Pith. sign in

REVIEW 3 major objections 4 minor 74 references

PAID: Pairwise Angular-Invariant Decomposition for Continual Test-Time Adaptation

T0 review · 3 major / 4 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 PAID claims that the pairwise angular structure of pretrained weights is a domain-invariant semantic prior, and preserves it during continual test-time adaptation by learning only magnitudes and an orthogonal rotation.

desk verdict Solid empirical CTTA results undermined by a false invariance claim: right multiplication by O does not preserve the pairwise angles the paper says it preserves. read the letter →

arxiv 2506.02453 v2 pith:LGPTCAUU submitted 2025-06-03 cs.CV

Kunyu Wang , Xueyang Fu , Yuanfei Bao , Chengjie Ge , Chengzhi Cao , Wei Zhai , Zheng-Jun Zha This is my paper
classification cs.CV
keywords continualtest-timeadaptationpairwiseangularstructureweightdecompositionHouseholderreflectionorthogonalmatrixdomain-invariantpriorvisiontransformer
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 argues that the pairwise angular structure of pretrained weight matrices is a domain-invariant semantic prior: across the fifteen corruptions of ImageNet-C it stays nearly constant, whereas magnitude and absolute angle drift. To exploit this, PAID decomposes each weight as $W = M \odot \hat{W}$ into a magnitude row and a unit-norm direction matrix, freezes $\hat{W}$, and learns only $M$ and an orthogonal matrix built from Householder reflections, so the updated weight is $M' \odot (\hat{W} O)$. On four continual test-time adaptation benchmarks, including ImageNet-C, CIFAR-100-C, and Cityscapes-to-ACDC, PAID reports lower error rates (or higher mIoU) than prior methods, making the case that preserving pairwise angles is a simple and effective principle for CTTA.

What carries the argument

The central object is the decomposition $W = M \odot \hat{W}$, where $M$ scales each column and $\hat{W}$ holds unit-norm direction vectors, together with a learnable orthogonal matrix $O \in \mathbb{R}^{k \times k}$ constructed as a chain of Householder reflections, $O = \prod_i (I - 2u_i u_i^\top)$ with $\|u_i\|=1$. A Householder reflection is the orthogonal reflection of a vector across a hyperplane. The paper's mechanism is to freeze $\hat{W}$ and learn only $M$ and $O$, so that adaptation is restricted to magnitude changes and a global rotation; the claim is that this rotation preserves the pairwise angular structure of the direction columns while still allowing the absolute orientation to adapt.

What would settle it

Pick any non-orthonormal direction matrix $\hat{W}$, multiply it on the right by the learned orthogonal matrix $O$, and compare the Gram matrix $G = \hat{W}^\top \hat{W}$ before and after: the new Gram matrix is $O^\top G O$, whose off-diagonal entries generally differ from $G$. If they differ, the pairwise angles have changed, and the claimed angular invariance is not what the method actually enforces.

Watch

Extended reading notes

Core claim

The paper's central claim is that the pairwise angular structure of pretrained weights encodes semantic information invariant to domain shift: it remains stable across the 15 corrupted domains of ImageNet-C, but changes when the model is fine-tuned on a semantically different dataset such as CIFAR-100. From this, the authors conclude that continual test-time adaptation should preserve that structure. PAID implements the conclusion by decomposing every linear weight into magnitude and direction, freezing the direction matrix, and introducing a learnable orthogonal matrix, parameterized as a product of Householder reflections, that rotates all directions together. The updated weights take the form $M' \odot (\hat{W} O)$, with only $M$ and $O$ learnable. The experiments report consistent improvements over previous state-of-the-art methods on four benchmarks, which the paper takes as evidence that the angular-geometry prior is a useful design principle for CTTA.

Load-bearing premise

Everything rests on the claim that rotating all weight directions together with an orthogonal matrix keeps the angles between every pair of directions unchanged, so freezing the direction matrix really does preserve the semantic structure the paper identifies.

Editorial extensions

If this is right

  • Only magnitudes and the orthogonal matrix need to be updated; the frozen direction matrix is enough for state-of-the-art CTTA performance.
  • Preventing arbitrary per-vector rotation appears to reduce catastrophic forgetting and error accumulation in long adaptation sequences, with the paper reporting stable 10-round results.
  • The method needs only about 500 source images to estimate the statistics used in the loss, and stays competitive for batch sizes as small as 4.
  • The number of Householder reflections, r, controls a capacity-stability trade-off, with r = 12 working best across benchmarks.

Reading between the lines

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

  • As written, the mathematics does not actually preserve the pairwise column angles: right multiplication by $O$ sends the Gram matrix to $O^\top G O$, which is not equal to $G$ for a generic orthogonal $O$ and non-orthonormal column set. A version that left-multiplies directions by a $d \times d$ orthogonal matrix would genuinely preserve the angles, and comparing the two would separate the geometr
  • The reported performance could alternatively come from the learnable input transform induced by $\hat{W} O$ or from magnitude scaling, rather than from true angular preservation; a provably angle-preserving parameterization would separate these explanations.
  • The paper's Fourier analogy suggests a broader principle: separating a 'phase-like' relational structure from an 'amplitude-like' magnitude, and adapting only the latter, could apply to other layer types and parameter spaces such as recurrent weights.
  • Since the pairwise-angle prior is defined purely from the source weights, PAID could be paired with other adaptation objectives; whether the gains persist under entropy minimization or pseudo-labeling is untested.
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

3 major / 4 minor

Summary. The paper addresses continual test-time adaptation (CTTA) by exploiting geometric structure in pretrained weights. The authors decompose each linear layer weight W into a magnitude M and a unit-direction matrix \hat W, argue from three motivation experiments that the pairwise angular structure of the columns of \hat W is domain-invariant, and propose PAID: freeze \hat W and update only M together with a learnable orthogonal matrix constructed from Householder reflections. The paper reports state-of-the-art results on four CTTA benchmarks. The central mathematical claim, however, is that the update \hat W \mapsto \hat W O with O \in R^{k\times k} preserves the pairwise column angles of \hat W; this claim is false, and the motivation experiments depend on the same questionable update. As a result, the core premise as implemented is not established.

Significance. If the central claim were correct, the idea that pretrained weights carry a domain-invariant pairwise angular prior would be a genuinely interesting and potentially useful principle for CTTA, especially in a parameter-efficient setting. The paper has several strengths: it covers a broad set of benchmarks (ImageNet-C, CIFAR-10/100-C, Cityscapes-to-ACDC), reports consistent gains over strong baselines, and includes detailed ablations and an analysis of source-data requirements. However, the significance rests entirely on the assertion that the proposed update actually preserves the pairwise angular structure of neuron columns. Since that assertion is false for the operation defined in Eq. (5), the method as described does not implement the stated prior, and the reported empirical results cannot be attributed to the claimed geometric invariance. The contribution is therefore not currently established.

major comments (3)
  1. [Section 3.2, Eqs. (5)–(8)] The update in Eq. (5), \hat W \mapsto \hat W O with O \in R^{k\times k}, does not preserve the pairwise angles between the columns of \hat W. The column Gram matrix transforms as G = \hat W^\top \hat W \mapsto O^\top G O, which equals G for every orthogonal O only if G is a scalar multiple of the identity; neither pretrained neuron columns nor the trained \hat W satisfy or are constrained to satisfy that condition. The inner-product preservation quoted in Eq. (7), \langle Ox, Oy \rangle = \langle x, y \rangle, applies to O acting on vectors in R^k, whereas the columns of \hat W lie in R^d and are not individually multiplied by O: each column of \hat W O is a linear combination of the original columns. A concrete two-column example shows the angle changing: for \hat W = [[1, 1/\sqrt{2}], [0, 1/\sqrt{2}]] and a 45-degree rotation O in R^2, the column angle changes from 45 degrees to 90 degrees. What right multiplication actually preserves is the row Gram matrix \hat W \hat W^\top, i.e., pairwise inner products between rows, which are not the quantities analyzed in Section 1.1. In addition, the columns of \hat W O need not have unit norm, so the decomposition W = M \odot (\hat W O) no longer separates magnitude from direction. PAID therefore implements a different operation from the one claimed, and the name 'pairwise angular-invariant' is not supported.
  2. [Section 1.1 / Appendix A] Experiment 1 is not an independent test of the claimed domain invariance. Appendix A states that the test-time adaptation results in Experiment 1 were obtained with PAID itself in a non-continual setting, so the stability of pairwise angular structure shown in Fig. 1 is a property of PAID's own update trajectory rather than of pretrained weights under general adaptation. Since the update rule in Eq. (5) does not actually preserve column angles (see the previous comment), the experiment cannot validate the premise. The same issue affects Experiment 2 and the ablations in Table 5: settings (3) and (5), described as 'preserving pairwise structure,' change the column Gram matrix, so the claimed causal role of angular-structure preservation is not supported.
  3. [Table 5 / Section 4.3] The ablation labeled 'Inject Orth.' is interpreted as evidence that preserving pairwise angular structure is beneficial, but with right multiplication the orthogonal reparameterization changes both column angles and column norms. The observed gains (44.1% and 42.2% mean errors for the relevant rows of Table 5) therefore cannot be attributed to the invariance that motivates PAID; they may come from any other effect of the constrained update. A correct implementation of global column rotation would require left multiplication by an element of O(d), or an equivalent operation on the d-dimensional column space, at which point the parameterization, the loss, and all reported experiments would need to be redone. As written, the central conclusion of the paper is not established by the experiments.
minor comments (4)
  1. [Eq. (1)] The notation M \odot \hat W with M \in R^{1\times k} and \hat W \in R^{d\times k} is not defined; please state explicitly that M broadcasts across the rows of \hat W.
  2. [Eqs. (2)–(3), Appendix A] The metric in Eq. (3) measures inverse chordal distance between unit vectors rather than the angle itself; since the text repeatedly refers to 'pairwise angular structure,' it would help to state the monotonic relationship or define the metric directly in terms of cosine similarity.
  3. [Section 3.2 / Eq. (8)] The role of the coefficient r is described inconsistently as the number of Householder matrices, the chain length, and 'orthogonal matrix coefficient'; please define it once and use the same term throughout.
  4. [Tables 1–4] The tables list 'Source Pseudo [27]' as a method, but [27] is the pseudo-label method; please label it consistently as 'Pseudo-label' and check the formatting in Tables 2 and 3, where the Gain rows appear misaligned.

Circularity Check

1 steps flagged · score 6.0 of 10

The paper's motivation experiment measures pairwise-angular stability under corruption using PAID itself, so the central 'discovery' is enforced by the method's constrained update rather than observed independently.

  1. self definitional [Section 1.1 (Experiment 1) and Appendix A]
    "Experiment 1 examines the domain invariance of geometric attributes in weight space ... Corruption, where the model performs test-time adaptation on each of the 15 corrupted domains in ImageNet-C. ... The TTA results correspond to our proposed method, PAID, applied under a non-continual setting where all linear layers are adapted."

    The main discovery—that pairwise angular structure is stable across corruptions—is quantified by ΔS between the pretrained weights and the weights produced by PAID. But PAID freezes the direction matrix and updates only magnitudes and an orthogonal matrix, which Eq. 5 defines as 'preserving pairwise angular structure'. Therefore ΔS≈0 under corruption is built into the update rule, not discovered empirically. This self-consistent stability result is then used as the load-bearing prior in Section 1.2 to justify preserving the same structure during CTTA. The motivation evidence consequently reduces to the method's own constraint and cannot independently validate the paper's central claim.

full rationale

The paper contains one load-bearing circular step: the claim that pairwise angular structure is domain-invariant is demonstrated by adapting with PAID, whose update is designed to preserve that structure. Appendix A confirms that the Experiment 1 corruption results use PAID itself, making the measured stability a by-construction consequence rather than an empirical discovery. This makes the central prior partially self-confirming, despite the independent CTTA benchmark comparisons. No fitted parameter is renamed as a prediction, and there is no self-citation chain or imported uniqueness theorem. The separate mathematical issue that right-multiplication by a general orthogonal O does not generally preserve column-wise pairwise angles (Eqs. 5–8) is a correctness concern, not a circularity, so it is not scored here beyond the self-definitional motivation loop.

Assumptions & free parameters 3 free parameters · 4 assumptions · 1 invented entities

The central design rests on two unvalidated assumptions: (1) the false claim that right multiplication by O preserves column angles, and (2) the domain-invariance of pairwise angular structure, validated circularly. The remaining hyperparameters are standard tuning choices. The claimed prior is not independently evidenced.

free parameters (3)
  • Orthogonal matrix chain length r = 12
    Set by ablation on ImageNet-C to 12 and reused across all benchmarks; a capacity-surrogate hyperparameter.
  • Loss balancing coefficient λ = 1.0/1.0/0.1/1.0
    Chosen per dataset (CIFAR10-C, CIFAR100-C, ImageNet-C, ACDC) with no derivation; affects the trade-off between mean and variance alignment.
  • Number of source images for statistics = 500
    Chosen to balance performance and practicality; the paper does not report the exact error at 0 source images, so the dependence on this parameter is not fully quantified.
assumptions (4)
  • ad hoc to paper For an orthogonal matrix O and any matrix \hat{W} with unit-norm columns, the matrix \hat{W} O has the same pairwise column angles as \hat{W}.
    Invoked after Eq 5 and Eq 7 to claim preservation of pairwise angular structure; false in general because the Gram matrix becomes O^T G O, not G.
  • domain assumption The pairwise angular structure of pre-trained weights is domain-invariant under corruption and encodes semantic information.
    Central hypothesis; supported only by the self-referential Experiment 1 and by citations to hyperspherical learning and OFT.
  • domain assumption Source feature statistics computed from 500 images are sufficient to represent the source domain.
    Used in Eq 9; sensitivity is ablated only above 500, not down to the critical zero-image regime.
  • standard math Householder reflections with r = k can represent any orthogonal matrix.
    Standard result; used to justify expressiveness of the parametrization.
invented entities (1)
  • Pairwise angular structure as a domain-invariant semantic prior
    purpose: To justify freezing the direction matrix and updating only orthogonal rotations
    The paper claims this prior exists in pretrained weights, but the only direct evidence (Experiment 1) is produced by PAID itself, so there is no external falsifiable handle.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PAID: Pairwise Angular-Invariant Decomposition for Continual Test-Time Adaptation." pith.science (2026). https://pith.science/paper/LGPTCAUU

@misc{pith2026250602453,
  author       = {Pith},
  title        = {Pith review of: PAID: Pairwise Angular-Invariant Decomposition for Continual Test-Time Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LGPTCAUU}},
  note         = {Machine review of arXiv:2506.02453}
}
read the original abstract

Continual Test-Time Adaptation (CTTA) aims to online adapt a pre-trained model to changing environments during inference. Most existing methods focus on exploiting target data, while overlooking another crucial source of information, the pre-trained weights, which encode underutilized domain-invariant priors. This paper takes the geometric attributes of pre-trained weights as a starting point, systematically analyzing three key components: magnitude, absolute angle, and pairwise angular structure. We find that the pairwise angular structure remains stable across diverse corrupted domains and encodes domain-invariant semantic information, suggesting it should be preserved during adaptation. Based on this insight, we propose PAID (Pairwise Angular-Invariant Decomposition), a prior-driven CTTA method that decomposes weight into magnitude and direction, and introduces a learnable orthogonal matrix via Householder reflections to globally rotate direction while preserving the pairwise angular structure. During adaptation, only the magnitudes and the orthogonal matrices are updated. PAID achieves consistent improvements over recent SOTA methods on four widely used CTTA benchmarks, demonstrating that preserving pairwise angular structure offers a simple yet effective principle for CTTA.

Figures

Figures reproduced from arXiv: 2506.02453 by the authors.

Figure 1
Figure 1. (Experiment 1) Visualization of cross-domain variation of three geometric properties. Pairwise angular structure remains stable under corruption but varies under semantic shift, suggesting it encodes semantic-relevant, domain-invariant information. In contrast, magnitude and absolute angle fluctuate irregularly across domains, reflecting domain-specific shifts. 1.1 Motivation Recent studies in hyperspherical learnin… view at source ↗
Figure 2
Figure 2. (Experiment 2) Illustration of three update strategies. (1) Scaling the magnitude of each vector; (2) Independently rotating each vector, altering their absolute angle; (3) Jointly rotating all vectors while preserving their pairwise angular structure. Magnitude + Direction (Preserving pairwise structure) Direction only (Preserving pairwise structure) Magnitude only Magnitude + Direction Direction only 4.3% 8.3% 19.… view at source ↗
Figure 3
Figure 3. (Experiment 2) Radar and bar charts showing classification error rates across 15 corruption domains and their mean. Comparisons between settings (5 vs 1, 3; 5 vs 4; 3 vs 2) show that pairwise angular structure is a domain-invariant component worth preserving, while magnitude and direction, when constrained by fixed angular structure, are domain-specific and beneficial for adaptation. exhibit irregular variations acr… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: (Experiment 3) Attention map visualizations for two setting pairs (5 vs 4; 3 vs 2). The comparison further supports the semantic relevance and domain invariance of pairwise angular structure in CTTA. capture core object information. This comparison reinforces the criti…
Figure 5
Figure 5. Figure 5: PAID decomposes pre-trained weights into magnitude and direction. To preserve pair￾wise angular structure, we introduce a learnable orthogonal matrix, enabling global rotation. Only magnitude and orthogonal matrices are updated during adaptation. Given a model fθ pre-t…
Figure 6
Figure 6. Figure 6: Ablation on (a) the coefficient of the injected orthogonal matrix, (b) the number of source [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Ablation on loss balancing coefficient λ [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

74 extracted references · 36 canonical work pages

  1. [1]

    Progressive feature alignment for unsupervised domain adaptation

    Chaoqi Chen, Weiping Xie, Wenbing Huang, Yu Rong, Xinghao Ding, Yue Huang, Tingyang Xu, and Junzhou Huang. Progressive feature alignment for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 627–636, 2019

    work page 2019

  2. [2]

    Harmonizing transfer- ability and discriminability for adapting object detectors

    Chaoqi Chen, Zebiao Zheng, Xinghao Ding, Yue Huang, and Qi Dou. Harmonizing transfer- ability and discriminability for adapting object detectors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8869–8878, 2020

    work page 2020

  3. [3]

    I3net: Implicit instance-invariant network for adapting one-stage object detectors

    Chaoqi Chen, Zebiao Zheng, Yue Huang, Xinghao Ding, and Yizhou Yu. I3net: Implicit instance-invariant network for adapting one-stage object detectors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12576–12585, 2021

    work page 2021

  4. [4]

    Contrastive test-time adaptation

    Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 295–305, 2022

    work page 2022

  5. [5]

    Amplitude-phase recombination: Rethinking robustness of convolutional neural networks in frequency domain

    Guangyao Chen, Peixi Peng, Li Ma, Jia Li, Lin Du, and Yonghong Tian. Amplitude-phase recombination: Rethinking robustness of convolutional neural networks in frequency domain. In Proceedings of the IEEE/CVF international conference on computer vision, pages 458–467, 2021

    work page 2021

  6. [6]

    The cityscapes dataset for semantic urban scene understanding

    Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213–3223, 2016

    2016

  7. [7]

    Imagenet: A large- scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large- scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009

    2009

  8. [8]

    Efficient adaptation of pre-trained vision transformer via householder transformation

    Wei Dong, Yuan Sun, Yiting Yang, Xing Zhang, Zhijun Lin, Qingsen Yan, Haokui Zhang, Peng Wang, Yang Yang, and Hengtao Shen. Efficient adaptation of pre-trained vision transformer via householder transformation. arXiv preprint arXiv:2410.22952, 2024

    arXiv

Show all 74 references
  1. [9]

    Efficient adaptation of large vision transformer via adapter re-composing

    Wei Dong, Dawei Yan, Zhijun Lin, and Peng Wang. Efficient adaptation of large vision transformer via adapter re-composing. Advances in Neural Information Processing Systems, 36:52548–52567, 2023

    2023

  2. [10]

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

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv...

    2010 arXiv

  3. [11]

    Test time adaptation through perturbation robustness

    François Fleuret et al. Test time adaptation through perturbation robustness. In NeurIPS 2021 Workshop on Distribution Shifts: Connecting Methods and Applications, 2021

    2021

  4. [12]

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

    Yulu Gan, Yan Bai, Yihang Lou, Xianzheng Ma, Renrui Zhang, Nian Shi, and Lin Luo. Decorate the newcomers: Visual domain prompt for continual test time adaptation. In Proceedings of the AAAI conference on artificial intelligence, volume 37, pages 7595–7603, 2023

    2023

  5. [13]

    Visual prompt tuning for test-time domain adaptation

    Yunhe Gao, Xingjian Shi, Yi Zhu, Hao Wang, Zhiqiang Tang, Xiong Zhou, Mu Li, and Dimitris N Metaxas. Visual prompt tuning for test-time domain adaptation. arXiv preprint arXiv:2210.04831, 2022

    2022 arXiv

  6. [14]

    Note: Robust continual test-time adaptation against temporal correlation

    Taesik Gong, Jongheon Jeong, Taewon Kim, Yewon Kim, Jinwoo Shin, and Sung-Ju Lee. Note: Robust continual test-time adaptation against temporal correlation. Advances in Neural Information Processing Systems, 35:27253–27266, 2022

    2022

  7. [15]

    Sotta: Robust test-time adaptation on noisy data streams

    Taesik Gong, Yewon Kim, Taeckyung Lee, Sorn Chottananurak, and Sung-Ju Lee. Sotta: Robust test-time adaptation on noisy data streams. Advances in Neural Information Processing Systems, 36:14070–14093, 2023. 10

    2023

  8. [16]

    Parameter-efficient fine-tuning for large models: A comprehensive survey

    Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient fine-tuning for large models: A comprehensive survey. arXiv preprint arXiv:2403.14608, 2024

    2024 arXiv

  9. [17]

    Benchmarking neural network robustness to common corruptions and perturbations

    Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019

    1903 arXiv

  10. [18]

    Parameter-efficient transfer learning for nlp

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790–2799. PMLR, 2019

    2019

  11. [19]

    Unitary triangularization of a nonsymmetric matrix

    Alston S Householder. Unitary triangularization of a nonsymmetric matrix. Journal of the ACM (JACM), 5(4):339–342, 1958

    1958

  12. [20]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022

    2022

  13. [21]

    Fsdr: Frequency space domain randomization for domain generalization

    Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6891–6902, 2021

    2021

  14. [22]

    Test-time classifier adjustment module for model-agnostic domain generalization

    Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427–2440, 2021

    2021

  15. [23]

    Test-time adaptation via self-training with nearest neighbor information

    Minguk Jang, Sae-Young Chung, and Hye Won Chung. Test-time adaptation via self-training with nearest neighbor information. arXiv preprint arXiv:2207.10792, 2022

    2022 arXiv

  16. [24]

    Feature stylization and domain-aware contrastive learning for domain generalization

    Seogkyu Jeon, Kibeom Hong, Pilhyeon Lee, Jewook Lee, and Hyeran Byun. Feature stylization and domain-aware contrastive learning for domain generalization. In Proceedings of the 29th ACM International Conference on Multimedia, pages 22–31, 2021

    2021

  17. [25]

    Visual prompt tuning

    Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European conference on computer vision, pages 709–727. Springer, 2022

    2022

  18. [26]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009

    2009

  19. [27]

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

    Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML , volume 3, page 896. Atlanta, 2013

    2013

  20. [28]

    Decompose, adjust, compose: Effective normalization by playing with frequency for domain generalization

    Sangrok Lee, Jongseong Bae, and Ha Young Kim. Decompose, adjust, compose: Effective normalization by playing with frequency for domain generalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11776–11785, 2023

    2023

  21. [29]

    A comprehensive survey on test-time adaptation under distribution shifts

    Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision, 133(1):31–64, 2025

    2025

  22. [30]

    Deep frequency filtering for domain generalization

    Shiqi Lin, Zhizheng Zhang, Zhipeng Huang, Yan Lu, Cuiling Lan, Peng Chu, Quanzeng You, Jiang Wang, Zicheng Liu, Amey Parulkar, et al. Deep frequency filtering for domain generalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages...

    2023

  23. [31]

    Continual-mae: Adaptive distribution masked autoencoders for continual test-time adaptation

    Jiaming Liu, Ran Xu, Senqiao Yang, Renrui Zhang, Qizhe Zhang, Zehui Chen, Yandong Guo, and Shanghang Zhang. Continual-mae: Adaptive distribution masked autoencoders for continual test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Rec...

    2024

  24. [32]

    Vida: Homeostatic visual domain adapter for continual test time adaptation

    Jiaming Liu, Senqiao Yang, Peidong Jia, Renrui Zhang, Ming Lu, Yandong Guo, Wei Xue, and Shanghang Zhang. Vida: Homeostatic visual domain adapter for continual test time adaptation. arXiv preprint arXiv:2306.04344, 2023. 11

    2023 arXiv

  25. [33]

    Dora: Weight-decomposed low-rank adaptation

    Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. Dora: Weight-decomposed low-rank adaptation. In Forty-first International Conference on Machine Learning, 2024

    2024

  26. [34]

    Learning towards minimum hyperspherical energy

    Weiyang Liu, Rongmei Lin, Zhen Liu, Lixin Liu, Zhiding Yu, Bo Dai, and Le Song. Learning towards minimum hyperspherical energy. Advances in neural information processing systems, 31, 2018

    2018

  27. [35]

    Orthogonal over-parameterized training

    Weiyang Liu, Rongmei Lin, Zhen Liu, James M Rehg, Liam Paull, Li Xiong, Le Song, and Adrian Weller. Orthogonal over-parameterized training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7251–7260, 2021

    2021

  28. [36]

    Decoupled networks

    Weiyang Liu, Zhen Liu, Zhiding Yu, Bo Dai, Rongmei Lin, Yisen Wang, James M Rehg, and Le Song. Decoupled networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2771–2779, 2018

    2018

  29. [37]

    Deep hyperspherical learning

    Weiyang Liu, Yan-Ming Zhang, Xingguo Li, Zhiding Yu, Bo Dai, Tuo Zhao, and Le Song. Deep hyperspherical learning. Advances in neural information processing systems, 30, 2017

    2017

  30. [38]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  31. [39]

    Universal test-time adaptation through weight ensembling, diversity weighting, and prior correction

    Robert A Marsden, Mario Döbler, and Bin Yang. Universal test-time adaptation through weight ensembling, diversity weighting, and prior correction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2555–2565, 2024

    2024

  32. [40]

    Comparing vision transformers and convolutional neural networks for image classification: A literature review

    José Maurício, Inês Domingues, and Jorge Bernardino. Comparing vision transformers and convolutional neural networks for image classification: A literature review. Applied Sciences, 13(9):5521, 2023

    2023

  33. [41]

    Image segmentation using deep learning: A survey

    Shervin Minaee, Yuri Boykov, Fatih Porikli, Antonio Plaza, Nasser Kehtarnavaz, and Demetri Terzopoulos. Image segmentation using deep learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(7):3523–3542, 2021

    2021

  34. [42]

    Actmad: Activation matching to align distributions for test-time-training

    Muhammad Jehanzeb Mirza, Pol Jané Soneira, Wei Lin, Mateusz Kozinski, Horst Possegger, and Horst Bischof. Actmad: Activation matching to align distributions for test-time-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2415...

    2023

  35. [43]

    Test-time model adaptation with only forward passes

    Shuaicheng Niu, Chunyan Miao, Guohao Chen, Pengcheng Wu, and Peilin Zhao. Test-time model adaptation with only forward passes. arXiv preprint arXiv:2404.01650, 2024

    2024 arXiv

  36. [44]

    Efficient test-time model adaptation without forgetting

    Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International confer- ence on machine learning, pages 16888–16905. PMLR, 2022

    2022

  37. [45]

    Towards stable test-time adaptation in dynamic wild world

    Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. arXiv preprint arXiv:2302.12400, 2023

    2023 arXiv

  38. [46]

    The fast Fourier transform

    Henri J Nussbaumer and Henri J Nussbaumer. The fast Fourier transform. Springer, 1982

    1982

  39. [47]

    Learning how to ask: Querying lms with mixtures of soft prompts

    Guanghui Qin and Jason Eisner. Learning how to ask: Querying lms with mixtures of soft prompts. arXiv preprint arXiv:2104.06599, 2021

    2021 arXiv

  40. [48]

    Controlling text-to-image diffusion by orthogonal finetuning

    Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Schölkopf. Controlling text-to-image diffusion by orthogonal finetuning. Advances in Neural Information Processing Systems, 36:79320–79362, 2023

    2023

  41. [49]

    Do imagenet classifiers generalize to imagenet? In International conference on machine learning, pages 5389–5400

    Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imagenet? In International conference on machine learning, pages 5389–5400. PMLR, 2019

    2019

  42. [50]

    Imagenet-21k pretraining for the masses

    Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972, 2021. 12

    2021 arXiv

  43. [51]

    Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding

    Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10765–10775, 2021

    2021

  44. [52]

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

    Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. Advances in Neural Information Processing Systems, 35:14274–14289, 2022

    2022

  45. [53]

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

    Junha Song, Jungsoo Lee, In So Kweon, and Sungha Choi. Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11920–11929, 2023

    2023

  46. [54]

    Hyperspherical consistency regularization

    Cheng Tan, Zhangyang Gao, Lirong Wu, Siyuan Li, and Stan Z Li. Hyperspherical consistency regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7244–7255, 2022

    2022

  47. [55]

    Tesla: Test- time self-learning with automatic adversarial augmentation

    Devavrat Tomar, Guillaume Vray, Behzad Bozorgtabar, and Jean-Philippe Thiran. Tesla: Test- time self-learning with automatic adversarial augmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20341–20350, 2023

    2023

  48. [56]

    Un-mixing test-time normalization statistics: Combatting label temporal correlation

    Devavrat Tomar, Guillaume Vray, Jean-Philippe Thiran, and Behzad Bozorgtabar. Un-mixing test-time normalization statistics: Combatting label temporal correlation. arXiv preprint arXiv:2401.08328, 2024

    2024 arXiv

  49. [57]

    Tent: Fully test-time adaptation by entropy minimization

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020

    2006 arXiv

  50. [58]

    Continual test-time domain adaptation

    Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 7201–7211, 2022

    2022

  51. [59]

    Feature alignment and uniformity for test time adaptation

    Shuai Wang, Daoan Zhang, Zipei Yan, Jianguo Zhang, and Rui Li. Feature alignment and uniformity for test time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20050–20060, 2023

    2023

  52. [60]

    In search of lost online test-time adaptation: A survey

    Zixin Wang, Yadan Luo, Liang Zheng, Zhuoxiao Chen, Sen Wang, and Zi Huang. In search of lost online test-time adaptation: A survey. International Journal of Computer Vision, pages 1–34, 2024

    2024

  53. [61]

    Beyond model adaptation at test time: A survey

    Zehao Xiao and Cees GM Snoek. Beyond model adaptation at test time: A survey. arXiv preprint arXiv:2411.03687, 2024

    2024 arXiv

  54. [62]

    Segformer: Simple and efficient design for semantic segmentation with transformers

    Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems, 34:12077–12090, 2021

    2021

  55. [63]

    A fourier-based framework for domain generalization

    Qinwei Xu, Ruipeng Zhang, Ya Zhang, Yanfeng Wang, and Qi Tian. A fourier-based framework for domain generalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14383–14392, 2021

    2021

  56. [64]

    A versatile framework for continual test-time domain adaptation: Balancing discriminability and generalizability

    Xu Yang, Xuan Chen, Moqi Li, Kun Wei, and Cheng Deng. A versatile framework for continual test-time domain adaptation: Balancing discriminability and generalizability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23731–23740, 2024

    2024

  57. [65]

    Fda: Fourier domain adaptation for semantic segmentation

    Yanchao Yang and Stefano Soatto. Fda: Fourier domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4085–4095, 2020

    2020

  58. [66]

    Robust test-time adaptation in dynamic scenarios

    Longhui Yuan, Binhui Xie, and Shuang Li. Robust test-time adaptation in dynamic scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 15922–15932, 2023. 13

    2023

  59. [67]

    Bridging the gap between low-rank and orthogonal adaptation via householder reflection adaptation

    Shen Yuan, Haotian Liu, and Hongteng Xu. Bridging the gap between low-rank and orthogonal adaptation via householder reflection adaptation. arXiv preprint arXiv:2405.17484, 2024

    2024 arXiv

  60. [68]

    Rethinking precision of pseudo label: Test-time adaptation via complementary learning

    Longbin Zeng, Jiayi Han, Liang Du, and Weiyang Ding. Rethinking precision of pseudo label: Test-time adaptation via complementary learning. Pattern Recognition Letters, 177:96–102, 2024

    2024

  61. [69]

    Memo: Test time robustness via adaptation and augmentation

    Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Advances in neural information processing systems, 35:38629–38642, 2022

    2022

  62. [70]

    Adalora: Adaptive budget allocation for parameter- efficient fine-tuning

    Qingru Zhang, Minshuo Chen, Alexander Bukharin, Nikos Karampatziakis, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adalora: Adaptive budget allocation for parameter- efficient fine-tuning. arXiv preprint arXiv:2303.10512, 2023

    2023 arXiv

  63. [71]

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

    Yunbei Zhang, Akshay Mehra, and Jihun Hamm. Dynamic domains, dynamic solutions: Dpcore for continual test-time adaptation. arXiv preprint arXiv:2406.10737, 2024

    2024 arXiv

  64. [72]

    Ot-vp: Optimal transport-guided visual prompting for test-time adaptation

    Yunbei Zhang, Akshay Mehra, and Jihun Hamm. Ot-vp: Optimal transport-guided visual prompting for test-time adaptation. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1122–1132. IEEE, 2025

    2025

  65. [73]

    Delta: degradation-free fully test-time adaptation

    Bowen Zhao, Chen Chen, and Shu-Tao Xia. Delta: degradation-free fully test-time adaptation. arXiv preprint arXiv:2301.13018, 2023

    2023 arXiv

  66. [74]

    10-round

    Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 111(3):257–276, 2023. 14 A Additional Details of Three Motivation Experiments In Experiment 1, we compute a weighted average of three statistics ...

    2023

Pith tools

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