Pith. sign in

REVIEW 4 major objections 5 minor 31 references

Beyond Accuracy: Metrics that Uncover What Makes a 'Good' Visual Descriptor

T0 review · 4 major / 5 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 claims that two label-free metrics—global alignment with a reference vision embedding and similarity to CLIP's training captions—can evaluate and rank text-based visual descriptor sets independently of classification accuracy.

desk verdict A clearly-written empirical study with one load-bearing bug in the Class Name condition that makes the headline ordering impossible under the stated algorithm. read the letter →

arxiv 2507.03542 v2 pith:L2NLX3UF submitted 2025-07-04 cs.CV

Ethan Lin , Linxi Zhao , Atharva Sehgal , Jennifer J. Sun This is my paper
classification cs.CV
keywords visualdescriptorsvision-languagemodelszero-shotclassificationrepresentationalalignmentCLIPembeddingspaceDINOv2descriptorevaluationmetricslabel-free
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 argues that classification accuracy is too blunt a yardstick for judging text-based visual descriptors, and proposes two label-free metrics to replace it. The first, which it calls DINO alignment, measures how well the similarity structure induced by a set of descriptors matches the image-similarity structure of a reference visual embedding, DINOv2-Base. The second, CLIP similarity, measures how well descriptors align with the captions and images in a large sample of CLIP's pretraining data. Across three benchmarks, the two metrics rank descriptor-generation methods in a consistent order—random token appends worst, iteratively refined descriptors better—even when accuracy differences are small, and the metrics improve as the ESCHER refinement loop runs. If the claims hold, descriptor evaluation and search no longer need labels or downstream accuracy as the primary signal.

What carries the argument

Two named metrics carry the argument. DINO_align is a mutual k-nearest-neighbor overlap score: images are projected into the descriptor space through CLIP's text embeddings, their k-nearest-neighbor graph is compared with the k-nearest-neighbor graph of DINOv2 image embeddings, and the average overlap is the alignment score; class names are removed from descriptors before computing it. CLIP_sim is the average, over top-matching captions retrieved in a five-million-caption sample of CLIP's pretraining data, of the text-to-image similarity of each matched caption-pair, with a companion frequency statistic counting captions above a similarity threshold. The first metric tests representational capacity; the second tests compatibility with the VLM's training distribution.

What would settle it

If swapping DINOv2-Base for a different self-supervised reference model changes the method ranking, or if a scrambled-token descriptor set scores as high as CBD on DINO_align, the metric is not measuring semantic quality.

Watch

Extended reading notes

Core claim

The core claim is that a good visual descriptor set is one whose induced representation space agrees with a strong self-supervised image embedding, and whose individual descriptors resemble visually grounded text in CLIP's training distribution. Rather than asking only whether a descriptor set classifies well, the paper measures, on CUB-200-2011, NABirds, and CIFAR-100, whether the similarity matrix produced by projecting images onto descriptor concepts correlates with DINOv2 image neighborhoods (global alignment) and whether each descriptor retrieves captions whose paired images are close in CLIP space (CLIP similarity). The authors report that after stripping class names from descriptors, alignment scores order the methods as DCLIP < WaffleCLIP < CBD ≤ ESCHER < Class Names on all datasets, that this order is stable regardless of accuracy, and that ESCHER iterations drive both alignment and accuracy upward on CIFAR-100 and CUB. They interpret the stable ordering as evidence that alignment captures semantic grounding that accuracy misses, and the negative frequency–similarity trend for CLIP similarity as evidence that rarer, more specific descriptors are more visually grounded.

Load-bearing premise

The rankings depend on trusting that DINOv2's notion of visual similarity is the right reference for what good descriptors should capture, and that a five-million-caption sample fairly represents CLIP's training distribution.

Editorial extensions

If this is right

  • Descriptor search can be guided by DINO_align and CLIP_sim without any labeled validation set, so new descriptor generation methods can be screened cheaply before evaluation.
  • A method like ESCHER can use alignment as an additional reward signal or early-stopping criterion, since both metrics track its refinement progress.
  • The finding that random-token descriptors score low on alignment despite near-comparable accuracy explains why accuracy alone can be misleading and motivates reporting alignment alongside accuracy.
  • Descriptor sets such as class names, which score high on alignment but inconsistently on CLIP_sim, can be diagnosed: they capture global structure but lack fine-grained visual grounding.
  • Because class names dominate alignment scores when left in, the paper's protocol of removing class names provides a standard for isolating the contribution of the descriptors themselves.

Reading between the lines

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

  • An immediate testable extension is to use DINO_align as an objective in evolutionary or gradient-based descriptor search: if the claimed signal is real, optimizing alignment directly should produce descriptor sets whose accuracy matches ESCHER's, without the critic loop.
  • The negative frequency–similarity correlation suggests a general 'specificity premium' in prompt design that may also apply to other CLIP-based applications, such as retrieval or text-to-image generation.
  • If DINO_align is a genuine measure of semantic grounding, it could serve as a proxy for human interpretability in concept-bottleneck models, letting developers audit whether a concept set is meaningful before deploying it.
  • The choice of reference embedding is a sensitivity knob: swapping DINOv2 for another self-supervised model would test whether the rankings reflect universal visual structure or DINOv2-specific geometry.
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 two label-free metrics for evaluating text-based visual descriptors used in zero-shot classification with vision-language models. The first, DINO_align, computes a mutual k-nearest-neighbor alignment between the descriptor-induced CLIP similarity space and a DINOv2 image embedding space. The second, CLIP_sim, measures how often and how strongly each descriptor matches image-text pairs in a sample of CLIP's pre-training data. The authors evaluate five descriptor sets (Class Name Prompt, CBD, ESCHER, DCLIP, WaffleCLIP) on CIFAR-100, CUB, and NABirds, report a consistent DINO_align ordering (DCLIP < WaffleCLIP < CBD <= ESCHER < Class Names), a negative frequency-similarity trend in CLIP_sim, and argue that ESCHER iterations improve both CLIP_sim and accuracy. The manuscript claims these metrics move beyond accuracy and provide an unsupervised signal for descriptor quality and refinement.

Significance. If the central claim is valid, the paper would contribute a useful and inexpensive way to evaluate descriptor sets without labels, with clear applications to descriptor search and interpretability in VLM-based classification. The idea of combining representational alignment and pre-training-data compatibility is timely, and the code release is a practical strength. However, the main empirical support is thin: the headline DINO_align ordering rests on an internally inconsistent computation for the Class Name Prompt, only three datasets and five descriptor sets are used, no uncertainty or significance estimates are reported, and the choice of DINOv2 as the reference space is not justified or stress-tested. The paper's significance is therefore conditional on a corrected computation and substantially stronger evaluation.

major comments (4)
  1. [Section 3.1, step 1; Section 4.3; Figure 3] The stated rule that class names are removed from all descriptors makes the reported Class Name Prompt result impossible. For the Class Name Prompt, descriptors have the form 'An image of a {class name}'; after removing the class name, every descriptor reduces to the constant template 'An image of a'. The combined descriptor set then contains either one unique text or identical columns in the text-embedding matrix, so the projected similarity matrix S = XY^T has no class-discriminative structure and DINO_align cannot be the highest value shown in Figure 3. The only way to obtain that ranking is if class-name tokens were retained for this condition, which is precisely the leakage that Supplementary Figure 9 shows destroys all consistent trends. Since Class Names is one of the five data points and anchors the claimed ordering, Figure 3 does not support the central claim as stated. The authors must either define a non-degenerate class-name-free prompt, recompute the scores, and report what happens to the ordering, or explicitly revise the claim if Class Names no longer ranks highest.
  2. [Section 4.1, Figure 3, Figure 4] The empirical basis for the central 'consistent' claims is too thin. Each descriptor set contributes a single point with no error bars, no repeated sampling, and no significance test for the DINO_align ordering or for the negative frequency-similarity trend in Figure 4. With only three datasets and five methods, the statement that alignment 'consistently reflects the quality of descriptor sets, regardless of their classification accuracy' is an overgeneralization. The authors should report variability across runs or bootstrap resamples, and provide a statistical test or confidence intervals for the claimed ordering and trend.
  3. [Section 4.1, Section 3.1] The choice of DINOv2-Base as the reference embedding space is not justified beyond calling it 'the current SOTA image embedding space', and no sensitivity analysis is provided. Since DINO_align is defined as agreement with DINOv2 geometry, the metric could simply be measuring compatibility with DINOv2 rather than general descriptor quality. This is load-bearing for the central claim, because if a different reference model (e.g., a supervised ViT, a CLIP image encoder, or a different DINOv2 size) yields a different ranking, then 'alignment' is not a unique or well-founded quality signal. The authors should provide a rationale grounded in the Platonic Representation Hypothesis with evidence, or a sensitivity analysis over several reference models.
  4. [Section 3.2, Section 4.1, Figure 5] CLIP_sim depends on several free parameters—similarity threshold tau=0.7, the top 5% caption fraction, and the number of descriptors/neighbors involved—whose sensitivity is not examined. The consistent negative frequency-similarity trend in Figure 4 is reported without a correlation coefficient or confidence interval, and the causal interpretation ('rare descriptors are more visually grounded') is speculative. This matters for the ESCHER convergence claim in Figure 5, where CLIP_sim is used as a quality signal; the authors should show that the reported trends are stable across reasonable parameter choices and sample sizes. In addition, because both CLIP_sim and the zero-shot accuracy are computed using the same CLIP model, the authors should discuss to what extent the observed correlation is an artifact of sharing the same embedding space.
minor comments (5)
  1. [Figure 4, Figure 9] There are typos in the figure text: 'Relevent' should be 'Relevant' in Figure 4, and 'wihtout' should be 'without' in Figure 9.
  2. [References] References [13] and [14] both point to Menon and Vondrick, and references [17] and [18] both point to Radford et al.; these duplicates should be merged.
  3. [Section 3.1] The definition of the alignment score does not state the distance metric used for k-nearest neighbors in the image and descriptor spaces, nor how ties are broken. This is especially important because a constant descriptor space (as produced by the class-name-free Class Name Prompt) would contain degenerate distances and arbitrary neighbor sets.
  4. [Section 3.2] The definition of Sim(d) divides by |I_d|, which excludes descriptors with zero caption matches. The paper should state explicitly how zero-frequency descriptors are treated when computing CLIP_sim averages and when plotting Figures 4, 6, 7, and 8.
  5. [Figure 5] The x-axis tick labels in Figure 5 appear truncated or duplicated (for example, '19', '18 18 20'), which makes it difficult to read the iteration numbers; the authors should regenerate the figure with clear axis labels.

Circularity Check

1 steps flagged · score 6.0 of 10

DINO_align's top-ranked Class Name Prompt is impossible under the stated class-name removal rule; the headline ordering rests on a leaked/self-definitional input.

  1. other [Section 3.1 (step 1), Section 4.2, Section 4.3, Supplementary C]
    "Section 3.1: "Remove class names from all descriptors to avoid bias." Section 4.2: "Class Name Prompt: Uses the standard zero-shot format 'An image of a {class name}'." Section 4.3: "The resulting scores follow a consistent pattern: DCLIP < WAFFLE CLIP < CBD <= ESCHER < CLASS NAME." Supp. C: "In contrast to the clearer trends observed when class names are removed, no consistent patterns emerge across methods or datasets.""

    Applying step 1 to the Class Name Prompt leaves the constant string 'An image of a ' for every class, so S=XY^T has identical columns and the M-KNN alignment cannot encode class-discriminative structure, let alone the maximum in Figure 3. The only computation that can place Class Names at the top retains the class-name token, exactly the bias the method says it removes. Supp. Fig. 9 confirms that retention destroys the consistent trend, so the Figure 3 ordering is not a valid output of the stated algorithm. The headline result 'CLASS NAME' highest is therefore an artifact of the unremoved input, and the claim that alignment 'consistently reflects descriptor quality' is supported by a data point that reduces to the very class-name signal the metric was designed to eliminate.

full rationale

The main circular/self-contradictory step is in the DINO_align validation: the Class Name Prompt's top rank is impossible under the stated class-name removal rule, implying class-name leakage and reducing that data point to the excluded input. For the other descriptor sets, DINO_align is computed from CLIP-projected descriptor similarities versus a DINO reference embedding and is not fitted to accuracy, so the metric is not circular in itself. CLIP_sim shares the CLIP model with the accuracy target, which limits its independence, but the reported relationship is an empirical correlation rather than an equation-level reduction; I do not score that as circular. The ESCHER citations are self-citations, but the paper runs new analyses rather than importing the ESCHER conclusions as evidence, so they are not load-bearing self-citation. The central ordering is partially invalidated by the Class Names inconsistency, warranting a score of 6 rather than a lower score.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claims rest on three domain assumptions (DINOv2 as ground truth, M-KNN validity in the projected space, and the 5M subsample) and on three heuristic hyperparameters (τ, top 5%, and k) that are not fitted to the target accuracy and are not ablated. No new physical or conceptual entities are introduced.

free parameters (3)
  • CLIP_sim similarity threshold τ = 0.7
    Hand-set threshold for counting a caption as a match; changes both Freq(d) and Sim(d); no sensitivity analysis. Section 3.2 and supplementary.
  • top-k caption fraction = top 5%
    Selects the caption pool for each descriptor; affects all CLIP_sim statistics.
  • M-KNN neighborhood size k = average images per class (per dataset)
    Heuristic from supplementary Section A; not justified formally and affects DINO_align scores.
assumptions (3)
  • domain assumption DINOv2 embedding space is a valid reference for meaningful visual structure.
    Assumed in Section 4.1 where DINOv2-Base is called current SOTA; DINO_align has no meaning without this.
  • domain assumption Mutual k-NN alignment with cosine similarity measures true representational similarity of descriptor spaces.
    Taken from Huh et al. [5] without re-validation in the descriptor-induced projection setting.
  • domain assumption A 5M random subsample of DFN2B approximates CLIP's full pre-training distribution for text similarity statistics.
    Stated in Section 4.1 and acknowledged as a limitation in Section D; all CLIP_sim results depend on it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Accuracy: Metrics that Uncover What Makes a 'Good' Visual Descriptor." pith.science (2026). https://pith.science/paper/L2NLX3UF

@misc{pith2026250703542,
  author       = {Pith},
  title        = {Pith review of: Beyond Accuracy: Metrics that Uncover What Makes a 'Good' Visual Descriptor},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L2NLX3UF}},
  note         = {Machine review of arXiv:2507.03542}
}
read the original abstract

Text-based visual descriptors--ranging from simple class names to more descriptive phrases--are widely used in visual concept discovery and image classification with vision-language models (VLMs). Their effectiveness, however, depends on a complex interplay of factors, including semantic clarity, presence in the VLM's pre-training data, and how well the descriptors serve as a meaningful representation space. In this work, we systematically analyze descriptor quality along two key dimensions: (1) representational capacity, and (2) relationship with VLM pre-training data. We evaluate a spectrum of descriptor generation methods, from zero-shot LLM-generated prompts to iteratively refined descriptors. Motivated by ideas from representation alignment and language understanding, we introduce two alignment-based metrics--Global Alignment and CLIP Similarity--that move beyond accuracy. These metrics shed light on how different descriptor generation strategies interact with foundation model properties, offering new ways to study descriptor effectiveness beyond accuracy evaluations.

Figures

Figures reproduced from arXiv: 2507.03542 by the authors.

Figure 1
Figure 1. Overview: Achieving high accuracy does not guarantee that a set of descriptors is “good”. Other factors like interpretabil￾ity may suffer. Global alignment and CLIP similarity can serve as new metrics for evaluating and understanding different sets of visual descriptors. constitutes a “good” textual visual descriptor? Descriptors can be obtained in many ways—from using class names di￾rectly [18], to leveraging large… view at source ↗
Figure 2
Figure 2. Overview of our descriptor evaluation metrics. (Left) DINOalign measures how well a set of descriptors preserves class structure by aligning the descriptor-induced similarity matrix with that of DINO image features. (Right) CLIPsim assesses how well descriptors align with visual content by retrieving related image-text pairs from CLIP’s pre-training dataset and averaging their similarity scores. We define two statis… view at source ↗
Figure 3
Figure 3. Classification accuracy versus alignment score across all datasets. Alignment consistently reflects the quality of descrip￾tor sets, regardless of their classification accuracy. Viewing classification by description as a form of seman￾tic projection—analogous to PCA—the consistently high alignment scores of class names across datasets suggest that they function like the first principal components: basis di￾rections … view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Distribution of descriptors by frequency and correspond [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 6
Figure 6. Figure 6: Classification accuracy versus CLIPsim across all datasets. 10 0 10 1 10 2 10 3 10 4 Num of Relevent Train Data (log scale) 0.15 0.20 0.25 0.30 0.35 CLIP Similarity [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Distribution of descriptors by frequency (number of [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Distribution of descriptors by frequency (number of [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: presents alignment and accuracy scores when class names are retained in all descriptors. In contrast to the clearer trends observed when class names are re￾moved, no consistent patterns emerge across methods or datasets—highlighting how strongly class name inclusion ca…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 20 canonical work pages

  1. [1]

    Flamingo: a Visual Language Model for Few-Shot Learning,

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, An- toine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Se- bastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sa- hand Sharifzadeh, Mikolaj Bink...

    work page

  2. [2]

    Vision-language models do not understand negation, 2025

    Kumail Alhamoud, Shaden Alshammari, Yonglong Tian, Guohao Li, Philip Torr, Yoon Kim, and Marzyeh Ghassemi. Vision-language models do not understand negation, 2025. 2

    work page 2025

  3. [3]

    Calmon, and Himabindu Lakkaraju

    Usha Bhalla, Alex Oesterling, Suraj Srinivas, Flavio P. Calmon, and Himabindu Lakkaraju. Interpreting CLIP with Sparse Linear Concept Embeddings (SpLiCE), 2024. arXiv:2402.10376. 1

    arXiv 2024

  4. [4]

    Evolving interpretable visual classifiers with large language models

    Mia Chiquier, Utkarsh Mall, and Carl V ondrick. Evolving interpretable visual classifiers with large language models. CoRR, 2024. 1, 2

    work page 2024

  5. [5]

    The platonic representation hypothesis, 2024

    Minyoung Huh, Brian Cheung, Tongzhou Wang, and Phillip Isola. The platonic representation hypothesis, 2024. 2

    work page 2024

  6. [6]

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

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

  7. [7]

    What’s ”up” with vision-language models? Investigating their strug- gle with spatial reasoning, 2023

    Amita Kamath, Jack Hessel, and Kai-Wei Chang. What’s ”up” with vision-language models? Investigating their strug- gle with spatial reasoning, 2023. arXiv:2310.19785 [cs]. 2

    arXiv 2023

  8. [8]

    Large language models struggle to learn long-tail knowledge, 2023

    Nikhil Kandpal, Haikang Deng, Adam Roberts, Eric Wal- lace, and Colin Raffel. Large language models struggle to learn long-tail knowledge, 2023. 2, 3

    work page 2023

Show all 31 references
  1. [9]

    Cifar- 100 (canadian institute for advanced research)

    Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton. Cifar- 100 (canadian institute for advanced research). 2009. 3

    2009

  2. [10]

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

    Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learn- ing with momentum distillation. Advances in neural infor- mation processing systems, 34:9694–9705, 2021. 2

    2021

  3. [11]

    Improved baselines with visual instruction tuning

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296–26306, 2024

    2024

  4. [12]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2

    2024

  5. [13]

    Visual classification via description from large language models, 2022

    Sachit Menon and Carl V ondrick. Visual classification via description from large language models, 2022. 1, 3

    2022

  6. [14]

    Visual Classification via Description from Large Language Models

    Sachit Menon and Carl V ondrick. Visual Classification via Description from Large Language Models. 2022. 2, 1

    2022

  7. [15]

    Slip: Self-supervision meets language-image pre- training

    Norman Mu, Alexander Kirillov, David Wagner, and Sain- ing Xie. Slip: Self-supervision meets language-image pre- training. In European conference on computer vision, pages 529–544. Springer, 2022. 2

    2022

  8. [16]

    Dinov2: Learning robust visual features with- out supervision, 2024

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mah- moud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michae...

    2024

  9. [17]

    Learning Transferable Visual Models From Natural Language Supervision, 2021

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision, 2021. Num- ber: arXi...

    2021 arXiv

  10. [18]

    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, and others. Learning transferable visual models from natural language supervision. In International conference on machine learn- ...

    2021

  11. [19]

    Sophia Koepke, Oriol Vinyals, Cordelia Schmid, and Zeynep Akata

    Karsten Roth, Jae Myung Kim, A. Sophia Koepke, Oriol Vinyals, Cordelia Schmid, and Zeynep Akata. Waffling around for performance: Visual classification with random words and broad concepts, 2023. 1, 2, 3

    2023

  12. [20]

    Sun, and Swarat Chaudhuri

    Atharva Sehgal, Patrick Yuan, Ziniu Hu, Yisong Yue, Jen- nifer J. Sun, and Swarat Chaudhuri. Self-evolving visual concept library using vision-language critics, 2025. 1, 2, 3

    2025

  13. [21]

    Love, Christo- pher J

    Ilia Sucholutsky, Lukas Muttenthaler, Adrian Weller, Andi Peng, Andreea Bobu, Been Kim, Bradley C. Love, Christo- pher J. Cueva, Erin Grant, Iris Groen, Jascha Achterberg, Joshua B. Tenenbaum, Katherine M. Collins, Katherine L. Hermann, Kerem Oktar, Klaus Greff, Martin N. Heba...

    2024

  14. [22]

    Understanding the emergence of multimodal representation alignment, 2025

    Megan Tjandrasuwita, Chanakya Ekbote, Liu Ziyin, and Paul Pu Liang. Understanding the emergence of multimodal representation alignment, 2025. 2

    2025

  15. [23]

    Eyes Wide Shut? Exploring the Vi- sual Shortcomings of Multimodal LLMs

    Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes Wide Shut? Exploring the Vi- sual Shortcomings of Multimodal LLMs. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9568–9578, Seattle, W A, USA, 2024. IEEE. 2

    2024

  16. [24]

    Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection

    Grant Van Horn, Steve Branson, Ryan Farrell, Scott Haber, Jessie Barry, Panos Ipeirotis, Pietro Perona, and Serge Be- longie. Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection. In Proceedings of t...

    2015

  17. [25]

    C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. Caltech-ucsd birds-200-2011. Technical Report CNS-TR- 2011-001, California Institute of Technology, 2011. 3

    2011

  18. [26]

    An Yan, Yu Wang, Yiwu Zhong, Chengyu Dong, Zexue He, Yujie Lu, William Wang, Jingbo Shang, and Julian J. McAuley. Learning concise and descriptive attributes for vi- sual recognition. CoRR, abs/2308.03685, 2023. 2

    2023 arXiv

  19. [27]

    Learning concise and descriptive attributes for visual recognition

    An Yan, Yu Wang, Yiwu Zhong, Chengyu Dong, Zexue He, Yujie Lu, William Yang Wang, Jingbo Shang, and Julian McAuley. Learning concise and descriptive attributes for visual recognition. In Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 3090–3100,

  20. [28]

    Language in a bottle: Language model guided concept bottlenecks for in- terpretable image classification, 2023

    Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for in- terpretable image classification, 2023. 2

    2023

  21. [29]

    Filip: Fine-grained interactive language-image pre-training

    Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. In International Conference on Learning Rep- resentations, 2022. 2

    2022

  22. [30]

    laysan albatross, which is a

    Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mo- jtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research, 2022. 2 Beyond Accuracy: Metrics that Uncover What Makes a ‘Good’ Visual Descr...

    2022

  23. [2022]

    arXiv:2204.14198 [cs]. 2

    arXiv

Pith tools

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