acceptodds
Withdrawn from ICLR 2027

BitRot: Understanding Quantization Fragility in Vision Transformers Under Distribution Shift

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Post-training quantization (PTQ) enables efficient deployment of Vision Transformers, but preserving clean ImageNet accuracy do not necessarily preserve performance under distribution shift. This is important in deployment settings where low-precision models may encounter corruptions, visual domain shifts, and naturally adversarial examples, yet the interaction between aggressive quantization and robustness under distribution shift remains underexplored. We systematically study this interaction across architectures, PTQ methods, pretraining strategies, model scales, activation precisions, calibration settings, and several forms of distribution shift. We find that W4A4 models can retain near-lossless in-distribution accuracy while suffering substantially larger degradation on shifted data, with the severity varying across distributions and model configurations. We further test whether this fragility is primarily explained by activation-statistics and calibration mismatch. Mixed calibration using samples from multiple shifted distributions provides only modest and inconsistent recovery, leaving a substantial gap to full precision. Motivated by these observations, we develop a post-hoc prediction-level model that treats quantization as a perturbation to full-precision classification margins and characterizes susceptibility through both the available margin and its sensitivity to accumulated quantization residuals. The resulting predictor tracks quantization-induced degradation across distributions, architectures, model scales, and activation bit-widths. At the sample level, examples with comparable full-precision margins but higher directional quantization sensitivity are consistently more likely to be misclassified after quantization. These results show that clean ImageNet accuracy alone can substantially underestimate the deployment cost of low-precision ViTs, and that robustness under distribution shift should be evaluated explicitly when assessing PTQ methods.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
Quantization, Vision Transformers, Robustness

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_1qYSOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how post-training quantization affects the robustness of Vision Transformers under distribution shifts. The authors evaluate several quantization methods across different architectures, model sizes, pretraining strategies, and datasets. They show that preserving clean ImageNet accuracy does not necessarily preserve performance under distribution shift. The paper also introduces a prediction-level model based on classification margins and quantization sensitivity to explain why certain samples are more affected by quantization. The experimental results provide useful insights into the robustness of low-precision Vision Transformers.

Critical strengths
  • The paper addresses an important and practically relevant problem. Most quantization methods focus on preserving clean accuracy, while their behavior under distribution shifts receives less attention. The authors provide a useful analysis showing why clean accuracy alone may not be sufficient for evaluating quantized models.

  • The experimental study is quite comprehensive. The authors consider multiple quantization methods, architectures, model sizes, activation precisions, and pretraining strategies. The inclusion of ImageNet-C, ImageNet-R, and ImageNet-A helps examine different types of distribution shifts.

  • I find the proposed margin-based analysis particularly interesting. Rather than only reporting accuracy degradation, the authors attempt to explain quantization fragility through classification margins and sensitivity to quantization errors. The reported correlations and sample-level analysis provide encouraging evidence for this explanation.

Critical weaknesses
  • While the proposed prediction-level model is interesting, I would like to see a stronger validation of its individual components. The model relies on a first-order Taylor approximation and a Gaussian approximation of accumulated quantization errors. The reported correlations are promising, but it is not entirely clear how much improvement comes from considering directional quantization sensitivity compared with simpler measures such as classification margins or quantization-error magnitude alone.

  • The analysis of calibration mismatch could be explored further. The authors show that activation-distribution differences are relatively small and that mixed calibration provides limited improvements. However, the activation statistics mainly describe overall distribution differences, while a small number of outliers or highly sensitive layers could still contribute significantly to the observed degradation. It would be useful to examine whether certain layers or activation regions are more responsible for the robustness loss. Also, the ImageNet-C experiments consider only four corruption types at severity 5, which limits the conclusions across corruption severities.

  • I noticed a few inconsistencies in the reported experimental results that should be clarified. For example, the W4A4 AdaLog accuracy for Swin-Base under Gaussian noise is 30.068% in Table 1, whereas Appendix Table 4 reports 41.978%. There are also some non-monotonic results when increasing activation precision, particularly for DeiT-Small. These may be reporting issues or differences in experimental configurations, but an explanation would help make the results more reliable.

Questions for authors
  1. How much additional predictive information does directional quantization sensitivity provide compared with the full-precision classification margin alone? Could the authors compare the proposed model against simpler predictors based on margin, quantization-error magnitude, and their combination? I think this would help establish the contribution of the proposed formulation.

  2. The paper suggests that activation-statistics mismatch alone cannot explain the robustness degradation. Could the authors provide a layer-wise analysis of quantization sensitivity or examine whether a small number of highly sensitive layers contribute disproportionately to the degradation? It would also be interesting to see whether the observed behavior remains consistent across different corruption severities.

  3. Could the authors clarify the differences between the results reported in Table 1 and Appendix Table 4, particularly for Swin-Base under Gaussian noise? Also, why does increasing activation precision sometimes lead to worse performance for certain configurations, such as DeiT-Small? Are these differences related to calibration, quantizer optimization, or experimental variability?

1 comment

We thank the reviewer for the time and effort to review our paper and providing valuable comments.

Comparison Against Simpler Predictors

We agree that directional sensitivity should be evaluated against simpler measures. We compared our model against four baselines in the same setting as our post-hoc model.

  1. Raw FP margin uses the negative mean correct-versus-runner-up logit margin over FP-correct samples.
  2. Normalized FP margin divides each sample’s margin by its logit standard deviation before averaging and negating.
  3. Activation quantization error measures normalized squared quantization residuals, averaged equally across quantized sites.
  4. Combined quantization error averages the activation and weight normalized squared errors with equal weights.
PredictorA (n=63)B (n=63)C (n=189)
Raw FP margin0.7683 / 0.73770.6820 / 0.68190.5888 / 0.5813
Normalized FP margin0.5192 / 0.43360.5983 / 0.57400.5059 / 0.4979
Activation Quantization error0.3625 / 0.3920−0.1803 / −0.15610.3636 / 0.4025
Combined Quantization error0.3579 / 0.3853−0.1842 / −0.16600.3568 / 0.3920
Proposed model0.9086 / 0.88720.8109 / 0.77530.8360 / 0.7954

Values report Spearman / Pearson correlation across seven datasets.

  • Setting A includes ViT scales across W4A3, W4A4, and W4A5.
  • Setting B includes ViT, DeiT, and Swin across scales at fixed W4A4.
  • Setting C includes all architectures, scales, and bit-widths.

Our model achieves higher correlations in all three pooled settings. At fixed W4A4, Spearman correlation increases from 0.6820 for raw classification margins to 0.8109.

Table 3 provides complementary sample-level evidence. Within narrow FP-margin bins, samples with higher first-order margin-contraction scores exhibit 4.19 to 27.04 percentage points higher misclassification rates. This supports the additional explanatory value of directional sensitivity among samples with comparable margins.

The combined-error baseline combines activation and weight errors rather than margins and error magnitude. We will make the baseline definitions and the scope of these findings explicit.

Robustness Across Corruption Severities

We extended the evaluation to severities 1, 3, and 4 for motion blur, fog, Gaussian noise, and contrast, using AdaLog W4A4 on ViT-Base, DeiT-Base, and Swin-Base. Each evaluation uses 50,000 images. The table reports the full-precision minus quantized top-1 accuracy gap in percentage points.

BackboneCorruptionSeverity 1Severity 3Severity 4
ViT-BaseMotion blur9.0116.5319.91
ViT-BaseFog11.1615.1116.35
ViT-BaseGaussian noise7.0413.1318.14
ViT-BaseContrast9.4117.1030.57
DeiT-BaseMotion blur6.869.979.81
DeiT-BaseFog4.496.797.45
DeiT-BaseGaussian noise5.218.4311.41
DeiT-BaseContrast5.507.8213.17
Swin-BaseMotion blur3.647.8710.28
Swin-BaseFog4.848.5910.27
Swin-BaseGaussian noise4.369.1913.49
Swin-BaseContrast4.728.2512.39
Mean6.3510.7314.44

We observe substantial gaps below severity 5. At severity 3, the gaps range from 6.79 to 17.10 percentage points. They generally increase with severity, although not strictly for every configuration. For example, the DeiT-Base motion-blur gap decreases slightly from 9.97 points at severity 3 to 9.81 at severity 4.

These results show that the degradation is not confined to the most severe corruptions. Its magnitude depends on corruption type, severity, and architecture. We will include these evaluations in the revision.

Table Discrepancies and Activation-Precision Anomalies

We identified two separate issues:

  1. The Swin-Base Gaussian Noise value of 41.978% in Table 4 was inadvertently copied from the APHQ-ViT result. The actual value is same as the value for Swin-Base Gaussian Noise value of 30.068% in Table 1.
  2. A subset of the W4A3 and W4A5 runs underlying Table 4 omitted the scale-reparameterization flag. We reran the affected configurations with reparameterization enabled as intended. The verified Swin-Base W4A4 AdaLog Gaussian Noise accuracy matches Table 1. The missing reparameterization also affected the DeiT-Small W4A5 checkpoint. The original comparison therefore did not isolate the effect of activation precision. We will replace the affected entries with the corrected evaluations. The corrected Setting A Spearman/Pearson correlations are 0.9086/0.8872, compared with approximately 0.857/0.836 in the submission. The qualitative findings remain unchanged. We acknowledge these reporting and configuration errors and will update Table 4, Figure 5a, and the associated discussion.
Reviewer_RKmdOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates the effects of post-training quantization (PTQ) on Vision Transformers under distribution shift. It evaluates multiple PTQ methods on ViT, DeiT, and Swin, and develops a post-hoc prediction model based on full-precision classification margins and their sensitivity to quantization errors. The observation that near-full-precision accuracy on clean data does not imply preserved robustness under distribution shift is valuable. However, the experiments remain concentrated on earlier image-classification backbones. The limited evaluation of larger VLMs, text-only and omni-modal models, recent quantization methods, and actual low-bit deployment does not sufficiently establish the applicability of the analysis to current large-model systems.

Critical strengths
  1. The problem selection is good : retaining clean ImageNet accuracy after quantization does not ensure that robustness under distribution shift is preserved. The experiments consider multiple PTQ methods, pretraining strategies, model scales, and activation precisions.

  2. The margin-and-sensitivity analysis provides an interpretable hypothesis for quantization fragility. Configuration-level correlations and comparisons between samples with similar margins but different sensitivities provide empirical support for this hypothesis.

Critical weaknesses
  1. Experiments on larger, more recent models are missing. The experiments are limited to ViT, DeiT, and Swin, without evaluating complete large VLMs. Additional experiments on high-capacity VLMs, such as LLaVA-1.5-13B [1] and Qwen3-VL-32B [2], as well as omni-modal models such as Qwen3-Omni [3], are needed to assess the scope of the conclusions. The authors should also explain how classification margins would be defined in these models, how gradient-based quantization-error estimation would be adapted, and how the computational cost scales with model size.

  2. Evaluation across additional modalities is missing. The design of the margin-perturbation and sensitivity model does not appear inherently specific to vision. The experiments with GPTQ, SmoothQuant, and QuaRot are still conducted on image classifiers. Experiments on text-only LLMs should be included.

  3. Actual low-bit GPU deployment has not been demonstrated. The paper studies the effects of numerical precision on robustness but does not report measured GPU memory usage or inference latency. Full-precision and quantized configurations should be compared under identical conditions, including on larger models. It is also unclear whether W4A4 uses native low-bit kernels or quantization simulated through floating-point operations. Please specify the inference backend, weight-packing format, activation kernels, and floating-point fallbacks.

  4. Comparisons with stronger, more recent quantization methods are missing. The evaluated methods mainly date from 2023–2025. Recent ViT PTQ methods, such as LS-ViT [4], should be included under matched precision and calibration budgets to determine whether the fragility findings still hold. For VLM and omni-modal extensions, VLM-PTQ [5] is relevant method to consider.

  5. The discussion of recent quantization research remains insufficient. The paper does not discuss the recent ViT, VLM, and omni-modal PTQ methods in [4–5]. The related work should explain how these methods affect the scope of the empirical conclusions and what the margin-based analysis adds to existing explanations of quantization reliability.

  6. An overview diagram of the analysis pipeline is missing. The paper lacks a figure that clearly connects calibration, full-precision and quantized execution, error estimation, classification-margin gradients, and misclassification-probability prediction. Such a diagram would make the procedure easier to understand.

[1] H. Liu, C. Li, Y. Li, and Y. J. Lee, “Improved Baselines with Visual Instruction Tuning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024.

[2] S. Bai et al., “Qwen3-VL Technical Report,” arXiv preprint arXiv:2511.21631, 2025.

[3] J. Xu et al., “Qwen3-Omni Technical Report,” arXiv preprint arXiv:2509.17765, 2025.

[4] H. Hwang, X. T. Nguyen, and H.-J. Lee, “LS-ViT: Least-Squares Hessian Based Block Reconstruction for Low-Bit Post-Training Quantization of Vision Transformers,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026. [5] J. Deng and K. Huang, “VLM-PTQ: Efficient Post-Training Quantization for Large Vision-Language Models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026.

Questions for authors
  1. Why is the analysis restricted to vision classifiers? How does the margin-and-sensitivity formulation differ for text-only or omni-modal Transformers? Can the same analysis predict quantization failures in video-understanding models?

  2. The experimental backbones and quantization comparisons mainly come from earlier model generations. Without experiments on larger, more recent models and recent PTQ methods, I remain unconvinced that the conclusions extend to current large-model deployment. Can the authors provide such evidence, including measured GPU deployment performance?

1 comment

We thank you for taking the time and effort to review our paper and providing valuable comments.

Scope and Potential Extension to VLMs & Other Modalities

The scope of our study is to evaluate Vision Transformers and explain why low-bit PTQ causes a disproportionate accuracy drop under different distribution shifts despite near-lossless clean accuracy. While evaluating large VLMs, text-only LLMs, and omni-modal models is beyond this scope, our margin-and-sensitivity formulation in Section 4 is not inherently restricted to vision classifiers and could be extended to these architectures in future work as follows.

Defining Margins Across Modalities: Our framework (Eqs. 1–12) operates on a scalar score gap mc(x)=fy(x)−fkc(x)(x)m_c(x) = f_y(x) - f_{k_c(x)}(x) between the full-precision prediction yy and top-CC competitors kc(x)k_c(x). For discriminative or multiple-choice tasks across text, image, audio, or video inputs, fk(x)f_k(x) could be defined as the candidate answer logit (for single-token outputs) or the conditional sequence log-likelihood log⁡Pθ(yk∣x)\log P_\theta(y_k \mid x) (for multi-token candidate responses).

For open-ended autoregressive generation, an analogous step-wise margin could be defined between the full-precision top token and competing tokens at a shared prefix y<ty_{<t}:

mc,t(x,y<t)=zyt−zkc,t,m_{c,t}(x,y_{<t}) = z_{y_t} - z_{k_c,t},

helping identify prefixes where quantization could flip a token and alter the continuation. Extending this analysis to complete generations is more complicated. Once a token changes, subsequent predictions depend on different prefixes, and a local token flip may preserve answer correctness, redirect reasoning, or prolong generation. We would therefore need to connect local margin crossings to sequence-level outcomes, with additional gradient evaluations across decoding steps.

Inference Backend and Why Latency Does Not Affect the Analysis

Following standard ViT PTQ benchmarks such as AdaLog (Table 1), all W4A4 evaluations use simulated quantization in PyTorch on an NVIDIA L40 GPU: selected weights (per-channel) and activations (per-tensor across linear inputs, attention logits/probabilities, and post-GELU features) are clipped and rounded to 4-bit uniform or logarithmic grids and dequantized for floating-point execution, while LayerNorm and residual operations remain in floating point (Section 2, Appendix A.5).

Measuring native low-bit GPU latency and memory requires a deployment backend and depends on kernel support, weight packing, and logarithmic activation handling. Our empirical findings and explanatory analysis concern the numerical effects of quantization within the evaluated simulation setup. For each method and backbone, we use the same calibrated checkpoint and execution procedure on clean and shifted inputs. Native deployment measurements would complement these accuracy comparisons.

We will clarify the simulation details in Section 3.1 and include an overview diagram of the Section 4 pipeline in the revision.

Comparison with Recent PTQ Methods: LS-ViT (CVPR 2026)

We evaluated LS-ViT (CVPR 2026) at W4A4 across ViT-Base, DeiT-Base, and Swin-Base using clean ImageNet calibration. Each cell reports top-1 accuracy (%) and, in parentheses, the drop relative to the full-precision reference in Table 1 (AccFP−AccLS−ViTAcc_{FP} - Acc_{LS-ViT}, in percentage points), in percentage points) across four severity-5 ImageNet-C corruptions, ImageNet-R, and ImageNet-A.

Backbone (W4A4)GaussianMotionFogContrastIN-RIN-A
ViT-Base42.84 (11.33)40.11 (12.21)53.35 (12.47)28.18 (8.42)53.19 (6.01)31.21 (18.45)
DeiT-Base35.03 (10.65)32.01 (3.13)54.80 (8.13)32.35 (16.73)41.78 (3.11)16.31 (10.80)
Swin-Base42.21 (6.75)41.80 (7.60)58.68 (8.43)40.86 (9.80)52.97 (6.34)37.24 (14.48)

Even with LS-ViT's least-squares Hessian block reconstruction, a substantial full-precision gap persists across all evaluated shifted conditions, ranging from 3.11 to 18.45 percentage points (e.g., 18.45 points on IN-A and 12.47 points on Fog for ViT-Base).

Furthermore, under mixed calibration (oracle exposure to shifted samples) on ViT-Base, LS-ViT reaches 46.19% on Gaussian Noise (+3.35 points over clean calibration), 54.60% on IN-R (+1.41 points), and 35.01% on IN-A (+3.80 points), leaving remaining full-precision gaps of 7.98, 4.60, and 14.66 percentage points, respectively. Thus, substantial degradation under shift persists with this recent PTQ method, including under mixed calibration. We will add the LS-ViT baseline to Table 1.

Reviewer_5RyhOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates the robustness of post-training quantized Vision Transformers under distribution shifts. The authors systematically evaluate several existing PTQ methods across different architectures, model sizes, activation bit-widths, pretraining strategies, and datasets, including ImageNet-C, ImageNet-R, and ImageNet-A. The results show that low-bit quantization can cause substantially larger accuracy degradation under distribution shifts than on clean ImageNet, while mixed-distribution calibration provides limited recovery. To explain this phenomenon, the authors introduce a probabilistic model based on first-order Taylor approximation, classification margins, and quantization error statistics to estimate quantization-induced misclassification probabilities. The model demonstrates correlations with empirical accuracy degradation across different experimental settings.

Critical strengths

1. Extensive Experimental Evaluation. The paper provides a broad empirical evaluation by reproducing multiple existing PTQ methods and testing them across different Vision Transformer architectures and publicly available distribution-shift benchmarks. These experiments provide useful comparisons of quantization robustness under different conditions.

2. Analysis of Quantization Fragility. Beyond reporting accuracy degradation, the authors investigate several factors affecting quantization robustness, including model scale, activation precision, pretraining strategies, and calibration data. These analyses provide additional empirical observations that may help understand the behavior of low-bit quantized models under distribution shifts.

Critical weaknesses

1. Unclear Motivation and Practical Significance. The motivation of this paper is not sufficiently justified. Although the authors propose a prediction-level model to estimate quantization-induced accuracy degradation, its practical significance remains unclear. Even if the model successfully predicts accuracy degradation, how can this information be used to improve quantization performance or guide the development of more robust quantization methods? The paper does not provide a convincing answer or demonstrate actionable insights from the proposed model.

2. Limited Research Contributions and Unfocused Presentation. The paper reads more like an extensive technical report than a research paper with clearly defined contributions. Most of the content consists of evaluations and analyses of existing quantization methods across different experimental settings, without a clear central research hypothesis or novel technical solution. If the primary contribution is the prediction-level model, the extensive empirical analysis does not sufficiently establish its necessity or significance.

3. Questionable Novelty of the Prediction Model. The proposed model largely relies on conventional quantization error analysis techniques, including first-order Taylor expansion, gradient-based error propagation, and Gaussian approximation. Similar techniques have already been extensively explored in prior studies, such as BRECQ (ICLR 2021) and Floating-Point Quantization Analysis of Multi-Layer Perceptron Artificial Neural Networks (2024). The authors should clearly explain what fundamentally new theoretical contribution distinguishes their formulation from existing quantization error propagation models.

4. Insufficient Validation of the Proposed Model. The experimental validation of the prediction model is not convincing. The authors mainly report correlations between predicted misclassification probabilities and empirical accuracy degradation, without comparing against simpler predictors based on classification margins or quantization error magnitudes. Therefore, it remains unclear whether the proposed probabilistic formulation provides any meaningful advantage over conventional sensitivity analysis.

5. Inconsistencies in Experimental Results. Some experimental results are inconsistent between the main paper and appendix. For example, the W4A4 AdaLog accuracy of Swin-Base on Gaussian Noise is reported as 30.068% in Table 1 but 41.978% in Table 4. Such substantial discrepancies raise concerns about the reliability of the experimental results and should be carefully verified.

Questions for authors

Given the fundamental concerns raised above, particularly regarding the unclear motivation, limited technical novelty, and insufficient research contributions, I believe these issues would be difficult to adequately address within the rebuttal period. Therefore, I am unlikely to change my overall recommendation unless the authors provide substantial evidence that directly addresses these concerns. Nevertheless, I would appreciate clarification on the following questions:

1. Motivation and Significance. What is the intended practical application of the proposed prediction-level model? Since the model requires access to target-distribution samples and computationally expensive gradient evaluations, what advantages does it offer over directly evaluating the quantized model? More importantly, how can its predictions provide actionable insights for improving quantization robustness?

2. Technical Novelty. What is the fundamental theoretical novelty of the proposed model compared with conventional quantization error propagation techniques? The formulation appears to be a straightforward combination of first-order Taylor expansion, gradient-based sensitivity analysis, and Gaussian approximation, all of which have been extensively investigated in prior research. Could the authors clarify which specific aspects constitute a novel technical contribution?

3. Necessity of the Proposed Model. Why is the proposed probabilistic formulation necessary? Have the authors compared it against simpler alternatives, such as classification-margin-based predictors or direct first-order margin perturbation estimates? Without such comparisons, it is difficult to determine whether the additional statistical modeling provides meaningful benefits.

4. Experimental Reliability. Could the authors explain the inconsistencies between Table 1 and Table 4, particularly the substantial differences in the reported W4A4 AdaLog results for Swin-Base? Which results are correct, and do these inconsistencies affect the conclusions of the paper?

2 comments

We thank the reviewer for their time and valuable comments.

Motivation and Purpose of the Model

We ask why low-bit PTQ causes disproportionate and heterogeneous degradation under distribution shift despite near-lossless clean accuracy. Our hypothesis is that fragility depends on the interaction between FP margins and directional quantization sensitivity. Section 3 examines calibration mismatch and activation statistics, so their limited explanatory power motivates Section 4.

Our model is a post-hoc explanatory framework. We use its predictions to test this hypothesis against observed failures and answer the question we are asking, not to replace direct accuracy evaluation. Target samples and gradients help investigate why configurations and samples differ in susceptibility. The practical implication is to assess shifted performance and decision-level sensitivity alongside clean accuracy and reconstruction error. We have not demonstrated a robustness-improving quantization algorithm and will clarify this distinction.

Distinction from Prior Quantization Error Models

We agree that Taylor expansion, gradient propagation, and Gaussian approximations are established. Our contribution is their formulation and empirical validation for explaining quantization-induced prediction failures under shift.

  • BRECQ uses a second-order task-loss approximation to motivate block reconstruction. We analyze competing classification margins and their boundary-crossing probabilities.
  • Al-Rikabi & Renczes (2024) analyze floating-point error propagation and binary-threshold crossing. We formulate joint multiclass crossing for low-bit ViTs, modeling clipping-induced mean shifts, rounding variability, and correlations between competing margins:
(\Sigma_m)_{cd}=g_c^\top\Sigma_e g_d,\qquad \pi_C=1-\Phi_C(\boldsymbol{\rho};\mathbf{R}).$$ We validate this explanation across distributions and configurations, and among samples with comparable FP margins. We will make this contribution and its relationship to established techniques explicit. ### Comparison Against Simpler Predictors We compared our model against four baselines using empirical AdaLog relative accuracy drop as the target: 1. **Raw FP margin:** Negative mean correct-versus-runner-up logit margin over FP-correct samples. 2. **Normalized FP margin:** The same score after dividing each sample's margin by its logit standard deviation. 3. **Activation Quantization error:** Mean normalized squared quantization residual, averaged equally across sites. 4. **Combined Quantization error:** Equal-weighted activation and weight normalized squared errors. | Predictor | A ($n=63$) | B ($n=63$) | C ($n=189$) | |:--|:--:|:--:|:--:| | Raw FP margin | 0.7683 / 0.7377 | 0.6820 / 0.6819 | 0.5888 / 0.5813 | | Normalized FP margin | 0.5192 / 0.4336 | 0.5983 / 0.5740 | 0.5059 / 0.4979 | | Activation Quantization error | 0.3625 / 0.3920 | −0.1803 / −0.1561 | 0.3636 / 0.4025 | | Combined Quantization error | 0.3579 / 0.3853 | −0.1842 / −0.1660 | 0.3568 / 0.3920 | | **Proposed model** | **0.9086 / 0.8872** | **0.8109 / 0.7753** | **0.8360 / 0.7954** | Values report Spearman $\rho$ / Pearson $r$ across seven datasets. - **A:** ViT scales × bit-widths (W4A3/4/5). - **B:** ViT/DeiT/Swin × scales at W4A4. - **C:** All architectures × scales × bit-widths. Our model achieves higher correlations in all three pooled settings, including fixed W4A4, where Spearman $\rho$ improves from $0.6820$ to $0.8109$ over raw margins. Table 3 provides complementary sample-level evidence: within narrow FP-margin bins, samples with higher first-order contraction scores $S_Q=-\Delta m_1^{lin}$ exhibit 4.19–27.04 percentage points higher misclassification rates. These results support directional sensitivity beyond the evaluated margin-only and error-only summaries. We will add the baseline comparisons to Section 4. ### Resolution of Table Discrepancies We identified two issues: 1. The Swin-Base Gaussian Noise value 41.978% in Table 4 was inadvertently copied from the APHQ-ViT result. 2. A subset of Table 4 scale/bit-width runs omitted the scale-reparameterization flag. We reran the affected configurations (W4A3 and W4A5) with reparameterization enabled as intended. The verified Swin-Base W4A4 AdaLog Gaussian Noise accuracy in Table 4 matches the value reported in Table 1 (30.068%). The corrected Setting A Spearman/Pearson correlations are 0.9086/0.8872, compared with approximately 0.857/0.836 in the submission. The qualitative findings remain unchanged. We acknowledge the reporting and configuration errors and will correct Table 4 (in the appendix) and Figure 5a, identifying the affected entries.
Reviewer_5RyhOpenReview ↗

Thank you for the clarification and here is my final assessment

I thank the authors for their detailed rebuttal and additional experiments. I appreciate their efforts to clarify the experimental inconsistencies and provide additional comparisons against simpler predictors. These responses have addressed several of my secondary concerns. However, my two fundamental concerns regarding practical significance and technical novelty remain unresolved.

1. Practical Significance. While I understand that the proposed model is intended as a post-hoc explanatory framework, I remain unconvinced of its practical value. I would find it substantially more meaningful if the analysis could guide quantization strategies to improve model robustness under distribution shifts, rather than merely predicting or explaining accuracy degradation. Without demonstrating such actionable benefits, I believe the overall contribution remains insufficient for ICLR.

2. Technical Novelty. I also remain unconvinced by the claimed methodological novelty. The proposed formulation largely builds upon established techniques, including Taylor approximation, gradient-based sensitivity analysis, and quantization error propagation. Similar principles have already been widely explored in prior quantization research, including mixed-precision bit-width allocation and optimization-guided quantization. Although the authors extend these ideas to multiclass margin-crossing probabilities, I do not consider this extension sufficiently novel to meet the ICLR acceptance bar.

As explicitly stated in my original review, these two concerns are fundamental to my assessment, and I would be unlikely to revise my recommendation without substantial evidence addressing them. While I acknowledge the authors' efforts and the improvements made during rebuttal, neither concern has been sufficiently resolved.

Therefore, I maintain my original rating as my final recommendation.

Comments on the paper

1 comment

Dear Reviewer,

I hope this message finds you well. I am writing to kindly remind you that the authors’ rebuttal has been posted and is now available for your review in the system.

We would greatly appreciate it if you could review the rebuttal and provide your assessment at your earliest convenience, as your input is important for us to move forward with a timely decision.

Thank you very much for your time and continued support.

ACs

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

Related papers

Discussion (0)

Sign in to comment.