acceptodds
Under review as a conference paper at ICLR 2027

Understanding LLM Quantization through Activation-Guided Compensation and Orthogonal Residuals

Abstract

Post-training weight-activation quantization reduces the memory and inference costs of large language models, but aggressive W4A4 quantization remains difficult because activation outliers degrade effective quantization resolution. Although weight optimization, channel-wise scaling, and orthogonal rotation mitigate this problem, the error components they address and their relationship remain unclear. Using an exact decomposition of local weight-activation quantization error into an activation-guided weight compensation term and an orthogonal residual, we bound the residual using persistent channel-wise outlier and regular activation quantities. This decomposition clarifies which error components can be addressed by weight compensation and which require transformation design. We then use the residual bounds to derive practical guidelines for randomized Hadamard rotation, sign-based probability amplification, and channel scaling. In particular, the analysis explains how random signs suppress constructive interference among persistent outlier channels and how probability amplification through independent sign randomization tightens the high-probability bound on the minimum outlier term across sign configurations. It also shows how second-moment balancing leads to an scaling rule while a further relaxation recovers SmoothQuant-style scaling. We evaluate these guidelines through backpropagation-free configurations across eight Llama and Mistral models, obtaining performance competitive with gradient-trained SpinQuant.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, post-training quantization, weight-activation quantization, quantization error analysis

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_nEy4OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper develops a theoretical framework for understanding the interaction between weight compensation and output-preserving transformations in low-bit weight-activation quantization of LLMs. The key idea is to exactly decompose the local reconstruction error into an activation-guided weight compensation (AGWC) term and an orthogonal residual. For a fixed transformation, the former can be reduced through weight optimization, while the latter cannot be eliminated by modifying the quantized weights and therefore motivates transformation design.

Critical strengths
  1. It provises a useful conceptual decomposition of weight-activation quantization error.
  2. The motivation is clear and well presented.
  3. The empirical comparison is reasonably controlled.
Critical weaknesses
  1. The theoretical analysis assumes no-clipping, dynamic per-token symmetric uniform activation quantization. In contrast, the experiments use activation clipping with a ratio of 0.9 and K/V clipping of 0.95, while K/V quantization is asymmetric. This does not invalidate the analysis, but it weakens the claim that the derived bounds directly explain the actual experimental pipeline. It would substantially strengthen the paper to either extend the argument to clipped quantizers or empirically demonstrate that the theoretically motivated quantities remain predictive under the practical setting.
  2. The paper demonstrates that the final recipe improves perplexity, but this is not the same as demonstrating that the proposed theoretical mechanism explains the improvement, such as whether J_co actually decreases substantially after random sign modulation; whether the predicted O(N2_co) vs. approximately O(N_co) behavior can be observed across layers/models; whether J_co or the derived bound correlates with layer reconstruction error or final perplexity; and so on.
  3. Theorem 2 bounds the minimum J_co among independently sampled sign configurations. However, in practice the configurations are selected using validation perplexity rather than J_co. Moreover, one configuration jointly determines multiple rotation types, sign vectors are shared across layers within each category, and the final model-level criterion aggregates the behavior of many layers. Hence, the theorem does not directly imply that the implemented model-level SPA procedure will attain the predicted improvement. The empirical results suggest that candidate search is useful, but the theoretical connection would be more convincing if the paper showed that low validation perplexity candidates also correspond to lower predicted residual quantities.
  4. The experimental scope is somewhat limited for an ICLR 2027 LLM quantization paper. Although eight models are tested, they are limited to Llama variants and one Mistral model, with model sizes ranging only from 1B to 13B. Moreover, the paper should include comparisons with more advanced methods.
Questions for authors
  1. How tight are the proposed residual bounds in actual LLM layers?
  2. Can the authors directly validate the constructive-interference hypothesis?
7 comments
AuthorsOpenReview ↗

Response to Reviewer nEy4 — 1/5

Thank you for your thorough assessment of our manuscript and for identifying specific ways to strengthen it. Your comments on the theoretical assumptions, empirical validation, SPA candidate selection, and experimental coverage have guided the revisions presented in our responses below.

We have revised the manuscript in response to the reviewers’ insightful comments and questions. The changes are highlighted in the updated PDF.

W1

The theoretical analysis assumes no-clipping, dynamic per-token symmetric uniform activation quantization. In contrast, the experiments use activation clipping with a ratio of 0.9 and K/V clipping of 0.95, while K/V quantization is asymmetric. This does not invalidate the analysis, but it weakens the claim that the derived bounds directly explain the actual experimental pipeline. It would substantially strengthen the paper to either extend the argument to clipped quantizers or empirically demonstrate that the theoretically motivated quantities remain predictive under the practical setting.

A-W1

We can extend our theoretical analysis to clipped symmetric quantization and asymmetric quantization. We have added these extensions and their proofs as Lemmas 1 and 2 in Appendix B.2 of the revised manuscript.

These extensions replace the activation quantization error bound in Eq. (34) of the proof of Theorem 1 (Appendix B.1). Specifically, with qA=2bA−1−1q_A=2^{b_A-1}-1, qA′=2bA−1q_A'=2^{b_A}-1, and clipping ratio ρ∈(0,1]\rho\in(0,1], the three settings give

\leq d\sum_{t=1}^{T}\\|(\mathbf Z)_{t,:}\\|_\infty^2 \times \begin{cases} \dfrac{1}{4q_A^2}, & \text{no-clipping symmetric},\\\\[6pt] \max\left\lbrace\dfrac{\rho}{2q_A},\\,1-\rho\right\rbrace^{2}, & \text{clipped symmetric (Lemma 1)},\\\\[6pt] \dfrac{1}{(q_A')^2}, & \text{asymmetric (Lemma 2)}. \end{cases}$$ The clipped asymmetric case can also be handled by combining the arguments of Lemmas 1 and 2. Substituting the corresponding bound for Eq. (34), the remainder of the proof proceeds unchanged. The residual bound therefore retains the same dependence on $\mathcal J_{\mathrm{co}}$ and $\mathcal J_{\mathrm{reg}}$, with only the quantizer-dependent coefficient modified. The subsequent analyses of these quantities remain applicable. ### Manuscript revision In response to your comment, we added the extensions to clipped symmetric and asymmetric quantization, together with their proofs, as Lemmas 1 and 2 in Appendix B.2 (pp. 17-18) of the revised manuscript.
AuthorsOpenReview ↗

Response to Reviewer nEy4 — 2/5

W4

The experimental scope is somewhat limited for an ICLR 2027 LLM quantization paper. Although eight models are tested, they are limited to Llama variants and one Mistral model, with model sizes ranging only from 1B to 13B. Moreover, the paper should include comparisons with more advanced methods.

A-W4

Regarding model coverage, additional experiments on Llama-30B, Qwen3-8B [1], and Qwen3-32B support the effectiveness of L2-SmoothRot on the Qwen3 family and at the 30B scale. Regarding comparisons with advanced methods, we compared L2-SmoothRot with OSTQuant [2] and FlatQuant [3], both of which learn their transforms via backpropagation, on Llama3-8B, and L2-SmoothRot outperformed OSTQuant but trailed FlatQuant in PPL and mean zero-shot accuracy, with calibration time and peak GPU memory between the two.

Expanded model coverage. We added W4A4KV4 results for Llama-30B, Qwen3-8B, and Qwen3-32B, reporting WikiText-2 PPL and six-task mean zero-shot accuracy (%). We updated Tables 1 and 3 in the revised manuscript. L2-SmoothRot achieves the lowest perplexity on both 30B-class models and the highest mean accuracy on both Qwen3 models. SpinQuant attains lower perplexity on Qwen3-8B. These results support the effectiveness of L2-SmoothRot on the Qwen3 family and at the 30B scale.

ModelQuaRotSpinQuantL2-SmoothRot
PPL ↓ / Acc. ↑PPL ↓ / Acc. ↑PPL ↓ / Acc. ↑
Llama-30B4.51 / 73.434.57 / 72.854.49 / 73.11
Qwen3-8B11.09 / 62.9010.57 / 63.6610.91 / 64.15
Qwen3-32B8.57 / 70.238.31 / 70.928.26 / 72.08

Comparison with advanced methods. On Llama3-8B under W4A4KV4, we evaluated OSTQuant and FlatQuant alongside L2-SmoothRot. L2-SmoothRot achieved lower WikiText-2 perplexity and higher mean zero-shot accuracy than OSTQuant. FlatQuant achieved the best results on both metrics. L2-SmoothRot fell between OSTQuant and FlatQuant in both transform calibration time and peak GPU memory. Backpropagation-free L2-SmoothRot achieved competitive results with these methods and required less calibration time than FlatQuant.

MethodBackpropPPL ↓Acc. ↑Calib. time (min)Calib. peak mem. (GiB)
OSTQuantYes7.3267.7424.635.49
FlatQuantYes7.0169.04150.97.11
L2-SmoothRotNo7.2068.1056.18.21

Manuscript revision

Based on these additional experiments, we updated Table 1 (p. 8) in the main text and Table 3 (p. 22) in Appendix D to include results for Llama-30B, Qwen3-8B, and Qwen3-32B.

[1] Yang et al., Qwen3 Technical Report.

[2] Hu et al., OSTQuant: Refining Large Language Model Quantization with Orthogonal and Scaling Transformations for Better Distribution Fitting.

[3] Sun et al., FlatQuant: Flatness Matters for LLM Quantization.

AuthorsOpenReview ↗

Response to Reviewer nEy4 — 3/5

Q1

How tight are the proposed residual bounds in actual LLM layers?

A-Q1

We measured the orthogonal residual and its Theorem 1 bound in selected FFN down-projection layers of Llama3-8B and found that the bounds are conservative. We used L2L_2 scaling and 4-bit RTN on 128 WikiText-2 validation samples.

To understand why, we evaluated the intermediate quantities in the proof. The spectral-norm inequality ∣AV⊤∣F2≤∣A∣F2∣WΛ∣22\\|AV^\top\\|_F^2 \leq \\|A\\|_F^2\\|W\Lambda\\|_2^2 produced the largest increase: approximately 55× in layer 16, compared with 1–3× for each remaining relaxation. This step uses the maximum weight amplification, which can substantially exceed the amplification actually experienced by activation errors.

The table reports per-token values averaged over 10 random-sign Hadamard configurations.

LayerResidualBound
04.14×10−34.14\times10^{-3}1.931.93
80.1660.16651.851.8
160.3860.38682.282.2
240.7070.70789.189.1
318.018.013.68×1033.68\times10^{3}

Q2

Can the authors directly validate the constructive-interference hypothesis?

A-Q2

Yes. Constructive interference occurs when same-sign contributions from multiple outlier channels add up at one Hadamard output coordinate, creating an activation peak. In the synthetic examples, fixed Hadamard produced the larger peak in the worst case and a single random sign draw did so in the ordinary case, with best-of-10 (corresponding to our SPA) suppressing both peaks. In Llama3-8B's FFN down-projection inputs, fixed Hadamard produced the larger peak in the layer-29 example and a single random sign draw did so in the layer-28 example, with best-of-10 again suppressing both peaks.

We added Figs. 3, 4, 6, and 7 to Appendix F of the revised manuscript to visualize constructive interference and peak suppression in these examples.

The table below reports the peak CO activation value, defined as the maximum across channels of the mean absolute activation over tokens.

InputBefore rotationFixed HadamardRandom signsBest of 10
Synthetic worst case1.00000.53450.28400.2506
Synthetic ordinary case1.00000.26730.33410.2506
Llama3-8B, layer 280.61300.07310.08100.0690
Llama3-8B, layer 291.81550.08800.07440.0778

W2

The paper demonstrates that the final recipe improves perplexity, but this is not the same as demonstrating that the proposed theoretical mechanism explains the improvement, such as whether J_co actually decreases substantially after random sign modulation; whether the predicted O(N2_co) vs. approximately O(N_co) behavior can be observed across layers/models; whether J_co or the derived bound correlates with layer reconstruction error or final perplexity; and so on.

A-W2 (1/2)

We conducted additional experiments and found that random sign modulation actually reduced JcoJ_{\mathrm{co}} relative to fixed Hadamard rotation. Regarding the dependence on NcoN_{\mathrm{co}}, we validated the quadratic growth of JcoJ_{\mathrm{co}} under fixed rotation and its approximately linear growth under random signs on synthetic inputs. Regarding the relationship of JcoJ_{\mathrm{co}} and JregJ_{\mathrm{reg}} to PPL, our rationale is that reducing these quantities can suppress activation peaks and reduce local quantization error, potentially improving PPL.

JcoJ_{\mathrm{co}} decrement. We measured JcoJ_{\mathrm{co}} on two synthetic inputs with 64 CO channels and at Llama3-8B's FFN down-projection inputs on 128 WikiText-2 samples. Fixed Hadamard uses no sign flips, random signs uses one random sign configuration, and best-of-10 mirrors SPA's search over 10 configurations using JcoJ_{\mathrm{co}} for selection in this analysis. The table reports JcoJ_{\mathrm{co}} relative to fixed Hadamard rotation. Random signs reduce JcoJ_{\mathrm{co}} in two examples, and best-of-10 reduces it relative to fixed rotation in all four, including the two where the random draw increases it.

The table below reports JcoJ_{\mathrm{co}} relative to fixed Hadamard rotation (normalized to 1).

InputFixed HadamardRandom signsBest of 10
Synthetic worst case1.0000.2820.220
Synthetic ordinary case1.0001.5630.879
Llama3-8B, layer 281.0001.0430.945
Llama3-8B, layer 291.0000.8180.856
Reviewer_4dujOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes L2-SmoothRot, a rotation-based method for 4-bit weight and activation quantization that combines signed online rotation, sign-based probability amplification, and L2 scaling. The authors derive theoretical bounds related to channel-wise activation outliers, providing a theoretical foundation for the proposed design choices. Experiments on Llama and Mistral models under the W4A4 setting demonstrate lower perplexity and competitive downstream accuracy compared with the evaluated baselines.

Critical strengths

The theoretical analysis provides a clear motivation for the proposed method. In particular, the bounds help explain the role of random signs in controlling outlier-related quantities and motivate searching over multiple sign configurations. This provides a useful theoretical perspective on the proposed design choices.

Critical weaknesses
  1. The connection between the theoretical bounds and actual improvements needs stronger empirical support. Theorem 2 establishes upper bounds on outlier-related quantities rather than directly demonstrating reductions in their observed values. A tighter upper bound alone does not establish that actual quantization error or downstream performance will improve. Numerical measurements comparing the bounds with the corresponding empirical quantities would help assess their tightness and clarify how the theoretical motivation relates to the reported performance gains.
  2. The description of backpropagation requires clarification. The abstract and introduction describe the proposed method as backpropagation-free, and Section 4.6 states that sign configurations are evaluated using forward passes only. However, the experimental section also describes training rotations on WikiText-2 for 100 steps following SpinQuant’s protocol. It is unclear from this paragraph whether this training procedure applies only to the SpinQuant baseline or also to the proposed method. The authors should explicitly distinguish the procedures used for each method and clarify whether any stage of L2-SmoothRot involves backpropagation.
  3. The practical advantages of the method need further analysis. Beyond the reported quantization performance, additional analysis would help establish the method’s practical benefits. In particular, the authors should report total calibration time, including candidate screening and reevaluation, since the proposed procedure evaluates multiple sign configurations. Direct measurements of activation outliers before and after transformation would also help demonstrate whether the method suppresses outliers more effectively than the baselines.
  4. Comparisons with FlatQuant and OSTQuant are needed. The empirical evaluation would be stronger with comparisons against FlatQuant and OSTQuant. In particular, the apparent connection between the proposed scaling approach and learnable diagonal transformations motivates a closer comparison with OSTQuant. These comparisons would help clarify the method’s relative advantages in quantization quality and calibration cost, as well as its methodological differences from these approaches.
  5. The component ablation should include zero-shot performance. Table 2 evaluates the individual components using perplexity only. Reporting zero-shot accuracy for the same configurations would help determine whether the improvements from signed online rotation, sign-based probability amplification, and L2 scaling also translate into better downstream performance.
Questions for authors
  1. How tight are the theoretical bounds in Theorem 2 in practice? Could the authors report the corresponding empirical quantities and measurements of channel-wise activation outliers to demonstrate the effectiveness of the proposed transformations?
  2. Which stages of the proposed method, if any, use backpropagation? Does the 100-step rotation-training procedure described in the experimental section apply exclusively to the SpinQuant baseline?
  3. What is the total calibration time, including candidate screening and reevaluation, and how does it compare with the baselines?
  4. How does the proposed method compare with FlatQuant and OSTQuant in quantization performance and calibration cost? How does its scaling approach differ from that of OSTQuant?
  5. How does zero-shot accuracy change across the component configurations evaluated in Table 2?
11 comments
AuthorsOpenReview ↗

Response to Reviewer 4duj — 1/5

Thank you for your careful review and constructive feedback on our work. Your comments have helped us clarify the presentation, strengthen the empirical support for our theoretical analysis, and expand the evaluation and comparisons with existing methods.

We have revised the manuscript in response to the reviewers’ insightful comments and questions. The changes are highlighted in the updated PDF.

Q2, W2

Which stages of the proposed method, if any, use backpropagation?

A2-1

L2-SmoothRot uses no backpropagation at any stage. Channel scaling, rotation candidate sampling and selection, and GPTAQ weight calibration are all gradient-free. Our backpropagation-free method achieves performance competitive with gradient-trained SpinQuant.

Does the 100-step rotation-training procedure described in the experimental section apply exclusively to the SpinQuant baseline?

A2-2

Yes. The 100-step rotation-training procedure on the WikiText-2 training split applies only to SpinQuant. Like L2-SmoothRot, QuaRot and SmoothRot use no backpropagation.

Manuscript revision

In response to your comment, we revised the description of rotation training in Section 5 (Quantization and calibration; lines 446–448 of the submitted manuscript) as follows:

Before

Following the rotation-training protocol of SpinQuant (Liu et al., 2025), we train rotations on WikiText-2 train for 100 steps with frozen 16-bit weights and 4-bit activations and K/V, then quantize the rotated weights with GPTAQ.

After

For SpinQuant (Liu et al., 2025), we train rotations via backpropagation on the WikiText-2 training split for 100 steps with frozen 16-bit weights and 4-bit activations and K/V, then quantize the rotated weights with GPTAQ. This rotation-training procedure applies exclusively to the SpinQuant baseline. L2-SmoothRot and the other baselines use no backpropagation at any stage.

AuthorsOpenReview ↗

Response to Reviewer 4duj — 2/5

Q1, W1

How tight are the theoretical bounds in Theorem 2 in practice?

A1-1

On synthetic inputs, Theorem 2's bounds were 1.51.5–44 times the measured JcoJ_{\mathrm{co}} in the ordinary case, while fixed rotation attained its bound in the constructed worst case.

Since Theorem 2 assumes a persistent sparse channel-wise outlier (CO) model in which each CO channel has a constant activation value across tokens, whereas real CO activations can vary across tokens, we measured JcoJ_{\mathrm{co}} and evaluated the bounds on synthetic inputs satisfying this model. We used two inputs with 64 CO channels. Best-of-10 (corresponding to SPA's search) takes the smallest measured JcoJ_{\mathrm{co}} among 10 random sign configurations for each input.

For the ordinary input, the bounds for fixed rotation, random signs, and best-of-10 are 4×4\times, 1.58×1.58\times, and 1.50×1.50\times their respective measured values. In the constructed worst case, the CO contributions align under fixed Hadamard rotation and attain its bound. Random signs disrupt this alignment, reducing JcoJ_{\mathrm{co}} from 146.29 to 37.63 on average, and best-of-10 reduces it further to 32.14.

RotationOrdinary-case JcoJ_{\mathrm{co}}Worst-case JcoJ_{\mathrm{co}}Theorem 2 bound
Fixed Hadamard36.57146.29146.29
Random signs38.2737.6360.61
Best of 1032.1432.1448.29

Could the authors report the corresponding empirical quantities and measurements of channel-wise activation outliers to demonstrate the effectiveness of the proposed transformations?

A1-2

Yes, our experiments showed that best-of-10 can lower the estimated JcoJ_{\mathrm{co}} relative to fixed rotation on Llama3-8B activations. The synthetic and real activation profiles also showed that best-of-10 can suppress CO interference peaks.

First, we estimated JcoJ_{\mathrm{co}} for COs detected at the inputs to Llama3-8B's FFN down-projection layers using WikiText-2. We report empirical JcoJ_{\mathrm{co}} relative to fixed rotation in the table below. In the layer-29 example, adding random signs reduces JcoJ_{\mathrm{co}} under both the illustrated draw and the best-of-10 configuration. In the layer-28 example, the illustrated random draw increases JcoJ_{\mathrm{co}}, and selecting a sign configuration via best-of-10 brings it below the fixed-rotation value.

We next show how each rotation redistributes CO activations using synthetic inputs and real activations. The peak is the largest mean absolute CO activation across channels, with each channel averaged over tokens. The peaks follow the same trends as JcoJ_{\mathrm{co}}. Fixed rotation gives the largest peak in the synthetic worst case and the layer-29 example, while random signs give the largest peak in the ordinary case and the layer-28 example. Best-of-10 reduces these peaks in all four examples.

InputMetricBefore rotationFixed HadamardRandom signsBest of 10
Synthetic worst caseJcoJ_{\mathrm{co}} ratio—1.0000.2820.220
Peak1.00000.53450.28400.2506
Synthetic ordinary caseJcoJ_{\mathrm{co}} ratio—1.0001.5630.879
Peak1.00000.26730.33410.2506
Llama3-8B, layer 28JcoJ_{\mathrm{co}} ratio—1.0001.0430.945
Peak0.61300.07310.08100.0690
Llama3-8B, layer 29JcoJ_{\mathrm{co}} ratio—1.0000.8180.856
Peak1.81550.08800.07440.0778

Manuscript revision

Based on these analyses, we added Appendix F (pp. 24-27) to the manuscript. It presents the experimental settings, comparisons of the synthetic JcoJ_{\mathrm{co}} measurements with Theorem 2's upper bounds, and measurements and visualizations of CO activations on both synthetic inputs and Llama3-8B.

AuthorsOpenReview ↗

Response to Reviewer 4duj — 3/5

Q5, W5

How does zero-shot accuracy change across the component configurations evaluated in Table 2?

A5

Adding SOR, SPA, or L2S individually to QuaRot improves mean zero-shot accuracy, as with WikiText-2 PPL. The accuracy gains are 1.75, 2.11, and 2.12 percentage points, respectively. Combining all three components achieves the highest accuracy (68.10%).

QuaRotSORSPAL2SPPL ↓Acc. (%) ↑
✓7.4565.71
✓✓7.3067.46
✓✓7.4067.82
✓✓7.3267.83
✓✓✓✓7.2068.10

Manuscript revision

We have added the six-task mean zero-shot accuracy for Llama3-8B under W4A4KV4 to Table 2 (p. 9).

Reviewer_HB6VOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper provides a unified theoretical framework for several existing techniques in weight-activation quantization and derives a practical method based on this framework. Specifically, the quantization error can be decomposed into two terms. The first term can be compensated for by weight optimization, while the second term is an orthogonal residual. Further analysis of the orthogonal residual motivates signed Hadamard rotations, sign-based probability amplification, and L2 channel scaling. Experiments on Llama and Mistral models show competitive performance compared with existing methods.

Critical strengths
  1. The theoretical part is extensive and reasonable, and is closely related to the proposed method.
  2. The paper is overall well organized. The framing of the theoretical part is clear.
  3. The experimental results support the proposed method, which is competitive with existing quantization methods.
Critical weaknesses
  1. There is a large gap between the theoretical analysis and the conducted experiments.

    • The core of the theory is the orthogonal residual, JcoJ_{\mathrm{co}}, and JregJ_{\mathrm{reg}}, but the paper does not show whether these quantities actually decrease with the proposed method. The authors also do not verify the relationship between these quantities and PPL.
    • SPA uses validation PPL to select candidates, which is different from the theoretical motivation that multiple samples are used to find a smaller JcoJ_{\mathrm{co}}. The performance improvement may therefore come simply from best-of-NN search based on PPL. Although the paper studies different candidate budgets and screening strategies, there is no ablation that distinguishes the proposed JcoJ_{\mathrm{co}}-based mechanism from a generic best-of-NN search effect.
    • The derived worst-case bound for fixed Hadamard rotation scales as O(Nco2)O(N_{\mathrm{co}}^2), while random signs improve the dependence to approximately linear in NcoN_{\mathrm{co}}, up to logarithmic factors. However, this behavior is only established theoretically and is not directly verified in the experiments.
  2. There are no experiments related to efficiency, either during calibration or at runtime. Since online operations are incorporated into the inference process, end-to-end runtime measurements are necessary. In addition, the calibration process includes 10 RTN evaluations and 3 GPTAQ evaluations. This introduces non-trivial calibration/search cost, but the paper does not report calibration time or compare this cost with the baselines.

  3. The theoretical analysis occupies a large portion of the main paper, leaving relatively limited space for a more comprehensive experimental evaluation. Since quantization is targeted at real-world use and deployment, experiments should be a critical part of the paper.

Questions for authors

Please see weakness 1-2.

11 comments
AuthorsOpenReview ↗

Response to Reviewer HB6V — 1/4

Thank you for the time and effort you devoted to reviewing our manuscript and for your detailed feedback. We appreciate your comments on the connection between theory and experiments, calibration and inference efficiency, and the balance between theoretical analysis and empirical evaluation.

We have revised the manuscript in response to the reviewers’ insightful comments and questions. The changes are highlighted in the updated PDF.

W3

The theoretical analysis occupies a large portion of the main paper, leaving relatively limited space for a more comprehensive experimental evaluation. Since quantization is targeted at real-world use and deployment, experiments should be a critical part of the paper.

A3

We agree that the field’s emphasis on practical performance is well placed and that experiments are essential for assessing it. At the same time, we believe that understanding why successful quantization methods work is important for advancing the field. Theoretical analysis can deepen that understanding and guide the design of future methods.

As our title suggests, this paper aims to develop a theoretical understanding of quantization techniques that have proved effective in practice. While we also present L2-SmoothRot, our primary contribution is a unified theoretical framework that explains how rotation and channel scaling, two widely used techniques for smoothing activations, complement each other in reducing residual error after weight compensation. We believe that developing this framework and explaining these complementary roles warrants the space devoted to the analysis in the main paper.

AuthorsOpenReview ↗

Response to Reviewer HB6V — 2/4

W1

The core of the theory is the orthogonal residual JcoJ_{\mathrm{co}}, and JregJ_{\mathrm{reg}} , but the paper does not show whether these quantities actually decrease with the proposed method. The authors also do not verify the relationship between these quantities and PPL.

A1-1

For JcoJ_{\mathrm{co}}, best-of-10 (corresponding to our SPA search) yielded a lower layer mean than fixed Hadamard in 26 of 30 Llama3-8B layers with COs. For JregJ_{\mathrm{reg}}, L2L_2 scaling yielded lower values than L∞L_\infty scaling in 12 of 32 layers, including the final eight layers, where JregJ_{\mathrm{reg}} is large under L∞L_\infty scaling. Regarding their relationship to PPL, our rationale is that reducing JcoJ_{\mathrm{co}} and JregJ_{\mathrm{reg}} can suppress activation peaks that contribute to local quantization error, a practical proxy for model performance used by existing methods.

JcoJ_{\mathrm{co}} decrement. We measured JcoJ_{\mathrm{co}} on two synthetic inputs with 64 CO channels and at Llama3-8B's FFN down-projection inputs on 128 WikiText-2 samples. Fixed Hadamard uses no sign flips, random signs uses one random sign configuration, and best-of-10 mirrors SPA's search over 10 configurations using JcoJ_{\mathrm{co}} for selection in this analysis. The table reports JcoJ_{\mathrm{co}} relative to fixed Hadamard rotation. Random signs reduce JcoJ_{\mathrm{co}} in two examples, and best-of-10 reduces it relative to fixed rotation in all four, including the two where the random draw increases it.

The table below reports JcoJ_{\mathrm{co}} relative to fixed Hadamard rotation (normalized to 1).

InputFixed HadamardRandom signsBest of 10
Synthetic worst case1.0000.2820.220
Synthetic ordinary case1.0001.5630.879
Llama3-8B, layer 281.0001.0430.945
Llama3-8B, layer 291.0000.8180.856

JregJ_{\mathrm{reg}} decrement. We measured JregJ_{\mathrm{reg}} under L2L_2 and SmoothQuant-style L∞L_\infty scaling [1] on the same 128 WikiText-2 samples and 10 random sign configurations. The effect varies across layers. In the final eight layers, where the L∞L_\infty baseline is largest, L2L_2 scaling reduces measured JregJ_{\mathrm{reg}} by about 10% to 55% in every layer.

LayersLayers with lower JregJ_{\mathrm{reg}} under L2L_2Range of L2/L∞L_2/L_\infty ratios
0–234 / 240.856–4.434
24–318 / 80.452–0.896
All12 / 320.452–4.434

Relationship to PPL. Although we have not directly tested how JcoJ_{\mathrm{co}} and JregJ_{\mathrm{reg}} relate to PPL, our rationale is as follows. The term ∑t=1T∣(XcoΛ−1DH⊤)t,:∣∞2\sum_{t=1}^{T}\\|(X_{\mathrm{co}}\Lambda^{-1}DH^\top)_{t,:}\\|_\infty^2 in JcoJ_{\mathrm{co}} captures per-token peaks of the CO component after rotation, whereas ∑t=1T∣(XregΛ−1DH⊤)t,:∣∞2\sum_{t=1}^{T}\\|(X_{\mathrm{reg}}\Lambda^{-1}DH^\top)_{t,:}\\|_\infty^2 in JregJ_{\mathrm{reg}} captures those of the regular component. Minimizing JcoJ_{\mathrm{co}} and JregJ_{\mathrm{reg}} encourages smaller values for these peak terms, which can help suppress peaks in the activations. Reducing these peaks can permit finer quantization steps at the same bit width when they dominate the quantization range [1]. By minimizing local reconstruction error during calibration, GPTQ [2], OmniQuant [3], and FlatQuant [4] use it as a practical proxy for model performance. We follow the same rationale in our analysis.

[1] Xiao et al., SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models.

[2] Frantar et al., GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers.

[3] Shao et al., OmniQuant: Omnidirectionally Calibrated Quantization for Large Language Models.

[4] Sun et al., FlatQuant: Flatness Matters for LLM Quantization.

AuthorsOpenReview ↗

Response to Reviewer HB6V — 3/4

W1 (continued)

The derived worst-case bound for fixed Hadamard rotation scales as O(Nco2)O(N_{\mathrm{co}}^2), while random signs improve the dependence to approximately linear in NcoN_{\mathrm{co}} , up to logarithmic factors. However, this behavior is only established theoretically and is not directly verified in the experiments.

A1-2

In response to your comment, we measured JcoJ_{\mathrm{co}} on synthetic inputs. In the constructed worst case, fixed Hadamard rotation attained its quadratic bound, while JcoJ_{\mathrm{co}} under random signs grew approximately linearly at larger CO counts.

We varied the number of CO channels (NcoN_{\mathrm{co}}) in a synthetic input satisfying the persistent sparse CO model. For unit CO magnitudes and no channel scaling, Theorem 2 gives the following upper bounds on JcoJ_{\mathrm{co}}.

J_{\mathrm{co}}^{\mathrm{fix}} &\le B_{\mathrm{fix}} = \frac{T N_{\mathrm{co}}^2}{d} & &\text{(fixed Hadamard)}, \\\\ J_{\mathrm{co}}^{\mathrm{rand}} &\le B_{\mathrm{rand}} = \frac{2T N_{\mathrm{co}}}{d}\bigl(\log(2d)+\log(1/\delta)\bigr) & &\text{(random signs)}, \\\\ J_{\mathrm{co}}^{\mathrm{best}} &\le B_{\mathrm{best}} = \frac{2T N_{\mathrm{co}}}{d}\bigl(\log(2d)+\tfrac{1}{10}\log(1/\delta)\bigr) & &\text{(best)}. \end{aligned}$$ Here, $T=512$ is the sequence length, $d=14{,}336$ is the input dimension, and $\delta=0.05$. The last two inequalities each hold with probability at least $1-\delta=0.95$. The single-draw bound $B_{\mathrm{rand}}$ becomes smaller than $B_{\mathrm{fix}}$ at $N_{\mathrm{co}}\geq 27$. Since Theorem 2 assumes a persistent sparse CO model with constant activations across tokens within each CO channel, we did not compute its bounds for Llama3-8B activations, which vary across tokens. However, in Appendix F we estimate $J_{\mathrm{co}}$ layer by layer by retaining only the activations in detected CO channels. The table reports $J_{\mathrm{co}}^{\mathrm{fix}}$ for fixed Hadamard rotation, $J_{\mathrm{co}}^{\mathrm{rand}}$ for random signs, and $J_{\mathrm{co}}^{\mathrm{best}}$ for the minimum over 10 random sign configurations, alongside the theoretical bounds. The results show that $J_{\mathrm{co}}^{\mathrm{fix}}$ grows quadratically with $N_{\mathrm{co}}$, matching its upper bound, whereas $J_{\mathrm{co}}^{\mathrm{rand}}$ and $J_{\mathrm{co}}^{\mathrm{best}}$ grow approximately linearly at larger CO counts. **Worst case.** | $N_{\mathrm{co}}$ | $J_{\mathrm{co}}^{\mathrm{fix}}$ | $B_{\mathrm{fix}}$ | $J_{\mathrm{co}}^{\mathrm{rand}}$ | $B_{\mathrm{rand}}$ | $J_{\mathrm{co}}^{\mathrm{best}}$ | $B_{\mathrm{best}}$ | |--:|--:|--:|--:|--:|--:|--:| | 8 | 2.29 | 2.29 | 2.29 | 7.58 | 2.29 | 6.04 | | 16 | 9.14 | 9.14 | 7.64 | 15.15 | 7.00 | 12.07 | | 32 | 36.57 | 36.57 | 17.01 | 30.31 | 14.29 | 24.14 | | 64 | 146.29 | 146.29 | 37.63 | 60.61 | 32.14 | 48.29 | | 128 | 585.14 | 585.14 | 77.11 | 121.23 | 63.00 | 96.58 | > SPA uses validation PPL to select candidates, which is different from the theoretical motivation that multiple samples are used to find a smaller $J_{\mathrm{co}}$. The performance improvement may therefore come simply from best-of-$N$ search based on PPL. Although the paper studies different candidate budgets and screening strategies, there is no ablation that distinguishes the proposed $J_{\mathrm{co}}$-based mechanism from a generic best-of-$N$ search effect. ### A1-3 As you mentioned, SPA can be interpreted as a best-of-$N$ search over random sign configurations using validation PPL. Theorem 2 shows that more independent draws tighten the high-probability bound on the minimum $J_{\mathrm{co}}$ in the candidate set. Although we did not directly test whether selecting candidates with lower validation PPL also yields lower $J_{\mathrm{co}}$, our rationale is as follows. The activation profiles in Figures 3, 4, 6, and 7, added to Appendix F of the revised manuscript, show that smaller CO peaks accompany lower $J_{\mathrm{co}}$. Lower validation PPL may reflect smaller local quantization errors. As discussed in SmoothQuant [1], activation outliers can substantially increase quantization error by expanding the quantization range. Thus, smaller local errors may partly reflect more effective smoothing of CO peaks, which is associated with lower $J_{\mathrm{co}}$ in our measurements. ### Manuscript revision We added empirical measurements of $J_{\mathrm{co}}$ and $J_{\mathrm{reg}}$ in Appendices G (pp. 27-30) and H (pp. 31-32), respectively. [1] Xiao et al., SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models.

Comments on the paper

1 comment
Area_Chair_znHkOpenReview ↗

Engage in discussion

Reviewers: please review the authors’ rebuttal and consider revising your assessment/score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.