Pith. sign in

REVIEW 3 major objections 7 minor 51 references

QuARI: Query Adaptive Retrieval Improvement

T0 review · 3 major / 7 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 QuARI claims that a per-query linear projection of frozen image features beats costly reranking methods on two large-scale retrieval benchmarks.

desk verdict A genuinely new per-query low-rank projection idea with strong ILIAS results, but the INQUIRE text-to-image gains may be inflated by training on an iNaturalist-derived dataset without an overlap check. read the letter →

arxiv 2505.21647 v1 pith:N42VA5NJ submitted 2025-05-27 cs.CV cs.LG

Eric Xing , Abby Stylianou , Robert Pless , Nathan Jacobs This is my paper
classification cs.CVcs.LG
keywords query-adaptiveretrievalhypernetworklow-rankprojectionvision-languageembeddingsinstancetext-to-imagelarge-scalesemi-positivesamples
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that retrieval accuracy can be improved without a bigger encoder or slow re-ranking by computing a query-specific linear projection of an already-computed embedding space. The proposed model, QuARI, reads a query embedding and outputs a low-rank matrix that is applied to every gallery embedding before similarity search, so the cost of adaptation is paid once per query rather than per image. On the ILIAS and INQUIRE benchmarks the authors report consistent gains over frozen vision-language features, over a static task-adaptation projection, and over reranking methods that include large vision-language models, at lower computational cost. If the result holds, it gives a practical recipe for specializing frozen general-purpose features at scale: keep the encoder fixed, and let a small transformer reshape the space around each query.

What carries the argument

The load-bearing mechanism is the query adaptation network, a transformer-based hypernetwork that generates the weights of the projection. It tokenizes the projection matrix as $2r$ zero-initialized column tokens, splits them into $U$ and $V$ tokens, and iteratively refines them for $L$ steps against a query-conditioned control token; the final matrix is $T=\sum_{j=1}^{r} \mathrm{MLP}_u(u_j)\,\mathrm{MLP}_v(v_j)^\top$. The training objective is a symmetric contrastive loss over transformed query-image pairs, with two additions: noise injected into queries to bridge the text-image modality gap, and semi-positive samples taken from precomputed nearest neighbours so that visually similar but incorrect images are not treated as hard negatives.

What would settle it

Check whether near-duplicate images exist between the BioTrove training subset and the INQUIRE gallery and query images; if a nontrivial fraction of INQUIRE evaluation images appears in training, the claimed mechanism is confounded. A cleaner experiment would train QuARI on a biodiversity dataset with no images shared with INQUIRE and see whether the mAP@50 gains remain.

Watch

Extended reading notes

Core claim

QuARI's central claim is that per-query reshaping of gallery embeddings is enough to close much of the gap between generic vision-language features and expensive specialized reranking. Given a query embedding $q$, the query adaptation network produces a transformed query $q'$ and a rank-64 transformation $T = \sum_{j=1}^{r} u_j v_j^\top$, which is applied to every database embedding before cosine similarity search. Because $T$ is linear and low-rank, the adapted database can be scored with one matrix multiplication per image from precomputed features. The authors show that this consistently improves image-to-image and text-to-image retrieval on ILIAS and INQUIRE, that it beats static task adaptation by a wide margin, and that it outperforms open-source vision-language-model rerankers while using orders of magnitude less query-time computation.

Load-bearing premise

The load-bearing premise is that training on a random subset of a biodiversity dataset does not give QuARI meaningful prior exposure to the INQUIRE evaluation images, so the measured INQUIRE gains come from the query-adaptation mechanism rather than from having trained on the evaluation domain.

Editorial extensions

If this is right

  • Because the transform is linear and low-rank, a database of frozen features can be adapted per query with a single matrix multiplication per image, making it feasible to rerank millions of candidates rather than only a small top-k.
  • The INQUIRE results suggest that expert-domain text-to-image search over large natural-world collections can be improved by training a small hypernetwork on captioned biodiversity data, without fine-tuning the encoder.
  • The ILIAS results suggest that static task adaptation is not sufficient: the same learned projection applied to all queries leaves large gains on the table compared with query-specific projections.
  • Ablations attribute much of the gain to noise injection and semi-positive sampling; removing noise hurts image-to-image retrieval more than removing iterative generation or semi-positives.
  • QuARI is competitive with closed-source large vision-language rerankers on INQUIRE top-100 while costing orders of magnitude less per query.

Reading between the lines

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

  • Editorial inference: if the per-query projection mechanism is the real driver, the same hypernetwork recipe could be applied to any frozen retrieval embedding space, such as audio or video, by training only on paired examples from that domain and without re-encoding the gallery.
  • Editorial inference: the reported INQUIRE gains may partly reflect domain overlap, because the training set is drawn from a biodiversity dataset whose image source also underlies INQUIRE; a clean test is to train on a disjoint natural-world domain or to measure image-level duplicate overlap between training and evaluation sets.
  • Editorial inference: the semi-positive soft-target objective generalizes the standard one-positive-per-anchor contrastive loss and could be tested in other metric-learning settings where an anchor has multiple plausible matches.
  • Editorial inference: because the adaptation is linear, its ceiling is set by what the frozen features already encode; non-linear but still cheap query-driven feature gating is the natural next step, as the paper's limitation discussion also suggests.
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

3 major / 7 minor

Summary. The paper proposes QuARI, a transformer-based hypernetwork that takes a query embedding as input and predicts both a transformed query embedding and a low-rank (r=64) linear transformation of the gallery feature space. The transformation is applied to frozen CLIP/SigLIP2 embeddings so that retrieval can be performed per query without fine-tuning the backbone. Training uses a symmetric contrastive loss, augmented with LinCIR-style query noise and 'semi-positive' samples mined from backbone nearest neighbors. The method is evaluated on ILIAS (image-to-image and text-to-image, at 5M and 100M scale) and INQUIRE (expert text-to-image retrieval), with comparisons to frozen backbones, static task adaptation, and several re-ranking baselines, including large VLMs. The paper reports large improvements, ablation studies, t-SNE visualizations, and runtime comparisons for top-k re-ranking.

Significance. If the reported results hold, query-adaptive linear reshaping of frozen VLM features is a simple and potentially useful complement to expensive re-ranking: the method is architecture-agnostic, uses only global embeddings, and the ablation study in Table 4 indicates that each component (iterative refinement, semi-positives, query noise) contributes. The paper's strengths include consistent gains on ILIAS across multiple backbones, direct comparison to task adaptation, and a runtime comparison for re-ranking. However, the central text-to-image claim rests on INQUIRE, and the training set includes BioTrove, an iNaturalist-derived dataset, while INQUIRE is also built from iNaturalist; the absence of any overlap analysis is a serious external-validity concern. In addition, the full-corpus efficiency claim is not backed by measurements at 100M scale, and some numerical tables are internally inconsistent. These issues make the current evidence for the abstract's strong claims incomplete.

major comments (3)
  1. [Section 4, Training Datasets and Evaluation Datasets; Section 6, Limitations] The INQUIRE evaluation may not be external to the training distribution. Section 4 states that QuARI is trained on a random 5M subset of BioTrove, which is assembled from iNaturalist observations, while the INQUIRE benchmark is built on the iNaturalist 2024 dataset. The manuscript never reports image-level deduplication, near-duplicate analysis, species-level overlap, or any distributional comparison between the BioTrove training subset and the INQUIRE evaluation images. Because the same source platform and taxonomic annotation structure underlie both datasets, the large INQUIRE gains in Table 2 and Table 3b could reflect training on the evaluation domain (e.g., learning species-name-to-image subspaces) rather than a general query-adaptation mechanism. The ILIAS results provide some out-of-domain support for instance retrieval, but they do not validate the expert ecological text-to-image claim. The paper should either provide an overlap analysis (exact and near-duplicate, plus species-level overlap) or retrain QuARI without BioTrove/iNaturalist data and show that the INQUIRE gains persist. Section 6, which lists limitations, does not acknowledge this potential contamination.
  2. [Section 5.5 and Figure 4; Abstract and Section 1] The abstract claims the transformation can be applied 'with minimal computational cost to millions of image embeddings,' but no experiment measures the cost of applying QuARI to a full 100M gallery. Figure 4 reports time only for re-ranking tasks over 1,000 image pairs (ILIAS) and 100 text-image pairs (INQUIRE). For full-corpus retrieval, a per-query low-rank transformation with rank r=64 must be applied to every gallery embedding (or, equivalently, the query must be transformed and all embeddings re-scored in a way that accounts for per-image normalization after transformation), which is a fundamentally different and much larger cost than the top-k setting shown. The paper should report end-to-end latency and total FLOPs for the full 100M ILIAS task, including the cost of transforming or rescoring all candidate embeddings, and should temper the 'orders of magnitude faster' claim to the re-ranking setting where it is actually measured.
  3. [Tables 2, 3b, and 4] The INQUIRE numbers for the SigLIP2 backbone are inconsistent across tables. Table 2 lists only OAI CLIP ViT-B, OAI CLIP ViT-L, SigLIP ViT-L, and SigLIP SoViT-400m, with no SigLIP2 row. Table 4 reports a SigLIP2 baseline of 37.2 mAP@50 and a QuARI result of 50.7 mAP@50 on INQUIRE, while Table 3b reports SigLIP2 + QuARI as 45.6 mAP@50 on INQUIRE under 'Top-100 Re-ranking.' The paper should clarify which task each table reports, include all backbones in Table 2, and reconcile the 45.6 versus 50.7 discrepancy. As written, the reader cannot determine the actual SigLIP2 + QuARI INQUIRE performance, which is directly relevant to the central claim.
minor comments (7)
  1. [Section 4, Training Datasets] The text says QuARI is trained on a random subset of 5M BioTrove samples but then says Qwen2.5-VL is used to caption only a 500K subset of BioTrove. This leaves unclear what labels or captions are used for the remaining 4.5M samples, or whether the effective BioTrove training set is 500K rather than 5M. Please specify the exact training data composition and the number of samples per source.
  2. [Section 3.3, Eq. (9)-(10)] The notation for transformed database embeddings is confusing: the text defines d_j^i as the i-th target image embedding transformed by the j-th transformation, but Eq. (9) uses superscripts and subscripts inconsistently. The sentence 'We only consider similarities where...' also needs to be reconciled with the formula, particularly which transformation is applied to which image when computing S_{i,j} for the semi-positive term.
  3. [Section 4, Evaluation Datasets] Please clarify whether the ILIAS text-to-image results and the INQUIRE results use the same query text preprocessing, and whether ILIAS text queries are the official instance-name queries or the 'what is this?' style queries; the paper does not state this, which affects comparability with prior ILIAS baselines.
  4. [Tables 1-4] All main results are reported as single numbers with no variance or number of seeds. Given the abstract's phrase 'consistently outperforms,' at least one table with mean and standard deviation across three or more training runs would substantially strengthen the reproducibility claims.
  5. [Section 5.3, Table 4] The 'Fine-tuned SigLIP2' baseline is not described in enough detail: it is unclear what data it is fine-tuned on, for how many steps, and with what learning rate. Since the ablation's purpose is to show QuARI beats simple fine-tuning, these details are necessary for a fair comparison.
  6. [Section 5.4, Figure 3] The t-SNE visualizations are used to claim that ground-truth responses are 'mapped much closer' to the query, but t-SNE preserves local neighborhoods only approximately and distances in the embedding are not quantitatively meaningful. A quantitative nearest-neighbor distance or rank statistic before and after adaptation would be more convincing.
  7. [Section 3.3, Eq. (8)] The notation for the noise term, q_i <- q_i + U[0,1] × N[0,1], is ambiguous: it should state whether the multiplication is elementwise and whether the uniform and normal samples are per-dimension, and it should cite the exact LinCIR formulation being borrowed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: QuARI's per-query projections are learned functions evaluated on external benchmarks, not reduced to their inputs by construction.

full rationale

QuARI is an empirical learning method, not a derivational chain that assumes its own conclusion. The hypernetwork Hθ is trained with a contrastive loss on paired text–image data (Section 3.3), and the resulting query-specific linear projections are evaluated on held-out external benchmarks, ILIAS and INQUIRE (Section 4). There is no equation in which a reported prediction is equivalent to a fitted parameter by construction, and no load-bearing self-citation: the authors' own prior work appears only in a broader-impacts example (Ref. [35]) and is not used to justify the method. The known concern that BioTrove (an iNaturalist-derived dataset) may overlap with the iNaturalist-based INQUIRE evaluation set is a potential training-domain contamination or correctness risk, not a circularity of the derivation: the reported INQUIRE numbers could be inflated by in-domain training, but the method itself is not defined in terms of the benchmark metric, and the ILIAS results provide an independent external evaluation. Because no specific reduction of the claimed result to its inputs can be quoted, the appropriate circularity score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical entities. It relies on one strong domain assumption (sufficient information in frozen VLM features), several ad hoc design choices (zero-init tokens, semi-positive mining, query noise), and standard linear algebra and transformer machinery.

free parameters (6)
  • low-rank rank r = 64
    Sets expressiveness of the projection; chosen by hand in Section 3.2.
  • number of semi-positive samples = top 2 of 100 nearest neighbors
    Defines soft target structure in Section 3.3; chosen without sensitivity analysis.
  • query noise distribution = U[0,1] x N[0,1] (LinCIR)
    Borrowed from LinCIR as a modality-gap regularizer; its scale is not tuned or ablated independently.
  • contrastive temperature tau = 0.07
    Set in Appendix A; common default.
  • number of iterative refinement steps L = not specified
    Controls depth of transformation refinement in Section 3.2, but the paper never states its value.
  • transformer layers, batch size, learning rate = 4 layers, batch 320, cosine 1e-5 to 2e-7
    Training hyperparameters listed in Appendix A and Section 3.3.
assumptions (5)
  • domain assumption VLM embeddings contain enough fine-grained information for instance retrieval; a linear transform can expose it.
    Section 6 admits this assumption and notes it is the main limitation.
  • domain assumption Contrastive training on COCO, CC12M, and BioTrove transfers to ILIAS and INQUIRE.
    Section 3.3 and Section 4 assume these datasets are sufficient training signal.
  • ad hoc to paper Zero-initialized tokens and iterative transformer refinement converge to useful projections.
    Section 3.2 introduces this design without analysis of alternative initializations or convergence.
  • ad hoc to paper Semi-positive samples from backbone nearest neighbors improve generalization.
    Section 3.3 postulates this training addition; no theory or sensitivity analysis is given.
  • ad hoc to paper LinCIR-style query noise bridges the text-image modality gap.
    Section 3.3 borrows this augmentation; the ablation shows it is critical, but its distribution is not derived.

how reviews work

0 comments
Cite this review

Pith. "Pith review of QuARI: Query Adaptive Retrieval Improvement." pith.science (2026). https://pith.science/paper/N42VA5NJ

@misc{pith2026250521647,
  author       = {Pith},
  title        = {Pith review of: QuARI: Query Adaptive Retrieval Improvement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N42VA5NJ}},
  note         = {Machine review of arXiv:2505.21647}
}
read the original abstract

Massive-scale pretraining has made vision-language models increasingly popular for image-to-image and text-to-image retrieval across a broad collection of domains. However, these models do not perform well when used for challenging retrieval tasks, such as instance retrieval in very large-scale image collections. Recent work has shown that linear transformations of VLM features trained for instance retrieval can improve performance by emphasizing subspaces that relate to the domain of interest. In this paper, we explore a more extreme version of this specialization by learning to map a given query to a query-specific feature space transformation. Because this transformation is linear, it can be applied with minimal computational cost to millions of image embeddings, making it effective for large-scale retrieval or re-ranking. Results show that this method consistently outperforms state-of-the-art alternatives, including those that require many orders of magnitude more computation at query time.

Figures

Figures reproduced from arXiv: 2505.21647 by the authors.

Figure 1
Figure 1. We propose a new query-specific approach to retrieval, QuARI. QuARI dynamically [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. An overview of our query adaptation network. A zero initialization of the transformation [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. t-SNE visualizations comparing original features and QuARI features. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Comparison of re-ranking performance and inference cost for image-to-image retrieval on [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 36 canonical work pages

  1. [1]

    Zico Kolter

    Victor Akinwande, Mohammad Sadegh Norouzzadeh, Devin Willmott, Anna Bair, Madan Ravi Ganesh, and J. Zico Kolter. Hyperclip: Adapting vision-language models with hypernetworks,

    work page

  2. [2]

    Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization, 2016. URL https://arxiv.org/abs/1607.06450

    arXiv 2016

  3. [3]

    Recognition in terra incognita

    Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. InProceedings of the European Conference on Computer Vision (ECCV), September 2018

    work page 2018

  4. [4]

    Pinto, Alexander Kolesnikov, Xiao Wang, et al

    Lucas Beyer, Andreas Steiner, André S. Pinto, Alexander Kolesnikov, Xiao Wang, et al. Paligemma: A versatile 3b vision–language model for transfer. Technical report, Google Research, 2024. URLhttps://ai.google.dev/gemma/docs/paligemma

    work page 2024

  5. [5]

    Gritsenko, Matthias Minderer, Charles Blundell, Razvan Pascanu, and Jovana Mitrovi´c

    Ioana Bica, Anastasija Ili´c, Matthias Bauer, Goker Erdogan, Matko Bošnjak, Christos Kaplanis, Alexey A. Gritsenko, Matthias Minderer, Charles Blundell, Razvan Pascanu, and Jovana Mitrovi´c. Improving fine-grained understanding in image-text pre-training, 2024. URL https: //arxiv.org/abs/2401.09865

    arXiv 2024

  6. [6]

    Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts

    Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558–3568, 2021. 10

    work page 2021

  7. [7]

    Transformers as meta-learners for implicit neural representa- tions

    Yinbo Chen and Xiaolong Wang. Transformers as meta-learners for implicit neural representa- tions. InEuropean Conference on Computer Vision, 2022

    work page 2022

  8. [8]

    Total recall: Automatic query expansion with a generative feature model for object retrieval

    Ondˇrej Chum, James Philbin, Josef Sivic, Michael Isard, and Andrew Zisserman. Total recall: Automatic query expansion with a generative feature model for object retrieval. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1–8,

    work page

Show all 51 references
  1. [9]

    Wenliang Dai, Junnan Li, Dongxu Li, Anthony M. H. Tiong, Junqi Zhao, et al. Instructblip: Towards general-purpose vision-language models with instruction tuning. InAdvances in Neural Information Processing Systems (NeurIPS, pages 17246–17262, 2023

    2023

  2. [10]

    Detecting mis- information in photos utilizing reverse image search

    Vinh Dang, Thanh-Son Nguyen, Minh-Triet Tran, and Duc-Tien Dang-Nguyen. Detecting mis- information in photos utilizing reverse image search. InProceedings of the 2024 International Conference on Multimedia Retrieval, pages 1321–1323, 2024

    2024

  3. [11]

    Language-only training of zero-shot composed image retrieval

    Geonmo Gu, Sanghyuk Chun, Wonjae Kim, Yoohoon Kang, and Sangdoo Yun. Language-only training of zero-shot composed image retrieval. InConference on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  4. [12]

    David Ha, Andrew Dai, and Quoc V . Le. Hypernetworks.arXiv preprint arXiv:1609.09106, 2016

    2016 arXiv

  5. [13]

    OpenCLIP

    Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. OpenCLIP. https://doi.org/10.5281/zenodo.5143773, July 2021

    doi

  6. [14]

    Colbert: Efficient and effective passage search via contextu- alized late interaction over bert

    Omar Khattab and Matei Zaharia. Colbert: Efficient and effective passage search via contextu- alized late interaction over bert. InSIGIR, 2020

    2020

  7. [15]

    Region-centric image–language pretraining for open-vocabulary detection

    Dahun Kim, Anelia Angelova, and Weicheng Kuo. Region-centric image–language pretraining for open-vocabulary detection. InProceedings of the European Conference on Computer Vision (ECCV), 2024

    2024

  8. [16]

    Ilias: Instance-level image retrieval at scale

    Giorgos Kordopatis-Zilos, Vladan Stojni ´c, Anna Manko, Pavel Šuma, Nikolaos-Antonios Ypsilantis, Nikos Efthymiadis, Zakaria Laskar, Jiˇrí Matas, Ondˇrej Chum, and Giorgos Tolias. Ilias: Instance-level image retrieval at scale. InProceedings of the IEEE/CVF Conference on Compu...

    2025

  9. [17]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InProceedings of the International Conference on Machine Learning (ICML), volume 202, pages 10439–10460, 2023

    2023

  10. [18]

    Vila: On pre-training for visual language models

    Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23178–2...

    2024

  11. [19]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. InProceedings of the European Conference on Computer Vision (ECCV), pages 740–755. Springer, 2014

    2014

  12. [20]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems (NeurIPS), pages 16112–16128, 2023

    2023

  13. [21]

    Improved baselines with visual instruction tuning

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22926–22936, 2024

    2024

  14. [22]

    Llava -1.6: Release notes and model card

    Haotian Liu, Chunyuan Li, and the LLaV A Team. Llava -1.6: Release notes and model card. https://llava-vl.github.io/blog/2024-01-30-llava-next/ , 2025. Accessed 14 May 2025. 11

    2024

  15. [23]

    Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

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

    2017 arXiv

  16. [24]

    Mini-model adaptation: Efficiently extending pretrained models to new languages via aligned shallow training

    Kelly Marchisio, Patrick Lewis, Yihong Chen, and Mikel Artetxe. Mini-model adaptation: Efficiently extending pretrained models to new languages via aligned shallow training. In Findings of the Association for Computational Linguistics: ACL 2023, pages 5474–5490, 2023

    2023

  17. [25]

    Gpt-4 technical report

    OpenAI. Gpt-4 technical report. Technical report, OpenAI, 2023. URL https://arxiv.org/ abs/2303.08774. Describes the multimodal GPT-4V variant

    2023 arXiv

  18. [26]

    Introducing gpt -4o

    OpenAI. Introducing gpt -4o. https://openai.com/index/ gpt-4o-and-more-tools-to-chatgpt-free/, 2025. Official launch post

    2025

  19. [27]

    Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas...

    2023 arXiv

  20. [28]

    Object retrieval with large vocabularies and fast spatial matching

    James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Object retrieval with large vocabularies and fast spatial matching. In2007 IEEE conference on computer vision and pattern recognition, pages 1–8. IEEE, 2007

    2007

  21. [29]

    Rerankers and two-stage retrieval

    Pinecone. Rerankers and two-stage retrieval. https://www.pinecone.io/learn/series/ rag/rerankers/, 2023

    2023

  22. [30]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, and et al. Learning transferable visual models from natural language supervision. InICML, 2021

    2021

  23. [31]

    Ape: Aligning pretrained encoders to quickly learn aligned multimodal representations

    Elan Rosenfeld, Preetum Nakkiran, Hadi Pouransari, Oncel Tuzel, and Fartash Faghri. Ape: Aligning pretrained encoders to quickly learn aligned multimodal representations. InAdvances in Neural Information Processing Systems (NeurIPS), 2022. URL https://arxiv.org/abs/ 2210.03927

    2022 arXiv

  24. [32]

    Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, and et al. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  25. [33]

    Global features are all you need for image retrieval and reranking

    Shihao Shao, Kaifeng Chen, Arjun Karpur, Qinghua Cui, André Araujo, and Bingyi Cao. Global features are all you need for image retrieval and reranking. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11036–11046, 2023

    2023

  26. [34]

    CNN features off-the-shelf: an astounding baseline for recognition

    Ali Sharif Razavian, Hossein Azizpour, Josephine Sullivan, and Stefan Carlsson. CNN features off-the-shelf: an astounding baseline for recognition. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 806–813, 2014

    2014

  27. [35]

    Hotels-50k: A global hotel recognition dataset

    Abby Stylianou, Hong Xuan, Maya Shende, Jonathan Brandt, Richard Souvenir, and Robert Pless. Hotels-50k: A global hotel recognition dataset. InThe AAAI Conference on Artificial Intelligence (AAAI), January 2019

    2019

  28. [36]

    Ames: Asymmetric and memory-efficient similarity estimation for instance-level retrieval

    Pavel Suma, Giorgos Kordopatis-Zilos, Ahmet Iscen, and Giorgos Tolias. Ames: Asymmetric and memory-efficient similarity estimation for instance-level retrieval. InEuropean Conference on Computer Vision (ECCV), 2024

    2024

  29. [37]

    Instance-level image retrieval using reranking transformers

    Fuwen Tan, Jiangbo Yuan, and Vicente Ordonez. Instance-level image retrieval using reranking transformers. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021

    2021

  30. [38]

    Qwen2.5-vl, January 2025

    Qwen Team. Qwen2.5-vl, January 2025. URL https://qwenlm.github.io/blog/qwen2. 5-vl/. 12

    2025

  31. [39]

    Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li

    Bart Thomee, David A. Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: the new data in multimedia research.Commun. ACM, 59(2):64–73, January 2016. ISSN 0001-0782. doi: 10.1145/2812802. URL https: //doi.org/10.1145/2812802

    2016 doi

  32. [40]

    Siglip 2: Multilingual vision- language encoders with improved semantic understanding, localization, and dense features,

    Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alab- dulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, Olivier Hénaff, Jeremiah Harmsen, Andreas Steiner, and Xiaohua Zhai. Siglip 2: Multilingual vision- language ...

  33. [41]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. InProceedings of the 31st International Conference on Neural Information Processing Systems, NIPS’17, page 6000–6010, Red Ho...

    2017

  34. [42]

    Inquire: A natural world text-to-image retrieval benchmark.Advances in Neural Information Processing Systems, 37:126500–126514, 2024

    Edward Vendrow, Omiros Pantazis, Alexander Shepard, Gabriel Brostow, Kate Jones, Oisin Mac Aodha, Sara Beery, and Grant Van Horn. Inquire: A natural world text-to-image retrieval benchmark.Advances in Neural Information Processing Systems, 37:126500–126514, 2024

    2024

  35. [43]

    Density-aware chamfer distance as a comprehensive metric for point cloud completion

    Tong Wu, Liang Pan, Junzhe Zhang, Tai WANG, Ziwei Liu, and Dahua Lin. Density-aware chamfer distance as a comprehensive metric for point cloud completion. InIn Advances in Neural Information Processing Systems (NeurIPS), 2021

    2021

  36. [44]

    Biotrove: A large curated image dataset enabling ai for biodiversity.Advances in Neural Information Processing Systems, 37:102101–102120, 2024

    Chih-Hsuan Yang, Benjamin Feuer, Talukder Jubery, Zi Deng, Andre Nakkab, Md Zahid Hasan, Shivani Chiranjeevi, Kelly Marshall, Nirmal Baishnab, Asheesh Singh, et al. Biotrove: A large curated image dataset enabling ai for biodiversity.Advances in Neural Information Processing S...

    2024

  37. [45]

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

    Lewei Yao, Runhui Huang, Lu Hou, and et al. Filip: Fine-grained interactive language–image pre-training. InICLR, 2022

    2022

  38. [46]

    Towards universal image embeddings: A large-scale dataset and challenge for generic image representations

    Nikolaos-Antonios Ypsilantis, Kaifeng Chen, Bingyi Cao, Mário Lipovský, Pelin Dogan- Schönberger, Grzegorz Makosa, Boris Bluntschli, Mojtaba Seyedhosseini, Ondˇrej Chum, and André Araujo. Towards universal image embeddings: A large-scale dataset and challenge for generic image...

    2023

  39. [47]

    Sigmoid loss for language image pre-training

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11975–11986, October 2023

    2023

  40. [48]

    Regionclip: Region-based language-image pretraining

    Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. Regionclip: Region-based language-image pretraining. InAdvances in Neural Information Processing Systems (NeurIPS), 2021. 13

    2021

  41. [2007]

    URLhttps://api.semanticscholar.org/CorpusID:570516

  42. [2024]

    URLhttps://arxiv.org/abs/2412.16777

    arXiv

  43. [2025]

    URLhttps://arxiv.org/abs/2502.14786

    arXiv

Pith tools

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