Test-time Correlation Alignment

Jiabao Lu; Linjing You; Xiayuan Huang

REVIEW 4 major objections 5 minor 31 references

Test-time Correlation Alignment

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a model can adapt at test time, with no backpropagation and no source data, by aligning the feature correlation of its most confident test predictions with the feature correlation of the test stream, and that this…

desk verdict A genuinely cheap, backprop-free TTA plug-in with consistent empirical gains, wrapped in a theory that doesn't survive a close read. read the letter →

arxiv 2505.00533 v2 pith:R7KBCVN4 submitted 2025-05-01 cs.LG cs.AI

Linjing You , Jiabao Lu , Xiayuan Huang This is my paper · ORCID

classification cs.LGcs.AI

keywords test-timeadaptationcorrelationalignmentpseudo-sourcecovariancebackpropagation-freedistributionshiftdomaingeneralization

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 usually means updating a model with gradients on unlabeled test data, which is slow, memory-hungry, and prone to forgetting the source domain. This paper tries to establish a cheaper route: treat the test samples the model is most confident about as a stand-in 'pseudo-source,' compute their feature covariance, and linearly transform the rest of the test features to match it. The theoretical part argues that high-certainty test features approximate the true source covariance (Theorem 3.5), and that shrinking the distance between test covariance and this pseudo-source covariance lowers a bound on test classification error (Corollary 3.7). Practically, the resulting LinearTCA needs no model update and nearly no extra memory or time, and the LinearTCA+ variant is reported to improve its chosen base TTA method in every benchmark cell tested. A sympathetic reader would care because, if this stands, correlation-level alignment—previously unavailable without source data—becomes a cheap and attachable test-time fix.

What carries the argument

The load-bearing object is the pseudo-source covariance $\hat{\Sigma}_s = \frac{1}{\hat{n}_s - 1}(\hat{Z}_s^T \hat{Z}_s - \frac{1}{\hat{n}_s} \mathbf{1}^T \hat{Z}_s^T \hat{Z}_s \mathbf{1})$, built from the k test embeddings whose predictions have the smallest uncertainty $\|\hat{y}_t - p_t\|$ (Eqs. 15-16). It substitutes for the unavailable source covariance and turns correlation alignment into a test-time operation. The second piece of machinery is the closed-form linear transform $W = U_t \Lambda_t^{1/2} \hat{U}_s^T \hat{\Lambda}_s^{-1/2}$ (Eq. 19), which whitens the test covariance and recolors it to match the pseudo-source covariance, applied as $Z'_t = (Z_t - \mu_t)W + \hat{\mu}_s$ so that both second-order correlation and first-order mean are aligned in a single matrix multiplication. The argument is carried by the inequality chain in Corollary 3.7 (Eq. 14), which converts the Theorem 3.5 covariance bound into an explicit test-error bound and shows that reducing $\|\Sigma_t - \hat{\Sigma}_s\|$ is the operation that lowers classification error.

What would settle it

On a heavily corrupted benchmark such as CIFAR-10-C at severity 5, compute the ground-truth accuracy of the k highest-certainty test predictions the method selects as its pseudo-source; if that accuracy is no higher than the average test accuracy, the certainty-for-correctness proxy fails and the premise of Theorem 3.5 collapses. One can also measure, per corruption type, whether the correlation distance from the pseudo-source covariance to the true source covariance actually shrinks as certainty increases, to locate where the guarantee stops holding.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is that CORAL-style correlation alignment is feasible at test time without source data, because the source correlation can be replaced by a pseudo-source correlation computed from the k test instances with the smallest prediction uncertainty $\omega_i = \|\hat{y}_i - p_i\|$. Theorem 3.5 bounds the distance between this pseudo-source covariance and the true source covariance by a quantity driven by prediction uncertainty together with the output-error terms $\varepsilon(h_\theta(X_t))$ and $\varepsilon(h_\theta(X_s))$; Theorem 3.6 bounds test error by the source-test covariance distance; Corollary 3.7 chains the two, so that aligning $\|\Sigma_t - \hat{\Sigma}_s\|$ provably lowers the test-error bound. The algorithms realize this with a closed-form linear transformation $W = U_t \Lambda_t^{1/2} \hat{U}_s^T \hat{\Lambda}_s^{-1/2}$ applied as $Z'_t = (Z_t - \mu_t)W + \hat{\mu}_s$: LinearTCA applies it directly to the frozen network's embeddings, while LinearTCA+ applies it on top of any existing TTA method's adapted embeddings. Empirically, LinearTCA+ raises accuracy over its base method in every cell of the main comparison table (for instance 87.83 to 88.77 on PACS with ResNet-18, and 76.93 to 77.13 on CIFAR-10-C), with LinearTCA itself using roughly 4% of the peak GPU memory and 0.6% of the runtime of the strongest baseline.

Load-bearing premise

The load-bearing premise is that the test samples the model is most confident about are also the ones it classifies correctly — certainty is used as a proxy for error — so under heavy corruption, where confidently wrong predictions are common, the pseudo-source stops resembling the source domain and the alignment moves features toward the wrong target.

Editorial extensions

If this is right

Test-time adaptation becomes a single forward pass: a frozen model plus one matrix multiplication on its embeddings, so devices without the memory for backpropagation can still adapt on the fly.
Any existing backpropagation-based TTA method can be boosted by post-processing its outputs with correlation alignment; the paper reports LinearTCA+ surpassing the best baseline on every dataset and backbone in Table 1.
Because model parameters never change, source-domain knowledge is retained: measured source accuracy after adaptation drops by far less than with gradient-based methods, and even improves on PACS.
The theory connects test error to an observable quantity, the covariance distance between the test stream and the high-certainty pseudo-source, giving practitioners a monitorable signal for whether adaptation is working.
The method stays effective with batch size 1 and small buffers, since covariance estimates accumulate incrementally over the stream rather than requiring a large batch.

Reading between the lines

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

If certainty truly proxies correctness, the same pseudo-source construction could be reused for other source-free statistics, such as class-conditional covariances, higher-order moments, or correlation targets for regression and segmentation heads, not just global feature covariance.
The paper's own analysis shows linear alignment saturates on nonlinear shifts; a natural next step it does not take is to choose between linear and nonlinear alignment on the fly by monitoring whether the covariance distance $\|\Sigma_t - \hat{\Sigma}_s\|$ actually decreases under the transform.
A direct test of the crux would examine each corruption type at severity 5 and check whether the selected high-certainty predictions are genuinely more accurate than average; where confidently wrong predictions cluster, the pseudo-source premise breaks and the method should degrade predictably.
The reported positive backward transfer on PACS hints that pseudo-source alignment can act as a regularizer rather than a distortion; verifying whether this survives when the pseudo-source is built from confidently misclassified samples would sharpen the method's failure boundary.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

The practical core is the real news: LinearTCA+ is a backpropagation-free, plug-and-play module that improves a chosen base TTA method in every reported benchmark cell, and the efficiency numbers (near-zero added memory, 0.6% of AdaNPC's runtime on ViT) are striking if they reproduce. The two ideas — constructing a pseudo-source from high-certainty test features, and aligning test covariance to it with a closed-form linear transform without model updates — are genuinely not present in the papers they cite, and the plug-in formulation is clean. The forgetting-resistance results in Table 4 also line up with the no-update design, and the batch-size robustness in Table 8 is a nice bonus. Credit where it's due: the empirical core looks consistent and valuable.

The load-bearing theory, though, does not hold up. The stress-test note is right: the proof of Eq. (11) divides by the Jacobian norm as though Jg were a scalar. It is a matrix; the correct bound is governed by the smallest singular value, and the stated inequality is not implied. So the Corollary 3.7 guarantee is not established. The density condition in Assumption 3.2 is also practically vacuous for image inputs (the pi_{dI} r^{dI} term with d_I = 3072 is negligible unless r is unphysical), and the paper never instantiates r* or the other constants. That makes the central theoretical claim a heuristic at best.

There are additional, smaller soft spots. Hyperparameters (notably k) are tuned on the test set across datasets, which inflates the reported numbers; this is common in TTA papers but should be stated openly. The Appendix D description of the algorithm — "we first obtain the embeddings of all test data" — does not match the online streaming protocol in the main text; that discrepancy needs to be resolved. And Section 5.3's prose overclaims slightly: LinearTCA+ does consistently boost its base, but LinearTCA itself trails strong baselines on corruption benchmarks, which the paper later acknowledges in Section 5.5.

Who this is for: anyone working on test-time adaptation, especially on efficiency or forgetting constraints. The method is likely to be useful regardless of the theoretical wrapper. The paper deserves a serious referee, but the version sent to review should be a major revision: fix or reframe the theory, specify the optimization details, make the online/offline protocol unambiguous, and report error bars. I would not desk-reject this.

Referee Report

4 major / 5 minor

Summary. The paper proposes Test-time Correlation Alignment (TCA), a backpropagation-free test-time adaptation approach. The core idea is to use the k test embeddings with the smallest prediction uncertainty as a pseudo-source, compute their covariance as a surrogate for the unavailable source covariance, and then apply a linear transform to align the test covariance to this pseudo-source covariance. The authors provide a theoretical analysis (Theorem 3.5, Theorem 3.6, Corollary 3.7) intended to show that high-certainty test instances approximate the source correlation and that correlation alignment reduces a test-error bound. They instantiate this as LinearTCA and as a plug-and-play module LinearTCA+ for existing TTA methods. The experimental section reports accuracy gains across PACS, OfficeHome, DomainNet, CIFAR-10/100-C, ImageNet-C, and CLIP, together with low memory/time overhead and improved forgetting resistance.

Significance. If the theoretical guarantee and the closed-form transform were correct, the paper would make a useful contribution: a simple, cheap, and model-agnostic TTA module that addresses correlation alignment without source data. The empirical package is a genuine strength: Table 1 and the appendix tables consistently show LinearTCA+ improving its chosen base method in every reported cell, Table 2 shows near-zero added memory and runtime, Table 4 supports the forgetting-resistance claim, and the code is publicly available. However, the theoretical derivation currently contains a load-bearing linear-algebra error, the closed-form transform does not satisfy the stated objective, and the connection between confidence and accuracy is asserted rather than proved. These issues are fixable in principle, but as written the central claimed guarantee is not established.

major comments (4)

[Appendix B.1, Eqs. (24)-(27) and Theorem 3.5] The proof of Theorem 3.5 treats the Jacobian Jg(Zs) as a scalar when passing from Pt - Ps = Jg(Zs)dZs + o(dZs) to dZs = (Pt - Ps - o)/Jg(Zs). This step is not a valid matrix operation. The correct inversion uses a pseudo-inverse and gives ||dZs|| <= (||Pt-Ps|| + ||o||)/sigma_min(Jg), not the displayed expression with ||Jg(Zs)|| in the denominator; the inequality ||Pt-Ps|| <= ||Jg|| ||dZs|| only yields a lower bound on ||dZs||. Therefore Eq. (11), Eq. (12), and Corollary 3.7 are not implied by the stated proof. The theorem is likely repairable by replacing ||Jg(Zs)||_F with sigma_min(Jg), but the guarantee as written is unsupported.
[Section 4.2, Eq. (19)] The claimed closed-form solution W = U_t Lambda_t^{1/2} \hat U_s^T \hat Lambda_s^{-1/2} does not satisfy the stated objective W^T Sigma_t W = \hat Sigma_s. In the scalar case the formula gives W = sqrt(sigma_t/sigma_s), whereas the required transform is sqrt(sigma_s/sigma_t). A correct form is W = U_t Lambda_t^{-1/2} \hat U_s \hat Lambda_s^{1/2} \hat U_s^T, equivalently Sigma_t^{-1/2} \hat Sigma_s^{1/2}. Since the authors report using gradient descent rather than Eq. (19) in the experiments, this error may not affect the empirical numbers, but the derivation as presented is incorrect.
[Theorem 3.5, Eq. (12), and Remark after Corollary 3.7] The bound in Eq. (12) contains the hypothesis errors epsilon(h_theta(X_t)) and epsilon(h_theta(X_s)) inside the term A, but the selection rule in Section 4.1 minimizes only \|\hat Y_t - P_t\|. The Remark asserts, citing prior empirical studies, that higher output certainty implies lower output error; this is an unproved assumption, not a consequence of the stated conditions. Without a formal condition linking confidence to accuracy, the conclusion that high-certainty selection makes the pseudo-source covariance close to the source covariance is not a theorem. Figure 3a gives indirect support, but it does not evaluate the actual k-selection procedure used in Algorithm 1.
[Assumption 3.2 and Theorem 3.5] The probability bound in Theorem 3.5 depends on the ball-volume term pi_{d_I} r^{d_I} with d_I the input dimension, which is 3072 for the image benchmarks. Unless r is taken to be unrealistically large, this term is astronomically small and the stated probability is effectively zero; the paper never instantiates r*, c_t, or mu_- for any experiment. The guarantee is therefore vacuous in the tested regime. The authors should either formulate the assumption at a lower-dimensional feature level with appropriate hypotheses, or explicitly discuss parameter ranges in which the probability is non-negligible.

minor comments (5)

[Appendix B.1, Eq. (43)] Eq. (43) in the appendix states a bound on \|Sigma_s - Sigma_t\|_F, while Theorem 3.5 and Eq. (12) in the main text state a bound on \|Sigma_s - \hat Sigma_s\|_F; these notations should be aligned.
[Theorem 3.5 and Eq. (12)] The phrase "one-hot encoding of P_t" is ambiguous; it should read "one-hot encoding of argmax(P_t)" for consistency with Section 4.1.
[Appendix B.1, Eq. (38)] The symbol d is used both for the embedding dimension and for the range of Sigma'_s in the Hoeffding bound; use distinct notation to avoid confusion.
[Section 4.2, footnote 1] For the gradient-descent implementation of W, the paper reports only a learning rate; please specify the optimizer and the number of iterations for reproducibility.
[Conclusion] The word "chanllenges" should be "challenges", and the abstract's "methods significantly outperforms" should agree in number.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the theoretical bound is a conditional sufficient condition, not an identity, and the empirical claims are checked against real source data and labels.

full rationale

The derivation chain is not circular. Theorem 3.5 defines the pseudo-source covariance \hat\Sigma_s as the covariance of the k test features with smallest uncertainty \|\hat Y_t - P_t\|, and then proves a bound on \|\hat\Sigma_s - \Sigma_s\| in terms of that same uncertainty plus hypothesis errors \epsilon(h_\theta(X_t)), \epsilon(h_\theta(X_s)) and a sampling term. That is a sufficient condition, not an equivalence by construction: the bound could in principle be large even when uncertainty is small, and the paper explicitly invokes external empirical evidence that high certainty correlates with low error, rather than defining certainty as error. Corollary 3.7 is only a triangle inequality application, so no conclusion is reused as an input. The transformation W is solved to make W^T\Sigma_t W equal to \hat\Sigma_s, but the claims that this improves accuracy and reduces distance to the true source are validated against real labels and source data in Table 1, Table 7 and Figure 3, so they are not forced by the optimization objective. The only self-citation in the paper (You et al., ICASSP 2025) appears in the introduction as one of several routine TTA citations and is not load-bearing. The scalar-Jacobian inversion in Appendix B.1 is a mathematical proof defect that would undermine the stated guarantee, but it is a correctness issue, not a circularity: it does not make any predicted quantity equal to its input by definition. No fitted parameter is renamed a prediction, no uniqueness theorem is imported from the authors, and no ansatz is smuggled in through self-citation.

Assumptions & free parameters 4 free parameters · 6 assumptions · 1 invented entities

The contribution rests on three modeling assumptions (density, Lipschitz, Taylor linearity), one empirical heuristic (certainty implies correctness) imported from prior work, and two cited bounds (Shen et al. 2018, Gui et al. 2024). No new physical entities are postulated; the pseudo-source correlation is a measurement-based construct with a direct empirical handle (Figure 3a). The method's free parameters are the pseudo-source size k, the W-optimization learning rate, an unspecified optimizer and step count for the W update, and the per-benchmark choice of base method for LinearTCA+.

free parameters (4)

k (pseudo-source size) = 5 to 300 per dataset and backbone, e.g., fkTCA=30 on PACS ResNet-18 and ResNet-50, fkTCA=10 on OfficeHome ResNet-18…
Tuned against test accuracy per dataset; controls the bias-variance balance of the pseudo-source covariance estimate and is the method's only named hyperparameter in the main text.
W-optimization learning rate = 1e-3
Footnote 1: fixed learning rate for the gradient-based solve of Eq. (18); stated without a search or sensitivity study.
Optimizer and step count for W
Not specified in the paper; the closed-form CORAL solution of Eq. (19) is replaced by an unspecified gradient routine, so exact behavior depends on unstated choices.
Base method for LinearTCA+ = TSD on PACS R18/R50, OfficeHome R50 and CIFAR-10-C; TIPI on PACS ViT and OfficeHome ViT; TEA on OfficeHome R18; EATA…
Selected as the best-performing baseline per benchmark before applying the plug-in module; an oracle choice that the reported gains depend on.

assumptions (6)

domain assumption Assumption 3.2: strong density condition
Assumes source and test densities are comparable with bounded ratios and that every test point has a ball of radius up to r_t with at least c_t lambda of source mass within it. Load-bearing for the Chernoff bound in the proof of Theorem 3.5, but for image inputs (d_I = 3072) the volume term pi_{dI} r^{dI} makes the stated probability negligible for r < 1.
standard math Assumption 3.3: L-Lipschitz encoder
Used to convert input-space proximity to feature-space proximity via Eq. (32). Standard in the cited domain adaptation literature (Mansour et al. 2009).
domain assumption Assumption 3.4: first-order Taylor approximation of the decoder
Assumes g is approximately affine within feature radius r*. The bound in Eq. (12) divides by ||J_g(Z_s)||, requiring a non-degenerate Jacobian; saturated last layers would make the bound blow up.
domain assumption Certainty implies correctness
The Remark after Theorem 3.5 states that prior studies empirically suggest high output certainty implies low output error; the selection rule of Section 4.1 uses this to justify the pseudo-source. Not proven in the paper and known to fail for confidently wrong predictions.
standard math Lemma B.2 (Theorem 1 of Shen et al. 2018)
Wasserstein-distance domain adaptation bound used as the backbone of Theorem 3.6.
standard math Generalization bound of Gui et al. 2024 (Lemma 7)
VC-dimension generalization bound used in the proof of Theorem 3.6.

invented entities (1)

Pseudo-source correlation (high-certainty test subset as a stand-in for the source) independent evidence
purpose: Replaces the unavailable source covariance Sigma_s with a covariance computed from the k most certain test instances, enabling CORAL-style alignment at test time (Eqs. 15 and 16).
It is a directly measurable construct with an empirical handle: Figure 3a shows the uncertainty-correlation distance relationship on real data, so the entity is falsifiable outside the paper's own model. It is not a new physical quantity.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Test-time Correlation Alignment." pith.science (2026). https://pith.science/paper/R7KBCVN4

@misc{pith2026250500533,
  author       = {Pith},
  title        = {Pith review of: Test-time Correlation Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R7KBCVN4}},
  note         = {Machine review of arXiv:2505.00533}
}

read the original abstract

Deep neural networks often degrade under distribution shifts. Although domain adaptation offers a solution, privacy constraints often prevent access to source data, making Test-Time Adaptation (TTA, which adapts using only unlabeled test data) increasingly attractive. However, current TTA methods still face practical challenges: (1) a primary focus on instance-wise alignment, overlooking CORrelation ALignment (CORAL) due to missing source correlations; (2) complex backpropagation operations for model updating, resulting in overhead computation and (3) domain forgetting. To address these challenges, we provide a theoretical analysis to investigate the feasibility of Test-time Correlation Alignment (TCA), demonstrating that correlation alignment between high-certainty instances and test instances can enhance test performances with a theoretical guarantee. Based on this, we propose two simple yet effective algorithms: LinearTCA and LinearTCA+. LinearTCA applies a simple linear transformation to achieve both instance and correlation alignment without additional model updates, while LinearTCA+ serves as a plug-and-play module that can easily boost existing TTA methods. Extensive experiments validate our theoretical insights and show that TCA methods significantly outperforms baselines across various tasks, benchmarks and backbones. Notably, LinearTCA achieves higher accuracy with only 4% GPU memory and 0.6% computation time compared to the best TTA baseline. It also outperforms existing methods on CLIP over 1.86%.

Figures

Figures reproduced from arXiv: 2505.00533 by the authors.

**Figure 1.** Illustration of key limitations in existing TTA methods. (a) Correlation distance increases with domain shifts. (b) Computation time and peak GPU memory usage on CIFAR-10-C, showing high overhead of existing methods. (c) Source domain performance after test-time adaptation, revealing challenges in retaining source knowledge. 1. Introduction Deep neural networks (DNNs) have significantly advanced numerous tasks in … view at source ↗

**Figure 2.** The pipeline of our proposed LinearTCA method. During testing, we first obtain original embeddings and predictions using the source model. Based on the certainty of the original predictions, we select a subset embeddings to form a “pseudo-source domain”. A linear transformation is then applied to align the correlations of the original embeddings with those of the pseudo-source domain, ultimately producing the fin… view at source ↗

**Figure 3.** Experimental validation of theories. (a) Average uncertainty and correlation distance to source domain of each group, groups with lower uncertainty exhibit smaller correlation distances. (b) Relationships between ACC, correlation distance to the source, and correlation distance to the pseudo-source, both ACC and ∥Σt − Σs∥ are strongly linearly related to ∥Σt − Σˆ s∥. 5. Experiments 5.1. Experimental settings Follow… view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Analysis of TCA. (a) When the test domain (yellow) undergoes a nearly linear shift from the source domain (blue), after adaptation by LinearTCA, the transformed test domain (red) is well-aligned with the source. (b) In the case of a nonlinear shift, although partial al…

**Figure 5.** Figure 5: Adaptation process of LinearTCA to datasets with linear shifts. 29 [PITH_FULL_IMAGE:figures/full_fig_p029_5.png]

**Figure 6.** Figure 6: Adaptation process of LinearTCA to datasets with nonlinear shifts 30 [PITH_FULL_IMAGE:figures/full_fig_p030_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 22 canonical work pages

[1]

Let 2 exp − 2kϵ d2 =σ, then: ϵ =−log(σ 2 ) 2k (39) With a probability of at least 1−σ, we have: ∥Σ′ s− Σs∥F <√ϵ = r log(2/δ) 2k (40) By combining Equations (37) and (40), we obtain: ∥Σs− ˆΣs∥F≤∥ Σs− Σ′ s∥F +∥Σ′ s− ˆΣs∥F≤ r log(2/δ) 2k + 2∥Zs∥F∥dZs∥F +∥dZs∥2 F (41) We can further expand Equation (41) by applying Equation (33). However, since we cannot dete...

work page 2017
[6]

ISBN 9781713845393

Curran Associates Inc. ISBN 9781713845393. Jang, M., Chung, S.-Y ., and Chung, H. W. Test-time adapta- tion via self-training with nearest neighbor information. arXiv preprint arXiv:2207.10792,

arXiv
[8]

Domain adaptation: Learning bounds and algorithms

Mansour, Y ., Mohri, M., and Rostamizadeh, A. Domain adaptation: Learning bounds and algorithms. arXiv preprint arXiv:0902.3430,

arXiv
[10]

Moment matching for multi-source domain adaptation

Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., and Wang, B. Moment matching for multi-source domain adaptation. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 1406–1415,

work page 2019
[12]

Towards real-world test-time adaptation: Tri-net self-training with balanced normaliza- tion

Su, Y ., Xu, X., and Jia, K. Towards real-world test-time adaptation: Tri-net self-training with balanced normaliza- tion. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp. 15126–15135, 2024a. Su, Y ., Xu, X., Li, T., and Jia, K. Revisiting realistic test- time training: Sequential inference and adaptation by an- chored clus...

work page 2016
[14]

Probabilistic contrastive test-time adaptation

You, L., Lu, J., and Huang, X. Probabilistic contrastive test-time adaptation. In ICASSP 2025 - 2025 IEEE In- ternational Conference on Acoustics, Speech and Sig- nal Processing (ICASSP), pp. 1–5,

work page 2025
[16]

integrates CORAL into Graph Neural Network (GNN) to generate the domain-invariant features. Although CORAL has achieved significant success in domain adaptation (DA), its application in test-time adaptation (TTA) is constrained by privacy and resource limitations, which make it infeasible to compute the source correlation. This limitation significantly ha...

work page 2024
[17]

Although BP-Free TTA methods are computationally efficient, their image corruption adaptation capabilities are often limited

utilize k-nearest neighbor (kNN) to refine predictions. Although BP-Free TTA methods are computationally efficient, their image corruption adaptation capabilities are often limited. Backpropagation (BP)-Based TTA: This group encompasses certain pseudo-labeling methods (Zeng et al., 2024), consistency training methods (Sinha et al., 2023), and clustering-b...

work page 2024

Show all 31 references

[18]

filters unreliable features or predictions with high entropy, as lower entropy correlates with higher accuracy, and applies a consistency filter to refine instances further. Consistency training methods aim to enhance the stability of network predictions or features by address...

2023
[19]

removes noisy instances with large gradients and encourages model weights to converge toward a flat minimum, enhancing robustness against residual noise. Generally, BP-Based TTA methods demonstrate superior domain adaptation capabilities compared to BP-Free methods, but they t...

2023
[22]

2: Output: Final predictionsP ′ T

Algorithm 1 LinearTCA Algorithm 1: Input: Test instancesXt, source modelhθ. 2: Output: Final predictionsP ′ T . 3: If use LinearTCA+: Updateθ by Equation (1) 4: Obtain embeddings and predictions: ˆPt,Zt =hθ(Xt) 5: Selectk high-certainty embeddings: ˆZs ={Zt[i]|ωi t≤ωk min} 6: ...

2020
[23]

27.43 33.56 21.57 43.64 40.48 51.26 51.29 68.18 54.52 66.65 87.50 27.59 67.06 48.86 72.37 50.80 BN (Schneider et al., 2020)66.05 68.22 56.83 82.34 57.86 79.78 82.32 74.99 74.30 78.85 87.22 81.80 70.31 73.61 71.00 73.70 T3A (Iwasawa & Matsuo, 2021)44.16 50.32 29.64 56.98 49.02 ...

2020
[24]

30.81 37.09 24.71 38.07 41.66 51.97 51.17 68.49 60.52 66.79 86.19 28.25 65.19 38.95 71.66 50.77 BN (Schneider et al., 2020)61.98 63.05 56.25 82.58 54.49 80.11 82.61 74.16 72.36 79.28 87.04 81.06 67.16 71.27 70.22 72.24 T3A (Iwasawa & Matsuo, 2021)45.34 49.51 36.76 39.10 46.88 ...

2020
[25]

37.25 44.31 39.94 83.16 70.31 83.54 85.80 87.15 85.06 79.19 92.75 29.73 84.73 84.68 84.58 71.48 BN (Schneider et al., 2020)0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 T3A (Iwasawa & Matsuo, 2021)47.84 52.78 51.52 83.16 73.06 83.35 85.66 87.0...

2020
[26]

10.46 12.49 3.36 34.44 23.63 38.10 42.67 39.25 33.01 32.84 55.78 11.55 46.48 34.88 46.15 31.01 BN (Schneider et al., 2020)39.78 39.81 29.95 56.18 40.92 54.71 58.68 48.52 49.59 46.79 61.89 48.63 50.26 54.61 45.37 48.38 T3A (Iwasawa & Matsuo, 2021)10.51 11.59 3.93 36.77 26.94 40...

2020
[27]

17.23 19.42 9.77 35.34 31.87 39.15 41.98 41.99 38.68 32.00 54.56 11.18 47.57 42.51 47.02 34.02 BN (Schneider et al., 2020)42.09 42.22 31.37 56.23 42.36 54.61 57.22 48.43 49.61 45.29 60.06 45.07 50.52 55.09 45.96 48.41 T3A (Iwasawa & Matsuo, 2021)18.46 20.58 10.98 37.34 34.81 4...

2020
[28]

21.71 24.74 19.53 62.41 43.14 61.13 67.65 66.34 67.48 54.03 77.43 33.26 60.09 60.48 56.17 51.71 BN (Schneider et al., 2020)0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 T3A (Iwasawa & Matsuo, 2021)24.32 27.25 23.66 65.04 47.67 63.22 69.56 67.4...

2020
[29]

1.54 2.27 1.48 11.44 8.68 11.12 17.62 10.64 16.21 14.02 51.52 3.44 16.49 23.35 30.67 14.70 BN (Schneider et al., 2020)13.65 14.84 14.17 11.95 13.04 23.34 33.89 29.18 28.42 40.80 58.11 12.09 38.92 44.35 37.08 27.59 T3A (Iwasawa & Matsuo, 2021)1.61 2.35 1.65 10.57 8.20 10.12 17....

2020
[30]

3.00 3.70 2.64 17.91 9.74 14.71 22.45 16.60 23.06 24.01 59.12 5.38 16.51 20.87 32.63 18.15 BN (Schneider et al., 2020)16.32 17.09 16.97 15.23 15.54 26.64 39.38 34.46 33.45 48.43 65.67 17.08 44.62 49.49 40.47 32.06 T3A (Iwasawa & Matsuo, 2021)2.97 3.38 2.65 17.05 9.37 13.69 22....

2020
[31]

35.09 32.16 35.88 31.42 25.31 39.45 31.55 24.47 30.13 54.74 64.48 48.98 34.20 53.17 56.45 39.83 BN (Schneider et al., 2020)0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 T3A (Iwasawa & Matsuo, 2021)27.87 28.15 30.27 32.42 27.00 40.66 33.52 25.9...

2020
[2006]

Active test-time adaptation: Theoretical analyses and an algorithm

Gui, S., Li, X., and Ji, S. Active test-time adaptation: Theoretical analyses and an algorithm. arXiv preprint arXiv:2404.05094,

arXiv
[2015]

Pseudo-label : The simple and efficient semi- supervised learning method for deep neural networks

Lee, D.-H. Pseudo-label : The simple and efficient semi- supervised learning method for deep neural networks. ICML 2013 Workshop : Challenges in Representation Learning (WREPL), 07

2013
[2016]

and Dietterich, T

Hendrycks, D. and Dietterich, T. G. Benchmarking neural network robustness to common corruptions and pertur- bations. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9,

2019
[2017]

Tent: Fully test-time adaptation by entropy minimiza- tion

Wang, D., Shelhamer, E., Liu, S., Olshausen, B., and Darrell, T. Tent: Fully test-time adaptation by entropy minimiza- tion. arXiv preprint arXiv:2006.10726,

2006 arXiv
[2018]

(Theorem 1 in (Shen et al., 2018)) LetH be anL-Lipschitz continuous hypothesis class with VC-dimension dv

as follows: Lemma B.2. (Theorem 1 in (Shen et al., 2018)) LetH be anL-Lipschitz continuous hypothesis class with VC-dimension dv. Given two domain distributions, Ds and Dt, letγ = minh∈H{ϵs(h(t)) +ϵt(h(t))}. The risk of hypothesis ˆh on the test domain is then bounded by: ϵt(ˆ...

2018
[2019]

Qian, Q., Qin, Y ., Luo, J., Wang, Y ., and Wu, F

doi: 10.1109/ICCV .2019.00149. Qian, Q., Qin, Y ., Luo, J., Wang, Y ., and Wu, F. Deep dis- criminative transfer learning network for cross-machine fault diagnosis. Mechanical Systems and Signal Process- ing, 186:109884,

2019
[2021]

An image is worth 16x16 words: Trans- formers for image recognition at scale

Dosovitskiy, A. An image is worth 16x16 words: Trans- formers for image recognition at scale. arXiv preprint arXiv:2010.11929,

2010 arXiv
[2022]

Robust and high-order correlation alignment for unsupervised domain adaptation

Cheng, Z., Chen, C., Chen, Z., Fang, K., and Jin, X. Robust and high-order correlation alignment for unsupervised domain adaptation. Neural Computing and Applications, 33:6891–6903, 2021a. Cheng, Z., Chen, C., Chen, Z., Fang, K., and Jin, X. Robust and high-order correlation a...

arXiv
[2023]

Test- time model adaptation with only forward passes

Niu, S., Miao, C., Chen, G., Wu, P., and Zhao, P. Test- time model adaptation with only forward passes. arXiv preprint arXiv:2404.01650,

arXiv
[2024]

and Sun, Y

Hardt, M. and Sun, Y . Test-time training on nearest neighbors for large language models. arXiv preprint arXiv:2305.18466,

arXiv
[2025]

Yuan, L., Xie, B., and Li, S

doi: 10.1109/ ICASSP49660.2025.10890260. Yuan, L., Xie, B., and Li, S. Robust test-time adaptation in dynamic scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 15922–15932,

2025

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [1]

Let 2 exp − 2kϵ d2 =σ, then: ϵ =−log(σ 2 ) 2k (39) With a probability of at least 1−σ, we have: ∥Σ′ s− Σs∥F <√ϵ = r log(2/δ) 2k (40) By combining Equations (37) and (40), we obtain: ∥Σs− ˆΣs∥F≤∥ Σs− Σ′ s∥F +∥Σ′ s− ˆΣs∥F≤ r log(2/δ) 2k + 2∥Zs∥F∥dZs∥F +∥dZs∥2 F (41) We can further expand Equation (41) by applying Equation (33). However, since we cannot dete...

work page 2017

[2] [6]

ISBN 9781713845393

Curran Associates Inc. ISBN 9781713845393. Jang, M., Chung, S.-Y ., and Chung, H. W. Test-time adapta- tion via self-training with nearest neighbor information. arXiv preprint arXiv:2207.10792,

arXiv

[3] [8]

Domain adaptation: Learning bounds and algorithms

Mansour, Y ., Mohri, M., and Rostamizadeh, A. Domain adaptation: Learning bounds and algorithms. arXiv preprint arXiv:0902.3430,

arXiv

[4] [10]

Moment matching for multi-source domain adaptation

Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., and Wang, B. Moment matching for multi-source domain adaptation. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 1406–1415,

work page 2019

[5] [12]

Towards real-world test-time adaptation: Tri-net self-training with balanced normaliza- tion

Su, Y ., Xu, X., and Jia, K. Towards real-world test-time adaptation: Tri-net self-training with balanced normaliza- tion. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp. 15126–15135, 2024a. Su, Y ., Xu, X., Li, T., and Jia, K. Revisiting realistic test- time training: Sequential inference and adaptation by an- chored clus...

work page 2016

[6] [14]

Probabilistic contrastive test-time adaptation

You, L., Lu, J., and Huang, X. Probabilistic contrastive test-time adaptation. In ICASSP 2025 - 2025 IEEE In- ternational Conference on Acoustics, Speech and Sig- nal Processing (ICASSP), pp. 1–5,

work page 2025

[7] [16]

integrates CORAL into Graph Neural Network (GNN) to generate the domain-invariant features. Although CORAL has achieved significant success in domain adaptation (DA), its application in test-time adaptation (TTA) is constrained by privacy and resource limitations, which make it infeasible to compute the source correlation. This limitation significantly ha...

work page 2024

[8] [17]

Although BP-Free TTA methods are computationally efficient, their image corruption adaptation capabilities are often limited

utilize k-nearest neighbor (kNN) to refine predictions. Although BP-Free TTA methods are computationally efficient, their image corruption adaptation capabilities are often limited. Backpropagation (BP)-Based TTA: This group encompasses certain pseudo-labeling methods (Zeng et al., 2024), consistency training methods (Sinha et al., 2023), and clustering-b...

work page 2024