pith. sign in

arxiv: 2606.28142 · v2 · pith:YWXTKEUVnew · submitted 2026-06-26 · 💻 cs.LG

MixTTA: Low-Rank Cross-Channel Mixing for Reliable Test-Time Adaptation

Mansoo Jung , Youngwook Kim , Jungwoo Lee This is my paper

Pith reviewed 2026-07-01 06:35 UTC · model grok-4.3

classification 💻 cs.LG
keywords test-time adaptationdistribution shiftnormalization layerslow-rank transformationcross-channel mixingaffine parametersdomain adaptation
0
0 comments X

The pith

Per-channel affine parameters cannot correct cross-channel structural changes from distribution shifts.

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

Standard test-time adaptation updates per-channel affine parameters of normalization layers, which perform only axis-aligned scaling and shifting. This geometric limitation leaves them unable to address cross-channel structural changes induced by distribution shifts. MixTTA introduces a lightweight plug-in module that adds a low-rank cross-channel transformation to enable inter-channel mixing at each layer. Decoupling Projection enforces separation from the diagonal affine path, while Spectral Projection prevents rank-1 collapse in non-stationary streams. The module integrates into existing normalization-based TTA methods and yields consistent gains in both standard and wild settings while reducing adaptation failures.

Core claim

Test-time adaptation methods update the affine parameters of normalization layers to adapt models under distribution shifts. However, per-channel affine parameters perform axis-aligned scaling and shifting, making them geometrically incapable of correcting cross-channel structural changes induced by distribution shift. MixTTA equips normalization layers with a low-rank cross-channel transformation, enabling inter-channel mixing at each layer. Decoupling Projection enforces strict separation from the diagonal affine path, and Spectral Projection prevents rank-1 collapse under non-stationary test streams. MixTTA integrates seamlessly into any existing normalization-based TTA method and produce

What carries the argument

low-rank cross-channel transformation module with Decoupling Projection and Spectral Projection

If this is right

  • MixTTA integrates seamlessly into any existing normalization-based TTA method.
  • It produces consistent improvements over strong baselines in standard and wild TTA settings.
  • It mitigates adaptation failure under challenging conditions.
  • The projections ensure the low-rank branch captures only cross-channel interactions.

Where Pith is reading between the lines

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

  • The same low-rank mixing idea could be tested on layers other than normalization to address distribution shifts more broadly.
  • Cross-channel dependencies may prove central to robustness in other adaptation settings such as continual learning.
  • Higher-rank variants of the transformation could be explored if the low-rank constraint turns out to be the main performance limiter.

Load-bearing premise

The Decoupling Projection enforces strict separation from the diagonal affine path and the Spectral Projection prevents rank-1 collapse under non-stationary test streams.

What would settle it

An experiment that removes the Spectral Projection and observes rank-1 collapse together with degraded adaptation performance under non-stationary test streams would falsify the claim that the projection prevents such collapse.

Figures

Figures reproduced from arXiv: 2606.28142 by Jungwoo Lee, Mansoo Jung, Youngwook Kim.

Figure 1
Figure 1. Figure 1: (a) Correlation distance comparison across three layers of a pre-trained ViT on ImageNet-C Gaussian noise under varying severity levels. (b) and (c) provide overviews of Tent [40] and our proposed MixTTA, respectively, where each colored circle repre￾sents a feature dimension. Tent applies only channel-wise affine modulation without mixing across dimensions (colors remain unchanged while intensities vary),… view at source ↗
Figure 2
Figure 2. Figure 2: Overall structure of MixTTA module, which operates within normalization layers of the backbone network. Standardized features are modulated through both the affine and residual low-rank branches. Decoupling Projection ensures strict separation between diagonal and off-diagonal updates, while Spectral Projection filters collapse￾prone directions during adaptation. a diagonal initialization. To address this,… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of (a) top-1 accuracy, (b) the condition number κ of the fea￾ture covariance at an intermediate layer, and (c) || diag(∆)||2 for adapted models on ImageNet-C snow (severity 5) under imbalanced label shift, across different projection configurations. Fig. 3a shows that enabling both DP and SP yields the most stable accuracy trajectory, preventing the performance collapse observed in other configu… view at source ↗
Figure 4
Figure 4. Figure 4: Layer-wise and spectral analysis of feature correlation changes on ImageNet-C Gaussian noise at severity level 5. (a) Layer-wise correlation distances. (b) Cumulative energy of the singular-value spectrum at blocks.0.norm2. Fig. 4b further examines the spectral structure of the MixTTA-induced cor￾relation shift. Specifically, we plot how rapidly the cumulative spectral energy of the MixTTA-induced correcti… view at source ↗
read the original abstract

Test-Time Adaptation (TTA) methods commonly update the affine parameters of normalization layers to adapt deployed models under distribution shifts. However, per-channel affine parameters perform axis-aligned scaling and shifting, making them geometrically incapable of correcting cross-channel structural changes induced by distribution shift. To address this limitation, we propose MixTTA, a lightweight plug-in module that equips normalization layers with a low-rank cross-channel transformation, enabling inter-channel mixing at each layer. To ensure that the low-rank branch captures only cross-channel interactions, we also propose Decoupling Projection that enforces strict separation from the diagonal affine path, along with Spectral Projection that prevents rank-1 collapse under non-stationary test streams. MixTTA can be seamlessly integrated into any existing normalization-based TTA method. Experiments in both standard and wild TTA settings show consistent improvements over strong baselines while mitigating adaptation failure under challenging conditions. The source code is publicly available at https://github.com/delta6189/MixTTA.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 2 minor

Summary. The paper argues that per-channel affine parameters in normalization layers for test-time adaptation (TTA) are geometrically restricted to axis-aligned scaling/shifting and thus cannot correct cross-channel structural changes under distribution shift. It proposes MixTTA, a lightweight low-rank cross-channel transformation module that plugs into existing normalization-based TTA methods, together with a Decoupling Projection to enforce separation from the diagonal affine path and a Spectral Projection to avoid rank-1 collapse on non-stationary streams. Experiments in standard and wild TTA settings report consistent gains over baselines and reduced adaptation failure; source code is released.

Significance. If the geometric limitation holds and the two projections mathematically guarantee that the low-rank branch captures only off-diagonal interactions, the method could improve robustness of TTA without substantial overhead. Public code is a clear strength that enables direct verification of the claimed gains.

major comments (2)
  1. [§3.2] §3.2 (Decoupling Projection): the claim of 'strict separation' from the diagonal affine path is load-bearing for the central geometric argument, yet the manuscript provides no explicit proof or invariant showing that the low-rank term is prevented from acquiring non-zero diagonal entries under SGD on non-stationary test streams; if the projection is realized only as a soft regularizer or post-hoc mask, the separation can fail and the claimed distinction collapses.
  2. [§3.3] §3.3 (Spectral Projection): the mechanism that 'prevents rank-1 collapse' is described at a high level; without a concrete bound or eigenvalue analysis showing stability across the reported non-stationary streams, it is unclear whether the low-rank branch remains expressive or reverts to the axis-aligned case the paper seeks to avoid.
minor comments (2)
  1. [Table 2] Table 2 and Figure 4: error bars or standard deviations across the N runs are not reported, making it difficult to assess whether the reported gains are statistically reliable.
  2. Notation: the low-rank matrix is introduced with different symbols in the text and in the algorithm box; consistent notation would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the geometric arguments and the two projection mechanisms. We address each major comment below and will revise the manuscript to strengthen the presentation with additional formal details.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (Decoupling Projection): the claim of 'strict separation' from the diagonal affine path is load-bearing for the central geometric argument, yet the manuscript provides no explicit proof or invariant showing that the low-rank term is prevented from acquiring non-zero diagonal entries under SGD on non-stationary test streams; if the projection is realized only as a soft regularizer or post-hoc mask, the separation can fail and the claimed distinction collapses.

    Authors: The Decoupling Projection is realized as a hard orthogonal projection applied immediately after every gradient step: the low-rank matrix is explicitly masked to zero its diagonal entries before being added to the affine path. This operation is deterministic and independent of the loss landscape or stream non-stationarity, thereby maintaining the invariant that the low-rank branch contains only off-diagonal terms by construction. We will insert a short lemma and proof of the invariant (showing that the projection commutes with the update and preserves the zero-diagonal property) into the revised §3.2. revision: yes

  2. Referee: [§3.3] §3.3 (Spectral Projection): the mechanism that 'prevents rank-1 collapse' is described at a high level; without a concrete bound or eigenvalue analysis showing stability across the reported non-stationary streams, it is unclear whether the low-rank branch remains expressive or reverts to the axis-aligned case the paper seeks to avoid.

    Authors: The Spectral Projection clamps the singular values of the low-rank matrix to a minimum positive threshold ε after each update, which directly enforces a lower bound on the smallest singular value and thereby precludes rank-1 collapse. We will augment §3.3 with (i) the explicit eigenvalue bound relating ε to the Frobenius norm of the update and (ii) a short stability argument for non-stationary streams, together with additional plots of singular-value trajectories on the wild TTA benchmarks. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain; method proposal is additive and self-contained

full rationale

The paper states a geometric limitation of per-channel affine parameters (axis-aligned scaling/shifting) as an independent observation, then proposes MixTTA as an additive low-rank module plus explicit Decoupling and Spectral Projections to enable and restrict cross-channel mixing. No equations, fitted parameters, or results are shown reducing a claimed prediction or first-principles outcome to the inputs by construction. The separation property is a deliberate design feature of the proposed projections rather than a tautological redefinition. No load-bearing self-citations or uniqueness theorems from prior author work appear in the provided text. The derivation chain is a standard method proposal and does not exhibit any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

The central claim rests on the assumption that low-rank matrices suffice to model cross-channel interactions and that the two new projections achieve the stated separation and rank preservation without introducing new failure modes.

axioms (1)
  • domain assumption Low-rank cross-channel transformations are sufficient to correct structural changes induced by distribution shift in normalization layers.
    Invoked to justify the choice of low-rank branch over full matrices.
invented entities (2)
  • Decoupling Projection no independent evidence
    purpose: Enforces strict separation between low-rank cross-channel path and diagonal affine parameters.
    New mechanism introduced to prevent interference with existing affine path.
  • Spectral Projection no independent evidence
    purpose: Prevents rank-1 collapse of the low-rank branch under non-stationary test streams.
    New mechanism introduced to maintain mixing capacity during streaming adaptation.

pith-pipeline@v0.9.1-grok · 5702 in / 1367 out tokens · 34020 ms · 2026-07-01T06:35:35.419768+00:00 · methodology

Review history (2 revisions) →

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

75 extracted references · 5 canonical work pages · 2 internal anchors

  1. [1]

    2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Deep Residual Learning for Image Recognition , author=. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=. 2016 , organization=

    2016

  2. [2]

    Proceedings of the European conference on computer vision (ECCV) , pages=

    Group normalization , author=. Proceedings of the European conference on computer vision (ECCV) , pages=

  3. [3]

    GitHub repository , doi =

    Ross Wightman , title =. GitHub repository , doi =. 2019 , publisher =

    2019

  4. [4]

    International Conference on Learning Representations , year =

    Benchmarking Neural Network Robustness to Common Corruptions and Perturbations , author =. International Conference on Learning Representations , year =

  5. [5]

    International conference on machine learning , pages=

    Unsupervised domain adaptation by backpropagation , author=. International conference on machine learning , pages=. 2015 , organization=

    2015

  6. [6]

    European conference on computer vision , pages=

    Deep coral: Correlation alignment for deep domain adaptation , author=. European conference on computer vision , pages=. 2016 , organization=

    2016

  7. [7]

    Advances in Neural Information Processing Systems , volume=

    Test-time classifier adjustment module for model-agnostic domain generalization , author=. Advances in Neural Information Processing Systems , volume=

  8. [8]

    International Conference on Learning Representations , year =

    Tent: Fully Test-Time Adaptation by Entropy Minimization , author =. International Conference on Learning Representations , year =

  9. [9]

    International conference on machine learning , pages=

    Efficient test-time model adaptation without forgetting , author=. International conference on machine learning , pages=. 2022 , organization=

    2022

  10. [10]

    International Conference on Learning Representations , year =

    Towards Stable Test-Time Adaptation in Dynamic Wild World , author =. International Conference on Learning Representations , year =

  11. [11]

    The Twelfth International Conference on Learning Representations , year=

    Entropy is not Enough for Test-time Adaptation: From the Perspective of Disentangled Factors , author=. The Twelfth International Conference on Learning Representations , year=

  12. [12]

    Advances in neural information processing systems , volume=

    Memo: Test time robustness via adaptation and augmentation , author=. Advances in neural information processing systems , volume=

  13. [13]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Image style transfer using convolutional neural networks , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  14. [14]

    International Conference on Learning Representations , year=

    Minimal-Entropy Correlation Alignment for Unsupervised Deep Domain Adaptation , author=. International Conference on Learning Representations , year=

  15. [15]

    HoMM: Higher-Order Moment Matching for Unsupervised Domain Adaptation , booktitle =

    Chao Chen and Zhihang Fu and Zhihong Chen and Sheng Jin and Zhaowei Cheng and Xinyu Jin and Xian. HoMM: Higher-Order Moment Matching for Unsupervised Domain Adaptation , booktitle =. 2020 , url =. doi:10.1609/AAAI.V34I04.5745 , timestamp =

    work page doi:10.1609/aaai.v34i04.5745 2020

  16. [16]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Domain adaptation by mixture of alignments of second-or higher-order scatter tensors , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  17. [17]

    Advances in neural information processing systems , volume=

    Universal style transfer via feature transforms , author=. Advances in neural information processing systems , volume=

  18. [18]

    Ieee Access , volume=

    Deep correlation multimodal neural style transfer , author=. Ieee Access , volume=. 2021 , publisher=

    2021

  19. [19]

    Deep Learning using Rectified Linear Units (ReLU)

    Deep learning using rectified linear units (relu) , author=. arXiv preprint arXiv:1803.08375 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  20. [20]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Joint geometrical and statistical alignment for visual domain adaptation , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  21. [21]

    International conference on machine learning , pages=

    Domain adaptation with asymmetrically-relaxed distribution alignment , author=. International conference on machine learning , pages=. 2019 , organization=

    2019

  22. [22]

    International conference on machine learning , pages=

    Batch normalization: Accelerating deep network training by reducing internal covariate shift , author=. International conference on machine learning , pages=. 2015 , organization=

    2015

  23. [23]

    Layer Normalization

    Layer normalization , author=. arXiv preprint arXiv:1607.06450 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  24. [24]

    International Conference on Learning Representations , year =

    LoRA: Low-Rank Adaptation of Large Language Models , author =. International Conference on Learning Representations , year =

  25. [25]

    International conference on machine learning , pages=

    Learning transferable features with deep adaptation networks , author=. International conference on machine learning , pages=. 2015 , organization=

    2015

  26. [26]

    International conference on machine learning , pages=

    Parameter-efficient transfer learning for NLP , author=. International conference on machine learning , pages=. 2019 , organization=

    2019

  27. [27]

    Proceedings of the 2021 conference on empirical methods in natural language processing , pages=

    The power of scale for parameter-efficient prompt tuning , author=. Proceedings of the 2021 conference on empirical methods in natural language processing , pages=

    2021

  28. [28]

    International Conference on Learning Representations (ICLR) , year =

    Central Moment Discrepancy (CMD) for Domain-Invariant Representation Learning , author =. International Conference on Learning Representations (ICLR) , year =

  29. [29]

    International Conference on Machine Learning (ICML) , pages =

    Deep Transfer Learning with Joint Adaptation Networks , author =. International Conference on Machine Learning (ICML) , pages =

  30. [30]

    Journal of machine learning research , volume=

    Domain-adversarial training of neural networks , author=. Journal of machine learning research , volume=

  31. [31]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Adversarial Discriminative Domain Adaptation , author =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  32. [32]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Maximum Classifier Discrepancy for Unsupervised Domain Adaptation , author =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  33. [33]

    Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics (EACL) , year =

    AdapterFusion: Non-Destructive Task Composition for Transfer Learning , author =. Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics (EACL) , year =

  34. [34]

    European Conference on Computer Vision (ECCV) , year =

    Visual Prompt Tuning , author =. European Conference on Computer Vision (ECCV) , year =

  35. [35]

    Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL) , pages =

    Intrinsic Dimensionality Explains the Effectiveness of Language Model Fine-Tuning , author =. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL) , pages =

  36. [36]

    Advances in neural information processing systems , volume=

    Qlora: Efficient finetuning of quantized llms , author=. Advances in neural information processing systems , volume=

  37. [37]

    Proceedings of the thirteenth international conference on artificial intelligence and statistics , pages=

    Understanding the difficulty of training deep feedforward neural networks , author=. Proceedings of the thirteenth international conference on artificial intelligence and statistics , pages=. 2010 , organization=

    2010

  38. [38]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Projection metric learning on Grassmann manifold with application to video based face recognition , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  39. [39]

    SIAM Journal on Matrix Analysis and Applications , volume=

    Schubert varieties and distances between subspaces of different dimensions , author=. SIAM Journal on Matrix Analysis and Applications , volume=. 2016 , publisher=

    2016

  40. [40]

    Proceedings of the 25th international conference on Machine learning , pages=

    Grassmann discriminant analysis: a unifying view on subspace-based learning , author=. Proceedings of the 25th international conference on Machine learning , pages=

  41. [41]

    2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages=

    Test-time low rank adaptation via confidence maximization for zero-shot generalization of vision-language models , author=. 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages=. 2025 , organization=

    2025

  42. [42]

    Kojima, J

    Lora-ttt: Low-rank test-time training for vision-language models , author=. arXiv preprint arXiv:2502.02069 , year=

    work page arXiv

  43. [43]

    International Conference on Machine Learning , pages=

    Test-time Correlation Alignment , author=. International Conference on Machine Learning , pages=. 2025 , organization=

    2025

  44. [44]

    International Conference on Learning Representations , year=

    Spectral Normalization for Generative Adversarial Networks , author=. International Conference on Learning Representations , year=

  45. [45]

    International journal of computer vision , volume=

    Imagenet large scale visual recognition challenge , author=. International journal of computer vision , volume=. 2015 , publisher=

    2015

  46. [46]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Deep residual learning for image recognition , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  47. [47]

    9th International Conference on Learning Representations,

    Alexey Dosovitskiy and Lucas Beyer and Alexander Kolesnikov and Dirk Weissenborn and Xiaohua Zhai and Thomas Unterthiner and Mostafa Dehghani and Matthias Minderer and Georg Heigold and Sylvain Gelly and Jakob Uszkoreit and Neil Houlsby , title =. 9th International Conference on Learning Representations,. 2021 , url =

    2021

  48. [48]

    International conference on machine learning , pages=

    On the importance of initialization and momentum in deep learning , author=. International conference on machine learning , pages=. 2013 , organization=

    2013

  49. [49]

    Advances in neural information processing systems , volume=

    Learning robust global representations by penalizing local predictive power , author=. Advances in neural information processing systems , volume=

  50. [50]

    International Conference on Machine Learning , pages=

    Beyond Entropy: Region Confidence Proxy for Wild Test-Time Adaptation , author=. International Conference on Machine Learning , pages=. 2025 , organization=

    2025

  51. [51]

    Proceedings of the Asian Conference on Computer Vision , pages=

    Learning dual hierarchical representation for 3D surface reconstruction , author=. Proceedings of the Asian Conference on Computer Vision , pages=

  52. [52]

    IEEE Access , year=

    Leveraging Single Positive Class Label Supervision for Weakly Supervised Semantic Segmentation , author=. IEEE Access , year=

  53. [53]

    IEEE Access , year=

    A plug-in curriculum scheduler for improved deformable medical image registration , author=. IEEE Access , year=

  54. [54]

    IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , volume=

    Instance-Dependent Multilabel Noise Generation for Multilabel Remote Sensing Image Classification , author=. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , volume=. 2024 , publisher=

    2024

  55. [55]

    ICT Express , year=

    EDAS: Effective Data Augmentation Strategies for test-time adaptation , author=. ICT Express , year=

  56. [56]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Continual test-time domain adaptation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  57. [57]

    Advances in Neural Information Processing Systems , volume=

    Swad: Domain generalization by seeking flat minima , author=. Advances in Neural Information Processing Systems , volume=

  58. [58]

    Yuge Shi and Jeffrey Seely and Philip H. S. Torr and Siddharth Narayanaswamy and Awni Y. Hannun and Nicolas Usunier and Gabriel Synnaeve , title =. The Tenth International Conference on Learning Representations,. 2022 , url =

    2022

  59. [59]

    International conference on machine learning , pages=

    Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation , author=. International conference on machine learning , pages=. 2020 , organization=

    2020

  60. [60]

    Advances in neural information processing systems , volume=

    Exploiting the intrinsic neighborhood structure for source-free domain adaptation , author=. Advances in neural information processing systems , volume=

  61. [61]

    Proceedings of the 61st ACM/IEEE Design Automation Conference , pages=

    Rl-ptq: Rl-based mixed precision quantization for hybrid vision transformers , author=. Proceedings of the 61st ACM/IEEE Design Automation Conference , pages=

  62. [62]

    IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems , year=

    Autonomous Model Quantization Framework for Hybrid Vision Transformers based on Reinforcement Learning , author=. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems , year=

  63. [63]

    Proceedings of the IEEE international conference on computer vision , pages=

    Mask r-cnn , author=. Proceedings of the IEEE international conference on computer vision , pages=

  64. [64]

    The Twelfth International Conference on Learning Representations,

    Jiaming Liu and Senqiao Yang and Peidong Jia and Renrui Zhang and Ming Lu and Yandong Guo and Wei Xue and Shanghang Zhang , title =. The Twelfth International Conference on Learning Representations,. 2024 , url =

    2024

  65. [65]

    The Thirteenth International Conference on Learning Representations,

    Qingyang Zhang and Yatao Bian and Xinke Kong and Peilin Zhao and Changqing Zhang , title =. The Thirteenth International Conference on Learning Representations,. 2025 , url =

    2025

  66. [66]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Neural Collapse in Test-Time Adaptation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  67. [67]

    Advances in Neural Information Processing Systems , volume=

    Note: Robust continual test-time adaptation against temporal correlation , author=. Advances in Neural Information Processing Systems , volume=

  68. [68]

    MoETTA: Test-Time Adaptation Under Mixed Distribution Shifts with MoE-LayerNorm , booktitle =

    Xiao Fan and Jingyan Jiang and Zhaoru Chen and Fanding Huang and Xiao Chen and Qinting Jiang and Bowen Zhang and Xing Tang and Zhi Wang , editor =. MoETTA: Test-Time Adaptation Under Mixed Distribution Shifts with MoE-LayerNorm , booktitle =. 2026 , url =. doi:10.1609/AAAI.V40I25.39243 , timestamp =

    work page doi:10.1609/aaai.v40i25.39243 2026

  69. [69]

    Advances in Neural Information Processing Systems , volume=

    Feature-based instance neighbor discovery: Advanced stable test-time adaptation in dynamic world , author=. Advances in Neural Information Processing Systems , volume=

  70. [70]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Learning to generate gradients for test-time adaptation via test-time training layers , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  71. [71]

    BECoTTA: Input-dependent Online Blending of Experts for Continual Test-time Adaptation , booktitle =

    Daeun Lee and Jaehong Yoon and Sung Ju Hwang , editor =. BECoTTA: Input-dependent Online Blending of Experts for Continual Test-time Adaptation , booktitle =. 2024 , url =

    2024

  72. [72]

    Advances in Neural Information Processing Systems , volume=

    Rethinking Entropy in Test-Time Adaptation: The Missing Piece from Energy Duality , author=. Advances in Neural Information Processing Systems , volume=

  73. [73]

    Advances in Neural Information Processing Systems , volume=

    Partition-then-adapt: Combating prediction bias for reliable multi-modal test-time adaptation , author=. Advances in Neural Information Processing Systems , volume=

  74. [74]

    International Conference on Learning Representations , volume=

    Dimension agnostic neural processes , author=. International Conference on Learning Representations , volume=

  75. [75]

    International Conference on Machine Learning , pages=

    Boost-and-Skip: A Simple Guidance-Free Diffusion for Minority Generation , author=. International Conference on Machine Learning , pages=. 2025 , organization=

    2025