Ranked Entropy Minimization for Continual Test-Time Adaptation

Jaemin Na; Jisu Han; Wonjun Hwang

REVIEW 4 major objections 5 minor 1 cited by

Ranked Entropy Minimization for Continual Test-Time Adaptation

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Ranked entropy minimization keeps continual test-time adaptation from collapsing into a single class, and on ImageNet-C it beats the previous state of the art by 3.3 error points.

desk verdict A legitimate empirical CTTA contribution with a real efficiency win; the unproven monotonicity premise is the main risk, but it is addressable and does not sink the paper. read the letter →

arxiv 2505.16441 v1 pith:CJICLYCW submitted 2025-05-22 cs.CV cs.LG

Jisu Han , Jaemin Na , Wonjun Hwang This is my paper

classification cs.CVcs.LG

keywords continualtest-timeadaptationentropyminimizationmodelcollapseattentionmaskingconsistencyregularizationrankinglossvisiontransformerdistributionshift

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that the instability of test-time entropy minimization—model collapse to a single class—can be fixed by imposing an explicit rank order on prediction difficulty. It proposes Ranked Entropy Minimization (REM), which progressively masks the most object-relevant patches of a ViT image using attention scores, so each image generates a ladder of progressively harder views. Two losses use that ladder: a masked consistency loss pulls harder views toward the original prediction, and an entropy ranking loss keeps the entropy order with a margin. The paper reports that one model updating only normalization layers, with three forward passes and 0.03M trainable parameters, reaches mean errors of 39.2 on ImageNet-C, 9.4 on CIFAR10-C, and 23.4 on CIFAR100-C. That is a 16.6-point gain over the source model and a 3.3-point gain over the previous state of the art, leaving a 3.5-point gap to the supervised upper bound.

What carries the argument

The central object is the explicit mask chain. For each image, the attention scores $A=\sum_{h=1}^H \mathrm{Softmax}(Q_{h,\mathrm{cls}}K_{h,\mathrm{img}}^{\top}/\sqrt{d})$ over the ViT's image tokens are sorted descending, and the top-$m$ proportion of tokens is masked for a set of increasing ratios, usually $\{0, 5\%, 10\%, 15\%\}$. That creates a ladder of prediction difficulty within a single image. The two losses that carry the argument are the masked consistency loss $L_{\mathrm{MCL}}=\sum_{i<j} H(f_t(x_j), \mathrm{sg}(f_t(x_i)))$, which makes harder views agree with easier views, and the entropy ranking loss $L_{\mathrm{ERL}}=\sum_{i<j} \max(0, S(f_t(x_i))-\mathrm{sg}(S(f_t(x_j)))+m)$, which preserves the order of entropies with margin $m$. The stop-gradient operation keeps the original prediction as a fixed target so the model cannot collapse to the masked view's trivial solution, and Figure 3's empirical monotonicity of entropy and error with mask ratio is what licenses both losses.

What would settle it

Run the Figure 3 protocol at severity 5 on all 15 ImageNet-C corruptions and check whether the entropy order $S(f_t(x_{m_i})) \le S(f_t(x_{m_j}))$ for $m_i < m_j$ holds per sample; a corruption where most images violate the order, or a single image class where masking the foreground makes the prediction more confident, would invert the entropy ranking loss and could drive the model toward the wrong target.

Watch

Extended reading notes

Core claim

The paper's central claim is that model collapse in continual test-time adaptation is not an unavoidable price of entropy minimization but a consequence of minimizing entropy with no ordering information. REM constructs, for each test image, a chain of masked versions with increasing mask ratios, using the ViT's class-token attention to erase the most attended (foreground) patches first. Since harder views should be less accurate and more uncertain, the masked views form a ranked difficulty structure, and the paper treats this ranking as the training signal. The masked consistency loss aligns each harder view to the original prediction while the entropy ranking loss preserves the order of entropies across the chain, with stop-gradient keeping the targets fixed. The paper's experiments, including online label-shift, batch-size-1, mixed-shift, CLIP, CNN, and non-ViT transformer settings, are offered as evidence that this ranked objective gives the stability of teacher-student consistency methods at the computational cost of plain entropy minimization.

Load-bearing premise

The method rests on the premise that masking the most-attended image patches always raises prediction entropy and lowers accuracy, for every corruption type and at every adaptation stage, so the masked views really are ordered by difficulty.

Editorial extensions

If this is right

Because REM updates only normalization layers and runs three forward passes, the reported ImageNet-C error of 39.2 comes with roughly a third of the wall-clock time of Continual-MAE and 0.03% of its trainable parameters.
The method transfers beyond the CTTA protocol: it gains 3.9 points over DeYO under online label shift, 8.3 points at batch size 1, and 3.0 points on mixed corruption shifts at severity 5.
REM works as a plug-in on top of CLIP and on top of weight-averaged adaptation, and it extends to CNNs when attention maps are replaced by feature-activation or Grad-CAM maps.
The calibration analysis reports ECE 8.7% versus 12.6% for Tent and 14.6% for ViDA at similar error levels, indicating the ranked objective reduces overconfidence rather than just fitting a collapse.

Reading between the lines

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

The authors' failure-case analysis shows that easy corruptions such as brightness produce small output discrepancies between masked and unmasked views, which suggests a testable extension: scale each loss term by a per-domain difficulty estimate so easy domains do not adapt too slowly.
If the monotonicity premise holds, the same ranked-ladder idea could replace augmentation policies in other self-supervised or semi-supervised objectives wherever a cheap difficulty axis exists, since the paper's mechanism needs only a scalar ordering, not class labels.
A strict reading of the Section 5 caveat implies that any deployment corruption where foreground masking makes predictions more confident would invert the ranking loss and actively push the model the wrong way, so the practical range of REM is bounded by how often ImageNet-C's monotonicity extends to other corruptions.
The paper leaves unproven whether the ranking is monotone for every image or only in aggregate; an editor-level prediction is that per-sample violations concentrate on images with already-ambiguous foregrounds, which is where a confidence-gated mask ratio would be a natural safeguard.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Quick take: this is a legitimate empirical contribution to CTTA, and the efficiency story is the strongest part. The ranked-masking idea is new as far as I know: build a chain of increasingly masked views of the same image, align the high-mask predictions to the low-mask ones with stop-gradient, and add a loss that keeps the entropy ranking monotone. The authors show consistent gains over Continual-MAE on ImageNet-C, CIFAR10-C, and CIFAR100-C while updating only normalization layers of a single ViT-B/16 with three forward passes and 0.03M trainable parameters. The ablations and the forward/backward transfer analysis are well designed, and the code is available. I believe the empirical claims are for real.

The soft spots, in order of importance. First, the monotonicity premise—that masking top-attention patches monotonically increases entropy and error—is genuinely load-bearing. It is what makes Eq. 3 and Eq. 4 point in the right direction, and the paper's own Section 5 admits it lacks theoretical support and that counterexamples exist. Figure 3 only shows aggregate trends. Because the losses act on every sample at every online step, per-image violations could actively hurt. The authors need to provide a per-corruption, per-stage violation analysis, or at least a robust variant (e.g., a slack or a soft ranking that tolerates a small fraction of violations). This is the main thing I would ask for in revision.

Second, the main Tables 1-3 have no error bars. Some margins over Continual-MAE are substantial (3.3% on ImageNet-C), but per-domain numbers on CIFAR100-C are within a few points, so variance matters. Third, COME is cited in related work but not included in the main comparison; the appendix only compares to ROID/CMT under non-default settings, and those methods are mentioned without proper citations in the reference list. Authors should add COME to the main table and fix the citation gaps. Fourth, the hyperparameters (λ, margin, mask ratios) are tuned on the same ImageNet-C benchmark, which is a mild circularity—they do show robustness, so I don't consider it serious.

Overall, the paper is a solid, within-subfield advance, not a paradigm shift. It deserves a serious referee. I would send it to review with a request for the per-corruption monotonicity test, error bars, and fuller baseline comparison. The likely outcome is a strong paper after those additions.

Referee Report

4 major / 5 minor

Summary. The paper proposes Ranked Entropy Minimization (REM) for continual test-time adaptation (CTTA). The method uses the self-attention structure of a ViT to progressively mask high-attention patches, producing a chain of masked views of increasing difficulty. Two losses are introduced: a masked consistency loss (MCL, Eq. 3) that aligns higher-mask predictions to lower-mask targets, and an entropy ranking loss (ERL, Eq. 4) that enforces lower entropy for lower-mask predictions. The method updates only normalization layers of a single model with three forward passes. Experiments on ImageNet-to-ImageNetC, CIFAR10-to-CIFAR10C, and CIFAR100-to-CIFAR100C report mean errors of 39.2, 9.4, and 23.4, respectively, claimed to surpass Continual-MAE by 3.3% on ImageNetC and to close much of the stability-efficiency gap in CTTA. The paper also includes ablations, efficiency comparisons, online TTA experiments, CLIP adaptation, time-constrained scenarios, and an explicit limitation section.

Significance. If the empirical claims hold, REM makes a meaningful contribution to CTTA by showing that a single-model, normalization-only update scheme with a masked consistency/ranking objective can match or beat heavier teacher-student and parameter-restoration methods, while being substantially more efficient. The paper is honest about its main assumption, releases code, and provides a broad evaluation suite spanning CTTA, online TTA, CLIP-based adaptation, and efficiency benchmarking. The proposed loss pair is simple and plausible, and the model-collapse motivation (Figure 2) is well taken. However, the central regularity assumption on which the losses depend is self-admittedly unproven, and the main SOTA comparisons are reported without statistical uncertainty and without a recently published, directly relevant baseline (COME). These issues currently prevent the significance from being fully established.

major comments (4)

[§3.2, Fig. 3, Eqs. (3)-(4), §5]
[Tables 1-3, §4.2]
[§2.1, Tables 1-3]
[§4.1, Appendix J]

minor comments (5)

[§4.5]
[§3.2 / Notation]
[Figures 9 and 10]
[Appendix H, Table 16]
[§4.2 / Table 1]

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported CTTA gains are empirical and not reducible to the loss definitions or to fitted parameters.

full rationale

The paper does not reduce its reported accuracies to any fitted input or to a self-citation. The two loss terms, Eq. (3) masked consistency loss and Eq. (4) entropy ranking loss, encode the masking-difficulty premise as a training objective, but the benchmark errors in Tables 1-3 are measured after online adaptation and are not obtainable from the loss equations alone. The monotonicity premise is load-bearing: Section 3.2 uses Figure 3 to claim that masking objects increases entropy and error with mask ratio, and both losses assume this rank structure. However, the paper enters this premise as an empirical regularity, not as a definition, and Section 5 explicitly flags it as unproven: it is 'not yet fully supported by rigorous theoretical proof' and 'counterexamples arising from the diversity of images still pose significant challenges.' Appendix I's TVD analysis further shows per-domain heterogeneity (e.g., Gaussian 5.54 vs. Brightness 1.69 on CIFAR100C), which is a correctness or robustness risk for the method, not a circularity in the derivation. Hyperparameters such as lambda, margin, and mask ratios are selected using the same ImageNetC benchmark (Appendix J), which is a benchmark-selection concern, but those selected values are not the predicted errors, and no equation turns them into the final accuracy. There is no load-bearing self-citation: the masking procedure cites external prior work (Bolya et al.; Son et al.) and the ranking-loss motivation cites external calibration work (Moon et al.; Noh et al.). No equation, fitted parameter, or cited result is equivalent by construction to the reported performance, so the derivation chain is self-contained and the score is 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central empirical result depends on three domain assumptions: attention-guided masking creates a valid difficulty ranking, this ranking is monotone in entropy and error, and normalization-layer updates suffice. No free parameter is fitted to the benchmark error rates, but hyperparameters are tuned on the same ImageNet-C benchmark, which slightly raises the circularity burden.

free parameters (3)

lambda (Eq. 5) = 1.0
Weight balancing MCL and ERL; selected by hyperparameter grid search in Appendix J.
margin m (Eq. 4) = 0
Set to zero after sensitivity analysis; ERL becomes max(0, S_i - S_j).
mask chain MN = {0, 5%, 10%}
Adopted from Table 18 as an accuracy/compute trade-off; note that the best reported N=3 +5% configuration was not used.

assumptions (3)

domain assumption Masking the top attention patches monotonically increases prediction entropy and error for every corruption and at every adaptation stage.
Load-bearing premise for the ranking loss; empirically supported by Figure 3 but self-admittedly without proof in Section 5.
domain assumption ViT class-token attention scores identify object or foreground locations containing domain-invariant features.
Used to build the mask chain in Eq. 2; if attention tracks noise or background, the difficulty ladder is invalid.
domain assumption Updating only normalization layers of a single model is sufficient to adapt while avoiding catastrophic forgetting.
The method inherits the Tent/EM training scheme; the stability and efficiency claims depend on this assumption.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Ranked Entropy Minimization for Continual Test-Time Adaptation." pith.science (2026). https://pith.science/paper/CJICLYCW

@misc{pith2026250516441,
  author       = {Pith},
  title        = {Pith review of: Ranked Entropy Minimization for Continual Test-Time Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CJICLYCW}},
  note         = {Machine review of arXiv:2505.16441}
}

read the original abstract

Test-time adaptation aims to adapt to realistic environments in an online manner by learning during test time. Entropy minimization has emerged as a principal strategy for test-time adaptation due to its efficiency and adaptability. Nevertheless, it remains underexplored in continual test-time adaptation, where stability is more important. We observe that the entropy minimization method often suffers from model collapse, where the model converges to predicting a single class for all images due to a trivial solution. We propose ranked entropy minimization to mitigate the stability problem of the entropy minimization method and extend its applicability to continuous scenarios. Our approach explicitly structures the prediction difficulty through a progressive masking strategy. Specifically, it gradually aligns the model's probability distributions across different levels of prediction difficulty while preserving the rank order of entropy. The proposed method is extensively evaluated across various benchmarks, demonstrating its effectiveness through empirical results. Our code is available at https://github.com/pilsHan/rem

Figures

Figures reproduced from arXiv: 2505.16441 by the authors.

**Figure 1.** Our Intuition. We explicitly raise the prediction difficulty of the input images through the masking strategy. Based on the intuition that increased difficulty decreases prediction accuracy and increases entropy, we attempt to maintain a rank ordering of entropy while improving consistency from original to masked predictions. Our approach addresses the problem of model collapse in entropy minimization methods in a… view at source ↗

**Figure 2.** Observation on model collapse in the entropy minimization approach. (a) Under the CTTA scenario, the EM approach (Tent) undergoes significant performance degradation at a critical point (adaptation order T3, Impulse noise). (b) This phenomenon occurs because the model learns constant representations that do not depend on input images, leading to a collapse in prediction diversity. This is evidenced by class probabil… view at source ↗

**Figure 3.** Empirical study according to masking ratio. We report the changes in error and entropy as the masking ratio increases. Both entropy and error exhibit a monotone increasing trend with respect to the masking ratio, and we observe that linearity becomes more pronounced, especially in regions with lower masking ratios. masks content containing domain-invariant information for augmentation. To this end, we focus on the s… view at source ↗

Figures from the paper (8 more)

**Figure 4.** Figure 4 [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Adaptability analysis on ImageNetC under Gaussian noise corruption [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 7.** Figure 7: Masked image visualization. We compare the predictions and entropy of REM, Tent, and Source and visualize the results of our masking strategy. Each column represents images with masking ratios of 0, 10%, and 20% for each method, while each row shows the true label on t…

**Figure 8.** Figure 8: Grad CAM visualization. We compare attention maps to identify the pixels contributing to predictions. high learning rates achieves the best performance of 42.1% on Seen domains, while slower adaptation yields the best performance of 41.4% on Unseen domains. Such a tren…

**Figure 9.** Figure 9: Effect of each component. We present the results of REM and compare them with variations where MCL and ERL are removed [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]

**Figure 10.** Figure 10: Effect of masking strategy. We present the results of REM (foreground masking) in comparison with different masking strategies, including background masking, and random masking. model’s predictions and to gain insights into its decisionmaking process. Each row corres…

**Figure 11.** Figure 11: Conceptual illustration comparing CTTA frameworks. (a) Entropy minimization approach updates only the normalization layer for the target domain while utilizing a single model. (b) Consistency regularization method employs both a student model, which updates all parame…

**Figure 12.** Figure 12: Hyperparameter sensitivity analysis 16 [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

Multi-Cache Enhanced Prototype Learning for Test-Time Generalization of Vision-Language Models
cs.CV 2025-08 unverdicted novelty 5.0 of 10

The submitted full text does not match the abstract, so the manuscript cannot be assessed as a coherent preprint.

Reference graph

Works this paper leans on

12 extracted references · 11 canonical work pages · cited by 1 Pith paper

[1]

Batch Size 1 Gaussianshot impulsedefocusglass motionzoomsnowfrost fog brightnesscontrastelastic trans pixelatejpegMean↑ Gain Source (Dosovitskiy et al., 2021)9.5 6.8 8.2 29.0 23.5 33.9 27.1 15.9 26.5 47.2 54.7 44.1 30.5 44.5 47.8 29.9 0.0 MEMO (Zhang et al.,

work page 2021
[6]

Experiments on Online Test-Time Adaptation Scenario In addition to CTTA scenarios, our method is readily applicable to a wide range of TTA scenarios

63.8 75.4 49.9 63.0 REM (Ours) 64.3 75.2 49.7 63.1 C. Experiments on Online Test-Time Adaptation Scenario In addition to CTTA scenarios, our method is readily applicable to a wide range of TTA scenarios. To evaluate its effectiveness in a more challenging setting, we compare our approach against EM-based state-of-the-art methods, including MEMO (Zhang et ...

work page 2022
[7]

Classification accuracy (%) on ImageNetC (severity level5) under online imbalanced label shifts (imbalance ratio = ∞). Label Shifts Gaussianshot impulsedefocusglass motionzoomsnowfrost fog brightnesscontrastelastic trans pixelatejpegMean↑ Gain Source (Dosovitskiy et al., 2021)9.4 6.7 8.3 29.1 23.4 34.0 27.0 15.8 26.3 47.4 54.7 43.9 30.5 44.5 47.6 29.9 0.0...

work page 2021
[8]

Similar to the label shift scenario, our method achieves the best performance across all domains except for elastic transform, resulting in an 8.3% performance improvement

Table 8 shows the results for TTA under a batch size of 1, demonstrating the robustness of our method in scenarios where batch statistics cannot be effectively leveraged. Similar to the label shift scenario, our method achieves the best performance across all domains except for elastic transform, resulting in an 8.3% performance improvement. Mixed Distrib...

work page 2021
[11]

Following W ATT, we report the TTA results from the CLIP model to CIFAR (Krizhevsky et al.,

We compare the performance of our method against CLIP (Radford et al., 2021), Tent (Wang et al., 2021), TPT (Shu et al., 2022), CLIPArTT (Hakim et al., 2024), and W ATT (Osowiechi et al., 2024). Following W ATT, we report the TTA results from the CLIP model to CIFAR (Krizhevsky et al.,

work page 2021
[14]

Calibration Error Analysis We investigate the mitigation of model collapse by analyzing the issue of overconfidence through model calibration error

Mean error rate (%) on ImageNetC using CNN architectures Model CoTTA EATA EcoTTA BECoTTA REM (FA) REM (Grad-CAM) WideResNet-28 16.2 18.6 16.8 - 16.9 16.5 WideResNet-40 - 37.1 36.4 35.5 34.5 34.6 G. Calibration Error Analysis We investigate the mitigation of model collapse by analyzing the issue of overconfidence through model calibration error. We observe...

work page 2015
[15]

ImageNet-C Source Tent SAR ViDA REM ECE (%) ↓ 5.3 12.6 10.3 14.6 8.7 Error (%) ↓ 55.8 51.0 45.2 43.4 39.2 H

Comparison of ECE and error rates on ImageNetC. ImageNet-C Source Tent SAR ViDA REM ECE (%) ↓ 5.3 12.6 10.3 14.6 8.7 Error (%) ↓ 55.8 51.0 45.2 43.4 39.2 H. Comparison with Augmentation-based EM approaches Recent augmentation-based EM methods (Marsden et al., 2024; Lee & Chang,

work page
[2009]

and various domain adaptation and generalization benchmarks, including VisDA (Peng et al., 2018), Office- Home (Venkateswara et al., 2017), PACS (Li et al., 2017), and VLCS (Fang et al., 2013). As a result, our method achieves competitive performance compared to the previous state-of-the-art W ATT, without requiring additional inner-loop training processe...

work page 2018

Show all 12 references

[2014]

Hakim, G. A. V ., Osowiechi, D., Noori, M., Cheragha- likhani, M., Bahri, A., Yazdanpanah, M., Ayed, I. B., and Desrosiers, C. Clipartt: Light-weight adaptation of clip to new domains at test time. arXiv preprint arXiv:2405.00754,

arXiv
[2021]

Zeno’s paradoxes

Huggett, N. Zeno’s paradoxes. In Zalta, E. N. (ed.), The Stanford Encyclopedia of Philosophy . Meta- physics Research Lab, Stanford University, 2018 edi- tion,

2018
[2022]

We compare EATA (Niu et al.,

under the time-constrained protocol (Alfarra et al., 2024). We compare EATA (Niu et al.,

2024
[2024]

Visda: A synthetic-to-real benchmark for visual domain adaptation

Peng, X., Usman, B., Kaushik, N., Wang, D., Hoffman, J., and Saenko, K. Visda: A synthetic-to-real benchmark for visual domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pp. 2021–2026,

2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Batch Size 1 Gaussianshot impulsedefocusglass motionzoomsnowfrost fog brightnesscontrastelastic trans pixelatejpegMean↑ Gain Source (Dosovitskiy et al., 2021)9.5 6.8 8.2 29.0 23.5 33.9 27.1 15.9 26.5 47.2 54.7 44.1 30.5 44.5 47.8 29.9 0.0 MEMO (Zhang et al.,

work page 2021

[2] [6]

Experiments on Online Test-Time Adaptation Scenario In addition to CTTA scenarios, our method is readily applicable to a wide range of TTA scenarios

63.8 75.4 49.9 63.0 REM (Ours) 64.3 75.2 49.7 63.1 C. Experiments on Online Test-Time Adaptation Scenario In addition to CTTA scenarios, our method is readily applicable to a wide range of TTA scenarios. To evaluate its effectiveness in a more challenging setting, we compare our approach against EM-based state-of-the-art methods, including MEMO (Zhang et ...

work page 2022

[3] [7]

Classification accuracy (%) on ImageNetC (severity level5) under online imbalanced label shifts (imbalance ratio = ∞). Label Shifts Gaussianshot impulsedefocusglass motionzoomsnowfrost fog brightnesscontrastelastic trans pixelatejpegMean↑ Gain Source (Dosovitskiy et al., 2021)9.4 6.7 8.3 29.1 23.4 34.0 27.0 15.8 26.3 47.4 54.7 43.9 30.5 44.5 47.6 29.9 0.0...

work page 2021

[4] [8]

Similar to the label shift scenario, our method achieves the best performance across all domains except for elastic transform, resulting in an 8.3% performance improvement

Table 8 shows the results for TTA under a batch size of 1, demonstrating the robustness of our method in scenarios where batch statistics cannot be effectively leveraged. Similar to the label shift scenario, our method achieves the best performance across all domains except for elastic transform, resulting in an 8.3% performance improvement. Mixed Distrib...

work page 2021

[5] [11]

Following W ATT, we report the TTA results from the CLIP model to CIFAR (Krizhevsky et al.,

We compare the performance of our method against CLIP (Radford et al., 2021), Tent (Wang et al., 2021), TPT (Shu et al., 2022), CLIPArTT (Hakim et al., 2024), and W ATT (Osowiechi et al., 2024). Following W ATT, we report the TTA results from the CLIP model to CIFAR (Krizhevsky et al.,

work page 2021

[6] [14]

Calibration Error Analysis We investigate the mitigation of model collapse by analyzing the issue of overconfidence through model calibration error

Mean error rate (%) on ImageNetC using CNN architectures Model CoTTA EATA EcoTTA BECoTTA REM (FA) REM (Grad-CAM) WideResNet-28 16.2 18.6 16.8 - 16.9 16.5 WideResNet-40 - 37.1 36.4 35.5 34.5 34.6 G. Calibration Error Analysis We investigate the mitigation of model collapse by analyzing the issue of overconfidence through model calibration error. We observe...

work page 2015

[7] [15]

ImageNet-C Source Tent SAR ViDA REM ECE (%) ↓ 5.3 12.6 10.3 14.6 8.7 Error (%) ↓ 55.8 51.0 45.2 43.4 39.2 H

Comparison of ECE and error rates on ImageNetC. ImageNet-C Source Tent SAR ViDA REM ECE (%) ↓ 5.3 12.6 10.3 14.6 8.7 Error (%) ↓ 55.8 51.0 45.2 43.4 39.2 H. Comparison with Augmentation-based EM approaches Recent augmentation-based EM methods (Marsden et al., 2024; Lee & Chang,

work page

[8] [2009]

and various domain adaptation and generalization benchmarks, including VisDA (Peng et al., 2018), Office- Home (Venkateswara et al., 2017), PACS (Li et al., 2017), and VLCS (Fang et al., 2013). As a result, our method achieves competitive performance compared to the previous state-of-the-art W ATT, without requiring additional inner-loop training processe...

work page 2018