Pith. sign in

REVIEW 5 major objections 7 minor 67 references

A Closer Look at Conditional Prompt Tuning for Vision-Language Models

T0 review · 5 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper argues that prompt tuning for vision-language models should condition dynamic prompts on textual class names rather than visual image information, and that this choice raises base-to-new harmonic accuracy by about 2.6 points on…

desk verdict A cheap, plausible prompt-tuning plugin with a genuinely surprising random-noise result, but the central 'text is the key' claim is confounded by per-class versus task-level conditioning and needs a missing control. read the letter →

arxiv 2506.23856 v1 pith:YECUCIEQ submitted 2025-06-30 cs.CV

Ji Zhang , Shihan Wu , Lianli Gao , Jingkuan Song , Nicu Sebe , Heng Tao Shen This is my paper
classification cs.CV
keywords PrompttuningConditionalVision-languagemodelsFew-shotlearningBase-newtradeoffClass-adaptivepromptsTextualclassinformationCLIP
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

Prompt tuning adapts CLIP-style vision-language models to a task by learning a few prompt vectors, but tuned prompts tend to overfit base classes and lose accuracy on new classes. The paper identifies this base-new tradeoff and questions the usual remedy of conditioning prompts on per-image visual features. It reports that prompts conditioned on fixed random noise outperform the image-conditioned version, while prompts conditioned on class-name text embeddings match the best class-level visual oracle. The proposed CaPT method learns class-adaptive prompt offsets from frozen textual class embeddings through a small Meta-Net, plus a margin-adjusted loss that stops similar base classes from being pushed too far apart. Plugged into five existing unconditional prompt-tuning methods, CaPT raises average harmonic base-to-new accuracy by about 2.6 points with roughly 0.9 per-mille extra training time, and combining it with DePT sets the best reported harmonic mean on the 11 datasets.

What carries the argument

The load-bearing object is the class-conditional prompt $c_i(e_i) = \{v_1 + h_\alpha(e_i), \ldots, v_l + h_\alpha(e_i), [CLS]_i\}$, where $e_i = \varphi(\text{class name})$ is a frozen textual class embedding, $h_\alpha$ is a two-layer MLP Meta-Net, and the $v$'s are the learned context vectors. The offset $h_\alpha(e_i)$ lets a prompt trained on base classes generate modified prompts for unseen new classes at inference, because class names are always available. The margin-ITM loss uses pairwise cosine similarities among base-class text embeddings to subtract a class-similarity margin from the classification logits, keeping similar classes from being pushed too far apart in the learned feature space.

What would settle it

Train CaPT on a set of base classes, then evaluate on new classes whose name embeddings have been deliberately pushed far from every base-class embedding, for example by adding large random perturbations to the class-name vectors. If harmonic accuracy collapses to the level of the same prompt without CaPT, the claim that textual class conditioning drives the gain is falsified. A cheaper correlational check is to plot each dataset's CaPT gain against the average cosine distance between that dataset's new-class embeddings and the base-class embedding centroid; the paper's mechanism predicts a clear negative relationship.

Watch

Extended reading notes

Core claim

The paper's controlled comparison replaces the image-conditioning in CoCoOp with four alternatives: fixed random noise, visual task prototypes, visual class prototypes, and textual class embeddings. It finds that image-conditioning is not the source of the new-class generalization gain; even random noise does no worse, class-level visual information helps more, and textual class information matches the oracle class-level visual variant. The central claim is therefore that dynamic prompts conditioned on Textual Class Information (TCI) are the key to overcoming the Base-New Tradeoff in prompt tuning. CaPT operationalizes this by passing frozen class-name embeddings through a lightweight MLP Meta-Net that outputs class-adaptive offsets to learned prompt vectors, and a margin-adjusted Image-Text Matching loss that prevents over-separation of semantically similar base classes.

Load-bearing premise

The Meta-Net is assumed to turn embeddings of class names it has never seen into useful prompt adjustments; if a new class's name embedding sits far outside the range of the base classes, the conditioning signal becomes indistinguishable from noise.

Editorial extensions

If this is right

  • Adding CaPT to CoOp raises average harmonic base-to-new accuracy from 74.84 to 77.44, and the same plugin lifts four other unconditional baselines by between 0.78 and 3.89 H points.
  • Because new-task class names are available at inference, the same learned Meta-Net can adapt a base-trained prompt to unseen classes without seeing any of their images.
  • Integrating CaPT with DePT yields DeCaPT, which reaches 79.63 average H on 11 datasets, 3.49 points above the Bayes conditional prompt-tuning scheme.
  • CaPT's gains persist across 1, 2, 4, 8 and 16 shots, and in the 1-shot setting it is the only method tested that stays above CLIP's zero-shot average accuracy.
  • The plugin adds only about 0.9 per-mille extra training time when applied to the five tested baselines, making the improvement nearly computation-free.

Reading between the lines

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

  • If class-name text is what drives generalization, per-image conditioning in existing conditional prompt-tuning methods could likely be dropped entirely at inference, saving compute without hurting accuracy; the paper does not make this claim directly.
  • The mechanism predicts a testable correlation: CaPT's gain on a new class should shrink as that class's name embedding moves farther from the base-class embedding distribution, a prediction the paper does not examine.
  • The margin-ITM principle, using semantic similarity to avoid over-separation of classes, may transfer beyond prompts to other parameter-efficient tuning techniques such as adapters or low-rank updates, though the paper only tests it for prompt tuning.
  • Because the Meta-Net is trained only on base-class name embeddings, the method's behavior on class vocabularies far outside that distribution, such as specialized medical or technical terms, remains an open untested boundary.
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

5 major / 7 minor

Summary. This paper studies conditional prompt tuning for CLIP. It reports a diagnostic study comparing CoCoOp variants conditioned on fixed random noise, visual task information, visual class information, and textual class information, concluding that textual class information (TCI) is the key to overcoming the base-new tradeoff (BNT). It then proposes Class-adaptive Prompt Tuning (CaPT), which uses a Word2Vec/CLIP-text-encoder to extract per-class embeddings, feeds them through a lightweight Meta-Net to produce prompt offsets, and trains with a margin-adjusted ITM loss. CaPT is evaluated as a plug-in on five unconditional baselines across 11 datasets and three transfer settings, reporting average harmonic-mean improvements of about 2.6 points at negligible computational cost. Combining CaPT with the authors' DePT yields DeCaPT, reported to outperform Bayes by 3.49 average H.

Significance. If the claims hold, CaPT would be a simple, cheap and broadly applicable plugin for improving base-to-new generalization in prompt tuning, and the diagnostic finding about VII versus TCI would redirect the design of conditional prompting. The paper's strengths are its clear exposition, the breadth of experiments (base-to-new, cross-dataset, cross-domain), the low induced overhead, and the public code release. However, the central causal claim is currently under-supported because the comparison confounds conditioning source with per-class granularity, and the quantitative evidence is weakened by test-set hyperparameter selection, missing variance estimates, and a few internally inconsistent table entries.

major comments (5)
  1. [Sec. 3.2, Fig. 2(d)] The experimental design does not isolate the effect of textual semantics from per-class conditioning. CoCoOp-FRN uses one fixed random vector shared by all classes, while CoCoOp-TCI uses a distinct text embedding per class; CoCoOp-VCI also uses a distinct vector per class and reaches H=77.32, close to CoCoOp-TCI's 76.91. A per-class fixed random-vector variant is required to establish that TCI is the key; without it, class-level granularity alone is a plausible explanation. Please add such a control (e.g., CoCoOp-FCN) and report per-variant standard deviations, since all differences are within about one H point.
  2. [Sec. 4.3.1, Fig. 5] The scaling weight epsilon and the number of training epochs are selected by sweeping on the same 11 evaluation datasets and picking the best average (epsilon=0.31, epochs=10). This is test-set model selection; the reported average gains may be optimistic. Please use a per-dataset validation split or report results with a pre-registered fixed epsilon and epoch count, and quantify sensitivity to these choices.
  3. [Table 2, Food101 ProDA row] The H=98.98 entry for ProDA on Food101 is arithmetically impossible given Base=90.37 and New=89.59, since the harmonic mean always lies between the two accuracies. This suggests a data entry error and undermines confidence in the SOTA comparison table; all entries in Tables 1, 2, A1, and A2 should be rechecked.
  4. [Sec. 4.2, Table 1] The abstract says CaPT consistently improves all baselines, but the paper admits failure cases (OxfordPets), and Table 1 contains per-dataset drops (e.g., KgCoOp ImageNet H 73.25 to 72.95; DePT OxfordPets H 95.78 to 95.68). Because no standard deviations are reported despite three runs, the 2.6% average gain cannot be assessed for significance. Please report per-run results or confidence intervals and qualify consistent as on average.
  5. [Eq. (16)] The fast adaptation to new classes relies on the assumption that the Meta-Net trained on base-class TCI produces useful offsets for unseen class-name embeddings. This is not guaranteed by the training objective; if new-class TCI lies outside the base-class embedding manifold, the prompt offset becomes an uninformative perturbation. The paper does not characterize this. Please report the distribution of cosine similarities between base and new TCI, or show the per-dataset relationship between H gain and TCI distance, and discuss failure cases like OxfordPets in that light.
minor comments (7)
  1. [Sec. 3.2, Eq. (9)] The equation writes v_l(µ_base) = v_l + h_α(µ_base) but should refer to p_i^base, not µ_base; this is a notation typo that obscures the definition.
  2. [Sec. 3.3, Eq. (13)] The text contains the typo 'µbaase' and inconsistently uses v_l versus v for the context vectors; please standardize the notation.
  3. [Related Work] The word 'methdos' appears in the first paragraph of Section 2; it should be 'methods'.
  4. [Sec. 3.3] The phrase 'an margin-adjusted' should be 'a margin-adjusted'.
  5. [Table 6] All rows in the '+CaPT' part of Table 6 are labeled '+DePT', which appears to be a copy-paste error from the DePT paper; the labels should read '+CaPT'.
  6. [Sec. 4.3.1, Impact of the Scaling Weight] The sentence 'DePT establishes the best results ... when ϵ=0.31' should refer to CaPT, not DePT, since this paragraph is ablating CaPT.
  7. [References] References [7] and [45] are duplicate entries for the same paper (Zhou et al., IJCV 2022); please consolidate.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: CaPT's effectiveness is an empirical, externally benchmarked result; the TCI-key claim is an ablative observation, not a definitional consequence.

full rationale

The derivation chain is not circular. Section 3.2 defines four CoCoOp variants by explicitly replacing the conditioning input (Eqs. 5, 7, 9, 11); the conclusion that TCI conditioning outperforms VII conditioning is an empirical comparison, not an identity. CaPT (Eqs. 12-16) defines TCI-conditioned prompts and a margin-ITM loss, and the reported gains are measured against published baselines (e.g., Bayes in Table 2) rather than being forced by construction. No fitted parameter is renamed as a prediction: the choice ε=0.31 comes from a sweep on the same 11 datasets (Fig. 5), which is a test-set selection concern, but it does not make the reported H ACC equal to the input by construction. The DePT baseline is a self-citation, yet CaPT's plugin claim is also demonstrated on four external methods (CoOp, KgCoOp, ProGrad, DAPT), and DeCaPT is a combination, not a restatement of inputs. The noted confound in Fig. 2(d), where TCI versus VCI changes both the condition modality and the class-level granularity simultaneously, is an experimental isolation gap rather than a definitional reduction. The impossible ProDA Food101 H entry and the mislabeled '+DePT' rows in Table 6 are reporting errors that affect reliability, not circularity. Under the hard rule requiring a quoted equation-level reduction or a fitted-parameter-as-prediction, no qualifying circular step exists.

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

No physical or ontological entities are introduced; Meta-Net is a conventional two-layer MLP. The central claim rests on assumptions about CLIP's frozen text-image alignment, the informativeness of class names for unseen classes, the use of text cosine similarity as a proxy for visual class similarity in the margin loss, and the fairness of the 10-epoch protocol. The only numerically fitted hyperparameter is the margin scaling epsilon, selected on the same 11 datasets used for the headline results.

free parameters (4)
  • margin-ITM scaling weight epsilon = 0.31
    Chosen by sweeping {0.25,...,0.39} on the same 11 datasets used for the final H ACC (Fig. 5 left); reported gains partly include this selection.
  • Training epochs = 10
    Selected because H ACC saturates at epoch 10 on the same benchmark (Fig. 5 right); baselines are also retrained to 10 epochs, so comparisons are fair, but the value is benchmark-tuned.
  • Meta-Net bottleneck dimension = 32
    Fixed across datasets, but the architecture is chosen after an ablation on the same 11 datasets (Table 5); not reported as tuned, so there is a small risk of selection on the test benchmark.
  • TCI extractor choice = CLIP text encoder (called Word2Vec in the text)
    Selected from Table 4 ablation on the same 11 datasets; the CLIP text encoder gives the highest H and is then used as the TCI extractor.
assumptions (4)
  • domain assumption CLIP's frozen text and image encoders provide a shared feature space in which class-name conditioning can transfer to unseen classes.
    Used throughout Sec. 3.1-3.3; if class-name embeddings are not aligned with visual class structure, TCI conditioning cannot help.
  • domain assumption Cosine similarity between TCI embeddings of base classes is a reliable proxy for visual inter-class similarity in the margin-ITM loss.
    The margin loss in Eq. 17 uses suv = epsilon * cosine similarity of class-name embeddings; if text similarity does not track visual similarity, the margin adjustment can distort embeddings.
  • domain assumption New class names are available at inference and are informative enough to generate useful prompts.
    Eq. 16 builds new-class prompts from class-name TCI only; if new class names are semantically uninformative, the conditioning signal is noise.
  • ad hoc to paper Ten training epochs is a fair and sufficient schedule for all compared methods.
    Introduced in Sec. 4.1 to equalize baselines, but it deviates from some original papers' schedules, which affects the SOTA comparison in Table 2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Closer Look at Conditional Prompt Tuning for Vision-Language Models." pith.science (2026). https://pith.science/paper/YECUCIEQ

@misc{pith2026250623856,
  author       = {Pith},
  title        = {Pith review of: A Closer Look at Conditional Prompt Tuning for Vision-Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YECUCIEQ}},
  note         = {Machine review of arXiv:2506.23856}
}
read the original abstract

Despite the great promise of Prompt Tuning (PT) in adapting large Vision-Language Pretrained Models (VLPMs) to downstream tasks, they often struggle to overcome the Base-New Tradeoff (BNT) dilemma: as VLPMs are better tuned to a base task, their ability to generalize to new tasks diminishes. Recent work on conditional PT addresses this problem by replacing static prompts with dynamic Visual Image Information (VII)-conditioned prompts, improving the model's generalization to new tasks to some extent. In this work, we first identify a critical issue with existing conditional PT methods: using VII as the "condition" of prompts yields suboptimal performance, and even random noise-conditioned prompts can outperform the VII-conditioned counterparts. On further analysis, we find that learning dynamic prompts conditioned on Textual Class Information (TCI) is the key to solving the BNT problem. Motivated by this, we then propose Class-adaptive Prompt Tuning (CaPT), which enables fast adaptation of tuned models to new classes by learning TCI-conditioned prompts from base classes. Remarkably, CaPT can be used as a plugin to mitigate the BNT problem for existing unconditional PT schemes. Extensive experiments on 11 datasets show that CaPT consistently improves the performance of five strong unconditional PT baselines with negligible additional computational cost. Additionally, by integrating CaPT with our recently proposed DePT framework, we devise a new conditional PT approach, termed DeCaPT, which outperforms the H ACC of the state-of-the-art conditional PT scheme by 3.49%, averaged over the 11 datasets. Code: https://github.com/Koorye/CaPT.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

67 extracted references · 67 canonical work pages

  1. [1]

    Dept: Decoupled prompt tuning

    Zhang J, Wu S, Gao L, Shen H, Song J. Dept: Decoupled prompt tuning. CVPR. 2024

    work page 2024

  2. [2]

    Learning trans- ferable visual models from natural language supervision

    Radford A, Kim JW, Hallacy C, Ramesh A, Goh G, Agarwal S, et al. Learning trans- ferable visual models from natural language supervision. In: ICML; 2021. p. 8748–8763

    work page 2021

  3. [3]

    A closer look at few-shot classification again

    Luo X, Wu H, Zhang J, Gao L, Xu J, Song J. A closer look at few-shot classification again. In: ICML. PMLR; 2023. p. 23103–23123

    work page 2023

  4. [4]

    Bayesian prompt learning for image- language model generalization

    Derakhshani MM, Sanchez E, Bulat A, da Costa VGT, Snoek CG, Tzimiropoulos G, et al. Bayesian prompt learning for image- language model generalization. In: CVPR

    work page

  5. [5]

    Maple: Multi-modal prompt learning

    Khattak MU, Rasheed H, Maaz M, Khan S, Khan FS. Maple: Multi-modal prompt learning. In: CVPR; 2023. p. 19113–19122

    work page 2023

  6. [6]

    Conditional prompt learning for vision-language models

    Zhou K, Yang J, Loy CC, et al. Conditional prompt learning for vision-language models. In: CVPR; 2022. p. 16816–16825

    work page 2022

  7. [8]

    Visual-language prompt tuning with knowledge-guided con- text optimization

    Yao H, Zhang R, Xu C. Visual-language prompt tuning with knowledge-guided con- text optimization. In: CVPR; 2023. p. 6757– 6767

    work page 2023

  8. [9]

    Prompt-aligned gradient for prompt tuning

    Zhu B, Niu Y, Han Y, Wu Y, Zhang H. Prompt-aligned gradient for prompt tuning. ICCV. 2023

    work page 2023

Show all 67 references
  1. [10]

    Distribution-aware prompt tuning for vision-language models

    Cho E, Kim J, Kim HJ. Distribution-aware prompt tuning for vision-language models. In: ICCV; 2023. p. 22004–22013

    2023

  2. [11]

    Context-aware Alignment and Mutual Mask- ing for 3D-Language Pre-training

    Jin Z, Hayat M, Yang Y, Guo Y, Lei Y. Context-aware Alignment and Mutual Mask- ing for 3D-Language Pre-training. In: CVPR

  3. [12]

    Recent advances in natural language processing via large pre- trained language models: A survey

    Min B, Ross H, Sulem E, Veyseh APB, Nguyen TH, Sainz O, et al. Recent advances in natural language processing via large pre- trained language models: A survey. ACM Computing Surveys. 2023;56(2):1–40

    2023

  4. [13]

    Nat- ural language processing: State of the art, current trends and challenges

    Khurana D, Koli A, Khatter K, Singh S. Nat- ural language processing: State of the art, current trends and challenges. Multimedia tools and applications. 2023;82(3):3713–3744

    2023

  5. [14]

    Vilt: Vision-and- language transformer without convolution or region supervision

    Kim W, Son B, Kim I. Vilt: Vision-and- language transformer without convolution or region supervision. In: ICML. PMLR; 2021. p. 5583–5594

    2021

  6. [15]

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

    Jia C, Yang Y, Xia Y, Chen YT, Parekh Z, Pham H, et al. Scaling up visual and vision- language representation learning with noisy text supervision. In: ICML; 2021. p. 4904– 4916

    2021

  7. [16]

    WenLan: Bridging vision and language by large-scale multi-modal pre- training

    Huo Y, Zhang M, Liu G, Lu H, Gao Y, Yang G, et al. WenLan: Bridging vision and language by large-scale multi-modal pre- training. arXiv preprint arXiv:210306561. 2021

    2021

  8. [17]

    Align before fuse: Vision and language representation learn- ing with momentum distillation

    Li J, Selvaraju R, Gotmare A, Joty S, Xiong C, Hoi SCH. Align before fuse: Vision and language representation learn- ing with momentum distillation. NeurIPS. 2021;34:9694–9705. 14

    2021

  9. [18]

    Disentangled Multiplex Graph Represen- tation Learning

    Mo Y, Lei Y, Shen J, Shi X, Shen HT, Zhu X. Disentangled Multiplex Graph Represen- tation Learning. In: ICML. vol. 202; 2023. p. 24983–25005

    2023

  10. [19]

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

    Lu J, Batra D, Parikh D, Lee S. Vilbert: Pretraining task-agnostic visiolinguistic rep- resentations for vision-and-language tasks. NeurIPS. 2019;32

    2019

  11. [20]

    X-clip: End-to-end multi-grained contrastive learning for video-text retrieval

    Ma Y, Xu G, Sun X, Yan M, Zhang J, Ji R. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In: ACM MM; 2022. p. 638–647

    2022

  12. [21]

    Complementarity-aware space learning for video-text retrieval

    Zhu J, Zeng P, Gao L, Li G, Liao D, Song J. Complementarity-aware space learning for video-text retrieval. IEEE Transactions on Circuits and Systems for Video Technology. 2023

    2023

  13. [22]

    Denseclip: Language- guided dense prediction with context-aware prompting

    Rao Y, Zhao W, Chen G, Tang Y, Zhu Z, Huang G, et al. Denseclip: Language- guided dense prediction with context-aware prompting. In: CVPR; 2022. p. 18082–18091

    2022

  14. [23]

    Extract free dense labels from clip

    Zhou C, Loy CC, Dai B. Extract free dense labels from clip. In: ECCV. Springer; 2022. p. 696–712

    2022

  15. [24]

    Clip-nerf: Text-and-image driven manipula- tion of neural radiance fields

    Wang C, Chai M, He M, Chen D, Liao J. Clip-nerf: Text-and-image driven manipula- tion of neural radiance fields. In: CVPR

  16. [25]

    Styleclip: Text-driven manipulation of stylegan imagery

    Patashnik O, Wu Z, Shechtman E, Cohen- Or D, Lischinski D. Styleclip: Text-driven manipulation of stylegan imagery. In: CVPR

  17. [26]

    Parameter-efficient transfer learning for NLP

    Houlsby N, Giurgiu A, Jastrzebski S, Mor- rone B, De Laroussilhe Q, Gesmundo A, et al. Parameter-efficient transfer learning for NLP. In: ICML; 2019. p. 2790–2799

    2019

  18. [27]

    Learning a universal tem- plate for few-shot dataset generalization

    Triantafillou E, Larochelle H, Zemel R, Dumoulin V. Learning a universal tem- plate for few-shot dataset generalization. In: ICML; 2021. p. 10424–10433

    2021

  19. [28]

    Reliable Few-shot Learning under Dual Noises

    Zhang J, Song J, Gao L, Sebe N, Shen HT. Reliable Few-shot Learning under Dual Noises. arXiv preprint arXiv:250616330. 2025

    2025

  20. [29]

    Prefix-Tuning: Optimiz- ing Continuous Prompts for Generation

    Li XL, Liang P. Prefix-Tuning: Optimiz- ing Continuous Prompts for Generation. In: ACL; 2021. p. 4582–4597

    2021

  21. [30]

    Skip Tuning: Pre-trained Vision- Language Models are Effective and Efficient Adapters Themselves

    Wu S, Zhang J, Zeng P, Gao L, Song J, Shen HT. Skip Tuning: Pre-trained Vision- Language Models are Effective and Efficient Adapters Themselves. In: CVPR; 2025. p. 14723–14732

    2025

  22. [31]

    Lora: Low-rank adapta- tion of large language models

    Hu EJ, Shen Y, Wallis P, Allen-Zhu Z, Li Y, Wang S, et al. Lora: Low-rank adapta- tion of large language models. arXiv preprint arXiv:210609685. 2021

    2021

  23. [32]

    Learning to decompose visual features with latent textual prompts

    Wang F, Li M, Lin X, Lv H, Schwing AG, Ji H. Learning to decompose visual features with latent textual prompts. ICLR. 2022

    2022

  24. [33]

    Prompt learning with optimal transport for vision-language models

    Chen G, Yao W, Song X, Li X, Rao Y, Zhang K. Prompt learning with optimal transport for vision-language models. ICLR. 2022

    2022

  25. [34]

    Consistent Prompt Tuning for Generalized Category Discovery

    Yang M, Yin J, Gu Y, Deng C, Zhang H, Zhu H. Consistent Prompt Tuning for Generalized Category Discovery. International Journal of Computer Vision. 2025;p. 1–28

    2025

  26. [35]

    Progressive visual prompt learn- ing with contrastive feature re-formation

    Xu C, Zhu Y, Shen H, Chen B, Liao Y, Chen X, et al. Progressive visual prompt learn- ing with contrastive feature re-formation. International Journal of Computer Vision. 2025;133(2):511–526

    2025

  27. [36]

    HybridPrompt: Domain-Aware Prompting for Cross-Domain Few-Shot Learning

    Wu J, Zhang T, Zhang Y. HybridPrompt: Domain-Aware Prompting for Cross-Domain Few-Shot Learning. International Journal of Computer Vision. 2024;132(12):5681–5697

    2024

  28. [37]

    DETA: Denoised Task Adaptation for Few- Shot Learning

    Zhang J, Gao L, Luo X, Shen H, Song J. DETA: Denoised Task Adaptation for Few- Shot Learning. CVPR. 2024

    2024

  29. [38]

    Meta-fdmixup: Cross- domain few-shot learning guided by labeled target data

    Fu Y, Fu Y, Jiang YG. Meta-fdmixup: Cross- domain few-shot learning guided by labeled target data. In: ACM MM; 2021. p. 5326– 5334. 15

    2021

  30. [39]

    StyleAdv: Meta Style Adversarial Training for Cross- Domain Few-Shot Learning

    Fu Y, Xie Y, Fu Y, Jiang YG. StyleAdv: Meta Style Adversarial Training for Cross- Domain Few-Shot Learning. In: CVPR; 2023. p. 24575–24584

    2023

  31. [40]

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

    Zhang R, Zhang W, Fang R, Gao P, Li K, Dai J, et al. Tip-adapter: Training-free adaption of clip for few-shot classification. In: ECCV. Springer; 2022. p. 493–510

    2022

  32. [41]

    Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners

    Zhang R, Hu X, Li B, Huang S, Deng H, Qiao Y, et al. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. In: CVPR; 2023. p. 15211– 15222

    2023

  33. [42]

    Visual prompt tuning

    Jia M, Tang L, Chen BC, Cardie C, Belongie S, Hariharan B, et al. Visual prompt tuning. In: ECCV; 2022. p. 709–727

    2022

  34. [43]

    Diversity-Aware Meta Visual Prompting

    Huang Q, Dong X, Chen D, Zhang W, Wang F, Hua G, et al. Diversity-Aware Meta Visual Prompting. In: CVPR; 2023. p. 10878–10887

    2023

  35. [44]

    Self-regulating Prompts: Foundational Model Adaptation without For- getting

    Khattak MU, Wasim ST, Naseer M, Khan S, Yang MH, Khan FS. Self-regulating Prompts: Foundational Model Adaptation without For- getting. In: ICCV; 2023. p. 15190–15200

    2023

  36. [45]

    Learn- ing to prompt for vision-language models

    Zhou K, Yang J, Loy CC, Liu Z. Learn- ing to prompt for vision-language models. International Journal of Computer Vision. 2022;130(9):2337–2348

    2022

  37. [46]

    FastText.zip: Compress- ing text classification models

    Joulin A, Grave E, Bojanowski P, Douze M, J´ egou H, Mikolov T. FastText.zip: Compress- ing text classification models. arXiv preprint arXiv:161203651. 2016

    2016

  38. [47]

    Wikipedia2Vec: An Efficient Toolkit for Learning and Visual- izing the Embeddings of Words and Entities from Wikipedia

    Yamada I, Asai A, Sakuma J, Shindo H, Takeda H, Takefuji Y, et al. Wikipedia2Vec: An Efficient Toolkit for Learning and Visual- izing the Embeddings of Words and Entities from Wikipedia. In: EMNLP. ACL; 2020. p. 23–30

    2020

  39. [48]

    Glove: Global vectors for word representation

    Pennington J, Socher R, Manning CD. Glove: Global vectors for word representation. In: EMNLP; 2014. p. 1532–1543

    2014

  40. [49]

    Clip-adapter: Bet- ter vision-language models with feature adapters

    Gao P, Geng S, Zhang R, Ma T, Fang R, Zhang Y, et al. Clip-adapter: Bet- ter vision-language models with feature adapters. International Journal of Computer Vision. 2024;132(2):581–595

    2024

  41. [50]

    Imagenet: A large-scale hierarchical image database

    Deng J, Dong W, Socher R, Li LJ, Li K, Fei- Fei L. Imagenet: A large-scale hierarchical image database. In: CVPR. Ieee; 2009. p. 248–255

    2009

  42. [51]

    Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories

    Fei-Fei L, Fergus R, Perona P. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In: CVPR W. IEEE; 2004. p. 178–178

    2004

  43. [52]

    Cats and dogs

    Parkhi OM, Vedaldi A, Zisserman A, Jawa- har C. Cats and dogs. In: CVPR. IEEE; 2012. p. 3498–3505

    2012

  44. [53]

    3d object representations for fine-grained cate- gorization

    Krause J, Stark M, Deng J, Fei-Fei L. 3d object representations for fine-grained cate- gorization. In: ICCVW; 2013. p. 554–561

    2013

  45. [54]

    Automated flower classification over a large number of classes

    Nilsback ME, Zisserman A. Automated flower classification over a large number of classes. In: ICVGIP. IEEE; 2008. p. 722–729

    2008

  46. [55]

    Food- 101–mining discriminative components with random forests

    Bossard L, Guillaumin M, Van Gool L. Food- 101–mining discriminative components with random forests. In: ECCV. Springer; 2014. p. 446–461

    2014

  47. [56]

    Fine-grained visual classification of aircraft

    Maji S, Rahtu E, Kannala J, Blaschko M, Vedaldi A. Fine-grained visual classification of aircraft. arXiv preprint arXiv:13065151. 2013

    2013

  48. [57]

    Eurosat: A novel dataset and deep learning benchmark for land use and land cover clas- sification

    Helber P, Bischke B, Dengel A, Borth D. Eurosat: A novel dataset and deep learning benchmark for land use and land cover clas- sification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing. 2019;12(7):2217–2226

    2019

  49. [58]

    UCF101: A dataset of 101 human actions classes from videos in the wild

    Soomro K, Zamir AR, Shah M. UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:12120402. 2012;. 16

    2012

  50. [59]

    Describing textures in the wild

    Cimpoi M, Maji S, Kokkinos I, Mohamed S, Vedaldi A. Describing textures in the wild. In: CVPR; 2014. p. 3606–3613

    2014

  51. [60]

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

    Xiao J, Hays J, Ehinger KA, Oliva A, Tor- ralba A. Sun database: Large-scale scene recognition from abbey to zoo. In: CVPR. IEEE; 2010. p. 3485–3492

    2010

  52. [61]

    Do imagenet classifiers generalize to ima- genet? In: ICML

    Recht B, Roelofs R, Schmidt L, Shankar V. Do imagenet classifiers generalize to ima- genet? In: ICML. PMLR; 2019. p. 5389–5400

    2019

  53. [62]

    Learning robust global representations by penalizing local predictive power

    Wang H, Ge S, Lipton Z, Xing EP. Learning robust global representations by penalizing local predictive power. NeurIPS. 2019;32

    2019

  54. [63]

    Generating natural adversarial examples with universal perturbations for text classifi- cation

    Gao H, Zhang H, Yang X, Li W, Gao F, Wen Q. Generating natural adversarial examples with universal perturbations for text classifi- cation. Neurocomputing. 2022;471:175–182

    2022

  55. [64]

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

    Hendrycks D, Basart S, Mu N, Kadavath S, Wang F, Dorundo E, et al. The many faces of robustness: A critical analysis of out-of- distribution generalization. In: ICCV; 2021. p. 8340–8349

    2021

  56. [65]

    Self-regulating prompts: Foundational model adaptation without for- getting

    Khattak MU, Wasim ST, Naseer M, Khan S, Yang MH, Khan FS. Self-regulating prompts: Foundational model adaptation without for- getting. In: ICCV; 2023. p. 15190–15200

    2023

  57. [66]

    Prompt distribution learning

    Lu Y, Liu J, Zhang Y, Liu Y, Tian X. Prompt distribution learning. In: CVPR; 2022. p. 5206–5215

    2022

  58. [67]

    Black box few-shot adaptation for vision- language models

    Ouali Y, Bulat A, Matinez B, Tzimiropoulos G. Black box few-shot adaptation for vision- language models. In: ICCV; 2023. p. 15534– 15546

    2023

  59. [68]

    Read-only prompt optimization for vision- language few-shot learning

    Lee D, Song S, Suh J, Choi J, Lee S, Kim HJ. Read-only prompt optimization for vision- language few-shot learning. In: ICCV; 2023. p. 1401–1411. Appendix A Additional Results Cross-Dataset Generalization. The cross- dataset generalization setting assesses whether models traine...

    2023

Pith tools

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