Pith. sign in

REVIEW 3 major objections 4 minor 66 references

The paper claims that a decoder-only multimodal LLM, using 16 special tokens each covering a slice of the vocabulary, can produce sparse lexical and dense embeddings in one forward pass, with sparse performance within about one point of den

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 · deepseek-v4-flash

2026-08-04 04:13 UTC pith:DC2MD6UQ

load-bearing objection Genuinely new idea—decoder-only MLLM emitting dense and sparse embeddings via partitioned special tokens—but the headline parity claim is undercut by a 1.9-point appendix discrepancy and single-seed numbers. the 3 major comments →

arxiv 2608.02583 v1 pith:DC2MD6UQ submitted 2026-08-03 cs.CV cs.AIcs.CLcs.IR

UEmbed: Unified Sparse and Dense Multimodal Embeddings

Tingyu Song , Mingxin Li , Yanzhao Zhang , Dingkun Long , Pengjun Xie , Zhijie Nie , Yilun Zhao , Shu Wu This is my paper
classification cs.CV cs.AIcs.CLcs.IR
keywords unified embeddingslearned sparse retrievaldecoder-onlymultimodal retrievaldense retrievalspecial tokensvocabulary partitionMMEB-v2
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.

The paper tries to establish that sparse lexical retrieval, historically tied to bidirectional encoder architectures, can be produced natively by a decoder-only multimodal model without auxiliary cross-modal modules. It achieves this by appending 16 learnable special tokens, each assigned a disjoint slice of the vocabulary via k-means clustering, and training the model to output a sparse vector that is nearly as effective as its dense vector. On MMEB-v2, the 9B model scores 71.8 (dense) and 71.0 (sparse), a gap under one point, and the paper reports the first sparse model evaluated on that benchmark. A sympathetic reader would care because, if right, sparse retrieval gains the efficiency and interpretability of lexical search while keeping semantic power, and it works across text, images, video, and documents in one backbone.

Core claim

The central finding is that the information bottleneck of causal attention for sparse retrieval can be circumvented by partitioning the vocabulary into N disjoint subsets and assigning each to a dedicated special token at the end of the input. Each token's causal hidden state produces sparse weights only over its assigned slice, and the N slices are concatenated into the full sparse vector; the dense embedding comes from the EOS hidden state preceding the special tokens. Trained with a joint dense/sparse contrastive objective, UEmbed-9B reaches 71.8 (dense) and 71.0 (sparse) on MMEB-v2, closing the dense-sparse gap to within one point at every scale tested. The paper also shows that joint tr

What carries the argument

The partitioned sparse head: 16 learnable special tokens appended after the input, each assigned a disjoint subset of the (accent-stripped, lowercased, whitespace-collapsed) vocabulary via k-means clustering. Each token computes log(1 + ReLU(W_t^T h_s + b_t)) over its slice, and the slices are concatenated into the full sparse vector; the dense embedding is the hidden state of the EOS token immediately before the special tokens. This replaces SPLADE's max-pooling over all token states, which causal masking forbids, with a fixed set of causal readout positions.

Load-bearing premise

The argument assumes that 16 causal readout tokens, each restricted to a fixed vocabulary slice, have enough representational capacity to replace the bidirectional max-pooling over all token states used by SPLADE; the paper offers empirical ablations but no capacity argument, so if 16 fixed slices cannot jointly express the needed sparse vector, sparse mode would degrade.

What would settle it

Take the released 9B checkpoint and compute sparse vectors for a set of queries rich in rare or technical terms; compare the number of relevant activated terms against SPLADE-v3 on the same queries. If the fixed 16-slice head consistently misses terms SPLADE finds, the capacity bet is falsified. Alternatively, train a variant with N=64 while keeping the contrastive sequence length constant (e.g., by reusing input-token hidden states) and check whether sparse performance improves; if it does, the paper's N=16 choice is a capacity limit rather than a sweet spot.

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

If this is right

  • If the central claim holds, any causal LLM can be adapted to produce sparse retrieval vectors in a single forward pass, removing the need to re-train backbones as bidirectional encoders.
  • Hybrid scoring over the model's dense and sparse modes requires no extra encoding pass and measurably improves text and visual-document retrieval.
  • Multimodal sparse retrieval can be served from one backbone without auxiliary cross-modal modules, simplifying extension to new modalities.
  • The method scales consistently from 2B to 9B parameters, with sparse performance staying within one point of dense at every scale tested.
  • The sparse representation is natively compatible with inverted-index search and high-throughput autoregressive serving stacks, lowering deployment cost for lexical retrieval.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The 16-token partition is a capacity bet: the paper's observed drop at N=32 may reflect inflated contrastive sequence lengths rather than a true optimum, so increasing N while controlling sequence length is a testable extension that could push sparse performance higher.
  • Because the dense mode is the hidden state of a single EOS token, it is a weaker representation than the ensemble of 16 sparse heads; stronger dense pooling could widen the sparse-dense gap again, so the reported parity may partly reflect an underpowered dense baseline.
  • The English-Chinese training distribution limits the sparse vocabulary's language coverage; a more language-diverse corpus would test whether the partition mechanism generalizes or collapses to a few dominant languages.
  • The smaller gain on video suggests a flat sparse vector has capacity limits for temporally dense data; hierarchical or multi-vector sparse representations are a natural next step the paper leaves implicit.

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

3 major / 4 minor

Summary. The paper proposes UEmbed, a family of decoder-only multimodal embedding models (2B, 4B, 9B) that produce both dense and learned-sparse representations in a single causal forward pass. Sparse representations are generated by appending N=16 learnable special tokens, each responsible for a disjoint k-means-clustered subset of a compressed vocabulary, and concatenating their per-subset sparse predictions. The dense representation is taken from the EOS hidden state. Training uses a combined InfoNCE objective with FLOPS regularizers on publicly available text, image, video, and visual-document data. On MMEB-v2, UEmbed-9B reaches 71.8 dense and 71.0 sparse, and the paper claims the dense-sparse gap is at most 1.0 point at all scales, establishing a new state of the art in sparse multimodal retrieval. On BEIR, UEmbed remains competitive with dense and sparse baselines. Additional experiments analyze joint-training robustness, vocabulary partitioning, sparse temperature, the number of special tokens, hybrid scoring, serving efficiency, and agentic-search cost.

Significance. If the claims hold, the architecture is a meaningful advance: it is one of the first decoder-only models to natively support both dense and sparse retrieval for multimodal inputs, eliminates auxiliary cross-modal modules, and is compatible with high-throughput causal serving stacks and inverted indices. The paper's strengths include extensive evaluations on MMEB-v2 and BEIR, controlled ablations of each design choice, release of models at three scales, and explicit discussion of limitations (language bias, vocabulary artifacts, video gap). The central quantitative claim—that dense and sparse modes are within one point across scales—is the load-bearing result, and it is currently supported only by single-seed numbers that appear to be internally inconsistent with the paper's own appendix.

major comments (3)
  1. [Appendix B.4, Table 10 vs. Table 1] Table 1 reports UEmbed-2B dense 66.5 / sparse 65.5 on MMEB-v2, while Table 10 in B.4 reports dense 64.6 / sparse 63.4 for the same model name, and both the main training description (A.2) and the appendix header state random seed 42. The dense discrepancy is 1.9 points—larger than the claimed maximum dense-sparse gap of 1.0 point. The paper must either explicitly state that Table 10 uses the 500k-instance ablation mixture from §5.1 (and therefore is not directly comparable to Table 1), or, if the same training setup was used, the run-to-run variance is large enough to undermine the central 'within one point' parity claim. Please report multi-seed means and variances, or clearly label the appendix models as ablation-subset models.
  2. [Abstract and §4.2, Table 1] The abstract says UEmbed 'outperforming multimodal embedding models trained on publicly available data (e.g., RzenEmbed)'. This is true for dense mode (71.8 vs. 71.1 for RzenEmbed-V2-7B), but false for sparse mode: UEmbed-9B sparse scores 71.0, 0.1 below RzenEmbed-V2-7B (71.1). The body text in §4.2 carefully restricts the claim to dense mode, but the abstract's blanket wording is contradicted by Table 1. Please qualify the claim by mode, or revise the abstract to reflect the dense-only comparison.
  3. [§4.2 and §5.1] The central parity result relies on single-seed runs: A.2 fixes random seed 42, and no variance or confidence intervals are reported anywhere. In addition, the hyperparameters that most affect the sparse mode—N, sparse temperature tau_s, and the hybrid beta weights—were tuned on MMEB-v1 or held-out splits of the same benchmark family (B.3). This tuning dependence weakens the claim that the MMEB-v2 numbers are unbiased. Please report at least three seeds for the main models, or clearly separate tuned-ablations from final evaluations and justify the absence of variance estimates.
minor comments (4)
  1. [§5.1, Table 3] The SPLADE baseline shares the Qwen3.5 backbone but adopts bidirectional attention. Please explain how bidirectional attention is realized in this decoder-only backbone (e.g., full attention masking, weight conversion, or architectural modification), as this is essential for reproducing the +3.2/+2.1 gains.
  2. [§3.2] The vocabulary compression reduces size from 248,320 to 184,016, but it is not stated which tokenizer/vocabulary this corresponds to (presumably the Qwen3.5 tokenizer). Please state the vocabulary source explicitly.
  3. [Figure 5] The training-dynamics figure has a single y-axis labeled 'Metric Value' for three subplots with very different scales (loss, max-neg gap, mean-neg gap). Please use separate axes or normalize so the stabilization claim is visually interpretable.
  4. [Appendix B.3] The hybrid beta values are tuned per modality on a held-out split, but it is not stated whether the held-out split is from MMEB-v2, MMEB-v1, or another benchmark. Please specify, since this affects the validity of the hybrid-scoring results in Table 5.

Circularity Check

0 steps flagged

No circular derivation: central claims rest on external MMEB-v2/BEIR scores; self-citations and hyperparameter tuning are not load-bearing.

full rationale

The paper's derivation chain is an architecture (causal sparse heads over k-means vocabulary partitions) plus a joint training objective (Eq. 5), evaluated on external benchmarks MMEB-v2 and BEIR. No fitted parameter is renamed as a prediction: N=16 and tau_s=32 are tuned in ablations (Sec. 5.1) and then used in final models whose MMEB-v2 scores are measured, not derived. The sparse vector is defined as the concatenation of subset head outputs (Eqs. 3-4), so the dense-vs-sparse gap is an empirical comparison of two outputs of the same checkpoint, not a quantity forced by the construction. Self-citations (GME, UniME, Qwen3-VL-Embedding teacher, GVE monitoring) are inputs or baselines with publicly available checkpoints/scores; they do not carry the central argument, and the teacher is explicitly not claimed to be beaten (Qwen3-VL-Embedding-8B outscores UEmbed-9B). The k-means partition is introduced in this paper as a design choice, not imported as an external theorem, so there is no uniqueness or ansatz-smuggling chain. One non-circular concern should be flagged: Appendix B.4/Table 10 reports UEmbed-2B dense 64.6/sparse 63.4 on MMEB-v2 while Table 1 reports 66.5/65.5 for the same configuration, a 1.9-point dense shift that exceeds the claimed max 1.0 gap. This is a reproducibility/consistency issue, not a circular reduction, and does not change the circularity score.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The model rests on a pretrained MLLM and benchmark assumptions; there is no mathematical derivation. Most free parameters are hyperparameters swept on MMEB-v1 or held-out splits; the sparse-head weights themselves are learned, not fitted constants.

free parameters (5)
  • N (number of special tokens) = 16
    Swept over {2,4,8,16,32} on MMEB-v1; N=16 chosen as default (Section 5.1, Figure 2c).
  • sparse temperature tau_s = 32
    Swept over {2,4,8,16,32,64} on MMEB-v1; tau_s=32 chosen as default (Section 5.1, Figure 2b).
  • loss weights lambda, alpha_q, alpha_d = lambda=1.0, alpha_q=alpha_d=1e-4
    Chosen by hand to balance joint dense/sparse training and FLOPS regularization (Eq. 5, Appendix A.2).
  • hybrid interpolation weights beta per modality = Text 5e-8, Image 5e-4, Video 1e-4, VisDoc 7e-4
    Tuned on a held-out split of each modality; used to report hybrid scoring gains (Appendix B.3).
  • dense temperature tau = 0.03
    Standard contrastive temperature, fixed for all runs (Appendix A.2).
axioms (5)
  • domain assumption Pretrained Qwen3.5 MLLM supplies sufficiently strong visual and text representations for both heads
    All results are inherited from this backbone; the paper does not pretrain from scratch (Section 4.1, Appendix A.2).
  • domain assumption In-batch negatives plus teacher-mined hard negatives give a valid contrastive signal
    InfoNCE loss (Eq. 1) assumes negatives are informative; text data rely on batch negatives of size 256.
  • domain assumption k-means on token embeddings yields coherent semantic vocabulary subsets
    The method relies on partition quality; the ablation shows a small but consistent benefit over random split (Section 5.1, Table 4).
  • ad hoc to paper Vocabulary compression is lossless for retrieval after max-retention scoring
    Canonical merging of 248k to 184k tokens assumes no retrieval-critical distinction is lost (Section 3.2).
  • domain assumption MMEB-v2, the selected BEIR subsets, and BrowseComp-Plus are faithful proxies for retrieval quality
    Central claims are benchmark-based; there is no external deployment validation beyond the BrowseComp-Plus agent study (Sections 4 and 5.3).
invented entities (1)
  • N=16 appended special tokens <s1>...<s16> no independent evidence
    purpose: Causal readout positions; each predicts sparse weights over one vocabulary partition
    New architectural tokens; no falsifiable handle outside the paper. Their contribution is evidenced only by in-paper ablations (Figure 2c).

pith-pipeline@v1.3.0-daily-deepseek · 18945 in / 15775 out tokens · 171816 ms · 2026-08-04T04:13:01.417509+00:00 · methodology

0 comments
read the original abstract

Sparse retrieval underpins modern search systems, from web search to retrieval-augmented generation. Existing work has introduced Learned Sparse Retrieval (LSR) to push beyond exact lexical matching toward richer semantics. Yet LSR has so far remained tied to encoder-style bidirectional architectures, and its extension to multimodal settings still relies heavily on auxiliary cross-modal modules. To address these limitations, we introduce UEmbed (Unified Embedding), a decoder-only multimodal embedding model that produces both sparse lexical and dense representations in one causal forward pass. UEmbed appends N learnable special tokens to the input and partitions the vocabulary into N disjoint subsets. Each token's causal hidden state predicts sparse weights over its assigned subset, and the N subsets are concatenated into the full sparse vector. Trained on public data, we release UEmbed at 2B, 4B, and 9B scales. UEmbed-9B reaches 71.8 (dense) and 71.0 (sparse) on MMEB-v2, outperforming multimodal embedding models trained on publicly available data (e.g., RzenEmbed). On BEIR, UEmbed also remains competitive with strong dense and sparse baselines. Furthermore, we demonstrate the practical utility of UEmbed across three dimensions: effectiveness, efficiency, and agentic applications. Overall, UEmbed offers a new paradigm: it unifies dense and sparse embeddings in one model, while further extending sparse retrieval to unify text and multimodal inputs.

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

66 extracted references · 17 linked inside Pith

  1. [1]

    5-vl technical report , author=

    Qwen2. 5-vl technical report , author=. arXiv preprint arXiv:2502.13923 , year=

    Pith/arXiv arXiv

  2. [2]

    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

  3. [3]

    arXiv preprint arXiv:2312.11805 , year=

    Gemini: a family of highly capable multimodal models , author=. arXiv preprint arXiv:2312.11805 , year=

    Pith/arXiv arXiv

  4. [4]

    Gemini-2.0 , url =

    Google , month =. Gemini-2.0 , url =

  5. [5]

    2024 , eprint=

    GPT-4o System Card , author=. 2024 , eprint=

    2024

  6. [6]

    arXiv preprint arXiv:2601.03267 , year=

    Openai gpt-5 system card , author=. arXiv preprint arXiv:2601.03267 , year=

    Pith/arXiv arXiv

  7. [7]

    arXiv preprint arXiv:2511.21631 , year=

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

    Pith/arXiv arXiv

  8. [8]

    2009 , publisher=

    The probabilistic relevance framework: BM25 and beyond , author=. 2009 , publisher=

    2009

  9. [9]

    SPLADE: Sparse Lexical and Expansion Model for First Stage Ranking , year =

    Formal, Thibault and Piwowarski, Benjamin and Clinchant, St\'. SPLADE: Sparse Lexical and Expansion Model for First Stage Ranking , year =. Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =

  10. [10]

    SPLADE v2: Sparse Lexical and Expansion Model for Information Retrieval , publisher =

    Formal, Thibault and Lassance, Carlos and Piwowarski, Benjamin and Clinchant, Stéphane , keywords =. SPLADE v2: Sparse Lexical and Expansion Model for Information Retrieval , publisher =. 2021 , copyright =. doi:10.48550/ARXIV.2109.10086 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2109.10086 2021

  11. [11]

    From Distillation to Hard Negative Sampling: Making Sparse Neural IR Models More Effective , year =

    Formal, Thibault and Lassance, Carlos and Piwowarski, Benjamin and Clinchant, St\'. From Distillation to Hard Negative Sampling: Making Sparse Neural IR Models More Effective , year =. Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =. doi:10.1145/3477495.3531857 , abstract =

    arXiv

  12. [12]

    arXiv preprint arXiv:2403.06789 , year=

    SPLADE-v3: New baselines for SPLADE , author=. arXiv preprint arXiv:2403.06789 , year=

    Pith/arXiv arXiv

  13. [13]

    Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

    An efficiency study for splade models , author=. Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

  14. [14]

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

    Unified embeddings for multimodal retrieval via frozen LLMs , author=. Findings of the Association for Computational Linguistics: EACL 2024 , pages=

    2024

  15. [15]

    arXiv preprint arXiv:2603.13277 , year=

    Learning retrieval models with sparse autoencoders , author=. arXiv preprint arXiv:2603.13277 , year=

    arXiv

  16. [16]

    Findings of the association for computational linguistics: ACL 2024 , pages=

    M3-embedding: Multi-linguality, multi-functionality, multi-granularity text embeddings through self-knowledge distillation , author=. Findings of the association for computational linguistics: ACL 2024 , pages=

    2024

  17. [17]

    arXiv preprint arXiv:2601.01684 , year=

    LACONIC: Dense-Level Effectiveness for Scalable Sparse Retrieval via a Two-Phase Training Curriculum , author=. arXiv preprint arXiv:2601.01684 , year=

    arXiv

  18. [18]

    Proceedings of the 45th international ACM SIGIR conference on research and development in information retrieval , pages=

    From distillation to hard negative sampling: Making sparse neural ir models more effective , author=. Proceedings of the 45th international ACM SIGIR conference on research and development in information retrieval , pages=

  19. [19]

    arXiv preprint arXiv:2404.05961 , year=

    Llm2vec: Large language models are secretly powerful text encoders , author=. arXiv preprint arXiv:2404.05961 , year=

    Pith/arXiv arXiv

  20. [20]

    arXiv preprint arXiv:2408.11119 , year=

    Mistral-splade: Llms for better learned sparse retrieval , author=. arXiv preprint arXiv:2408.11119 , year=

    Pith/arXiv arXiv

  21. [21]

    European Conference on Information Retrieval , pages=

    Multimodal learned sparse retrieval with probabilistic expansion control , author=. European Conference on Information Retrieval , pages=. 2024 , organization=

    2024

  22. [22]

    Proceedings of the 34th ACM International Conference on Information and Knowledge Management , pages=

    Sparse and Dense Retrievers Learn Better Together: Joint Sparse-Dense Optimization for Text-Image Retrieval , author=. Proceedings of the 34th ACM International Conference on Information and Knowledge Management , pages=

  23. [23]

    VisualSparta: An embarrassingly simple approach to large-scale text-to-image search with weighted bag-of-words , author=. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) , pages=

  24. [24]

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

    Stair: Learning sparse text and image representation in grounded tokens , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    2023

  25. [25]

    Advances in neural information processing systems , volume=

    Laion-5b: An open large-scale dataset for training next generation image-text models , author=. Advances in neural information processing systems , volume=

  26. [26]

    arXiv preprint arXiv:2111.02114 , year=

    Laion-400m: Open dataset of clip-filtered 400 million image-text pairs , author=. arXiv preprint arXiv:2111.02114 , year=

    Pith/arXiv arXiv

  27. [27]

    Computer Vision--ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13 , pages=

    Microsoft coco: Common objects in context , author=. Computer Vision--ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13 , pages=. 2014 , organization=

    2014

  28. [28]

    Proceedings of the IEEE international conference on computer vision , pages=

    Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models , author=. Proceedings of the IEEE international conference on computer vision , pages=

  29. [29]

    Direct preference optimization of video large multimodal models from language model reward , author=. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    2025

  30. [30]

    International Conference on Learning Representations , volume=

    Visrag: Vision-based retrieval-augmented generation on multi-modality documents , author=. International Conference on Learning Representations , volume=

  31. [31]

    International Conference on Learning Representations , volume=

    Colpali: Efficient document retrieval with vision language models , author=. International Conference on Learning Representations , volume=

  32. [32]

    2025 , url=

    Song, Tingyu and Gan, Guo and Shang, Mingsheng and Zhao, Yilun , booktitle=. 2025 , url=

    2025

  33. [33]

    2025 , url=

    Weller, Orion and Chang, Benjamin and MacAvaney, Sean and Lo, Kyle and Cohan, Arman and Van Durme, Benjamin and Lawrie, Dawn and Soldaini, Luca , booktitle=. 2025 , url=

    2025

  34. [34]

    arXiv preprint arXiv:2402.14334 , year=

    Instructir: A benchmark for instruction following of information retrieval models , author=. arXiv preprint arXiv:2402.14334 , year=

    Pith/arXiv arXiv

  35. [35]

    2026 , eprint=

    MRMR: A Realistic and Expert-Level Multidisciplinary Benchmark for Reasoning-Intensive Multimodal Retrieval , author=. 2026 , eprint=

    2026

  36. [36]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track , year=

    SSRB: Direct Natural Language Querying to Massive Heterogeneous Semi-Structured Data , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track , year=

  37. [37]

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

    Beyond Global Similarity: Towards Fine-Grained, Multi-Condition Multimodal Retrieval , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=. 2603.01082 , archivePrefix=

    arXiv

  38. [38]

    Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2) , year=

    BEIR: A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models , author=. Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2) , year=

  39. [39]

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

    MTEB: Massive Text Embedding Benchmark , author=. Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics , pages=

  40. [40]

    ArXiv , year=

    BrowseComp-Plus: A More Fair and Transparent Evaluation Benchmark of Deep-Research Agent , author=. ArXiv , year=

  41. [41]

    Proceedings of the 33rd ACM International Conference on Multimedia , pages=

    Breaking the modality barrier: Universal embedding learning with multimodal llms , author=. Proceedings of the 33rd ACM International Conference on Multimedia , pages=

  42. [42]

    Xiao, Zilin and Ma, Qi and Gu, Mengting and Chen, Chun-Cheng Jason and Chen, Xintao and Ordonez, Vicente and Mohan, Vijai , journal=

  43. [43]

    arXiv preprint arXiv:2510.05014 , year=

    Think Then Embed: Generative Context Improves Multimodal Embedding , author=. arXiv preprint arXiv:2510.05014 , year=

    arXiv

  44. [44]

    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 , journal=

  45. [45]

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

    mme5: Improving multimodal multilingual embeddings via high-quality synthetic data , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  46. [46]

    arXiv preprint arXiv:2507.14902 , year=

    U-marvel: Unveiling key factors for universal multimodal retrieval via embedding learning with mllms , author=. arXiv preprint arXiv:2507.14902 , year=

    arXiv

  47. [47]

    The Thirteenth International Conference on Learning Representations , year=

    VLM2Vec: Training Vision-Language Models for Massive Multimodal Embedding Tasks , author=. The Thirteenth International Conference on Learning Representations , year=

  48. [48]

    arXiv preprint arXiv:2507.04590 , year=

    Vlm2vec-v2: Advancing multimodal embedding for videos, images, and visual documents , author=. arXiv preprint arXiv:2507.04590 , year=

    Pith/arXiv arXiv

  49. [49]

    Zhang, Xin and Zhang, Yanzhao and Xie, Wen and Li, Mingxin and Dai, Ziqi and Long, Dingkun and Xie, Pengjun and Zhang, Meishan and Li, Wenjie and Zhang, Min , journal=

  50. [50]

    arXiv preprint arXiv:2510.27571 , year=

    Towards Universal Video Retrieval: Generalizing Video Embedding via Synthesized Multimodal Pyramid Curriculum , author=. arXiv preprint arXiv:2510.27571 , year=

    arXiv

  51. [51]

    arXiv preprint arXiv:2601.04720 , year=

    Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking , author=. arXiv preprint arXiv:2601.04720 , year=

    Pith/arXiv arXiv

  52. [52]

    arXiv preprint arXiv:2510.27350 , year=

    Rzenembed: Towards comprehensive multimodal retrieval , author=. arXiv preprint arXiv:2510.27350 , year=

    arXiv

  53. [53]

    arXiv preprint arXiv:2602.13823 , year=

    Embed-rl: Reinforcement learning for reasoning-driven multimodal embeddings , author=. arXiv preprint arXiv:2602.13823 , year=

    arXiv

  54. [54]

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

    One Embedder, Any Task: Instruction-Finetuned Text Embeddings , author=. Findings of the Association for Computational Linguistics: ACL 2023 , pages=

    2023

  55. [55]

    arXiv preprint arXiv:2401.00368 , year=

    Improving text embeddings with large language models , author=. arXiv preprint arXiv:2401.00368 , year=

    Pith/arXiv arXiv

  56. [56]

    arXiv preprint arXiv:2402.09906 , year=

    Generative representational instruction tuning , author=. arXiv preprint arXiv:2402.09906 , year=

    Pith/arXiv arXiv

  57. [57]

    International Conference on Learning Representations , volume=

    Repetition improves language model embeddings , author=. International Conference on Learning Representations , volume=

  58. [58]

    2025 , url =

    OpenSearch-AI/Ops-MM-embedding-v1-7B , author =. 2025 , url =

    2025

  59. [59]

    Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year=

    Efficient Memory Management for Large Language Model Serving with PagedAttention , author=. Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year=

  60. [60]

    Advances in neural information processing systems , volume=

    Sglang: Efficient execution of structured language model programs , author=. Advances in neural information processing systems , volume=

  61. [61]

    NLTK : The Natural Language Toolkit

    Bird, Steven and Loper, Edward. NLTK : The Natural Language Toolkit. Proceedings of the ACL Interactive Poster and Demonstration Sessions. 2004

    2004

  62. [62]

    arXiv preprint arXiv:2004.05665 , year=

    Minimizing flops to learn efficient sparse representations , author=. arXiv preprint arXiv:2004.05665 , year=

    Pith/arXiv arXiv 2004

  63. [63]

    arXiv preprint arXiv:2602.21456 , year=

    Revisiting Text Ranking in Deep Research , author=. arXiv preprint arXiv:2602.21456 , year=

    Pith/arXiv arXiv

  64. [64]

    and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle=

    Hu, Edward J. and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle=. 2022 , url=

    2022

  65. [65]

    2025 , howpublished=

    Tongyi DeepResearch Team , title=. 2025 , howpublished=

    2025

  66. [66]

    Bert: Pre-training of deep bidirectional transformers for language understanding , 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