Pith. sign in

REVIEW 3 major objections 3 minor 63 references

Learn Faster and Remember More: Balancing Exploration and Exploitation for Continual Test-time Adaptation

T0 review · 3 major / 3 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read This paper proposes a mean-teacher method that balances exploration and exploitation in continual test-time adaptation via intermediate-feature alignment and historical-anchor replay.

desk verdict The BEE method combines known ideas into a plausible CTTA recipe, but the abstract offers no evidence for its headline claims and the MCR mechanism may actually slow adaptation rather than speed it. read the letter →

arxiv 2508.12643 v1 pith:WPJ37H2Q submitted 2025-08-18 cs.CV

classification cs.CV
keywords continualtest-timeadaptationmeanteacherexploration–exploitationbalancemulti-levelconsistencyregularizationintermediatefeaturealignmentanchorreplaycatastrophicforgettingdomainshift
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper addresses continual test-time adaptation (CTTA), the setting where a pretrained model keeps adapting to a stream of changing target domains during inference. Its central claim is that an ideal CTTA method needs both exploration—fast adaptation to the current domain—and exploitation—retaining knowledge of past domains for reuse when similar domains return. To that end the paper proposes BEE, a mean-teacher framework with two mechanisms: a Multi-level Consistency Regularization (MCR) loss that aligns intermediate student–teacher features to accelerate adaptation, and a Complementary Anchor Replay (CAR) mechanism that reuses historical checkpoints to recover complementary knowledge from previous domains. The paper reports that BEE outperforms existing methods on several benchmarks, evidence for the claim that balancing exploration and exploitation this way improves CTTA.

What carries the argument

The load-bearing machinery is a mean-teacher setup (student adapts online, teacher is an exponential moving average of student weights) together with two mechanisms. MCR—Multi-level Consistency Regularization—is a loss that matches student and teacher activations at intermediate layers, not merely final predictions, on the grounds that domain shifts show up in shallow features; this is meant to make exploration fast. CAR—Complementary Anchor Replay—stores historical checkpoints as anchors and reuses them to recover complementary knowledge of earlier domains, which is meant to preserve exploitation. The two mechanisms together are what the paper says balances exploration and exploitation.

What would settle it

An experiment would settle it: apply a domain shift that changes only the deepest classifier features while leaving shallow features untouched, and compare BEE to the same method with MCR removed. If there is no speed-up in adaptation, the paper's shallow-feature premise fails; on recurring-domain benchmarks, a version with CAR disabled should show fast forgetting if CAR is really the exploitation mechanism.

Watch

Extended reading notes

Core claim

The paper's central claim is that continual test-time adaptation gets faster, not slower, when you act on intermediate features and hold on to old checkpoints. Domain shifts, it argues, hit shallow features first, so aligning student and teacher at multiple hidden layers (MCR) lets the model explore a new domain quickly; meanwhile, replaying historical anchors (CAR) lets it exploit knowledge from similar past domains instead of forgetting them. On the paper's own terms, the two mechanisms together give a mean-teacher framework that outperforms existing CTTA methods on several benchmarks.

Load-bearing premise

The load-bearing premise is that domain shifts mainly disturb shallow features, so aligning intermediate student–teacher features is the right fast route to adaptation; if shallow features are not the main bottleneck, the MCR mechanism loses its rationale.

Editorial extensions

If this is right

  • If the central claim is right, adaptation to each new domain should be noticeably faster because corrections are applied at intermediate layers instead of being back-propagated only through deep predictions.
  • Models that keep historical anchors should handle revisited domains better, giving a concrete route to avoiding catastrophic forgetting without storing target data.
  • The mean-teacher design itself becomes the balancing mechanism: the student explores by adapting, while the teacher averages history and stabilizes the objective.
  • Benchmark comparisons in CTTA should report not only final accuracy but also early adaptation speed, since that is the quantity MCR is designed to improve.

Reading between the lines

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

  • Editorial inference: if MCR works because shallow features carry the shift, the MCR loss could double as a cheap domain-shift detector—a spike in intermediate misalignment would signal that the current domain has changed.
  • Editorial inference: the anchor idea suggests a memory-budget law: with a fixed set of anchors, performance on a future domain should depend on how well the anchor set covers the feature-space region of that domain, which could be tested by pruning anchors.
  • Editorial inference: if the balance claim generalizes, a CTTA method could be tuned by one trade-off parameter between exploration and exploitation losses rather than by task-specific engineering.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The manuscript (arXiv:2508.12643) proposes a continual test-time adaptation (CTTA) method called BEE, built on a mean-teacher framework with two components: a Multi-level Consistency Regularization (MCR) loss that aligns intermediate student and teacher features to accelerate adaptation, and a Complementary Anchor Replay (CAR) mechanism that replays historical checkpoints to retain knowledge of previous domains. The abstract claims that BEE significantly outperforms state-of-the-art methods on several benchmarks. The full text provided to this referee consists only of the abstract; no experimental details, algorithmic specification, or additional technical content is available for evaluation.

Significance. If the claimed results hold, BEE would address a real gap in CTTA—balancing rapid adaptation to new domains against retention of previously learned knowledge—and the two mechanisms (MCR and CAR) would constitute a plausible design contribution. The paper also makes a substantive empirical claim about state-of-the-art performance. However, because the manuscript as provided contains no datasets, metrics, error bars, ablations, or baseline comparisons, the significance cannot currently be assessed beyond the plausibility of the proposed ideas. The conceptual framing, especially the diagnosis that domain shift affects shallow features and that historical checkpoints retain complementary knowledge, is interesting and worth investigating.

major comments (3)
  1. [Abstract, experimental claim] The central claim that the proposed method "significantly outperforms state-of-the-art methods on several benchmarks" is unsupported by any experimental evidence in the provided text. There are no dataset names, evaluation metrics, error bars, number of runs, baseline results, or ablation studies. This is a load-bearing omission because the paper's contribution is explicitly empirical; without these details, the central claim cannot be verified or reproduced. The authors should provide the full experimental setup and results, including a comparison table against state-of-the-art CTTA methods and an ablation isolating MCR and CAR.
  2. [Abstract, paragraph 2, MCR mechanism] The MCR loss aligns intermediate features of the student and teacher models in a mean-teacher framework. Since the teacher is typically an exponential moving average of the student, it lags behind the student immediately after a domain shift. Forcing the student's intermediate features to match the lagging teacher can act as a consistency regularizer that penalizes the very feature updates needed for fast exploration, potentially slowing adaptation rather than accelerating it. The abstract does not explain how this is avoided—for example, through gradient blocking, asymmetric backpropagation, target-batch normalization statistics, or a warm-up phase. The authors should describe the MCR update rule in detail and provide an experiment that directly tests adaptation speed in the first few batches after a shift, comparing MCR against a variant that aligns features only after several update steps.
  3. [Abstract, premise about shallow features] The paper's rationale for MCR rests on the premise that "domain shifts typically affect shallow features, which are inefficient to be adjusted from deep predictions." This premise is stated without supporting evidence or a reference, and it is not obvious that aligning shallow features is the most effective way to accelerate adaptation. If shallow features are not the primary bottleneck, MCR may have little effect, and the claimed gains would need to come from CAR or other implementation details. The authors should either cite empirical or theoretical support for this premise or report an experiment that directly tests whether MCR accelerates adaptation more than deep-layer-only consistency.
minor comments (3)
  1. [Abstract, writing] The phrase "dilatory exploration" is unusual; consider using "slow exploration" or "sluggish adaptation" for clarity.
  2. [Abstract, capitalization] In "strikes an appropriate Balance between Exploration and Exploitation," the capitalization of "Balance," "Exploration," and "Exploitation" is inconsistent with the rest of the abstract and should be corrected for style consistency.
  3. [Abstract, missing details] The abstract should name at least one or two benchmark datasets in common use for CTTA (e.g., CIFAR-10/100-C, ImageNet-C) to make the experimental claim more concrete.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity detected; BEE is presented as an empirical training scheme with no derivation that reduces to its own outputs.

full rationale

The supplied text contains only the abstract, which describes the proposed BEE method as a mean-teacher framework with two mechanisms: Multi-level Consistency Regularization (MCR) and Complementary Anchor Replay (CAR). The claims are that MCR aligns intermediate features to accelerate exploration and that CAR reuses historical checkpoints to retain complementary knowledge. These are algorithmic interventions rather than quantities derived from the benchmark results, and no equation, fitted parameter, or uniqueness theorem is presented in the visible text. The stated premise that domain shifts affect shallow features is a modeling assumption, not a circular definition of the method's success. No self-citation is visible, no prediction is constructed from a fitted input, and there is no derivation chain that reduces to the paper's own inputs. Whether hyperparameters were tuned on test benchmarks would be an overfitting or correctness concern, not a circularity concern, and the text provides no evidence for it. Accordingly, the honest finding is no significant circularity, with score 0.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

No free parameters or invented entities appear in the abstract. The method's burden is carried by four domain assumptions about the CTTA setting, feature-hierarchy sensitivity to domain shift, and the usefulness of historical checkpoints.

assumptions (4)
  • domain assumption A source pre-trained model is available and only unlabeled target data arrives during inference.
    Abstract states the CTTA setup directly; this is a task constraint, not a derived result.
  • domain assumption Target domains change continually and similar domains can recur later.
    Abstract says the method must handle 'similar domains in the future', which assumes recurrence or reuse of historical knowledge.
  • domain assumption Domain shifts affect shallow features more than deep outputs, so adjusting shallow representations is necessary for fast adaptation.
    Abstract's first challenge states this as a motivation for MCR. If false, the multi-level consistency mechanism loses its main justification.
  • ad hoc to paper Historical checkpoints retain complementary knowledge that can be replayed without destabilizing adaptation to the current domain.
    Abstract's CAR mechanism depends on this assumption; no evidence is given in the abstract.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learn Faster and Remember More: Balancing Exploration and Exploitation for Continual Test-time Adaptation." pith.science (2026). https://pith.science/paper/WPJ37H2Q

@misc{pith2026250812643,
  author       = {Pith},
  title        = {Pith review of: Learn Faster and Remember More: Balancing Exploration and Exploitation for Continual Test-time Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WPJ37H2Q}},
  note         = {Machine review of arXiv:2508.12643}
}
read the original abstract

Continual Test-Time Adaptation (CTTA) aims to adapt a source pre-trained model to continually changing target domains during inference. As a fundamental principle, an ideal CTTA method should rapidly adapt to new domains (exploration) while retaining and exploiting knowledge from previously encountered domains to handle similar domains in the future. Despite significant advances, balancing exploration and exploitation in CTTA is still challenging: 1) Existing methods focus on adjusting predictions based on deep-layer outputs of neural networks. However, domain shifts typically affect shallow features, which are inefficient to be adjusted from deep predictions, leading to dilatory exploration; 2) A single model inevitably forgets knowledge of previous domains during the exploration, making it incapable of exploiting historical knowledge to handle similar future domains. To address these challenges, this paper proposes a mean teacher framework that strikes an appropriate Balance between Exploration and Exploitation (BEE) during the CTTA process. For the former challenge, we introduce a Multi-level Consistency Regularization (MCR) loss that aligns the intermediate features of the student and teacher models, accelerating adaptation to the current domain. For the latter challenge, we employ a Complementary Anchor Replay (CAR) mechanism to reuse historical checkpoints (anchors), recovering complementary knowledge for diverse domains. Experiments show that our method significantly outperforms state-of-the-art methods on several benchmarks, demonstrating its effectiveness for CTTA tasks.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

63 extracted references · 52 canonical work pages

  1. [1]

    J. Wang, C. Lan, C. Liu, Y. Ouyang, T. Qin, W. Lu, Y. Chen, W. Zeng, and S. Y. Philip, ``Generalizing to unseen domains: A survey on domain generalization,'' IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 8, pp. 8052--8072, 2022

  2. [2]

    K. Zhou, Z. Liu, Y. Qiao, T. Xiang, and C. C. Loy, ``Domain generalization: A survey,'' IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 4, pp. 4396--4415, 2022

  3. [4]

    Janai, F

    J. Janai, F. G \"u ney, A. Behl, A. Geiger et al., ``Computer vision for autonomous vehicles: Problems, datasets and state of the art,'' Foundations and Trends in Computer Graphics and Vision, vol. 12, no. 1--3, pp. 1--308, 2020

  4. [5]

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

  5. [6]

    D. Wang, E. Shelhamer, S. Liu, B. Olshausen, and T. Darrell, ``Tent: Fully test-time adaptation by entropy minimization,'' in International Conference on Learning Representations, 2021

  6. [7]

    Jang, S.-Y

    M. Jang, S.-Y. Chung, and H. W. Chung, ``Test-time adaptation via self-training with nearest neighbor information,'' in International Conference on Learning Representations, 2023

  7. [8]

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

  8. [9]

    T. H. Hoang, M. Vo, and M. Do, ``Persistent test-time adaptation in recurring testing scenarios,'' Advances in Neural Information Processing Systems, vol. 37, pp. 123\,402--123\,442, 2024

Show all 63 references
  1. [10]

    S. Niu, J. Wu, Y. Zhang, Y. Chen, S. Zheng, P. Zhao, and M. Tan, ``Efficient test-time model adaptation without forgetting,'' in International Conference on Machine Learning. 1em plus 0.5em minus 0.4em PMLR, 2022, pp. 16\,888--16\,905

  2. [11]

    J. Song, J. Lee, I. S. Kweon, and S. Choi, ``Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization,'' in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 11\,920--11\,929

  3. [12]

    D. Lee, J. Yoon, and S. J. Hwang, ``Becotta: Input-dependent online blending of experts for continual test-time adaptation,'' in International Conference on Machine Learning, 2024

  4. [13]

    D \"o bler, R

    M. D \"o bler, R. A. Marsden, and B. Yang, ``Robust mean teacher for continual and gradual test-time adaptation,'' in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 7704--7714

  5. [14]

    Y. Gan, Y. Bai, Y. Lou, X. Ma, R. Zhang, N. Shi, and L. Luo, ``Decorate the newcomers: Visual domain prompt for continual test time adaptation,'' in Association for the Advancement of Artificial Intelligence, vol. 37, no. 6, 2023, pp. 7595--7603

  6. [15]

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

  7. [16]

    J. Liu, R. Xu, S. Yang, R. Zhang, Q. Zhang, Z. Chen, Y. Guo, and S. Zhang, ``Continual-mae: Adaptive distribution masked autoencoders for continual test-time adaptation,'' in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 28\,653--28\,663

  8. [17]

    Y. Yu, S. Shin, S. Back, M. Ko, S. Noh, and K. Lee, ``Domain-specific block selection and paired-view pseudo-labeling for online test-time adaptation,'' in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 22\,723--22\,732

  9. [18]

    Zhang, A

    R. Zhang, A. Cheng, Y. Luo, G. Dai, H. Yang, J. Liu, R. Xu, L. Du, Y. Du, Y. Jiang et al., ``Decomposing the neurons: Activation sparsity via mixture of experts for continual test time adaptation,'' arXiv preprint arXiv:2405.16486, 2024

  10. [19]

    A. M. Saxe, Y. Bansal, J. Dapello, M. Advani, A. Kolchinsky, B. D. Tracey, and D. D. Cox, ``On the information bottleneck theory of deep learning,'' Journal of Statistical Mechanics: Theory and Experiment, vol. 2019, no. 12, p. 124020, 2019

  11. [20]

    D. H. Wolpert and W. G. Macready, ``No free lunch theorems for optimization,'' IEEE transactions on evolutionary computation, vol. 1, no. 1, pp. 67--82, 1997

  12. [21]

    Z. Nado, S. Padhy, D. Sculley, A. D'Amour, B. Lakshminarayanan, and J. Snoek, ``Evaluating prediction-time batch normalization for robustness under covariate shift,'' arXiv preprint arXiv:2006.10963, 2020

  13. [22]

    F. You, J. Li, and Z. Zhao, ``Test-time batch statistics calibration for covariate shift,'' arXiv preprint arXiv:2110.04065, 2021

  14. [23]

    J. Hong, L. Lyu, J. Zhou, and M. Spranger, ``Mecta: Memory-economic continual test-time model adaptation,'' in International Conference on Learning Representations, 2023

  15. [24]

    Z. Su, J. Guo, K. Yao, X. Yang, Q. Wang, and K. Huang, ``Unraveling batch normalization for realistic test-time adaptation,'' in Association for the Advancement of Artificial Intelligence, vol. 38, no. 13, 2024, pp. 15\,136--15\,144

  16. [25]

    T. Gong, J. Jeong, T. Kim, Y. Kim, J. Shin, and S.-J. Lee, ``Note: Robust continual test-time adaptation against temporal correlation,'' Advances in Neural Information Processing Systems, vol. 35, pp. 27\,253--27\,266, 2022

  17. [26]

    Y. Su, X. Xu, and K. Jia, ``Towards real-world test-time adaptation: Tri-net self-training with balanced normalization,'' in Association for the Advancement of Artificial Intelligence, vol. 38, no. 13, 2024, pp. 15\,126--15\,135

  18. [27]

    Grandvalet and Y

    Y. Grandvalet and Y. Bengio, ``Semi-supervised learning by entropy minimization,'' Advances in Neural Information Processing Systems, vol. 17, 2004

  19. [28]

    Saito, D

    K. Saito, D. Kim, S. Sclaroff, T. Darrell, and K. Saenko, ``Semi-supervised domain adaptation via minimax entropy,'' in International Conference on Computer Vision, 2019, pp. 8050--8058

  20. [29]

    D. Chen, D. Wang, T. Darrell, and S. Ebrahimi, ``Contrastive test-time adaptation,'' in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 295--305

  21. [30]

    Z. Zhu, X. Hong, Z. Ma, W. Zhuang, Y. Ma, Y. Dai, and Y. Wang, ``Reshaping the online data buffering and organizing mechanism for continual test-time adaptation,'' in European Conference on Computer Vision. 1em plus 0.5em minus 0.4em Springer, 2024, pp. 415--433

  22. [31]

    Tarvainen and H

    A. Tarvainen and H. Valpola, ``Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results,'' Advances in Neural Information Processing Systems, vol. 30, 2017

  23. [32]

    S. Yang, J. Wu, J. Liu, X. Li, Q. Zhang, M. Pan, Y. Gan, Z. Chen, and S. Zhang, ``Exploring sparse visual prompt for domain adaptive dense prediction,'' in Association for the Advancement of Artificial Intelligence, vol. 38, no. 15, 2024, pp. 16\,334--16\,342

  24. [33]

    Grill, F

    J.-B. Grill, F. Strub, F. Altch \'e , C. Tallec, P. Richemond, E. Buchatskaya, C. Doersch, B. Avila Pires, Z. Guo, M. Gheshlaghi Azar et al., ``Bootstrap your own latent-a new approach to self-supervised learning,'' Advances in Neural Information Processing Systems, vol. 33, p...

  25. [34]

    Caron, H

    M. Caron, H. Touvron, I. Misra, H. J \'e gou, J. Mairal, P. Bojanowski, and A. Joulin, ``Emerging properties in self-supervised vision transformers,'' in International Conference on Computer Vision, 2021, pp. 9650--9660

  26. [35]

    Chen and K

    X. Chen and K. He, ``Exploring simple siamese representation learning,'' in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 15\,750--15\,758

  27. [36]

    Baevski, W.-N

    A. Baevski, W.-N. Hsu, Q. Xu, A. Babu, J. Gu, and M. Auli, ``Data2vec: A general framework for self-supervised learning in speech, vision and language,'' in International Conference on Machine Learning. 1em plus 0.5em minus 0.4em PMLR, 2022, pp. 1298--1312

  28. [37]

    K. Song, J. Xie, S. Zhang, and Z. Luo, ``Multi-mode online knowledge distillation for self-supervised visual representation learning,'' in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 11\,848--11\,857

  29. [38]

    E. Fini, P. Astolfi, K. Alahari, X. Alameda-Pineda, J. Mairal, M. Nabi, and E. Ricci, ``Semi-supervised learning made simple with self-supervised clustering,'' in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 3187--3197

  30. [39]

    Zhang and C

    Y. Zhang and C. Wu, ``Unsupervised camouflaged object segmentation as domain adaptation,'' in International Conference on Computer Vision, 2023, pp. 4334--4344

  31. [40]

    S. K. Ainsworth, J. Hayase, and S. Srinivasa, ``Git re-basin: Merging models modulo permutation symmetries,'' arXiv preprint arXiv:2209.04836, 2022

  32. [41]

    M. Li, S. Gururangan, T. Dettmers, M. Lewis, T. Althoff, N. A. Smith, and L. Zettlemoyer, ``Branch-train-merge: Embarrassingly parallel training of expert language models,'' arXiv preprint arXiv:2208.03306, 2022

  33. [42]

    W. Li, Y. Peng, M. Zhang, L. Ding, H. Hu, and L. Shen, ``Deep model fusion: A survey,'' arXiv preprint arXiv:2309.15698, 2023

  34. [43]

    H. Wang, M. Yurochkin, Y. Sun, D. Papailiopoulos, and Y. Khazaeni, ``Federated learning with matched averaging,'' arXiv preprint arXiv:2002.06440, 2020

  35. [44]

    Wortsman, G

    M. Wortsman, G. Ilharco, S. Y. Gadre, R. Roelofs, R. Gontijo-Lopes, A. S. Morcos, H. Namkoong, A. Farhadi, Y. Carmon, S. Kornblith et al., ``Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time,'' in International Con...

  36. [45]

    Izmailov, D

    P. Izmailov, D. Podoprikhin, T. Garipov, D. Vetrov, and A. G. Wilson, ``Averaging weights leads to wider optima and better generalization,'' arXiv preprint arXiv:1803.05407, 2018

  37. [46]

    Ilharco, M

    G. Ilharco, M. T. Ribeiro, M. Wortsman, L. Schmidt, H. Hajishirzi, and A. Farhadi, ``Editing models with task arithmetic,'' in International Conference on Learning Representations, 2023

  38. [47]

    M. S. Matena and C. A. Raffel, ``Merging models with fisher-weighted averaging,'' Advances in Neural Information Processing Systems, vol. 35, pp. 17\,703--17\,716, 2022

  39. [48]

    D.-H. Jang, S. Yun, and D. Han, ``Model stock: All we need is just a few fine-tuned models,'' in European Conference on Computer Vision. 1em plus 0.5em minus 0.4em Springer, 2024, pp. 207--223

  40. [49]

    X. Jin, X. Ren, D. Preotiuc-Pietro, and P. Cheng, ``Dataless knowledge fusion by merging weights of language models,'' in International Conference on Learning Representations, 2023

  41. [50]

    Yadav, D

    P. Yadav, D. Tam, L. Choshen, C. A. Raffel, and M. Bansal, ``Ties-merging: Resolving interference when merging models,'' Advances in Neural Information Processing Systems, vol. 36, 2024

  42. [51]

    Marczak, B

    D. Marczak, B. Twardowski, T. Trzci \'n ski, and S. Cygert, ``Magmax: Leveraging model merging for seamless continual learning,'' in European Conference on Computer Vision. 1em plus 0.5em minus 0.4em Springer, 2024, pp. 379--395

  43. [52]

    S. Choi, S. Yang, S. Choi, and S. Yun, ``Improving test-time adaptation via shift-agnostic weight regularization and nearest source prototypes,'' in European Conference on Computer Vision. 1em plus 0.5em minus 0.4em Springer, 2022, pp. 440--458

  44. [53]

    Van Den Oord, O

    A. Van Den Oord, O. Vinyals et al., ``Neural discrete representation learning,'' Advances in Neural Information Processing Systems, vol. 30, 2017

  45. [54]

    Caron, I

    M. Caron, I. Misra, J. Mairal, P. Goyal, P. Bojanowski, and A. Joulin, ``Unsupervised learning of visual features by contrasting cluster assignments,'' Advances in Neural Information Processing Systems, vol. 33, pp. 9912--9924, 2020

  46. [55]

    Oquab, T

    M. Oquab, T. Darcet, T. Moutakanni, H. Vo, M. Szafraniec, V. Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby et al., ``Dinov2: Learning robust visual features without supervision,'' arXiv preprint arXiv:2304.07193, 2023

  47. [56]

    Cuturi, ``Sinkhorn distances: Lightspeed computation of optimal transport,'' Advances in Neural Information Processing Systems, vol

    M. Cuturi, ``Sinkhorn distances: Lightspeed computation of optimal transport,'' Advances in Neural Information Processing Systems, vol. 26, 2013

  48. [57]

    Brahma and P

    D. Brahma and P. Rai, ``A probabilistic framework for lifelong test-time adaptation,'' in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 3582--3591

  49. [58]

    Press, S

    O. Press, S. Schneider, M. K \"u mmerer, and M. Bethge, ``Rdumb: A simple approach that questions our progress in continual test-time adaptation,'' Advances in Neural Information Processing Systems, vol. 36, 2024

  50. [59]

    Ahmad and S

    S. Ahmad and S. Purdy, ``Real-time anomaly detection for streaming analytics,'' arXiv preprint arXiv:1607.02480, 2016

  51. [60]

    Zagoruyko, ``Wide residual networks,'' arXiv preprint arXiv:1605.07146, 2016

    S. Zagoruyko, ``Wide residual networks,'' arXiv preprint arXiv:1605.07146, 2016

  52. [61]

    S. Xie, R. Girshick, P. Doll \'a r, Z. Tu, and K. He, ``Aggregated residual transformations for deep neural networks,'' in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017, pp. 1492--1500

  53. [62]

    K. He, X. Zhang, S. Ren, and J. Sun, ``Deep residual learning for image recognition,'' in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016, pp. 770--778

  54. [63]

    Croce, M

    F. Croce, M. Andriushchenko, V. Sehwag, E. Debenedetti, N. Flammarion, M. Chiang, P. Mittal, and M. Hein, ``Robustbench: a standardized adversarial robustness benchmark,'' arXiv preprint arXiv:2010.09670, 2020

  55. [64]

    11em plus .33em minus .07em 4000 4000 100 4000 4000 500 `\.=1000 = #1 \@IEEEnotcompsoconly \@IEEEcompsoconly #1 * [1] 0pt [0pt][0pt] #1 * [1] 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEcompsocnotconfonly \@IEEEauthorblockAstyle \@IEEEcompsocnotconfonly \@IEEEco...

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.