Pith. sign in

REVIEW 3 major objections 3 minor 51 references

Unifying Uniform and Binary-coding Quantization for Accurate Compression of Large Language Models

T0 review · 3 major / 3 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read UniQuanF claims that LLM quantization can keep both the easy optimization of uniform schemes and the adaptive levels of binary-coding schemes, and that the merged quantizer deploys with no extra cost.

desk verdict Solid, incremental LLM quantization paper with a correct unification theorem; the experiments are convincing for the chosen baselines but miss key comparisons and leave a train-deploy code-assignment mismatch unquantified. read the letter →

arxiv 2506.03781 v2 pith:HDTEO7UZ submitted 2025-06-04 cs.CL

Seungcheol Park , Jeongin Bae , Beomseok Kwon , Minjun Kim , Byeongwook Kim , Se Jung Kwon , U Kang , Dongsoo Lee This is my paper
classification cs.CL
keywords LLMquantizationbinary-codinguniformpost-trainingFlexRoundALTERNATINGnon-uniformlevelsunificationtheorem
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Low-bit quantization usually forces a choice between two schemes: uniform quantization (UQ), whose evenly spaced levels make optimization easy, and binary-coding quantization (BCQ), whose flexible levels adapt to the weight distribution but are hard to optimize. This paper claims the choice is unnecessary. UniQuanF inserts FlexRound's learnable affine transformation in front of ALTERNATING's non-uniform binary-level mapper, so the quantizer gets UQ's steep error reduction and BCQ's distribution-shaped levels at the same time. After optimization, a unification theorem folds the two inference steps into one standard BCQ step, so deployment runs at exactly the memory and compute cost of plain BCQ. On GSM8K the merged quantizer beats the best single-scheme method by up to 4.60% at the same bit width.

What carries the argument

The carrying object is the unified quantizer $Q^*_{IF}(w; \Theta_{IF}) = D_R(M^*_B(T_F(w; \Theta_F); \Theta_B); \Theta_R)$, a three-stage pipeline. FlexRound's transformation $T_F(w) = w \oslash (\Delta \cdot s \cdot s_r) + z_U \mathbf{1}$ moves weights into a space where UQ-style optimization is easy; ALTERNATING's mapper $M^*_B$ assigns each transformed weight to the nearest node of the binary tree of levels centered at $z_B$ with widths $\alpha_1, \dots, \alpha_k$; FlexRound's affine detransformation $D_R$ returns to the original weight space. Three mechanisms make this practical: unified initialization, which nests ALTERNATING's alternating updates inside FlexRound's grid search so $\alpha$ and $z_B$ are found jointly with $\Delta$ and $z_U$; local and periodic mapping, which remaps each weight only every $p$ steps and checks only the previous level and its two neighbors, exploiting the observed stability of index assignments (99.97% unchanged per step on 3-bit Llama-3 8B); and the unification theorem, which merges the two affine stages into one BCQ parameter set through the identity $\alpha^* = \Delta\alpha$, $z^*_B = \Delta(z_B - z_U)$.

What would settle it

Quantize a model where index assignments are known to be restless — for example a 2-bit large model — and count per-step level-index changes exactly as Table 2 does. If a non-negligible fraction of weights change more than one level between remaps, the local remap will pin weights to wrong levels for up to $p$ steps, and the accuracy gap between UniQuanF and a full-remap variant should shrink or vanish; that would show the locality premise, not the unification, is carrying the result.

Watch

Extended reading notes

Core claim

The paper's central claim is that the two things that make UQ and BCQ separately useful live in different stages of the same three-stage quantization process $Q = D(M(T(w)))$: UQ's strength is its parameterized transformation and detransformation, while BCQ's strength is its non-uniform mapping $M_B$. UniQuan composes them as $D_U(M_B(T_U(w)))$, and UniQuanF instantiates this with FlexRound's transformation $T_F$ and detransformation $D_R$ wrapped around ALTERNATING's mapper. Joint optimization of the unified parameters, aided by unified initialization and a local, periodic remapping of weights to levels, produces quantized models that beat both parents. The load-bearing identity is the unification theorem: for any binary code matrix $C$, $R_B(C; \Theta^*_B) = D_R(R_B(C; \Theta_B); \Theta_R)$ holds exactly with $\alpha^* = \Delta\alpha$ and $z^*_B = \Delta(z_B - z_U)$, meaning the two-step training-time reconstruction is precisely a single BCQ reconstruction at deployment, with no extra parameters stored and no extra arithmetic executed.

Load-bearing premise

Local and periodic mapping trusts that during optimization each weight's chosen quantization level stays fixed from step to step and, when it changes, moves only to an adjacent level — a pattern measured on a single configuration, 3-bit Llama-3 8B.

Editorial extensions

If this is right

  • A UniQuanF-quantized 3-bit or 4-bit model is stored and executed exactly like a plain BCQ model — one binary code matrix $C$ and one parameter set $\Theta^*_B$ — so the accuracy gain carries no memory or latency penalty at deployment.
  • Because the theorem only requires that the UQ component's transformation and detransformation be affine, the framework extends in theory to AWQ, OmniQuant, and GPTQ-style compensation layered onto BCQ levels, not just FlexRound.
  • The gains concentrate exactly where the parents fail: ALTERNATING alone scores near 0% on GSM8K and FlexRound alone reaches 54.13% at 3 bits, while UniQuanF reaches 58.73%.
  • At matched accuracy, UniQuanF is not slower to train at 3 bits: it needs 15,905 seconds to surpass FlexRound's best accuracy, which itself costs 16,868 seconds.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the locality premise holds beyond 3-bit Llama-3 8B — weights' level indices stable per step, changes only adjacent — then the remapping period $p$ could be tuned per layer from the measured change rate; the paper's own finding that $p = 2$ or $4$ beats $p = 1$ already suggests sparser remapping stabilizes rather than harms optimization.
  • The unification theorem is a closure property — an affine image of a binary-coding codebook is another binary-coding codebook — so any future quantization scheme whose reconstruction is affine in its code admits the same zero-cost merge; whether non-affine mappers admit an approximate theorem with a small explicit approximation cost is an open question the paper does not address.
  • Because the deployed artifact is a BCQ codebook, its parameter overhead is $16(k+1)/g$ bits per weight; on Llama-3 70B, where group size $g = 128$ is used, that overhead is an order of magnitude larger than for channel-wise groups ($g > 4000$), so small-group deployments should count the codebook parameters against the memory budget.
  • The paper validates one pair experimentally (FlexRound plus ALTERNATING) and only sketches the others theoretically; a direct test of the framework's promise is to swap in a different mapper — say a deeper binary tree or a different bit width — and rerun the GSM8K comparison.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. The paper proposes UniQuanF, a post-training quantization method for LLMs that combines the flexible affine transformation of FlexRound with the non-uniform binary-coding levels of ALTERNATING/BCQ. The training-time quantizer (Eq. 3) uses FlexRound's element-wise and row-wise scales inside the BCQ mapping, and a unification theorem (Theorem 1, Appendix E.5) is used to rewrite the two-step inference process as a single BCQ reconstruction with adjusted parameters. The authors add unified initialization, local and periodic mapping, and report experiments on Mistral 7B, Llama-3 8B, and Llama-3 70B on MMLU, WikiText2, and GSM8K, claiming up to 4.60% higher accuracy over the included baselines.

Significance. If the central claim holds, the paper makes a useful and practical contribution: it provides an exact algebraic bridge between a uniform-quantization-style training procedure and a BCQ-style deployment format, and it attacks a real limitation of BCQ methods by transferring UQ optimization techniques. The unification theorem proof is correct for a fixed binary code, the ablation study in Table 5 supports all three proposed components, the analysis of local index changes in Table 2 motivates the remapping scheme, and the authors make the code public. However, the main accuracy claim is currently supported only by comparisons against a narrow baseline set, and there is a train-deploy inconsistency in how binary codes are selected. These issues need to be resolved before the empirical and conceptual claims are fully established.

major comments (3)
  1. [Algorithm 1, Eq. (3)-(4), Appendix E.5, Table 15] The deployed binary codes are not the codes optimized during training. During optimization (Algorithm 1, lines 4-7), Eq. (3) selects C_train = argmin_{C'} ||TF(w) - (C'α + z_B 1_g)|| with TF(w) = w⊘(Δ·s·s_r) + z_U 1_g, so after multiplying by Δ the selected code minimizes ||w/(s·s_r) - (C'α* + z*_B 1_g)||. At deployment (Algorithm 1, line 11, Eq. (4)), the code is instead C_deploy = argmin_{C'} ||w - (C'α* + z*_B 1_g)||. These two argmins coincide only when s = s_r = 1. Table 15 shows that in the 3-bit Llama-3 8B model the learned scales change the mapped index for about 3.96% of weights (3.86% by one level and about 0.10% by two or more). Consequently, the model obtained after line 11 is not the model whose block-reconstruction loss was minimized in lines 5-7, and the unification theorem, which is correct as an algebraic identity for a fixed C, does not by itself transfer the optimized accuracy to the deployed quantizer. Please either modify the algorithm so that training and deployment optimize the same code-assignment rule, or quantify the reconstruction gap and justify why the trained parameters remain optimal for the recomputed codes.
  2. [Tables 3-4, Section 4.1] The main comparison omits two standard UQ baselines, GPTQ and AWQ, even though both are discussed in Section 5 and Appendix B and are supported by the same inference kernels (Table 8). The abstract and Section 1 claim that UniQuanF outperforms existing UQ and BCQ methods, but Tables 3 and 4 include only RTN, OmniQuant, FlexRound, and ALTERNATING. Please add GPTQ and AWQ under the same evaluation and kernel settings, or restrict the empirical claim to the included baselines.
  3. [Tables 3-4, Appendix D] The main accuracy tables report a single number per configuration even though Appendix D states that results are averaged over random seeds 0, 1, and 2 (and seed 0 for 70B). Several margins are small, for example 3-bit Mistral 7B MMLU 53.68 vs FlexRound 53.45 and 4-bit Llama-3 70B WikiText2 3.19 vs 3.31, so without standard deviations or per-seed values the reader cannot judge whether the reported improvements are systematic. Please report variance or per-seed results for the main tables.
minor comments (3)
  1. [Appendix E.1, Algorithm 3] Algorithm 2 line 6 calls general-alternating(¯w, z_B, G, T), but Algorithm 3's input list omits z_B even though lines 4 and 6 use it; this is a typo that should be fixed.
  2. [Section 3.3, Table 2] The locality statistics in Table 2 are reported for a single configuration (3-bit Llama-3 8B), while the local and periodic mapping is then used for all models and bit-widths. Since Algorithm 4 assumes index changes are local and sparse, a sentence or a small table showing that this property holds for other configurations would strengthen the generality claim.
  3. [Appendix F.4, Table 13] Fixed-maximum clipping strategy gives 54.34% MMLU on 3-bit Llama-3 8B, higher than the Fixed-minimum strategy (53.46%) listed as the default in Table 10; please clarify how the default clipping strategy was selected and whether the main-text results would change if the better strategy were used.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the unification theorem is a direct algebraic identity, and the reported accuracy gains are empirical rather than constructed from the fitted parameters.

full rationale

The claimed derivation chain is self-contained. The unified quantizer in Eq. 3 composes FlexRound's affine transform, BCQ's nearest-code mapping, and FlexRound's affine detransform. Theorem 1 is proved in Appendix E.5 by direct substitution from the definitions of R_B and D_R: D_R(R_B(C; Θ_B); Θ_R) = Δ((Cα + z_B 1) − z_U 1) = C(Δα) + Δ(z_B − z_U)1 = R_B(C; {Δα, Δ(z_B − z_U)}). This is an algebraic identity, not a fitted equality; the paper supplies the proof rather than importing it from an unverified self-citation. The optimized parameters Θ_IF are fitted to a block-wise reconstruction loss (Algorithm 1, lines 5–7), while the headline accuracies are downstream task measurements; no reported benchmark number is defined as a function of the fitted parameters. The main self-citation, LUT-GEMM (Park et al., 2024a), supports kernel latency and BCQ expressiveness, and its latency tables are reproduced in Appendix B as externally checkable measurements, so the citation does not carry the derivation. The train/deploy recomputation of C (Algorithm 1, line 11 vs. Eq. 3) is a genuine consistency question: with learned s and sr, the deployed argmin in the original weight space may differ from the trained argmin in the transformed space. That is a correctness concern, not a circular one. Likewise, calibrating on 128 GSM8K sequences and then evaluating on GSM8K is a benchmark-calibration concern, not a case where the accuracy is forced by construction. The Limitations section honestly scopes the experimental validation to FlexRound + ALTERNATING; it does not reveal a circular dependency. Under the quoted-equation standard, no circular step is exhibited.

Assumptions & free parameters 10 free parameters · 5 assumptions · 0 invented entities

The ledger shows that UniQuanF's accuracy rests on a learnable per-weight rescaling (s, sr) plus BCQ parameters (alpha, zB) and UQ clipping parameters (Delta, zU), all fitted to calibration data, together with several empirical premises about the stability of the mapping, the validity of the straight-through estimator, and the speed of the self-cited LUT-GEMM kernel.

free parameters (10)
  • Per-weight scale s = learned via SGD, one scalar per weight
    Element-wise division factors in TF (Eq. 1) that let each weight explore different quantization levels; trained on the block-wise reconstruction loss.
  • Row-wise scale s_r = learned via SGD, one scalar per row
    Row-wise division factor in TF (Eq. 1) trained along with s.
  • BCQ scale factors alpha = learned per weight group, k values
    Binary-coding scale factors optimized via alternating update and SGD; define the non-uniform quantization levels.
  • BCQ shift z_B = learned per weight group
    Center of the binary tree of quantization levels; initialized to (2^k - 1)/2 and refined by alternating update or grid search.
  • UQ scale Delta and zero-point z_U = selected by grid search over G iterations
    Clipping range parameters inherited from FlexRound; chosen by minimizing quantization error in unified initialization.
  • Gradient filtering threshold tau = set to min(alpha)
    Weights with transformed-space mapping error larger than the smallest BCQ scale have their gradients zeroed; this rule is justified by analogy to UQ clipping, not derived.
  • Remapping period p = 2 (or 4)
    Hyperparameter controlling how often local mapping is refreshed; chosen via sensitivity analysis (Appendix F.3).
  • Learning rates for Theta_F and Theta_B = 0.005 and 0.0005
    Fixed SGD learning rates for the two parameter groups; not tuned per model.
  • Grid search iterations G and alternating update iterations T = G in {1,30}, T = 15
    Search depths for unified initialization; selected by sensitivity analysis.
  • Clipping strategy = Fixed-minimum or Fixed-maximum
    The grid search adjusts the clipping range in three ways (Appendix E.2); the better strategy is chosen per experiment (Table 13).
assumptions (5)
  • domain assumption Block-wise output reconstruction loss on a small calibration set (128 sequences of length 2048) transfers to the evaluation benchmarks.
    The method optimizes each block to match the unquantized block's outputs on C4 or GSM8K samples, then reports MMLU, WikiText2, and GSM8K accuracy; this standard PTQ assumption is invoked throughout Section 4.
  • domain assumption Straight-through estimator is valid for the mapping function when mapping errors are below tau = min(alpha).
    Appendix C.1 uses STE to propagate gradients through non-differentiable MB and filters weights with large mapping errors; no theoretical guarantee beyond the UQ analogy.
  • domain assumption Index changes of mapped quantization levels are rare and local during optimization.
    Table 2 reports that per-step 99.97% of indices are stable and changes are mostly to adjacent levels; this empirical premise underpins Algorithm 4.
  • domain assumption LUT-GEMM kernel supports BCQ and UQ at the same speed, making the deployment-cost comparison valid.
    The claim of no extra deployment cost relies on the BCQ kernel from Park et al. (2024a), cited in Section 2.1 and Appendix B; the latency tables are reproduced from that self-cited paper.
  • domain assumption The pretrained weight matrices can be well approximated per-row (or per-128 group) by the unified quantizer.
    The paper chooses row-wise weight groups (group size 128 for Llama-3 70B) without studying other groupings; if weights are not locally structured, the quantizer's capacity is insufficient.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unifying Uniform and Binary-coding Quantization for Accurate Compression of Large Language Models." pith.science (2026). https://pith.science/paper/HDTEO7UZ

@misc{pith2026250603781,
  author       = {Pith},
  title        = {Pith review of: Unifying Uniform and Binary-coding Quantization for Accurate Compression of Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HDTEO7UZ}},
  note         = {Machine review of arXiv:2506.03781}
}
read the original abstract

How can we quantize large language models while preserving accuracy? Quantization is essential for deploying large language models (LLMs) efficiently. Binary-coding quantization (BCQ) and uniform quantization (UQ) are promising quantization schemes that have strong expressiveness and optimizability, respectively. However, neither scheme leverages both advantages. In this paper, we propose UniQuanF (Unified Quantization with Flexible Mapping), an accurate quantization method for LLMs. UniQuanF harnesses both strong expressiveness and optimizability by unifying the flexible mapping technique in UQ and non-uniform quantization levels of BCQ. We propose unified initialization, and local and periodic mapping techniques to optimize the parameters in UniQuanF precisely. After optimization, our unification theorem removes computational and memory overhead, allowing us to utilize the superior accuracy of UniQuanF without extra deployment costs induced by the unification. Experimental results demonstrate that UniQuanF outperforms existing UQ and BCQ methods, achieving up to 4.60% higher accuracy on GSM8K benchmark.

Figures

Figures reproduced from arXiv: 2506.03781 by the authors.

Figure 1
Figure 1. A comparison of (a) uniform, (b) binary-coding, and (c) unified quantization schemes. UQ has strong [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustrations of quantization levels (q) as￾signed for a weight group w under (a) UQ and (b) BCQ schemes, where wm and wM are the minimum and the maximum weights in w, respectively. UQ has evenly￾spaced quantization levels within a clipping range while BCQ has non-uniform quantization levels determined by its scale factors α and a shifting factor zB. in Section 5. Finally, we conclude the paper with a summary of our… view at source ↗
Figure 3
Figure 3. Quantization processes for a weight w in (a) UQ, (b) BCQ, and (c) UniQuan schemes. Blue-colored processes are parameterized functions which are the source of UQ’s strong optimizability and BCQ’s powerful expressiveness, respectively. UniQuan takes the advantages of both UQ and BCQ schemes by combining the parameterized functions in both schemes. See Section 3.1 for details. form quantization (UQ) and binary-coding q… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: (a) Quantization levels learned by UniQuan [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: An illustration of a Transformer architecture [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Change of average accuracy on 0-shot and 5-shot MMLU benchmarks with regard to the change [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: Index difference of mapped quantization lev [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 29 canonical work pages

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Croci, Bo Li, Martin Jaggi, Dan Alistarh, Torsten Hoefler, and James Hensman

    Saleh Ashkboos, Amirkeivan Mohtashami, Maximilian L. Croci, Bo Li, Martin Jaggi, Dan Alistarh, Torsten Hoefler, and James Hensman. 2024. Quarot: Outlier-free 4-bit inference in rotated llms. CoRR, abs/2404.00456

    arXiv 2024

  4. [4]

    Courville

    Yoshua Bengio, Nicholas L \' e onard, and Aaron C. Courville. 2013. Estimating or propagating gradients through stochastic neurons for conditional computation. CoRR, abs/1308.3432

    arXiv 2013

  5. [5]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877--1901

    2020

  6. [6]

    Ikhyun Cho and U Kang. 2022. Pea-kd: Parameter-efficient and accurate knowledge distillation on bert. PLOS ONE, 17(2)

    work page 2022

  7. [7]

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168

    arXiv 2021

  8. [8]

    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 Systems, 35:30318--30332

    2022

Show all 51 references
  1. [9]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    2024 arXiv

  2. [10]

    Vage Egiazarian, Andrei Panferov, Denis Kuznedelev, Elias Frantar, Artem Babenko, and Dan Alistarh. 2024. Extreme compression of large language models via additive quantization. In Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27...

    2024

  3. [11]

    Elias Frantar and Dan Alistarh. 2023. Sparsegpt: Massive language models can be accurately pruned in one-shot. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA , volume 202 of Proceedings of Machine Learning Research, pages 103...

    2023

  4. [12]

    Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. 2023. https://openreview.net/forum?id=tcbBPnfwxS OPTQ : Accurate quantization for generative pre-trained transformers . In The Eleventh International Conference on Learning Representations

    2023

  5. [13]

    Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac'h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang S...

    2023 doi

  6. [14]

    Yuxian Gu, Li Dong, Furu Wei, and Minlie Huang. 2024. https://openreview.net/forum?id=5h0qf7IBZZ Minillm: Knowledge distillation of large language models . In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024 . OpenReview.net

    2024

  7. [15]

    Suyog Gupta, Ankur Agrawal, Kailash Gopalakrishnan, and Pritish Narayanan. 2015. Deep learning with limited numerical precision. In International conference on machine learning, pages 1737--1746. PMLR

    2015

  8. [16]

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

    2021

  9. [17]

    Jun - Gi Jang, Chun Quan, Hyun Dong Lee, and U Kang. 2023. Falcon: lightweight and accurate convolution based on depthwise separable convolution. Knowl. Inf. Syst., 65(5):2225--2249

    2023

  10. [18]

    Hyojin Jeon, Seungcheol Park, Jin-Gee Kim, and U. Kang. 2023. Pet: Parameter-efficient knowledge distillation on transformer. PLOS ONE, 18(7)

    2023

  11. [19]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de Las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, L \' e lio Renard Lavaud, Marie - Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, ...

    2023 arXiv

  12. [20]

    Junghun Kim, Jinhong Jung, and U. Kang. 2021. Compressing deep graph convolution network with multi-staged knowledge distillation. PLOS ONE, 16

    2021

  13. [21]

    Minjun Kim, Jaehyeon Choi, Jongkeun Lee, Wonjin Cho, and U Kang. 2025 a . Zero-shot quantization: A comprehensive survey. In IJCAI

    2025

  14. [22]

    Minjun Kim, Jongjin Kim, and U Kang. 2025 b . Synq: Accurate zero-shot quantization by synthesis-aware fine-tuning. In ICLR

    2025

  15. [23]

    Sehoon Kim, Coleman Richard Charles Hooper, Amir Gholami, Zhen Dong, Xiuyu Li, Sheng Shen, Michael W Mahoney, and Kurt Keutzer. 2024. Squeezellm: Dense-and-sparse quantization. In Forty-first International Conference on Machine Learning

    2024

  16. [24]

    Se Jung Kwon, Jeonghoon Kim, Jeongin Bae, Kang Min Yoo, Jin - Hwa Kim, Baeseong Park, Byeongwook Kim, Jung - Woo Ha, Nako Sung, and Dongsoo Lee. 2022. https://doi.org/10.18653/V1/2022.FINDINGS-EMNLP.240 Alphatuning: Quantization-aware parameter-efficient adaptation of large-sc...

    2022 doi

  17. [25]

    Hyun Dong Lee, Seongmin Lee, and U. Kang. 2021. Auber: Automated bert regularization. PLOS ONE, 16(6)

    2021

  18. [26]

    Jung Hyun Lee, Jeonghoon Kim, Se Jung Kwon, and Dongsoo Lee. 2023. https://proceedings.mlr.press/v202/lee23h.html Flexround: Learnable rounding based on element-wise division for post-training quantization . In International Conference on Machine Learning, ICML 2023, 23-29 Jul...

    2023

  19. [27]

    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 Systems, 6:87--100

    2024

  20. [28]

    Yifei Liu, Jicheng Wen, Yang Wang, Shengyu Ye, Li Lyna Zhang, Ting Cao, Cheng Li, and Mao Yang. 2024 a . VPTQ: extreme low-bit vector post-training quantization for large language models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing...

    2024

  21. [29]

    Zechun Liu, Changsheng Zhao, Igor Fedorov, Bilge Soran, Dhruv Choudhary, Raghuraman Krishnamoorthi, Vikas Chandra, Yuandong Tian, and Tijmen Blankevoort. 2024 b . https://doi.org/10.48550/ARXIV.2405.16406 Spinquant: LLM quantization with learned rotations . CoRR, abs/2405.16406

    2024 doi arXiv

  22. [30]

    Xinyin Ma, Gongfan Fang, and Xinchao Wang. 2023. Llm-pruner: On the structural pruning of large language models. In NeurIPS

    2023

  23. [31]

    Xin Men, Mingyu Xu, Qingyu Zhang, Bingning Wang, Hongyu Lin, Yaojie Lu, Xianpei Han, and Weipeng Chen. 2024. Shortgpt: Layers in large language models are more redundant than you expect. arXiv

    2024

  24. [32]

    Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. 2017. Pointer sentinel mixture models. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings . OpenReview.net

    2017

  25. [33]

    Gunho Park, Baeseong Park, Minsub Kim, Sungjae Lee, Jeonghoon Kim, Beomseok Kwon, Se Jung Kwon, Byeongwook Kim, Youngjoo Lee, and Dongsoo Lee. 2024 a . LUT-GEMM: quantized matrix multiplication based on luts for efficient inference in large-scale generative language models. In...

    2024

  26. [34]

    Seungcheol Park, Hojun Choi, and U Kang. 2024 b . Accurate retraining-free pruning for pretrained encoder-based language models. In ICLR

    2024

  27. [35]

    Seungcheol Park, Jaehyeon Choi, Sojin Lee, and U Kang. 2024 c . A comprehensive survey of compression algorithms for language models. arXiv preprint arXiv:2401.15347

    2024 arXiv

  28. [36]

    Seungcheol Park, Sojin Lee, Jongjin Kim, Jinsik Lee, Hyunjik Jo, and U Kang. 2025. Accurate sublayer pruning for large language models by exploiting latency and tunability information. In IJCAI

    2025

  29. [37]

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. 2019. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32

    2019

  30. [38]

    Tairen Piao, Ikhyun Cho, and U Kang. 2022. Sensimix: Sensitivity-aware 8-bit index & 1-bit value mixed precision quantization for bert compression. PloS one, 17(4):e0265621

    2022

  31. [39]

    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. J. Mach. Learn. Res., 21:140:1--140:67

    2020

  32. [40]

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

    2024

  33. [41]

    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 arXiv

  34. [42]

    Albert Tseng, Jerry Chee, Qingyao Sun, Volodymyr Kuleshov, and Christopher De Sa. 2024. Quip \# : Even better LLM quantization with hadamard incoherence and lattice codebooks. In Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, ...

    2024

  35. [43]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30

    2017

  36. [44]

    Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. 2020. Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Proce...

    2020

  37. [45]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R \' e mi Louf, Morgan Funtowicz, and Jamie Brew. 2019. https://arxiv.org/abs/1910.03771 Huggingface's transformers: State-of-the-art natural language processin...

    2019 arXiv

  38. [46]

    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. In International Conference on Machine Learning, pages 38087--38099. PMLR

    2023

  39. [47]

    Chen Xu, Jianqiang Yao, Zhouchen Lin, Wenwu Ou, Yuanbin Cao, Zhirong Wang, and Hongbin Zha. 2018. https://openreview.net/forum?id=S19dR9x0b Alternating multi-bit quantization for recurrent neural networks . In 6th International Conference on Learning Representations, ICLR 2018...

    2018

  40. [48]

    Jaemin Yoo, Minyong Cho, Taebum Kim, and U Kang. 2019. Knowledge extraction with no observable data. In NeurIPS

    2019

  41. [49]

    Haoran You, Yipin Guo, Yichao Fu, Wei Zhou, Huihong Shi, Xiaofan Zhang, Souvik Kundu, Amir Yazdanbakhsh, and Yingyan Celine Lin. 2024. Shiftaddllm: Accelerating pretrained llms via post-training multiplication-less reparameterization. arXiv preprint arXiv:2405.04532

    2024 arXiv

  42. [50]

    Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. 2022. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068

    2022 arXiv

  43. [51]

    Longguang Zhong, Fanqi Wan, Ruijun Chen, Xiaojun Quan, and Liangzhi Li. 2024. Blockpruner: Fine-grained pruning for large language models. arXiv

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.