Pith. sign in

REVIEW 3 major objections 4 minor 46 references

MuRA: Multi-Rank Adaptation for Efficient and Effective Test-Time Vision-Language Generalization

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper claims that fixed LoRA rank, not model capacity, is the bottleneck in test-time adaptation, and that a token-level router across SVD-initialized ranks solves it.

desk verdict Useful, well-ablated TTA method with overclaimed theory and thin motivating correlation. read the letter →

arxiv 2608.03885 v1 pith:3GPV7MON submitted 2026-08-04 cs.CV

Gengyuan Liu , Nanzhou Wang , Chang Liu , Qinwen Wu , Zhenhao Wang , Jiacong Wang , Bokui Chen , Xiangyang Ji This is my paper
classification cs.CV
keywords Test-TimeAdaptationVision-LanguageModelsLow-RankDynamicRankRoutingDomainGeneralizationMixture-of-ExpertsCLIPMulti-RankOrthogonalDecomposition
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's claim is that the fixed rank used by LoRA-based test-time adaptation is a structural bottleneck: one rank cannot fit both visually simple and complex inputs, so any static choice either underfits or overfits. It reports an empirical correlation (R²=0.913 across six datasets) between image entropy and the optimal LoRA rank, then builds MuRA, which routes each token to a soft mixture of LoRA modules with ranks {2, 4, 8, 16, 32}. The modules are initialized by singular-value decomposition of the pretrained weights, the router is updated continuously across test samples while the adapters are reset, and the paper argues via lemmas that this routing is necessary and stable. On ImageNet OOD and cross-domain benchmarks, MuRA reports the highest average accuracy among compared TTA methods while using far less memory and higher throughput than prompt-tuning baselines. If correct, this makes dynamic rank routing, rather than larger static capacity, the key to efficient test-time VLM adaptation.

What carries the argument

The central mechanism is the trio Multi-Rank Orthogonal Decomposition (MROD), Unified Component Fusion (UCF), and Continuous Router Updating (CRU). MROD takes the singular value decomposition of a pretrained weight W, initializing rank-r_i adapters as A_i B_i^T plus a frozen residual R_i, with A_i B_i^T orthogonal to R_i. UCF averages the residuals into R-bar and uses a zero-initialized softmax router pi(h_l)=softmax(W_r h_l) to output o = sum_i pi_i A_i B_i^T h_l + R-bar h_l. CRU resets the adapters to their SVD-initialized values after each sample but keeps the router weights, stabilizing the learned complexity-to-capacity mapping. The theoretical engine is Lemma 4.2's expected logit updat

What would settle it

A decisive check: run MuRA and, at each router step, compare the true gradient inner product <g, f_i(h_l) - f-bar(h_l)> with the hypothetical loss gap loss_i(x) - loss_pi(x). If they systematically disagree in sign on a continuous domain stream, Lemma 4.2's necessity claim is refuted; an oracle static-rank baseline with matched parameters would then determine whether the empirical gains come from routing or simply from extra capacity.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that representational capacity in test-time adaptation should be allocated per token by complexity. The authors decompose the deepest-layer weight into orthogonal principal components and initialize five LoRA experts of increasing rank; a zero-initialized softmax router produces token-level mixture weights, and only the router persists across samples. They argue theoretically that one router step promotes an expert exactly when its complexity-capacity mismatch is below the routing-weighted average, and that keeping the router while resetting experts keeps the routing landscape stationary. The result is that MuRA's accuracy peaks at the deepe

Load-bearing premise

The proof that dynamic routing is necessary assumes that a certain gradient re-weighting of experts can be approximated by comparing each expert's loss to the mixture loss, along with a quadratic 'distance to ideal capacity' risk model; both are heuristics that are never derived or bounded, so if the approximation fails the necessity theorem does not follow.

Editorial extensions

If this is right

  • TTA methods should stop treating rank as a fixed hyperparameter; the paper's results imply that a token-level complexity-to-rank mapping is learnable and yields higher accuracy across domains.
  • SVD-based initialization of LoRA modules becomes the appropriate starting point for short-horizon adaptation, since zero-initialized LoRA is unstable under one-step test-time updates.
  • Keeping the router while resetting the adapters is a safe way to accumulate knowledge across a test stream; the paper shows this avoids catastrophic forgetting and routing collapse.
  • Adapting at the deepest visual layer with dynamic ranks gives a better accuracy/memory trade-off than inserting static-rank adapters in shallower layers, because it uses the shortest backpropagation path.
  • The gains are not explained by parameter count alone: a single-rank baseline with matched trainable parameters is slower, heavier, and less accurate than MuRA.

Reading between the lines

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

  • The entropy-rank correlation is reported at dataset level with six points; a per-image or per-token regression would be a sharper test of whether token-level routing is truly driven by complexity alignment or partly by ensembling many experts.
  • The router's rank-utilization entropy (the paper's own diagnostic) could be used as an online confidence signal to prune rarely used ranks during a stream, further cutting compute in a way the paper does not explore.
  • The CRU stability argument assumes the expert adapters are reset every sample; relaxing that assumption, for example by letting the residual drift slowly, would map exactly where routing collapse reappears, a boundary the paper leaves untested.
  • The same multi-rank routing idea could be applied to the text encoder or to autoregressive vision-language decoders, but the paper only validates the visual encoder.
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

3 major / 4 minor

Summary. The paper identifies a perceived bottleneck in test-time adaptation (TTA) of vision-language models: static LoRA rank configurations that cannot adapt to varying input complexity. It proposes MuRA, which uses multi-rank LoRA modules initialized by SVD (MROD), a token-level softmax router (UCF), and a continuous router-updating scheme (CRU). The paper reports state-of-the-art accuracy on ImageNet OOD and cross-domain benchmarks, with reduced memory and higher throughput relative to prior TTA methods, and claims rigorous theoretical proof of the necessity and stability of dynamic routing. Empirical sections include ablations, efficiency comparisons, scalability to ViT-L/14 and ResNet-50, and a continual TTA stream.

Significance. If the empirical results hold, MuRA is a practically valuable contribution: it shows a concrete accuracy-per-memory improvement over existing knowledge-adaptive TTA, with a matched-parameter single-rank baseline indicating that the gains are not simply from increased parameter count. The ablations are systematic and the efficiency comparison against a matched-parameter baseline is a good practice. The scalability experiments to ViT-L/14 and CNN backbones are also informative. However, the theoretical justification, which the paper explicitly lists as a main contribution, is not rigorous as stated, and the motivating entropy-rank correlation is based on six in-sample points. The core empirical method may still be sound, but the paper's framing as providing mathematical proof needs substantial revision.

major comments (3)
  1. [Section 4, Lemma 4.2] The claimed proof of the necessity of dynamic routing is not a proof. The derivation depends on (i) the first-order approximation <g, f_i(h_l) - f_bar(h_l)> ≈ ell_i(x) - ell_pi(x), which is asserted without derivation or error bound, and (ii) an ad hoc local-risk model R_i(H)=alpha_i + lambda_i d(H, mu_i)^2. Moreover, the displayed expected-update expression is proportional to the squared-distance term only if alpha_i and lambda_i are equal across experts; otherwise the alpha and lambda terms do not cancel, and the conclusion about complexity-capacity mismatch does not follow. No such assumption is stated or justified. Thus Lemma 4.2 does not establish that a static rank forces an inevitable compromise; at best it is a plausibility heuristic. The abstract and contribution list claim 'rigorous theoretical justifications mathematically proving the necessity'; this should be substantially r
  2. [Section 4, Lemma 4.3] Lemma 4.3 only bounds the probability change over a single gradient step by the Lipschitz continuity of softmax; this is immediate from differentiability and does not depend on the CRU mechanism. Stability of CRU over a long test stream requires controlling the cumulative effect of many updates, e.g., via total variation or contraction of the iterates. Bounded per-step gradients do not prevent unbounded drift over an unbounded stream. The remark's conclusion that 'Lemma 4.3 theoretically guarantees that accumulating gradients ... is mathematically stable' and that CRU avoids 'catastrophic forgetting or semantic drift' is unsupported. The bound also omits the norm of h_l, which enters the Lipschitz constant of the map W_r -> softmax(W_r h_l). Please provide a multi-step stability analysis or soften the claim to a one-step bound.
  3. [Section 1, Figure 3] The central motivation—that static rank is a fundamental bottleneck—rests on a correlation (R^2=0.913, N=6) between dataset-level image entropy and the optimal rank. The rank-selection protocol is not described; if optimal ranks are obtained by test-set accuracy of the same benchmark, the correlation is in-sample and likely overfitted. With six points, no error bars, and no multiple runs, the linear relation is fragile, and the figure suggests EuroSAT is a high-leverage point. Moreover, MuRA routes at token level within a single stream, whereas the correlation is established only at dataset level. Please specify the rank-selection protocol, provide cross-validated or hold-out evidence, and, if possible, report a sample- or token-level entropy-rank relationship.
minor comments (4)
  1. [General] No code is released and tables report single numbers without standard deviations or seed information. For a paper claiming state-of-the-art results, this limits reproducibility and prevents assessing the significance of the reported gains.
  2. [Section 4] The notation is confusing: W_r is defined as a k x d matrix, but z_i=(W^i_r)^T h_l suggests W^i_r is a vector. Clarify whether rows or columns of W_r are used, and align the notation in Lemmas 4.1 and 4.2.
  3. [Section 3.1 / Eq. (3)] The value of rho (the entropy percentile) is not given in the main text or implementation details; please specify it. Also clarify how tau is computed exactly across the augmented views.
  4. [Table 4] The caption says 'Single-Rank' baselines, but it is not stated what rank is used or whether the parameter count is matched across depths. Since the depth comparison is used to support the 'deepest layer' claim, please describe the baselines' configuration and parameter counts.

Circularity Check

2 steps flagged · score 6.0 of 10

Lemma 4.2's 'proof' of the necessity of dynamic routing is self-definitional: the conclusion is baked into the mismatch-based risk model; the static-rank bottleneck is also supported only by an in-sample entropy-rank correlation.

  1. self definitional [Section 4, Lemma 4.2 and its remark]
    "we define ... local risk R_i(H)=α_i+λ_i d(H,μ_i)^2, where d(H,μ_i) measures complexity-capacity mismatch. ... using the first-order approximation ⟨g,f_i(h_l)−f_bar(h_l)⟩≈ℓ_i(x)−ℓ_π(x) ... Lemma 4.2 ... E_x∈Ω_i[Δz_i]∝π_i[∑_j π_j d(H(Ω_i),μ_j)^2 − d(H(Ω_i),μ_i)^2]. ... The update promotes an expert if and only if its complexity-capacity mismatch is below the routing-weighted average mismatch across all experts."

    The 'necessity' theorem does not follow from the loss gradient alone. The expected-update formula is obtained by (i) an unproven first-order approximation ⟨g,f_i−f_bar⟩≈ℓ_i−ℓ_π and (ii) positing R_i(H)=α_i+λ_i d(H,μ_i)^2. With that risk model, the sign of E[Δz_i] is by construction negative when d_i^2 is above the π-weighted average d_j^2 and positive when below; the 'promotes lower-mismatch expert' conclusion is a restatement of the assumed definition, not a derivation of the necessity of dynamic routing. The formula is also only up to proportionality, so no quantitative content is added. Consequently, the paper's claim to have 'mathematically prov[ed] the necessity' of MuRA's adaptive mechanism is circular: the theorem's output is encoded in its input risk model.

  2. fitted input called prediction [Section 1 (Introduction), Figure 3 and the paragraph beginning 'Through systematic evaluation...']
    "we discover a strong linear correlation (R2 = 0.913) between the visual complexity of an input—quantified by image entropy [35] across RGB channels—and its optimal LoRA rank (Figure 3, Right)."

    The 'optimal LoRA rank' for each dataset is obtained by scanning static ranks on exactly those six datasets (Figure 3, Left). The linear regression of log2(rank) on image entropy is then fit to those same six optimal-rank points, and R^2=0.913 is reported. This is an in-sample coefficient of determination: it measures how well the fitted line explains the very data used to choose the ranks, not how well entropy predicts optimal rank on unseen data. The paper then uses this fitted 'discovery' as the evidence for the fundamental static-rank bottleneck and as the motivation for dynamic routing. The 'prediction' that high-entropy inputs require high ranks is therefore statistically forced by the fitting procedure rather than independently validated.

full rationale

The paper's main empirical contribution—MuRA's accuracy and efficiency on external benchmarks—is self-contained and not circular: the method is compared against published baselines and matched-parameter controls, and the core routing mechanism is trained on test data rather than derived from the criticized theorem. However, the paper's central theoretical claim is partially circular. Lemma 4.2 is presented as a 'rigorous mathematical proof' of the necessity of dynamic rank routing, but its conclusion is effectively built into the assumed local-risk model R_i(H)=α_i+λ_i d(H,μ_i)^2: low mismatch experts are favored because the risk was defined that way, and the expected-update formula is only proportional, hiding unquantified constants. The connected claim that static rank forces an 'inevitable optimization compromise' is supported by an in-sample correlation (Figure 3) computed on the same six datasets used to select the optimal ranks, with no held-out validation. I did not count Lemma 4.3 as circular because it is a standard per-step Lipschitz bound; the remark overreaches by calling it a guarantee of long-term stability, but the bound itself is not definitionally equivalent to its conclusion. I also found no load-bearing self-citation chain: the self-citations [18,31,32] appear only as related-work context, and the MROD initialization explicitly follows the external PiSSA work. Overall, the empirical results may stand, but the 'necessity' proof and the 'fundamental bottleneck' framing reduce, at least in part, to definitions and an in-sample fit, warranting a score of 6.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The central method depends on a manually chosen rank set, borrowed entropy-loss hyperparameters, an in-sample entropy-rank correlation, and two unproven or ad hoc theoretical moves in Lemma 4.2 and Lemma 4.3. No new physical entities are introduced. The experimental method is largely independent of the theory, which is why soundness is moderate rather than low.

free parameters (7)
  • Rank configuration = {2,4,8,16,32}
    Chosen by ablation (Figure 6a), not derived; the paper notes future work will derive boundaries from singular-value decay.
  • Number of augmented views = 63
    Set in Section 5.1 following TTL-style augmentation; no sensitivity analysis.
  • Adapter learning rate = 6e-3
    Manual choice with AdamW; no sweep reported.
  • Router learning rate = 1e-4
    Manual choice; separate learning rate for the router.
  • Entropy percentile rho = not reported
    Loss threshold tau is the rho-percentile of augmented-view entropies; rho is unspecified in the main text.
  • Adaptation depth = deepest visual layer
    Selected by ablation (Table 4); used because it gives the shortest backpropagation path.
  • Attention matrices updated = Q,K,V,O
    Selected by ablation (Figure 6b); updating the complete attention block maximizes gain.
assumptions (6)
  • standard math SVD decomposition W=USV^T exists and the initialization W = A_i B_i^T + R_i holds for each rank.
    Used in Section 3.2 to justify MROD initialization.
  • domain assumption Image entropy across RGB channels is a monotone proxy for visual complexity and optimal adaptation rank.
    Empirically asserted in Figure 3 with six dataset-level points and no held-out validation.
  • domain assumption Entropy minimization on augmented views is a valid training signal for unlabeled test-time adaptation.
    Borrowed from TTL [14] and used as the loss in Section 3.1.
  • ad hoc to paper The first-order approximation <g, f_i(h_l) - f_bar(h_l)> approximately equals l_i(x) - l_pi(x) holds with negligible error.
    Invoked before Lemma 4.2 with no derivation or error bound.
  • ad hoc to paper Expert loss is well modeled by R_i(H)=alpha_i + lambda_i d(H, mu_i)^2.
    Introduced in Section 4 so that the necessity conclusion follows from the definition of capacity mismatch.
  • domain assumption Gradients are bounded during test-time adaptation, so one-step Lipschitz boundedness implies stable accumulation of router updates.
    Lemma 4.3 assumes bounded gradients but does not prove that accumulated router movement avoids semantic drift or collapse.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MuRA: Multi-Rank Adaptation for Efficient and Effective Test-Time Vision-Language Generalization." pith.science (2026). https://pith.science/paper/3GPV7MON

@misc{pith2026260803885,
  author       = {Pith},
  title        = {Pith review of: MuRA: Multi-Rank Adaptation for Efficient and Effective Test-Time Vision-Language Generalization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3GPV7MON}},
  note         = {Machine review of arXiv:2608.03885}
}
read the original abstract

Vision-language models exhibit remarkable zero-shot capabilities but suffer significant performance degradation under distribution shifts. While test-time adaptation (TTA) via Low-Rank Adaptation offers a parameter-efficient solution, we identify a fundamental bottleneck in current methods: the reliance on static rank configurations. Because visual inputs inherently possess varying information densities, a fixed rank forces an inevitable optimization compromise, leading to underfitting on complex scenes and overfitting on simple ones. To bridge this gap, we propose Multi-Rank Adaptation (MuRA), a novel framework that dynamically selects and fuses adaptation modules of varying capacities based on token-level visual complexity. MuRA synergizes Multi-Rank Orthogonal Decomposition to provide a superior, knowledge-preserving initialization, and Unified Component Fusion with Continuous Router Updating to sustainably learn semantic-to-rank mappings. Furthermore, we provide rigorous theoretical justifications mathematically proving the necessity and gradient stability of this adaptive mechanism. Crucially, MuRA's dynamic design uniquely thrives at the deepest visual layer, capitalizing on the shortest gradient backpropagation path. Extensive experiments demonstrate that MuRA achieves state-of-the-art accuracy across extensive domain generalization and cross-dataset benchmarks while significantly reducing both computational and memory overhead.

Figures

Figures reproduced from arXiv: 2608.03885 by the authors.

Figure 1
Figure 1. Overview of test-time adaptation approaches in vision-language models. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Effectiveness and efficiency of MuRA. (Left) MuRA [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. The stabilizing effect of MROD-Initialization and [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Overall Architecture of Multi-Rank Adaptation (MuRA). MuRA is an adaptive framework integrated into the deepest [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: Analysis of (a) different rank configurations, where [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 8
Figure 8. Figure 8: MuRA dynamically stabilizes and adapts its rank [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 7
Figure 7. Figure 7: Visualization of attention maps and top-3 predic [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 12 canonical work pages

  1. [1]

    Jameel Abdul Samadh, Mohammad Hanan Gani, Noor Hussein, Muhammad Uzair Khattak, Muhammad Muzammal Naseer, Fahad Shahbaz Khan, and Salman H Khan. 2023. Align your prompts: Test-time prompting with distribution alignment for zero-shot generalization.Advances in Neural Information Processing Systems 36 (2023), 80396–80413

    work page 2023

  2. [2]

    Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. 2014. Food-101–mining discriminative components with random forests. InEuropean conference on com- puter vision. Springer, 446–461

    2014

  3. [3]

    Xinyu Chen, Haotian Zhai, Can Zhang, Xiupeng Shi, and Ruirui Li. 2025. Multi- Cache Enhanced Prototype Learning for Test-Time Generalization of Vision- Language Models. InProceedings of the IEEE/CVF International Conference on Computer Vision. 2281–2291

    work page 2025

  4. [4]

    Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and An- drea Vedaldi. 2014. Describing textures in the wild. InProceedings of the IEEE conference on computer vision and pattern recognition. 3606–3613

    2014

  5. [5]

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Imagenet: A large-scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition. Ieee, 248–255

    2009

  6. [6]

    Xinqi Fan, Xueli Chen, Luoxiao Yang, Chuin Hong Yap, Rizwan Qureshi, Qi Dou, Moi Hoon Yap, and Mubarak Shah. 2025. Test-Time Retrieval-Augmented Adap- tation for Vision-Language Models. InProceedings of the IEEE/CVF International Conference on Computer Vision. 8810–8819

    2025

  7. [7]

    Li Fei-Fei, Rob Fergus, and Pietro Perona. 2004. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In2004 conference on computer vision and pattern recognition workshop. IEEE, 178–178

    2004

  8. [8]

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

    work page 2023

Show all 46 references
  1. [9]

    Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. 2024. Clip-adapter: Better vision-language models with feature adapters.International Journal of Computer Vision132, 2 (2024), 581–595

    2024

  2. [10]

    Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. 2019. 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 Sensing12, 7 (2019), 2217–2226

    2019

  3. [11]

    Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. 2021. The many faces of robustness: A critical analysis of out-of-distribution generalization. InProceedings of the IEEE/CVF internatio...

    2021

  4. [12]

    Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song

  5. [13]

    Zhaohong Huang, Yuxin Zhang, Jingjing Xie, Fei Chao, and Rongrong Ji. 2025. GS-Bias: Global-Spatial Bias Learner for Single-Image Test-Time Adaptation of Vision-Language Models.arXiv preprint arXiv:2507.11969(2025)

    2025 arXiv

  6. [14]

    Raza Imam, Hanan Gani, Muhammad Huzaifa, and Karthik Nandakumar. 2025. Test-time low rank adaptation via confidence maximization for zero-shot gen- eralization of vision-language models. In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV). IEEE, 5449–5459

    2025

  7. [15]

    Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. 1991. Adaptive mixtures of local experts.Neural computation3, 1 (1991), 79–87

    1991

  8. [16]

    Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, and Eric Xing. 2024. Efficient test-time adaptation of vision-language models. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 14162–14171

    2024

  9. [17]

    Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 2013. 3d object repre- sentations for fine-grained categorization. InProceedings of the IEEE international conference on computer vision workshops. 554–561

    2013

  10. [18]

    Weixian Lei, Jiacong Wang, Haochen Wang, Xiangtai Li, Jun Hao Liew, Jiashi Feng, and Zilong Huang. 2025. The scalability of simplicity: Empirical analysis of vision- language learning with a single transformer.arXiv preprint arXiv:2504.10462 (2025)

    2025 arXiv

  11. [19]

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

    2022

  12. [20]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual in- struction tuning.Advances in neural information processing systems36 (2023), 34892–34916

    2023

  13. [21]

    Ilya Loshchilov and Frank Hutter. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101(2017)

    2017 arXiv

  14. [22]

    Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi

  15. [23]

    Fanxu Meng, Zhaohui Wang, and Muhan Zhang. 2024. Pissa: Principal singular values and singular vectors adaptation of large language models.Advances in Neural Information Processing Systems37 (2024), 121038–121072

    2024

  16. [24]

    Maria-Elena Nilsback and Andrew Zisserman. 2008. Automated flower classifica- tion over a large number of classes. In2008 Sixth Indian conference on computer vision, graphics & image processing. IEEE, 722–729

    2008

  17. [25]

    Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. 2012. Cats and dogs. In2012 IEEE conference on computer vision and pattern recognition. IEEE, 3498–3505

    2012

  18. [26]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. InInternational conference on machine learnin...

    2021

  19. [27]

    Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. 2019. Do imagenet classifiers generalize to imagenet?. InInternational conference on machine learning. PMLR, 5389–5400

    2019

  20. [28]

    Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anand- kumar, and Chaowei Xiao. 2022. Test-time prompt tuning for zero-shot gener- alization in vision-language models.Advances in Neural Information Processing Systems35 (2022), 14274–14289

    2022

  21. [29]

    Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. 2012. Ucf101: A dataset of 101 human actions classes from videos in the wild.arXiv preprint arXiv:1212.0402(2012)

    2012 arXiv

  22. [30]

    Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. 2019. Learning robust global representations by penalizing local predictive power.Advances in neural information processing systems32 (2019)

    2019

  23. [31]

    Jiacong Wang, Zijian Kang, Haochen Wang, Haiyong Jiang, Jiawen Li, Bohong Wu, Ya Wang, Jiao Ran, Xiao Liang, Chao Feng, et al. 2025. Vgr: Visual grounded reasoning.arXiv preprint arXiv:2506.11991(2025)

    2025 arXiv

  24. [32]

    Jiacong Wang, Bohong Wu, Haiyong Jiang, Xun Zhou, Xin Xiao, Haoyuan Guo, and Jun Xiao. 2024. World to code: Multi-modal data generation via self-instructed compositional captioning and filtering.arXiv preprint arXiv:2409.20424(2024)

    2024 arXiv

  25. [33]

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. 2024. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191(2024)

    2024 arXiv

  26. [34]

    Xiangyu Wu, Dongming Jiang, Feng Yu, Yueying Tian, Jiaqi Tang, Qing-Guo Chen, Yang Yang, and Jianfeng Lu. 2026. Adaptive Debiasing Tsallis Entropy for Test-Time Adaptation.arXiv preprint arXiv:2602.11743(2026)

    2026

  27. [35]

    Yue Wu, Yicong Zhou, George Saveriades, Sos Agaian, Joseph P Noonan, and Premkumar Natarajan. 2013. Local Shannon entropy measure with statistical tests for image randomness.Information Sciences222 (2013), 323–342

    2013

  28. [36]

    Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba

  29. [37]

    Maxime Zanella and Ismail Ben Ayed. 2024. On the test-time zero-shot gen- eralization of vision-language models: Do we really need prompt learning?. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion. 23783–23793

    2024

  30. [38]

    Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. 2022. Unified vision and language prompt learning.arXiv preprint arXiv:2210.07225 (2022)

    2022 arXiv

  31. [39]

    Ce Zhang, Simon Stepputtis, Katia Sycara, and Yaqi Xie. 2024. Dual prototype evolving for test-time generalization of vision-language models.Advances in Neural Information Processing Systems37 (2024), 32111–32136

    2024

  32. [40]

    Renrui Zhang, Rongyao Fang, Wei Zhang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. 2021. Tip-adapter: Training-free clip-adapter for better vision-language modeling.arXiv preprint arXiv:2111.03930(2021)

    2021 arXiv

  33. [41]

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. 2022. Conditional prompt learning for vision-language models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 16816–16825

    2022

  34. [42]

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. 2022. Learning to prompt for vision-language models.International Journal of Computer Vision 130, 9 (2022), 2337–2348

    2022

  35. [43]

    Lihua Zhou, Mao Ye, Shuaifeng Li, Nianxin Li, Xiatian Zhu, Lei Deng, Hongbin Liu, and Zhen Lei. 2025. Bayesian test-time adaptation for vision-language models. InProceedings of the Computer Vision and Pattern Recognition Conference. 29999–30009

    2025

  36. [2010]

    In2010 IEEE computer society conference on computer vision and pattern recognition

    Sun database: Large-scale scene recognition from abbey to zoo. In2010 IEEE computer society conference on computer vision and pattern recognition. IEEE, 3485–3492

  37. [2013]

    Fine-grained visual classification of aircraft.arXiv preprint arXiv:1306.5151 (2013)

    2013 arXiv

  38. [2021]

    InProceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Natural adversarial examples. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 15262–15271

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.