Pith. sign in

REVIEW 4 major objections 4 minor 50 references

SloMo-Fast: Slow-Momentum and Fast-Adaptive Teachers for Source-Free Continual Test-Time Adaptation

T0 review · 4 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read SloMo-Fast claims a source-free dual-teacher framework that adapts quickly to new domains and retains knowledge of old ones, outperforming existing methods across 11 continual test-time adaptation settings, including a new cyclic-repetition

desk verdict Dual-teacher CTTA idea is real, but the cyclic benchmark formula is a no-op and the cyclic tables contain implausible entries; the SOTA claims are not supported as submitted. read the letter →

arxiv 2511.18468 v2 pith:HCHSM6EL submitted 2025-11-23 cs.LG cs.CV

Md Akil Raihan Iftee , Mir Sazzat Hossain , Rakibul Hasan Rajib , Tariq Iqbal , Md Mofijul Islam , M Ashraful Amin , Amin Ahsan Ali , AKM Mahbubur Rahman This is my paper
classification cs.LGcs.CV
keywords source-freecontinualtest-timeadaptationdual-teacherframeworkdynamicclassprototypescycliccatastrophicforgettingpseudo-labelreliabilitycontrastivelearningonline
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

Continual test-time adaptation lets a deployed model keep learning from a stream of unlabeled, shifting domains, but models typically forget old domains, adapt too slowly to new ones, and rely on source data. SloMo-Fast is an attempt to solve all three at once with two teachers: a fast teacher that chases the current domain and a slow teacher that holds onto general, cross-domain knowledge. The paper claims this pairing, guided by class prototypes built on the fly from confident test features, consistently beats existing methods across 11 settings, including a new cyclic-repetition benchmark that mimics real-world recurring shifts such as weather or lighting cycles. If true, it would make test-time adaptation practical in privacy-sensitive, source-free deployment.

What carries the argument

The load-bearing mechanism is the test-time class prototype: class-specific priority queues that keep only T1 features with low prediction entropy (confidence) and high prediction sensitivity (stability under augmentation), then merge them into entropy-weighted prototypes. Those prototypes anchor T2's contrastive and MSE losses, letting the slow teacher learn domain-invariant features; a stochastic restoration step partially resets weights to the pretrained model to guard against catastrophic forgetting.

What would settle it

Run the cyclic benchmark with a labeled holdout stream and compute the true error of T1's pseudo-labels as they enter the priority queues; if natural contamination approaches the 20–40% injected-noise range, SloMo-Fast's advantage over simpler baselines should erode. Alternatively, start the cycle with a domain where T1 is wrong on a majority of high-confidence samples and check whether the error rate fails to recover on the second cycle.

Watch

Extended reading notes

Core claim

At the center of SloMo-Fast is a division of labor. The fast teacher (T1) is refreshed by an exponential moving average of the student, so it tracks the current domain cheaply and without backpropagation. The slow teacher (T2) is trained only through its batch-normalization layers on three objectives: a contrastive loss that pulls sample features toward dynamically generated class prototypes, an MSE loss that aligns features to the nearest prototype, and an information-maximization loss that keeps predictions confident but diverse. The prototypes themselves come from T1: only features with low entropy and high prediction sensitivity are stored in per-class priority queues and averaged with e

Load-bearing premise

The entire method leans on the assumption that the fast teacher's high-confidence, low-sensitivity features are genuinely correct: if early pseudo-labels are wrong, the prototypes they seed will mislead the contrastive and MSE losses and lock in the error—the paper's own robustness study (Fig. A.6) tests only injected noise, not naturally occurring contamination under cyclic shift.

Editorial extensions

If this is right

  • Models could be updated continuously in the field without access to the original training data or held-out prototypes, easing privacy and storage constraints.
  • When a previously seen domain returns, the model should recover quickly instead of re-adapting from scratch—directly testable on the new Cyclic-TTA benchmark.
  • The BN-only variant, which trains only 4.9% of parameters, is reported to be competitive with methods that update all parameters, suggesting lightweight adaptation is sufficient.
  • The adaptation-rate metric (time-to-plateau, average positive slope, stability) offers a standardized way to measure how fast an adapter reacts to domain shifts.
  • The paper's across-the-board gains on CIFAR-10C, CIFAR-100C, and ImageNet-C imply the approach transfers across corruption types and model architectures.

Reading between the lines

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

  • The central assumption—early T1 pseudo-labels are reliable enough to seed prototypes—is tested only with injected label noise; a natural test would measure the actual contamination rate of the priority queues under cyclic shift, since the contrastive and MSE losses will amplify whatever errors enter them.
  • The slow/fast decomposition suggests a general recipe for online learning under drift: keep an agile copy for the present and a slowly moving copy as a memory anchor; this could transfer to robotic continual learning or personalized medical imaging where domains recur.
  • The periodic removal of the lowest-entropy feature from each queue is a pragmatic heuristic; one could test whether diversity-aware replacement (e.g., rewarding feature spread) improves robustness when early domains are unrepresentative.
  • Cyclic-TTA, as formulated with variable cycle repetition rate, could also be used to stress-test reset-based baselines—methods that periodically reinitialize will lose exactly the knowledge the benchmark rewards retaining.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes SloMo-Fast, a source-free continual test-time adaptation (CTTA) method based on a dual-teacher and student architecture. The Fast-Teacher (T1) is updated via EMA of the student and supplies high-confidence features for dynamic class-prototype construction; the Slow-Teacher (T2) is trained with contrastive, MSE, and information-maximization losses to retain long-term knowledge. The paper also introduces a new evaluation protocol, Cyclic-TTA, intended to model recurring domain shifts. The central claim is that SloMo-Fast consistently outperforms state-of-the-art methods across 11 CTTA settings on CIFAR10-C, CIFAR100-C, ImageNet-C, ImageNet-R, and ImageNet-Sketch, with a mean error rate of 34.7%. The supplementary material provides extensive per-domain tables, ablations, and robustness experiments.

Significance. If the claims were validated, the dual-teacher design and test-time prototype generation would be a useful contribution to source-free CTTA, and the Cyclic-TTA benchmark would fill a real gap by testing domain recurrence. The paper is also commendable for the breadth of its evaluation and the detail of its supplementary tables. However, the central claims are undermined by a degenerate definition of the new benchmark, internally inconsistent reported numbers, and hyperparameter selection on the same benchmarks used for evaluation. These issues are load-bearing: they affect the headline SOTA claim and the novelty of the proposed setting. The strengths of the method cannot be assessed reliably until these problems are resolved.

major comments (4)
  1. [§3.4] The Cyclic-TTA domain index formula is degenerate. Eq. c(t) = ((t−1) mod (K·r)) mod K + 1 reduces mathematically to c(t) = (t−1 mod K) + 1 for every integer r, because (a mod (K·r)) mod K = a mod K. Thus r has no effect on the stream; the promised 'variable-length intervals before a domain reappears' do not exist. The benchmark is exactly a fixed-period repetition with period K, already covered by ordinary continual settings. This invalidates the novelty of Cyclic-TTA and the interpretation of all cyclic results as reflecting variable recurrence.
  2. [Table A.34] The per-domain cyclic results for SloMo-Fast* on CIFAR100-C are implausibly flat: defocus 27.01, glass 27.04, motion 26.89, snow 26.98, fog 26.89. The source model's per-domain errors (Table A.8) for these corruptions are 29.3, 54.1, 30.8, 39.4, and 30.5 respectively, spanning a wide range. Independently measured online-adaptation errors would not collapse to within 0.2 across such different corruptions unless the entries are aggregated or copied. Since the cyclic SOTA numbers (27.5% mean error, Table 1) are part of the headline claim, this inconsistency makes the cyclic results unreliable.
  3. [§4.6 / Fig. 7 / Fig. A.5] Hyperparameters (σ, δ, α, λ) are selected using sensitivity curves computed on the same CIFAR10-C, CIFAR100-C, and ImageNet-C benchmarks whose results are then reported as the method's performance. §4.6 states the 'optimal hyperparameters' are chosen from these figures. This is selection on the evaluation set, which can inflate reported margins and makes the comparisons to baselines (which may not have been tuned on the same folds) unfair. The method should be evaluated with hyperparameters fixed a priori or chosen on a separate validation split.
  4. [§4.4, Eq. (17)] The Adaptation Rate metric is defined as APS/TTP − λ·STD with a free parameter λ that is not specified or justified, and the moving-average window k used to compute TTP, APS, and STD is also not defined. The claim in Fig. 3 that SloMo-Fast has the 'best adaptation rate' is therefore not a reproducible or falsifiable comparison. Without a fixed metric, the figure only shows a qualitative curve, not a quantitative superiority result.
minor comments (4)
  1. [Table 1] There are typographical errors in the table, e.g., the IN-C Episodic row for TENT reads '57.36.70' rather than a valid number, and '6.70' appears without a decimal point consistent with adjacent entries. These errors obscure the reported baseline numbers.
  2. [Table A.18] The SloMo-Fast row for ImageNet-C Digital shows a Cycle 1 average of 25.01, but Table A.32 reports brightness 33.61 and contrast 56.98 for the same group, whose average is 45.30. The table header also appears to list 'CIFAR100-C CIFAR10-C' with swapped column ordering. Such inconsistencies call into question the accuracy of the aggregated tables.
  3. [Throughout] Notation is inconsistent: 'ROID' is sometimes written 'Roid', 'SloMo-Fast' appears as 'Slomo-Fast', and Figure A.3's caption refers to 'PA' instead of SloMo-Fast. The heading of Section C.5 contains the typo 'Catastrophic Fogetting'.
  4. [§A.3 / §4.2] The definitions of some settings in the supplementary material are terse. For example, 'Continual-Cross Group' and 'Hard2Easy' would benefit from explicit stream construction formulas, as the Cyclic-TTA definition in the main text is the only one given formally.

Circularity Check

1 steps flagged · score 3.0 of 10

Cyclic-TTA degenerates to a fixed-period stream because r drops out by modular arithmetic, so the 'novel' benchmark is a relabeling of an existing continual setting; the core SloMo-Fast derivation is otherwise not circular.

  1. renaming known result [Section 3.4, Cyclic Test-Time Adaptation (Cyclic-TTA), c(t) definition]
    "At each test-time step t, the model receives a sample x_t ∼ p(x|G_{c(t)}), where the group index cycles as c(t) = ((t−1) mod (K·r)) mod K + 1, and r denotes the cycle repetition rate, allowing variable-length intervals before a domain reappears."

    By modular arithmetic, (a mod (K·r)) mod K = a mod K for integer a, so c(t) = (t−1 mod K) + 1 independent of r. The claimed 'variable-length intervals before a domain reappears' do not exist; the stream is exactly the fixed-period repetition with period K, which is already a standard continual setting. Thus Cyclic-TTA is not a new benchmark with variable recurrence—it is a relabeling of the ordinary fixed-order stream, and results presented as testing a novel cyclic regime are not supported as new.

full rationale

The central SloMo-Fast derivation is not circular. The dual-teacher losses (Eqs. 2, 10, 11, 12), prototype construction from T1 features with entropy/PLPD filtering (Eqs. 7–8), and the final student+T2 ensemble (Eq. 14) are self-contained: they are defined independently of the reported test errors, and the reported accuracy is measured on external corruption benchmarks rather than fitted to the target labels. No load-bearing self-citation chain was found; the method cites external prior work (DeYo for PLPD, RMT for contrastive alignment, CoTTA for stochastic restoration) without making its central claim reduce to those citations. However, one contribution-level reduction is present: the Cyclic-TTA benchmark equation collapses to a fixed-period stream, so the 'novel' benchmark claim and the associated cyclic-domain SOTA numbers do not test a genuinely new setting. Additionally, the per-domain cyclic errors for SloMo-Fast* in Table A.34 are implausibly flat (e.g., defocus 27.01, glass 27.04, motion 26.89, snow 26.98, fog 26.89) compared to the source model's large per-domain error range (9.3–73.0 in Table A.8), suggesting those entries may not be independently measured—a correctness concern, though not itself a circularity. Hyperparameters in Fig. 7 appear to be evaluated on the same corruption benchmarks as the headline results without an explicit validation split, which is a leakage risk but is not demonstrated to be a fitted-input-called-prediction step. Overall, the method's core adaptation pipeline is independent; the score reflects the degenerate/renamed Cyclic-TTA setting and the resulting unsupported novelty, not a collapse of the main prediction chain.

Assumptions & free parameters 10 free parameters · 6 assumptions · 2 invented entities

The paper's claims rest on several tuned hyperparameters, many of which are selected on the same benchmarks used for final evaluation. The Cyclic-TTA benchmark is a new evaluation construct without external validation. No new physical entities are postulated; the invented entities are methodological.

free parameters (10)
  • EMA smoothing factor α = 0.01
    Fast-teacher update rate Eq. 3; tuned per Fig. 7 on CIFAR10-C.
  • Entropy threshold σ = 0.5
    Prototype selection criterion Eq. 5/7; sensitivity analysis on CIFAR10-C (Fig. 7).
  • PLPD/sensitivity threshold δ = 0.2
    Prototype stability criterion Eq. 6/7; tuned on CIFAR10-C.
  • Loss weights λ_cl, λ_mse, λ_im = 1, 1, 1
    Weights for T2 objective Eq. 13; sensitivity analysis Fig. A.5 shows 1 is best on the evaluated datasets.
  • Priority queue size K = 10
    Main text uses 10, but ablation Table A.5 does not list 10 and shows better performance at 25 (CIFAR10-C) and 100 (CIFAR100-C).
  • Contrastive temperature τ = unspecified
    Eq. 10 uses τ but no value is reported.
  • Stochastic restoration probability = unspecified
    CoTTA-style restoration of source weights (§3.2) requires a restore probability that is not stated.
  • Queue maintenance interval p = unspecified
    Periodic removal of lowest-entropy element (Algorithm 1, §3.2); p not specified.
  • Adaptation-rate stability regularizer λ = unspecified
    Metric Eq. 17 uses λ·STD; value not given.
  • Cycle repetition rate r = unspecified
    Cyclic-TTA definition Eq. 17; r is not instantiated in experiments.
assumptions (6)
  • domain assumption Entropy threshold σ and sensitivity threshold δ select reliable pseudo-labels
    Section 3.2 Eq. 5-7; the core premise that low-entropy, stable features are trustworthy under shift.
  • domain assumption Class prototypes from target features are valid proxies for source class boundaries
    Section 3.2 Eq. 8; the method replaces source prototypes with queues from T1 without proof that these anchors remain class-discriminative.
  • domain assumption Batch-mean softmax approximates near-uniform prior for prior correction
    Eq. 15-16; assumes information maximization ensures uniform class distribution, which may fail on imbalanced batches.
  • domain assumption BN-only updates of the slow teacher preserve enough adaptation capacity
    Section 3/Table 3; T2 updates only batch norm layers, yet is claimed to learn domain-invariant features.
  • ad hoc to paper Cyclic corruption groups reflect realistic recurring domains
    Section 3.4; the Cyclic-TTA benchmark assumes weather/digital groups recur with periodicity; no real-world validation is provided.
  • ad hoc to paper Adaptation Rate metric APS/TTP - λ·STD is a valid measure of adaptation speed
    Eq. 17 (Section 4.4); introduced without derivation or specification of λ or window size k.
invented entities (2)
  • Cyclic-TTA benchmark
    purpose: Evaluation setting where corruption groups repeat (noise, blur, weather, digital, distortion) to test domain recurrence
    No external validation that this construction matches real recurring deployment; the definition's formula is degenerate for integer r.
  • Fast-Teacher / Slow-Teacher pair
    purpose: Dual model components to balance adaptation speed and knowledge retention
    The benefit of the two-teacher separation is demonstrated only on the paper's own benchmarks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SloMo-Fast: Slow-Momentum and Fast-Adaptive Teachers for Source-Free Continual Test-Time Adaptation." pith.science (2026). https://pith.science/paper/HCHSM6EL

@misc{pith2026251118468,
  author       = {Pith},
  title        = {Pith review of: SloMo-Fast: Slow-Momentum and Fast-Adaptive Teachers for Source-Free Continual Test-Time Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HCHSM6EL}},
  note         = {Machine review of arXiv:2511.18468}
}
read the original abstract

Continual Test-Time Adaptation (CTTA) is crucial for deploying models in real-world applications with unseen, evolving target domains. Existing CTTA methods, however, often rely on source data or prototypes, limiting their applicability in privacy-sensitive and resource-constrained settings. Although several methods attempt to mitigate catastrophic forgetting, they often fail to preserve long-term domain-specific knowledge across many domain shifts. Moreover, their relatively slow adaptation rates during domain transitions can cause error accumulation, allowing mistakes to propagate before effective adaptation occurs. To address these challenges, we propose SloMo-Fast, a source-free, dual-teacher CTTA framework designed for enhanced quick adaptability and generalization. It includes two complementary teachers: the Slow-Teacher, which exhibits slow forgetting and retains long-term knowledge of previously encountered domains to ensure robust generalization, and the Fast-Teacher rapidly adapts to new domains while accumulating and integrating knowledge across them. This framework preserves knowledge of past domains and adapts efficiently to new ones. Our extensive experiments show that SloMo-Fast consistently outperforms state-of-the-art methods across Cyclic Test-Time Adaptation (Cyclic-TTA), a CTTA benchmark that simulates recurring domain shifts, along with ten other CTTA settings, highlighting its ability to both adapt and generalize across evolving, revisited domains.

Figures

Figures reproduced from arXiv: 2511.18468 by the authors.

Figure 1
Figure 1. Overview of current CTTA approaches utilizing teacher [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The SloMo-FAST framework comprises a dual-teacher and student model. The fast teacher [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Adaptation rate over a cyclic domain shift. SloMo-Fast [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Validation accuracy on the initial domain after sequential [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: Impact of batch size on error rates (%) in the CIFAR10C [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 5
Figure 5. Figure 5: t-SNE visualization of feature representations and class [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 2 linked inside Pith

  1. [1]

    Recognition in terra incognita

    Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. InProceedings of the European confer- ence on computer vision (ECCV), pages 456–473, 2018. 2

    2018

  2. [2]

    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, 19

    2023

  3. [3]

    Improved test-time adaptation for domain generalization

    Liang Chen, Yong Zhang, Yibing Song, Ying Shan, and Lingqiao Liu. Improved test-time adaptation for domain generalization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24172– 24182, 2023. 1

    2023

  4. [4]

    Feature augmentation based test- time adaptation

    Younggeol Cho, Youngrae Kim, Junho Yoon, Seunghoon Hong, and Dongman Lee. Feature augmentation based test- time adaptation. 2024. 2

    2024

  5. [5]

    Adaptive energy alignment for accelerating test-time adapta- tion

    Wonjeong Choi, Do-Yeon Kim, Jungwuk Park, Jungmoon Lee, Younghyun Park, Dong-Jun Han, and Jaekyun Moon. Adaptive energy alignment for accelerating test-time adapta- tion. InThe Thirteenth International Conference on Learn- ing Representations. 2

  6. [6]

    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 Computer Vision and Pattern Recognition Conference, pages 30566–30576, 2025. 1

    2025

  7. [7]

    Learning to gen- erate gradients for test-time adaptation via test-time training layers

    Qi Deng, Shuaicheng Niu, Ronghao Zhang, Yaofo Chen, Runhao Zeng, Jian Chen, and Xiping Hu. Learning to gen- erate gradients for test-time adaptation via test-time training layers. InProceedings of the AAAI Conference on Artificial Intelligence, pages 16235–16243, 2025. 2, 3

    2025

  8. [8]

    Marsden, and Bin Yang

    Mario D ¨obler, Robert A. Marsden, and Bin Yang. Robust mean teacher for continual and gradual test-time adaptation. InCVPR, pages 7704–7714, 2023. 2, 3, 5, 1, 19

    2023

Show all 50 references
  1. [9]

    Ranked entropy minimization for continual test-time adaptation.Interna- tional conference on machine learning, 2025

    Jisu Han, Jaemin Na, and Wonjun Hwang. Ranked entropy minimization for continual test-time adaptation.Interna- tional conference on machine learning, 2025. 2

    2025

  2. [10]

    Benchmarking neu- ral network robustness to common corruptions and perturba- tions.arXiv preprint arXiv:1903.12261, 2019

    Dan Hendrycks and Thomas Dietterich. Benchmarking neu- ral network robustness to common corruptions and perturba- tions.arXiv preprint arXiv:1903.12261, 2019. 6

    1903 arXiv

  3. [11]

    The many faces of robust- ness: A critical analysis of out-of-distribution generalization

    Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kada- vath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robust- ness: A critical analysis of out-of-distribution generalization. InProceedings of the IEEE/CVF internationa...

    2021

  4. [12]

    Test-time classifier adjustment module for model-agnostic domain generaliza- tion.Advances in Neural Information Processing Systems, 34:2427–2440, 2021

    Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generaliza- tion.Advances in Neural Information Processing Systems, 34:2427–2440, 2021. 1

    2021

  5. [13]

    Test-time adaptable neural networks for robust medical image segmentation.Medical Image Analysis, 68: 101907, 2021

    Neerav Karani, Ertunc Erdil, Krishna Chaitanya, and Ender Konukoglu. Test-time adaptable neural networks for robust medical image segmentation.Medical Image Analysis, 68: 101907, 2021. 2

    2021

  6. [14]

    Universal source-free domain adaptation

    Jogendra Nath Kundu, Naveen Venkat, R Venkatesh Babu, et al. Universal source-free domain adaptation. InProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4544–4553, 2020. 1

    2020

  7. [15]

    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. pages 27072–27093, 2024. 3

    2024

  8. [16]

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

    Jonghyun Lee, Dahuin Jung, Saehyung Lee, Junsung Park, Juhyeon Shin, Uiwon Hwang, and Sungroh Yoon. Entropy is not enough for test-time adaptation: From the perspective of disentangled factors. InICLR, 2024. 1, 2, 3, 4

    2024

  9. [17]

    Stationary latent weight inference for unreliable observations from online test-time adaptation

    Jae-Hong Lee and Joon-Hyuk Chang. Stationary latent weight inference for unreliable observations from online test-time adaptation. InForty-first International Conference on Machine Learning. 1

  10. [18]

    Continual momentum filtering on parameter space for online test-time adaptation

    Jae-Hong Lee and Joon-Hyuk Chang. Continual momentum filtering on parameter space for online test-time adaptation. InICLR, 2024. 3

    2024

  11. [19]

    Source-free active domain adaptation via energy-based lo- cality preserving transfer

    Xinyao Li, Zhekai Du, Jingjing Li, Lei Zhu, and Ke Lu. Source-free active domain adaptation via energy-based lo- cality preserving transfer. InProceedings of the 30th ACM international conference on multimedia, pages 5802–5810,

  12. [20]

    Do we really need to access the source data? source hypothesis transfer for un- supervised domain adaptation

    Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for un- supervised domain adaptation. InInternational conference on machine learning, pages 6028–6039. PMLR, 2020. 5

    2020

  13. [21]

    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 Reco...

    2024

  14. [22]

    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. InICLR, 2024. 1, 2, 3, 19

    2024

  15. [23]

    Palm: Pushing adaptive learning rate mechanisms for continual test-time adaptation

    Sarthak Kumar Maharana, Baoming Zhang, and Yunhui Guo. Palm: Pushing adaptive learning rate mechanisms for continual test-time adaptation. InProceedings of the AAAI Conference on Artificial Intelligence, pages 19378–19386,

  16. [24]

    Kitting in the wild through online domain adaptation

    Massimiliano Mancini, Hakan Karaoguz, Elisa Ricci, Patric Jensfelt, and Barbara Caputo. Kitting in the wild through online domain adaptation. In2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1103–1109. IEEE, 2018. 1

    2018

  17. [25]

    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. 2, 3, 5, 6, 1, 9, 19

    2024

  18. [26]

    Marsden, Mario D ¨obler, and Bin Yang

    Robert A. Marsden, Mario D ¨obler, and Bin Yang. Introduc- ing intermediate domains for effective self-training during test-time. pages 1–10, 2024. 3

    2024

  19. [27]

    Maintaining consistent inter-class topology in continual test-time adaptation

    Chenggong Ni, Fan Lyu, Jiayao Tan, Fuyuan Hu, Rui Yao, and Tao Zhou. Maintaining consistent inter-class topology in continual test-time adaptation. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 15319–15328, 2025. 2, 3, 19

    2025

  20. [28]

    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,

  21. [29]

    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. InProceedings of the 39th International Conference on Machine Learning (ICML), pages 16888–16905. PMLR, 2022. 1, 2, 19

    2022

  22. [30]

    Towards stable test-time adaptation in dynamic wild world.arXiv preprint arXiv:2302.12400, 2023

    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. 2

    2023 arXiv

  23. [31]

    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,

  24. [32]

    Classifier adap- tation at prediction time

    Amelie Royer and Christoph H Lampert. Classifier adap- tation at prediction time. InProceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition, pages 1401–1409, 2015. 5

    2015

  25. [33]

    Memory- efficient pseudo-labeling for online source-free universal do- main adaptation using a gaussian mixture model

    Pascal Schlachter, Simon Wagner, and Bin Yang. Memory- efficient pseudo-labeling for online source-free universal do- main adaptation using a gaussian mixture model. In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 6425–6434. IEEE, 2025. 1

    2025

  26. [34]

    Ar-tta: A simple method for real- world continual test-time adaptation

    Damian S ´ojka, Sebastian Cygert, Bartłomiej Twardowski, and Tomasz Trzci ´nski. Ar-tta: A simple method for real- world continual test-time adaptation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 3491–3495, 2023. 1, 2

    2023

  27. [35]

    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. InCVPR, pages 11920– 11929, 2023. 3

    2023

  28. [36]

    Parameter-selective continual test- time adaptation

    Jiaxu Tian and Fan Lyu. Parameter-selective continual test- time adaptation. InProceedings of the Asian Conference on Computer Vision, pages 1384–1400, 2024. 3

    2024

  29. [37]

    Ol- shausen, and Trevor Darrell

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, B. Ol- shausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. InICLR, 2021. 1, 2, 8, 19

    2021

  30. [38]

    Learning robust global representations by penalizing local predictive power.Advances in Neural Information Pro- cessing Systems, 32, 2019

    Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power.Advances in Neural Information Pro- cessing Systems, 32, 2019. 6

    2019

  31. [39]

    Con- tinual test-time domain adaptation

    Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Con- tinual test-time domain adaptation. InCVPR, pages 7191– 7201, 2022. 1, 2, 3, 5, 9, 19

    2022

  32. [40]

    Symmetric cross entropy for robust learn- ing with noisy labels

    Yisen Wang, Xingjun Ma, Zaiyi Chen, Yuan Luo, Jinfeng Yi, and James Bailey. Symmetric cross entropy for robust learn- ing with noisy labels. InProceedings of the IEEE/CVF in- ternational conference on computer vision, pages 322–330,

  33. [41]

    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, page...

    2024

  34. [42]

    Multi-label test-time adaptation with bound en- tropy minimization.International Conference on Learning Representations, 2025

    Xiangyu Wu, Feng Yu, Qing-Guo Chen, Yang Yang, and Jianfeng Lu. Multi-label test-time adaptation with bound en- tropy minimization.International Conference on Learning Representations, 2025. 2

    2025

  35. [43]

    Revisiting source-free domain adaptation: a new perspective via uncertainty control

    Gezheng Xu, Hui Guo, Li Yi, Charles Ling, Boyu Wang, and Grace Yi. Revisiting source-free domain adaptation: a new perspective via uncertainty control. InThe Thirteenth Inter- national Conference on Learning Representations, 2024. 1

    2024

  36. [44]

    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 Thirty-Second International Joint Con- ference on Artificial Intelligence, pages 1649–1657, 2023. 1

    2023

  37. [45]

    Navigating continual test-time adaptation with symbiosis knowledge

    Xu Yang, Moqi Li, Jie Yin, Kun Wei, and Cheng Deng. Navigating continual test-time adaptation with symbiosis knowledge. InProceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pages 5326– 5334, 2024. 1

    2024

  38. [46]

    Domain-specific block se- lection and paired-view pseudo-labeling for online test-time adaptation

    Yeonguk Yu, Sungho Shin, Seunghyeok Back, Minhwan Ko, Sangjun Noh, and Kyoobin Lee. Domain-specific block se- lection and paired-view pseudo-labeling for online test-time adaptation. InCVPR, pages 22723–22732, 2024. 2

    2024

  39. [47]

    Robust test-time adaptation in dynamic scenarios

    Longhui Yuan, Binhui Xie, and Shuang Li. Robust test-time adaptation in dynamic scenarios. InCVPR, pages 15922– 15932, 2023. 2, 3, 19

    2023

  40. [48]

    Tea: Test-time energy adaptation

    Yige Yuan, Bingbing Xu, Liang Hou, Fei Sun, Huawei Shen, and Xueqi Cheng. Tea: Test-time energy adaptation. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23901–23911, 2024. 2

    2024

  41. [49]

    MEMO: test time robustness via adaptation and augmentation

    Marvin Zhang, Sergey Levine, and Chelsea Finn. MEMO: test time robustness via adaptation and augmentation. In NeurIPS, pages 38629–38642, 2022. 1

    2022

  42. [50]

    Reshaping the on- line data buffering and organizing mechanism for continual test-time adaptation

    Zhilin Zhu, Xiaopeng Hong, Zhiheng Ma, Weijun Zhuang, Yaohui Ma, Yong Dai, and Yaowei Wang. Reshaping the on- line data buffering and organizing mechanism for continual test-time adaptation. InEuropean Conference on Computer Vision, pages 415–433. Springer, 2024. 1, 3, 19 SloM...

    2024

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.