Pith. sign in

REVIEW 4 major objections 5 minor 35 references

Attn-Adapter: Attention Is All You Need for Online Few-shot Learner of Vision-Language Model

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read The paper proposes Attn-Adapter, an online few-shot learner whose two cross-attention modules refine CLIP's category and image embeddings, reporting accuracy gains over prior online baselines without retraining the CLIP backbone.

desk verdict A genuinely new dual-attention adapter for online few-shot CLIP, but the local-feature definition is missing and two evaluation tables disagree, so the SOTA claim is not yet verifiable. read the letter →

arxiv 2509.03895 v1 pith:KBCZ3K5P submitted 2025-09-04 cs.CV

classification cs.CV
keywords few-shotlearningvision-languagemodelsCLIPcross-attentiononlineadaptationlocal-globalfeaturefusioncross-datasetgeneralizationbase-to-novel
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 claims that a frozen CLIP model can be adapted to a new task with only a handful of labeled images, no prompt tuning, and no backbone updates, provided the adaptation is done by two lightweight attention modules. The first module treats text-derived class descriptions as queries and few-shot support images as keys and values, pulling each category embedding toward the support examples that actually match it. The second module treats the global image embedding as a query over local image features, sharpening the embedding before classification. On the paper's experiments, this dual attention scheme improves over online baselines such as Tip-Adapter and Meta-Adapter on cross-dataset transfer and on new (novel) categories, while training in about the same time as Meta-Adapter. If correct, this gives a practical recipe for adapting CLIP to new domains in minutes rather than hours.

What carries the argument

The load-bearing object is a dual cross-attention architecture built on top of frozen CLIP features. In the Memory Attn-Adapter, class embeddings from the text encoder are the queries and the support image embeddings are the keys and values; after softmax attention, the support features are aggregated into F̂ and merged into the category embedding via a learned projector p. In the Local-Global Attn-Adapter, the global image embedding g is the query and local features l are the keys and values, giving l̂, which is merged into f = g + p(g) ⊙ l̂. Both adapters use MLP projections only for keys and queries, leaving the zero-shot values unchanged, so the original CLIP features remain reusable. Th

What would settle it

Take the best reported setting (ViT-B/16 on ImageNet) and rerun Attn-Adapter with the local features l replaced by random patch embeddings, or by the global feature g itself; if accuracy on ImageNet-A and ImageNet-Sketch does not drop substantially, the Local-Global adapter is not doing the work the paper attributes to it.

Watch

Extended reading notes

Core claim

Attn-Adapter's central claim is that the adjustments Tip-Adapter and Meta-Adapter make to CLIP's image-class similarities — a hand-tuned exponential modulation and a learned residual — can be replaced by two learned attention updates trained end-to-end on the few-shot support set. The Memory Attn-Adapter computes ŵ = w + p(w) ⊙ F̂, merging zero-shot class embeddings with cross-attention-aggregated support embeddings through a learned projector p. The Local-Global Attn-Adapter computes f = g + p(g) ⊙ l̂, merging the global image embedding with local features through cross-attention. Training uses cross-entropy on the few-shot classes plus an L2 regularizer pulling f back toward g. The paper r

Load-bearing premise

Everything rests on the local features l in the Local-Global Attn-Adapter being informative and aligned with the global image feature g; the paper never states how l is produced, so if the chosen CLIP internal features do not capture the discriminative parts of the image, the cross-attention fusion cannot be what drives the reported gains.

Editorial extensions

If this is right

  • Few-shot adaptation can move online: Attn-Adapter trains in about 21 minutes on ImageNet (10 epochs) versus 17+ hours for offline prompt-learning methods like CLIP-Adapter or CoOp.
  • A single adaptation on ImageNet transfers to other datasets: the paper reports an average cross-dataset gain of +3.64 over Meta-Adapter across ten datasets, and +6.15 when hyperparameters are tuned per dataset.
  • The method acts as a cheap robustness fix: after ImageNet adaptation it improves over zero-shot CLIP by 10+ points on ImageNet-A and ImageNet-Sketch and by 12+ on ImageNet-V2 (ViT-B/16).
  • The gains are not limited to one architecture: reported improvements grow from ResNet-50 to ViT-B/16 and RN50×64, suggesting the adapter benefits from richer visual encoders.
  • In base-to-novel splits the gain concentrates on novel categories rather than memorized base classes, e.g. ImageNet novel accuracy rises from Meta-Adapter's 40.8 to 43.6 and UCF101 from 47.7 to 52.2.

Reading between the lines

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

  • Beyond the paper: the local features l are never defined in the manuscript; a natural test is whether they should be CLIP's last-layer patch tokens, and whether the cross-attention in Eq. (5)–(6) needs spatial filtering to work.
  • Beyond the paper: the Memory Attn-Adapter is a soft aggregation over support examples, so the same equations could accept non-image support (e.g., class descriptions or region masks); the paper does not try this.
  • Beyond the paper: the 21-minute ImageNet training, with the L2 anchor to g, hints that Attn-Adapter could be run as a streaming or test-time adapter as new labeled samples arrive; the paper evaluates only fixed support sets.
  • Beyond the paper: the mechanism could be composed with cache-based methods (e.g., refine support keys with the Memory Adapter before a Tip-Adapter-style lookup); such a combination is not explored.
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 Attn-Adapter, an online few-shot CLIP adaptation method with two attention modules: a Memory Attn-Adapter that refines category embeddings from support features via cross-attention, and a Local-Global Attn-Adapter that refines image embeddings by attending to local features. The method is evaluated on cross-dataset and cross-category few-shot benchmarks, comparing mainly against Tip-Adapter and Meta-Adapter. The abstract claims state-of-the-art efficiency and accuracy with stable scaling across CLIP backbones. The reported results show gains of +3.64% average over Meta-Adapter in Table 1 and large improvements on domain-shifted ImageNet variants in Table 2, plus a backbone-compatibility study in Table 5.

Significance. If the central claim is valid, Attn-Adapter would be a useful lightweight alternative to prompt-tuning for few-shot CLIP, since it avoids backbone updates and adds only modest overhead. The paper addresses a relevant problem and includes experiments across multiple datasets and backbones, which is valuable. However, the manuscript as submitted is not reproducible: the local features l used in the second component are never defined, the equations are dimensionally ambiguous, and several important experimental settings are missing. Because the Local-Global Attn-Adapter is one of the two core contributions, the SOTA claim cannot currently be verified from the text. The paper ships no code, no supplementary material despite references to one, and no error bars. These omissions are fixable, so the central idea is not discredited, but the manuscript needs substantial revision before the claims can be accepted.

major comments (4)
  1. [§3.2.2, Eq. (5)–(6)] The local features l are never defined. The text states that 'global and local features (g, l)' are passed through the Local-Global Attn-Adapter, but it does not specify whether l is the CLS-patch sequence for ViT, an intermediate feature map for ResNet, which encoder layer is used, or what shape l has. Since Eq. (5) uses l as keys and values and Eq. (6) uses the result to refine the global embedding, the entire second component is uncomputable as written. If l is not carefully aligned with g, the reported gains in Tables 1 and 5 would not follow from the described mechanism. Please define l precisely for each backbone and state the softmax dimension in Eq. (5).
  2. [§3.2.1, Eq. (4) and §3.2.2, Eq. (5)] The dimensionality of the cross-attention equations is ambiguous. In Eq. (2), F_i is described as a D×K matrix, while in Eq. (4) the expression F^T σ(MLP_K(F) MLP_Q(w)^⊤ / √D) only makes sense with another convention. Similarly, Eq. (5) uses l^T σ(MLP_K(l) MLP_Q(g)^T / √D), but the expected softmax dimension is not stated. Without consistent shapes and a clear description of the batch/head handling, the method cannot be implemented. Please rewrite the equations with explicit dimensions and provide details of the MLP layers, number of attention heads, hidden sizes, temperature τ, and regularization weight λ.
  3. [Tables 2, 3, and 6] The experimental protocol is incompletely reported. Table 2 lists Attn-Adapter on ImageNet-V2 but omits Tip-Adapter and Meta-Adapter on that dataset, making the claimed '+12% on ImageNet-V2' untestable. Tables 3 and 6 appear to use different base/novel splits: Table 3 follows Meta-Adapter [26] while Table 6 follows CoCoOp's framework [34], yet no cross-reference or explanation is given for the differences in reported base/novel numbers (e.g., ImageNet Base 87.35 vs 81.4). The absence of seeds or error bars further weakens the claim that Attn-Adapter consistently outperforms baselines. Please report the exact split used for each table and add variance or at least multiple seeds.
  4. [Table 5, Backbone Compatibility] The backbone scaling claim is contradicted by the numerical results: Attn-Adapter obtains 50.35% on RN50, 58.17% on RN101, 56.42% on ViT-B/32, 55.65% on ViT-B/16, 63.69% on RN50×16, and 66.27% on RN50×64. ViT-B/16 (55.65%) is lower than ViT-B/32 (56.42%) and RN101 (58.17%), which is inconsistent with the stated 'consistent upward trend' and the broader narrative that the method scales smoothly with backbone strength. This may be a typo or a sign of an implementation issue; please check and clarify, since the scaling claim is part of the abstract.
minor comments (5)
  1. [§3.2.2, first sentence] Typo: 'the Memory Attn-Adapter aggregates the local and global features' should be 'the Local-Global Attn-Adapter'.
  2. [§7, Supplementary Materials] The text refers to 'Supplementary Material' for training-time and further evaluations, but no supplementary material is included in the uploaded manuscript. Either include it or remove the references.
  3. [Figure 1] The figure is mentioned in §3.2 but not present in the text. A diagram would help clarify how the two adapters are connected during fine-tuning and inference.
  4. [§3.1, Eq. (2)] The notation L_j for one-hot labels is used in the Tip-Adapter equation but not clearly defined; also the indexing of F_i and L_i is confusing when later F is used as a full matrix.
  5. [Related Work] Reference [1] appears to be a self-citation to an arXiv preprint; ensure that all citations are properly placed and that the distinction between offline and online methods is accurate.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: Attn-Adapter is trained on base classes and evaluated on held-out categories/datasets; the reported gains are not constructed from fitted values. Minor self-citations are not load-bearing. The undefined local features l are a reproducibility gap, not a circular derivation.

full rationale

Attn-Adapter's derivation chain is not circular. The two adapters (Memory Attn-Adapter, Eq. (4), and Local-Global Attn-Adapter, Eqs. (5)-(6)) are trained with the contrastive and L2 losses (Eqs. (7)-(8)) on the base split, then evaluated on novel categories and on unseen datasets (Tables 1-3, Secs. 4.1-4.2). No fitted hyperparameter or learned weight is renamed as a prediction: the reported cross-dataset and cross-category numbers are held-out evaluations, not in-sample fits. The paper's self-citations [1] and [22] appear only in the Related Work enumeration of offline/online methods and do no load-bearing work; the split design and training settings are attributed to Meta-Adapter [26], an external source. No uniqueness theorem or prior-work ansatz is invoked to force the architecture. The most serious issue is that the local feature l in Eq. (5) is never defined (which CLIP layer, patch-token resolution, or feature-map shape), and no code is released, making the central mechanism unverifiable as written. That is a reproducibility and correctness risk, not a circularity: the equations are not equivalent to their inputs by construction, and the empirical claim is externally benchmarked. Hence score 1.

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

The central claim depends on the quality of frozen CLIP features, the transferability of base-class training to novel classes and datasets, and the availability of informative local features. The architecture adds trainable adapters but no new physical entities.

free parameters (5)
  • learnable projector p = not reported
    Balances category/support and global/local embeddings in Eqs. (4)-(6); learned on base classes.
  • MLP key/query weights (MLP_K, MLP_Q) = not reported
    Linear projections for attention in Eqs. (4)-(5); learned during few-shot fine-tuning.
  • temperature tau = not reported
    Softmax temperature in the cross-entropy loss in Eq. (7); value is not stated.
  • regularization weight lambda = not reported
    Weights the L2 loss in Eq. (8); no value is reported.
  • attention heads and hidden dimensions = not reported
    Architecture hyperparameters for the multi-head attention modules; not specified, required for reproduction.
assumptions (4)
  • domain assumption CLIP text embeddings w_i and image embeddings f encode semantically useful representations for the target classes.
    The whole method operates on frozen CLIP features and never updates the backbone.
  • domain assumption Cross-entropy training on base classes transfers to novel classes and to new datasets.
    The adapters are tuned on ImageNet or base splits and evaluated frozen on novel classes and datasets.
  • domain assumption Local features l are well-defined and complementary to global features.
    Eqs. (5)-(6) use l without specifying how it is extracted from CLIP; if l is uninformative, the Local-Global Adapter cannot help.
  • standard math Standard matrix multiplication and softmax operations as defined in Eqs. (1)-(8).
    Attention formulas assume standard linear algebra and normalization.
invented entities (2)
  • Memory Attn-Adapter
    purpose: Cross-attention module that refines category embeddings using support embeddings as keys and values.
    A trainable module; no external evidence beyond this paper's reported benchmarks.
  • Local-Global Attn-Adapter
    purpose: Cross-attention module that enriches image embeddings by integrating local and global features.
    A trainable module; no independent verification outside the reported experiments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Attn-Adapter: Attention Is All You Need for Online Few-shot Learner of Vision-Language Model." pith.science (2026). https://pith.science/paper/KBCZ3K5P

@misc{pith2026250903895,
  author       = {Pith},
  title        = {Pith review of: Attn-Adapter: Attention Is All You Need for Online Few-shot Learner of Vision-Language Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KBCZ3K5P}},
  note         = {Machine review of arXiv:2509.03895}
}
read the original abstract

Contrastive vision-language models excel in zero-shot image recognition but face challenges in few-shot scenarios due to computationally intensive offline fine-tuning using prompt learning, which risks overfitting. To overcome these limitations, we propose Attn-Adapter, a novel online few-shot learning framework that enhances CLIP's adaptability via a dual attention mechanism. Our design incorporates dataset-specific information through two components: the Memory Attn-Adapter, which refines category embeddings using support examples, and the Local-Global Attn-Adapter, which enriches image embeddings by integrating local and global features. This architecture enables dynamic adaptation from a few labeled samples without retraining the base model. Attn-Adapter outperforms state-of-the-art methods in cross-category and cross-dataset generalization, maintaining efficient inference and scaling across CLIP backbones.

Figures

Figures reproduced from arXiv: 2509.03895 by the authors.

Figure 1
Figure 1. Illustration of the Attn-Adapter model, which utilizes a trainable network with two attention-based components to adjust the [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 24 canonical work pages

  1. [26]

    Meta-adapter: An online few-shot learner for vision-language model

    Lin Song, Ruoyi Xue, Hang Wang, Hongbin Sun, Yixiao Ge, Ying Shan, et al. Meta-adapter: An online few-shot learner for vision-language model. Advances in Neural Information Processing Systems, 36:55361–55374, 2023. 1, 2, 3, 4, 5

  2. [34]

    Conditional prompt learning for vision-language models

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816–16825, 2022. 1, 2, 4, 5, 6

  3. [1]

    Accelerating Conditional Prompt Learning via Masked Image Modeling for Vision-Language Models

    Phuoc-Nguyen Bui, Khanh-Binh Nguyen, and Hyunseung Choo. Accelerating conditional prompt learning via masked image modeling for vision-language models. arXiv preprint arXiv:2508.04942, 2025. 1

  4. [2]

    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. 3

  5. [3]

    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 Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255. Ieee, 2009. 3, 4

  6. [4]

    An image is worth 16x16 words: Trans- formers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4

  7. [5]

    Learning to prompt for open-vocabulary object Table 5

    Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object Table 5. Quantitative results on ImageNet of different models utilized various vision backbones. Model RN50 RN101 ViT-B/32 ViT-B/16 RN50×16 RN50×64 Zero-shot CLIP 32.82 39.22 40.10 45.77 50.10 54.67 Tip-Adapter 36.51 42.42 43.71 49.84 53.08 5...

  8. [6]

    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 Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshop, pages 178–178. IEEE, 2004. 3

Show all 35 references
  1. [7]

    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. arXiv preprint arXiv:2110.04544, 2021. 1, 2, 5

  2. [8]

    Open- vocabulary object detection via vision and language knowl- edge distillation

    Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open- vocabulary object detection via vision and language knowl- edge distillation. arXiv preprint arXiv:2104.13921 , 2021. 2

  3. [9]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770–778, 2016. 4

  4. [10]

    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. 3

  5. [11]

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

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

  6. [12]

    Natural adversarial examples

    Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262–15271, 2021. 3, 4

  7. [13]

    Gallop: Learning global and local prompts for vision-language models

    Marc Lafon, Elias Ramzi, Clément Rambour, Nicolas Aude- bert, and Nicolas Thome. Gallop: Learning global and local prompts for vision-language models. In European Conference on Computer Vision, pages 264–282. Springer, 2025. 2, 3

  8. [14]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 4

  9. [15]

    Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks

    Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 2

  10. [16]

    Fine-grained visual clas- sification of aircraft

    Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual clas- sification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 3

  11. [17]

    Locoop: Few-shot out-of-distribution detection via prompt learning

    Atsuyuki Miyai, Qing Yu, Go Irie, and Kiyoharu Aizawa. Locoop: Few-shot out-of-distribution detection via prompt learning. Advances in Neural Information Processing Systems, 36, 2024. 3

  12. [18]

    Debiasing, calibrating, and improving semi-supervised learning performance via simple ensemble projector

    Khanh-Binh Nguyen. Debiasing, calibrating, and improving semi-supervised learning performance via simple ensemble projector. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2441–2451, 2024. 1

  13. [19]

    Sequencematch: Revisiting the design of weak-strong augmentations for semi-supervised learning

    Khanh-Binh Nguyen. Sequencematch: Revisiting the design of weak-strong augmentations for semi-supervised learning. In Proceedings of the IEEE/CVF winter conference on appli- cations of computer vision, pages 96–106, 2024. 1

  14. [20]

    On calibration of prompt learning using temperature scaling

    Khanh-Binh Nguyen and Chae Jung Park. On calibration of prompt learning using temperature scaling. IEEE Access,

  15. [21]

    Boosting semi- supervised learning by bridging high and low-confidence predictions

    Khanh-Binh Nguyen and Joon-Sung Yang. Boosting semi- supervised learning by bridging high and low-confidence predictions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1028–1038, 2023. 1

  16. [22]

    Adaptive cache enhancement for test- time adaptation of vision-language models

    Khanh-Binh Nguyen, Phuoc-Nguyen Bui, Hyunseung Choo, and Duc Thanh Nguyen. Adaptive cache enhancement for test- time adaptation of vision-language models. arXiv preprint arXiv:2508.07570, 2025. 2

  17. [23]

    Proto-clip: Vision-language prototypical network for few-shot learning

    Kamalesh Palanisamy, Yu-Wei Chao, Xinya Du, Yu Xiang, et al. Proto-clip: Vision-language prototypical network for few-shot learning. In 2024 IEEE/RSJ International Confer- ence on Intelligent Robots and Systems (IROS), pages 2594–

  18. [24]

    Cats and dogs

    Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages 3498–3505. IEEE, 2012. 3

  19. [25]

    Learning transferable visual models from natural language supervision

    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 supervision. In International Conference on Machine Learning , p...

  20. [27]

    Ucf101: A dataset of 101 human actions classes from videos in the wild

    Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 3

  21. [28]

    Lxmert: Learning cross-modality encoder representations from transformers

    Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 2

  22. [29]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3

  23. [30]

    Learning robust global representations by penalizing local predictive power

    Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. Advances in Neural Information Processing Systems, 32, 2019. 3, 4

  24. [31]

    Sun database: Large-scale scene recog- nition from abbey to zoo

    Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recog- nition from abbey to zoo. In Proceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition, pages 3485–3492. IEEE, 2010. 3

  25. [32]

    Class-aware visual prompt tuning for vision-language pre-trained model

    Yinghui Xing, Qirui Wu, De Cheng, Shizhou Zhang, Guo- qiang Liang, and Yanning Zhang. Class-aware visual prompt tuning for vision-language pre-trained model. arXiv preprint arXiv:2208.08340, 2022. 2

  26. [33]

    Tip- adapter: Training-free clip-adapter for better vision-language modeling

    Renrui Zhang, Rongyao Fang, Wei Zhang, Peng Gao, Kun- chang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip- adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint arXiv:2111.03930, 2021. 1, 2, 3, 4

  27. [35]

    Learning to prompt for vision-language models

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. Interna- tional Journal of Computer Vision, 130(9):2337–2348, 2022. 1, 5

Pith tools

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