Pith. sign in

REVIEW 4 major objections 4 minor 26 references

Attributes Should Come from Images, Not Class Names: Distribution-Conditioned Attribute Selection for Vision-Language Models

T0 review · 4 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read LLM descriptors for image classes lean on the class name, not on visual evidence; selecting attributes from images fixes the flaw.

desk verdict A useful diagnostic and a solid few-shot method, but the strong causal claim needs an anchor-prompt control before it fully lands. read the letter →

arxiv 2607.18695 v1 pith:K2JE7TLP submitted 2026-07-21 cs.CV cs.AIcs.LGeess.IV

Gautam Rajendrakumar Gare , Jia Shi , Zhiqiu Lin , Deepak Pathak , John Galeotti , Deva Ramanan This is my paper
classification cs.CVcs.AIcs.LGeess.IV
keywords vision-languagemodelsattributeselectionzero-shotclassificationinterpretabilitydistributionshiftpromptensemblingfew-shotlearningclass-nameconfound
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper shows that descriptor-based prompting for vision-language models—asking a language model to list attributes of each class—rides entirely on the class name: remove it and ImageNet accuracy collapses from 59.5% to 15.5%. The authors argue that LLM descriptors are conditioned on the label, not on any image, so they describe the concept in general and mislead precisely when the data shifts. Their fix is to select attributes from the target image collection itself, scoring a large candidate pool in the model's joint embedding space and keeping the top attributes per class. This restores meaningful attribute-only accuracy, yields readable per-class word lists, and extends to few-shot classification, shift description, and localization.

What carries the argument

Distribution-conditioned attribute selection: a frozen CLIP encodes target images and a large attribute pool into a joint embedding space; scaled cosine similarities form an attribute feature vector per image; per-class attribute weights come from the mean feature or a linear probe; and the top-k attributes per class define the class's prompt set. The pool remains fixed and only the ranking changes, so the selected attributes follow the target distribution instead of the general concept captured by the class name.

What would settle it

A direct test: take shifted images such as ImageNet-Sketch strawberries, prompt an LLM for descriptors using only the class name, and check whether a CLIP-based attribute detector scores those descriptors (e.g., 'red', 'ripe') as present in the shifted images. If they do score high, the class-name confound is not the whole explanation. Alternatively, if the matched-pool selection gain disappears when the linear probe is replaced by a label-free scoring method, the selection mechanism would be partly an artifact of the probe.

Watch

Extended reading notes

Core claim

The paper establishes the class-name confound: LLM-generated descriptors carry little visual evidence on their own, so their apparent gains cannot be read as attribute-level evidence about the image. The proposed remedy—distribution-conditioned attribute selection—scores a fixed pool of attribute strings against the target images with a frozen vision-language model and keeps the top-scoring attributes per class. Holding the pool identical, image-conditioned selection lifts class-name-free accuracy from 15.5% to 19.4% on ImageNet; a larger pool reaches 23.8% with five attributes per class and 45.5% with the full weighted pool. The selected attributes beat gradient-based prompt tuning at one i

Load-bearing premise

The load-bearing assumption is that prompting CLIP with a bare attribute string faithfully measures whether that attribute is visually present in the image; if bare words are intrinsically hard for CLIP to ground, part of the 44-point collapse reflects measurement failure rather than absence of visual evidence.

Editorial extensions

If this is right

  • Any work claiming attribute-level interpretability from descriptor prompts should report class-name-free accuracy, since the standard protocol cannot separate attribute evidence from prompt ensembling.
  • Image-conditioned attribute selection is a strong extreme few-shot learner, beating prompt tuning by 3 points at one image per class while being a readable list of words.
  • Mean attribute profiles of datasets describe distribution shift in words, and the top rising attributes double as domain-specific prompt templates.
  • The mechanism extends a vision-language model to class vocabularies its text encoder cannot parse, and localizes objects without using class names in the query.

Reading between the lines

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

  • The 44-point collapse might partly reflect that CLIP struggles to ground bare attribute words without a class anchor; the matched-pool reselection partially rules this out, but that comparison uses labeled training images and a linear probe, so the diagnosis is not fully isolated.
  • If attribute selection is genuinely data-driven, it could enable test-time intervention—editing a classifier by deleting or reweighting words—and continuous monitoring of domain drift in deployed systems.
  • The accuracy plateau beyond four shots suggests the limiting factor is the fixed attribute vocabulary; a universal, expanded pool could extend the method's reach into the medium-shot regime.
  • The same selection mechanism could be applied to concept-bottleneck pipelines, choosing concept pools against the deployment distribution rather than generating them from labels.
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 / 4 minor

Summary. The paper argues that LLM-generated attribute descriptors used in VLM prompting (e.g., Menon & Vondrick) are conditioned on the class name rather than on image content. The central evidence is that removing the class name from the prompt collapses ImageNet accuracy from 59.5% to 15.5% (Table 1). The authors then propose a method for selecting attributes from the target image collection using a frozen CLIP, scoring a large fixed attribute pool against images and keeping the top-k attributes per class. This image-conditioned selection is shown to improve class-name-free accuracy, especially with a larger pool (23.8% top-5, 45.5% uncapped), to transfer across distribution shifts, to be competitive in extreme few-shot settings (1–4 shots), and to enable applications such as describing distribution shift in words, handling unparseable vocabularies, and attribute-guided localization. A matched-pool experiment (selecting from the LLM's own descriptor pool) attributes the gain to the selection mechanism rather than the vocabulary.

Significance. If the central diagnostic is correct, the paper makes an important and field-relevant point: descriptor-prompting methods that claim attribute-level interpretability may be largely riding the class name, and their output should not be read as image-grounded evidence. The proposed distribution-conditioned attribute selection is a simple, training-light, and interpretable alternative that shows practical value in extreme few-shot classification and in generating readable dataset summaries. The paper includes an honest limitations section and a matched-pool comparison that is a strong idea. However, the headline diagnostic is currently supported by an incomplete control: the bare-prompt protocol conflates removal of the class name with removal of syntactic context, and the matched-pool reselection does not fully isolate the class-name effect. With additional controls, the paper's claims could be made rigorous.

major comments (4)
  1. [Section 3, Table 1 (bottom block)] The 44-point collapse (59.5% to 15.5%) is the load-bearing evidence for the claim that LLM descriptors carry little visual evidence. But the comparison changes two variables at once: the class name is removed and the prompt is reduced from '{class name}, which {has/is} {descriptor}' to a bare descriptor string. A missing control is a class-agnostic anchor shared across all classes, e.g., 'object which {has/is} {descriptor}' or 'a photo of {descriptor}'. If such an anchor restores most of the gap, then the descriptors themselves do encode attribute-level evidence, and the headline claim would be overstated. The matched-pool reselection in Sec. 5.2 does not settle this: it changes the attribute strings and adds a linear probe, so it does not directly test whether bareness is the cause of the collapse for the original descriptors. Please add this control and adjust the claim accordingly.
  2. [Section 5.2, matched-pool reselection (Table 1 bottom block)] The 3.9-point gain from reselecting the LLM pool is attributed to 'the selection mechanism' rather than to the probe or the added labeled data. However, the comparison is not apples-to-apples: the LLM descriptor baseline uses all descriptors with no training, while the image-selected pipeline uses a linear probe trained on labeled ImageNet attribute features and then selects only the top-5 per class. The probe could exploit label information more effectively than LLM class-name conditioning, which would make the gain partly attributable to the probe rather than to image-conditioning alone. To strengthen the causal claim, please also report attribute-only accuracy using the mean-feature scoring variant (no training, as described in Sec. 4) for the same pool, or otherwise ablate the contribution of the linear probe.
  3. [Section 6.1, Table 5] The domain-attribute templates are computed from unlabeled images of the target shifted dataset and then evaluated on that same dataset. This is a transductive setup: the method uses the test distribution to construct prompts, which can inflate accuracy. The paper states 'no class-specific information leaks,' but distribution-level information does leak. Please state explicitly that this is a transductive adaptation setting, and either evaluate on a held-out split of each shifted dataset or discuss the implications for the claimed gains. This is important because Table 5 is used to claim complementarity between domain-level and class-level attributes.
  4. [General reproducibility] No error bars or repeated-run statistics are reported for any of the core tables. Several conclusions rest on small margins (e.g., Table 4: 60.13 vs 57.15 at 1 shot; Table 1: 59.53 vs 59.47 for top-5 with class name). The linear probe training and any randomness in attribute preprocessing could produce variance. Please report standard deviations over at least a few seeds, or state if the results are deterministic.
minor comments (4)
  1. [Section 5.5, Table 4] The CoOp and WiSE-FT baseline numbers are presented without specifying the exact setup (e.g., official code, hyperparameters, number of runs). Please clarify whether these numbers are reproduced or taken from prior work, and include the standard protocol details.
  2. [Section 6.2] The color-variant fruit experiment does not specify how the attribute prompts for the fruit dataset are selected (e.g., from the 100 training images per class?). Please clarify the selection setup so that the comparison with zero-shot CLIP is reproducible.
  3. [Throughout] The terms 'attribute' and 'descriptor' are sometimes used interchangeably (e.g., 'LLM descriptors' vs 'selected attributes'). While understandable, a brief clarification of the distinction (generated vs. selected) would improve readability.
  4. [Section 4, Eq. (2)] The notation for the weighted variant ($\omega_{c,j} = w_{c,j}$) appears only later in Sec. 5.3; consider defining both variants at first use to avoid confusion.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the central class-name-confound claim is tested by a matched-pool control on an external descriptor pool; self-cited attribute pools and one transductive template experiment are caveats, not circular reductions.

full rationale

The paper's central diagnostic (Sec. 3) is an empirical intervention, not a definitional identity: it compares the standard prompt "{class name}, which {has/is} {descriptor}" with the bare "{descriptor}" prompt and reports a 59.5→15.5 collapse. This is falsifiable, and the conclusion that gains 'ride the class name' rests on an explicit operational assumption (attribute evidence should survive removal of the class name). That assumption could be wrong—CLIP may embed bare multi-word fragments poorly—so this is a validity/control concern, not circularity. The key non-circular control is the matched-pool experiment in Table 1 (bottom block) and Sec. 5.2: reselecting from the identical Menon & Vondrick GPT-3 pool, with the same frozen CLIP and same bare evaluation protocol, lifts accuracy from 15.5% to 19.4%. Since the pool, model, template, and test sets are held fixed and only the selection mechanism changes, the gain is not constructed by reusing the fitted quantity. Attributes are selected on ImageNet training images only and transferred unchanged to the four shifted variants, so no test labels or test images enter the selection loop; the shared CLIP embedding space is the model under test, not a fitted parameter. The few-shot comparison with CoOp and the linear probe are trained on the same labeled shots and evaluated on held-out ImageNet, which is standard and non-circular. Minor self-citation exists: the VAW [15] and LSA [16] attribute pools are co-authored by Zhiqiu Lin, and Table 1's larger-pool gains (23.8%) and Table 2's 45.5% rely on them. However, the central mechanism is established on the external GPT-3 pool, so the self-citation is not load-bearing. One additional caveat is Sec. 6.1/Table 5, where top-rising domain attributes are computed from unlabeled images of the target distribution and evaluated on that same distribution; this is transductive and can overstate template gains, but it is not a circular reduction and is not the paper's central claim. Overall, no step reduces by construction to its own input.

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

No new physical entities or latent mechanisms are introduced; the 'attribute feature space' is a re-expression of CLIP embeddings, not a new entity. The main free choices are the attribute budget and the domain-template count, neither fitted to the target benchmark.

free parameters (2)
  • top-k attribute budget k = 5, 10, 100, full pool
    Chosen by hand; headline attribute-only numbers use top-5 (23.8%), while uncapped reaches 45.5%; accuracy is monotone in k. The choice affects the reported headline result.
  • domain template attribute count = 5 rising attributes per dataset
    Chosen by hand for Table 5; the number of attributes ensembled in the domain prompt template is not derived from data.
assumptions (5)
  • domain assumption CLIP cosine similarity in the joint embedding space is a valid proxy for attribute presence in an image
    Used throughout the scoring in Eq. (1) and for classification in Eq. (2); if false, both selection and evaluation fail.
  • domain assumption Bare attribute strings, without a class name or template, faithfully probe attribute evidence
    The Section 3 diagnosis interprets the 59.5-to-15.5 collapse as evidence that descriptors carry little visual evidence; this assumes the drop is not an artifact of prompt format.
  • domain assumption The VAW+LSA+LLM attribute pool covers relevant visual attributes for the evaluated classes
    Used as the candidate vocabulary; the paper itself shows pool choice matters (random strings trail by 5.7 points, Appendix Table 7).
  • domain assumption ImageNet labels used for attribute selection are correct and representative
    All selection uses ImageNet training labels; incorrect labels would corrupt class-attribute weights.
  • domain assumption Linear probe trained on attribute features of a few images per class yields rankings that transfer to the test distribution
    Drives the few-shot results in Table 4; no theoretical guarantee is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Attributes Should Come from Images, Not Class Names: Distribution-Conditioned Attribute Selection for Vision-Language Models." pith.science (2026). https://pith.science/paper/K2JE7TLP

@misc{pith2026260718695,
  author       = {Pith},
  title        = {Pith review of: Attributes Should Come from Images, Not Class Names: Distribution-Conditioned Attribute Selection for Vision-Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K2JE7TLP}},
  note         = {Machine review of arXiv:2607.18695}
}
read the original abstract

A popular route to interpretable zero-shot classification asks a large language model (LLM) to describe each class name and prompts CLIP with the resulting descriptors. We show that these descriptors carry little visual evidence of their own: removing the class name from the prompt collapses ImageNet accuracy from 59.5% to 15.5%. The diagnosis is that the descriptors are conditioned on the label rather than on the images, so they describe the concept in general and mislead exactly when the data shifts; an LLM insists that strawberries are red, but every strawberry in ImageNet-Sketch is a colorless line drawing. We therefore select attributes from the target image collection instead: we score a large attribute pool against the images in CLIP's joint embedding space and keep the top-scoring attributes per class. Selected this way, class-name-free attribute prompts reach 23.8% on ImageNet (against 15.5% for LLM descriptors), the gain holds on four shifted ImageNet variants, and reselecting from the LLM's own pool isolates the selection mechanism as the cause. With one image per class, the selected attributes outperform the prompt-tuning method CoOp by 3 points while fitting in under a minute instead of 14 hours, with no learned soft prompt to obscure the decision. Because the attribute set is chosen by the data, it doubles as a readable summary of a dataset, which we use to describe distribution shift in words.

Figures

Figures reproduced from arXiv: 2607.18695 by the authors.

Figure 1
Figure 1. Class-name-conditioned descriptors fail under distribution shift; [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Distribution-conditioned attribute selection. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Attribute features trade statistical independence for named, seman [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Distribution shift, described in words. Top-10 rising and falling attributes (change in mean attribute score over unlabeled images, y-axis) of ImageNet-Sketch (left panel) and ImageNet-R (right panel) relative to ImageNet. Sketch gains “colorless”, “gray”, and “cartoon…
Figure 5
Figure 5. Figure 5: Examples from the collected color-variant fruit test set. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: CLIP zero-shot predictions latch onto color. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Class-name-free localization from selected attributes. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 1 canonical work pages

  1. [1]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    Akata,Z.,Perronnin,F.,Harchaoui,Z.,Schmid,C.:Label-embeddingforattribute- based classification. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 819–826 (2013) 4

    2013

  2. [2]

    Advances in Neural In- formation Processing Systems37(2 2024).https://doi.org/10.52202/079017- 2678,https://arxiv.org/pdf/2402.103764

    Bhalla, U., Oesterling, A., Srinivas, S., Calmon, F.P., Lakkaraju, H.: Interpreting CLIP with Sparse Linear Concept Embeddings (SpLiCE). Advances in Neural In- formation Processing Systems37(2 2024).https://doi.org/10.52202/079017- 2678,https://arxiv.org/pdf/2402.103764

    arXiv 2024

  3. [3]

    In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP)

    Chen, C., Zhang, B., Cao, L., Shen, J., Gunter, T., Jose, A., Toshev, A., Zheng, Y., Shlens, J., Pang, R., Yang, Y.: STAIR: Learning sparse text and image repre- sentation in grounded tokens. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP). pp. 15079–15094 (2023) 5

    2023

  4. [4]

    In: Inter- national Conference on Learning Representations (ICLR) (2023) 10

    Dunlap, L., Mohri, C., Guillory, D., Zhang, H., Darrell, T., Gonzalez, J.E., Raghu- nathan, A., Rohrbach, A.: Using language to extend to unseen domains. In: Inter- national Conference on Learning Representations (ICLR) (2023) 10

    2023

  5. [5]

    In: 2009 IEEE Conference on Computer Vision and Pattern Recognition

    Farhadi, A., Endres, I., Hoiem, D., Forsyth, D.: Describing objects by their at- tributes. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 1778–1785. IEEE (2009) 4

    2009

  6. [6]

    He,K.,Zhang,X.,Ren,S.,Sun,J.:Deepresiduallearningforimagerecognition.In: Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition. vol. 2016-December, pp. 770–778. IEEE Computer Society (12 2016).https://doi.org/10.1109/CVPR.2016.90,http://image-net.org/ challenges/LSVRC/2015/6

    doi:10.1109/cvpr.2016.90 2016

  7. [7]

    In: Proceed- ings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., Song, D., Steinhardt, J., Gilmer, J.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceed- ings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 8340–8349 (2021) 6, 9

    2021

  8. [8]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Hendrycks, D., Zhao, K., Basart, S., Steinhardt, J., Song, D.: Natural adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 15262–15271 (2021) 6

    2021

Show all 26 references
  1. [9]

    Advances in Neural Information Processing Systems4(January), 3464–3472 (9 2014),https://arxiv.org/pdf/1409.43274

    Jayaraman, D., Grauman, K.: Zero Shot Recognition with Unreliable Attributes. Advances in Neural Information Processing Systems4(January), 3464–3472 (9 2014),https://arxiv.org/pdf/1409.43274

    2014

  2. [10]

    In: Proceedings of the 37th International Conference on Machine Learning (ICML)

    Koh, P.W., Nguyen, T., Tang, Y.S., Mussmann, S., Pierson, E., Kim, B., Liang, P.: Concept bottleneck models. In: Proceedings of the 37th International Conference on Machine Learning (ICML). vol. 119. PMLR (2020) 4, 19

    2020

  3. [11]

    In: 2009 IEEE Conference on Com- puter Vision and Pattern Recognition

    Lampert, C.H., Nickisch, H., Harmeling, S.: Learning to detect unseen object classes by between-class attribute transfer. In: 2009 IEEE Conference on Com- puter Vision and Pattern Recognition. pp. 951–958. IEEE (2009) 4

    2009

  4. [12]

    In: Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2) (2021) 10

    Lin, Z., Shi, J., Pathak, D., Ramanan, D.: The CLEAR benchmark: Continual learning on real-world imagery. In: Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2) (2021) 10

    2021

  5. [13]

    In: International Conference on Learning Representations (ICLR) (2023) 1, 2, 3, 4, 5, 6, 8

    Menon, S., Vondrick, C.: Visual classification via description from large language models. In: International Conference on Learning Representations (ICLR) (2023) 1, 2, 3, 4, 5, 6, 8

    2023

  6. [14]

    11th International Conference on Learning Representations, ICLR 2023 (4 2023),https://arxiv.org/pdf/2304.061294 16 Gare et

    Oikarinen, T., Das, S., Nguyen, L.M., Weng, T.W.: Label-Free Concept Bottleneck Models. 11th International Conference on Learning Representations, ICLR 2023 (4 2023),https://arxiv.org/pdf/2304.061294 16 Gare et. al

    2023 arXiv

  7. [15]

    Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition pp

    Pham, K., Kafle, K., Lin, Z., Ding, Z., Cohen, S., Tran, Q., Shrivastava, A.: Learning to Predict Visual Attributes in the Wild. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition pp. 13013–13023 (6 2021).https://doi.org/10.48550/arx...

    2021 doi arXiv

  8. [16]

    Pham, K., Kafle, K., Lin, Z., Ding, Z., Cohen, S., Tran, Q., Shrivastava, A.: Im- proving Closed and Open-Vocabulary Attribute Prediction Using Transformers. Lecture Notes in Computer Science (including subseries Lecture Notes in Ar- tificial Intelligence and Lecture Notes in ...

    2022 doi

  9. [18]

    In: Proceedings of the 38th International Conference on Machine Learning (ICML)

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transfer- able visual models from natural language supervision. In: Proceedings of the 38th International Conference on M...

    2021

  10. [19]

    Recht, B., Roelofs, R., Schmidt, L., Shankar, V.: Do ImageNet Classifiers General- ize to ImageNet? 36th International Conference on Machine Learning, ICML 2019 2019-June, 9413–9424 (2 2019),https://arxiv.org/pdf/1902.108116

    2019 arXiv

  11. [20]

    Proceedings of the IEEE International Conference on Computer Vision pp

    Roth, K., Kim, J.M., Sophia Koepke, A., Vinyals, O., Schmid, C., Akata, Z.: Waf- fling around for Performance: Visual Classification with Random Words and Broad Concepts. Proceedings of the IEEE International Conference on Computer Vision pp. 15700–15711 (6 2023).https://doi.o...

    2023

  12. [21]

    com/datasets/kritikseth/fruit- and- vegetable- image- recognition(2020), kaggle 10, 13

    Seth, K.: Fruits and vegetables image recognition dataset.https://www.kaggle. com/datasets/kritikseth/fruit- and- vegetable- image- recognition(2020), kaggle 10, 13

    2020

  13. [22]

    Shen, S., Li, C., Hu, X., Yang, J., Xie, Y., Zhang, P., Gan, Z., Wang, L., Yuan, L., Liu, C., Keutzer, K., Darrell, T., Rohrbach, A., Gao, J.: K-LITE: Learning TransferableVisualModelswithExternalKnowledge.AdvancesinNeuralInforma- tion Processing Systems35(4 2022).https://doi....

    2022 arXiv

  14. [23]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The inaturalist species classification and detection dataset. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8769–8778 (2018) 12

    2018

  15. [24]

    Advances in Neural Information Processing Systems32(5 2019),https://arxiv.org/pdf/1905.135496

    Wang, H., Ge, S., Xing, E.P., Lipton, Z.C.: Learning Robust Global Represen- tations by Penalizing Local Predictive Power. Advances in Neural Information Processing Systems32(5 2019),https://arxiv.org/pdf/1905.135496

    2019 arXiv

  16. [25]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Wortsman, M., Ilharco, G., Kim, J.W., Li, M., Kornblith, S., Roelofs, R., Lopes, R.G., Hajishirzi, H., Farhadi, A., Namkoong, H., Schmidt, L.: Robust fine-tuning of zero-shot models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). ...

    2022

  17. [26]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Yang, Y., Panagopoulou, A., Zhou, S., Jin, D., Callison-Burch, C., Yatskar, M.: Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)....

    2023

  18. [27]

    {class name}, which is/has/{...} {attribute}

    Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision130(9), 2337–2348 (2022) 4, 8, 10 18 Gare et. al. Supplementary Material This supplementary material provides the experimental details referenced from...

    2022

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.