Beyond Entropy: Region Confidence Proxy for Wild Test-Time Adaptation

Ling-Yu Duan; Shixiang Tang; Xiaotong Li; Yichun Hu; Zixuan Hu

REVIEW 3 major objections 5 minor 3 cited by

Beyond Entropy: Region Confidence Proxy for Wild Test-Time Adaptation

T0 review · 3 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 This paper claims that replacing entropy minimization with a region-confidence proxy — two closed-form upper bounds derived from a Gaussian neighborhood — stabilizes and speeds up wild test-time adaptation.

desk verdict ReCAP is an empirically promising WTTA method whose central upper-bound theorem has a real hole; the empirical results may stand, but the paper's advertised guarantee does not. read the letter →

arxiv 2505.20704 v2 pith:CE7VRHBE submitted 2025-05-27 cs.CV

Zixuan Hu , Yichun Hu , Xiaotong Li , Shixiang Tang , Ling-Yu Duan This is my paper · ORCID

classification cs.CV

keywords wildtest-timeadaptationregionconfidenceentropyminimizationfinite-to-infiniteapproximationGaussianmodelinglocalpredictionconsistencyclosed-formupperbound

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

Test-time adaptation typically fine-tunes a model by minimizing prediction entropy on incoming test samples. This paper argues that under wild conditions — batches of one sample, mixed corruptions, and drifting label distributions — that objective creates conflicting update directions, because neighboring features often get inconsistent predictions. The paper proposes to optimize region confidence instead: the entropy averaged over a probabilistic neighborhood of each feature plus a penalty on prediction spread inside that neighborhood. Since direct computation is intractable, the authors derive two closed-form upper bounds, called Regional Entropy and Regional Instability, via a finite-to-infinite asymptotic approximation, so the whole region objective costs about the same as a plain entropy update. Experiments on ImageNet-C, ImageNet-R, and VisDA-2021 with both convolutional and vision-transformer backbones report consistent accuracy gains over entropy-based selection methods.

What carries the argument

The load-bearing object is the finite-to-infinite asymptotic approximation: a sampled average of a softmax probability over a region is first re-expressed, as the number of samples grows, as the ratio of two Gaussian expectation values, and the log of the softmax is bounded via Jensen's inequality using the Gaussian moment-generating function. This yields two closed-form surrogates — Regional Entropy ($L_{\mathrm{RE}}$) for the expected entropy over the region and Regional Instability ($L_{\mathrm{RI}}$) for the expected KL divergence from the region center — that together replace the region-confidence integral in the training loss. Because both are simple sums of exponentials of the logits, they simulate an infinite neighborhood of samples at negligible computational cost, which is what makes region-confidence optimization feasible for real-time test-time adaptation.

What would settle it

Take any trained classifier, fix a feature and its Gaussian neighborhood, draw many samples from that neighborhood, and compare the Monte Carlo estimate of the expected entropy with the closed-form Regional Entropy; one case where the formula is smaller than the estimate disproves the upper-bound claim.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the optimization dynamics, not just sample selection, are what limits wild test-time adaptation: entropy minimization driven by single samples amplifies local prediction disagreements and slows adaptation. Region confidence — defined as the integral of entropy over a Gaussian region in feature space plus a KL-divergence variance term — corrects both the bias and the dispersion of the update directions. The paper's main theoretical contribution is showing that this intractable region objective can be replaced by an upper-bounded proxy: applying Jensen's inequality and the moment-generating function of a Gaussian to the softmax logits turns the infinite region integral into the closed-form sums $L_{\mathrm{RE}}$ and $L_{\mathrm{RI}}$, which are computable with no sampling and no extra forward passes. Optimizing these bounds is the ReCAP training framework, and the paper reports that it outperforms existing wild-TTA methods and further improves selection-based methods when combined with them.

Load-bearing premise

The method stands on the claim that the average of the model's class probabilities over a local region can be replaced by a ratio of two Gaussian expectation values; if that replacement is wrong, the cheap Regional Entropy is not an upper bound on the true region confidence.

Editorial extensions

If this is right

Switching the training loss from per-sample entropy to the region-confidence proxy improves average accuracy on ImageNet-C across all three wild scenarios, for both ResNet and ViT backbones.
Because the proxy costs only about five percent more than plain entropy minimization, region-based optimization remains fast enough for online deployment on a single GPU.
The regional objective is orthogonal to sample-selection strategies: combining ReCAP with the entropy-based selectors SAR and DeYO yields additional gains, so the two lines of work can be stacked.
Region confidence can serve simultaneously as a reliability criterion and as the optimization loss, unifying sample selection and training in one framework.

Reading between the lines

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

We read the paper as implying a general recipe: any loss that integrates a function of logits over a feature-space neighborhood can be approximated by the same Gaussian-moment trick, so the approach could transfer to consistency regularization in semi-supervised learning or to adaptation losses beyond classification.
A testable extension would isolate the contribution of the variance term by sweeping its weight at higher resolution; the reported ablation already suggests both terms matter, but quantifying the trade-off would show practitioners where the gain actually comes from.
The approximation's validity hinges on the equality between a neighborhood-average softmax and a ratio of Gaussian expectations; a Monte-Carlo check of that identity across corruption types would reveal whether the upper-bound story or the loss's smoothing effect drives the empirical gains.

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

Dear Colleague,

First thing you should know: this paper proposes a new objective for wild test-time adaptation called region confidence, with a closed-form proxy ReCAP that replaces entropy minimization. The empirical story is real and worth taking seriously: on ImageNet-C, ImageNet-R, and VisDA-2021, ReCAP beats DeYO and other selection-based baselines by about +2 points, and the runtime is within 5% of plain Tent. The idea of modeling local regions as Gaussians and deriving a cheap proxy is genuinely new in this literature, and the ablations show both the bias and variance terms contribute. I have no quarrel with the empirical core.

But the theory does not hold up. Proposition 4.3's proof, in Appendix A.2, relies on Eq. 25, which asserts that the limiting average of softmax probabilities over a Gaussian region equals the ratio of Gaussian expectations of the numerator and denominator. That equality is false in general. The limit of (1/N)Σ p_j is E[p_j], not E[e^{a_j·z}] / E[Σ e^{a_j·z}]. The epsilon-limit argument that follows does not fix it; it just rewrites the same identification. The same pattern is used in the variance-term bound (Prop. 4.4). So the "upper-bounded proxy" advertised in the abstract is not established. This is a load-bearing issue, because the paper's novelty claim rests on the proxy being principled, not just another tuned loss.

The rest is softer. Hyperparameters λ, τ, τ_RE, L0 are tuned per benchmark, and the main tables report only means, not error bars. The code link is promised but not released. These are minor and fixable. The larger question is whether the authors can repair the derivation. If they can show the bound holds under plausible conditions, or if they reframe ReCAP as a heuristic approximation and lean on the empirical evidence, the contribution still has value.

Who gains from this? TTA researchers, especially those working on wild and online settings. I'd bring it to a reading group if we want to discuss the gap between heuristic design and formal guarantees. I would not cite the theory as it stands, but I might cite the empirical method with a caveat.

Recommendation: send it to peer review. The empirical results deserve scrutiny, and the theory is fixable or should be softened. A serious referee should flag Appendix A.2 as the main issue and ask for either a corrected proof or an explicit downgrade of the claim.

Referee Report

3 major / 5 minor

Summary. The paper proposes ReCAP, a new training objective for Wild Test-Time Adaptation (WTTA). ReCAP replaces entropy minimization with a region-confidence loss defined over a Gaussian-modeled local region in feature space, consisting of a bias term (expected entropy over the region) and a variance term (expected KL divergence from the region center). Because these expectations are intractable, the paper derives closed-form proxies L_RE and L_RI, claims they are upper bounds, and uses L_RE for sample selection and weighting in the overall loss. Experiments are reported on ImageNet-C, ImageNet-R, and VisDA-2021 under three wild scenarios (batch size 1, mixed domains, imbalanced label shift) with ResNet and ViT backbones, showing consistent gains over Tent, EATA, SAR, and DeYO, as well as improved runtime over selection-based baselines.

Significance. If the upper-bound results were correct, ReCAP would be a valuable contribution: a cheap, selection-compatible surrogate for region confidence with theoretical grounding. The empirical core is a genuine strength: the method is evaluated across three benchmarks, two architectures, three wild scenarios, and is shown to integrate with existing selection methods (SAR, DeYO), with a runtime table demonstrating only modest overhead. The finite-sampling inequality (Lemma A.1) and the negative-log-likelihood bound (Lemma A.2) are proved correctly, and the variance-term bound in Proposition 4.4 appears valid. However, the central theoretical claim of the paper, Proposition 4.3, is false: the asserted upper bound on the expected entropy is violated by a simple counterexample. Consequently, the advertised 'tractable and upper-bounded proxy' is not established, and the paper's current framing overstates its theoretical foundation. The method may still work as an empirical heuristic, but the claimed principled motivation needs substantial revision.

major comments (3)

[Appendix A.2, Eq. (25)] The proof of Proposition 4.3 identifies the limiting average of softmax probabilities, (1/N)Σ_k pθ(z_k)_j, with the ratio of Gaussian expectations E[e^{a_j·z̃+b_j}] / E[Σ_k e^{a_k·z̃+b_k}]. This is not an identity: the expectation of a ratio does not equal the ratio of expectations, and the limit of the sample average is E[pθ(z̃)_j], not the ratio defined in Eq. (25). The epsilon-limit argument in Eqs. (26) through (30) does not repair this, because it substitutes the ratio for E[pθ(z̃)_j] without justification. Lemma A.1 yields only E[L_ent] ≤ -Σ_j E[pθ(z̃)_j] E[log pθ(z̃)_j], and the replacement of E[pθ(z̃)_j] by the ratio of expectations is invalid.
[Proposition 4.3 / Eq. (7)] The asserted inequality EΩ[L_ent] ≤ L_RE is actually false, not merely unproven. For C=2, a1=0, a2=1, b=0, z=3, and Σ=1, the right-hand side evaluates to L_RE ≈ 0.180, while E_{z̃∼N(3,1)}[L_ent(pθ(z̃))] ≈ 0.223. This violates Proposition 4.3 under its stated hypotheses. The same issue propagates to the claim that L_RE is an upper bound on the bias term of region confidence and to the interpretation of the loss in Eq. (9) as minimizing an upper bound.
[Abstract and Section 4.2] Because Proposition 4.3 is false, the paper's central theoretical claim—that ReCAP provides a 'finite-to-infinite asymptotic approximation that transforms the intractable region confidence into a tractable and upper-bounded proxy'—is unsupported. The authors should either supply a valid bound with a correct proof, or explicitly present L_RE as a heuristic approximation and remove the upper-bound language from the abstract, contributions, and method description. The empirical evaluations may still support ReCAP as an effective heuristic, but the current manuscript overstates its theoretical foundation.

minor comments (5)

[Eqs. (7) and (8)] The symbol 'P' appears in the quadratic terms (e.g., '1/2 a_j P a_j^T' and '(a_i-a_j) P (a_i-a_j)^T') where the covariance 'Σ' is clearly intended; this should be corrected.
[Appendix A.2, Eq. (29)] The expression 'e^{(a_i-a_j)·z(b_i-b_j)}' is missing a plus sign; it should read 'e^{(a_i-a_j)·z+(b_i-b_j)}'.
[Appendix A.2, Eq. (25)] The notation pθ(z)_i is redefined in Eq. (25) to mean a ratio of Gaussian expectations, which conflicts with its earlier definition as the softmax probability in Eq. (3); this conflation is central to the proof error and should be avoided.
[Figure 2] The caption refers to '256 neighboring samples' but does not describe how these neighbors are sampled from the local region; please clarify the procedure.
[Section 5.1] The implementation details state 'batch size of 64 (except for batch size=1)' but do not explain how model updates are performed in the batch size 1 scenario; please specify whether updates are per-sample and how the Gaussian region is estimated in that setting.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ReCAP's loss and its closed-form bounds are derived from the pre-trained classifier's softmax and a Gaussian region model, not from fitted constants, target labels, or the benchmark results they are used to explain.

full rationale

The central derivation chain is self-contained relative to the model and the region model: Definition 3.1 defines region confidence as integrals of entropy and KL divergence over a local region; Eq. 3 writes the softmax probability; Eq. 4 models the region as N(z, tau*Sigma) with Sigma estimated from source data; Lemmas 4.1 and 4.2 (A.1 and A.2) supply inequalities proved with Jensen's inequality and the Gaussian moment-generating function; Propositions 4.3 and 4.4 combine these into closed-form surrogates L_RE and L_RI; and Eq. 9 optimizes these surrogates with a reliability gate. None of these steps fits a parameter to target labels or to the accuracy numbers reported in Tables 1-3. The only estimated input, Sigma, comes from source data, so using it in the bound is not circular. The self-citations (Hu et al. 2024 and Hu et al. 2025) appear only in related-work positioning and are not load-bearing for the derivation. A separate concern is that Eq. 25 in Appendix A.2 identifies the limiting sample average of softmax probabilities with a ratio of Gaussian moment expectations, which is not an identity; if that step fails, Proposition 4.3 is unproven. That is a mathematical soundness or rigor issue in the proof, not a circular reduction of the result to its inputs: the claimed bound is not defined to equal the quantity being bounded, and the empirical evaluation is an external benchmark comparison rather than a restatement of the training objective. Therefore no circular step is present.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim rests on four tuned hyperparameters and one unproven approximation (ratio of expectations as expectation of softmax). The Gaussian region model and the source-data covariance estimate add domain assumptions whose failure would weaken the method's theoretical grounding.

free parameters (5)

region scale tau = 1.2
Controls the covariance of the Gaussian region N(z, tau * Sigma); ablations (Fig. 4b) show performance peaks near 1.2 and degrades for tau=2.5.
variance-term weight lambda = 0.5
Trade-off between the bias term L_RE and the variance term L_RI in Eq. 9; Fig. 4a shows 0.5 is near-optimal.
selection threshold tau_RE = 0.8/1.0 * lnC (ResNet/ViT)
Filters samples with L_RE above this threshold; appendix sensitivity analysis (Fig. 7) shows performance peaks at these values.
weight offset L0 = 0.7/1.0 * lnC (ResNet/ViT)
Appears in the sample weight alpha(x); chosen jointly with tau_RE, no independent justification.
source-sample count for Sigma = 500
Diagonal covariance of the region is estimated from 500 source features ('e.g., 500 samples are enough'); the count is a choice not justified by theory.

assumptions (5)

ad hoc to paper Ratio-of-expectations approximation (Eq. 25 in Appendix A.2)
The proof of Prop 4.3 replaces E[softmax_j(z~)] by the ratio E[e^{a_j·z}] / E[Sigma_k e^{a_k·z}], which does not hold in general; this is the load-bearing step for the claimed upper bound.
ad hoc to paper Diagonal Gaussian region model (Eq. 4)
The local region is assumed multivariate Gaussian with diagonal covariance estimated from source features; no evidence that features are Gaussian or that off-diagonal correlations are negligible.
domain assumption Only normalization-layer affine parameters are adapted (Sec. 5.1)
Standard in TTA (Tent, SAR, DeYO); restricts the hypothesis class.
domain assumption Source data available for estimating Sigma (Sec. 4.1)
The region covariance uses a small source sample at test time; in strict source-free TTA this data may not exist.
standard math Jensen, Gaussian MGF, LLN, Fubini (Appendix A)
Standard background, correctly cited.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Beyond Entropy: Region Confidence Proxy for Wild Test-Time Adaptation." pith.science (2026). https://pith.science/paper/CE7VRHBE

@misc{pith2026250520704,
  author       = {Pith},
  title        = {Pith review of: Beyond Entropy: Region Confidence Proxy for Wild Test-Time Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CE7VRHBE}},
  note         = {Machine review of arXiv:2505.20704}
}

read the original abstract

Wild Test-Time Adaptation (WTTA) is proposed to adapt a source model to unseen domains under extreme data scarcity and multiple shifts. Previous approaches mainly focused on sample selection strategies, while overlooking the fundamental problem on underlying optimization. Initially, we critically analyze the widely-adopted entropy minimization framework in WTTA and uncover its significant limitations in noisy optimization dynamics that substantially hinder adaptation efficiency. Through our analysis, we identify region confidence as a superior alternative to traditional entropy, however, its direct optimization remains computationally prohibitive for real-time applications. In this paper, we introduce a novel region-integrated method ReCAP that bypasses the lengthy process. Specifically, we propose a probabilistic region modeling scheme that flexibly captures semantic changes in embedding space. Subsequently, we develop a finite-to-infinite asymptotic approximation that transforms the intractable region confidence into a tractable and upper-bounded proxy. These innovations significantly unlock the overlooked potential dynamics in local region in a concise solution. Our extensive experiments demonstrate the consistent superiority of ReCAP over existing methods across various datasets and wild scenarios.

Figures

Figures reproduced from arXiv: 2505.20704 by the authors.

**Figure 1.** (a) Illustration of Mild (Wang et al., 2020) and Wild (Niu et al., 2023) TTA settings. (b) Comparison of the adaptation process between mild and wild scenes on the Zoom domain of ImageNet-C dataset (Hendrycks & Dietterich, 2019). Different colors of points represent different predicted classes of samples in the local region. The results highlight that entropy minimization in the wild scenario causes significant loca… view at source ↗

**Figure 2.** Local consistency during the entropy minimization process under mild and wild (imbalanced label shift) scenarios. Consistency is measured by prediction discrepancies between each sample and its 256 neighboring samples. (a) shows the probability of inconsistent predictions in neighbors. (b) records the entropy and average KL Divergence between prediction probabilities of samples and their neighbors. (c) investigates … view at source ↗

**Figure 3.** Overview of our ReCAP. ReCAP performs probabilistic modeling to determine local regions in the latent space (Sec. 4.1). We further derives two closed-form upper bounds for the intractable bias and variance terms via a finite-to-infinite asymptotic approximation, offering an efficient proxy for optimizing Region Confidence without the lengthy sampling process (Sec. 4.2 & 4.3). Definition 3.1. (Region Confidence) Let … view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: (a) Performance with varying strengths λ of the variance term. (b) Performance with different ranges τ of the local region [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 7.** Figure 7: Performance under different selection boundary τRE for ResNet and ViT on ImageNet-C under label shifts [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]

**Figure 8.** Figure 8: The evolution of feature space under DeYO and ReCAP methods. The visualizations are conducted on ImageNet-C under labl shift scenario with ResNet50. ReCAP maintains a consistent performance advantage over prior state-of-the-art methods across a wide range of τRE values…

**Figure 9.** Figure 9: Visualizations of different corruption types in ImageNet corruption benchmark, which are taken from the original paper of ImageNet-C (Hendrycks & Dietterich, 2019). 19 [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]

**Figure 10.** Figure 10: Visualizations of different style shift types in ImageNet-R benchmark, which are taken from the original paper of ImageNet-R (Hendrycks et al., 2021). E. Related Work E.1. Consistency Learning Consistency learning is a key paradigm in semi-supervised learning (Berthel…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

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

Adaptive Dual Uncertainty Optimization: Boosting Monocular 3D Object Detection under Test-Time Shifts
cs.CV 2025-08 conditional novelty 6.0 of 10

DUO adapts monocular 3D detectors at test time by combining a label-free conjugate focal loss for semantic uncertainty with a semantic-guided normal-field consistency loss for geometric uncertainty.
Uncertainty-Aware Spatial Color Correlation for Low-Light Image Enhancement
cs.CV 2025-08 conditional novelty 5.0 of 10

U2CLLIE is a lightweight network for brightening dark images using entropy-guided dual-domain denoising and causal correlation modules, with small PSNR/SSIM gains and mixed LPIPS results.
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

17 extracted references · 12 canonical work pages · cited by 3 Pith papers

[1]

Wild Scenes on ImageNet-R and VisDA-2021 We conduct additional experiments on WTTA scenarios using the ImageNet-R (Hendrycks et al.,

B.1. Wild Scenes on ImageNet-R and VisDA-2021 We conduct additional experiments on WTTA scenarios using the ImageNet-R (Hendrycks et al.,

work page 2021
[6]

Rethinking Distributional Matching Based Domain Adaptation

Li, B., Wang, Y ., Che, T., Zhang, S., Zhao, S., Xu, P., Zhou, W., Bengio, Y ., and Keutzer, K. Rethinking distribu- tional matching based domain adaptation.arXiv preprint arXiv:2006.13352,

work page Pith review arXiv 2006
[7]

Theoretical analysis of self-training with deep networks on unlabeled data

Wei, C., Shen, K., Chen, Y ., and Ma, T. Theoretical analysis of self-training with deep networks on unlabeled data. arXiv preprint arXiv:2010.03622,

arXiv 2010
[8]

The visualizations are conducted on ImageNet-C under labl shift scenario with ResNet50

The evolution of feature space under DeYO and ReCAP methods. The visualizations are conducted on ImageNet-C under labl shift scenario with ResNet50. ReCAP maintains a consistent performance advantage over prior state-of-the-art methods across a wide range of τRE values, showcasing its robustness to variations in the sample selection boundary. C.2. Effecti...

work page 2008
[9]

19 Beyond Entropy: Region Confidence Proxy for Wild Test-Time Adaptation Figure

Visualizations of different corruption types in ImageNet corruption benchmark, which are taken from the original paper of ImageNet-C (Hendrycks & Dietterich, 2019). 19 Beyond Entropy: Region Confidence Proxy for Wild Test-Time Adaptation Figure

work page 2019
[10]

and VisDA-2021 (Bashkirova et al.,

work page 2021
[11]

All evaluations are performed using the same implementation details as outlined in the main paper

This rigorous testing environment ensures a comprehensive assessment of model robustness under real-world conditions. All evaluations are performed using the same implementation details as outlined in the main paper. Tab. 5 presents the results on ImageNet-R for batch size = 1 and imbalanced label distribution shift scenarios. Consistent with the findings...

work page 2021
[13]

18 Beyond Entropy: Region Confidence Proxy for Wild Test-Time Adaptation D.2

prioritizes samples with dominant shape information and applies a dual selection criterion to identify more reliable samples for adaptation. 18 Beyond Entropy: Region Confidence Proxy for Wild Test-Time Adaptation D.2. More Details on Dataset In this paper, we primarily evaluate the out-of-distribution (OOD) generalization ability of all methods using a w...

work page 2019

Show all 17 references

[14]

These artistic variations introduce unique challenges in terms of visual style, texture, and color distribution, which are notably different from the original ImageNet images

and VisDA-2021 (Bashkirova et al., 2022), to further validate the robustness and adaptability of our method across different types of distribution shifts.ImageNet-Rconsists of 30,000 images representing artistic renditions of 200 classes from ImageNet, with each image showcasi...

2021
[16]

Visualizations of different style shift types in ImageNet-R benchmark, which are taken from the original paper of ImageNet-R (Hendrycks et al., 2021). E. Related Work E.1. Consistency Learning Consistency learning is a key paradigm in semi-supervised learning (Berthelot et al....

2021
[17]

Second, consistency learning can act as a regularization technique by introducing data augmentation (Sajjadi et al., 2016)

select samples by evaluating the variation in pseudo-label probabilities under different augmentations, using this as a selection indicator. Second, consistency learning can act as a regularization technique by introducing data augmentation (Sajjadi et al., 2016). By requiring...

2016
[2009]

An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929,

Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929,

2010 arXiv
[2020]

Back to the source: Diffusion-driven test-time adaptation.arXiv preprint arXiv:2207.03442,

Gao, J., Zhang, J., Liu, X., Darrell, T., Shelhamer, E., and Wang, D. Back to the source: Diffusion-driven test-time adaptation.arXiv preprint arXiv:2207.03442,

arXiv
[2021]

and VisDA-2021 (Bashkirova et al., 2022). Both datasets present an array of distribution shifts and variations in data styles that extend beyond the typical corruptions found in ImageNet-C (Hendrycks & Dietterich, 2019), thereby providing a more comprehensive evaluation framew...

2021
[2022]

U., Teterwak, P., et al

Bashkirova, D., Hendrycks, D., Kim, D., Liao, H., Mishra, S., Rajagopalan, C., Saenko, K., Saito, K., Tayyab, B. U., Teterwak, P., et al. Visda-2021 competition: Universal domain adaptation to improve performance on out-of- distribution data. InNeurIPS 2021 Competitions and De...

2021
[2023]

Mixnorm: Test-time adaptation through online normalization estimation.arXiv preprint arXiv:2110.11478,

Hu, X., Uzunbas, G., Chen, S., Wang, R., Shah, A., Neva- tia, R., and Lim, S.-N. Mixnorm: Test-time adaptation through online normalization estimation.arXiv preprint arXiv:2110.11478,

arXiv
[2024]

Seva: Leveraging single- step ensemble of vicinal augmentations for test-time adap- tation.arXiv preprint arXiv:2505.04087,

Hu, Z., Hu, Y ., and Duan, L.-Y . Seva: Leveraging single- step ensemble of vicinal augmentations for test-time adap- tation.arXiv preprint arXiv:2505.04087,

arXiv

Pith tools

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

[1] [1]

Wild Scenes on ImageNet-R and VisDA-2021 We conduct additional experiments on WTTA scenarios using the ImageNet-R (Hendrycks et al.,

B.1. Wild Scenes on ImageNet-R and VisDA-2021 We conduct additional experiments on WTTA scenarios using the ImageNet-R (Hendrycks et al.,

work page 2021

[2] [6]

Rethinking Distributional Matching Based Domain Adaptation

Li, B., Wang, Y ., Che, T., Zhang, S., Zhao, S., Xu, P., Zhou, W., Bengio, Y ., and Keutzer, K. Rethinking distribu- tional matching based domain adaptation.arXiv preprint arXiv:2006.13352,

work page Pith review arXiv 2006

[3] [7]

Theoretical analysis of self-training with deep networks on unlabeled data

Wei, C., Shen, K., Chen, Y ., and Ma, T. Theoretical analysis of self-training with deep networks on unlabeled data. arXiv preprint arXiv:2010.03622,

arXiv 2010

[4] [8]

The visualizations are conducted on ImageNet-C under labl shift scenario with ResNet50

The evolution of feature space under DeYO and ReCAP methods. The visualizations are conducted on ImageNet-C under labl shift scenario with ResNet50. ReCAP maintains a consistent performance advantage over prior state-of-the-art methods across a wide range of τRE values, showcasing its robustness to variations in the sample selection boundary. C.2. Effecti...

work page 2008

[5] [9]

19 Beyond Entropy: Region Confidence Proxy for Wild Test-Time Adaptation Figure

Visualizations of different corruption types in ImageNet corruption benchmark, which are taken from the original paper of ImageNet-C (Hendrycks & Dietterich, 2019). 19 Beyond Entropy: Region Confidence Proxy for Wild Test-Time Adaptation Figure

work page 2019

[6] [10]

and VisDA-2021 (Bashkirova et al.,

work page 2021

[7] [11]

All evaluations are performed using the same implementation details as outlined in the main paper

This rigorous testing environment ensures a comprehensive assessment of model robustness under real-world conditions. All evaluations are performed using the same implementation details as outlined in the main paper. Tab. 5 presents the results on ImageNet-R for batch size = 1 and imbalanced label distribution shift scenarios. Consistent with the findings...

work page 2021

[8] [13]

18 Beyond Entropy: Region Confidence Proxy for Wild Test-Time Adaptation D.2

prioritizes samples with dominant shape information and applies a dual selection criterion to identify more reliable samples for adaptation. 18 Beyond Entropy: Region Confidence Proxy for Wild Test-Time Adaptation D.2. More Details on Dataset In this paper, we primarily evaluate the out-of-distribution (OOD) generalization ability of all methods using a w...

work page 2019