Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Grouped Sequency-arranged Rotation: Optimizing Rotation Transformation for Quantization for Free

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read Reordering a Hadamard matrix by sequency and applying it in blocks makes training-free 2-bit LLM quantization competitive with optimized methods.

desk verdict The method looks genuinely useful at 2-bit, but the paper's own numbers don't isolate the sequency ordering it claims is load-bearing. read the letter →

arxiv 2505.03810 v2 pith:GTE2ZNWH submitted 2025-05-02 cs.LG cs.AIcs.CL

Euntae Choi , Sumin Song , Woosang Lim , Sungjoo Yoo This is my paper · ORCID
classification cs.LGcs.AIcs.CL
keywords post-trainingquantizationLLMrotationtransformationWalsh-Hadamardtransformsequencyordering2-bitweightsgroupoutliermitigation
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

The paper claims that a zero-training change to how LLM weights are rotated before quantization can match, and in some settings beat, optimization-based 2-bit quantization methods. Its central move is to replace the standard randomized Hadamard rotation with a Walsh matrix—the same Hadamard matrix with rows reordered by sequency, the number of sign changes per row—and to apply it in blocks aligned with the quantization group. On Llama-2-7B with 2-bit weights and 16-bit activations, this Grouped Sequency-arranged Rotation (GSR) lowers WikiText-2 perplexity from 20.29 with the standard rotation to 11.59, within 0.62 of the trained OSTQuant's 10.97, while raising average zero-shot accuracy from 32.06 percent to 42.44 percent. Under 2-bit weights and 4-bit activations, GSR gives perplexity 15.23, better than OSTQuant's 16.16. If it holds, this matters because 2-bit quantization would roughly quarter the memory footprint of deployed models without requiring the optimization phase that current low-bit methods need.

What carries the argument

The central object is the Walsh matrix: a Hadamard matrix with rows permuted by bit-reversal and Gray-code so that the sequency of each row, the number of sign flips, is in ascending order. Sequency is treated as the discrete analogue of frequency, so filters of similar sequency produce rotated weight groups with smaller within-group outlier spread. GSR combines this ordering with a block-diagonal rotation whose block size equals the quantization group size, so each quantization group is rotated independently and remaining outliers stay inside one group. A supporting identity, Observation #1, shows that under group quantization each row group of the rotated weight is generated by the corresponding group of filters in the front rotation matrix and by all filters of the rear matrix, which is what lets the paper attribute the improvement to the arrangement of sequency filters.

What would settle it

Take a weight matrix from Llama-2-7B, rotate it separately with the global Hadamard and the global Walsh matrix, and measure the maximum absolute value within each quantization group before quantizing. If the Walsh-rotated groups do not have smaller within-group outlier magnitudes, or if the per-group quantization error is not lower, the stated sequency mechanism is falsified.

Watch

Extended reading notes

Core claim

The central discovery is that the ordering of rows inside a rotation matrix matters for group quantization, not just the rotation as a whole. Replacing the natural-order Hadamard matrix used in QuaRot with a Walsh matrix—the same entries permuted so that sequency increases linearly—reduces within-group variance of the rotated weights, and applying this Walsh rotation as a block-diagonal matrix whose block size matches the quantization group confines outliers to a single group. The paper shows that on Llama-2-7B, this training-free GSR achieves WikiText-2 PPL 11.59 and 42.44 percent average zero-shot accuracy at W2A16, compared with 20.29 and 32.06 percent for the standard global randomized Hadamard rotation, and that at W2A4 it reaches PPL 15.23, lower than the optimization-based OSTQuant's 16.16. The same change also improves SpinQuant and OSTQuant when it replaces their rotation, for example cutting OSTQuant's W2 PPL from 10.97 to 9.03.

Load-bearing premise

The argument stands on the claim, stated rather than derived, that sequency behaves like frequency in filtering, so that sorting rotation filters by sequency produces rotated weight groups with fewer massive outliers.

Editorial extensions

If this is right

  • Training-free 2-bit weight quantization of Llama-2-7B reaches WikiText-2 PPL 11.59, within 0.62 of the optimization-based OSTQuant's 10.97, while skipping the training phase entirely.
  • Under W2A4, GSR applied to QuaRot reaches PPL 15.23, outperforming OSTQuant's 16.16 with fewer resources.
  • GSR also strengthens learned-rotation methods when used as their rotation: OSTQuant with GSR improves W2 PPL from 10.97 to 9.03 and zero-shot accuracy from 45.52 percent to 50.51 percent.
  • The sequency reordering alone, without local blocking, improves QuaRot's W2 PPL from 20.29 to 15.38, showing that row ordering and block structure make separable contributions.

Reading between the lines

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

  • The paper does not test whether the benefit persists if the block size of the rotation and the quantization group size are decoupled; if the mechanism is right, misaligning them should create a sharp performance drop, which would be a direct test of the confinement claim.
  • The authors note in the appendix that local rotation on R4 disables the fast Hadamard transform and requires storing the full matrix in memory; this practical cost is not reflected in the headline perplexity numbers.
  • Sequency ordering is one specific permutation of Hadamard rows, and the paper does not explore other permutations that might further reduce within-group outliers for particular weight distributions.
  • The improvement is claimed only for extremely low-bit group quantization; the paper's own limitation statement says that at larger bit-widths the quantization error becomes small enough that sequency alignment shows no visible gain.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes Grouped Sequency-arranged Rotation (GSR), a training-free modification of the Hadamard rotations used in LLM post-training quantization. GSR replaces the standard Hadamard matrix with a sequency-ordered Walsh matrix, optionally in block-diagonal (local) form with block size equal to the quantization group size. The authors evaluate GSR on Llama-2-7B at W2A16 and W2A4 within QuaRot, SpinQuant, and OSTQuant pipelines, reporting improved WikiText-2 perplexity and zero-shot accuracy compared to the baseline rotations, and claiming performance comparable to optimization-based OSTQuant without training.

Significance. If the empirical claims hold, GSR is a practically attractive contribution: it has no learned parameters, is a drop-in replacement for the R1 rotation, and it also improves learned-rotation pipelines such as SpinQuant and OSTQuant. The evaluation is consistent across three quantization pipelines and includes ablations, which is a strength. However, the causal claim that sequency ordering is responsible for the improvement is not currently established, because the Hadamard baselines are randomized while the Walsh variants are not. The theoretical justification in Section 3.2 is an intuition rather than a derivation. The contribution may be salvageable with additional control experiments, but the central mechanism needs to be tested directly.

major comments (4)
  1. [Appendix A.1, Table 2] The comparisons labeled as isolating the sequency arrangement are confounded with randomization. Appendix A.1 states that all Hadamard matrices are randomized while Walsh matrices use the original Hadamard matrix. Therefore each Hadamard-versus-Walsh comparison in Table 2, including GH vs. GW and LH vs. GSR, changes two things at once: the row sequency order and the presence of random sign flips. Random sign flips are not a negligible detail; they are the incoherence mechanism used by QuIP# and QuaRot. The statement in Section 3.2 that randomization keeps the sequency arrangement 'with no significant changes' is not a proof. To support the paper's central claim, the authors should add the missing control cells: unrandomized global Hadamard, randomized global Walsh, unrandomized local Hadamard, and randomized local Walsh.
  2. [Section 3.2, Eq. (4) and Observation #1] Observation #1 is a matrix-product identity, not a derivation that sequency ordering reduces quantization error. The text moves from 'smaller sequency variance within each column group' to 'the Walsh matrix will produce rotated weight groups with fewer massive outliers' without a formal argument connecting sequency variance to the quantization error of a group quantizer (RTN or GPTQ). A concrete bound, or at least a variance calculation on the entries of the rotated weight groups, is needed before the title's claim that the transformation is 'optimized' for quantization can be accepted. Alternatively, the paper should explicitly reframe Section 3.2 as an empirical hypothesis rather than a justification.
  3. [Section 3.2, paragraph after Observation #1] The claim that 'For Rr, the arrangement has no impact as long as the set of sequency values is equal' is not justified and appears inconsistent with Eq. (4). Since W' = Rf^{-1} W Rr, the row arrangement of Rr changes which linear combination of columns of the intermediate matrix is taken in each row of W'. Unless the quantizer is invariant under the resulting column transformation, the element multiset of a row quantization group can change. The authors should either prove the invariance for group quantization or run a control that swaps only the rear rotation Rr between Hadamard and Walsh order.
  4. [Table 2, zero-shot columns] The zero-shot averages are not directly comparable across methods because they are computed over different task sets. QuaRot and SpinQuant use eight tasks including both LAMBADA variants, while OSTQuant uses nine tasks including BoolQ, OpenBookQA, and SIQA. The headline claim that GSR is 'comparable' to OSTQuant is therefore mainly supported by the WikiText-2 PPL column. The authors should either report per-overlap-task averages for cross-method comparisons or clearly mark that the zero-shot columns refer to different task sets.
minor comments (5)
  1. [Section 4] There is a typo: 'Walsh matirx' should be 'Walsh matrix'.
  2. [Eq. (3)] The Walsh matrix is written as an entrywise ±1 matrix, but for it to be a rotation matrix it must be scaled by 1/sqrt(G). This normalization should be stated explicitly in the definition of Hwal.
  3. [Section 3.2, comparing RHT and Walsh] The description that randomization 'flips the signs of diagonal elements in a Hadamard matrix' is imprecise. The randomized Hadamard transform multiplies the Hadamard matrix by a random diagonal sign matrix, i.e., H D or D H, which is not the same as flipping only the diagonal entries of H. Since this distinction is directly relevant to the missing control cells, the wording should be corrected.
  4. [References] Several references have formatting issues: 'Yonatan Bisk, ..., and 1 others', 'Hugo Touvron, ..., and 1 others', and 'Le Dinh Chon Tam and R.Y . Goulet' have spacing/author-list problems. These should be cleaned up.
  5. [Appendix A.2] The ablation on R4 mentions that local rotation disables the fast Hadamard transform and requires storing an FP32 tensor during inference. Since GSR is applied to R1 in the main experiments, clarify whether the same inference-time overhead applies to R1 or whether R1 is fully fused into the weights offline.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: GSR is a fixed, parameter-free rotation construction evaluated on external benchmarks, with no fitted input renamed as prediction and no load-bearing self-citation chain.

full rationale

The paper's central claim is that replacing the Hadamard rotation with a sequency-ordered Walsh rotation, optionally in block-diagonal form, reduces 2-bit group-quantization error on Llama-2-7B. Nothing in the derivation reduces by construction to its inputs. GSR contains no learned or fitted parameters: the only design choice is the block size G, which is matched to the quantization group size, and the resulting PPL/accuracy numbers are measured on external benchmarks (WikiText-2, ARC, HellaSwag, PIQA, etc.). The mechanism discussion in Section 3.2 is a qualitative signal-processing-style argument ('sequency is analogous to frequency... the Walsh matrix will produce rotated weight groups with fewer massive outliers'), and Observation #1 in Eq. (5) is a matrix-product identity used to frame the argument, not an assumption of the conclusion. The paper cites prior methods (QuaRot, SpinQuant, OSTQuant, QuIP#, DuQuant, etc.) for context and baselines, but none of these citations is by the present authors, and none is invoked as an unverified uniqueness theorem to rule out alternatives. The Walsh matrix is defined via a fixed bit-reversal/Gray-code permutation of the Sylvester Hadamard matrix, so the method is fully specified without reference to the outcome it predicts. The only notable issue is an experimental confound noted in Appendix A.1: 'All Hadamard matrices are randomized... When constructing Walsh matrices, the original Hadamard matrix is used,' which means GH-versus-GW and LH-versus-GSR comparisons change both sequency order and row-randomization simultaneously. That is a potential threat to the causal attribution of the improvement, but it is a correctness/experimental-design concern, not a circularity: the reported numbers are not mathematically forced by the inputs, and the paper does not fit a parameter and then present the same fitted value as a prediction. Under the hard rules requiring quoted, explicit reduction to one's own inputs, no circular step can be exhibited.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

No fitted parameters except block size G=128, which is taken from the quantization group. The main burden is the heuristic that sequency ordering reduces outliers; this is an assumption rather than a derived theorem, and prior local-rotation work is imported.

free parameters (1)
  • Walsh block size G = 128
    Block size is set equal to the GPTQ quantization group size used in all baselines. No sweep over G is reported, so performance depends on this hand-chosen alignment.
assumptions (4)
  • standard math Hadamard and Walsh matrices are orthogonal up to scaling, so row permutations preserve rotation properties.
    Used to construct R_GSR in Section 3.1.
  • domain assumption Rows of the rotation matrix can be associated with frequency (sequency), and lower within-group sequency variance causes fewer quantization outliers.
    Central mechanism in Section 3.2; asserted on analogy with frequency-domain filtering, not proven.
  • domain assumption Local (block-diagonal) rotation confines outlier effects within blocks and improves quantization robustness.
    Taken from DuQuant and DUAROT (Lin et al. 2024; Xiang et al. 2025) and used in the GSR design.
  • domain assumption Quantization group boundaries align with the blocks of R_GSR.
    Needed for Observation #1 to apply; assumed by setting G equal to the group size.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Grouped Sequency-arranged Rotation: Optimizing Rotation Transformation for Quantization for Free." pith.science (2026). https://pith.science/paper/GTE2ZNWH

@misc{pith2026250503810,
  author       = {Pith},
  title        = {Pith review of: Grouped Sequency-arranged Rotation: Optimizing Rotation Transformation for Quantization for Free},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GTE2ZNWH}},
  note         = {Machine review of arXiv:2505.03810}
}
read the original abstract

Large Language Models (LLMs) face deployment challenges due to high computational costs, and while Post-Training Quantization (PTQ) offers a solution, existing rotation-based methods struggle at very low bit-widths like 2-bit. We introduce a novel, training-free approach to construct an improved rotation matrix, addressing the limitations of current methods. The key contributions include leveraging the Walsh-Hadamard transform with sequency ordering, which clusters similar frequency components to reduce quantization error compared to standard Hadamard matrices, significantly improving performance. Furthermore, we propose a Grouped Sequency-arranged Rotation (GSR) using block-diagonal matrices with smaller Walsh blocks, effectively isolating outlier impacts and achieving performance comparable to optimization-based methods without requiring any training. Our method demonstrates robust performance on reasoning tasks and Perplexity (PPL) score on WikiText-2. Our method also enhances results even when applied over existing learned rotation techniques.

Figures

Figures reproduced from arXiv: 2505.03810 by the authors.

Figure 1
Figure 1. Overall diagram of rotation scheme. We ap [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of global and local rotation strate [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. When Local Variance Optimality Is Not Enough: RoPE-Aligned Q/K Rotations for Dynamic 4-Bit Quantisation

    cs.LG 2026-08 conditional novelty 6.0 of 10

    Under dynamic W4A4KV4 quantisation, a head-shared RoPE-aligned pairwise rotation that exactly minimizes a pooled variance surrogate still yields higher perplexity than full-head Hadamard mixing in all evaluated comparisons.

Reference graph

Works this paper leans on

21 extracted references · 7 canonical work pages · cited by 1 Pith paper

  1. [1]

    Saleh Ashkboos, Amirkeivan Mohtashami, Maximilian L Croci, Bo Li, Martin Jaggi, Dan Alistarh, Torsten Hoefler, and James Hensman. 2024. https://arxiv.org/abs/2405.04517 Quarot: Outlier-free 4-bit inference in rotated llms . In Thirty-eighth Conference on Neural Information Processing Systems

    arXiv 2024

  2. [2]

    Yonatan Bisk, Rowan Zellers, Jianfeng Gao, Yejin Choi, and 1 others. 2020. Piqa: Reasoning about physical commonsense in natural language. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 7432--7439

    2020

  3. [3]

    Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. 2019. https://doi.org/10.18653/v1/N19-1300 B ool Q : Exploring the surprising difficulty of natural yes/no questions . In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language...

    doi:10.18653/v1/n19-1300 2019

  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:1803.05457v1

    arXiv 2018

  5. [5]

    Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. 2022. Optq: Accurate quantization for generative pre-trained transformers. In The Eleventh International Conference on Learning Representations

    work page 2022

  6. [6]

    Xing Hu, Yuan Cheng, Dawei Yang, Zhixuan Chen, Zukang Xu, JiangyongYu, XUCHEN, Zhihang Yuan, Zhe jiang, and Sifan Zhou. 2025. https://openreview.net/forum?id=rAcgDBdKnP OSTQ uant: Refining large language model quantization with orthogonal and scaling transformations for better distribution fitting . In The Thirteenth International Conference on Learning R...

    work page 2025

  7. [7]

    Haokun Lin, Haobo Xu, Yichen Wu, Jingzhi Cui, Yingtao Zhang, Linzhan Mou, Linqi Song, Zhenan Sun, and Ying Wei. 2024. Duquant: Distributing outliers via dual transformation makes stronger quantized llms. Advances in Neural Information Processing Systems, 37:87766--87800

    work page 2024

  8. [8]

    Zechun Liu, Changsheng Zhao, Igor Fedorov, Bilge Soran, Dhruv Choudhary, Raghuraman Krishnamoorthi, Vikas Chandra, Yuandong Tian, and Tijmen Blankevoort. 2025. https://openreview.net/forum?id=ogO6DGE6FZ Spinquant: Llm quantization with learned rotations . In The Thirteenth International Conference on Learning Representations

    work page 2025

Show all 21 references
  1. [9]

    Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. 2017. https://openreview.net/forum?id=Byj72udxe Pointer sentinel mixture models . In International Conference on Learning Representations

    2017

  2. [10]

    Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. 2018. https://doi.org/10.18653/v1/D18-1260 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 P...

    2018 doi

  3. [11]

    Denis Paperno, Germ \'a n Kruszewski, Angeliki Lazaridou, Ngoc Quan Pham, Raffaella Bernardi, Sandro Pezzelle, Marco Baroni, Gemma Boleda, and Raquel Fern \'a ndez. 2016. https://doi.org/10.18653/v1/P16-1144 The LAMBADA dataset: Word prediction requiring a broad discourse cont...

    2016 doi

  4. [12]

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

    2021

  5. [13]

    Maarten Sap, Hannah Rashkin, Derek Chen, Ronan Le Bras, and Yejin Choi. 2019. https://doi.org/10.18653/v1/D19-1454 Social IQ a: Commonsense reasoning about social interactions . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9...

    2019 doi

  6. [14]

    Le Dinh Chon Tam and R.Y. Goulet. 1972. https://doi.org/10.1109/T-C.1972.223524 On arithmetical shift for walsh functions . IEEE Transactions on Computers, C-21(12):1451--1452

    1972

  7. [15]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, and 1 others. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288

    2023 arXiv

  8. [16]

    Albert Tseng, Jerry Chee, Qingyao Sun, Volodymyr Kuleshov, and Christopher De Sa. 2024. https://openreview.net/forum?id=9BrydUVcoe Qu IP \ \#\ : Even better LLM quantization with hadamard incoherence and lattice codebooks . In Forty-first International Conference on Machine Learning

    2024

  9. [17]

    Jingyang Xiang and Sai Qian Zhang. 2024. Dfrot: Achieving outlier-free and massive activation-free for rotated llms with refined rotation. arXiv preprint arXiv:2412.00648

    2024 arXiv

  10. [18]

    Jingyang Xiang, Ying Zhang, Chi Ma, Yujie Wang, yulei, LiuChuan, Wei Lin, and Yong Liu. 2025. https://openreview.net/forum?id=oHBS7R6JcP Duarot: Dual rotation for advanced outlier mitigation in rotated LLM s

    2025

  11. [19]

    Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. https://doi.org/10.18653/v1/P19-1472 H ella S wag: Can a machine really finish your sentence? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4791--4...

    2019 doi

  12. [20]

    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...

  13. [21]

    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 gl...

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.