Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

BayesTTA: Continual-Temporal Test-Time Adaptation for Vision-Language Models via Gaussian Discriminant Analysis

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Modeling each CLIP class as a drifting Gaussian lets a vision-language model adapt to gradual test-time drift without storing data.

desk verdict Real gains on gradual-shift benchmarks, but the 'adaptive' covariance test runs only once at t=1—worth a serious referee, not a desk reject. read the letter →

arxiv 2507.08607 v1 pith:OECYDTR5 submitted 2025-07-11 cs.CV

Shuang Cui , Jinglin Xu , Yi Li , Xiongxin Tang , Jiangmeng Li , Jiahuan Zhou , Fanjiang Xu , Fuchun Sun
show 1 more author
Hui Xiong
This is my paper
classification cs.CV
keywords continualtest-timeadaptationtemporaldistributionshiftvision-languagemodelsCLIPGaussiandiscriminantanalysisincrementalexpectation-maximizationcovariancehomogeneitytestnormalization-layer
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 aims to establish that gradual, temporally correlated shifts in test data—daylight fading, seasonal changes, sensor aging—need a different treatment from the sudden corruption shifts most continual test-time adaptation studies consider. It introduces a formal setting, CT-TTA, in which the test distribution drifts slowly under a bounded KL-divergence step, and proposes BayesTTA to operate there. The method models CLIP's visual embeddings per class as drifting Gaussian clouds, updates their means and covariances online with incremental EM without storing raw data, and uses Gaussian discriminant analysis to produce calibrated pseudo-labels that supervise lightweight updates of normalization layers only. On a new four-dataset CT-TTA benchmark and on ten standard TTA datasets, the paper reports consistent accuracy gains over prior methods, including a +14.27% improvement over the strongest baseline on RMNIST with ViT-B/16, while running about 32x faster than a leading prompt-tuning method.

What carries the argument

The load-bearing object is the online class-conditional Gaussian mixture on CLIP visual embeddings, $p_t(z\mid Y=k)=\mathcal{N}(\mu_t^k,\Sigma_t^k)$ (Eq. 4), with the GDA posterior $P_t(Y=k\mid z_i^t)\propto \mathcal{N}(z_i^t\mid\mu_t^k,\Sigma_t^k)\pi_t^k$ used for calibration. Incremental EM propagates the sufficient statistics $s_t^k$, $\mu_t^k$, $\Sigma_t^k$, and $\pi_t^k$ (Eqs. 19-24), so historical information survives without memory replay. A PCA-projected, F-corrected Box's M test (Eqs. 10-14) decides homogeneous versus heterogeneous covariance, hence LDA versus QDA discriminant scores (Eq. 25); those are fused with CLIP sketch logits as $\ell^{\text{adapt}}_k(z)=\ell^{\text{sketch}}_k(z)+\alpha D_k(z)$ (Eq. 27). Final pseudo-labels minimize soft cross-entropy (Eq. 29) over normalization-layer parameters only, with EMA smoothing (Eq. 30). The mechanism is that calibrated generative pseudo-labels replace entropy heuristics, sufficient statistics replace caches, and normalization-layer updates replace full-model tuning.

What would settle it

On fMoW and Yearbook, compute per-class CLIP feature embeddings and run a multivariate normality test based on skewness and kurtosis statistics for each temporal domain; if the Gaussian assumption is strongly rejected at many timesteps and BayesTTA still beats cache- and entropy-based baselines by the claimed margins, then the accuracy gains cannot be attributed to the GDA calibration and some other component, such as logit fusion or normalization-layer adaptation, is doing the work. Alternatively, run BayesTTA on a deliberately multimodal test stream, such as CIFAR10-C with interleaved corruption types, where the per-class feature densities are clearly non-Gaussian; a sharp collapse in accuracy would confirm the assumption is load-bearing.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that BayesTTA lets a frozen CLIP backbone track a changing test distribution by maintaining a time-varying generative model of its own embeddings. Each class is represented by a Gaussian with a drifting mean and covariance; incremental EM updates the sufficient statistics from each new batch without ever storing image data, which is what preserves long-range memory. A covariance homogeneity test, built on PCA and a regularized Box's M statistic, decides between one shared covariance (LDA) and per-class covariances (QDA), so the model is not locked into the identity-covariance assumption of earlier baselines. The resulting GDA discriminant scores are fused with CLIP's zero-shot text-similarity logits into pseudo-labels, and those pseudo-labels supervise a soft cross-entropy update applied only to LayerNorm or BatchNorm parameters, smoothed by an exponential moving average. The paper reports that this closed loop largely prevents error accumulation and forgetting: on fMoW, Yearbook, RMNIST, and CIFAR10-Light it consistently beats state-of-the-art CTTA and TTA baselines and stays stable across three consecutive long-term adaptation rounds.

Load-bearing premise

BayesTTA's gains rest on the premise that CLIP's visual embeddings for each class are reasonably close to a multivariate Gaussian, so the GDA posterior used to generate pseudo-labels remains calibrated; the paper cites prior support for this but provides no per-dataset normality diagnostics on the hardest domains, fMoW and Yearbook.

Editorial extensions

If this is right

  • On gradual-drift streams, cache-based and per-batch-reset adapters should underperform methods that carry long-range sufficient statistics, because memory is the binding constraint.
  • Entropy-based confidence is not the only viable supervision signal; a generative posterior fused with zero-shot logits can yield more reliable pseudo-labels under temporal drift.
  • A very small parameter footprint (normalization layers only, under 0.05% of ViT-B/16) suffices to realign VLM representations when the supervision is well-calibrated, making online adaptation practical on a single GPU.
  • The CT-TTA benchmark with weighted temporal-domain accuracy and multi-round evaluation provides a reproducible way to compare methods on long-horizon drift rather than single abrupt corruption.
  • The method's gains on ten standard TTA datasets imply the same Gaussian-mixture and GDA machinery transfers beyond the gradually-evolving setting to ordinary test-time adaptation.

Reading between the lines

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

  • Editorial extension: if the class-conditional Gaussian assumption is the real source of the gains, then datasets whose classes are strongly multimodal should expose it; a direct normality diagnostic on fMoW and Yearbook embeddings would tell whether the mechanism or the fusion is doing the work.
  • Editorial extension: because only normalization layers are updated, BayesTTA is complementary to prompt-based adaptation and could be stacked with textual prompt tuning rather than compared against it.
  • Editorial extension: the bounded-drift formulation suggests a theoretical route the paper does not take—bounding the regret of incremental EM under a KL drift budget and linking the covariance homogeneity test to model-selection guarantees.
  • Editorial extension: a natural testable extension is to apply BayesTTA to sudden-switch sequences such as CIFAR10-C, where the homogeneous or heterogeneous covariance test and the Gaussian assumption are least likely to hold; if it still holds up there, the method is more general than its motivation suggests.
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.

Referee Report

4 major / 6 minor

Summary. The paper introduces Continual-Temporal Test-Time Adaptation (CT-TTA), a formulation for gradually evolving distribution shifts, and proposes BayesTTA, which models CLIP visual embeddings as class-conditional Gaussians via incremental EM, adaptively selects covariance structures through PCA plus Box's M test, and uses Gaussian discriminant analysis (GDA) to produce pseudo-labels that supervise lightweight updates of normalization layers. The authors construct a CT-TTA benchmark from fMoW, Yearbook, RMNIST, and CIFAR10-Light, and also evaluate on ten standard TTA datasets, reporting consistent gains over prior TTA/CTTA methods and substantially lower inference cost than prompt-tuning baselines.

Significance. If the reported results hold, this is a useful contribution: it formalizes a realistic temporal-drift setting, provides a public benchmark and code, and demonstrates that a lightweight GDA-based adaptation loop can outperform prompt-tuning and cache-based baselines on gradual shifts. The paper is commendable for its breadth of experiments, ablations, and efficiency analysis. The central novelty is the claim of adaptive covariance-structure selection, and that claim is currently undermined by an implementation detail: the covariance homogeneity test is executed only at the first time step. The experimental support would also be substantially stronger with multi-seed error bars, given that the gains on fMoW and Yearbook are small. These issues are fixable within the manuscript's scope.

major comments (4)
  1. [Algorithm 1, lines 6–8; Section IV-D] The covariance homogeneity test is performed only when t=1 (Algorithm 1, lines 6–8), so the LDA/QDA structure is fixed after the first batch and never re-evaluated as the test distribution evolves. This contradicts the abstract and Section IV-D, which promise that BayesTTA "adaptively selects covariance structures through statistical hypothesis testing" and "aligns its generative assumptions with the empirical structure of incoming test data." The ablation in Table VI shows the test is load-bearing: removing it drops RMNIST accuracy from 75.07% to 17.58% and fMoW from 15.39% to 14.33%. To support the adaptive claim, the test should be re-run periodically or on a sliding window, or the paper must justify why a one-time decision suffices under temporal drift; the authors should also report which covariance structure is selected for each dataset and backbone and show that the decision is stable over time.
  2. [Section IV-D; Table I] On fMoW, the t=1 decision is made from a single batch of 128 images over 62 classes, i.e., roughly two soft samples per class. Box's M test relies on per-class covariance estimates, and with such small effective sample sizes the test outcome may be unstable or arbitrary even after PCA truncation to d=10. The paper provides no diagnostic for this. Please report per-class effective sample sizes and a sensitivity analysis (e.g., bootstrapping the first batch, varying the significance level κ, or varying the number of PCA components d) to demonstrate that the covariance-structure decision is statistically reliable and that the final accuracies do not hinge on a lucky first batch.
  3. [Tables I–IV] All main accuracy tables report single-run results without standard deviations or number of seeds. Some of the headline gains are small: on fMoW the improvement over the strongest baseline is +1.81% (ResNet-50) and +1.74% (ViT-B/16), and on Yearbook the gain over WATT is +0.92% (ResNet-50) and +0.21% (ViT-B/16). These margins are within plausible run-to-run noise for online adaptation with SGD and EMA updates. Please report mean±std over at least three seeds for the main CT-TTA comparisons, or specify why the pipeline is fully deterministic and seed-insensitive.
  4. [Section IV-B, Eq. (4)] The entire pseudo-labeling and adaptation loop rests on the assumption that CLIP visual embeddings follow class-conditional Gaussian mixtures. This assumption is cited from prior work but is not validated for the datasets where the method is claimed to work, especially fMoW and Yearbook, where the domain gap is severe and the Gaussianity premise is least obvious. The paper would be much stronger with per-dataset diagnostics: e.g., multivariate normality tests on a held-out subset, calibration curves of GDA posteriors against oracle labels, or a sensitivity study in which the Gaussian density in Eq. (4) is replaced by a heavier-tailed or nonparametric alternative to show that the gains are not an artifact of the Gaussian assumption.
minor comments (6)
  1. [Appendix references] The text refers to Appendix A for the derivation of Eq. (25) and Appendix B for dataset summaries and samples, but these appendices are not present in the submitted version; please include them.
  2. [Eq. (13)] The formula for d2 is ambiguous: "d2 = d1 + 2 λ + ε" could be read as d2 = d1 + 2λ + ε or d2 = d1 + 2/(λ+ε). Please clarify the intended expression, since it directly affects the F-distribution degrees of freedom used in the test.
  3. [Section V-D, Figure 7] The sensitivity analysis covers lr, β, and α, but two key hyperparameters of the covariance homogeneity test, d (number of PCA components) and κ (significance level), are fixed without any ablation. Please add experiments or justify these choices.
  4. [Section V-A] The experimental protocol does not specify how batches are sampled from each temporal domain (e.g., how many batches per year/angle/light level) and whether the model is evaluated on the same batch it just adapted on; a precise protocol description would improve reproducibility.
  5. [Table V] Results for DOTA, HisTPT, and BCA are reported "as presented in their original publications"; please confirm that those numbers were obtained with the same backbone, prompt template, batch size, and evaluation protocol used for the other baselines, or note any discrepancies.
  6. [Fig. 9] The efficiency comparison mixes batch size 128 for CLIP, WATT, and BayesTTA with batch size 1 for TPT, TDA, DPE, and DynaPrompt; the text acknowledges this, but the "32x faster" claim would be clearer with a per-image or per-epoch cost breakdown in addition to total wall-clock time.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the evaluation is external and the central Gaussian-mixture assumption is a stated modeling premise supported by external references, not a self-referential reduction.

full rationale

BayesTTA's central modeling premise is the class-conditional Gaussian assumption (Eq. 4, Sec. IV-B), introduced explicitly as an assumption and supported by external textbook and prior-method citations ([25], [28], [29]); it is not derived from the paper's own outputs. The GDA parameters are estimated online from unlabeled test features via incremental EM with soft responsibilities (Eqs. 18-22), which is a standard self-training loop in test-time adaptation and is not presented as a derivation of ground truth. Final accuracies are measured against held-out labels on fMoW, Yearbook, RMNIST, CIFAR10-Light, and ten standard TTA datasets, so the evaluation is external and not forced by construction. No equation reduces a predicted quantity to a fitted parameter or to a self-citation chain. One implementation detail—Algorithm 1 runs the covariance homogeneity test only at t=1 (lines 6-8)—weakens the claim of adaptive covariance selection and is a legitimate correctness/engineering concern, but it is not a circularity because the test decision is not the quantity being evaluated against ground truth. Accordingly, no step in the derivation chain is equivalent to its own inputs by definition.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the Gaussian mixture assumption for CLIP embeddings and a set of manually chosen hyperparameters, including the fusion weight, learning rate, EMA decay, regularization terms, and PCA dimension. No new physical or conceptual entities are introduced. The covariance homogeneity test and incremental EM are standard tools adapted to the temporal setting.

free parameters (7)
  • alpha (fusion weight, Eq. 27) = not stated; sensitivity analysis on CIFAR10-Light
    Balances CLIP sketch logits and GDA discriminant scores in the adapted logits; tuned on the benchmark dataset.
  • lr (learning rate for self-paced optimization) = not stated; sensitivity around 0.005
    Controls the update strength of normalization layers during adaptation.
  • beta (EMA decay rate, Eq. 30) = not stated; sensitivity around 0.99
    Smooths parameter updates over time and stabilizes adaptation.
  • sigma^2_prior (Eq. 23) = 0.1
    Fixed heuristic for ridge regularization of covariance estimates.
  • epsilon (Eq. 23) = not stated
    Ridge regularization strength to ensure positive definite covariance matrices.
  • d (number of PCA components, Section IV-D) = 10
    Dimension of the projected feature space used for the covariance homogeneity test.
  • kappa (significance level, Section IV-D) = 0.05
    Threshold for rejecting the null hypothesis of homogeneous covariance in Box's M test.
assumptions (5)
  • domain assumption Class-conditional visual embeddings follow Gaussian mixture distributions (Eq. 4, Section IV-B).
    This is the core modeling assumption that makes GDA-based inference valid. It is adopted from prior work [25], [28] but not verified on the CT-TTA benchmark datasets.
  • domain assumption Bounded KL drift between consecutive test distributions (Eq. 2, Section IV-A).
    Formalizes the temporal smoothness of CT-TTA but is never used in the algorithm; it only motivates the setting.
  • standard math Box's M test with F-approximation is valid in the small-sample, high-dimensional PCA subspace.
    The paper invokes the F-corrected version of Box's M test [63] but does not validate the approximation on the small per-class sample sizes encountered during early test-time batches.
  • domain assumption Online EM converges and tracks non-stationary distributions under gradual drift (Section IV-E).
    The paper cites Cappe and Moulines [64] for online EM behavior, but the assumptions of that theory may not hold under the class-label uncertainty and drifting feature distributions in CT-TTA.
  • ad hoc to paper PCA truncation to d=10 components preserves the class-discriminative structure needed for GDA.
    The choice of 10 components is arbitrary and not justified by an eigenvalue analysis or a reconstruction error criterion; the covariance homogeneity test is conducted in this reduced space rather than the original feature space.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BayesTTA: Continual-Temporal Test-Time Adaptation for Vision-Language Models via Gaussian Discriminant Analysis." pith.science (2026). https://pith.science/paper/OECYDTR5

@misc{pith2026250708607,
  author       = {Pith},
  title        = {Pith review of: BayesTTA: Continual-Temporal Test-Time Adaptation for Vision-Language Models via Gaussian Discriminant Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OECYDTR5}},
  note         = {Machine review of arXiv:2507.08607}
}
read the original abstract

Vision-language models (VLMs) such as CLIP achieve strong zero-shot recognition but degrade significantly under \textit{temporally evolving distribution shifts} common in real-world scenarios (e.g., gradual illumination or seasonal changes). Existing continual test-time adaptation (CTTA) methods are typically built around sudden and severe distribution shifts and neglect temporal continuity, leading to three core defects: limited memory cache restricts long-range distribution modeling, causing catastrophic forgetting; entropy-based confidence becomes unreliable under temporal drift, worsening error accumulation; and static visual representations misalign with evolving inputs. We formalize this practical problem as \textit{Continual-Temporal Test-Time Adaptation (CT-TTA)}, where test distributions evolve gradually over time. To address it, we propose \textit{BayesTTA}, a Bayesian adaptation framework that enforces temporally consistent predictions and dynamically aligns visual representations. Specifically, BayesTTA incrementally estimates class-conditional Gaussian mixture distributions without storing raw data, adaptively selects covariance structures through statistical hypothesis testing, and performs calibrated inference using Gaussian discriminant analysis (GDA). These calibrated predictions supervise self-paced adaptation of normalization layers, ensuring efficient and stable representation alignment. We establish a comprehensive CT-TTA benchmark across four temporally evolving datasets and further evaluate generalization on ten standard TTA datasets. Extensive experiments show that BayesTTA consistently outperforms state-of-the-art methods, achieving significant gains while maintaining efficiency. Code is available at \href{https://github.com/cuishuang99/BayesTTA}{https://github.com/cuishuang99/BayesTTA}.

Figures

Figures reproduced from arXiv: 2507.08607 by the authors.

Figure 1
Figure 1. Test-time adaptation paradigms: conventional CTTA [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Empirical motivation for CT-TTA. Left: per-domain [PITH_FULL_IMAGE:figures/full_fig_p001_2.png] view at source ↗
Figure 3
Figure 3. Overview of BayesTTA: (1) Zero-shot Sketch Generation, (2) Covariance Homogeneity Test, (3) Debiased Temporal [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Performance comparisons in long-term CT-TTA adaptation. The horizontal axis represents the round number, and the [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: t-SNE visualization of BayesTTA on RMNIST across temporal test steps. Top: As digit rotation gradually evolves from [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 8
Figure 8. Figure 8: The impact of batch size. CLIP-ViT-B/16TPT TDAWATTDPE DynaPrompt BayesTTA 0 5000 10000 15000 20000 80 85 90 95 150 9078 593 18026 2730 13747 150 285 9078 593 18026 2730 13747 285 Time (s) Acc (%) Time Accuracy [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 7
Figure 7. Figure 7: Hyperparameter sensitivity analysis of BayesTTA. [PITH_FULL_IMAGE:figures/full_fig_p013_7.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

  1. Adapting Vision-Language Models Without Labels: A Comprehensive Survey

    cs.LG 2025-08 conditional novelty 5.0 of 10

    A survey that organizes unsupervised vision-language model adaptation by unlabeled-data availability into four paradigms: data-free transfer, domain transfer, episodic test-time, and online test-time adaptation.

Reference graph

Works this paper leans on

86 extracted references · 61 canonical work pages · cited by 1 Pith paper

  1. [1]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al., “Learning transferable visual models from natural language supervision,” in International conference on machine learning . PMLR, 2021, pp. 8748–8763

    2021

  2. [2]

    Scaling up visual and vision-language representation learning with noisy text supervision,

    C. Jia, Y . Yang, Y . Xia, Y .-T. Chen, Z. Parekh, H. Pham, Q. Le, Y .-H. Sung, Z. Li, and T. Duerig, “Scaling up visual and vision-language representation learning with noisy text supervision,” in International conference on machine learning . PMLR, 2021, pp. 4904–4916

    2021

  3. [3]

    Laion-400m: Open dataset of clip-filtered 400 million image-text pairs,

    C. Schuhmann, R. Vencu, R. Beaumont, R. Kaczmarczyk, C. Mullis, A. Katta, T. Coombes, J. Jitsev, and A. Komatsuzaki, “Laion-400m: Open dataset of clip-filtered 400 million image-text pairs,”arXiv preprint arXiv:2111.02114, 2021

    arXiv 2021

  4. [4]

    Laion- 5b: An open large-scale dataset for training next generation image-text models,

    C. Schuhmann, R. Beaumont, R. Vencu, C. Gordon, R. Wightman, M. Cherti, T. Coombes, A. Katta, C. Mullis, M. Wortsmanet al., “Laion- 5b: An open large-scale dataset for training next generation image-text models,” Advances in Neural Information Processing Systems , vol. 35, pp. 25 278–25 294, 2022

    2022

  5. [5]

    Learning to prompt for vision- language models,

    K. Zhou, J. Yang, C. C. Loy, and Z. Liu, “Learning to prompt for vision- language models,” International Journal of Computer Vision , vol. 130, no. 9, pp. 2337–2348, 2022

    2022

  6. [6]

    Vision-language models for vision tasks: A survey,

    J. Zhang, J. Huang, S. Jin, and S. Lu, “Vision-language models for vision tasks: A survey,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  7. [7]

    Tent: Fully test-time adaptation by entropy minimization,

    D. Wang, E. Shelhamer, S. Liu, B. Olshausen, and T. Darrell, “Tent: Fully test-time adaptation by entropy minimization,” in International Conference on Learning Representations , 2021

    2021

  8. [8]

    Continual test-time domain adaptation,

    Q. Wang, O. Fink, L. Van Gool, and D. Dai, “Continual test-time domain adaptation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 7201–7211

    work page 2022

Show all 86 references
  1. [9]

    Towards stable test-time adaptation in dynamic wild world,

    S. Niu, J. Wu, Y . Zhang, Z. Wen, Y . Chen, P. Zhao, and M. Tan, “Towards stable test-time adaptation in dynamic wild world,” in The Eleventh International Conference on Learning Representations , 2023

    2023

  2. [10]

    Robust test-time adaptation in dynamic scenarios,

    L. Yuan, B. Xie, and S. Li, “Robust test-time adaptation in dynamic scenarios,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 15 922–15 932

    2023

  3. [11]

    Uncertainty-calibrated test-time model adaptation without forgetting,

    M. Tan, G. Chen, J. Wu, Y . Zhang, Y . Chen, P. Zhao, and S. Niu, “Uncertainty-calibrated test-time model adaptation without forgetting,” IEEE Transactions on Pattern Analysis and Machine Intelligence , 2025

    2025

  4. [12]

    A comprehensive survey on test-time adaptation under distribution shifts,

    J. Liang, R. He, and T. Tan, “A comprehensive survey on test-time adaptation under distribution shifts,” International Journal of Computer Vision, vol. 133, no. 1, pp. 31–64, 2025

    2025

  5. [13]

    Historical test- time prompt tuning for vision foundation models,

    J. Zhang, J. Huang, X. Zhang, L. Shao, and S. Lu, “Historical test- time prompt tuning for vision foundation models,” in The Thirty-eighth Annual Conference on Neural Information Processing Systems , 2024

    2024

  6. [14]

    Dynaprompt: Dynamic test-time prompt tuning,

    Z. Xiao, S. Yan, J. Hong, J. Cai, X. Jiang, Y . Hu, J. Shen, C. Wang, and C. G. M. Snoek, “Dynaprompt: Dynamic test-time prompt tuning,” in The Thirteenth International Conference on Learning Representations , 2025

    2025

  7. [15]

    Benchmarking neural network ro- bustness to common corruptions and perturbations,

    D. Hendrycks and T. Dietterich, “Benchmarking neural network ro- bustness to common corruptions and perturbations,” in International Conference on Learning Representations , 2019

    2019

  8. [16]

    Chemical gas sensor drift compensation using classifier ensembles,

    A. Vergara, S. Vembu, T. Ayhan, M. A. Ryan, M. L. Homer, and R. Huerta, “Chemical gas sensor drift compensation using classifier ensembles,” Sensors and Actuators B: Chemical , vol. 166, pp. 320–329, 2012

    2012

  9. [17]

    Continuous manifold based adaptation for evolving visual domains,

    J. Hoffman, T. Darrell, and K. Saenko, “Continuous manifold based adaptation for evolving visual domains,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2014, pp. 867– 874

    2014

  10. [18]

    Wild- time: A benchmark of in-the-wild distribution shift over time,

    H. Yao, C. Choi, B. Cao, Y . Lee, P. W. W. Koh, and C. Finn, “Wild- time: A benchmark of in-the-wild distribution shift over time,” Advances in Neural Information Processing Systems , vol. 35, pp. 10 309–10 324, 2022

    2022

  11. [19]

    Evolving standardization for continual domain generalization over temporal drift,

    M. Xie, S. Li, L. Yuan, C. Liu, and Z. Dai, “Evolving standardization for continual domain generalization over temporal drift,” Advances in Neural Information Processing Systems , vol. 36, pp. 21 983–22 002, 2023

    2023

  12. [20]

    Foresee what you will learn: data augmentation for domain generalization in non- stationary environment,

    Q. Zeng, W. Wang, F. Zhou, C. Ling, and B. Wang, “Foresee what you will learn: data augmentation for domain generalization in non- stationary environment,” in Proceedings of the AAAI conference on artificial intelligence, vol. 37, no. 9, 2023, pp. 11 147–11 155

    2023

  13. [21]

    E. O. Brigham, The fast Fourier transform and its applications . Prentice-Hall, Inc., 1988

    1988

  14. [22]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, G. Hinton et al. , “Learning multiple layers of features from tiny images,” Citeseer, Toronto, ON, Canada, Technical Report, 2009

    2009

  15. [23]

    Efficient test-time adaptation of vision-language models,

    A. Karmanov, D. Guan, S. Lu, A. El Saddik, and E. Xing, “Efficient test-time adaptation of vision-language models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 14 162–14 171

    2024

  16. [24]

    Hastie, R

    T. Hastie, R. Tibshirani, and J. Friedman, The Elements of Statistical Learning: Data Mining, Inference, and Prediction , 2nd ed. Springer, 2009

    2009

  17. [25]

    C. M. Bishop and N. M. Nasrabadi, Pattern recognition and machine learning. Springer, 2006, vol. 4, no. 4. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 15

    2006

  18. [26]

    I. T. Jolliffe, Principal component analysis for special types of data . Springer, 2002

    2002

  19. [27]

    A general distribution theory for a class of likelihood criteria,

    G. E. Box, “A general distribution theory for a class of likelihood criteria,” Biometrika, vol. 36, no. 3/4, pp. 317–346, 1949

    1949

  20. [28]

    A hard- to-beat baseline for training-free clip-based adaptation,

    Z. Wang, J. Liang, L. Sheng, R. He, Z. Wang, and T. Tan, “A hard- to-beat baseline for training-free clip-based adaptation,” in The Twelfth International Conference on Learning Representations , 2024

    2024

  21. [29]

    Dota: Distributional test-time adaptation of vision-language models,

    Z. Han, J. Yang, J. Li, Q. Hu, Q. Xu, M. Z. Shou, and C. Zhang, “Dota: Distributional test-time adaptation of vision-language models,” arXiv preprint arXiv:2409.19375 , 2024

    2024

  22. [30]

    Free on the fly: Enhancing flexibility in test-time adaptation with online em,

    Q. Dai and S. Yang, “Free on the fly: Enhancing flexibility in test-time adaptation with online em,” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 9538–9548

    2025

  23. [31]

    Cris: Clip-driven referring image segmentation,

    Z. Wang, Y . Lu, Q. Li, X. Tao, Y . Guo, M. Gong, and T. Liu, “Cris: Clip-driven referring image segmentation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 11 686–11 695

    2022

  24. [32]

    Zegclip: Towards adapting clip for zero-shot semantic segmentation,

    Z. Zhou, Y . Lei, B. Zhang, L. Liu, and Y . Liu, “Zegclip: Towards adapting clip for zero-shot semantic segmentation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 11 175–11 185

    2023

  25. [33]

    Actionclip: Adapting language-image pretrained models for video action recognition,

    M. Wang, J. Xing, J. Mei, Y . Liu, and Y . Jiang, “Actionclip: Adapting language-image pretrained models for video action recognition,” IEEE Transactions on Neural Networks and Learning Systems , 2023

    2023

  26. [34]

    Clip-guided prototype modulating for few-shot action recognition,

    X. Wang, S. Zhang, J. Cen, C. Gao, Y . Zhang, D. Zhao, and N. Sang, “Clip-guided prototype modulating for few-shot action recognition,” International Journal of Computer Vision , vol. 132, no. 6, pp. 1899– 1912, 2024

    1912

  27. [35]

    Conditional prompt learning for vision-language models,

    K. Zhou, J. Yang, C. C. Loy, and Z. Liu, “Conditional prompt learning for vision-language models,” in Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) , June 2022, pp. 16 816–16 825

    2022

  28. [36]

    Maple: Multi-modal prompt learning,

    M. U. Khattak, H. Rasheed, M. Maaz, S. Khan, and F. S. Khan, “Maple: Multi-modal prompt learning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2023, pp. 19 113–19 122

    2023

  29. [37]

    Clip-adapter: Better vision-language models with feature adapters,

    P. Gao, S. Geng, R. Zhang, T. Ma, R. Fang, Y . Zhang, H. Li, and Y . Qiao, “Clip-adapter: Better vision-language models with feature adapters,” International Journal of Computer Vision , vol. 132, no. 2, pp. 581–595, 2024

    2024

  30. [38]

    Tip-adapter: Training-free adaption of clip for few-shot classification,

    R. Zhang, W. Zhang, R. Fang, P. Gao, K. Li, J. Dai, Y . Qiao, and H. Li, “Tip-adapter: Training-free adaption of clip for few-shot classification,” in European conference on computer vision . Springer, 2022, pp. 493– 510

    2022

  31. [39]

    Delta: Degradation-free fully test- time adaptation,

    B. Zhao, C. Chen, and S.-T. Xia, “Delta: Degradation-free fully test- time adaptation,” in The Eleventh International Conference on Learning Representations, 2022

    2022

  32. [40]

    Memo: Test time robustness via adaptation and augmentation,

    M. Zhang, S. Levine, and C. Finn, “Memo: Test time robustness via adaptation and augmentation,” Advances in Neural Information Processing Systems, vol. 35, pp. 38 629–38 642, 2022

    2022

  33. [41]

    Ef- ficient test-time model adaptation without forgetting,

    S. Niu, J. Wu, Y . Zhang, Y . Chen, S. Zheng, P. Zhao, and M. Tan, “Ef- ficient test-time model adaptation without forgetting,” in International conference on machine learning . PMLR, 2022, pp. 16 888–16 905

    2022

  34. [42]

    Domainadaptor: A novel approach to test-time adaptation,

    J. Zhang, L. Qi, Y . Shi, and Y . Gao, “Domainadaptor: A novel approach to test-time adaptation,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 18 971–18 981

    2023

  35. [43]

    Test time adaptation for blind image quality assessment,

    S. Roy, S. Mitra, S. Biswas, and R. Soundararajan, “Test time adaptation for blind image quality assessment,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 16 742–16 751

    2023

  36. [44]

    Entropy is not enough for test-time adaptation: From the perspective of disentangled factors,

    J. Lee, D. Jung, S. Lee, J. Park, J. Shin, U. Hwang, and S. Yoon, “Entropy is not enough for test-time adaptation: From the perspective of disentangled factors,” in The Twelfth International Conference on Learning Representations, 2024

    2024

  37. [45]

    Test-time prompt tuning for zero-shot generalization in vision-language models,

    M. Shu, W. Nie, D.-A. Huang, Z. Yu, T. Goldstein, A. Anandkumar, and C. Xiao, “Test-time prompt tuning for zero-shot generalization in vision-language models,” Advances in Neural Information Processing Systems, vol. 35, pp. 14 274–14 289, 2022

    2022

  38. [46]

    Robust test-time adaptation for zero-shot prompt tuning,

    D.-C. Zhang, Z. Zhou, and Y .-F. Li, “Robust test-time adaptation for zero-shot prompt tuning,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 15, 2024, pp. 16 714–16 722

    2024

  39. [47]

    Towards robustness prompt tuning with fully test-time adaptation for clip’s zero-shot generalization,

    R. Wang, H. Zuo, Z. Fang, and J. Lu, “Towards robustness prompt tuning with fully test-time adaptation for clip’s zero-shot generalization,” in Proceedings of the 32nd ACM International Conference on Multimedia , 2024, pp. 8604–8612

    2024

  40. [48]

    W ATT: Weight average test time adaptation of CLIP,

    D. OSOWIECHI, M. Noori, G. A. V . Hakim, M. Yazdanpanah, A. Bahri, M. Cheraghalikhani, S. Dastani, F. Beizaee, I. B. Ayed, and C. Desrosiers, “W ATT: Weight average test time adaptation of CLIP,” in The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  41. [49]

    Dual prototype evolving for test-time generalization of vision-language models,

    C. Zhang, S. Stepputtis, K. Sycara, and Y . Xie, “Dual prototype evolving for test-time generalization of vision-language models,” Advances in Neural Information Processing Systems , vol. 37, pp. 32 111–32 136, 2024

    2024

  42. [50]

    Bayesian test-time adaptation for vision-language models,

    L. Zhou, M. Ye, S. Li, N. Li, X. Zhu, L. Deng, H. Liu, and Z. Lei, “Bayesian test-time adaptation for vision-language models,” in Proceed- ings of the Computer Vision and Pattern Recognition Conference , 2025, pp. 29 999–30 009

    2025

  43. [51]

    Robust mean teacher for continual and gradual test-time adaptation,

    M. D ¨obler, R. A. Marsden, and B. Yang, “Robust mean teacher for continual and gradual test-time adaptation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 7704–7714

    2023

  44. [52]

    Decorate the newcomers: Visual domain prompt for continual test time adaptation,

    Y . Gan, Y . Bai, Y . Lou, X. Ma, R. Zhang, N. Shi, and L. Luo, “Decorate the newcomers: Visual domain prompt for continual test time adaptation,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 6, 2023, pp. 7595–7603

    2023

  45. [53]

    Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization,

    J. Song, J. Lee, I. S. Kweon, and S. Choi, “Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization,” in Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 11 920–11 929

    2023

  46. [54]

    ViDA: Homeostatic visual domain adapter for continual test time adaptation,

    J. Liu, S. Yang, P. Jia, R. Zhang, M. Lu, Y . Guo, W. Xue, and S. Zhang, “ViDA: Homeostatic visual domain adapter for continual test time adaptation,” in The Twelfth International Conference on Learning Representations, 2024

    2024

  47. [55]

    Dynamic domains, dynamic solutions: Dpcore for continual test-time adaptation,

    Y . Zhang, A. Mehra, and J. Hamm, “Dynamic domains, dynamic solutions: Dpcore for continual test-time adaptation,” arXiv preprint arXiv:2406.10737, 2024

    2024 arXiv

  48. [56]

    Training for the future: A simple gradient interpolation loss to generalize along time,

    A. Nasery, S. Thakur, V . Piratla, A. De, and S. Sarawagi, “Training for the future: A simple gradient interpolation loss to generalize along time,” Advances in Neural Information Processing Systems , vol. 34, pp. 19 198–19 209, 2021

    2021

  49. [57]

    Temporal domain generalization with drift-aware dynamic neural networks,

    G. Bai, C. Ling, and L. Zhao, “Temporal domain generalization with drift-aware dynamic neural networks,” in The Eleventh International Conference on Learning Representations , 2023

    2023

  50. [58]

    Generalizing to evolving domains with latent structure-aware sequential autoencoder,

    T. Qin, S. Wang, and H. Li, “Generalizing to evolving domains with latent structure-aware sequential autoencoder,” in International Confer- ence on Machine Learning . PMLR, 2022, pp. 18 062–18 082

    2022

  51. [59]

    Enhancing evolving domain generalization through dynamic latent representations,

    B. Xie, Y . Chen, J. Wang, K. Zhou, B. Han, W. Meng, and J. Cheng, “Enhancing evolving domain generalization through dynamic latent representations,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 14, 2024, pp. 16 040–16 048

    2024

  52. [60]

    Latent trajectory learning for limited timestamps under distribution shift over time,

    Q. Zeng, C. Shui, L.-K. Huang, P. Liu, X. Chen, C. Ling, and B. Wang, “Latent trajectory learning for limited timestamps under distribution shift over time,” in The Twelfth International Conference on Learning Representations, 2024

    2024

  53. [61]

    Weight diffusion for future: Learn to generalize in non-stationary environments,

    M. Xie, S. Li, B. Xie, C. Liu, J. Liang, Z. Sun, K. Feng, and C. Zhu, “Weight diffusion for future: Learn to generalize in non-stationary environments,” Advances in Neural Information Processing Systems , vol. 37, pp. 6367–6392, 2024

    2024

  54. [62]

    Kullback, Information theory and statistics

    S. Kullback, Information theory and statistics . Courier Corporation, 1997

    1997

  55. [63]

    An extension of box’s results on the use of the f distribution in multivariate analysis,

    S. Geisser and S. W. Greenhouse, “An extension of box’s results on the use of the f distribution in multivariate analysis,” The Annals of Mathematical Statistics, pp. 885–891, 1958

    1958

  56. [64]

    On-line expectation–maximization algo- rithm for latent data models,

    O. Capp ´e and E. Moulines, “On-line expectation–maximization algo- rithm for latent data models,” Journal of the Royal Statistical Society Series B: Statistical Methodology , vol. 71, no. 3, pp. 593–613, 2009

    2009

  57. [65]

    A generalized inverse for matrices,

    R. Penrose, “A generalized inverse for matrices,” in Mathematical proceedings of the Cambridge philosophical society , vol. 51, no. 3. Cambridge University Press, 1955, pp. 406–413

    1955

  58. [66]

    Layer normalization,

    J. L. Ba, J. R. Kiros, and G. E. Hinton, “Layer normalization,” arXiv preprint arXiv:1607.06450, 2016

    2016 arXiv

  59. [67]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift,

    S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” in International conference on machine learning . pmlr, 2015, pp. 448–456

    2015

  60. [68]

    Acceleration of stochastic approxima- tion by averaging,

    B. T. Polyak and A. B. Juditsky, “Acceleration of stochastic approxima- tion by averaging,” SIAM journal on control and optimization , vol. 30, no. 4, pp. 838–855, 1992

    1992

  61. [69]

    A century of portraits: A visual historical record of american high school yearbooks,

    S. Ginosar, K. Rakelly, S. Sachs, B. Yin, and A. A. Efros, “A century of portraits: A visual historical record of american high school yearbooks,” in Proceedings of the IEEE International Conference on Computer Vision Workshops, 2015, pp. 1–7

    2015

  62. [70]

    The mnist database of handwritten digit images for machine learning research [best of the web],

    L. Deng, “The mnist database of handwritten digit images for machine learning research [best of the web],” IEEE signal processing magazine , vol. 29, no. 6, pp. 141–142, 2012. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 16

    2012

  63. [71]

    Fine- grained visual classification of aircraft,

    S. Maji, E. Rahtu, J. Kannala, M. Blaschko, and A. Vedaldi, “Fine- grained visual classification of aircraft,” arXiv preprint arXiv:1306.5151, 2013

    2013 arXiv

  64. [72]

    Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories,

    L. Fei-Fei, R. Fergus, and P. Perona, “Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories,” in 2004 conference on computer vision and pattern recognition workshop . IEEE, 2004, pp. 178–178

    2004

  65. [73]

    3d object representations for fine-grained categorization,

    J. Krause, M. Stark, J. Deng, and L. Fei-Fei, “3d object representations for fine-grained categorization,” in Proceedings of the IEEE interna- tional conference on computer vision workshops , 2013, pp. 554–561

    2013

  66. [74]

    Describing textures in the wild,

    M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi, “Describing textures in the wild,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2014, pp. 3606–3613

    2014

  67. [75]

    Eurosat: A novel dataset and deep learning benchmark for land use and land cover classi- fication,

    P. Helber, B. Bischke, A. Dengel, and D. Borth, “Eurosat: A novel dataset and deep learning benchmark for land use and land cover classi- fication,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , vol. 12, no. 7, pp. 2217–2226, 2019

    2019

  68. [76]

    Automated flower classification over a large number of classes,

    M.-E. Nilsback and A. Zisserman, “Automated flower classification over a large number of classes,” in 2008 Sixth Indian conference on computer vision, graphics & image processing . IEEE, 2008, pp. 722–729

    2008

  69. [77]

    Food-101–mining discriminative components with random forests,

    L. Bossard, M. Guillaumin, and L. Van Gool, “Food-101–mining discriminative components with random forests,” in Computer vision– ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13 . Springer, 2014, pp. 446–461

    2014

  70. [78]

    Cats and dogs,

    O. M. Parkhi, A. Vedaldi, A. Zisserman, and C. Jawahar, “Cats and dogs,” in 2012 IEEE conference on computer vision and pattern recog- nition. IEEE, 2012, pp. 3498–3505

    2012

  71. [79]

    Sun database: Large-scale scene recognition from abbey to zoo,

    J. Xiao, J. Hays, K. A. Ehinger, A. Oliva, and A. Torralba, “Sun database: Large-scale scene recognition from abbey to zoo,” in 2010 IEEE computer society conference on computer vision and pattern recognition. IEEE, 2010, pp. 3485–3492

    2010

  72. [80]

    Ucf101: A dataset of 101 human actions classes from videos in the wild,

    K. Soomro, A. R. Zamir, and M. Shah, “Ucf101: A dataset of 101 human actions classes from videos in the wild,”arXiv preprint arXiv:1212.0402, 2012

    2012 arXiv

  73. [81]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770–778

    2016

  74. [82]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly et al., “An image is worth 16x16 words: Transformers for image recognition at scale,” in International Conference on Learning Representations , 2021

    2021

  75. [83]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in Neural Information Processing Systems , vol. 30, 2017

    2017

  76. [84]

    Visualizing data using t-sne

    L. Van der Maaten and G. Hinton, “Visualizing data using t-sne.” Journal of machine learning research , vol. 9, no. 11, 2008

    2008

  77. [85]

    Grad-cam: Visual explanations from deep networks via gradient-based localization,

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-cam: Visual explanations from deep networks via gradient-based localization,” in Proceedings of the IEEE international conference on computer vision , 2017, pp. 618–626. Shuang Cui received the B...

    2017

  78. [2013]

    With extensive experience in optical simulation calculations and optical software devel- opment, she has actively engaged in research and development in this field

    She currently holds the position of vice pro- fessor at the Institute of Software, Chinese Academy of Sciences. With extensive experience in optical simulation calculations and optical software devel- opment, she has actively engaged in research and development in this field. ...

    2013

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.