Pith. sign in

REVIEW 4 major objections 6 minor 47 references

EvdCLIP: Improving Vision-Language Retrieval with Entity Visual Descriptions from Large Language Models

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that enriching text queries with LLM-generated visual descriptions of entities, filtered by a trained rewriter, consistently improves CLIP-based vision-language retrieval on standard benchmarks and on short, entity-rich…

desk verdict Solid, practical retrieval paper with a plausible but not yet isolated causal claim: the gains are real in the tables, but the EVD content is never tested without the rewriting pipeline. read the letter →

arxiv 2505.18594 v1 pith:M3WSTZWA submitted 2025-05-24 cs.CV cs.IR

GuangHao Meng , Sunan He , Jinpeng Wang , Tao Dai , Letian Zhang , Jieming Zhu , Qing Li , Gang Wang
show 2 more authors
Rui Zhang Yong Jiang
This is my paper · ORCID
classification cs.CVcs.IR
keywords vision-languageretrievalentityvisualdescriptionsCLIPqueryrewritinglargelanguagemodelspreferenceoptimizationimage-textknowledge-enhanced
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 vision-language retrieval fails when the text encoder lacks visual knowledge of the entities in a query, such as what a mixing board or a camp of tents looks like. It proposes EvdCLIP, which uses a large language model to write Entity Visual Descriptions (EVDs), short lists of visual features like shape, color, and parts, and stores them in a knowledge base of about 10,000 entities. A trainable EVD-aware Rewriter (EaRW) integrates the relevant descriptions into each query while filtering noise, so the enhanced query aligns better with images. The paper reports consistent recall gains over CLIP and over prior description-enhancement methods on Flickr30K and MSCOCO, with the largest gains on short, entity-rich business queries such as wallpaper search. The central claim is that injecting LLM-written visual knowledge into queries, rather than into the model weights, improves cross-modal alignment.

What carries the argument

The load-bearing object is the EVD knowledge base: a mapping from roughly 10,237 visual entities to lists of visual descriptions generated by an LLM, focusing on color, shape, parts, and quantity. The second component is the EVD-aware Rewriter (EaRW), initialized from T5-large, which turns a raw query into an EVD-enhanced query. EaRW is warmed up by supervised fine-tuning on a dataset where CLIP's retrieval scores rank candidate rewritten queries as pseudo-labels, then aligned to the retriever with Preference Rank Optimization, a listwise extension of Bradley-Terry preference learning. During CLIP fine-tuning, rewritten and original queries are randomly mixed, and at inference the final score averages the original and rewritten query scores.

What would settle it

Shuffle the EVD knowledge base so each entity receives another entity's visual descriptions, then re-run the rewrite-and-retrieval pipeline and compare recall to the correctly aligned version. If recall does not drop, the specific entity-to-description mapping is not what carries the gain.

Watch

Extended reading notes

Core claim

The central discovery is that a frozen CLIP model can be made to retrieve more accurately by rewriting each text query to include compact visual descriptions of its entities. The descriptions are generated offline by an LLM, collected into an EVD knowledge base, and selected per query by a learned rewriter that is trained first by supervised fine-tuning on CLIP-scored pseudo-labels and then by preference rank optimization against retrieval feedback. On Flickr30K and MSCOCO, EvdCLIP consistently outperforms CLIP on every reported metric, with R@1 gains up to about 1.6 points, and the method also lifts CoCa and EVA-02-CLIP backbones. On four categories of a Chinese commercial image-text dataset it beats CLIP everywhere, and the paper attributes the largest margins, around 6.7 points at higher recalls on wallpaper, to queries that are short, vague, and entity-rich. The paper also shows that EVDs allow the system to recognise novel concepts and to correct retrieval bias by editing the stored descriptions.

Load-bearing premise

The method assumes that the entities appearing in a query can be reliably extracted and matched to the 10,237-entry knowledge base, but the paper does not specify the entity-linking algorithm, disambiguation, or what happens for unmatched entities.

Editorial extensions

If this is right

  • If the central claim holds, any dual-encoder retriever can be upgraded at query time by adding a frozen rewriter plus an EVD table, without changing the encoder weights.
  • Short, underspecified, entity-heavy queries, the typical failure mode in commercial search, stand to gain the most, as the wallpaper results suggest.
  • EVD gives operators a direct control handle: editing an entity's stored visual description changes retrieval behavior, enabling bias correction and knowledge updates without retraining.
  • The reported gains are additive on top of strong backbones such as CLIP, CoCa, and EVA-02-CLIP, implying the visual-knowledge signal is complementary to scale and better training techniques.

Reading between the lines

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

  • The paper leaves open how entities are linked in arbitrary queries; a natural extension is to test whether retrieval gains correlate with entity-linking confidence, which would isolate the rewriter's contribution from the knowledge base's.
  • If the mechanism generalizes, query-side augmentation could be a cheaper alternative to fine-grained model scaling for distinguishing visually similar concepts, since it adds a lightweight text-to-text model rather than retraining the encoder.
  • A testable extension is to apply the trained EaRW to retrieval models never seen during training; because the rewriter only touches text, it may transfer across CLIP variants without the reported fine-tuning on each backbone.
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 / 6 minor

Summary. This paper proposes EvdCLIP, a method for image-text retrieval that augments text queries with entity visual descriptions (EVDs) generated offline by large language models. The method builds a knowledge base of visual descriptions for roughly 10,000 visual entities, extracts entities from queries, retrieves their EVDs, and uses a trainable T5-based rewriter (EaRW) to integrate EVDs into the query. EaRW is trained on a dataset of query rewrites scored by CLIP, with supervised fine-tuning followed by preference rank optimization. At inference, the rewritten query is scored together with the original query. Experiments on Flickr30K, MSCOCO, and a Huawei business dataset report consistent improvements over CLIP, CoCa, and EVA-02 baselines, with ablations over description sources and LLMs.

Significance. If the reported gains are attributable to EVD content, the paper would be a useful contribution: it gives a concrete recipe for injecting LLM-derived visual knowledge into retrieval, shows compatibility with multiple CLIP-style backbones, and includes an interesting business-domain application. The paper is careful to compare against WordNet-based descriptions and to vary the LLM source. However, the central attribution claim is not yet established because the experimental design does not separate the effect of EVD content from the effect of learned query rewriting, and the entity-linking step is described only at the level of 'we retrieve the entities.' The improvements are small relative to likely run-to-run variance, and no significance testing is reported.

major comments (4)
  1. [EVD-aware Rewriter and Main Results] The paper's central claim that EVD content improves retrieval is underdetermined. EaRW is trained on rewrites selected by CLIP similarity and is then used both to produce the queries for fine-tuning CLIP and for inference; no ablation removes EVD content while keeping the rewriting pipeline. Without such a control (e.g., a rewriter trained on generic caption-like LLM rewrites without EVD, evaluated under the same protocol), the 1-3 point gains reported in Table 1 could be caused by the act of learned rewriting rather than by visual descriptions.
  2. [EVD-aware Rewriter] The entity-linking step is unspecified. The text states 'we first retrieve the entities ei' but gives no algorithm, disambiguation strategy, or fallback for queries whose entities are not in the 10,237-entry knowledge base. Because this step determines whether EVDs are actually injected, the paper should describe the linking method and report its coverage and accuracy on the test queries.
  3. [Table 1 and Main Results] The reported improvements are small (typically 1-2 recall points at R@1) and no error bars or significance tests are provided. Given the small margins, the claim that EvdCLIP 'consistently outperforms' baselines needs support from multiple seeds or a statistical test.
  4. [Preference Alignment, Eq. (5)-(6)] The reward function in the preference optimization is described inconsistently. The text says image-text similarity scores from the retrieval system serve as rewards, but then defines r(·) as 'the normalized log probability of the rewrite generated in PRO.' The exact reward used in Eqs. (5)-(6) should be specified unambiguously, since it is central to the EaRW training procedure.
minor comments (6)
  1. [Introduction] The phrase 'EVD knowlege base' should be 'EVD knowledge base'.
  2. [Figure 1 caption] The caption says 'camping of tents' while the text consistently uses 'camp of tents'; please make the wording consistent.
  3. [Methodology, Eq. (1)-(2)] The summation over j in the denominators should start at 1 (or the batch indexing should be clarified), since the batch is defined as N samples.
  4. [Implementation Details] There is a typo 'insted' for 'instead', and the abbreviation 'RPO' is introduced here while the method is called 'PRO' earlier in the paper.
  5. [Table 1] The backbone name 'EV A-02-CLIP' should be 'EVA-02-CLIP' for consistency with the text.
  6. [Related Work] The sentence 'Knowledge-CLIP (Pan et al. 2022) and ACP (Pan et al. 2022)' appears to cite the same reference for two different methods; please clarify whether the two methods have distinct references.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: EaRW is trained with CLIP feedback and CLIP is fine-tuned on EaRW outputs, but the evaluation is external and held-out; the missing no-EVD control is an attribution gap, not a definitional reduction.

full rationale

The paper's derivation chain is empirical, not definitional. The EVD knowledge base is generated offline by LLMs from a predefined entity set; the EaRW rewriter is trained on CLIP-similarity pseudo-labels and then used to rewrite queries during CLIP fine-tuning. This is a self-training-style loop, but it does not make the reported recall gains true by construction: the evaluation uses held-out Flickr30K, MSCOCO, and Huawei test splits and compares against fine-tuned CLIP, CoCa, EVA-02-CLIP, and description-enhancement baselines. The fact that EaRW is optimized with CLIP feedback means the improvements are not clean evidence that EVD content per se causes the gains; a no-EVD rewriting control would be needed to separate generic rewriting from entity-specific visual descriptions. That is a missing ablation or attribution gap, not a circular reduction. Self-citations in related work (e.g., Wang et al. 2022-2024, Zhu et al. 2024, Peng et al. 2024) are not load-bearing for the central claim; no uniqueness theorem or ansatz is imported from the authors' prior work. No equation in the paper defines the output in terms of the input, and no fitted parameter is renamed as a prediction. Therefore no significant circularity is identified.

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

No new natural kinds or physical entities are introduced; EVD and EaRW are engineered components, not invented entities requiring independent evidence. The central experiment rests on unstated modeling choices around entity matching, CLIP-reward-based pseudo-labeling, and coverage assumptions from the EVD base built on training captions.

free parameters (5)
  • low-frequency entity filter threshold = Unspecified
    The paper states 'we filter out low-frequency entities' but gives no threshold or coverage analysis; this directly affects the 10,237-entity EVD base.
  • query rewriting probability p = 0.6
    Probability of applying EaRW during CLIP fine-tuning; chosen by hand.
  • SFT loss weight beta = 0.2
    Weight balancing PRO and SFT losses in Eq. 7.
  • preference rank length k = 5
    Number of EVD-enhanced query candidates ranked per query in preference optimization.
  • inference score averaging weight = Unspecified
    The paper says averaging original and EVD-enhanced scores but does not give the weighting.
assumptions (5)
  • domain assumption CLIP similarity scores are a valid reward for query rewriting quality.
    Used to construct DEQR pseudo-labels and PRO rewards; if CLIP is biased, EaRW learns to exploit CLIP instead of adding visual knowledge.
  • domain assumption LLM-generated visual descriptions from text-only prompts are accurate for entities appearing in images.
    The EVD base is generated offline by ChatGPT or PanGu without image grounding; hallucinated descriptions must be filtered by EaRW.
  • domain assumption Entities in arbitrary queries can be matched to the EVD knowledge base.
    The paper says 'we first retrieve the entities ei' but provides no matching or disambiguation algorithm.
  • domain assumption Entity coverage from training captions transfers to test queries.
    EVD base is built from Flickr30k and MSCOCO training splits; test queries may contain unseen entities after low-frequency filtering.
  • standard math Preference Rank Optimization (Eq. 6) correctly generalizes pairwise preferences to listwise rankings.
    Taken as an external result from Song et al. 2024.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EvdCLIP: Improving Vision-Language Retrieval with Entity Visual Descriptions from Large Language Models." pith.science (2026). https://pith.science/paper/M3WSTZWA

@misc{pith2026250518594,
  author       = {Pith},
  title        = {Pith review of: EvdCLIP: Improving Vision-Language Retrieval with Entity Visual Descriptions from Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M3WSTZWA}},
  note         = {Machine review of arXiv:2505.18594}
}
read the original abstract

Vision-language retrieval (VLR) has attracted significant attention in both academia and industry, which involves using text (or images) as queries to retrieve corresponding images (or text). However, existing methods often neglect the rich visual semantics knowledge of entities, thus leading to incorrect retrieval results. To address this problem, we propose the Entity Visual Description enhanced CLIP (EvdCLIP), designed to leverage the visual knowledge of entities to enrich queries. Specifically, since humans recognize entities through visual cues, we employ a large language model (LLM) to generate Entity Visual Descriptions (EVDs) as alignment cues to complement textual data. These EVDs are then integrated into raw queries to create visually-rich, EVD-enhanced queries. Furthermore, recognizing that EVD-enhanced queries may introduce noise or low-quality expansions, we develop a novel, trainable EVD-aware Rewriter (EaRW) for vision-language retrieval tasks. EaRW utilizes EVD knowledge and the generative capabilities of the language model to effectively rewrite queries. With our specialized training strategy, EaRW can generate high-quality and low-noise EVD-enhanced queries. Extensive quantitative and qualitative experiments on image-text retrieval benchmarks validate the superiority of EvdCLIP on vision-language retrieval tasks.

Figures

Figures reproduced from arXiv: 2505.18594 by the authors.

Figure 1
Figure 1. Illustration of entity visual descriptions (EVD) en [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Challenges of EVD integration to VLR. (a) Noise [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The overall architecture of EvdCLIP comprises two components: EVD offline generation via LLMs and EVD [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visualization examples of retrieval. We present im [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: Examples of Huawei Wallpaper Retrieval. The left [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 26 canonical work pages

  1. [1]

    K.; and Huang, F

    An, B.; Zhu, S.; Panaitescu-Liess, M.-A.; Mummadi, C. K.; and Huang, F. 2023. More context, less distraction: Improving zero-shot inference of clip by inferring and describing spurious features. In Workshop on Efficient Systems for Foundation Models@ ICML2023

    work page 2023

  2. [2]

    D.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A.; et al

    Brown, T.; Mann, B.; Ryder, N.; Subbiah, M.; Kaplan, J. D.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A.; et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33: 1877--1901

    2020

  3. [3]

    Cheng, M.; Sun, Y.; Wang, L.; Zhu, X.; Yao, K.; Chen, J.; Song, G.; Han, J.; Liu, J.; Ding, E.; et al. 2022. ViSTA: vision and scene text aggregation for cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5184--5193

    work page 2022

  4. [4]

    E.; et al

    Chiang, W.-L.; Li, Z.; Lin, Z.; Sheng, Y.; Wu, Z.; Zhang, H.; Zheng, L.; Zhuang, S.; Zhuang, Y.; Gonzalez, J. E.; et al. 2023. Vicuna: An open-source chatbot impressing gpt-4 with 90\ See https://vicuna. lmsys. org (accessed 14 April 2023)

    work page 2023

  5. [5]

    E.; and Yeung-Levy, S

    Dunlap, L.; Zhang, Y.; Wang, X.; Zhong, R.; Darrell, T.; Steinhardt, J.; Gonzalez, J. E.; and Yeung-Levy, S. 2024. Describing differences in image sets with natural language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 24199--24208

    work page 2024

  6. [6]

    Huang, Y.; Tang, J.; Chen, Z.; Zhang, R.; Zhang, X.; Chen, W.; Zhao, Z.; Zhao, Z.; Lv, T.; Hu, Z.; et al. 2024. Structure-CLIP: Towards Scene Graph Knowledge to Enhance Multi-Modal Structured Representations. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 2417--2425

    work page 2024

  7. [7]

    Kilgarriff, A. 2000. Wordnet: An electronic lexical database

    work page 2000

  8. [8]

    Kim, W.; Son, B.; and Kim, I. 2021. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, 5583--5594. PMLR

    work page 2021

Show all 47 references
  1. [9]

    P.; and Ba, J

    Kingma, D. P.; and Ba, J. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

    2014 arXiv

  2. [10]

    Li, J.; Li, D.; Xiong, C.; and Hoi, S. 2022. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, 12888--12900. PMLR

    2022

  3. [11]

    Li, J.; Selvaraju, R.; Gotmare, A.; Joty, S.; Xiong, C.; and Hoi, S. C. H. 2021. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34: 9694--9705

    2021

  4. [12]

    Lin, T.-Y.; Maire, M.; Belongie, S.; Hays, J.; Perona, P.; Ramanan, D.; Doll \'a r, P.; and Zitnick, C. L. 2014. Microsoft coco: Common objects in context. In Computer Vision--ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 1...

    2014

  5. [13]

    Liu, H.; Li, C.; Wu, Q.; and Lee, Y. J. 2023. Visual Instruction Tuning

    2023

  6. [14]

    Ma, H.; Zhao, H.; Lin, Z.; Kale, A.; Wang, Z.; Yu, T.; Gu, J.; Choudhary, S.; and Xie, X. 2022. Ei-clip: Entity-aware interventional contrastive learning for e-commerce cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,...

    2022

  7. [15]

    Ma, X.; Gong, Y.; He, P.; Zhao, H.; and Duan, N. 2023. Query rewriting for retrieval-augmented large language models. arXiv preprint arXiv:2305.14283

    2023 arXiv

  8. [16]

    Maniparambil, M.; Vorster, C.; Molloy, D.; Murphy, N.; McGuinness, K.; and O'Connor, N. E. 2023. Enhancing clip with gpt-4: Harnessing visual descriptions as prompts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 262--271

    2023

  9. [17]

    Menon, S.; and Vondrick, C. 2022. Visual Classification via Description from Large Language Models. arXiv preprint arXiv:2210.07183

    2022 arXiv

  10. [18]

    OpenAI, T. 2022. Chatgpt: Optimizing language models for dialogue. OpenAI

    2022

  11. [19]

    Ordonez, V.; Kulkarni, G.; and Berg, T. 2011. Im2text: Describing images using 1 million captioned photographs. Advances in neural information processing systems, 24

    2011

  12. [20]

    Pan, X.; Ye, T.; Han, D.; Song, S.; and Huang, G. 2022. Contrastive Language-Image Pre-Training with Knowledge Graphs. arXiv preprint arXiv:2210.08901

    2022 arXiv

  13. [21]

    Peng, W.; Li, G.; Jiang, Y.; Wang, Z.; Ou, D.; Zeng, X.; Xu, D.; Xu, T.; and Chen, E. 2024. Large language model based long-tail query rewriting in taobao search. In Companion Proceedings of the ACM on Web Conference 2024, 20--28

    2024

  14. [22]

    A.; Wang, L.; Cervantes, C

    Plummer, B. A.; Wang, L.; Cervantes, C. M.; Caicedo, J. C.; Hockenmaier, J.; and Lazebnik, S. 2015. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, 2641--2649

    2015

  15. [23]

    Pratt, S.; Covert, I.; Liu, R.; and Farhadi, A. 2023. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 15691--15701

    2023

  16. [24]

    Qi, Z.; Khorram, S.; and Li, F. 2019. Visualizing Deep Networks by Optimizing with Integrated Gradients. In CVPR Workshops, volume 2, 1--4

    2019

  17. [25]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al

    Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, 8748--8763. PMLR

    2021

  18. [26]

    Raffel, C.; Shazeer, N.; Roberts, A.; Lee, K.; Narang, S.; Matena, M.; Zhou, Y.; Li, W.; and Liu, P. J. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140): 1--67

    2020

  19. [27]

    Shen, X.; Zhang, R.; Zhao, X.; Zhu, J.; and Xiao, X. 2024. PMG: Personalized Multimodal Generation with Large Language Models. In Proceedings of the ACM on Web Conference 2024, 3833--3843

    2024

  20. [28]

    Song, F.; Yu, B.; Li, M.; Yu, H.; Huang, F.; Li, Y.; and Wang, H. 2024. Preference ranking optimization for human alignment. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 18990--18998

    2024

  21. [29]

    Sun, Q.; Fang, Y.; Wu, L.; Wang, X.; and Cao, Y. 2023. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389

    2023 arXiv

  22. [30]

    Touvron, H.; Martin, L.; Stone, K.; Albert, P.; Almahairi, A.; Babaei, Y.; Bashlykov, N.; Batra, S.; Bhargava, P.; Bhosale, S.; et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288

    2023 arXiv

  23. [31]

    Wang, H.; He, D.; Wu, W.; Xia, B.; Yang, M.; Li, F.; Yu, Y.; Ji, Z.; Ding, E.; and Wang, J. 2022 a . Coder: Coupled diversity-sensitive momentum contrastive learning for image-text retrieval. In Computer Vision--ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23...

    2022

  24. [32]

    Wang, J.; Chen, B.; Liao, D.; Zeng, Z.; Li, G.; Xia, S.-T.; and Xu, J. 2022 b . Hybrid contrastive quantization for efficient cross-view video retrieval. In Proceedings of the ACM Web Conference 2022, 3020--3030

    2022

  25. [33]

    Wang, J.; Ge, Y.; Cai, G.; Yan, R.; Lin, X.; Shan, Y.; Qie, X.; and Shou, M. Z. 2022 c . Object-aware video-language pre-training for retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 3313--3322

    2022

  26. [34]

    Wang, J.; Zeng, Z.; Chen, B.; Wang, Y.; Liao, D.; Li, G.; Wang, Y.; and Xia, S.-T. 2024. Hugs Bring Double Benefits: Unsupervised Cross-Modal Hashing with Multi-granularity Aligned Transformers. International Journal of Computer Vision, 1--33

    2024

  27. [35]

    Wang, J.; Zeng, Z.; Wang, Y.; Wang, Y.; Lu, X.; Li, T.; Yuan, J.; Zhang, R.; Zheng, H.-T.; and Xia, S.-T. 2023. MISSRec: Pre-training and transferring multi-modal interest-aware sequence representation for recommendation. In Proceedings of the 31st ACM International Conference...

    2023

  28. [36]

    Xu, J.; Mei, T.; Yao, T.; and Rui, Y. 2016. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, 5288--5296

    2016

  29. [37]

    Yang, J.; Duan, J.; Tran, S.; Xu, Y.; Chanda, S.; Chen, L.; Zeng, B.; Chilimbi, T.; and Huang, J. 2022. Vision-language pre-training with triple contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 15671--15680

    2022

  30. [38]

    Yang, Y.; Panagopoulou, A.; Zhou, S.; Jin, D.; Callison-Burch, C.; and Yatskar, M. 2023. 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...

    2023

  31. [39]

    Yao, L.; Han, J.; Wen, Y.; Liang, X.; Xu, D.; Zhang, W.; Li, Z.; Xu, C.; and Xu, H. 2022. Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. arXiv preprint arXiv:2209.09407

    2022 arXiv

  32. [40]

    Yu, J.; Wang, Z.; Vasudevan, V.; Yeung, L.; Seyedhosseini, M.; and Wu, Y. 2022. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917

    2022 arXiv

  33. [41]

    Zeng, A.; Liu, X.; Du, Z.; Wang, Z.; Lai, H.; Ding, M.; Yang, Z.; Xu, Y.; Zheng, W.; Xia, X.; et al. 2022. Glm-130b: An open bilingual pre-trained model. arXiv preprint arXiv:2210.02414

    2022 arXiv

  34. [42]

    Zeng, W.; Ren, X.; Su, T.; Wang, H.; Liao, Y.; Wang, Z.; Jiang, X.; Yang, Z.; Wang, K.; Zhang, X.; et al. 2021. Pangu- : Large-scale autoregressive pretrained Chinese language models with auto-parallel computation. arXiv preprint arXiv:2104.12369

    2021 arXiv

  35. [43]

    Zhao, M.; Wang, J.; Liao, D.; Wang, Y.; Duan, H.; and Zhou, S. 2023. Keyword-Based Diverse Image Retrieval by Semantics-aware Contrastive Learning and Transformer. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieva...

    2023

  36. [44]

    Zhu, D.; Chen, J.; Shen, X.; Li, X.; and Elhoseiny, M. 2023. MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models. arXiv preprint arXiv:2304.10592

    2023 arXiv

  37. [45]

    Zhu, J.; Zhou, X.; Wu, C.; Zhang, R.; and Dong, Z. 2024. Multimodal Pretraining and Generation for Recommendation: A Tutorial. In Companion Proceedings of the ACM on Web Conference 2024, 1272--1275

    2024

  38. [46]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  39. [47]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.