acceptodds
Under review as a conference paper at ICLR 2027

Beyond Exact Equivalence: Rethinking Quantization Preprocessing

Abstract

Most quantization preprocessing methods search for transformations that exactly preserve the pretrained model’s full-precision function. However, the model that matters is the one after quantization: why should optimization remain confined to equivalence before rounding? We challenge this restriction and formulate quantization preprocessing as optimization over a broader space of representations, where exact equivalence is only one possible constraint. Our framework accommodates flexible offline preprocessing, nonlinear parameter transformations, and direct weight adaptation while keeping deployment constraints explicit. We instantiate this perspective with an equivalence-preserving structured initialization followed by global quantization-aware distillation, allowing the full-precision parameters to deviate from the original model to better preserve its behavior after quantization. Across multiple model families, this relaxation consistently enhances post-quantization language-modeling quality compared to matched post-training quantization baselines, without requiring preprocessing-specific online transformations at deployment. These results establish a broader principle for low-bit compression: exact equivalence can provide a safe starting point, but should not define the space in which the quantized model is optimized.

Primary area:
foundation or frontier models, including LLMs
Keywords:
model quantization, quantization preprocessing, low-bit compression, quantization-aware distillation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_R5FHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper answers the question whether exact FP equivalence is necesary when learning rotation transformation matrix throughout the quantization process. The paper propose a QAT framework that generate a fixed equivalence-preserving Hadamard matrix at initialization, then use global quantization-aware distillation (QAD). The update are optimized directly into the merged weight instead of updating the rotation matrix to preserve the equivalence. Experiment for W4A4 setting show improvements across many architecture over rotation only optimization.

Critical strengths
  1. Consistent improvement: the proposed method improve performance across 4 architectire, using perplexity and downstream zero-shot accuracy.

  2. The motivation is intuitive and makes sense. The experiments about direct-weight optimization and rotation only are meaningful.

Critical weaknesses
  1. Limited novelty. The proposed algorithm basically combines fixed Hadamard rotation preprocessing with normal QAF framework using STE and KL loss. Moreover, while the motivation make sense, existing QAT methods already optimize full-precision weights without functional equivalence (e.g. LLM-QAT, RCP). The two proposition are fairly simple, so theoretical contribution isn't strong.

  2. The experiment results are not unexpected, because direct-weight optimization has more degree of freedom for optimization, while the rotation-only optimization has constraints.

  3. Lack of competisons with existing QAT baselines. The paper only compare with same baseline but using rotation-only optimization.

Questions for authors

Please answer the concerns in weakness section

3 comments
AuthorsOpenReview ↗

Response to Reviewer R5FH

We thank the reviewer for recognizing the improvements across four architectures and the value of the direct-weight versus rotation-only comparison. We address these three concerns in turn.

W1 / Q1(a): Novelty is limited, and existing QAT already updates full-precision weights without equivalence.

Answer: Latent-weight QAT and fused-weight adaptation predate this work. The narrower question here is whether, after the same equivalence-preserving initialization, retaining exact equivalence helps or restricts the finite-budget endpoint.

The submission does not claim that direct fused-weight adaptation is new or first. LLM-QAT/BitDistiller/EfficientQAT perform latent-weight QAT, and RCP fuses rotation into the weights while jointly optimizing them with learnable clipping and non-uniform quantization parameters. Our distinction is to compare, from the same equivalence-preserving starting representation, a parameterization that maintains exact equivalence against one that releases it. The propositions formalize why releasing the constraint enlarges the admissible search space and why strict improvement is possible; the experiments test whether that distinction matters under a finite optimization budget.

W2 / Q1(b): The direct-weight arm has more degrees of freedom, so the improvement is expected.

Answer: Releasing exact equivalence does add optimization degrees of freedom. That is precisely the design variable we test: do those degrees of freedom produce a reproducible finite-budget benefit from the same rotation-derived initialization, or does greater freedom merely appear beneficial by construction?

We do not claim parameter-count matching or universal superiority. We measure the finite-budget gain, reproducibility across seeds, and whether the drift exhibits error compensation rather than arbitrary movement.

ModelRelaxed interaction termSequence-paired 95% bootstrap CIConstrained mean PPL ± SDRelaxed mean PPL ± SDPaired relaxed − constrained PPLConstrained mean six-task ± SDRelaxed mean six-task ± SD
LLaMA-2-7B-0.4392[-0.4823, -0.4092]8.8543 ± 0.03798.2124 ± 0.5059-0.64190.5899 ± 0.00570.6044 ± 0.0030

The interaction interval is a sequence-paired 95% bootstrap interval over the same held-out C4 set. The seed study covers LLaMA-2-7B and TinyLlama-1.1B only; no LLaMA-3 seed-variance claim is made. It uses forward KL, 800/32 C4 train/validation sequences, 500 updates, batch 64, 32,000 rank-samples, W4A4/G128/nohad evaluation, initializer seed 0, three predeclared training/data-order seeds, and the final checkpoint for both arms.

ModelConstrained PPL ± SDRelaxed PPL ± SDPaired ΔPPLConstrained six-task ± SDRelaxed six-task ± SD
LLaMA-2-7B8.8543 ± 0.03798.2124 ± 0.5059-0.64190.5899 ± 0.00570.6044 ± 0.0030
TinyLlama-1.1B13.6363 ± 0.081311.7167 ± 0.0106-1.91960.4659 ± 0.00600.4924 ± 0.0006

The table reports rows with all three seeds completed. TinyLlama uses a similar budget and evaluator family but a different model and setup; it supports the observed direction, not model-scale generalization. The seeds vary training and data-order randomness only, so initializer variance is not estimated.

W3 / Q1(c): Existing QAT baselines are missing.

Answer: External QAT methods provide important context, but they cannot serve as protocol-matched controls for the specific intervention studied here. We therefore distinguish method positioning from empirical comparison and test the equivalence constraint using a rotation-constrained control with the same initializer.

RCP targets W2/3A4KV4 with online rotations and a learnable non-uniform quantizer, unlike our W4A4/G128/nohad setting with a fixed uniform quantizer and matched C4/update budget. LLM-QAT/EfficientQAT likewise differ in data, bit configuration, trainable objects, and compute. Because we do not have protocol-compatible local reproductions of these external methods, we do not present paper-reported numbers as same-setting baselines; any contextual comparison is source-separated with model/data/bits/quantizer/graph/compute differences explicit. We do not propose direct-weight QAT as a new algorithm, nor claim that additional degrees of freedom always improve performance. Our claim is the narrower finite-budget consequence of maintaining exact equivalence after the same rotation-derived initialization.

Dear Reviewer R5FH,

As the discussion period is approaching its end, we would like to kindly follow up and ask whether our rebuttal has addressed your concerns. We have provided additional experimental results and clarifications in response to your comments and would greatly appreciate your feedback.

Best regards, Authors

Thank you for the rebuttal. The new experiments are helpful, solve some of my concerns.

But my novelty concern remains. The gap seem a bit too narrow for me, and it is unclear whether the gains would persist against more flexible transformations, such as affine transformations that preserve the model’s full-precision outputs. While existing QAT methods are difficult to compare across settings, the authors could still test an existing QAT method with and without the proposed initialization under the same protocol. This would show whether the initialization also helps beyond the paper’s own QAD recipe.

I am keeping my original rating.

Reviewer_AQHfOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper revisits the exact-equivalence constraint in quantization preprocessing, arguing that allowing full-precision deviations may lead to better post-quantization fidelity. The authors formulate a relaxed optimization framework and provide theoretical analysis to justify this perspective. The proposed method combines fixed Hadamard rotation initialization with global quantization-aware distillation, allowing direct weight adaptation without enforcing exact equivalence. Experiments on four language models demonstrate improvements over rotation-only baselines.

Critical strengths
  1. Clear and intuitive insight. The paper challenges the necessity of exact equivalence in quantization preprocessing and provides a straightforward and meaningful perspective on optimizing quantized models.
  2. Well-developed theoretical analysis. The paper presents a formal optimization framework with clearly stated assumptions and detailed proofs. The theoretical arguments are generally sound and provide a solid foundation for the proposed perspective.
Critical weaknesses
  1. Insufficient discussion of related work. The literature review is relatively limited, particularly given the broad scope of the proposed framework. A more comprehensive analysis of existing quantization preprocessing and adaptation methods is needed to better establish the novelty and positioning of this work.
  2. Limited methodological novelty. The proposed method mainly combines fixed Hadamard rotation with standard quantization-aware distillation. Both components are well-established, and the overall design introduces limited new algorithmic elements.
  3. Incomplete empirical validation of the central insight. Although the experiments demonstrate the effectiveness of relaxed optimization, they do not directly examine the relationship between full-precision drift and quantization error, leaving the proposed error-compensation interpretation insufficiently supported.
  4. Limited incremental contribution. Although the paper provides a useful conceptual perspective on exact equivalence, its substantive contribution remains relatively modest. The broader formulation does not lead to a substantially new optimization strategy or deeper insights into when relaxing equivalence is beneficial.
Questions for authors
  1. Could the authors clarify how the proposed framework differs from existing quantization preprocessing and adaptation approaches, particularly in terms of its conceptual and methodological contributions?
  2. Is there empirical evidence that the full-precision drift introduced by quantization-aware distillation compensates for quantization error in practice?
2 comments
AuthorsOpenReview ↗

Response to Reviewer AQHf

We thank the reviewer for the positive assessment of the framing and formalization. We address the four weaknesses and two questions directly below.

W1 / Q1: Related Work is too narrow. How does the framework differ conceptually and methodologically from existing preprocessing and adaptation?

Answer: The neighboring methods are important, but they do not make our question redundant. We study whether exact equivalence should remain enforced after the same equivalence-preserving initialization, separating the role of the constraint from the choice of QAT primitive.

A useful taxonomy is: (i) QuaRot/SpinQuant optimize equivalence-preserving rotation families; (ii) LLM-QAT/BitDistiller/EfficientQAT establish latent-weight QAT; (iii) RCP performs fused-weight QAT after fixed rotation, using learnable clipping and non-uniform quantization; and (iv) ParoQuant/post-rotation adaptation occupy the preprocessing-adaptation boundary. Our contribution is not another QAT component. We instead ask whether exact equivalence should remain enforced during adaptation from the same initialization.

W2: Fixed Hadamard plus standard QAD offers limited methodological novelty.

Answer: We somewhat agree that the individual building blocks, including Hadamard initialization, STE-based QAD, and latent-weight adaptation, are established. However, that observation does not exhaust the contribution: our focus is the experimentally testable consequence of maintaining exact equivalence after the same equivalence-preserving initialization.

In particular, RCP fuses rotation into the weights and jointly optimizes them with learnable clipping and non-uniform quantization parameters. We do not position the work as a new general-purpose QAT algorithm; instead, we formalize a representation- and constraint-level question and provide controlled finite-budget evidence for its effect.

W3 / Q2: There is no direct evidence that full-precision drift compensates quantization error.

Answer: To provide W4 evidence for this question, we measured the interaction 2VEμ[⟨eFPc,eQc⟩]\frac{2}{V}\mathbb E_\mu[\langle e_{\mathrm{FP}}^c,e_Q^c\rangle] directly on 32 held-out C4 sequences, with eFPc=C(Fa−F0)e_{\mathrm{FP}}^c=C(F_a-F_0), eQc=C(F^a−Fa)e_Q^c=C(\widehat F_a-F_a), and C=IV−V−111⊤C=I_V-V^{-1}\mathbf 1\mathbf 1^\top (which removes the vocabulary-wide mean logit). The table reports the interaction, its sequence-paired 95% bootstrap interval, and rμc(a)=V−1Eμ∥C(F^a−F0)∥22r_\mu^c(a)=V^{-1}\mathbb{E}_\mu\|C(\widehat F_a-F_0)\|_2^2, the mean centered squared-logit error to the teacher.

ModelSettingQADInteraction termPaired 95% bootstrap CIrμc(a)r_\mu^c(a) (centered logit error) ↓
LLaMA-2-7BFixed initializerNo-0.0005[-0.0008, -0.0002]3.0462
LLaMA-2-7BEquivalence-constrainedYes0.0003[0.0002, 0.0004]1.5732
LLaMA-2-7BRelaxedYes-0.4392[-0.4823, -0.4092]1.1139
LLaMA-3-8BFixed initializerNo0.0009[0.0005, 0.0013]4.1059
LLaMA-3-8BEquivalence-constrainedYes-0.0007[-0.0015, 0.0004]2.3962
LLaMA-3-8BRelaxedYes-0.3140[-0.3363, -0.2950]1.6487

For both relaxed settings, the negative interaction and CI excluding zero are consistent with partial compensation. This diagnostic is limited to W4A4/G128/nohad on two models and should be interpreted as evidence consistent with the compensation hypothesis, not as a universal causal law.

W4: The broader formulation does not yet explain when relaxation is beneficial.

Answer: The evidence supports a conditional conclusion: exact equivalence is useful for initialization, but maintaining it throughout finite-budget optimization can exclude compensating directions. The W3 comparison extends this test to a lower-bit deployable setting, while the seed study measures how consistently the effect appears under the fixed W4A4 protocol. The measured W3 boundary is:

ModelFixed initializer (PPL / six-task mean)Equivalence-constrained FKL (PPL / six-task mean)Relaxed FKL (PPL / six-task mean)
LLaMA-2-7B7903.7326 / 0.294711.5227 / 0.52239.2364 / 0.5620
LLaMA-3-8B7422.5008 / 0.302622.6612 / 0.488116.1739 / 0.5447

These are finite-budget results under the tested W3A4/G128/nohad protocol; they do not establish a universal rule across models, bit widths, or quantizers.

Dear Reviewer AQHf,

As the discussion period is approaching its end, we would like to kindly follow up and ask whether our rebuttal has addressed your concerns. We have provided additional experimental results and clarifications in response to your comments and would greatly appreciate your feedback.

Best regards, Authors

Reviewer_3wdSOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper revisits the requirement that quantization preprocessing preserve the full-precision function. It formulates representation selection as optimizing post-quantization fidelity subject to deployment constraints, distinguishes full-precision deviation from quantized error, and instantiates the idea with an equivalence-preserving structured initialization followed by global quantization-aware distillation (QAD). Experiments cover LLaMA-2/3, Mistral, and Qwen2.5 with RTN W4A4, using 800 C4 sequences for training. Results show gains over rotation-initialization and PTQ controls in several settings, although the broader design space is not yet extensively validated.

Critical strengths
  1. Treating post-quantization behavior, rather than full-precision functional equivalence, as the ultimate objective is a clear and worthwhile framing.
  2. The paper distinguishes offline preprocessing, deployment-graph constraints, and subsequent adaptation, and evaluates the proposed instantiation on four model families.
Critical weaknesses
  1. The formal result that relaxing the feasible set cannot worsen the optimum follows directly from set inclusion. The existence example for strict improvement does not explain why the proposed QAD should work reliably for real Transformers. Stronger mechanism analysis or component-isolating ablations would help.
  2. The main experiments focus on RTN W4A4, roughly 800 C4 sequences per model, and a small set of models. The paper does not establish whether gains persist at lower bit widths, with different calibration distributions, against stronger PTQ baselines, or under other deployment constraints. Variance across random seeds is also missing.
  3. The actual method is primarily structured initialization followed by conventional global quantization-aware distillation. The paper does not establish sufficient algorithmic novelty relative to existing QAT/PTQ distillation approaches, nor isolate its contribution with equal-training-budget controls.
Questions for authors
  1. Although the appendix includes lower-bit diagnostics such as W3A4, could you add a deployable W3 comparison and a stronger PTQ baseline to the main evaluation to test whether the gains are specific to RTN W4A4?
  2. How much variance does QAD exhibit across initialization and training seeds? Which components account for the gains relative to simply extending PTQ or distillation training?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 3wdS

Thank you for the constructive comments. We address the questions below.

W1: Set inclusion and the existence example do not explain why QAD works in real Transformers.

Answer: We agree that feasible-set inclusion alone does not predict finite-budget optimizer success. We therefore measured the centered interaction 2V−1Eμ[⟨eFPc,eQc⟩]2V^{-1}\mathbb E_\mu[\langle e_{\mathrm{FP}}^c,e_Q^c\rangle] on 32 held-out C4 sequences, where eFPc=C(Fa−F0)e_{\mathrm{FP}}^c=C(F_a-F_0), eQc=C(F^a−Fa)e_Q^c=C(\widehat F_a-F_a), and C=IV−V−111⊤C=I_V-V^{-1}\mathbf 1\mathbf 1^\top removes the vocabulary-wide mean logit. The interaction and its paired 95% bootstrap CI were:

ModelSettingInteraction (95% paired CI)
LLaMA-2-7BFixed initializer-0.0005 [-0.0008, -0.0002]
LLaMA-2-7BEquivalence-constrained0.0003 [0.0002, 0.0004]
LLaMA-2-7BRelaxed-0.4392 [-0.4823, -0.4092]
LLaMA-3-8BFixed initializer0.0009 [0.0005, 0.0013]
LLaMA-3-8BEquivalence-constrained-0.0007 [-0.0015, 0.0004]
LLaMA-3-8BRelaxed-0.3140 [-0.3363, -0.2950]

Both relaxed settings show a negative interaction with a CI excluding zero, consistent with partial compensation. This is setting-specific evidence, not a causal or universal mechanism claim.

W2 / Q1: The evaluation is concentrated on RTN W4A4. Can you add deployable W3 and a stronger PTQ baseline?

Answer: We add a deployable RTN W3A4/G128/nohad comparison:

ModelMethodQADGraph / deploymentWikiText-2 PPL ↓Six-task mean ↑
LLaMA-2-7BFixed initializerNonohad / RTN7903.73260.2947
LLaMA-2-7BEquivalence-constrained FKLYesnohad / RTN11.52270.5223
LLaMA-2-7BRelaxed FKLYesnohad / RTN9.23640.5620
LLaMA-3-8BFixed initializerNonohad / RTN7422.50080.3026
LLaMA-3-8BEquivalence-constrained FKLYesnohad / RTN22.66120.4881
LLaMA-3-8BRelaxed FKLYesnohad / RTN16.17390.5447

All six rows are completed measurements under the same W3 protocol.

Separate W4A4 comparison. For SpinQuant, W16A4 is the rotated source representation before the final GPTQ-W4A4/G128 endpoint; relaxed FKL is optimized and evaluated directly at W4A4.

ModelMethodDeployment graphOnline HadamardPipeline before final endpointQuantizerWikiText-2 PPL ↓
LLaMA-2-7BSpinQuant (local reproduction)hadYesW16A4 SpinQuant rotationGPTQ-W4A4G1285.99
LLaMA-2-7BSpinQuant (local reproduction)nohadNoW16A4 SpinQuant rotationGPTQ-W4A4G12813.45
LLaMA-2-7BRelaxed FKLnohadNoW4A4 direct-weight optimizationRTN-W4A4G1287.8436
LLaMA-3-8BSpinQuant (local reproduction)hadYesW16A4 SpinQuant rotationGPTQ-W4A4G1287.56
LLaMA-3-8BSpinQuant (local reproduction)nohadNoW16A4 SpinQuant rotationGPTQ-W4A4G12839.73
LLaMA-3-8BRelaxed FKLnohadNoW4A4 direct-weight optimizationRTN-W4A4G12812.1814

The SpinQuant rows are local reproductions; separately reported paper values are 5.9/9.2 and 7.1/18.6. In the nohad scope, relaxed FKL RTN gives lower PPL than local SpinQuant GPTQ: 7.8436 vs. 13.45 and 12.1814 vs. 39.73. This is not a calibration- or compute-matched ranking; the relaxed rows are RTN, not GPTQ, endpoints.

W3 / Q2: The method uses conventional QAD; equal-budget controls and seed variance are missing. Which component produces the gain?

Answer: Our contribution is not a new STE, KL loss, or optimizer. We test whether exact equivalence should be maintained or released from the same initializer. The seed study uses forward KL, 800/32 C4 sequences, 500 updates, batch 64, 32,000 rank-samples, W4A4/G128/nohad evaluation, initializer seed 0, three predeclared training/data-order seeds, and the final checkpoint for both arms.

ModelConstrained mean PPL ± SDRelaxed mean PPL ± SDPaired relaxed − constrained PPLConstrained mean six-task ± SDRelaxed mean six-task ± SD
LLaMA-2-7B8.8543 ± 0.03798.2124 ± 0.5059-0.64190.5899 ± 0.00570.6044 ± 0.0030
TinyLlama-1.1B (small-model replication)13.6363 ± 0.081311.7167 ± 0.0106-1.91960.4659 ± 0.00600.4924 ± 0.0006

The table reports mean, sample SD, and paired relaxed-minus-constrained PPL for rows with all three seeds completed. Because the rebuttal period was short, we completed this study only for LLaMA-2-7B and TinyLlama-1.1B; no LLaMA-3 seed-variance claim is made. TinyLlama uses a similar budget and evaluator family but a different model and setup. It supports the observed direction, not model-scale generalization. The seeds vary training and data-order randomness only; initializer variance is not estimated.

Dear Reviewer 3wdS,

As the discussion period is approaching its end, we would like to kindly follow up and ask whether our rebuttal has addressed your concernes. We have provided additional experimental results and clarifiactions in response to your comments and would greatly appreciate your feedback.

Best regards, Authors

The W3 results and constrained-vs-relaxed seed study are helpful. However, the interaction analysis remains correlational, the SpinQuant comparison is not matched (RTN vs GPTQ; SpinQuant is stronger under had), and the method is still largely structured initialization plus conventional QAD. My main concerns therefore remain, and I keep my rating of 2 (weak reject).

Dear 3wdS,

Thank you for your feedback and for taking the time to consider our additional results.

We would like to clarify that, although the quantizers are not matched, GPTQ generally outperforms RTN under comparable conditions, meaning that this difference would typically favor SpinQuant rather than our method. Moreover, SpinQuant-had requires additional online transformations, whereas our method does not. We therefore believe that SpinQuant-nohad is the more relevant comparison for the deployment setting.

While we acknowledge the remaining limitations, we believe that our additional W3A4 results, multi-seed experiments, and error-compensation analysis provide meaningful evidence addressing several of your original concerns.

Given these additional results and clarifications, we would sincerely appreciate it if you could reconsider your rating, particularly in light of the experimental evidence provided during the rebuttal.

Best regards, Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Author Remarks

Dear Area Chairs and Reviewers,

We sincerely thank the reviewers and Area Chairs for their thoughtful feedback and consideration.

As the discussion period comes to a close, we would like to emphasize the central contribution of our work and summarize the additional evidence provided during the rebuttal.

1. Contribution and novelty

Our contribution is not a new QAT optimizer, loss, or quantization primitive. Rather, we investigate a fundamental assumption in quantization preprocessing: whether exact full-precision equivalence must be preserved throughout optimization.

By comparing equivalence-constrained and relaxed optimization from the same equivalence-preserving initialization under matched experimental settings, we directly examine the practical consequences of this constraint. We believe this provides a distinct, experimentally grounded contribution beyond combining existing QAT components.

2. Additional experimental evidence

To address the reviewers’ concerns, we provided:

  • Lower-bit validation: Deployable W3A4 results showing that relaxed optimization consistently improves over the equivalence-constrained baseline on LLaMA-2-7B (PPL 9.24 vs. 11.52) and LLaMA-3-8B (16.17 vs. 22.66), with corresponding improvements in downstream accuracy.
  • Error-compensation evidence: Held-out logit analysis showing negative interactions between full-precision drift and quantization error, consistent with partial compensation.
  • Reproducibility: Three-seed W4A4 experiments on LLaMA-2-7B and TinyLlama-1.1B, showing improvements in mean PPL under the tested protocol.
  • Stronger PTQ comparison: Additional results against SpinQuant+GPTQ, with deployment differences and limitations explicitly documented.

3. Remaining limitations and significance

We acknowledge that the cross-method comparisons are not fully matched and that the error-compensation analysis does not establish causality. Nevertheless, the controlled constrained-versus-relaxed experiments provide direct evidence that maintaining exact equivalence can restrict PTQ performance under practical, finite-budget settings.

We believe this finding contributes a useful design insight beyond the specific QAD implementation.

We would be grateful if the reviewers and Area Chairs could reconsider the contribution of our work in light of these additional results and clarifications.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.