup that quantizes only the gate and up-projection layers to 2 bits while keeping other layers at higher precision. Low-rank adapters are then trained on those layers through logit distillation using 10k synthetic samples and no labeled data. On Qwen3-4B this recovers 80-95% of original accuracy on nine of twelve benchmarks. Roofline analysis across model families shows the quantization also delivers throughput gains, and synthetic data matches curated labeled data for the recovery task.

Core claim

Recover-LoRA trains low-rank adapters on the 2-bit quantized gate and up-projection layers via logit distillation with synthetic data, recovering 80-95% of the accuracy lost to quantization on most benchmarks while requiring only 10k synthetic samples and no access to original labeled training data.

What carries the argument

Recover-LoRA: low-rank adapters trained by logit distillation on synthetic data to correct errors from 2-bit quantization of gate and up-projection layers

If this is right

  • W4/W2-GateUp mixed precision yields 7.5-23.3% TPS improvement over uniform W4 across 4B-20B models and two hardware platforms.
  • Recovery reaches 80-95% on nine of twelve benchmarks with only 10k synthetic samples.
  • Synthetic data performs comparably to curated labeled data for the distillation-based recovery.
  • The recovered model generalizes to out-of-distribution evaluation tasks.

Where Pith is reading between the lines

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

  • The selective layer choice could be paired with other compression methods to further reduce memory use on edge devices.
  • The same distillation setup might extend to recovering from other forms of layer-wise corruption beyond quantization.
  • Scaling the number of synthetic samples or varying their generation method could be tested to see if recovery rates improve on the remaining three benchmarks.

Load-bearing premise

Logit distillation on synthetic data generated without access to the original training distribution can reliably restore performance lost from 2-bit quantization of the gate and up-projection layers.

What would settle it

Running the same Recover-LoRA procedure on Qwen3-4B but measuring whether accuracy recovery falls below 80% on the same nine benchmarks when synthetic data is replaced by data drawn from a markedly different distribution.

read the original abstract

Aggressive weight quantization to 2-bit precision offers substantial throughput and memory gains for large language model (LLM) inference, but typically incurs severe accuracy degradation. These gains are particularly relevant for edge and on-device deployment, where memory capacity and bandwidth are primary constraints. In this work, we extend Recover-LoRA -- a lightweight, data-free accuracy recovery method originally developed for general model weight corruption -- to the setting of ultra-low-bit quantization. We propose a selective mixed-precision strategy in which only gate and up projection layers of the MLP are quantized to 2-bit (W2), while all other linear layers remain at higher precision, yielding a mixed-precision GateUp configuration. We demonstrate via roofline analysis across three model families (4B--20B) and two hardware platforms that a W4/W2-GateUp deployment (4-bit base with 2-bit gate/up) delivers 7.5--23.3\% TPS improvement over uniform W4 depending on model and context length, while confining quantization error to a predictable subset of layers. We then apply Recover-LoRA -- training low-rank adapters on the quantized layers via logit distillation with synthetic data -- to recover accuracy lost from 2-bit quantization of the gate and up layers. In a case study on Qwen3-4B, Recover-LoRA achieves 80--95\% accuracy recovery on 9 of 12 benchmarks, using only 10k synthetic training samples and no labeled data. We further demonstrate that synthetic data performs comparably to curated labeled data for distillation-based recovery, and that recovery generalizes to out-of-distribution evaluation tasks. Our results present Recover-LoRA as a practical post-quantization accuracy recovery tool for aggressive weight compression in deployment settings.

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 / 0 minor

Summary. The paper extends Recover-LoRA to post-quantization recovery for LLMs, proposing a mixed-precision W4/W2-GateUp strategy that quantizes only the gate and up-projection layers of MLPs to 2 bits. It applies low-rank adapters trained via logit distillation on 10k synthetic samples (no labeled data) and reports 80--95% accuracy recovery on 9 of 12 benchmarks for Qwen3-4B, plus 7.5--23.3% TPS gains over uniform W4 via roofline analysis on 4B--20B models across two hardware platforms.

Significance. If the recovery results hold under rigorous validation, the method would offer a practical, low-data route to aggressive compression for edge deployment while confining error to a predictable subset of layers. The roofline analysis across model families and platforms is a concrete strength that grounds the throughput claims.

major comments (2)
  1. [Abstract / Experimental Results] Abstract and § on experimental results: the headline recovery figures (80--95% on 9/12 benchmarks with 10k synthetic samples) are presented without any description of the quantization procedure for the gate/up layers, the synthetic data generation process, baseline comparisons, error bars, or evaluation protocol, rendering it impossible to determine whether the numbers support the central recovery claim.
  2. [Method / Distillation Experiments] Method and § on distillation: the claim that logit distillation on synthetic data restores performance lost specifically from 2-bit gate/up quantization rests on the untested assumption that the synthetic activations overlap with the high-error regimes of those layers; no activation-distribution analysis or ablation is supplied to address this load-bearing point.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We agree that the original submission would benefit from greater detail on experimental procedures and additional analysis to support the core claims. We have revised the manuscript to address both major comments.

read point-by-point responses

  1. Referee: [Abstract / Experimental Results] Abstract and § on experimental results: the headline recovery figures (80--95% on 9/12 benchmarks with 10k synthetic samples) are presented without any description of the quantization procedure for the gate/up layers, the synthetic data generation process, baseline comparisons, error bars, or evaluation protocol, rendering it impossible to determine whether the numbers support the central recovery claim.

    Authors: We agree that the abstract and experimental results section require additional detail for reproducibility and to substantiate the recovery claims. In the revised manuscript we have expanded both sections to include: (1) a precise description of the 2-bit quantization procedure applied to the gate and up-projection layers (including the quantizer and scaling method); (2) the synthetic data generation process (prompt-based generation from the unquantized model with diversity controls); (3) explicit baseline comparisons (uniform W4, W2 on all layers, and alternative recovery techniques); (4) error bars computed over three independent runs with different random seeds; and (5) a clarified evaluation protocol specifying the 12 benchmarks, the exact recovery metric (recovered accuracy relative to the FP16 baseline), and the train/eval split. These changes appear in the updated Abstract and the Experimental Results section. revision: yes

  2. Referee: [Method / Distillation Experiments] Method and § on distillation: the claim that logit distillation on synthetic data restores performance lost specifically from 2-bit gate/up quantization rests on the untested assumption that the synthetic activations overlap with the high-error regimes of those layers; no activation-distribution analysis or ablation is supplied to address this load-bearing point.

    Authors: We acknowledge that the original manuscript does not contain an explicit activation-distribution analysis or ablation study directly testing overlap between synthetic activations and the high-error regimes induced by 2-bit gate/up quantization. This is a substantive observation. In the revised version we have added a new subsection under Method that reports activation histograms and KL-divergence statistics between synthetic and real activations for the gate and up layers, together with an ablation that substitutes real calibration data for the synthetic set. The added results show substantial distributional overlap in the regions where quantization error is largest, thereby supporting the original claim while addressing the referee's concern. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical recovery measurements

full rationale

The paper reports measured accuracy recovery percentages (80-95% on 9/12 benchmarks) from applying an existing method to a new quantization setting, using standard benchmarks and 10k synthetic samples. No equations, fitted parameters, or derivations are presented that reduce to inputs by construction. The reference to Recover-LoRA as prior work is a normal citation of an earlier method and does not serve as the load-bearing justification for the reported empirical outcomes, which stand as independent measurements.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the untested premise that synthetic-data distillation can substitute for labeled data when recovering from 2-bit gate/up quantization error; no free parameters or invented entities are visible in the abstract.

axioms (1)
  • domain assumption Logit distillation on synthetic data can recover accuracy lost from 2-bit quantization of gate and up-projection layers
    This premise is required for the accuracy-recovery claim to hold without real labeled data.

pith-pipeline@v0.9.1-grok · 5875 in / 1260 out tokens · 32300 ms · 2026-06-28T10:38:47.741804+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

56 extracted references · 7 canonical work pages · 5 internal anchors

  1. [1]

    Recover-

    Das, Devleena and Patwari, Rajeev and Sirasao, Ashish , booktitle=. Recover-

  2. [2]

    Lin, Ji and Tang, Jiaming and Tang, Haotian and Yang, Shang and Chen, Wei-Ming and Wang, Wei-Chen and Xiao, Guangxuan and Dang, Xingyu and Gan, Chuang and Han, Song , booktitle=

  3. [3]

    Frantar, Elias and Ashkboos, Saleh and Hoefler, Torsten and Alistarh, Dan , booktitle=

  4. [4]

    Transactions of the Association for Computational Linguistics , volume=

    A Survey on Model Compression for Large Language Models , author=. Transactions of the Association for Computational Linguistics , volume=

  5. [5]

    Liu, Zechun and Oguz, Barlas and Zhao, Changsheng and Chang, Ernie and Stock, Pierre and Mehdad, Yashar and Shi, Yangyang and Krishnamoorthi, Raghuraman and Chandra, Vikas , journal=

  6. [6]

    Ke, Wenjing and Li, Zhe and Li, Dong and Tian, Lu and Barsoum, Emad , booktitle=

  7. [7]

    Hu, Edward J and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle=

  8. [8]

    Dettmers, Tim and Pagnoni, Artidoro and Holtzman, Ari and Zettlemoyer, Luke , booktitle=

  9. [9]

    Xu, Yuhui and Xie, Lingxi and Gu, Xiaotao and Chen, Xin and Chang, Heng and Zhang, Hengheng and Chen, Zhengsu and Zhang, Xiaopeng and Tian, Qi , booktitle=

  10. [10]

    Tseng, Albert and Chee, Jerry and Sun, Qingyao and Kuleshov, Volodymyr and De Sa, Christopher , booktitle=

  11. [11]

    Egiazarian, Vage and Panferov, Andrei and Kuznedelev, Denis and Frantar, Elias and Babenko, Artem and Alistarh, Dan , booktitle=

  12. [12]

    Dong, Zhen and Yao, Zhewei and Gholami, Amir and Mahoney, Michael W and Keutzer, Kurt , booktitle=

  13. [13]

    Wang, Kuan and Liu, Zhijian and Lin, Yujun and Lin, Ji and Han, Song , booktitle=

  14. [14]

    Lee, Changhun and Jin, Jungyu and Kim, Taesu and Kim, Hyungjun and Park, Eunhyeok , booktitle=

  15. [16]

    Grattafiori, Aaron and Dubey, Abhimanyu and Jauhri, Abhinav and Pandey, Abhinav and Kadian, Abhishek and Al-Dahle, Ahmad and Letman, Aiesha and Mathur, Akhil and Schelten, Alan and Vaughan, Alex and others , journal=. The

  16. [18]

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

  17. [19]

    International Conference on Learning Representations (ICLR) , year=

    Measuring Massive Multitask Language Understanding , author=. International Conference on Learning Representations (ICLR) , year=

  18. [20]

    Think You Have Solved Question Answering? Try

    Clark, Peter and Cowhey, Isaac and Etzioni, Oren and Khot, Tushar and Sabharwal, Ashish and Schoenick, Carissa and Tafjord, Oyvind , journal=. Think You Have Solved Question Answering? Try

  19. [21]

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

  20. [22]

    Bisk, Yonatan and Zellers, Rowan and Le Bras, Ronan and Gao, Jianfeng and Choi, Yejin , booktitle=

  21. [23]

    Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP) , year=

    Can a Suit of Armor Conduct Electricity? A New Dataset for Open Book Question Answering , author=. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP) , year=

    2018

  22. [24]

    Clark, Christopher and Lee, Kenton and Chang, Ming-Wei and Kwiatkowski, Tom and Collins, Michael and Toutanova, Kristina , booktitle=

  23. [25]

    Forecasting

    Patwari, Rajeev and Sirasao, Ashish and Das, Devleena , journal=. Forecasting

  24. [26]

    DeepSeek-AI , journal=

  25. [27]

    Diagnosing

    Cim, Musa and Topcu, Burak and Kandemir, Mahmut , booktitle=. Diagnosing

  26. [28]

    Proceedings of the 40th International Conference on Machine Learning (ICML) , year=

    The Case for 4-Bit Precision: k-Bit Inference Scaling Laws , author=. Proceedings of the 40th International Conference on Machine Learning (ICML) , year=

  27. [29]

    ParoQuant: Pairwise Rotation Quantization for Efficient Reasoning

    Liang, Yesheng and Chen, Haisheng and Zhang, Zihan and Han, Song and Liu, Zhijian , booktitle=. ParoQuant: Pairwise Rotation Quantization for Efficient Reasoning

  28. [30]

    PIQA : Reasoning about physical commonsense in natural language

    Yonatan Bisk, Rowan Zellers, Ronan Le Bras, Jianfeng Gao, and Yejin Choi. PIQA : Reasoning about physical commonsense in natural language. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 7432--7439, 2020

    2020

  29. [31]

    Diagnosing FP4 inference: A layer-wise and block-wise sensitivity analysis of NVFP4 and MXFP4

    Musa Cim, Burak Topcu, and Mahmut Kandemir. Diagnosing FP4 inference: A layer-wise and block-wise sensitivity analysis of NVFP4 and MXFP4 . In Workshop on Scientific Methods for Understanding Deep Learning (Sci4DL), 2026

    2026

  30. [32]

    BoolQ : Exploring the surprising difficulty of natural yes/no questions

    Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. BoolQ : Exploring the surprising difficulty of natural yes/no questions. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), 2019

    2019

  31. [33]

    Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge

    Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. 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

  32. [34]

    Recover- LoRA : Data-free accuracy recovery of degraded language models via low-rank adaptation

    Devleena Das, Rajeev Patwari, and Ashish Sirasao. Recover- LoRA : Data-free accuracy recovery of degraded language models via low-rank adaptation. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 2377--2386, 2025

    2025

  33. [35]

    DeepSeek-V3 Technical Report

    DeepSeek-AI. DeepSeek-V3 technical report. arXiv preprint arXiv:2412.19437, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  34. [36]

    The case for 4-bit precision: k-bit inference scaling laws

    Tim Dettmers and Luke Zettlemoyer. The case for 4-bit precision: k-bit inference scaling laws. In Proceedings of the 40th International Conference on Machine Learning (ICML), 2023

    2023

  35. [37]

    QLoRA : Efficient finetuning of quantized LLMs

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. QLoRA : Efficient finetuning of quantized LLMs . In Advances in Neural Information Processing Systems (NeurIPS), volume 36, pages 10088--10115, 2023

    2023

  36. [38]

    HAWQ : Hessian AWare quantization of neural networks with mixed-precision

    Zhen Dong, Zhewei Yao, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. HAWQ : Hessian AWare quantization of neural networks with mixed-precision. In International Conference on Computer Vision (ICCV), 2019

    2019

  37. [39]

    AQLM : Extreme compression of large language models via additive quantization

    Vage Egiazarian, Andrei Panferov, Denis Kuznedelev, Elias Frantar, Artem Babenko, and Dan Alistarh. AQLM : Extreme compression of large language models via additive quantization. In International Conference on Machine Learning (ICML), 2024

    2024

  38. [40]

    GPTQ : Accurate post-training quantization for generative pre-trained transformers

    Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. GPTQ : Accurate post-training quantization for generative pre-trained transformers. In International Conference on Learning Representations (ICLR), 2023

    2023

  39. [41]

    The Llama 3 Herd of Models

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

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  40. [42]

    Measuring massive multitask language understanding

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. In International Conference on Learning Representations (ICLR), 2021

    2021

  41. [43]

    Distilling the Knowledge in a Neural Network

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  42. [44]

    LoRA : Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA : Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2022

    2022

  43. [45]

    DL-QAT : Weight-decomposed low-rank quantization-aware training for large language models

    Wenjing Ke, Zhe Li, Dong Li, Lu Tian, and Emad Barsoum. DL-QAT : Weight-decomposed low-rank quantization-aware training for large language models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track, 2024

    2024

  44. [46]

    OWQ : Outlier-aware weight quantization for efficient fine-tuning and inference of large language models

    Changhun Lee, Jungyu Jin, Taesu Kim, Hyungjun Kim, and Eunhyeok Park. OWQ : Outlier-aware weight quantization for efficient fine-tuning and inference of large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, 2024

    2024

  45. [47]

    Paroquant: Pairwise rotation quantization for efficient reasoning LLM inference

    Yesheng Liang, Haisheng Chen, Zihan Zhang, Song Han, and Zhijian Liu. Paroquant: Pairwise rotation quantization for efficient reasoning LLM inference. In International Conference on Learning Representations (ICLR), 2026

    2026

  46. [48]

    AWQ : Activation-aware weight quantization for on-device LLM compression and acceleration

    Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Wei-Ming Chen, Wei-Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, and Song Han. AWQ : Activation-aware weight quantization for on-device LLM compression and acceleration. In Proceedings of Machine Learning and Systems, volume 6, pages 87--100, 2024

    2024

  47. [49]

    Llm-qat: Data-free quantization aware training for large language models.CoRR, abs/2305.17888,

    Zechun Liu, Barlas Oguz, Changsheng Zhao, Ernie Chang, Pierre Stock, Yashar Mehdad, Yangyang Shi, Raghuraman Krishnamoorthi, and Vikas Chandra. LLM-QAT : Data-free quantization aware training for large language models. arXiv preprint arXiv:2305.17888, 2023

    work page arXiv 2023

  48. [50]

    Can a suit of armor conduct electricity? a new dataset for open book question answering

    Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new dataset for open book question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2018

    2018

  49. [51]

    Patwari, A

    Rajeev Patwari, Ashish Sirasao, and Devleena Das. Forecasting LLM inference performance via hardware-agnostic analytical modeling. arXiv preprint arXiv:2508.00904, 2025

    work page arXiv 2025

  50. [52]

    WinoGrande : An adversarial Winograd schema challenge at scale

    Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. WinoGrande : An adversarial Winograd schema challenge at scale. Communications of the ACM, 64 0 (9): 0 99--106, 2021

    2021

  51. [53]

    QuIP \# : Even better LLM quantization with hadamard incoherence and lattice codebooks

    Albert Tseng, Jerry Chee, Qingyao Sun, Volodymyr Kuleshov, and Christopher De Sa. QuIP \# : Even better LLM quantization with hadamard incoherence and lattice codebooks. In International Conference on Machine Learning (ICML), 2024

    2024

  52. [54]

    HAQ : Hardware-aware automated quantization with mixed precision

    Kuan Wang, Zhijian Liu, Yujun Lin, Ji Lin, and Song Han. HAQ : Hardware-aware automated quantization with mixed precision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019

    2019

  53. [55]

    QA-LoRA : Quantization-aware low-rank adaptation of large language models

    Yuhui Xu, Lingxi Xie, Xiaotao Gu, Xin Chen, Heng Chang, Hengheng Zhang, Zhengsu Chen, Xiaopeng Zhang, and Qi Tian. QA-LoRA : Quantization-aware low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2024

    2024

  54. [56]

    Qwen3 Technical Report

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Wang, Bowen Zheng, Chengyuan Yu, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  55. [57]

    HellaSwag : Can a machine really finish your sentence? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), 2019

    Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. HellaSwag : Can a machine really finish your sentence? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), 2019

    2019

  56. [58]

    A survey on model compression for large language models

    Xunyu Zhu, Jian Li, Yong Liu, Can Ma, and Weiping Wang. A survey on model compression for large language models. Transactions of the Association for Computational Linguistics, 12: 0 1556--1577, 2024

    2024