Pith. sign in

REVIEW 2 major objections 6 minor 36 references

Continual test-time adaptation can be organized into three method families that trade accuracy against forgetting and compute under unlabeled, non-stationary streams.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-10 12:02 UTC pith:OMWCYM6F

load-bearing objection Solid, usable CTTA survey: clean three-family taxonomy, honest about CSC limits, worth citing and reading; not a new algorithm. the 2 major comments →

arxiv 2607.08164 v1 pith:OMWCYM6F submitted 2026-07-09 cs.CV

Continual Test-Time Adaptation in Computer Vision: Methods, Benchmarks, and Future Directions

Sarthak Kumar Maharana , Shambhavi Mishra , Yunbei Zhang , Shuaicheng Niu , Taki Hasan Rafi , Jihun Hamm , Marco Pedersoli , Jose Dolz
show 1 more author
Yunhui Guo
This is my paper
classification cs.CV
keywords continual test-time adaptationdistribution shiftcatastrophic forgettingerror accumulationentropy minimizationteacher-studentparameter-efficient adaptationcomputer vision
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Deep models fail when the world keeps changing after deployment, because training and test data no longer match. Continual Test-Time Adaptation (CTTA) is the setting in which a pretrained model must update itself on a stream of unlabeled, evolving target data with no source data, no labels, and no knowledge of when domains change. The two failure modes that make this hard are catastrophic forgetting of source knowledge and error accumulation from noisy self-training signals. This survey formally defines the problem, maps the different temporal patterns of domain shift used in evaluation, and groups existing methods into three families: optimization-based, parameter-efficient, and architecture-based. Comparative results on standard corruption and driving benchmarks show clear accuracy–efficiency–stability trade-offs, and the paper sketches how the same ideas must scale to foundation models and black-box systems.

Core claim

The paper establishes that CTTA is a distinct problem from ordinary test-time adaptation and continual learning, and that the existing literature can be systematically organized by a hierarchical taxonomy of three families—optimization-based (entropy minimization, pseudo-labeling, parameter restoration), parameter-efficient (normalization statistics and adaptive layer selection), and architecture-based (teacher–student, adapters, visual prompts, masked modeling)—each of which attacks forgetting and error accumulation differently under continual unlabeled domain shifts.

What carries the argument

The hierarchical taxonomy of CTTA methods. It partitions the literature by what is adapted and how updates are controlled, giving a common language for comparing objectives, parameter budgets, and architectural add-ons under the same continual stream.

Load-bearing premise

The claim that standard fixed-order, closed-set corruption benchmarks with large i.i.d. batches are a faithful enough proxy for the real-world non-stationary streams the survey aims to address.

What would settle it

A head-to-head evaluation of the same methods under continually dynamic change (varying domain durations and order), temporal correlation, and open-set conditions that reverse the accuracy ranking obtained on the standard Continual Structured Change protocol.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 6 minor

Summary. This survey formally defines Continual Test-Time Adaptation (CTTA) as online, source-free adaptation of a pretrained model to a non-stationary stream of unlabeled target domains without known task boundaries, with the dual goals of mitigating catastrophic forgetting of source knowledge and error accumulation from noisy self-training signals. It analyzes evaluation protocols by domain-shift pattern (CSC, gradual transitions, PTTA, CDC, recurring patterns), proposes a hierarchical taxonomy of methods into optimization-based (entropy minimization, pseudo-labeling, topological consistency, parameter restoration), parameter-efficient (normalization-layer adaptation, adaptive parameter selection), and architecture-based (teacher–student, adapters, visual prompting, masked modeling) families, reviews representative methods, collates comparative results on CIFAR-C, ImageNet-C, and Cityscapes→ACDC, and outlines future directions including foundation models, black-box adaptation, and non-vision modalities.

Significance. The paper fills a clear gap relative to broader TTA surveys by treating CTTA as a first-class setting with its own failure modes, protocols, and design constraints. The three-family taxonomy is consistently applied and useful for organizing a rapidly growing literature; the formal problem definition and the structured analysis of domain-shift patterns in §3 are particularly valuable. Comparative tables and the explicit discussion of which conclusions are robust versus CSC-protocol-dependent (§7.6.4) give practitioners actionable guidance. The open repository and broader-impact statement further strengthen the contribution. If the organizational claim holds—as it does under the paper’s stated scope—the survey is a timely reference and roadmap for the field.

major comments (2)
  1. §7.4, Tables 4–6: Comparative numbers are collated from original papers and selected unified re-implementations (Wang et al., 2024b; Döbler et al., 2023). Not every method appears under a single controlled protocol (batch size, learning rate, number of rounds, backbone). The paper should state more explicitly, per table or in a short protocol appendix, which entries come from a unified re-run versus the original paper, and flag any known protocol mismatches so that cross-method rankings are not over-interpreted.
  2. §4 and Figure 3: Several methods (e.g., CoTTA, RMT, ViDA, PSMT) legitimately span multiple taxonomy branches. The hierarchical claim would be stronger if the paper stated an explicit primary-assignment rule (e.g., “primary family = dominant mechanism for anti-forgetting”) and applied it uniformly, with multi-category membership only as a secondary tag. Without that rule, category-level “best in family” statements in the tables are slightly under-specified.
minor comments (6)
  1. Figure 2: The 2026* bar (24 publications, “up to Feb 2026”) is incomplete relative to other years; either extend the cutoff consistently or mark the bar more clearly as partial so growth trends are not misread.
  2. §2.3 / Table 1: The CL vs. CTTA comparison is clear, but the notation for source collections S = {p(x_s^{(k)}, y_s^{(k)})}_{k=1}^N is under-used later; a brief note that standard CTTA uses a single source model would reduce confusion.
  3. Table 3: The Src. column encoding (S / F / R / –) is dense; a one-line legend under the table (already partially present) would help readers scan memory/teacher/source assumptions more quickly.
  4. §7.5, Table 7: Night remains the hardest condition; a short qualitative note on why LayerNorm-based SegFormer breaks BN-centric methods (already stated in text) could be echoed in the table caption for self-contained reading.
  5. Scattered typos and formatting: e.g., “thatunifiesandsystematicallydiscusses” (Need for this survey), “OpenReview:https: // openreview. net/...” spacing, and occasional missing spaces after periods. A careful copy-edit pass would help.
  6. §8.2–8.3: LLM/MLLM and strict black-box directions are timely; citing a few more concurrent works (where available) or stating the literature cutoff date would help readers judge coverage.

Circularity Check

0 steps flagged

No significant circularity: literature survey organizes existing methods without deriving predictions from fitted inputs or load-bearing self-citations.

full rationale

This is a survey paper that formally defines CTTA, proposes an organizational taxonomy of three method families drawn from the literature, collates published experimental numbers on standard benchmarks (CIFAR-C, ImageNet-C, ACDC), and discusses open directions. There are no first-principles derivations, no fitted parameters presented as predictions, no uniqueness theorems, and no ansatzes smuggled via citation. Author self-citations (e.g., PALM, DPCore, OT-VP) appear only as representative instances inside the taxonomy and tables; they are not used to justify a novel central claim. Comparative results simply restate numbers from original papers or unified benchmarks. The derivation chain is therefore self-contained organizational synthesis with no reduction of outputs to inputs by construction. Score 0 is the honest finding.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 1 invented entities

As a survey the paper inherits the standard CTTA problem statement (source-free, unlabeled, online, no task boundaries) and the closed-set covariate-shift assumption used by almost all cited methods. No free parameters are fitted by the survey itself; the free parameters that appear belong to the individual methods being reviewed. Invented entities are limited to the three-family taxonomy, which is an organizational construct rather than a physical postulate.

axioms (3)
  • domain assumption Source and target share the same closed label space; only covariate shift is considered (p(y|x) fixed, p(x) changes).
    Stated in §2.4 and Table 2; semantic/open-set shift is declared out of scope.
  • domain assumption Adaptation must be strictly online (single forward-backward pass per batch) with no access to source data or ground-truth labels.
    Core CTTA definition in §2.1–2.2; distinguishes CTTA from standard continual learning and offline DA.
  • domain assumption The CSC protocol (fixed sequential corruptions, severity 5, large batches) is a valid primary evaluation setting for comparing methods.
    Adopted throughout §7; the paper itself notes that more dynamic protocols (CDC, PTTA) are harder and less covered.
invented entities (1)
  • Three-family hierarchical taxonomy (optimization-based / parameter-efficient / architecture-based) no independent evidence
    purpose: Organize the CTTA literature and enable systematic comparison.
    Introduced in §4 and Figure 3; purely organizational, no independent physical claim.

pith-pipeline@v1.1.0-grok45 · 49964 in / 2536 out tokens · 29120 ms · 2026-07-10T12:02:31.268884+00:00 · methodology

0 comments
read the original abstract

Deep neural nets achieve remarkable performance when training and test data share the same distribution, but this assumption frequently breaks in real-world deployment, where data undergoes continual distributional shifts. Continual Test-Time Adaptation (CTTA) addresses this challenge by adapting pretrained models to non-stationary target distributions on-the-fly, without access to source data or labeled targets, while mitigating two critical failure modes: catastrophic forgetting of source knowledge and error accumulation from noisy pseudo-labels over extended time horizons. In this comprehensive survey, we formally define the CTTA problem, analyze the diverse continual domain shift patterns that characterize different evaluation protocols, and propose a hierarchical taxonomy that categorizes existing methods into three families: optimization-based strategies (entropy minimization, pseudo-labeling, parameter restoration), parameter-efficient methods (normalization layer adaptation, adaptive parameter selection), and architecture-based approaches (teacher-student frameworks, adapters, visual prompting, masked modeling). We systematically review representative methods within each category and present comparative benchmarks and experimental results across standard evaluation settings. Finally, we discuss limitations of current approaches and highlight emerging research directions, including adaptation of foundation models and black-box systems, providing a roadmap for future research in robust continual test-time adaptation. We encourage visiting our repository at [https://github.com/sarthaxxxxx/Awesome-Continual-Test-Time-Adaptation](https://github.com/sarthaxxxxx/Awesome-Continual-Test-Time-Adaptation)

Figures

Figures reproduced from arXiv: 2607.08164 by Jihun Hamm, Jose Dolz, Marco Pedersoli, Sarthak Kumar Maharana, Shambhavi Mishra, Shuaicheng Niu, Taki Hasan Rafi, Yunbei Zhang, Yunhui Guo.

Figure 1
Figure 1. Figure 1: Comparison of adaptation paradigms under distribution shift. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Growth of continual test-time adaptation and related research [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Hierarchical taxonomy of representative CTTA methods. Methods are organized into three main [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Timeline of methods from 2020 to 2025. The field evolved from foundational TTA works (TTT, [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Illustration of estimating BatchNorm statistics and feature modulation at test-time. For an input [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Teacher-Student Framework: The teacher model is updated via an exponential moving average [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Adapters: Parallel branch architecture for CTTA. Lightweight adapter modules are inserted along [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visual Prompting: Input space adaptation for CTTA. The model backbone remains completely [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Masked Modeling: Uncertainty-guided reconstruction for CTTA. [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: ImageNet-C corruption examples. The 15 corruption types are grouped into four categories: Noise (Gaussian, Shot, Impulse), Blur (Defocus, Glass, Motion, Zoom), Weather (Snow, Frost, Fog, Bright￾ness), and Digital (Contrast, Elastic, Pixelate, JPEG). 27 [PITH_FULL_IMAGE:figures/full_fig_p027_10.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

36 extracted references · 36 canonical work pages · 19 internal anchors

  1. [1]

    GPT-4 Technical Report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774,

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    Towards open world recognition

    Abhijit Bendale and Terrance Boult. Towards open world recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pp. 1893–1902,

    work page 1902

  3. [3]

    On the Opportunities and Risks of Foundation Models

    36 Published in Transactions on Machine Learning Research (07/2026) Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models.arXiv preprint arXiv:2108.07258,

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  4. [4]

    Parameter-free online test-time adaptation.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 8334–8343,

    work page 2022

  5. [5]

    The intriguing role of module criticality in the generalization of deep networks

    Niladri S Chatterji, Behnam Neyshabur, and Hanie Sedghi. The intriguing role of module criticality in the generalization of deep networks.arXiv preprint arXiv:1912.00528,

    work page internal anchor Pith review Pith/arXiv arXiv 1912

  6. [6]

    On Tiny Episodic Memories in Continual Learning

    Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K Dokania, Philip HS Torr, and Marc’Aurelio Ranzato. On tiny episodic memories in continual learning.arXiv preprint arXiv:1902.10486,

    work page internal anchor Pith review Pith/arXiv arXiv 1902

  7. [7]

    Imagenet: A large-scale hierarchical image database

    37 Published in Transactions on Machine Learning Research (07/2026) Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition, pp. 248–255. Ieee,

    work page 2026

  8. [8]

    A survey on in-context learning

    Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Baobao Chang, Xu Sun, Lei Li, and Zhifang Sui. A survey on in-context learning. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics,

    work page 2024

  9. [9]

    Deep Think with Confidence

    Yichao Fu, Xuewei Wang, Yuandong Tian, and Jiawei Zhao. Deep think with confidence.arXiv preprint arXiv:2508.15260,

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    MIT press,

    38 Published in Transactions on Machine Learning Research (07/2026) Ian Goodfellow, Yoshua Bengio, and Aaron Courville.Deep learning. MIT press,

    work page 2026

  11. [11]

    An Empirical Investigation of Catastrophic Forgetting in Gradient-Based Neural Networks

    Ian J Goodfellow, Mehdi Mirza, Da Xiao, Aaron Courville, and Yoshua Bengio. An empirical investigation of catastrophic forgetting in gradient-based neural networks.arXiv preprint arXiv:1312.6211,

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    MixNorm: Test-Time Adaptation Through Online Normalization Estimation

    Xuefeng Hu, Gokhan Uzunbas, Sirius Chen, Rui Wang, Ashish Shah, Ram Nevatia, and Ser-Nam Lim. Mixnorm: Test-time adaptation through online normalization estimation.arXiv preprint arXiv:2110.11478,

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    Pcotta: Continual test-time adaptation for multi-task point cloud understanding

    Jincen Jiang, Qianyu Zhou, Yuhang Li, Xinkui Zhao, Meili Wang, Lizhuang Ma, Jian Chang, Jian J Zhang, and Xuequan Lu. Pcotta: Continual test-time adaptation for multi-task point cloud understanding. Advances in Neural Information Processing Systems, 37:96229–96253, 2024a. 39 Published in Transactions on Machine Learning Research (07/2026) JincenJiang, Qia...

    work page 2026

  14. [14]

    Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation

    40 Published in Transactions on Machine Learning Research (07/2026) Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. InInternational conference on machine learning, pp. 6028–

    work page 2026

  15. [15]

    Continual test-time adaptation for end-to-end speech recognition on noisy speech

    Guan-Ting Lin, Wei-Ping Huang, and Hung yi Leeg. Continual test-time adaptation for end-to-end speech recognition on noisy speech. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics,

    work page 2024

  16. [16]

    ETTRL: Balancing Exploration and Exploitation in LLM Test-Time Reinforcement Learning Via Entropy Mechanism

    Jia Liu, ChangYi He, YingQiao Lin, MingMin Yang, FeiYang Shen, and ShaoGuo Liu. Ettrl: Balancing exploration and exploitation in llm test-time reinforcement learning via entropy mechanism.arXiv preprint arXiv:2508.11356, 2025a. Jiaming Liu, Ran Xu, Senqiao Yang, Renrui Zhang, Qizhe Zhang, Zehui Chen, Yandong Guo, and Shanghang Zhang. Continual-mae: Adapti...

    work page internal anchor Pith review Pith/arXiv arXiv

  17. [17]

    Ctta-t: Continual test-time adaptation for text understanding via teacher-student with a domain-aware and generalized teacher.arXiv preprint arXiv:2512.18321, 2025b

    Tianlun Liu, Zhiliang Tian, Zhen Huang, Xingzhi Zhou, Wanlong Yu, Tianle Liu, Feng Liu, and Dongsheng Li. Ctta-t: Continual test-time adaptation for text understanding via teacher-student with a domain-aware and generalized teacher.arXiv preprint arXiv:2512.18321, 2025b. David Lopez-Paz and Marc’Aurelio Ranzato. Gradient episodic memory for continual lear...

    work page arXiv

  18. [18]

    EmbodiTTA: Resource-Efficient Test-Time Adaptation for Embodied Visual Systems

    Xiao Ma, Young D Kwon, and Dong Ma. On-demand test-time adaptation for edge devices.arXiv preprint arXiv:2505.00986,

    work page internal anchor Pith review Pith/arXiv arXiv

  19. [19]

    Audio-Visual Continual Test-Time Adaptation without Forgetting

    Sarthak Kumar Maharana, Baoming Zhang, and Yunhui Guo. Palm: Pushing adaptive learning rate mecha- nisms for continual test-time adaptation. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pp. 19378–19386, 2025a. Sarthak Kumar Maharana, Baoming Zhang, Leonid Karlinsky, Rogerio Feris, and Yunhui Guo. Batclip: Bimodal online test...

    work page internal anchor Pith review Pith/arXiv arXiv

  20. [20]

    Semantic anchor transport: Robust test-time adaptation for vision-language models,

    41 Published in Transactions on Machine Learning Research (07/2026) Shambhavi Mishra, Julio Silva-Rodriguez, Ismail Ben Ayed, Marco Pedersoli, and Jose Dolz. Semantic anchor transport: Robust test-time adaptation for vision-language models,

    work page 2026

  21. [21]

    Effective restoration of source knowledge in continual test time adaptation

    Fahim Faisal Niloy, Sk Miraj Ahmed, Dripta S Raychaudhuri, Samet Oymak, and Amit K Roy-Chowdhury. Effective restoration of source knowledge in continual test time adaptation. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp. 2091–2100,

    work page 2091

  22. [22]

    Lifelong domain adaptive 3d human pose estimation

    42 Published in Transactions on Machine Learning Research (07/2026) Qucheng Peng, Hongfei Xue, Pu Wang, and Chen Chen. Lifelong domain adaptive 3d human pose estimation. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pp. 8358–8366,

    work page 2026

  23. [23]

    If your data distribution shifts, use self-learning

    Evgenia Rusak, Steffen Schneider, George Pachitariu, Luisa Eck, Peter Gehler, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. If your data distribution shifts, use self-learning.arXiv preprint arXiv:2104.12928,

    work page internal anchor Pith review Pith/arXiv arXiv

  24. [24]

    Empirical Analysis of the Hessian of Over-Parametrized Neural Networks

    Levent Sagun, Utku Evci, V Ugur Guney, Yann Dauphin, and Leon Bottou. Empirical analysis of the hessian of over-parametrized neural networks.arXiv preprint arXiv:1706.04454,

    work page internal anchor Pith review Pith/arXiv arXiv

  25. [25]

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

    43 Published in Transactions on Machine Learning Research (07/2026) 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, pp. 11920–11929,

    work page 2026

  26. [26]

    Less is More: Pseudo-Label Filtering for Continual Test-Time Adaptation

    Jiayao Tan, Fan Lyu, Chenggong Ni, Tingliang Feng, Fuyuan Hu, Zhang Zhang, Shaochuang Zhao, and Liang Wang. Less is more: Pseudo-label filtering for continual test-time adaptation.arXiv preprint arXiv:2406.02609,

    work page internal anchor Pith review Pith/arXiv arXiv

  27. [27]

    LLaMA: Open and Efficient Foundation Language Models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971,

    work page internal anchor Pith review Pith/arXiv arXiv

  28. [28]

    Instance Normalization: The Missing Ingredient for Fast Stylization

    Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization.arXiv preprint arXiv:1607.08022,

    work page internal anchor Pith review Pith/arXiv arXiv

  29. [29]

    Test-time adaptation for geospatial point cloud semantic segmentation with distinct domain shifts.ISPRS Journal of Photogrammetry and Remote Sensing, 229: 422–435, 2025d

    44 Published in Transactions on Machine Learning Research (07/2026) Puzuo Wang, Wei Yao, Jie Shao, and Zhiyi He. Test-time adaptation for geospatial point cloud semantic segmentation with distinct domain shifts.ISPRS Journal of Photogrammetry and Remote Sensing, 229: 422–435, 2025d. Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time d...

    work page 2026

  30. [30]

    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 internal anchor Pith review Pith/arXiv arXiv

  31. [31]

    Robust test-time adaptation in dynamic scenarios

    45 Published in Transactions on Machine Learning Research (07/2026) Longhui Yuan, Binhui Xie, and Shuang Li. Robust test-time adaptation in dynamic scenarios. InCVPR,

    work page 2026

  32. [32]

    Wide Residual Networks

    Sergey Zagoruyko and Nikos Komodakis. Wide residual networks.arXiv preprint arXiv:1605.07146,

    work page internal anchor Pith review Pith/arXiv arXiv

  33. [33]

    Right Question is Already Half the Answer: Fully Unsupervised LLM Reasoning Incentivization

    Qingyang Zhang, Haitao Wu, Changqing Zhang, Peilin Zhao, and Yatao Bian. Right question is already half the answer: Fully unsupervised llm reasoning incentivization.arXiv preprint arXiv:2504.05812, 2025a. YufeiZhang, YichengXu, HongxinWei, ZhipingLin, XiaofengZou, CenChen, andHuipingZhuang. Analytic continualtest-timeadaptationformulti-modalitycorruption....

    work page internal anchor Pith review Pith/arXiv arXiv 1929

  34. [34]

    Learning to Reason without External Rewards

    Xuandong Zhao, Zhewei Kang, Aosong Feng, Sergey Levine, and Dawn Song. Learning to reason without external rewards.arXiv preprint arXiv:2505.19590, 2025b. Jiahuan Zhou, Chao Zhu, Zhenyu Cui, Zichen Liu, Xu Zou, and Gang Hua. Class-aware domain knowledge fusion and fission for continual test-time adaptation. InThe Thirty-ninth Annual Conference on Neural I...

    work page internal anchor Pith review Pith/arXiv arXiv

  35. [35]

    TTRL: Test-time reinforcement learning

    46 Published in Transactions on Machine Learning Research (07/2026) Yuxin Zuo, Kaiyan Zhang, Li Sheng, Shang Qu, Ganqu Cui, Xuekai Zhu, Haozhan Li, Yuchen Zhang, Xinwei Long, Ermo Hua, Biqing Qi, Youbang Sun, Zhiyuan Ma, Lifan Yuan, Ning Ding, and Bowen Zhou. TTRL: Test-time reinforcement learning. InThe Thirty-ninth Annual Conference on Neural Informatio...

    work page 2026

  36. [36]

    While this capability addresses genuine and pressing challenges in robust deployment, it also introduces a set of ethical, safety, and societal considerations that we discuss below

    47 Published in Transactions on Machine Learning Research (07/2026) A Broader Impact Statement Our comprehensive survey on continual test-time adaptation (CTTA) focuses on self-adapting models to continual target or test distributions. While this capability addresses genuine and pressing challenges in robust deployment, it also introduces a set of ethical...

    work page 2026