pith. sign in

arxiv: 2605.18608 · v1 · pith:62HN7BNLnew · submitted 2026-05-18 · 💻 cs.CV

Dance Across Shifts: Forward-Facilitation Continual Test-Time Adaptation through Dynamic Style Bridging

Zhilin Zhu , Yabin Wang , Zhiheng Ma , Yaguang Song , Yaowei Wang , Xiaopeng Hong This is my paper

Pith reviewed 2026-05-20 10:44 UTC · model grok-4.3

classification 💻 cs.CV
keywords continual test-time adaptationdynamic style bridgingforward-facilitationdistribution shiftsclass exemplarstest-time adaptationcomputer vision
0
0 comments X

The pith

A forward-facilitation approach builds pre-deployment class exemplars and dynamically bridges them with new data styles at input, statistical, and representation levels to supply reliable supervision during continual distribution shifts.

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

The paper aims to replace rigid backward alignment of new data to source-derived surrogates with a forward process that keeps generated proxies up to date. Before deployment a compact set of class exemplars is created; at test time a multi-level bridging step injects the styles of arriving batches into those proxies without changing their class semantics. The updated proxies then serve as on-demand supervisory signals for adaptation. A reader would care because perception models deployed in changing environments could maintain performance across successive shifts without repeated access to the original training data.

Core claim

The central claim is that a compact knowledge base of generated class exemplars, updated during test time by a multi-level bridging mechanism that injects incoming data styles at the input, statistical, and representation levels while preserving original semantics, yields reliable supervisory signals and thereby enables stable adaptation under continual distribution shifts.

What carries the argument

The multi-level bridging mechanism that injects incoming data styles into pre-generated class-exemplar proxies at input, statistical, and representation levels while preserving proxy semantics.

If this is right

  • Reliable on-demand supervisory signals become available from the updated proxies.
  • Adaptation remains stable as distribution shifts continue over time.
  • Substantial and consistent gains appear over recent state-of-the-art CTTA methods on standard benchmarks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same bridging idea could be tested in settings where only a small memory budget is allowed for stored exemplars.
  • If the proxies remain semantically stable, the approach might combine naturally with memory-efficient continual-learning techniques that also store class representatives.
  • Real-time systems facing weather or lighting changes could adopt the method to avoid full retraining cycles.

Load-bearing premise

The bridging operations can add new data styles to the proxies without distorting their original class semantics or introducing large generative bias.

What would settle it

On a standard CTTA benchmark, if the adapted proxies produce lower source-domain accuracy than the unadapted ones, or if the method shows no consistent gain over recent baselines across multiple shift sequences, the central claim would be falsified.

Figures

Figures reproduced from arXiv: 2605.18608 by Xiaopeng Hong, Yabin Wang, Yaguang Song, Yaowei Wang, Zhiheng Ma, Zhilin Zhu.

Figure 1
Figure 1. Figure 1: Illustration of the considered CTTA problem and com [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The illustration of our framework. We construct in advance a compact set of proxies containing synthetic knowledge that encap [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison between different methods in terms of aver [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation study on the size of the knowledge base. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of synthetic images. We show source images (top row) followed by synthetic samples generated by BigGAN, Stable [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
read the original abstract

Continual Test-Time Adaptation (CTTA) aims to empower perception systems to handle dynamic distribution shifts encountered after deployment. Existing methods predominantly follow a backward-alignment paradigm, which rigidly aligns incoming data with supervisory surrogates derived from the source domain. Consequently, they struggle with unreliable supervision and evolving distribution shifts. To overcome these limitations, we introduce a novel forward-facilitation paradigm through a method termed Dynamic Style Bridging. Prior to deployment, we construct a compact knowledge base of generated class exemplars. During test time, to mitigate inherent generative bias and adapt these proxies to incoming data, we propose a multi-level bridging mechanism. This mechanism dynamically injects the proxies with incoming data styles at the input, statistical, and representation levels, while preserving the original semantics of the proxies. These high-fidelity proxies are then used to provide reliable, on-demand supervisory signals, enabling stable adaptation under continual shifts. Extensive experiments across standard CTTA benchmarks demonstrate that our method achieves consistent and substantial improvements over recent state-of-the-art approaches. Code is available at \href{https://github.com/z1358/DAS}.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces a forward-facilitation paradigm for Continual Test-Time Adaptation (CTTA) via Dynamic Style Bridging. Prior to deployment, a compact knowledge base of generated class exemplars is constructed. At test time, a multi-level bridging mechanism injects incoming data styles into these proxies at input, statistical, and representation levels while aiming to preserve original semantics and mitigate generative bias; the adapted proxies then supply on-demand supervisory signals for stable adaptation under continual shifts. Extensive experiments on standard CTTA benchmarks are reported to yield consistent improvements over recent state-of-the-art methods, with code released.

Significance. If the multi-level bridging demonstrably preserves semantics while enabling reliable forward supervision, the work could meaningfully advance CTTA by addressing limitations of backward-alignment approaches under evolving shifts. The public code release supports reproducibility and is a clear strength.

major comments (2)
  1. [Method] Method section (multi-level bridging description): the central claim that the mechanism 'preserves the original semantics of the proxies' while dynamically injecting styles at three levels lacks isolated validation. No per-level ablation results, no direct semantic-fidelity metrics (e.g., class-conditional cosine similarity or LPIPS between original and bridged proxies), and no analysis of behavior as shifts accumulate across test-time steps are provided; observed accuracy gains could therefore arise from generic regularization rather than the claimed forward-facilitation property.
  2. [Experiments] Experimental section (benchmark results): while aggregate improvements over SOTA are stated, the manuscript does not report whether the gains remain when the bridging mechanism is replaced by simpler proxy regularization or when the number of continual steps increases; this weakens the load-bearing link between the proposed mechanism and the reported performance.
minor comments (2)
  1. [Method] Notation for the three bridging levels is introduced at a high level; explicit equations or pseudocode for each level would improve clarity and reproducibility.
  2. [Experiments] Figure captions and axis labels in the experimental plots should explicitly state the number of continual adaptation steps and the exact metrics used (e.g., mean accuracy, forgetting measure).

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major comment below and will incorporate revisions to strengthen the validation of our claims.

read point-by-point responses

  1. Referee: [Method] Method section (multi-level bridging description): the central claim that the mechanism 'preserves the original semantics of the proxies' while dynamically injecting styles at three levels lacks isolated validation. No per-level ablation results, no direct semantic-fidelity metrics (e.g., class-conditional cosine similarity or LPIPS between original and bridged proxies), and no analysis of behavior as shifts accumulate across test-time steps are provided; observed accuracy gains could therefore arise from generic regularization rather than the claimed forward-facilitation property.

    Authors: We agree that isolated validation would more directly support the claim of semantic preservation and forward-facilitation. The multi-level design (input, statistical, and representation) is intended to inject styles while retaining class semantics via the pre-generated proxies, but the original manuscript relies on end-to-end accuracy gains rather than per-level ablations or explicit fidelity metrics such as cosine similarity or LPIPS. We will add these analyses in the revision, including per-level ablation results, semantic-fidelity metrics on bridged vs. original proxies, and step-wise behavior under accumulating shifts to better isolate the mechanism from generic regularization effects. revision: yes

  2. Referee: [Experiments] Experimental section (benchmark results): while aggregate improvements over SOTA are stated, the manuscript does not report whether the gains remain when the bridging mechanism is replaced by simpler proxy regularization or when the number of continual steps increases; this weakens the load-bearing link between the proposed mechanism and the reported performance.

    Authors: We acknowledge that additional controls would strengthen the attribution of gains to the dynamic bridging mechanism. The reported results demonstrate consistent improvements on standard CTTA benchmarks, but we did not include a direct comparison to simpler proxy regularization or extended continual-step settings. We will add these experiments in the revised version: a variant with simpler regularization in place of multi-level bridging, and results on benchmarks with a larger number of test-time steps, to confirm that performance benefits persist and are tied to the proposed forward-facilitation approach. revision: yes

Circularity Check

0 steps flagged

No significant circularity: derivation relies on independent proxy construction and bridging mechanism

full rationale

The paper introduces a forward-facilitation paradigm for CTTA by first generating class exemplars pre-deployment and then applying a multi-level style bridging process (input, statistical, representation) to adapt them while claiming semantic preservation. This construction is presented as a design choice rather than a self-referential definition or fitted parameter renamed as prediction. No equations reduce the claimed supervisory signals or performance gains to quantities defined in terms of the outputs themselves. The abstract and method description contain no load-bearing self-citations, uniqueness theorems imported from prior author work, or ansatzes smuggled via citation. Experiments are framed as external validation on benchmarks rather than tautological outcomes. The derivation chain remains self-contained against the stated assumptions.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim depends on the ability to construct and adapt a compact knowledge base of generated class exemplars prior to deployment.

axioms (1)
  • domain assumption A compact knowledge base of generated class exemplars can be constructed prior to deployment
    Stated as the starting point for the method in the abstract.
invented entities (1)
  • Dynamic Style Bridging mechanism no independent evidence
    purpose: To adapt generated proxies to incoming data styles at input, statistical, and representation levels while preserving semantics
    Newly proposed component central to mitigating generative bias and providing supervision.

pith-pipeline@v0.9.0 · 5745 in / 1181 out tokens · 57718 ms · 2026-05-20T10:44:06.637378+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

87 extracted references · 87 canonical work pages

  1. [1]

    Shekoofeh Azizi, Simon Kornblith, Chitwan Saharia, Mo- hammad Norouzi, and David J. Fleet. Synthetic data from diffusion models improves imagenet classification.Transac- tions on Machine Learning Research, 2023. 3

    work page 2023

  2. [2]

    A probabilistic frame- work for lifelong test-time adaptation

    Dhanajit Brahma and Piyush Rai. A probabilistic frame- work for lifelong test-time adaptation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3582–3591, 2023. 2

    work page 2023

  3. [3]

    Large scale GAN training for high fidelity natural image synthe- sis

    Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthe- sis. InInternational Conference on Learning Representa- tions, 2019. 8

    work page 2019

  4. [4]

    SANTA: Source anchoring network and target alignment for continual test time adaptation.Transactions on Machine Learning Research, 2023

    Goirik Chakrabarty, Manogna Sreenivas, and Soma Biswas. SANTA: Source anchoring network and target alignment for continual test time adaptation.Transactions on Machine Learning Research, 2023. 2

    work page 2023

  5. [5]

    Contrastive test-time adaptation

    Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 295–305, 2022. 2

    work page 2022

  6. [6]

    Reducing class-wise confusion for incremen- tal learning with disentangled manifolds

    Huitong Chen, Yu Wang, Yan Fan, Guosong Jiang, and Qinghua Hu. Reducing class-wise confusion for incremen- tal learning with disentangled manifolds. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10121–10130, 2025. 2

    work page 2025

  7. [7]

    Spectral property- driven data augmentation for hyperspectral single-source do- main generalization

    Taiqin Chen, Yifeng Wang, Xiaochen Feng, Zhilin Zhu, Hao Sha, Yingjian Li, and Yongbing Zhang. Spectral property- driven data augmentation for hyperspectral single-source do- main generalization. InProceedings of the AAAI Conference on Artificial Intelligence, pages 3038–3046, 2026. 2

    work page 2026

  8. [8]

    Each test image deserves a specific prompt: Con- tinual test-time adaptation for 2d medical image segmenta- tion

    Ziyang Chen, Yongsheng Pan, Yiwen Ye, Mengkang Lu, and Yong Xia. Each test image deserves a specific prompt: Con- tinual test-time adaptation for 2d medical image segmenta- tion. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 11184–11193,

    work page

  9. [9]

    Multi-granularity class prototype topology distillation for class-incremental source- free unsupervised domain adaptation

    Peihua Deng, Jiehua Zhang, Xichun Sheng, Chenggang Yan, Yaoqi Sun, Ying Fu, and Liang Li. Multi-granularity class prototype topology distillation for class-incremental source- free unsupervised domain adaptation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 30566–30576, 2025. 1

    work page 2025

  10. [10]

    Marsden, and Bin Yang

    Mario D ¨obler, Robert A. Marsden, and Bin Yang. Robust mean teacher for continual and gradual test-time adaptation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7704–7714, 2023. 2, 3, 5, 6, 13, 14, 17

    work page 2023

  11. [11]

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

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale. InInternational Con- ference on Learning Representations, 2021. 1, 6, 13

    work page 2021

  12. [12]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. InIn- ternational Conference on Machine Learning, pages 12606– 12633. PMLR, 2024. 8

    work page 2024

  13. [13]

    Scaling laws of synthetic images for model training

    Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7382–7392, 2024. 3

    work page 2024

  14. [14]

    Dynamic sub-graph distillation for robust semi-supervised continual learning

    Yan Fan, Yu Wang, Pengfei Zhu, and Qinghua Hu. Dynamic sub-graph distillation for robust semi-supervised continual learning. InProceedings of the AAAI Conference on Artifi- cial Intelligence, pages 11927–11935, 2024. 3

    work page 2024

  15. [15]

    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. InProceedings of the AAAI Conference on Artificial Intel- ligence, pages 7595–7603, 2023. 1, 3, 16

    work page 2023

  16. [16]

    Back to the source: Diffusion- driven adaptation to test-time corruption

    Jin Gao, Jialing Zhang, Xihui Liu, Trevor Darrell, Evan Shel- hamer, and Dequan Wang. Back to the source: Diffusion- driven adaptation to test-time corruption. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11786–11796, 2023. 2, 3, 5, 13

    work page 2023

  17. [17]

    All for one, and one for all: Urbansyn dataset, the third musketeer of synthetic driv- ing scenes.Neurocomputing, 637:130038, 2025

    Jose L G ´omez, Manuel Silva, Antonio Seoane, Agn `es Borr´as, Mario Noriega, Germ ´an Ros, Jose A Iglesias- Guitian, and Antonio M L ´opez. All for one, and one for all: Urbansyn dataset, the third musketeer of synthetic driv- ing scenes.Neurocomputing, 637:130038, 2025. 16

    work page 2025

  18. [18]

    Sotta: Robust test-time adaptation on noisy data streams.Advances in Neural Information Pro- cessing Systems, 36:14070–14093, 2023

    Taesik Gong, Yewon Kim, Taeckyung Lee, Sorn Chottana- nurak, and Sung-Ju Lee. Sotta: Robust test-time adaptation on noisy data streams.Advances in Neural Information Pro- cessing Systems, 36:14070–14093, 2023. 2

    work page 2023

  19. [19]

    Everything to the synthetic: Diffusion-driven test- time adaptation via synthetic-domain alignment

    Jiayi Guo, Junhao Zhao, Chaoqun Du, Yulin Wang, Chun- jiang Ge, Zanlin Ni, Shiji Song, Humphrey Shi, and Gao Huang. Everything to the synthetic: Diffusion-driven test- time adaptation via synthetic-domain alignment. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 30503–30513, 2025. 2, 3, 5, 13

    work page 2025

  20. [20]

    Synthclip: Are we ready for a fully synthetic clip training?arXiv preprint arXiv:2402.01832, 2024

    Hasan Abed Al Kader Hammoud, Hani Itani, Fabio Pizzati, Philip Torr, Adel Bibi, and Bernard Ghanem. Synthclip: Are we ready for a fully synthetic clip training?arXiv preprint arXiv:2402.01832, 2024. 3

    work page arXiv 2024

  21. [21]

    Ranked entropy minimization for continual test-time adaptation

    Jisu Han, Jaemin Na, and Wonjun Hwang. Ranked entropy minimization for continual test-time adaptation. InInterna- tional Conference on Machine Learning, 2025. 5, 13, 15

    work page 2025

  22. [22]

    Benchmarking neu- ral network robustness to common corruptions and perturba- tions

    Dan Hendrycks and Thomas Dietterich. Benchmarking neu- ral network robustness to common corruptions and perturba- tions. InInternational Conference on Learning Representa- tions, 2019. 5

    work page 2019

  23. [23]

    Cosmic: Clique- oriented semantic multi-space integration for robust clip test- time adaptation

    Fanding Huang, Jingyan Jiang, Qinting Jiang, Hebei Li, Faisal Nadeem Khan, and Zhi Wang. Cosmic: Clique- oriented semantic multi-space integration for robust clip test- time adaptation. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, pages 9772–9781, 2025. 2

    work page 2025

  24. [24]

    Arbitrary style transfer in real-time with adaptive instance normalization

    Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. InProceed- ings of the IEEE/CVF International Conference on Com- puter Vision, pages 1501–1510, 2017. 4 9

    work page 2017

  25. [25]

    Rtracker: Recoverable track- ing via pn tree structured memory

    Yuqing Huang, Xin Li, Zikun Zhou, Yaowei Wang, Zhenyu He, and Ming-Hsuan Yang. Rtracker: Recoverable track- ing via pn tree structured memory. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19038–19047, 2024. 1

    work page 2024

  26. [26]

    Oxford University Press, 1990

    Terence Irwin.Aristotle’s first principles. Oxford University Press, 1990. 2

    work page 1990

  27. [27]

    Leveraging proxy of training data for test-time adaptation

    Juwon Kang, Nayeong Kim, Donghyeon Kwon, Jungseul Ok, and Suha Kwak. Leveraging proxy of training data for test-time adaptation. InInternational Conference on Ma- chine Learning, pages 15737–15752. PMLR, 2023. 2

    work page 2023

  28. [28]

    Wilds: A benchmark of in-the- wild distribution shifts

    Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubra- mani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, et al. Wilds: A benchmark of in-the- wild distribution shifts. InInternational Conference on Machine Learning, pages 5637–5664. PMLR, 2021. 1

    work page 2021

  29. [29]

    Becotta: Input-dependent online blending of experts for continual test-time adaptation

    Daeun Lee, Jaehong Yoon, and Sung Ju Hwang. Becotta: Input-dependent online blending of experts for continual test-time adaptation. InInternational Conference on Ma- chine Learning, pages 27072–27093. PMLR, 2024. 3

    work page 2024

  30. [30]

    Continual adaptation: Environment-conditional param- eter generation for object detection in dynamic scenarios

    Deng Li, Aming Wu, Yang Li, Yaowei Wang, and Yahong Han. Continual adaptation: Environment-conditional param- eter generation for object detection in dynamic scenarios. arXiv preprint arXiv:2506.24063, 2025. 2, 3

    work page arXiv 2025

  31. [31]

    A comprehensive sur- vey on test-time adaptation under distribution shifts.Inter- national Journal of Computer Vision, 133(1):31–64, 2025

    Jian Liang, Ran He, and Tieniu Tan. A comprehensive sur- vey on test-time adaptation under distribution shifts.Inter- national Journal of Computer Vision, 133(1):31–64, 2025. 2

    work page 2025

  32. [32]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28653–28663, 2024. 5, 6, 13, 16

    work page 2024

  33. [33]

    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. InInternational Conference on Learning Representations, 2024. 3, 13, 16

    work page 2024

  34. [34]

    Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34: 21808–21820, 2021

    Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34: 21808–21820, 2021. 2

    work page 2021

  35. [35]

    Visual prompt tun- ing in null space for continual learning.Advances in Neural Information Processing Systems, 37:7878–7901, 2024

    Yue Lu, Shizhou Zhang, De Cheng, Yinghui Xing, Nannan Wang, Peng Wang, and Yanning Zhang. Visual prompt tun- ing in null space for continual learning.Advances in Neural Information Processing Systems, 37:7878–7901, 2024. 1

    work page 2024

  36. [36]

    Surgeon: Memory-adaptive fully test-time adapta- tion via dynamic activation sparsity

    Ke Ma, Jiaqi Tang, Bin Guo, Fan Dang, Sicong Liu, Zhui Zhu, Lei Wu, Cheng Fang, Ying-Cong Chen, Zhiwen Yu, et al. Surgeon: Memory-adaptive fully test-time adapta- tion via dynamic activation sparsity. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 30514–30523, 2025. 2

    work page 2025

  37. [37]

    Univer- sal test-time adaptation through weight ensembling, diver- sity weighting, and prior correction

    Robert A Marsden, Mario D ¨obler, and Bin Yang. Univer- sal test-time adaptation through weight ensembling, diver- sity weighting, and prior correction. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2555–2565, 2024. 17

    work page 2024

  38. [38]

    Tipi: Test time adaptation with transforma- tion invariance

    A Tuan Nguyen, Thanh Nguyen-Tang, Ser-Nam Lim, and Philip HS Torr. Tipi: Test time adaptation with transforma- tion invariance. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24162– 24171, 2023. 2

    work page 2023

  39. [39]

    Maintaining consistent inter-class topol- ogy in continual test-time adaptation

    Chenggong Ni, Fan Lyu, Jiayao Tan, Fuyuan Hu, Rui Yao, and Tao Zhou. Maintaining consistent inter-class topol- ogy in continual test-time adaptation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15319–15328, 2025. 3, 5, 14

    work page 2025

  40. [40]

    Diffusion models for adversarial purification

    Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Animashree Anandkumar. Diffusion models for adversarial purification. InInternational Conference on Ma- chine Learning, pages 16805–16827. PMLR, 2022. 2

    work page 2022

  41. [41]

    Effec- tive restoration of source knowledge in continual test time adaptation

    Fahim Faisal Niloy, Sk Miraj Ahmed, Dripta S Raychaud- huri, Samet Oymak, and Amit K Roy-Chowdhury. Effec- tive restoration of source knowledge in continual test time adaptation. InProceedings of the IEEE/CVF Winter Confer- ence on Applications of Computer Vision, pages 2091–2100,

    work page 2091

  42. [42]

    Adaxpert: Adapting neural architecture for growing data

    Shuaicheng Niu, Jiaxiang Wu, Guanghui Xu, Yifan Zhang, Yong Guo, Peilin Zhao, Peng Wang, and Mingkui Tan. Adaxpert: Adapting neural architecture for growing data. In International conference on machine learning, pages 8184–

    work page

  43. [43]

    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. InInter- national Conference on Machine Learning, pages 16888– 16905. PMLR, 2022. 2, 5, 13, 15, 17

    work page 2022

  44. [44]

    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. InInternational Conference on Learning Representations, 2023. 2, 16, 17

    work page 2023

  45. [45]

    Test-time model adaptation with only forward passes.arXiv preprint arXiv:2404.01650, 2024

    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. 2

    work page arXiv 2024

  46. [46]

    Rdumb: A simple approach that questions our progress in continual test-time adaptation.Advances in Neural Information Processing Systems, 36:39915–39935,

    Ori Press, Steffen Schneider, Matthias K ¨ummerer, and Matthias Bethge. Rdumb: A simple approach that questions our progress in continual test-time adaptation.Advances in Neural Information Processing Systems, 36:39915–39935,

    work page

  47. [47]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684–10695, 2022. 4, 6, 8, 13, 14

    work page 2022

  48. [48]

    Acdc: The adverse conditions dataset with correspondences for se- mantic driving scene understanding

    Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for se- mantic driving scene understanding. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 10765–10775, 2021. 1

    work page 2021

  49. [49]

    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 10 Conference on Computer Vision and Pattern Recognition, pages 11920–11929, 2023. 3, 16

    work page 2023

  50. [50]

    Test-time training with self- supervision for generalization under distribution shifts

    Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self- supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229–

    work page

  51. [51]

    Uncertainty- calibrated test-time model adaptation without forgetting

    Mingkui Tan, Guohao Chen, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Peilin Zhao, and Shuaicheng Niu. Uncertainty- calibrated test-time model adaptation without forgetting. arXiv preprint arXiv:2403.11491, 2024. 1

    work page arXiv 2024

  52. [52]

    Stablerep: Synthetic images from text-to- image models make strong visual representation learners

    Yonglong Tian, Lijie Fan, Phillip Isola, Huiwen Chang, and Dilip Krishnan. Stablerep: Synthetic images from text-to- image models make strong visual representation learners. Advances in Neural Information Processing Systems, 36: 48382–48402, 2023. 3

    work page 2023

  53. [53]

    Learning vision from mod- els rivals learning vision from data

    Yonglong Tian, Lijie Fan, Kaifeng Chen, Dina Katabi, Dilip Krishnan, and Phillip Isola. Learning vision from mod- els rivals learning vision from data. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15887–15898, 2024. 3

    work page 2024

  54. [54]

    Gda: Generalized diffusion for robust test-time adaptation

    Yun-Yun Tsai, Fu-Chen Chen, Albert YC Chen, Junfeng Yang, Che-Chun Su, Min Sun, and Cheng-Hao Kuo. Gda: Generalized diffusion for robust test-time adaptation. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23242–23251, 2024. 2

    work page 2024

  55. [55]

    Preserving clusters in prompt learning for unsupervised domain adaptation

    Tung-Long Vuong, Hoang Phan, Vy V o, Anh Bui, Thanh- Toan Do, Trung Le, and Dinh Phung. Preserving clusters in prompt learning for unsupervised domain adaptation. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19974–19984, 2025. 2

    work page 2025

  56. [56]

    Tent: Fully test-time adaptation by entropy minimization

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Ol- shausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. InInternational Conference on Learning Representations, 2021. 2, 5, 13, 15, 16, 17

    work page 2021

  57. [57]

    Effortless active label- ing for long-term test-time adaptation

    Guowei Wang and Changxing Ding. Effortless active label- ing for long-term test-time adaptation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25633–25642, 2025. 3

    work page 2025

  58. [58]

    Paid: Pair- wise angular-invariant decomposition for continual test-time adaptation.arXiv preprint arXiv:2506.02453, 2025

    Kunyu Wang, Xueyang Fu, Yuanfei Bao, Chengjie Ge, Chengzhi Cao, Wei Zhai, and Zheng-Jun Zha. Paid: Pair- wise angular-invariant decomposition for continual test-time adaptation.arXiv preprint arXiv:2506.02453, 2025. 3, 16

    work page arXiv 2025

  59. [59]

    Efficient test-time adap- tive object detection via sensitivity-guided pruning

    Kunyu Wang, Xueyang Fu, Xin Lu, Chengjie Ge, Chengzhi Cao, Wei Zhai, and Zheng-Jun Zha. Efficient test-time adap- tive object detection via sensitivity-guided pruning. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10577–10586, 2025. 2

    work page 2025

  60. [60]

    Continual test-time domain adaptation

    Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201–7211, 2022. 1, 2, 3, 5, 13, 16, 17

    work page 2022

  61. [61]

    Feature alignment and uniformity for test time adap- tation

    Shuai Wang, Daoan Zhang, Zipei Yan, Jianguo Zhang, and Rui Li. Feature alignment and uniformity for test time adap- tation. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 20050–20060,

    work page

  62. [62]

    Isolation and impartial aggre- gation: A paradigm of incremental learning without interfer- ence

    Yabin Wang, Zhiheng Ma, Zhiwu Huang, Yaowei Wang, Zhou Su, and Xiaopeng Hong. Isolation and impartial aggre- gation: A paradigm of incremental learning without interfer- ence. InProceedings of the AAAI Conference on Artificial Intelligence, pages 10209–10217, 2023. 3

    work page 2023

  63. [63]

    Continual test-time domain adaptation via dynamic sample selection

    Yanshuo Wang, Jie Hong, Ali Cheraghian, Shafin Rahman, David Ahmedt-Aristizabal, Lars Petersson, and Mehrtash Harandi. Continual test-time domain adaptation via dynamic sample selection. InProceedings of the IEEE/CVF Win- ter Conference on Applications of Computer Vision (WACV), pages 1701–1710, 2024. 2

    work page 2024

  64. [64]

    Distribution align- ment for fully test-time adaptation with dynamic online data streams

    Ziqiang Wang, Zhixiang Chi, Yanan Wu, Li Gu, Zhi Liu, Konstantinos Plataniotis, and Yang Wang. Distribution align- ment for fully test-time adaptation with dynamic online data streams. InEuropean Conference on Computer Vision, pages 332–349. Springer, 2024. 2

    work page 2024

  65. [65]

    Pytorch image models.https : / / github

    Ross Wightman. Pytorch image models.https : / / github . com / rwightman / pytorch - image - models, 2019. 6

    work page 2019

  66. [66]

    Synthetic data is an elegant gift for continual vision-language models

    Bin Wu, Wuxuan Shi, Jinqiao Wang, and Mang Ye. Synthetic data is an elegant gift for continual vision-language models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2813–2823, 2025. 3

    work page 2025

  67. [67]

    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,

    work page arXiv

  68. [68]

    Alvarez, and Ping Luo

    Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. Segformer: Simple and effi- cient design for semantic segmentation with transformers. In Advances in Neural Information Processing Systems, 2021. 16

    work page 2021

  69. [69]

    D3still: Decoupled differential distillation for asymmetric image retrieval

    Yi Xie, Yihong Lin, Wenjie Cai, Xuemiao Xu, Huaidong Zhang, Yong Du, and Shengfeng He. D3still: Decoupled differential distillation for asymmetric image retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 17181–17190, 2024. 1

    work page 2024

  70. [70]

    Exploring sparse visual prompt for domain adaptive dense prediction

    Senqiao Yang, Jiarui Wu, Jiaming Liu, Xiaoqi Li, Qizhe Zhang, Mingjie Pan, Yulu Gan, Zehui Chen, and Shanghang Zhang. Exploring sparse visual prompt for domain adaptive dense prediction. InProceedings of the AAAI Conference on Artificial Intelligence, pages 16334–16342, 2024. 16

    work page 2024

  71. [71]

    Exploring safety supervision for continual test-time domain adaptation

    Xu Yang, Yanan Gu, Kun Wei, and Cheng Deng. Exploring safety supervision for continual test-time domain adaptation. InProceedings of the International Joint Conference on Ar- tificial Intelligence, pages 1649–1657, 2023. 2

    work page 2023

  72. [72]

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

    Xu Yang, Xuan Chen, Moqi Li, Kun Wei, and Cheng Deng. A versatile framework for continual test-time domain adap- tation: Balancing discriminability and generalizability. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 23731–23740, 2024. 1

    work page 2024

  73. [73]

    Fda: Fourier domain adaptation for semantic segmentation

    Yanchao Yang and Stefano Soatto. Fda: Fourier domain adaptation for semantic segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4085–4095, 2020. 4

    work page 2020

  74. [74]

    Socialized learning: Making each other better through multi-agent collaboration

    Xinjie Yao, Yu Wang, Pengfei Zhu, Wanyu Lin, Jialu Li, Weihao Li, and Qinghua Hu. Socialized learning: Making each other better through multi-agent collaboration. InIn- 11 ternational Conference on Machine Learning, pages 56927– 56945. PMLR, 2024. 2

    work page 2024

  75. [75]

    Jayeon Yoo, Dongkwan Lee, Inseop Chung, Donghyun Kim, and Nojun Kwak. What how and when should object detec- tors update in continually changing test domains? InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23354–23363, 2024. 2

    work page 2024

  76. [76]

    Robust test- time adaptation in dynamic scenarios

    Longhui Yuan, Binhui Xie, and Shuang Li. Robust test- time adaptation in dynamic scenarios. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15922–15932, 2023. 17

    work page 2023

  77. [77]

    Dca: Dividing and conquering amnesia in incremental object detection

    Aoting Zhang, Dongbao Yang, Chang Liu, Xiaopeng Hong, Miao Shang, and Yu Zhou. Dca: Dividing and conquering amnesia in incremental object detection. InProceedings of the AAAI Conference on Artificial Intelligence, pages 9851– 9859, 2025. 3

    work page 2025

  78. [78]

    Memo: Test time robustness via adaptation and augmentation.Ad- vances in Neural Information Processing Systems, 35: 38629–38642, 2022

    Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation.Ad- vances in Neural Information Processing Systems, 35: 38629–38642, 2022. 2

    work page 2022

  79. [79]

    Revisiting generative replay for class incremental object detection

    Shizhou Zhang, Xueqiang Lv, Yinghui Xing, Qirui Wu, Di Xu, and Yanning Zhang. Revisiting generative replay for class incremental object detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20340–20349, 2025. 3

    work page 2025

  80. [80]

    Dpcore: Dynamic prompt coreset for continual test- time adaptation

    Yunbei Zhang, Akshay Mehra, Shuaicheng Niu, and Jihun Hamm. Dpcore: Dynamic prompt coreset for continual test- time adaptation. InInternational Conference on Machine Learning. PMLR, 2025. 2, 5, 6, 13, 16, 17

    work page 2025

Showing first 80 references.