Pith. sign in

REVIEW 2 major objections 5 minor 61 references

Object-aware token merging cuts visual tokens by over 93% while improving multi-vector image retrieval.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-11 16:32 UTC pith:OZHUOGCX

load-bearing objection Solid systems paper: object-instance merge prior for late-interaction VL indexes, with real storage wins and controlled gains over adapted pooling/pruning. the 2 major comments →

arxiv 2607.04605 v1 pith:OZHUOGCX submitted 2026-07-06 cs.IR cs.AIcs.CLcs.CV

Do All Visual Tokens Matter Equally? Object-Evidence Preserving Token Merging for Vision-Language Retrieval

Suhyeong Park , Junha Jung , Jungwoo Park , Jaewoo Kang This is my paper
classification cs.IR cs.AIcs.CLcs.CV
keywords multi-vector retrievallate interactiontoken mergingobject-aware compressionvision-language retrievalMaxSimColPaliphrase grounding
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Multi-vector vision-language retrieval keeps many image tokens so query words can match objects, attributes, and regions by maximum similarity. That design is accurate but expensive to store and score. The paper argues that ordinary pruning or feature pooling can erase or collapse the very object-level evidence later queries need. SaMer merges post-projection image tokens into a small set of centroids, guided at training time by object annotations that discourage mixing different instances, then adapts only the shared projection layer. At test time no boxes or detectors are required. With 64 tokens the method removes most image-side vectors, shrinks ColPali storage by about 16 times, and raises recall on Flickr30K and MSCOCO while improving phrase grounding. The claim is that efficient late-interaction retrieval depends less on raw token count than on keeping query-selectable object evidence intact.

Core claim

Aggressive compression of image-side multi-vector tokens need not hurt—and can improve—late-interaction retrieval when merging is made object-aware: training-time object labels act only as a merge prior that discourages cross-instance collapse, the original MaxSim interface is kept, and only the shared projection is adapted, so that K=64 centroids preserve the evidence future query tokens select.

What carries the argument

SaMer: feature-spatial soft assignment of post-projector visual tokens into K normalized centroids, with an object-instance inconsistency penalty used only during training to reshape assignment weights, followed by projection-only adaptation under compressed MaxSim scoring.

Load-bearing premise

That a training-only object-instance merge prior learned from one dataset’s boxes and reflected only through the projection layer will, without any boxes at test time, keep the right MaxSim-selectable evidence on other natural and compositional images.

What would settle it

Run the same K=64 projection-only adaptation on Flickr30K-Entities but drop the object-instance penalty (feature-spatial merging only); if R@1 and BoxMass on Flickr30K and MSCOCO then fall to the level of H-Pool or SAP, the object-aware prior is not doing the claimed work.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper proposes SaMer, an object-aware post-projector token merging method for multi-vector vision-language retrieval (ColPali/ColQwen2-style MaxSim). It compresses N image tokens into K soft feature-spatial centroids while keeping the late-interaction interface unchanged. During training only, Flickr30K-Entities box labels define an instance-inconsistency penalty that reshapes soft assignment to discourage cross-instance collapse; at inference the method is annotation-free and uses only the adapted shared projection layer with frozen backbones. With K=64 the authors report >93% token reduction, 16.09× ColPali storage reduction, R@1 gains on Flickr30K (77.0→82.4) and MSCOCO (47.4→51.6), competitive ImageCoDe results, and improved phrase-level grounding (BoxMass, RegionHit, CoverageIoU) relative to pruning/pooling baselines under matched budgets and adaptation.

Significance. If the results hold, the paper makes a useful systems contribution to multi-vector VL retrieval: it reframes compression as evidence preservation rather than pure token reduction, shows that a training-only object-instance merge prior plus projection-only adaptation can improve both storage/scoring cost and R@1, and supplies controlled ablations (matched-budget adapted baselines, K-sweeps, merge components, centroid variants, merging-vs-pruning) plus grounding metrics that go beyond retrieval accuracy alone. The method is practically attractive because it preserves the original MaxSim interface, needs no detector at inference, and freezes the heavy backbones. Code is promised. These strengths matter for large-scale image indexes and retrieval-augmented VQA where image-side multi-vector storage is the bottleneck.

major comments (2)
  1. The central transfer claim (Methods §4.2–4.3; adaptation protocol §5.2) rests on a training-only Flickr30K-Entities instance prior that is never re-applied at inference. While MSCOCO and ImageCoDe gains and training-free SaMer vs. pruning (Table 7) give supporting evidence, the manuscript should more explicitly quantify how much of the adapted gain is prior-driven versus generic projection fine-tuning under the same merge rule. Table 2 already matches adaptation across compressors; a short additional control that adapts SaMer without the object penalty (or with shuffled instance labels) would make the load-bearing causal claim tighter.
  2. DocVQA is correctly labeled a boundary case (§5.3), but the abstract and conclusion still emphasize broad efficiency for multi-vector VL retrieval. Given that SaMer underperforms full ColPali/ColQwen2 on DocVQA R@1 and is not designed for sparse OCR/layout evidence, the claims should more carefully scope the operating regime to natural-image / object-centric retrieval rather than implying domain-general multi-vector compression.
minor comments (5)
  1. Eq. (2)–(6): γ, τs, and the stop-gradient hard assignment ci are introduced without a sensitivity study in the main text; a short appendix note on default values and robustness would help reproducibility.
  2. Figure 2 and Table 3: the K-budget curves are informative; adding error bars or multi-seed variance for the main K=64 setting would strengthen confidence in the reported R@1 deltas.
  3. Grounding metric definitions (Appendix C) are clear, but the main text should briefly state how the phrase relevance map ŒS is obtained from MaxSim (query-token vs. merged-token similarities) so readers need not reverse-engineer it from the appendix.
  4. Related work §2.2 cites concurrent retrieval-side compressors (HPC, SAP, H-Pool); ensure citation versions and method-specific settings for K=64 are fully aligned with Table 1 footnotes.
  5. Minor wording: abstract and intro repeat the 93%/16.09×/R@1 numbers almost verbatim; a slightly tighter abstract would free space for the DocVQA scope caveat.

Circularity Check

0 steps flagged

No significant circularity: empirical compression method with held-out retrieval/grounding evaluation; claims do not reduce by construction to fitted inputs.

full rationale

SaMer is a systems paper whose load-bearing claims are empirical: at K=64, object-aware soft merging plus projection-only adaptation improves R@1/nDCG and phrase grounding vs matched-budget pruning/pooling baselines while cutting image-side tokens and MaxSim cost. The merge prior (Eqs. 5–6) uses training-time instance labels only to reshape soft assignment under InfoNCE (Eq. 7); inference is bbox-free feature-spatial assignment. Storage ratios (e.g., 1030→64 ≈ 16.09×) are arithmetic consequences of the chosen budget, not predictions of independent observables. Grounding metrics (BoxMass, RegionHit, CoverageIoU) are post-hoc diagnostics, not the training objective. Adaptation is on Flickr30K-Entities train; evaluation includes held-out Flickr30K, cross-dataset MSCOCO, ImageCoDe, and DocVQA against external baselines (H-Pool, SAP, HPC, full ColPali/ColQwen2). No uniqueness theorem, self-citation chain, or ansatz is load-bearing for the central result. Ordinary method design (soft centroids, γ, τs, K) does not make the reported gains true by definition.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The claim rests on standard late-interaction retrieval plus a small set of design choices: soft feature-spatial centroids, a training-only instance-inconsistency penalty from boxes, and projection-only InfoNCE adaptation. Free parameters are ordinary method knobs (K, temperatures, γ, lr). No new physical entities; the invented construct is the object-aware merge prior itself.

free parameters (4)
  • K (merged token budget) = 64
    Primary operating point chosen as K=64 after budget sweep; central storage and accuracy claims are reported at this value.
  • γ (spatial coherence weight)
    Balances feature vs spatial terms in assignment distance; chosen by authors and not exhaustively justified as unique.
  • τs (soft-assignment temperature)
    Controls softness of token-to-centroid assignment; affects how much cross-instance mixing the prior can suppress.
  • InfoNCE temperature τ and projection learning rate = lr=2e-4
    Adaptation hyperparameters (lr 2e-4, weight decay 1e-5, cosine schedule) that shape the reported adapted gains.
axioms (4)
  • domain assumption MaxSim late interaction over token embeddings is the correct retrieval interface to preserve under compression.
    Preliminaries §3.1 and scoring Eqs. 1 and 4; the entire method is designed to keep this interface unchanged.
  • ad hoc to paper Object-instance labels derived from spatial overlap with training boxes form a useful merge prior that discourages harmful cross-instance collapse for future MaxSim queries.
    §4.2 defines P_inst from hard assignment and box labels; this is the paper-specific inductive bias, not a standard theorem.
  • domain assumption Adapting only the shared projection layer with frozen backbones is sufficient to realign compressed centroids with the retrieval space.
    §4.3; gains after FT vs w/o FT and vs other adapted compressors rest on this choice.
  • domain assumption Soft weighted centroids preserve selectable evidence better than hard mean/medoid or pure pruning under the same K.
    Supported empirically in Appendix Table 6 and merging-vs-pruning Table 7, treated as design premise in the main method.
invented entities (2)
  • Object-aware merge prior P_inst(i,k) no independent evidence
    purpose: Penalize soft assignments that mix different object instances into the same centroid during training.
    Defined in Eq. 5 from stop-gradient hard assignment and box-label histograms; not an external physical object but a new algorithmic construct central to the claim.
  • SaMer compressed representation R(I) of K soft centroids no independent evidence
    purpose: Replace N post-projector tokens in the multi-vector index while preserving MaxSim scoring.
    The operational output of the method; independent evidence is only the paper's own retrieval/grounding experiments.

pith-pipeline@v1.1.0-grok45 · 24431 in / 3053 out tokens · 35046 ms · 2026-07-11T16:32:06.435348+00:00 · methodology

0 comments
read the original abstract

Multi-vector vision-language retrieval preserves fine-grained visual evidence through maximum-similarity late interaction, but dense image-side tokens make storage and scoring expensive. Existing token compression methods reduce this cost, yet they can remove or collapse object- and region-level evidence that future query tokens may need to select. We propose SaMer, an object-aware token merging framework that compresses image-side post-projector tokens into $K$ representative centroids while preserving the original late-interaction interface. SaMer uses object annotations only during training as a merge prior to discourage cross-instance mixing, requires no ground-truth bounding boxes or detectors at inference time, and adapts only the shared projection layer with frozen vision and language backbones. With $K=64$, SaMer removes more than 93% of image-side tokens and reduces ColPali storage by $16.09\times$, while improving R@1 on Flickr30K and MSCOCO. These gains arise because object-aware merging preserves query-selectable object evidence that pruning or feature-only pooling can remove or collapse. SaMer also outperforms compression baselines and shows stronger phrase-level grounding, suggesting that efficient multi-vector retrieval depends not only on reducing token count, but on preserving the evidence future query tokens need to select.

Figures

Figures reproduced from arXiv: 2607.04605 by Jaewoo Kang, Jungwoo Park, Junha Jung, Suhyeong Park.

Figure 1
Figure 1. Figure 1: Overview of SaMer. Frozen vision and language encoders produce hidden states, a shared trainable projection layer [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Token budget and R@5 trade-off. Solid lines show [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Grounding comparison between full ColPali, com [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative grounding examples. Red denotes the phrase-relevance map and green denotes the ground-truth box for [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

61 extracted references · 16 linked inside Pith

  1. [1]

    Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

    ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT , author=. Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

  2. [2]

    Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    ColBERTv2: Effective and Efficient Retrieval via Lightweight Late Interaction , author=. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    2022

  3. [3]

    arXiv preprint arXiv:2407.01449 , year=

    ColPali: Efficient Document Retrieval with Vision Language Models , author=. arXiv preprint arXiv:2407.01449 , year=

    Pith/arXiv arXiv

  4. [4]

    arXiv preprint arXiv:2409.12191 , year=

    Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution , author=. arXiv preprint arXiv:2409.12191 , year=

    Pith/arXiv arXiv

  5. [5]

    International Conference on Learning Representations , year=

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale , author=. International Conference on Learning Representations , year=

  6. [6]

    arXiv preprint arXiv:2506.04997 , year=

    Towards Storage-Efficient Visual Document Retrieval: An Empirical Study on Reducing Patch-Level Embeddings , author=. arXiv preprint arXiv:2506.04997 , year=

    Pith/arXiv arXiv

  7. [7]

    International conference on machine learning , pages=

    Scaling up visual and vision-language representation learning with noisy text supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  8. [8]

    Advances in neural information processing systems , volume=

    Align before fuse: Vision and language representation learning with momentum distillation , author=. Advances in neural information processing systems , volume=

  9. [9]

    International conference on machine learning , pages=

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation , author=. International conference on machine learning , pages=. 2022 , organization=

    2022

  10. [10]

    International conference on machine learning , pages=

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models , author=. International conference on machine learning , pages=. 2023 , organization=

    2023

  11. [11]

    Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    COIL: Revisit exact lexical match in information retrieval with contextualized inverted list , author=. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    2021

  12. [12]

    Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

    Splade: Sparse lexical and expansion model for first stage ranking , author=. Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

  13. [13]

    arXiv preprint arXiv:2109.10086 , year=

    SPLADE v2: Sparse lexical and expansion model for information retrieval , author=. arXiv preprint arXiv:2109.10086 , year=

    Pith/arXiv arXiv

  14. [14]

    Proceedings of the 31st ACM International Conference on Information & Knowledge Management , pages=

    PLAID: an efficient engine for late interaction retrieval , author=. Proceedings of the 31st ACM International Conference on Information & Knowledge Management , pages=

  15. [15]

    Proceedings of the 48th international ACM SIGIR conference on research and development in information retrieval , pages=

    WARP: An efficient engine for multi-vector retrieval , author=. Proceedings of the 48th international ACM SIGIR conference on research and development in information retrieval , pages=

  16. [16]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Llava-prumerge: Adaptive token reduction for efficient large multimodal models , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  17. [17]

    Findings of the Association for Computational Linguistics: NAACL 2025 , pages=

    Lvpruning: An effective yet simple language-guided vision token pruning approach for multi-modal large language models , author=. Findings of the Association for Computational Linguistics: NAACL 2025 , pages=

    2025

  18. [18]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Visionzip: Longer is better but not necessary in vision language models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  19. [19]

    arXiv preprint arXiv:2410.08584 , year=

    Zipvl: Efficient large vision-language models with dynamic token sparsification , author=. arXiv preprint arXiv:2410.08584 , year=

    Pith/arXiv arXiv

  20. [20]

    arXiv preprint arXiv:2605.13178 , year=

    CLIP Tricks You: Training-free Token Pruning for Efficient Pixel Grounding in Large VIsion-Language Models , author=. arXiv preprint arXiv:2605.13178 , year=

    Pith/arXiv arXiv

  21. [21]

    The Eleventh International Conference on Learning Representations , year=

    Token Merging: Your ViT But Faster , author=. The Eleventh International Conference on Learning Representations , year=

  22. [22]

    arXiv preprint arXiv:2511.21631 , year=

    Qwen3-vl technical report , author=. arXiv preprint arXiv:2511.21631 , year=

    Pith/arXiv arXiv

  23. [23]

    Transactions of the Association for Computational Linguistics , volume=

    From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions , author=. Transactions of the Association for Computational Linguistics , volume=. 2014 , publisher=

    2014

  24. [24]

    arXiv preprint arXiv:1505.04870 , year =

    Flickr30k Entities: Collecting Region-to-Phrase Correspondences for Richer Image-to-Sentence Models , author =. arXiv preprint arXiv:1505.04870 , year =

    Pith/arXiv arXiv

  25. [25]

    European Conference on Computer Vision , year =

    Microsoft COCO: Common Objects in Context , author =. European Conference on Computer Vision , year =

  26. [26]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , year =

    DocVQA: A Dataset for VQA on Document Images , author =. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , year =

  27. [27]

    arXiv preprint arXiv:2203.15867 , year =

    Image Retrieval from Contextual Descriptions , author =. arXiv preprint arXiv:2203.15867 , year =

    Pith/arXiv arXiv

  28. [28]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

    GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

  29. [29]

    European Conference on Computer Vision , year =

    A-OKVQA: A Benchmark for Visual Question Answering using World Knowledge , author =. European Conference on Computer Vision , year =

  30. [30]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  31. [31]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Cherti, Mehdi and Beaumont, Romain and Wightman, Ross and Wortsman, Mitchell and Ilharco, Gabriel and Gordon, Cade and Schuhmann, Christoph and Schmidt, Ludwig and Jitsev, Jenia , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2023 , pages =

    2023

  32. [32]

    Demystifying

    Hu Xu and Saining Xie and Xiaoqing Tan and Po-Yao Huang and Russell Howes and Vasu Sharma and Shang-Wen Li and Gargi Ghosh and Luke Zettlemoyer and Christoph Feichtenhofer , booktitle=. Demystifying. 2024 , url=

    2024

  33. [33]

    arXiv preprint arXiv:2303.15389 , year =

    EVA-CLIP: Improved Training Techniques for CLIP at Scale , author =. arXiv preprint arXiv:2303.15389 , year =

    Pith/arXiv arXiv

  34. [34]

    arXiv preprint arXiv:2309.17425 , year =

    Data Filtering Networks , author =. arXiv preprint arXiv:2309.17425 , year =

    Pith/arXiv arXiv

  35. [35]

    arXiv preprint arXiv:2502.14786 , year =

    SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features , author =. arXiv preprint arXiv:2502.14786 , year =

    Pith/arXiv arXiv

  36. [36]

    2026 , url=

    Rui Meng and Ziyan Jiang and Ye Liu and Mingyi Su and Xinyi Yang and Yuepeng Fu and Can Qin and Raghuveer Thirukovalluru and Xuan Zhang and Zeyuan Chen and Ran Xu and Caiming Xiong and Yingbo Zhou and Wenhu Chen and Semih Yavuz , journal=. 2026 , url=

    2026

  37. [37]

    arXiv preprint arXiv:2412.16855 , year =

    GME: Improving Universal Multimodal Retrieval by Multimodal LLMs , author =. arXiv preprint arXiv:2412.16855 , year =

    Pith/arXiv arXiv

  38. [38]

    Shi Yu and Chaoyue Tang and Bokai Xu and Junbo Cui and Junhao Ran and Yukun Yan and Zhenghao Liu and Shuo Wang and Xu Han and Zhiyuan Liu and Maosong Sun , booktitle=. Vis. 2025 , url=

    2025

  39. [39]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Attention correctness in neural image captioning , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  40. [40]

    arXiv preprint arXiv:2404.19128 , year=

    Q-GroundCAM: Quantifying Grounding in Vision Language Models via GradCAM , author=. arXiv preprint arXiv:2404.19128 , year=

    Pith/arXiv arXiv

  41. [41]

    International Journal of Computer Vision , volume=

    Top-down neural attention by excitation backprop , author=. International Journal of Computer Vision , volume=. 2018 , publisher=

    2018

  42. [42]

    European Conference on Computer Vision , pages=

    Grounding of textual phrases in images by reconstruction , author=. European Conference on Computer Vision , pages=. 2016 , organization=

    2016

  43. [43]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Learning deep features for discriminative localization , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  44. [44]

    International journal of computer vision , volume=

    Grad-CAM: visual explanations from deep networks via gradient-based localization , author=. International journal of computer vision , volume=. 2020 , publisher=

    2020

  45. [45]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics , pages=

    PuMer: Pruning and Merging Tokens for Efficient Vision Language Models , author=. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics , pages=

  46. [46]

    Advances in Neural Information Processing Systems , year=

    TokenLearner: What Can 8 Learned Tokens Do for Images and Videos? , author=. Advances in Neural Information Processing Systems , year=

  47. [47]

    Advances in Neural Information Processing Systems , volume=

    DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsification , author=. Advances in Neural Information Processing Systems , volume=

  48. [48]

    International Conference on Learning Representations , year=

    Not All Patches are What You Need: Expediting Vision Transformers via Token Reorganizations , author=. International Conference on Learning Representations , year=

  49. [49]

    European Conference on Computer Vision , year=

    Adaptive Token Sampling For Efficient Vision Transformers , author=. European Conference on Computer Vision , year=

  50. [50]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , year=

    Token Pooling in Vision Transformers for Image Classification , author=. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , year=

  51. [51]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Patch Slimming for Efficient Vision Transformers , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  52. [52]

    European Conference on Computer Vision , pages=

    Stacked Cross Attention for Image-Text Matching , author=. European Conference on Computer Vision , pages=

  53. [53]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Visual Semantic Reasoning for Image-Text Matching , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  54. [54]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Similarity Reasoning and Filtration for Image-Text Matching , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  55. [55]

    Towards Storage-Efficient Visual Document Retrieval: An Empirical Study on Reducing Patch-Level Embeddings

    Ma, Yubo and Li, Jinsong and Zang, Yuhang and Wu, Xiaobao and Dong, Xiaoyi and Zhang, Pan and Cao, Yuhang and Duan, Haodong and Wang, Jiaqi and Cao, Yixin and Sun, Aixin. Towards Storage-Efficient Visual Document Retrieval: An Empirical Study on Reducing Patch-Level Embeddings. Findings of the Association for Computational Linguistics: ACL 2025. 2025. doi...

    doi:10.18653/v1/2025.findings-acl.1003 2025

  56. [56]

    arXiv preprint arXiv:2506.21601 , year=

    Hierarchical Patch Compression for ColPali: Efficient Multi-Vector Document Retrieval with Dynamic Pruning and Quantization , author=. arXiv preprint arXiv:2506.21601 , year=

    Pith/arXiv arXiv

  57. [57]

    2026 , eprint=

    Structural Anchor Pruning: Training-Free Multi-Vector Compression for Visual Document Retrieval , author=. 2026 , eprint=

    2026

  58. [58]

    arXiv preprint arXiv:2602.21202 , year=

    Multi-Vector Index Compression in Any Modality , author=. arXiv preprint arXiv:2602.21202 , year=

    arXiv

  59. [59]

    arXiv preprint arXiv:1807.03748 , year=

    Representation learning with contrastive predictive coding , author=. arXiv preprint arXiv:1807.03748 , year=

    Pith/arXiv arXiv

  60. [60]

    Advances in Neural Information Processing Systems , volume=

    Learning to merge tokens via decoupled embedding for efficient vision transformers , author=. Advances in Neural Information Processing Systems , volume=

  61. [61]

    2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=

    Tosa: Token merging with spatial awareness , author=. 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=. 2025 , organization=

    2025