Pith. sign in

REVIEW 4 major objections 5 minor 61 references

Free on the Fly: Enhancing Flexibility in Test-Time Adaptation with Online EM

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper introduces FreeTTA, an online EM algorithm that adapts a frozen CLIP model to streaming test data by modeling the test distribution as a Gaussian mixture, achieving average gains of 3.76% on cross-domain and 1.66% on…

desk verdict Useful empirical TTA method with consistent gains, but the EM derivation in Eq. 10/12 is not the claimed online M-step, so the 'explicit distribution modeling' claim is overstated. read the letter →

arxiv 2507.06973 v1 pith:S6R4B67C submitted 2025-07-09 cs.CV

classification cs.CV
keywords test-timeadaptationonlineEMGaussianmixturemodelvision-languageCLIPtraining-freeout-of-distributiongeneralizationstreaminginference
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

This paper introduces FreeTTA, an online expectation-maximization algorithm that adapts a frozen vision-language model to a streaming test set without any training, without storing past samples, and without modifying model parameters. The key idea is to treat the unlabeled test stream as a Gaussian mixture whose per-class means are initialized by the model's text embeddings, then update means, a shared covariance, and class priors sample by sample using posterior probabilities as soft assignments. The authors claim this is the first test-time adaptation method to explicitly model the target distribution while remaining training-free and universally available, and report stable average gains of 3.76% on cross-domain and 1.66% on out-of-distribution benchmarks over existing methods across 15 datasets. If correct, the method shows that a simple closed-form update can replace expensive prompt tuning and still improve CLIP's accuracy under domain shift.

What carries the argument

The central object is an online EM algorithm operating on a Gaussian mixture model of the test distribution. Class means start as CLIP's text embeddings for each class name or prompt, the shared covariance starts as the identity, and every arriving sample contributes a weighted posterior to update the class means, the shared covariance, and the class priors; the weight is $w(h)=e^{-\beta h}$, where $h$ is the self-entropy of the CLIP zero-shot prediction, so uncertain samples influence the update less. These updated parameters are converted into logits via $w_y=\Sigma^{-1}\mu_y$ and $b_y=\log P(y)-\frac{1}{2}\mu_y^{\top}\Sigma^{-1}\mu_y$, blended with the original CLIP logits by a hyperparameter $\alpha$.

What would settle it

Generate a synthetic test stream from a known Gaussian mixture, then compare FreeTTA's recursively estimated covariance with the exact online EM covariance (computed by re-weighting all past samples around the updated means). If the two diverge systematically while the mean estimates agree, the covariance recursion is not a faithful EM update, and the gains on real benchmarks would need another explanation, such as the mean-shifting or entropy weighting alone.

Watch

Extended reading notes

Core claim

FreeTTA claims to be the first test-time adaptation method for vision-language models that simultaneously satisfies three properties: it explicitly models the target data distribution, it requires no gradient-based training, and it neither accesses historical data nor stores test samples. The target distribution is assumed to be a Gaussian mixture with a shared covariance matrix; each incoming sample is softly assigned to classes by posterior probability, and these assignments update the class means, covariance, and priors through an online EM recursion. CLIP's zero-shot predictions serve as priors that initialize class means from text embeddings and re-weight each sample's influence by the self-entropy of its zero-shot prediction. In experiments on 15 datasets, the method reports average improvements of 3.76% over state-of-the-art baselines in cross-domain settings and 1.66% in out-of-distribution settings, while using a batch size of one.

Load-bearing premise

The paper's claim to 'explicitly model the test data distribution' rests on treating the recursive covariance update in Eq. (12) as a valid online EM M-step, but that update is an approximation of the exact EM covariance update rather than the exact one.

Editorial extensions

If this is right

  • A frozen CLIP model can be adapted online with no gradients and no memory of past samples, which is compatible with API-only access to the model.
  • The test stream itself becomes the training signal: relationships among samples improve individual predictions even when each sample is processed one at a time.
  • Because every update is a closed-form weighted average, adaptation is deterministic and cheap relative to prompt tuning, with only two hyperparameters to set.
  • Across 15 cross-domain and OOD datasets, the authors report consistent accuracy gains over both training-free methods and prompt-optimization methods.

Reading between the lines

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

  • A natural extension of the paper's scheme would be to replace CLIP's text embeddings with any zero-shot prototype source, potentially carrying the same online EM adaptation to audio, text, or multimodal frozen encoders.
  • The paper's accuracy tables do not directly show whether the learned covariance tracks the true test-stream covariance; a diagnostic comparing estimated vs. empirical covariance would isolate how much of the gain comes from distribution modeling versus from the mean updates and entropy weighting.
  • If the covariance recursion is the weak step, then the method's reported gains might be reproducible with a simpler diagonal-covariance or mean-only variant, which would reframe the contribution as confidence-weighted prototype shifting rather than full target-distribution modeling.
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 / 5 minor

Summary. The paper proposes FreeTTA, a test-time adaptation method for vision-language models (CLIP) that uses an online Expectation-Maximization (EM) procedure to fit a Gaussian mixture model to the test feature distribution. Class means are initialized from CLIP text embeddings, the shared covariance is initialized to the identity, and CLIP’s zero-shot confidence is used to weight each sample’s contribution. The method is training-free (no gradient updates) and does not store historical test samples. Experiments on 15 cross-domain and out-of-distribution benchmarks report consistent improvements over several TTA baselines, with the largest gains over methods that also claim training-free and availability properties.

Significance. If the claims hold, FreeTTA would be a simple, non-gradient TTA method that explicitly models the target distribution while being broadly applicable to API-only or privacy-sensitive settings. The empirical evaluation is extensive: 15 datasets, two backbones, and comparisons against a range of prompt-tuning, caching, and mean-shift baselines. The reported gains over MTA and ZERO are consistent across datasets, and the method’s design avoids backpropagation and test-feature storage. However, the paper’s central conceptual contribution rests on the claim that Eq. (10)/(12) is an online EM M-step for a shared-covariance GMM; this derivation has substantive gaps that affect the validity of the 'explicit target distribution modeling' narrative.

major comments (4)
  1. [Abstract and Sec. 3.2] The recursion in Eq. (10) is not the EM M-step for the assumed shared-covariance GMM. For a batch, the M-step covariance is the responsibility-weighted scatter of all samples around the updated class means. An exact online update must therefore include a correction term for previous samples whose contributions change when the means move; Eq. (10) omits this and instead adds only the current sample’s outer product around the updated mean. It also divides by n_t - 1 (or n'_t - 1) rather than the effective total responsibility. Moreover, under the text’s definition of n_t as 'the total number of samples up to the t-th step', the denominator is zero at t = 1, so the unweighted update is undefined; the issue is avoided only if n_t is implicitly redefined as the sum of pseudo-counts N_y. Because the paper’s central claim is that FreeTTA 'explicitly models the test data distribution' via online EM (Abstract; Sec. 1), this derivation gap is load-bearing. The authors should either derive a correct online EM recursion using sufficient statistics and show that it yields similar results, or revise the method description and claims to characterize it as a heuristic online Gaussian adaptation.
  2. [Abstract and Sec. 3.2] The abstract states that FreeTTA is a method that 'makes no assumptions', but Sec. 3.2 explicitly assumes that each class follows an independent Gaussian distribution with a shared covariance matrix. If 'no assumptions' is intended to mean no assumptions about data availability, historical storage, or supervision, the wording should be qualified accordingly; as written, it is internally inconsistent and potentially misleading about the method’s modeling assumptions.
  3. [Sec. 4.1 and Eq. (13)] The hyperparameters α and β are set to 0.2 and 4.5, respectively, but the paper provides no sensitivity analysis or explanation of how these values were selected. Since Eq. (13) interpolates between the CLIP zero-shot logits (α = 0) and the GDA logits, the reported improvements could be sensitive to these choices. Please report sensitivity over a small grid of (α, β) on a validation split or a subset of datasets, or clearly state that the values were chosen once and fixed across all benchmarks.
  4. [Sec. 3.3.3 and Eq. (13)] Computing w_y = Σ^{-1} μ_y in Eq. (13) requires inverting the d×d shared covariance matrix, which is updated online as a full matrix via outer products. A naive inversion is O(d^3) per step, which conflicts with the claimed 'training-free for efficiency' property in Sec. 1. The paper should specify the computational complexity and state whether incremental inverse updates (e.g., Sherman-Morrison) are used; if the covariance is restricted to a diagonal or low-rank form, that should be stated explicitly.
minor comments (5)
  1. [Eq. (5)] Equation (5) omits the class prior P(y) in the numerator and the sum over priors in the denominator, making it inconsistent with Eq. (4). Eq. (6) later includes log P(y), so Eq. (5) should be corrected or explicitly stated to assume equal priors.
  2. [Eq. (12)] The notation in the denominator of the π update in Eq. (12) is unclear: 'n_t−1′' should be replaced with the properly defined previous total n'_{t-1} or current total n'_t, and the definitions of N_y, n_t, and n'_t should be given precisely to avoid the division-by-zero ambiguity at the first step.
  3. [Table 3] The row labels '2-Mean Vectors Update', '2-Covariance Matrix Update', and '2-VLM priors' appear to be typesetting artifacts for 'w/o' (without). Please replace them with unambiguous labels such as 'w/o Mean Vectors Update'.
  4. [Sec. 2, Related Work] The EM algorithm is cited as '[? ]' in the text of Sec. 2; a proper reference should be provided.
  5. [Sec. 4.2, Tables 1 and 2] The sentence reporting 'average accuracy increases of 2.42% and 1.56%' for MTA and ZERO does not match the average gains computed from the tables (cross-domain gains are 3.79% and 3.76%, OOD gains are 2.79% and 1.66%). Please verify the arithmetic or clarify which specific averages are being reported.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: FreeTTA's online EM updates and final logit combination do not reduce by construction to their CLIP inputs.

full rationale

The paper's central derivation is an online EM/GDA procedure in which CLIP text embeddings initialize the class means, CLIP entropy weights the influence of each incoming sample, and the final logits combine CLIP's zero-shot logits with GDA logits built from online-updated means, covariance, and priors. The GDA term is not equal to the current CLIP prediction by construction: it depends on accumulated running statistics from earlier test samples, so the final classifier is a genuine function of the test stream rather than a renaming of the CLIP input. Initialization and confidence weighting are priors, not fitted predictions. The only self-citation in the paper, reference [9], appears in the related-work survey and is not load-bearing for the method's derivation. The evaluation is transductive, but that is standard test-time adaptation protocol and does not make the reported accuracies circular. The recursive covariance update in Eq. 10/12 is not the exact EM M-step because it omits the correction for previously observed samples around moved means and divides by n_t - 1; however, this is a mathematical approximation issue, not a case where a prediction reduces to its input by definition. Similarly, hyperparameters alpha and beta are fixed constants in the experiments; even if they were tuned, that would be an overfitting concern rather than a circularity of the derivation. No quoted equation or argument exhibits a fitted parameter renamed as a prediction or a claim equivalent to its own assumption.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the GMM assumption, on using CLIP text embeddings as class means and identity as covariance, on entropy as confidence, and on treating Eqs 10-12 as online EM. The covariance recursion is not the exact EM M-step, so the 'explicit distribution modeling' claim is only as strong as that approximation. Two hyperparameters (alpha and beta) are tuned constants. No new entities are introduced.

free parameters (2)
  • alpha = 0.2
    Blend coefficient between CLIP zero-shot logits and the fitted Gaussian model in Eq 13. Fixed for all datasets with no sensitivity analysis, so benchmark tuning is plausible.
  • beta = 4.5
    Exponent in the confidence weighting w(h)=exp(-beta*h) that downweights high-entropy samples in Eq 12. Fixed without ablation or validation split.
assumptions (5)
  • domain assumption Each class's test features are multivariate Gaussian with a shared covariance matrix.
    Invoked in Sec 3.2 Eq 3 and Eq 8; CLIP embeddings are not demonstrated to be Gaussian, and a single streaming sample sequence is used to estimate the parameters.
  • domain assumption CLIP text embeddings are valid initial class means and the identity matrix is an unbiased initial covariance.
    Eq 7 initialization; no empirical justification that text-feature centers align with image-feature class centers under this metric.
  • domain assumption CLIP self-entropy reliably measures sample confidence for downweighting.
    Eq 11 and the weight w(h) in Eq 12; no calibration analysis or ablation across beta values.
  • ad hoc to paper The recursive updates in Eq 10/12 implement the EM M-step sufficiently well.
    The covariance update is not the exact M-step: it omits the mean-movement correction and divides by n_t-1 rather than total responsibility, so the method is an approximation presented as EM.
  • domain assumption Online single-sample evaluation with batch size 1 is the correct TTA protocol.
    Adopted from TPT [39] in Sec 4.1; excludes methods that require test batches.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Free on the Fly: Enhancing Flexibility in Test-Time Adaptation with Online EM." pith.science (2026). https://pith.science/paper/S6R4B67C

@misc{pith2026250706973,
  author       = {Pith},
  title        = {Pith review of: Free on the Fly: Enhancing Flexibility in Test-Time Adaptation with Online EM},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S6R4B67C}},
  note         = {Machine review of arXiv:2507.06973}
}
read the original abstract

Vision-Language Models (VLMs) have become prominent in open-world image recognition for their strong generalization abilities. Yet, their effectiveness in practical applications is compromised by domain shifts and distributional changes, especially when test data distributions diverge from training data. Therefore, the paradigm of test-time adaptation (TTA) has emerged, enabling the use of online off-the-shelf data at test time, supporting independent sample predictions, and eliminating reliance on test annotations. Traditional TTA methods, however, often rely on costly training or optimization processes, or make unrealistic assumptions about accessing or storing historical training and test data. Instead, this study proposes FreeTTA, a training-free and universally available method that makes no assumptions, to enhance the flexibility of TTA. More importantly, FreeTTA is the first to explicitly model the test data distribution, enabling the use of intrinsic relationships among test samples to enhance predictions of individual samples without simultaneous access--a direction not previously explored. FreeTTA achieves these advantages by introducing an online EM algorithm that utilizes zero-shot predictions from VLMs as priors to iteratively compute the posterior probabilities of each online test sample and update parameters. Experiments demonstrate that FreeTTA achieves stable and significant improvements compared to state-of-the-art methods across 15 datasets in both cross-domain and out-of-distribution settings.

Figures

Figures reproduced from arXiv: 2507.06973 by the authors.

Figure 1
Figure 1. Comparison of key properties: target distribution mod￾eling, availability, and training-free efficiency. (a) Prompt-based methods fail the training-free criterion due to lengthy backpropa￾gation. (b) Other methods require access to source or additional test data, affecting availability. In contrast, (c) our FreeTTA satis￾fies all criteria, offering a universally available, training-free solu￾tion that efficiently mo… view at source ↗
Figure 2
Figure 2. The overall framework of our FreeTTA. Given a test sample xt, we use the frozen CLIP image encoder to extract the image feature, while the text encoder, using prompt templates, generates class feature vectors. The online EM algorithm is initialized with the text features as mean vectors and the identity matrix as the shared covariance matrix. It updates in two steps: the E-step calculates the posterior probability f… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 39 canonical work pages

  1. [1]

    Align your prompts: Test-time prompting with distribution align- ment for zero-shot generalization

    Jameel Abdul Samadh, Mohammad Hanan Gani, Noor Hus- sein, Muhammad Uzair Khattak, Muhammad Muzammal Naseer, Fahad Shahbaz Khan, and Salman H Khan. Align your prompts: Test-time prompting with distribution align- ment for zero-shot generalization. Advances in Neural Infor- mation Processing Systems, 36, 2024. 1, 2, 3, 4, 7, 8

  2. [2]

    Flamingo: a visual language model for few-shot learning

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Men- sch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716–23736,

  3. [3]

    An essay towards solving a problem in the doctrine of chances

    Thomas Bayes. An essay towards solving a problem in the doctrine of chances. Biometrika, 45(3-4):296–315, 1958. 2, 4, 5

  4. [4]

    Pattern recognition and machine learning

    Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning. Springer, 2006. 2, 4

  5. [5]

    Food-101–mining discriminative components with random forests

    Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101–mining discriminative components with random forests. In Computer vision–ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, pro- ceedings, part VI 13, pages 446–461. Springer, 2014. 6

  6. [6]

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

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

  7. [7]

    Gaussian yolov3: An accurate and fast object detec- tor using localization uncertainty for autonomous driving

    Jiwoong Choi, Dayoung Chun, Hyun Kim, and Hyuk-Jae Lee. Gaussian yolov3: An accurate and fast object detec- tor using localization uncertainty for autonomous driving. In Proceedings of the IEEE/CVF International conference on computer vision, pages 502–511, 2019. 3

  8. [8]

    Describing textures in the wild

    Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606–3613, 2014. 6

Show all 61 references
  1. [9]

    Curriculum point prompting for weakly-supervised referring image segmentation

    Qiyuan Dai and Sibei Yang. Curriculum point prompting for weakly-supervised referring image segmentation. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13711–13722, 2024. 3

  2. [10]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009. 6

  3. [11]

    Frustratingly easy test- time adaptation of vision-language models

    Matteo Farina, Gianni Franchi, Giovanni Iacca, Massim- iliano Mancini, and Elisa Ricci. Frustratingly easy test- time adaptation of vision-language models. arXiv preprint arXiv:2405.18330, 2024. 2, 7, 8

  4. [12]

    Learning gener- ative visual models from few training examples: An incre- mental bayesian approach tested on 101 object categories

    Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning gener- ative visual models from few training examples: An incre- mental bayesian approach tested on 101 object categories. In 2004 conference on computer vision and pattern recognition workshop, pages 178–178. IEEE, 2004. 6

  5. [13]

    Diverse data augmentation with diffusions for effective test-time prompt tuning

    Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 2704–2714, 2023. 1, 2, 3, 4, 7, 8

  6. [14]

    On the mathematical foundations of the- oretical statistics

    Ronald A Fisher. On the mathematical foundations of the- oretical statistics. Philosophical transactions of the Royal Society of London. Series A, containing papers of a mathe- matical or physical character, 222(594-604):309–368, 1922. 2, 4

  7. [15]

    Clip-adapter: Better vision-language models with feature adapters

    Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581–595, 2024. 3

  8. [16]

    Open-vocabulary object detection via vision and language knowledge distillation

    Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921 ,

  9. [17]

    Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification

    Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217–2226, 2019. 6

  10. [18]

    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. In Proceedings of the IEEE/CVF internation...

  11. [19]

    Natural adversarial examples

    Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Stein- hardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 15262–15271, 2021. 6

  12. [20]

    Scaling up visual and vision-language representa- tion learning with noisy text supervision

    Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representa- tion learning with noisy text supervision. In International conference on machine learning, pages 4904–4916. PMLR,

  13. [21]

    Efficient test-time adaptation of vision-language models

    Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, and Eric Xing. Efficient test-time adaptation of vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 14162–14171, 2024. 1, 2, 3, 4, 6, 7, 8

  14. [22]

    The kinetics hu- man action video dataset

    Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics hu- man action video dataset. arXiv preprint arXiv:1705.06950,

  15. [23]

    Maple: Multi-modal prompt learning

    Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113–19122, 2023. 1

  16. [24]

    3d object representations for fine-grained categorization

    Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on com- puter vision workshops, pages 554–561, 2013. 6 9

  17. [25]

    F-vlm: Open-vocabulary object detec- tion upon frozen vision and language models

    Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detec- tion upon frozen vision and language models. arXiv preprint arXiv:2209.15639, 2022. 3

  18. [26]

    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. arXiv preprint arXiv:2403.07366 ,

  19. [27]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Interna- tional conference on machine learning, pages 12888–12900. PMLR, 2022. 1, 2

  20. [28]

    Video test-time adaptation for action recognition

    Wei Lin, Muhammad Jehanzeb Mirza, Mateusz Kozinski, Horst Possegger, Hilde Kuehne, and Horst Bischof. Video test-time adaptation for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 22952–22961, 2023. 3

  21. [29]

    Gmm-unit: Unsuper- vised multi-domain and multi-modal image-to-image trans- lation via attribute gaussian mixture modeling

    Yahui Liu, Marco De Nadai, Jian Yao, Nicu Sebe, Bruno Lepri, and Xavier Alameda-Pineda. Gmm-unit: Unsuper- vised multi-domain and multi-modal image-to-image trans- lation via attribute gaussian mixture modeling. arXiv preprint arXiv:2003.06788, 2020. 3

  22. [30]

    Swapprompt: Test-time prompt adaptation for vision- language models

    Xiaosong Ma, Jie Zhang, Song Guo, and Wenchao Xu. Swapprompt: Test-time prompt adaptation for vision- language models. Advances in Neural Information Process- ing Systems, 36, 2024. 2

  23. [31]

    Fine-grained visual classi- fication of aircraft

    Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classi- fication of aircraft. arXiv preprint arXiv:1306.5151 , 2013. 6

  24. [32]

    The expectation-maximization algorithm

    Todd K Moon. The expectation-maximization algorithm. IEEE Signal processing magazine, 13(6):47–60, 1996. 2

  25. [33]

    Automated flower classification over a large number of classes

    Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & im- age processing, pages 722–729. IEEE, 2008. 6

  26. [34]

    Cats and dogs

    Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498–3505. IEEE, 2012. 6

  27. [35]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

  28. [36]

    Do imagenet classifiers generalize to im- agenet? In International conference on machine learning , pages 5389–5400

    Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to im- agenet? In International conference on machine learning , pages 5389–5400. PMLR, 2019. 6

  29. [37]

    Gaussian mixture models

    Douglas A Reynolds et al. Gaussian mixture models. Ency- clopedia of biometrics, 741(659-663), 2009. 3, 5

  30. [38]

    Probabilistic face embeddings

    Yichun Shi and Anil K Jain. Probabilistic face embeddings. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6902–6911, 2019. 3

  31. [39]

    Test- time prompt tuning for zero-shot generalization in vision- language models

    Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test- time prompt tuning for zero-shot generalization in vision- language models. Advances in Neural Information Process- ing Systems, 35:14274–14289, 2022. 1, 2, 3, 4, 6, 7, 8

  32. [40]

    Just shift it: Test-time prototype shifting for zero-shot gen- eralization with vision-language models

    Elaine Sui, Xiaohan Wang, and Serena Yeung-Levy. Just shift it: Test-time prototype shifting for zero-shot gen- eralization with vision-language models. arXiv preprint arXiv:2403.12952, 2024. 1, 2, 3

  33. [41]

    Test time adaptation with regularized loss for weakly supervised salient object detection

    Olga Veksler. Test time adaptation with regularized loss for weakly supervised salient object detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7360–7369, 2023. 3

  34. [42]

    Tent: Fully test-time adaptation by entropy minimization

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Ol- shausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726,

  35. [43]

    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

  36. [44]

    Dynamically instance- guided adaptation: A backward-free approach for test-time domain adaptive semantic segmentation

    Wei Wang, Zhun Zhong, Weijie Wang, Xi Chen, Charles Ling, Boyu Wang, and Nicu Sebe. Dynamically instance- guided adaptation: A backward-free approach for test-time domain adaptive semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern R...

  37. [45]

    Cris: Clip- driven referring image segmentation

    Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip- driven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686–11695, 2022. 3

  38. [46]

    Unsu- pervised learning of models for recognition

    Markus Weber, Max Welling, and Pietro Perona. Unsu- pervised learning of models for recognition. In Computer Vision-ECCV 2000: 6th European Conference on Computer Vision Dublin, Ireland, June 26–July 1, 2000 Proceedings, Part I 6, pages 18–32. Springer, 2000. 3

  39. [47]

    Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching

    Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7031–7040, 2023. 3

  40. [48]

    Sun database: Large-scale scene recognition from abbey to zoo

    Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer so- ciety conference on computer vision and pattern recognition, pages 3485–3492. IEEE, 2010. 6

  41. [49]

    Bridging vision and language en- coders: Parameter-efficient tuning for referring image seg- mentation

    Zunnan Xu, Zhihong Chen, Yong Zhang, Yibing Song, Xi- ang Wan, and Guanbin Li. Bridging vision and language en- coders: Parameter-efficient tuning for referring image seg- mentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17503–17512, 2023. 3

  42. [50]

    Bottom-up shift and reasoning for referring im- age segmentation

    Sibei Yang, Meng Xia, Guanbin Li, Hong-Yu Zhou, and Yizhou Yu. Bottom-up shift and reasoning for referring im- age segmentation. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 11266–11275, 2021. 3 10

  43. [51]

    Rapid network adaptation: Learning to adapt neural networks using test-time feedback

    Teresa Yeo, O ˘guzhan Fatih Kar, Zahra Sodagar, and Amir Zamir. Rapid network adaptation: Learning to adapt neural networks using test-time feedback. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 4674–4687, 2023. 3

  44. [52]

    C-tpt: Calibrated test-time prompt tuning for vision- language models via text feature dispersion

    Hee Suk Yoon, Eunseop Yoon, Joshua Tian Jin Tee, Mark Hasegawa-Johnson, Yingzhen Li, and Chang D Yoo. C-tpt: Calibrated test-time prompt tuning for vision- language models via text feature dispersion. arXiv preprint arXiv:2403.14119, 2024. 2, 3

  45. [53]

    Maxime Zanella and Ismail Ben Ayed. On the test-time zero- shot generalization of vision-language models: Do we really need prompt learning? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 23783–23793, 2024. 1, 2, 3, 7, 8

  46. [54]

    Memo: Test time robustness via adaptation and augmentation

    Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Ad- vances in neural information processing systems, 35:38629– 38642, 2022. 3

  47. [55]

    Tip- adapter: Training-free adaption of clip for few-shot classi- fication

    Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kun- chang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip- adapter: Training-free adaption of clip for few-shot classi- fication. In European conference on computer vision, pages 493–510. Springer, 2022. 1, 3

  48. [56]

    Boostadapter: Improving vision- language test-time adaptation via regional bootstrapping

    Taolin Zhang, Jinpeng Wang, Hang Guo, Tao Dai, Bin Chen, and Shu-Tao Xia. Boostadapter: Improving vision- language test-time adaptation via regional bootstrapping. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 2

  49. [57]

    Seg- mentation of brain mr images through a hidden markov ran- dom field model and the expectation-maximization algo- rithm

    Yongyue Zhang, Michael Brady, and Stephen Smith. Seg- mentation of brain mr images through a hidden markov ran- dom field model and the expectation-maximization algo- rithm. IEEE transactions on medical imaging, 20(1):45–57,

  50. [58]

    Dual memory networks: A versa- tile adaptation approach for vision-language models

    Yabin Zhang, Wenjie Zhu, Hui Tang, Zhiyuan Ma, Kaiyang Zhou, and Lei Zhang. Dual memory networks: A versa- tile adaptation approach for vision-language models. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 28718–28728, 2024. 1

  51. [59]

    Regionclip: Region- based language-image pretraining

    Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chun- yuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region- based language-image pretraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognitio...

  52. [60]

    Conditional prompt learning for vision-language mod- els

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language mod- els. In Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition , pages 16816–16825,

  53. [61]

    Learning to prompt for vision-language models

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. In- ternational Journal of Computer Vision, 130(9):2337–2348,

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.