Pith. sign in

REVIEW 3 major objections 4 minor 73 references

Generative Embedding Benchmark: How Much Information Survives in a Dense Embedding?

T0 review · 3 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read This paper proposes the Generative Embedding Benchmark (GEB), which scores dense embeddings by how much answer-relevant content a decoder can recover from them, arguing that separability-based benchmarks miss information bottlenecks.

desk verdict A genuinely useful benchmark idea and careful dataset, but the Qwen-family decoder makes the cross-model rankings—and part of the category story—untrustworthy as-is. read the letter →

arxiv 2608.06972 v1 pith:NFM6ZZFK submitted 2026-08-07 cs.CV

Yun Li , Biao Yang , Peixi Wu , Yunhao Zhou , Mingzhou Jiang , Wei Yuan , Fan Yang , Wenwu Ou This is my paper · ORCID
classification cs.CV
keywords generativeembeddingbenchmarkevaluationreadoutinformationbottleneckmultimodalembeddingsvisualquestionansweringseparabilityvsrecoverability
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 argues that standard embedding benchmarks, which score representations by their ability to separate correct answers from distractors in a gallery, do not tell us whether the content needed to generate an answer actually survives compression into a dense vector. To test this directly, the authors introduce the Generative Embedding Benchmark (GEB): a frozen embedding, together with the textual question, is fed to a decoder that must produce the answer without ever seeing the original image. Each of seven public embedding models receives the same decoder architecture, training data, and optimization budget, so answer quality becomes an operational measure of recoverable content. On the held-out test set, visual-only scores cluster tightly (28.25–33.21), while question-conditioned joint encoding lifts the best model to 65.56, still far below the 84.30 of a reference that sees the image. The paper's central conclusion is that strong retrieval performance does not imply preservation of the fine-grained information downstream generators need.

What carries the argument

The generative readout is the central mechanism: a frozen embedding is layer-normalized and projected by a two-layer MLP (width 1024) into the hidden space of Qwen3-0.6B, where it replaces a special <EMBED_PAD> token at the start of the first user turn. The decoder is fine-tuned end-to-end with a fixed recipe (one epoch, learning rate 1e-4, effective batch 512, assistant-token loss), and answer quality under this readout is defined as generative information. GEB adds a curated 2,700-item VQA set (1,800 dev, 900 test) balanced across natural images, scene text, and visual documents, with an answerability filter retaining items a reference VLM answers correctly.

What would settle it

Train the identical GEB readout recipe with a decoder from a different language-model family (for example, a Llama-based 0.6B model) and compare the new ranking against Table 4; if Qwen3-VL-Embedding models no longer lead while UME-R1 or VLM2Vec improve, then the original ranking is confounded by encoder–decoder family alignment rather than by information content alone.

Watch

Extended reading notes

Core claim

GEB measures what the authors call generative information: the answer-relevant content a fixed generative readout can recover from a frozen embedding. The readout is a two-layer adapter plus Qwen3-0.6B, trained on LLaVA-NeXT 738K with the embedding model frozen. In visual-only mode the image is embedded once and reused across questions; in vision-language joint mode the embedding is computed from image and question together. Controlled interventions show that matched embeddings outperform text-only, zero, and shuffled inputs by 11–43 points, and the two encoding modes produce different model rankings that do not reproduce the MMEB-V2 ordering. Category-level scores show natural-image content is far more recoverable than scene text or visual-document content, pinpointing where fixed-capacity embeddings lose exact strings, numbers, and layout.

Load-bearing premise

The decoder is assumed to be a neutral probe that reads all embedding spaces equally; in practice it is a Qwen-family model, and the top-scoring embedders are also Qwen3-VL models, so a family-specific alignment could inflate their scores.

Editorial extensions

If this is right

  • If GEB rankings are taken seriously, a model's position on separability benchmarks such as MMEB-V2 does not predict how much of its information a generator can use; models with lower retrieval scores can rank higher on generative readout.
  • The large visual-only-to-VL-joint gains (e.g., Qwen3-VL-Embedding-8B from 33.21 to 65.56) indicate that question-conditioned encoding reallocates embedding capacity toward query-relevant content, at the cost of reusability.
  • Scene text and visual documents show that fixed-capacity embeddings lose exact strings, numbers, and layout more readily than coarse scene semantics, so applications that rely on OCR-like detail need embeddings with adaptive capacity or explicit slots.
  • The control experiments imply that GEB scores are not explained by decoder language priors: shuffled or blank inputs drop scores by roughly 40 points in joint mode.
  • A lightweight query-conditioned adapter over a precomputed image embedding could reclaim part of the VL-joint selectivity without sacrificing offline precomputation.

Reading between the lines

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

  • Because the decoder and the top-scoring embedding models come from the same Qwen3-VL family, the cross-model ranking may partly reflect encoder–decoder hidden-space alignment; testing a decoder from a different family would reveal whether the ranking is an artifact of that alignment.
  • The answerability filter uses a single reference VLM (Gemini-3-Flash), so GEB measures recoverability within an answerable pool; sample-level scores may not transfer to items that this reference cannot answer.
  • One could use the GEB protocol to train embedding models directly: since the readout is differentiable, a loss on generative answer quality could shape embeddings toward preserving generator-relevant information, not just separability.
  • Category-level scores provide a per-question diagnostic that could be used to allocate embedding capacity dynamically—for instance, more slots or longer vectors for scene-text and document images.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper introduces the Generative Embedding Benchmark (GEB), which evaluates frozen multimodal embedding models by training a decoder to answer visual questions using only the embedding and the question text, without access to the original image or intermediate visual features. The benchmark has a curated 1,800-item development split and a 900-item held-out test split across natural images, scene text, and visual documents. Seven public embedding models are evaluated in visual-only and vision-language joint modes under a common decoder (Qwen3-0.6B) and training recipe. The main results are that visual-only scores span a narrow range (28.25–33.21), VL-joint encoding raises VLM-based model scores substantially (best 65.56 for Qwen3-VL-Embedding-8B), natural-image information is more recoverable than scene text or visual-document information, and GEB rankings differ from MMEB-V2 rankings. Control experiments show that matched, sample-specific embeddings outperform text-only, zero, and shuffled inputs.

Significance. The paper proposes a genuinely complementary evaluation axis to separability-based benchmarks: instead of asking whether an embedding can rank candidates, GEB asks whether answer-relevant content can be recovered from the embedding by a generative readout. The protocol is carefully staged, with development-set selection, a held-out test set, source-attrition reporting, image-identity leakage audits, and multiple control conditions. The authors also disclose important limitations, including the answerability-filter dependence and the fact that bootstrap intervals reflect item sampling rather than training-run variation. If the decoder-neutrality concern can be addressed, the benchmark would be a useful contribution to the community, and the paper's concrete observations about scene-text and visual-document recoverability would be valuable.

major comments (3)
  1. [§3.2 and Table 4] The cross-model rankings and the central claim that GEB exposes information bottlenecks rest on the assumption that the shared Qwen3-0.6B decoder is a neutral readout across embedding families. The top-scoring VL-joint models in Table 4 are Qwen3-VL-Embedding-2B and Qwen3-VL-Embedding-8B, which share the Qwen3-VL family with the decoder LM, while the readout-adequacy checks in Supplementary S2 only vary decoder size within the Qwen3 family and training-data scale. The paper also lacks an oracle control in which the same trained decoder receives image-derived visual features or image tokens from the same visual encoder family; without such a control, the low scene-text and visual-document scores (e.g., 23.67 and 13.42 for the best visual-only models) could reflect a decoder-side inability to reproduce exact strings from any input rather than an embedding-side information bottleneck. I request either a control that feeds original-image visual features through the same adapter/decoder pipeline as an upper bound, or a rerun of the main comparisons with a decoder from a different LM family, to separate decoder-family alignment from embedding content.
  2. [§5.3, Table 4, and Supplementary S4.2] Each model–mode row in Table 4 comes from a single training run, and the bootstrap intervals in S4.2 are explicitly stated to quantify test-item sampling variation, not optimization variation; the paper itself notes that a decoder retrained with a different seed could score outside the reported intervals. Under this uncertainty, several load-bearing comparisons are not established: the visual-only range from 28.25 to 33.21 and the 'nearly tied' characterization of Qwen3-VL-Embedding-8B versus 2B fall within the item-sampling CIs, and the VL-joint ordering changes relative to MMEB-V2 (e.g., Qwen3-VL-Embedding-2B at 44.87 versus UME-R1-2B at 51.45) could be training noise rather than a real difference. I ask for at least one independent seed for the main model–mode rows, or a formal seed-variance estimate, and for ranking claims to be restricted to differences that survive both item-sampling and seed variation.
  3. [§3.3 and §5.3] The visual-only and VL-joint conditions differ not only in whether the embedding is conditioned on the question but also in training-instance construction: visual-only uses one embedding shared across the full conversation, while VL-joint expands each conversation into turn-level instances, increasing the training set from about 736,900 to about 4,601,635 instances (S4.1). The gain attributed to query-conditioned encoding is therefore confounded with supervision format and data quantity. Although §5.3 acknowledges that 'this gain reflects the complete mode-specific pipeline rather than encoder conditioning alone,' the Discussion interprets the gain in terms of query-conditioned compression. I request an ablation that trains a visual-only decoder on the same turn-level instance format, reusing the same frozen image embedding for each turn, or otherwise equates the supervision budget, so that the mode comparison isolates encoder conditioning.
minor comments (4)
  1. [Table 4] In the visual-only block, the Scene and Doc entries for VLM2Vec-V2 are rendered as '11.6015.19'; a missing separator makes the value ambiguous and should be fixed.
  2. [§4.2] The sentence 'A leakage audit confirms that no image-identity group crosses the two splits: ... are all zero (Table 2)' refers to a table that actually reports per-source candidate attrition; the zero-overlap audit is described in Supplementary S3. Please correct the cross-reference.
  3. [§5.5] The comparison with MMEB-V2 would benefit from stating explicitly that the archived MMEB-V2 scores may come from different checkpoints or evaluation versions than the GEB runs, to avoid implying that the same model instances are being compared on the two benchmarks.
  4. [Supplementary S4.1] The two-slot late-fusion interface for CLIP and SigLIP is mentioned only briefly; a short example of how the two vectors are concatenated or otherwise combined in the adapter would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: GEB is an empirical benchmark with a held-out test split and external controls; its central claims do not reduce to their inputs.

full rationale

GEB is an evaluation protocol rather than a derivation. Generative information is defined operationally as answer-relevant content recoverable by a fixed generative readout, and the paper then measures that quantity; this is an explicit operationalization, not a claim derived from itself. The protocol is fixed on a development split, and the 900-item test split is held out, so no model comparison or control result is fitted to force a conclusion. Test scores are anchored to external benchmarks (MMEB-V2 rankings and source-task answerability) and to non-informative controls (zero, shuffled, blank, and text-only inputs). The release includes leakage audits, and the paper explicitly checks that test items do not appear in the decoder's training data. The strongest residual concern is that the Qwen3-0.6B decoder may favor Qwen-family embedding spaces, which is a validity threat for cross-model rankings, but it is not circularity: no parameter is fitted from the target outcome, no load-bearing premise is imported via self-citation, and no equation reduces a claimed result to the benchmark's own inputs. Accordingly, the circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim depends on the readout being both capable and neutral. The free parameters listed are hand-selected protocol components that determine the measured scores; none is fitted to produce a target result. The axioms capture assumptions about answerability filtering, metric validity, and readout neutrality that, if violated, would change the conclusions. No new physical or mathematical entities are postulated; the paper introduces an operational construct, 'generative information', and a benchmark artifact, which are measurement tools rather than invented entities.

free parameters (4)
  • Decoder LM choice = Qwen3-0.6B
    Hand-chosen readout language model. It shares the Qwen3 family with the top-scoring embedding models, which may make the adapter mapping easier for those embedders and bias cross-model comparisons. The paper does not control for model-family alignment.
  • Answerability filter model and thresholds = Gemini-3-Flash; binary correct or ANLS/consensus >= 0.5
    Candidates are retained only if Gemini-3-Flash answers correctly under each source metric. Per-source retention ranges from 39.4% to 95.6%, so the test pool is conditioned on this reference VLM and is not model-neutral.
  • Adapter width = 1024
    A fixed hidden width of 1024 projects each embedding into the decoder. This could impose a bottleneck that interacts differently with embedding dimensionalities (768 to 4096) across models.
  • Decoder training recipe = 1 epoch, LR 1e-4, batch 512, LLaVA-NeXT 738K
    One fixed recipe and one training trajectory per model; the recipe was selected on the development set, but its interaction with each embedding space is not measured, and no variance across training runs is reported.
assumptions (4)
  • domain assumption A single fixed decoder LM and training recipe provides a neutral readout across different embedding models.
    All models are compared using Qwen3-0.6B with the same training recipe. If the readout is not neutral across embedding families, the Table 4 rankings are confounded. This is the core fairness assumption of the comparative claim.
  • domain assumption Reference-VLM answerability implies the item is a fair test of recoverable content.
    Section 4.1 retains only items Gemini-3-Flash answers correctly. The filter may change the difficulty distribution and favor content styles that one reference model handles well; the paper acknowledges the pool is answerable-pool, not model-neutral.
  • domain assumption The decoder has enough capacity and training data to recover whatever information the embedding contains.
    Development ablations (Table 3, S2) show limited gains from a larger decoder, but this does not prove that the readout can extract all recoverable information for every model or question type.
  • domain assumption Source VQA metrics are valid for scoring the generative readout.
    Section 4.3 reuses ANLS, consensus, and relaxed accuracy. These metrics were designed for the original models and may not capture semantic correctness when the decoder sees only a compressed embedding.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generative Embedding Benchmark: How Much Information Survives in a Dense Embedding?." pith.science (2026). https://pith.science/paper/NFM6ZZFK

@misc{pith2026260806972,
  author       = {Pith},
  title        = {Pith review of: Generative Embedding Benchmark: How Much Information Survives in a Dense Embedding?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NFM6ZZFK}},
  note         = {Machine review of arXiv:2608.06972}
}
read the original abstract

Embeddings have emerged as a standard representational interface linking foundation models with downstream systems. Most embedding benchmarks assess representations through discriminative tasks or geometric criteria centered on separability in embedding space. However, strong performance on such evaluations does not establish whether content compressed into an embedding remains accessible to a downstream generator. To address this gap, we introduce the Generative Embedding Benchmark (GEB), in which a decoder answers questions using only a frozen embedding and question text, without access to the original image or intermediate visual features. Answer quality under this readout measures generative information: the answer-relevant content recoverable from an embedding. GEB includes a curated visual-question-answering dataset with a 1,800-item development split and a held-out 900-item test split covering natural images, scene text, and visual documents. Using a common decoder and training recipe, we evaluate seven public embedding models in visual-only and vision-language joint modes. On the test set, visual-only scores range from 28.25 to 33.21; with image-question joint encoding, all five VLM-based embedding models score higher, and the best reaches 65.56. Matched embeddings also outperform text-only inputs, zero embeddings, and shuffled embeddings. Natural-image information is much easier to recover than scene text or visual-document information, while a Qwen3-VL-2B reference with access to the original image reaches 84.30. Together, these results show that generative readout exposes information bottlenecks that separability-based evaluation does not capture.

Figures

Figures reproduced from arXiv: 2608.06972 by the authors.

Figure 1
Figure 1. Discriminative benchmarks measure separability [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. GEB encoding modes. The decoder always receives [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Two qualitative views of generative readout: (a) cross-model answers to two questions; (b) multi-question interrogation [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

73 extracted references · 48 canonical work pages

  1. [1]

    Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018) , year=

    SentEval: An Evaluation Toolkit for Universal Sentence Representations , author=. Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018) , year=

    work page 2018

  2. [2]

    Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics , pages=

    Muennighoff, Niklas and Tazi, Nouamane and Magne, Lo. Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics , pages=. 2023 , address=

    work page 2023

  3. [3]

    2025 , eprint=

    Enevoldsen, Kenneth and Chung, Isaac and Kerboua, Imene and Kardos, M. 2025 , eprint=

    work page 2025

  4. [4]

    2410.05160 , archivePrefix=

    Jiang, Ziyan and Meng, Rui and Yang, Xinyi and Yavuz, Semih and Zhou, Yingbo and Chen, Wenhu , year=. 2410.05160 , archivePrefix=

    arXiv

  5. [5]

    2025 , eprint=

    Xiao, Chenghao and Chung, Isaac and Kerboua, Imene and Stirling, Jamie and Zhang, Xin and Kardos, M. 2025 , eprint=

    work page 2025

  6. [6]

    2507.04590 , archivePrefix=

    Meng, Rui and Jiang, Ziyan and Liu, Ye and Su, Mingyi and Yang, Xinyi and Fu, Yuepeng and Qin, Can and Chen, Zeyuan and Xu, Ran and Xiong, Caiming and Zhou, Yingbo and Chen, Wenhu and Yavuz, Semih , year=. 2507.04590 , archivePrefix=

    arXiv

  7. [7]

    and Tang, Michael and others , year=

    Su, Hongjin and Yen, Howard and Xia, Mengzhou and Shi, Weijia and Muennighoff, Niklas and Wang, Han-yu and Liu, Haisu and Shi, Quan and Siegel, Zachary S. and Tang, Michael and others , year=. 2407.12883 , archivePrefix=

    arXiv

  8. [8]

    2024 , address=

    Mahajan, Yash and Bansal, Naman and Blanco, Eduardo and Karmaker, Santu , booktitle=. 2024 , address=

    work page 2024

Show all 73 references
  1. [9]

    2506.05176 , archivePrefix=

    Zhang, Yanzhao and Li, Mingxin and Long, Dingkun and Zhang, Xin and Lin, Huan and Yang, Baosong and Xie, Pengjun and Yang, An and Liu, Dayiheng and Lin, Junyang and Huang, Fei and Zhou, Jingren , year=. 2506.05176 , archivePrefix=

    arXiv

  2. [10]

    2601.04720 , archivePrefix=

    Li, Mingxin and Zhang, Yanzhao and Long, Dingkun and Chen, Keqin and Song, Sibo and Bai, Shuai and Yang, Zhibo and Xie, Pengjun and Yang, An and Liu, Dayiheng and Zhou, Jingren and Lin, Junyang , year=. 2601.04720 , archivePrefix=

    arXiv

  3. [11]

    2505.09388 , archivePrefix=

    Yang, An and Li, Anfeng and Yang, Baosong and Zhang, Beichen and Hui, Binyuan and Zheng, Bo and Yu, Bowen and Gao, Chang and others , year=. 2505.09388 , archivePrefix=

    arXiv

  4. [12]

    2511.21631 , archivePrefix=

    Bai, Shuai and Cai, Yuxuan and Chen, Ruizhe and Chen, Keqin and Chen, Xionghui and Cheng, Zesen and Deng, Lianghao and Ding, Wei and Gao, Chang and others , year=. 2511.21631 , archivePrefix=

    arXiv

  5. [13]

    2511.00405 , archivePrefix=

    Lan, Zhibin and Niu, Liqiang and Meng, Fandong and Zhou, Jie and Su, Jinsong , year=. 2511.00405 , archivePrefix=

  6. [14]

    2602.13823 , archivePrefix=

    Jiang, Haonan and Wang, Yuji and Zhu, Yongjie and Lu, Xin and Qin, Wenyu and Wang, Meng and Wan, Pengfei and Tang, Yansong , year=. 2602.13823 , archivePrefix=

  7. [15]

    What you can cram into a single

    Conneau, Alexis and Kruszewski, German and Lample, Guillaume and Barrault, Lo. What you can cram into a single. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=. 2018 , address=

    2018

  8. [16]

    A Structural Probe for Finding Syntax in Word Representations , author=. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) , pages=. 2019 , address=

    2019

  9. [17]

    Designing and Interpreting Probes with Control Tasks , author=. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) , pages=. 2019 , address=

    2019

  10. [18]

    Transactions of the Association for Computational Linguistics , volume=

    Amnesic Probing: Behavioral Explanation with Amnesic Counterfactuals , author=. Transactions of the Association for Computational Linguistics , volume=

  11. [19]

    Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages=

    Information-Theoretic Probing with Minimum Description Length , author=. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages=. 2020 , address=

    2020

  12. [20]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    Text Embeddings Reveal (Almost) As Much As Text , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=. 2023 , address=

    2023

  13. [21]

    Findings of the Association for Computational Linguistics: ACL 2023 , pages=

    Sentence Embedding Leaks More Information than You Expect: Generative Embedding Inversion Attack to Recover the Whole Sentence , author=. Findings of the Association for Computational Linguistics: ACL 2023 , pages=. 2023 , address=

    2023

  14. [22]

    2025 , address=

    Chen, Yiyi and Xu, Qiongkai and Bjerva, Johannes , booktitle=. 2025 , address=

    2025

  15. [23]

    2025 , eprint=

    Universal Zero-shot Embedding Inversion , author=. 2025 , eprint=

    2025

  16. [24]

    2026 , eprint=

    Zero2Text: Zero-Training Cross-Domain Inversion Attacks on Textual Embeddings , author=. 2026 , eprint=

    2026

  17. [25]

    Proceedings of the First Workshop on Large Language Model Memorization (L2M2) , pages=

    Memory Tokens: Large Language Models Can Generate Reversible Sentence Embeddings , author=. Proceedings of the First Workshop on Large Language Model Memorization (L2M2) , pages=. 2025 , address=

    2025

  18. [26]

    Training

    Su, Chang and Shi, Dengliang and Huang, Siyuan and Du, Jintao and Meng, Changhua and Cheng, Yu and Wang, Weiqiang and Lin, Zhouhan , booktitle=. Training. 2025 , address=

    2025

  19. [27]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    Adapting Language Models to Compress Contexts , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=. 2023 , address=

    2023

  20. [28]

    2024 , eprint=

    In-Context Autoencoder for Context Compression in a Large Language Model , author=. 2024 , eprint=

    2024

  21. [29]

    Advances in Neural Information Processing Systems , volume=

    Learning to Compress Prompts with Gist Tokens , author=. Advances in Neural Information Processing Systems , volume=. 2023 , url=

    2023

  22. [30]

    Cheng, Xin and Wang, Xun and Zhang, Xingxing and Ge, Tao and Chen, Si-Qing and Wei, Furu and Zhang, Huishuai and Zhao, Dongyan , year=. x. 2405.13792 , archivePrefix=

    arXiv

  23. [31]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Pretraining Context Compressor for Large Language Models with Embedding-Based Memory , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=. 2025 , address=

    2025

  24. [32]

    2023 , eprint=

    Duquenne, Paul-Ambroise and Schwenk, Holger and Sagot, Beno. 2023 , eprint=

    2023

  25. [33]

    2506.15841 , archivePrefix=

    Zhou, Zijian and Qu, Ao and Wu, Zhaoxuan and Kim, Sunghwan and Prakash, Alok and Rus, Daniela and Zhao, Jinhua and Low, Bryan Kian Hsiang and Liang, Paul Pu , year=. 2506.15841 , archivePrefix=

    arXiv

  26. [34]

    2509.24704 , archivePrefix=

    Zhang, Guibin and Fu, Muxin and Yan, Shuicheng , year=. 2509.24704 , archivePrefix=

  27. [35]

    2026 , eprint=

    Latent Collaboration in Multi-Agent Systems , author=. 2026 , eprint=

    2026

  28. [36]

    2023 , address=

    Jiang, Huiqiang and Wu, Qianhui and Lin, Chin-Yew and Yang, Yuqing and Qiu, Lili , booktitle=. 2023 , address=

    2023

  29. [37]

    Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    Pan, Zhuoshi and Wu, Qianhui and Jiang, Huiqiang and Xia, Menglin and Luo, Xufang and Zhang, Jue and Lin, Qingwei and R. Findings of the Association for Computational Linguistics: ACL 2024 , pages=. 2024 , address=

    2024

  30. [38]

    2025 , eprint=

    Generative Representational Instruction Tuning , author=. 2025 , eprint=

    2025

  31. [39]

    2404.05961 , archivePrefix=

    BehnamGhader, Parishad and Adlakha, Vaibhav and Mosbach, Marius and Bahdanau, Dzmitry and Chapados, Nicolas and Reddy, Siva , year=. 2404.05961 , archivePrefix=

    arXiv

  32. [40]

    2024 , eprint=

    Improving Text Embeddings with Large Language Models , author=. 2024 , eprint=

    2024

  33. [41]

    2405.17428 , archivePrefix=

    Lee, Chankyu and Roy, Rajarshi and Xu, Mengyao and Raiman, Jonathan and Shoeybi, Mohammad and Catanzaro, Bryan and Ping, Wei , year=. 2405.17428 , archivePrefix=

    arXiv

  34. [42]

    2026 , eprint=

    Compressing then Matching: An Efficient Pre-training Paradigm for Multimodal Embedding , author=. 2026 , eprint=

    2026

  35. [43]

    2024 , eprint=

    Multi-Modal Generative Embedding Model , author=. 2024 , eprint=

    2024

  36. [44]

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

    Lost in Embeddings: Information Loss in Vision--Language Models , author=. Findings of the Association for Computational Linguistics: EMNLP 2025 , pages=. 2025 , address=

    2025

  37. [45]

    Advances in Neural Information Processing Systems , volume=

    Flamingo: a Visual Language Model for Few-Shot Learning , author=. Advances in Neural Information Processing Systems , volume=. 2022 , url=

    2022

  38. [46]

    Proceedings of the 38th International Conference on Machine Learning , pages=

    Learning Transferable Visual Models From Natural Language Supervision , author=. Proceedings of the 38th International Conference on Machine Learning , pages=. 2021 , volume=. 2103.00020 , archivePrefix=

    2021 arXiv

  39. [47]

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

    Sigmoid Loss for Language Image Pre-Training , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=. 2023 , eprint=

    2023

  40. [48]

    2024 , url=

    Tong, Shengbang and Brown, Ellis and Wu, Penghao and Woo, Sanghyun and Middepogu, Manoj and Akula, Sai Charitha and Yang, Jihan and Yang, Shusheng and Iyer, Adithya and Pan, Xichen and Wang, Austin and Fergus, Rob and LeCun, Yann and Xie, Saining , booktitle=. 2024 , url=

    2024

  41. [49]

    Mathew, Minesh and Karatzas, Dimosthenis and Jawahar, C. V. , booktitle=

  42. [50]

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

    Mathew, Minesh and Bagal, Viraj and Tito, Rub. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=. 2022 , url=

    2022

  43. [51]

    Liu, Haotian and Li, Chunyuan and Li, Yuheng and Li, Bo and Zhang, Yuanhan and Shen, Sheng and Lee, Yong Jae , year=

  44. [52]

    2025 , url=

    Introducing. 2025 , url=

    2025

  45. [53]

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

    nocaps: novel object captioning at scale , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  46. [54]

    2306.13394 , archivePrefix=

    Fu, Chaoyou and Chen, Peixian and Shen, Yunhang and Qin, Yulei and Zhang, Mengdan and Lin, Xu and Yang, Jinrui and Zheng, Xiawu and Li, Ke and Sun, Xing and Wu, Yunsheng and Ji, Rongrong and Shan, Caifeng and He, Ran , year=. 2306.13394 , archivePrefix=

    arXiv

  47. [55]

    Liu, Yuliang and Li, Zhang and Huang, Mingxin and Yang, Biao and Yu, Wenwen and Li, Chunyuan and Yin, Xucheng and Liu, Cheng-Lin and Jin, Lianwen and Bai, Xiang , journal=

  48. [56]

    Singh, Amanpreet and Natarajan, Vivek and Shah, Meet and Jiang, Yu and Chen, Xinlei and Batra, Dhruv and Parikh, Devi and Rohrbach, Marcus , booktitle=. Towards

  49. [57]

    Biten, Ali Furkan and Tito, Ruben and Mafla, Andres and Gomez, Lluis and Rusinol, Marcal and Mathew, Minesh and Jawahar, C. V. and Valveny, Ernest and Karatzas, Dimosthenis , booktitle=

  50. [58]

    2022 , address=

    Masry, Ahmed and Long, Do Xuan and Tan, Jia Qing and Joty, Shafiq and Hoque, Enamul , booktitle=. 2022 , address=

    2022

  51. [59]

    2024 , howpublished=

    2024

  52. [60]

    2025 , address=

    Zhang, Kaichen and Li, Bo and Zhang, Peiyuan and Pu, Fanyi and Cahyono, Joshua Adrian and Hu, Kairui and Liu, Shuai and Zhang, Yuanhan and Yang, Jingkang and Li, Chunyuan and Liu, Ziwei , booktitle=. 2025 , address=

    2025

  53. [61]

    2002 , address=

    Papineni, Kishore and Roukos, Salim and Ward, Todd and Zhu, Wei-Jing , booktitle=. 2002 , address=

    2002

  54. [62]

    Lawrence and Parikh, Devi , booktitle=

    Vedantam, Ramakrishna and Zitnick, C. Lawrence and Parikh, Devi , booktitle=

  55. [63]

    Microsoft

    Chen, Xinlei and Fang, Hao and Lin, Tsung-Yi and Vedantam, Ramakrishna and Gupta, Saurabh and Doll. Microsoft. 2015 , eprint=

    2015

  56. [64]

    Proceedings of the Sixth Workshop on Statistical Machine Translation , pages=

    Meteor 1.3: Automatic Metric for Reliable Optimization and Evaluation of Machine Translation Systems , author=. Proceedings of the Sixth Workshop on Statistical Machine Translation , pages=. 2011 , address=

    2011

  57. [65]

    2004 , address=

    Lin, Chin-Yew , booktitle=. 2004 , address=

    2004

  58. [66]

    Findings of the Association for Computational Linguistics: EMNLP 2023 , pages=

    How Well Do Text Embedding Models Understand Syntax? , author=. Findings of the Association for Computational Linguistics: EMNLP 2023 , pages=. 2023 , address=

    2023

  59. [67]

    2023 , url=

    Li, Wei and Zhu, Linchao and Wen, Longyin and Yang, Yi , booktitle=. 2023 , url=. 2303.03032 , archivePrefix=

    2023 arXiv

  60. [68]

    Implicit Inversion turns

    D'Orazio, Antonio and Briglia, Maria Rosaria and Crisostomi, Donato and Loi, Dario and Rodol. Implicit Inversion turns. 2025 , eprint=

    2025

  61. [69]

    Proceedings of the 37th Annual Allerton Conference on Communication, Control, and Computing , pages=

    The Information Bottleneck Method , author=. Proceedings of the 37th Annual Allerton Conference on Communication, Control, and Computing , pages=

  62. [70]

    Huang, Haohang and Lu, Xuan and Su, Mingyi and Zhang, Xuan and Jiang, Ziyan and Nie, Ping and Zou, Kai and Pfister, Tomas and Chen, Wenhu and Zhang, Wei and Shen, Xiaoyu and Meng, Rui , journal=

  63. [71]

    International Conference on Learning Representations , year=

    A Theory of Usable Information under Computational Constraints , author=. International Conference on Learning Representations , year=

  64. [72]

    , journal=

    Mokady, Ron and Hertz, Amir and Bermano, Amit H. , journal=

  65. [73]

    Text-Only Training for Image Captioning using Noise-Injected

    Nukrai, David and Mokady, Ron and Globerson, Amir , booktitle=. Text-Only Training for Image Captioning using Noise-Injected. 2022 , publisher=

    2022

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.