acceptodds
Under review as a conference paper at ICLR 2027

Corruption Robustness in Post-Training Quantization of Vision Transformers: A Calibration Perspective

Abstract

Post-training quantization (PTQ) enables efficient deployment of pretrained models using only a small calibration set, but its robustness under distribution shifts remains underexplored. We observe that quantized models retain substantially less of their full-precision performance on corrupted inputs than on clean data, particularly under low-bit quantization. We study this problem from the perspective of calibration data and show that different transformations of the same fixed-budget calibration set lead to markedly different corruption robustness. Our analysis reveals that effective transformations tend to reduce activation energy, exhibiting a characteristic energy shift similar to that observed under common corruptions. We associate this reduction with two changes in self-attention: lower attention concentration and weaker directional alignment among value vectors associated with different tokens. Motivated by these observations, we introduce *spatial entropy of difference energy*, a model-independent score that measures how broadly a transformation distributes its changes across an image. This score strongly correlates with both activation-energy reduction and post-quantization corruption robustness, enabling transformation selection without model inference. Based on this proxy, we propose **Entropy-**, which selects multiple high-scoring transformations while preserving the original calibration budget. Entropy- requires neither corrupted target-domain samples nor modification of the underlying PTQ algorithm. Experiments across multiple Vision Transformer architectures and low-bit PTQ settings on ImageNet and ImageNet-C demonstrate consistent improvements in corruption robustness while maintaining competitive clean accuracy.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Post-Training Quantization, 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_FW1pOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on how calibrated data affects the damage robustness of low-bit training and quantization of visual Transformer. It links effective calibration transformation with the reduction of activation energy. It proposes Entropy-K, selecting the transformation by using the spatial entropy of image difference energy without model inference or increasing the number of calibration samples. Experiments covering different architectures and PTQ methods on ImageNet. The ablation studies and downstream task evaluations show that Entropy-K can improve damage robustness while typically maintaining similar clean data accuracy.

Critical strengths

Easy for practical application. Entropy-K does not require an increase in the number of calibration samples, acquisition of damaged data from the target domain, or modification of the underlying PTQ algorithm to improve damage robustness. Therefore, it is easy to integrate into existing quantization processes.

Extensive experimental verification. The experiments cover various architectures, PTQ methods, and low-bit settings, demonstrating the applicability of this method beyond a single model or quantization method.

Critical weaknesses

The incremental gains are limited. Entropy-K outperforms clean data calibration, but compared to stronger single transformations such as PRIME, its average damage robustness gains are smaller. These comparisons lack uncertainty estimates from repeated experiments, so it is difficult to determine whether entropy-guided selection provides stable and practically significant advantages.

The method innovation is limited. Existing works such as MetaAug have explored improving PTQ by calibrating data transformations. Entropy-K mainly adds a spatial entropy sorting heuristic for selecting and combining existing transformations. Although the selection indicators without model inference and the analysis for damage robustness have independent contributions, I believe its advancement compared to existing calibration data methods is still relatively incremental.

Questions for authors

i. What evidence separates the value of entropy-based ranking from simply choosing established strong augmentations? Does the advantage remain against a mixture selected using prior robustness knowledge, without entropy, under the same calibration budget?

ii. If the Gaussian-noise candidates in Appendix B were added to the pool, would Entropy-K select them? What rule excludes high-entropy but ineffective candidates without using corruption-test results?

iii. How stable are the modest gains over PRIME across calibration samples and random seeds? Can the authors report paired differences and uncertainty for both clean and corruption accuracy?

iv. Which data and metrics determined K, augmentation strengths, and the candidate pool? Were any of these choices informed by ImageNet-C results? Please distinguish the data used to determine these settings from those used for the final entropy-based ranking.

6 comments
AuthorsOpenReview ↗

Response to Reviewer FW1p (1)

We thank the reviewer for the constructive and detailed feedback. We address each point below.

W1 & Q1. Comparison with PRIME and robustness-oriented mixture.

Comparison with PRIME. Averaged over the seven architectures in Table 23 (Appendix K), Entropy-KK outperforms PRIME in both clean and corruption accuracy under all three bit-widths, whereas PRIME lowers clean accuracy below the clean-calibration baseline.

MethodW3A3 IDW3A3 CW3A3 Avg.W2A4 IDW2A4 CW2A4 Avg.W2A3 IDW2A3 CW2A3 Avg.
Baseline71.1940.5155.8567.2336.9652.0956.0525.3240.69
PRIME70.9242.7156.8266.5238.4552.4855.5027.6941.59
Entropy-KK71.4442.8357.1467.4038.8253.1156.6128.0842.35

While the corruption gains over PRIME are modest, Entropy-KK improves corruption robustness without sacrificing clean accuracy. The stability of these gains across calibration seeds is reported in our response to Q3.

Comparison with a prior-knowledge mixture. To separate the benefit of entropy-based ranking from simply choosing established strong augmentations, we evaluated Robust-3, a mixture of three established robustness-oriented augmentations (PRIME, AugMix, and HybridAugment++) selected without SentS_{\mathrm{ent}}, under the same calibration budget. Notably, these are also the three best single transformations under W3A3 in Table 1, so Robust-3 effectively reflects prior knowledge of ImageNet-C performance.

MethodW/AViT-S IDViT-S CDeiT-S IDDeiT-S CSwin-S IDSwin-S CAvg.
Baseline3/364.5031.1769.1838.1177.9144.7654.27
Robust-363.7833.1869.8040.3978.2046.3455.28
Entropy-364.6533.8269.9340.9478.2446.8555.74
Baseline2/342.0215.4252.4522.8767.7331.3738.64
Robust-341.6117.3055.2427.4768.5033.3240.57
Entropy-342.4717.7854.7626.9868.3634.1540.75

Entropy-3 achieves higher average accuracy than Robust-3 under both W3A3 and W2A3 (+0.46+0.46 and +0.18+0.18 points). Robust-3 and Entropy-3 share PRIME and AugMix and differ only in the third transformation: SentS_{\mathrm{ent}} selects MixUp, which is not typically regarded as a robustness-oriented augmentation, instead of HybridAugment++. This suggests that entropy-based ranking identifies effective calibration transformations beyond what prior robustness knowledge suggests, without using any corruption-test results.

In the revised manuscript, we added the comparison with Robust-3 in Section 6 (Table 2(a)) and Appendix G.3 (Table 16).

AuthorsOpenReview ↗

Response to Reviewer FW1p (2)

W2. Novelty and distinction from existing calibration augmentation methods.

Entropy-KK differs from existing calibration augmentation methods such as MetaAug (Pham et al., 2024) in both what the calibration data is designed for and how the transformations are obtained:

MetaAugEntropy-KK (ours)
GoalMitigate calibration-set overfittingCorruption robustness
EvaluationClean accuracy, CNNsClean and corruption accuracy, ViTs
TransformationsLearned by bi-level optimizationSelected from existing ones by SentS_{\mathrm{ent}}
Training / inferencePer model and bit-widthNone (once from input images)

Hence, the novelty of our work lies not in the ranking rule alone, but in the following, which existing calibration-data methods, including MetaAug, do not examine: (i) we show that low-bit PTQ of ViTs preserves substantially less FP performance under common corruptions than on clean inputs (Fig. 1), and we address this gap on the calibration-data side without modifying the PTQ algorithm or the deployed model; (ii) we identify activation-energy reduction as a shared characteristic of effective calibration transformations and corrupted inputs, and explain it through reduced attention concentration and value alignment (Section 4); (iii) we derive SentS_{\mathrm{ent}} from this analysis as a forward-free proxy, computed once from input images and reusable across architectures and bit-widths.

AuthorsOpenReview ↗

Response to Reviewer FW1p (3)

Q2. High-entropy Gaussian noise candidates.

As the reviewer pointed out, the three Gaussian-noise variants in Appendix B.4 (Sent=0.998S_{\mathrm{ent}}=0.998–0.9990.999) would rank above PRIME (0.9660.966) if added to the pool. However, their high score has a different origin. Augmentation methods transform the image itself (e.g., by mixing images or altering colors), so the resulting change T(x)−xT(x)-x depends on the image content, and SentS_{\mathrm{ent}} measures how this change is distributed. Additive noise is sampled independently of the image, so every patch receives nearly the same noise energy and Sent≈1S_{\mathrm{ent}}\approx 1 regardless of the content.

Such perturbations can be identified from the inputs alone by checking whether δ=T(x)−x\delta=T(x)-x depends on xx:

ρ(x,T)=⟨x~,δ~⟩∥x~∥2∥δ~∥2,D(T)=1∣Dcal∣∑x∈Dcal∣ρ(x,T)∣,\rho(x,T)=\frac{\langle \tilde{x},\tilde{\delta}\rangle}{\lVert\tilde{x}\rVert_2\lVert\tilde{\delta}\rVert_2},\qquad D(T)=\frac{1}{|\mathcal D_{\mathrm{cal}}|}\sum_{x\in\mathcal D_{\mathrm{cal}}}|\rho(x,T)|,

where x~\tilde{x} and δ~\tilde{\delta} denote the mean-centered input and perturbation. Candidates with D(T)<0.1D(T)<0.1, a simple threshold, are flagged before ranking (1,024 calibration images).

TransformationSentS_{\mathrm{ent}}D(T)D(T)Decision
Gaussian (4/255)0.99910.022Flagged
Gaussian (16/255)0.99860.046Flagged
Gaussian (32/255)0.99820.079Flagged
PRIME0.9660.583Passed
MixUp0.9520.687Passed
AugMix0.9050.431Passed
RandAugment0.8710.523Passed
HybridAugment++0.8710.405Passed
CutMix0.8260.339Passed
TrivialAugment0.7870.504Passed
RandomErasing0.6370.349Passed
RVT-PWA0.4690.120Passed

All three noise variants are flagged and all nine original candidates pass, so the selection is unchanged. RVT-PWA, the candidate closest to the threshold (D(T)=0.120D(T)=0.120), also yields negligible gains over clean calibration in Table 1 (−0.02-0.02 and +0.25+0.25 points under W3A3 and W2A3).

In the revised manuscript, we discussed this issue in Appendix B.5 (Table 6).

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

This paper studies how calibration data affects the corruption robustness of post-training quantized vision models. The authors show that different calibration transformations can lead to different robustness outcomes, and analyze the relationship between effective transformations, activation energy, and attention behaviors. Based on these observations, they propose Spatial Entropy as a model-free criterion for selecting calibration transformations and introduce Entropy-K.

Critical strengths
  1. Instead of focusing only on clean accuracy, the paper highlights the overlooked impact of calibration data on the OOD robustness of quantized models, which is valuable for building more reliable low-bit vision models.

  2. The paper provides a meaningful analysis of why certain calibration transformations improve robustness. The investigation of activation energy, attention behaviors, and Spatial Entropy offers useful insights into the relationship between calibration distributions and quantized model robustness.

Critical weaknesses
  1. The proposed model-independent design is appealing for reducing calibration cost, but calibration effectiveness is closely related to model-specific activation statistics and quantization configurations. More discussion or comparison with model-aware selection strategies would help clarify the trade-off between efficiency and effectiveness.

  2. Spatial Entropy plays a central role in the proposed method, yet its necessity is not fully established. It would be helpful to compare against simpler image-level criteria, such as overall perturbation magnitude or spatial coverage of changes, to understand whether Spatial Entropy provides additional useful signals beyond basic statistics.

  3. The analysis of activation energy and attention behaviors provides interesting insights, but the connection to Spatial Entropy remains somewhat indirect. It is unclear why an input-level spatial distribution metric can reflect the model-level changes observed in attention patterns.

  4. The empirical results show consistent improvements across several architectures and PTQ methods, but the stability of the entropy-based ranking is less clear.

Questions for authors
  1. Could the authors provide comparisons with model-aware selection strategies or discuss the performance gap between these two settings?

  2. Whether Spatial Entropy provides advantages over simpler statistics, such as perturbation magnitude or spatial coverage, for ranking transformations?

  3. How the input-level Spatial Entropy is connected to model-level behaviors (e.g., activation and attention patterns)?

6 comments
AuthorsOpenReview ↗

Response to Reviewer UvDt (1)

We thank the reviewer for the constructive and detailed feedback. We address each point below.

W1 & Q1. Comparison with model-aware selection.

To examine the trade-off between model independence and selection effectiveness, we additionally evaluated a model-aware strategy (Energy-KK) that selects the KK transformations with the lowest activation energy measured at the middle transformer block of each architecture, under the same calibration budget and evaluation protocol as the proposed method of Entropy-KK.

MethodW/AViT-S IDViT-S CDeiT-S IDDeiT-S CSwin-S IDSwin-S CAvg.
Entropy-33/364.6533.8269.9340.9478.2446.8555.74
Energy-363.7433.8269.9841.0578.3747.3755.72
Entropy-32/342.4717.7854.7626.9868.3634.1540.75
Energy-341.3017.8555.0328.1869.0634.7441.03

Entropy-3 performs on par with Energy-3 under W3A3 (55.7455.74 vs. 55.7255.72) and remains within 0.280.28 points under W2A3 (40.7540.75 vs. 41.0341.03), without using any model information. Energy-KK selects architecture-dependent transformations (MixUp, PRIME, and HybridAugment++ for ViT-S and DeiT-S; MixUp, HybridAugment++, and AugMix for Swin-S), suggesting that model-specific statistics can provide additional architecture-specific gains.

Note that Energy-KK requires forward passes and activation measurements for each architecture, whereas Entropy-KK uses only input-level statistics, computed once without model inference or a GPU and reused across all architectures and bit-widths (0.275 vs. 1.248 ms per image). Thus, the proposed method achieves comparable accuracy to the model-aware variant while substantially reducing computational overhead.

In the revised manuscript, we discussed these comparative results in Section 6 ("Comparison with model-aware selection") and Appendix G.4 (Table 17).

AuthorsOpenReview ↗

Response to Reviewer UvDt (2)

W2 & Q2. Comparison with perturbation magnitude and spatial coverage.

To support the advantage of the proposed SentS_{\mathrm{ent}}, we compared SentS_{\mathrm{ent}} with two simpler model-independent criteria: perturbation magnitude M(T)M(T), the mean relative change, and patch coverage Cpatch(T)C_{\mathrm{patch}}(T), the mean fraction of patches with a nonzero change. We evaluated the top-3 transformations of each criterion on ViT-S, DeiT-S, and Swin-S under W3A3 and W2A3 with the same calibration budget.

MethodW/AViT-S IDViT-S CDeiT-S IDDeiT-S CSwin-S IDSwin-S CAvg.
Coverage-33/364.6533.8269.9340.9478.2446.8555.74
Magnitude-364.1132.3470.5440.6278.3145.7155.27
Entropy-364.6533.8269.9340.9478.2446.8555.74
Coverage-32/342.4717.7854.7626.9868.3634.1540.75
Magnitude-342.2116.5555.6126.6869.0832.9640.52
Entropy-342.4717.7854.7626.9868.3634.1540.75

Magnitude-3 (RandAugment, HybridAugment++, CutMix) is lower than Entropy-3 (PRIME, MixUp, AugMix) by 0.470.47 and 0.230.23 points on average, and lower in corruption accuracy in all six configurations. The amount of perturbation alone therefore does not identify effective transformations.

SentS_{\mathrm{ent}} inherently accounts for coverage, since a transformation that modifies only a few patches has low entropy. In addition, it weights each patch by its share of the perturbation energy, so negligible changes are attenuated rather than counted as full modifications. In the original pool the two criteria select the same set, but coverage saturates among the top candidates (99.9995%, 99.7653%, and 99.1380%), whereas SentS_{\mathrm{ent}} still separates them. The two criteria can therefore lead to different selections on other pools: on the nine unseen candidates (response to Reviewer dxxm, W1), Coverage-3 selects FACT, ColorJitter, and AutoAugment, whereas Entropy-3 selects FACT, ColorJitter, and TokenMix.

MethodW/AViT-S IDViT-S CDeiT-S IDDeiT-S CAvg.
Coverage-33/364.5833.0969.7840.2651.93
Entropy-364.1533.6670.2740.5752.16
Coverage-32/342.2517.0954.9025.9135.04
Entropy-343.4817.3356.3826.6035.95

Entropy-3 is higher by 0.230.23 and 0.910.91 points on average, and higher in corruption accuracy in all four configurations.

In the revised manuscript, we added these comparative results in Appendix B.6 (Table 7) and Appendix G.5 (Table 19).

AuthorsOpenReview ↗

Response to Reviewer UvDt (3)

W3 & Q3. Connection between Spatial Entropy and model-level representations.

Why an input-level spatial metric reflects attention-level changes. Self-attention relies on a global structure among tokens: each query focuses on a few salient keys (high CAC_A, Eq. 6), and the aggregated value vectors share common directions (high ΓA\Gamma_A, Eq. 7). Since each input patch is independently projected into a token in ViTs, SentS_{\mathrm{ent}} measures how evenly a transformation perturbs these tokens. A localized change (low SentS_{\mathrm{ent}}) leaves most tokens largely unaffected and preserves this structure, whereas a spread change (high SentS_{\mathrm{ent}}) perturbs most tokens and tends to lower CAC_A and ΓA\Gamma_A, as observed under corrupted inputs (Fig. 4), which themselves have high spatial entropy (mean Sent=0.913S_{\mathrm{ent}}=0.913; Fig. 7). High-SentS_{\mathrm{ent}} transformations therefore expose the quantizer to corruption-like attention behavior during calibration.

Empirical evidence. We measured CAC_A and ΓA\Gamma_A of FP ViT-S and DeiT-S on clean and Entropy-KK-transformed calibration inputs.

ModelCAC_A CleanCAC_A Entropy-KKCAC_A Δ\DeltaΓA\Gamma_A CleanΓA\Gamma_A Entropy-KKΓA\Gamma_A Δ\Delta
ViT-S0.051750.05025−2.90%0.139370.13181−5.42%
DeiT-S0.114090.09845−13.71%0.113430.10865−4.21%

Entropy-KK reduces both CAC_A and ΓA\Gamma_A on both models, consistent with the behaviors under corruptions (Fig. 4).

We further measured the feature-difference entropy HfeatH_{\mathrm{feat}}, i.e., SentS_{\mathrm{ent}} computed on token-wise activation differences ∥al,i(T(x))−al,i(x)∥22\lVert a_{l,i}(T(x))-a_{l,i}(x)\rVert_2^2 instead of input differences, where al,i(x)a_{l,i}(x) is the token of the FP activation in Eq. (4) at the ii-th patch. Using FP ViT-S, the nine original candidates, and 1,024 calibration images, we computed Spearman correlations across transformations.

SentS_{\mathrm{ent}} vs. HfeatH_{\mathrm{feat}}HfeatH_{\mathrm{feat}} vs. El(T(x))/El(x)E_l(T(x))/E_l(x)
Spearman ρ\rho+0.70−0.73

Higher SentS_{\mathrm{ent}} is associated with more spatially distributed activation changes, which are in turn associated with lower activation-energy ratios El(T(x))/El(x)E_l(T(x))/E_l(x). These results support SentS_{\mathrm{ent}} as a forward-free proxy for model-level behavior.

In the revised manuscript, we added a discussion of this issue in the first paragraph of Section 5, and included the corresponding experimental analysis in Appendix B.7 (Tables 8 and 9).

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

This paper studies the larger accuracy degradation that low-bit post-training quantized Vision Transformers experience on corrupted inputs relative to clean ImageNet inputs. It proposes a fixed-budget calibration strategy, Entropy-K, which ranks candidate image transformations by the spatial entropy of their transformation-induced difference energy and uses the top-K transformations on disjoint calibration subsets. The paper further analyzes activation-energy shifts under corruptions and transformations, relates them to attention concentration and value-vector alignment, and evaluates the approach across ImageNet/ImageNet-C, seven ViT-family architectures, two low-bit settings, and several reconstruction-based PTQ methods.

Critical strengths
  1. The problem is relevant and practically motivated. Figure 1 clearly establishes the central empirical phenomenon: relative performance preservation after quantization is substantially worse on ImageNet-C than on ImageNet.
  2. The experiments cover a broad scope. Table 1 includes seven architectures spanning ViT, DeiT, and Swin families, two aggressive precision settings, and three strong reconstruction baselines.
Critical weaknesses
  1. The paper's central claim is that S_ent can pick good transformations without running the model, but the evidence for it, Figure 5, contains only 9 points. More importantly, the 9 transformations used to validate S_ent are also the candidate pool from which the top-3 are selected, and they are also the transformations whose performance is compared in the final results. A score that ranks a set of familiar augmentation methods well does not necessarily work equally well on new transformations. Could the authors test whether S_ent’s ranking remains consistent with ImageNet-C gains on a set of new transformations independent of the current 9 (or on different strength parameters of the same transformations)? Appendix I extends the candidates to 13 on COCO, but only shows the selected ones.
  2. All classification experiments in the paper fix K=3, and the basis for choosing K=3 is Table 2(a), which compares K=1/3/5/9 on the ImageNet-C test results of ViT-S under W3A3. This amounts to tuning a hyperparameter on the test set. Could the authors provide a rule for choosing K that does not depend on the test set, or show that the conclusions remain stable under several predetermined values of K across other architectures and bit-width settings?
  3. The only comparison is Random-3 (randomly choosing 3), but the random pool itself contains the three good transformations PRIME, AugMix, and MixUp, which dilutes the comparison. There is no choosing the 3 transformations with the lowest S_ent. If S_ent is truly useful, choosing the lowest-scoring ones should be clearly worse; this is the most direct test.
  4. Table 8 splits ImageNet-C into four categories: noise, blur, weather, and digital. Entropy-K’s improvements are concentrated on Noise. The claimed “general corruption robustness” therefore looks more like "noise robustness". Could the authors explain why the gains are concentrated on noise? After removing noise corruptions, does Entropy-K still show a stable advantage over the baseline and over single transformations?
Questions for authors

Please refer to the weakness part.

5 comments
AuthorsOpenReview ↗

Response to Reviewer dxxm (1)

We thank the reviewer for the constructive and detailed feedback. We address each point below.

W1. Generalization to unseen transformation candidates.

To examine whether SentS_{\mathrm{ent}} generalizes beyond the original pool, we evaluated nine new transformations that do not overlap with the original candidates. Using SentS_{\mathrm{ent}} alone, without model inference or corruption-test results, Entropy-KK (with K=3K=3 unchanged) selects FACT, ColorJitter, and TokenMix (marked with †), which differ entirely from the original selection of PRIME, MixUp, and AugMix.

MethodW/AViT-S IDViT-S CDeiT-S IDDeiT-S CAvg.
Baseline3/364.5031.1769.1838.1150.74
AutoAugment63.8632.1669.0239.0851.03
ColorJitter†64.5131.7769.2838.2850.96
GridMask60.2926.6868.6737.1148.19
FACT†63.5632.7669.5739.9551.46
Hide-and-Seek62.1729.2968.2137.5949.32
Patch-Gaussian64.6131.3669.3538.9251.06
YOCO64.5431.0669.0638.2350.72
FMix63.9930.0670.2939.8851.06
TokenMix†63.6331.3470.2340.2051.35
Entropy-KK64.1533.6670.2740.5752.16
Baseline2/342.0215.4252.4522.8733.19
AutoAugment41.6416.5252.0123.2533.36
ColorJitter†42.3115.6953.7524.0733.96
GridMask36.5212.3050.4221.4930.18
FACT†40.3016.8154.6325.7034.36
Hide-and-Seek40.3314.2652.7323.1632.62
Patch-Gaussian41.9315.4253.3123.9933.66
YOCO43.5915.7352.7222.6633.68
FMix42.3115.2656.7225.8235.03
TokenMix†42.3616.0256.6426.6935.43
Entropy-KK43.4817.3356.3826.6035.95

Entropy-KK achieves the highest average accuracy under both W3A3 and W2A3 (52.1652.16 and 35.9535.95), outperforming the best individual transformation (FACT, 51.4651.46; TokenMix, 35.4335.43). Moreover, SentS_{\mathrm{ent}} is positively correlated with ImageNet-C accuracy gains across the new transformations (ρ=0.62\rho=0.62), indicating that the entropy ranking remains informative beyond the original candidates. This relationship is also preserved when the transformation strength is varied to match perturbation magnitude (Appendix B.3, Table 4).

In the revised manuscript, these results are included in Section 6 ("Generalization to unseen candidates") and Appendix G.5 (Table 18 and Figure 10).

AuthorsOpenReview ↗

Response to Reviewer dxxm (2)

W2. Selection of KK and sensitivity analysis.

We apologize for omitting this detail from the manuscript. KK was not tuned on the evaluation set; instead, we selected KK on the four extra corruptions (speckle noise, Gaussian blur, spatter, and saturate) that the ImageNet-C benchmark (Hendrycks & Dietterich, 2019) provides as an official validation set for hyperparameter selection, which are disjoint from the 15 standard corruptions used for final evaluation. On ViT-S under W3A3, K∈{1,3,5,9}K \in \lbrace 1, 3, 5, 9 \rbrace achieved 39.65%, 40.19%, 39.66%, and 39.41%, respectively, and the selected K=3K=3 was then applied to all other settings without further tuning.

We further examine the sensitivity to KK on the standard ImageNet-C across ViT-S, DeiT-S, and Swin-S under W3A3, W2A4, and W2A3.

MethodW/AViT-S IDViT-S CDeiT-S IDDeiT-S CSwin-S IDSwin-S CAvg.
Entropy-13/363.3433.4169.5441.4177.7246.2855.28
Entropy-364.6533.8269.9340.9478.2446.8555.74
Entropy-564.2733.6769.8741.2078.1146.1455.54
Entropy-964.2633.1669.9740.8778.1046.0355.40
Entropy-12/457.2129.4865.5737.0674.5542.0250.98
Entropy-359.0429.9466.2637.4774.9542.8051.74
Entropy-558.8229.7165.9337.4574.7442.6951.56
Entropy-958.8929.4465.7937.3874.9842.7851.54
Entropy-12/340.9217.8753.8927.3167.2632.7540.00
Entropy-342.4717.7854.7626.9868.3634.1540.75
Entropy-541.8117.2155.1127.3668.3633.9240.63
Entropy-942.5417.5454.6126.8568.6133.5140.61

K=3K=3 achieves the highest average accuracy under all three bit-widths, consistent with the choice made on the validation set. Notably, K=3K=3 also outperforms K=9K=9, which uses all candidates without selection, indicating that entropy-guided selection is beneficial beyond simply combining all transformations.

In the revised manuscript, we clarified how KK is selected in Section 5 and added sensitivity analysis results in Section 6 and Appendix G.2 (Table 15).

AuthorsOpenReview ↗

Response to Reviewer dxxm (3)

W3. Effectiveness of entropy-guided selection.

In addition to Random-3 (Appendix G.1), we compare Entropy-3 with two baselines under the same calibration budget and evaluation protocol: Bottom-3, the three lowest-SentS_{\mathrm{ent}} candidates (RandomErasing, RVT-PWA, and TrivialAugment), and Robust-3, three established robustness-oriented augmentations (AugMix, PRIME, and HybridAugment++) chosen without entropy-based ranking.

MethodW/AViT-S IDViT-S CDeiT-S IDDeiT-S CSwin-S IDSwin-S CAvg.
Baseline3/364.5031.1769.1838.1177.9144.7654.27
Robust-363.7833.1869.8040.3978.2046.3455.28
Bottom-364.1131.9569.1838.6477.8045.3154.50
Entropy-364.6533.8269.9340.9478.2446.8555.74
Baseline2/342.0215.4252.4522.8767.7331.3738.64
Robust-341.6117.3055.2427.4768.5033.3240.57
Bottom-342.5115.9952.8923.9167.4931.8439.11
Entropy-342.4717.7854.7626.9868.3634.1540.75

Entropy-3 achieves the highest average accuracy under both W3A3 and W2A3, outperforming Bottom-3 by +1.24+1.24 and +1.64+1.64 points and Robust-3 by +0.46+0.46 and +0.18+0.18 points, respectively. In particular, Bottom-3 yields only marginal gains over the baseline (+0.23+0.23 and +0.47+0.47), whereas Entropy-3 improves it by +1.47+1.47 and +2.11+2.11 points. These results show that selecting high-entropy transformations is more effective than selecting low-entropy ones, supporting the usefulness of SentS_{\mathrm{ent}} beyond random selection.

In the revised manuscript, we included these comparison results in Section 6 (Table 2(a)) and Appendix G.3 (Table 16).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.