Pith. sign in

REVIEW 3 major objections 6 minor 39 references

The paper claims that retrieval computation belongs in a mid-to-late layer interval, and that looping that interval with shared weights and small learnable registers improves multimodal retrieval at lower latency than token reasoning.

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

2026-08-03 00:29 UTC pith:A4HZEKFL

load-bearing objection A solid, honest empirical paper worth refereeing, with the main caveat that the stage-localization rationale is not re-validated on the recurrent model and hyperparameters are tuned on the test set. the 3 major comments →

arxiv 2607.28751 v1 pith:A4HZEKFL submitted 2026-07-30 cs.CV

ReLoop-UME: Recurrent Depth with Learnable Retrieval Registers for Universal Multimodal Embedding

Shijie Wang , Xiangzhao Hao , Yueti Li , Guangyu Cao , Xinyu Tang , Haiyun Guo This is my paper
classification cs.CV
keywords universal multimodal embeddingrecurrent depthlearnable retrieval registerslayer-wise retrieval discriminabilityMMEB-V2multimodal retrievalparameter sharinglatent reasoning
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 sets out to establish that the best place to add computation to a universal multimodal embedding model is along depth, not along tokens. By measuring how well each layer of already-trained embedding models separates positive from negative query-candidate pairs, the authors find a consistent three-stage progression: input contextualization, a contiguous middle-to-late interval where retrieval discrimination forms, and final embedding mapping. Their method executes the early layers once, reuses the parameter-shared retrieval-forming block for four loops, and reads the embedding from a small set of learnable retrieval registers after the final loop. On MMEB-V2 this reaches 63.2 All at 2B scale and 65.9 at 7B, beating reasoning-token methods while running 44.9x faster than UME-R1 and 1.5x faster than PLUME. The paper also reports that video retrieval remains a weakness and that zero-shot transfer to MRMR is modest outside knowledge and theorem groups.

Core claim

The central claim is that retrieval-discriminative computation in UME models is localized to a mid-to-late layer interval, so recurrently reusing exactly that interval—with shared weights, a fixed token workspace, and a compact trainable register state—produces a better terminal embedding than either a single forward pass or token-level reasoning. The analysis uses a layer-wise positive-negative margin measured with an 80th-percentile negative quantile, which identifies layers 17-26 on 28-layer Qwen-VL backbones and layers 12-22 on Qwen3.5-2B as the Retrieval Formation Stage. ReLoop-UME appends five learnable retrieval registers to the multimodal sequence, runs the prefix once, applies the s

What carries the argument

The carrying mechanism is the localized recurrent block together with Learnable Retrieval Registers. Layer-wise retrieval discriminability S_l, defined as the positive similarity minus the 80th-percentile negative similarity at each layer, is used to split the network into prefix, retrieval-forming block, and suffix via a constrained three-segment change-point fit. The chosen block G is executed T times with shared parameters and unchanged attention masks and positions; the M registers are trainable tokens appended after content tokens so they attend to the entire input, persist across loops, and exchange evidence register-to-register, with only the final register read out after the suffix.

Load-bearing premise

The load-bearing premise is that the layer interval where positive-negative separation rises in an independently trained single-forward model is still the correct interval to loop once the architecture is trained recurrently; the split is fixed before recurrent training and is not re-validated during or after unrolling.

What would settle it

A decisive check: train ReLoop-UME with recurrence over a randomly chosen 10-layer interval matched for decoder applications, using the same T, M, and training recipe on MMEB-V2. If the random interval matches the 63.2 All result of layers 17-26, the localization claim is not load-bearing; if it is substantially worse, stage localization is doing the work.

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

If this is right

  • Given the three-stage progression, extra compute for retrieval can be added along depth with a fixed token workspace, avoiding the serial latency and intermediate-state dependence of explicit or latent reasoning tokens.
  • The stage-localization recipe transfers across Qwen-VL backbones: the same method improves MMEB-V2 All by +2.6, +2.2, and +2.4 points on Qwen2-VL-2B, Qwen3-VL-2B, and Qwen3.5-2B.
  • The 7B checkpoint trained only on MMEB-V2 transfers zero-shot to MRMR, leading Knowledge (71.6) and Theorem (39.1) groups with average 48.6, while remaining weak on Contradiction and Negation.
  • Recurrence over layers 17-26 with T=4 and M=5 yields 63.2 All at 201 ms/sample, 44.9x faster than UME-R1 and 1.5x faster than PLUME, placing the method on a favorable accuracy-efficiency curve.
  • Ablations show both localization and registers matter: looping the full stack (112 decoder applications) is 1.9 points worse than looping layers 17-26 (58 applications), and removing registers drops All from 63.2 to 61.8.

Where Pith is reading between the lines

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

  • If the localization-then-loop principle is generic, it should be testable as a drop-in depth intervention for any single-forward embedding model: measure the rising interval on a new backbone, loop it, and compare at matched decoder applications.
  • The register mechanism suggests a natural adaptive-depth controller: monitor how much the final register's representation changes across loops and stop when it saturates, making the extra computation input-dependent.
  • An alternative explanation for the gain is simply additional depth rather than stage specificity; a matched-depth non-recurrent baseline would help separate those mechanisms.
  • Because the token workspace stays fixed, the register idea could combine with document-region or video-frame-specific registers to address the local-detail and temporal-boundary failures the paper documents.

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 / 6 minor

Summary. The paper proposes ReLoop-UME, a universal multimodal embedding method that expands computation along model depth by recurrently reusing a parameter-shared block of layers identified as the 'retrieval formation stage' via layer-wise positive–negative similarity analysis on a single-forward UME model. Learnable Retrieval Registers serve as a fixed-size persistent workspace, and the final register is used as the embedding readout. The model is trained with terminal InfoNCE. On MMEB-V2, ReLoop-UME achieves All scores of 63.2 at ~2B and 65.9 at ~7B, outperforming UME-R1 and PLUME, with reported 44.9x and 1.5x latency reductions, respectively; zero-shot transfer to MRMR is also reported. Controlled ablations vary the recurrent interval, register count M, and recurrent depth T.

Significance. The paper proposes a novel and potentially efficient alternative to token-expansion reasoning in UME: depth-wise recurrence with a fixed token workspace. The layer-wise stage analysis is an interesting diagnostic that could inform future architecture design. The empirical gains are consistent across three Qwen-VL backbones (+2.2 to +2.6 points over single-forward), and the efficiency advantages are substantial. The paper includes a thorough ablation suite, a 2x2 factorial experiment, and detailed reproducibility settings. However, two methodological concerns—the untested transfer of the stage localization to the recurrent model and the selection of hyperparameters on the test benchmark—currently temper the strength of the claims. The central empirical result is nonetheless well supported by controlled comparisons.

major comments (3)
  1. [Section 3.2/3.3, Eq. (2)-(3)] The recurrent block boundaries are fixed from an independently trained single-forward UME model and are never re-validated on the trained recurrent model. The interval ablation (Table 3) supports 17–26 as a good fixed choice among the tested alternatives, but it does not establish that, after recurrent training, the largest sustained rise in S_l still lies inside the repeated block, nor that a split chosen by re-running the change-point criterion on the recurrent model would not be better. This directly affects the 'localize-then-loop' rationale and the accuracy–efficiency statement in Eq. (5). Please either re-run the S_l analysis on the recurrent model (e.g., after each loop or on the final hidden states) and report whether the discriminative rise is still concentrated in the reused interval, or explicitly state that the interval is an empirically effective fixed hyperparameter rather
  2. [Section 4.4 and Supplement H.3] The default hyperparameters (layers 17–26, M=5, T=4) are selected by scanning the MMEB-V2 All score on the test set. Tables 3–5 and the interaction analysis in Supplement H.3 are computed on the test benchmark, and the best configuration is then used for the main comparisons. This risks overfitting the benchmark and may inflate the reported gains. Please provide a validation-based selection procedure (e.g., a held-out split of the 78 tasks or a separate validation set) or a stability analysis showing that the chosen hyperparameters are insensitive to seeds and backbone variants. This is important for assessing the absolute improvement claims.
  3. [Section 4.3, Table 2] The headline speedups (44.9x vs UME-R1 and 1.5x vs PLUME) rest on latency measurements whose protocol is not fully specified. It is not stated whether the baselines are run using official released code or a reimplementation, whether the input preprocessing and batch sizes are identical across methods, or whether the reported latency includes all decoding steps for the autoregressive baselines. If the baselines are not run in a matched setting, the efficiency claims may be misleading. Please provide a detailed protocol (implementation provenance, warm-up, input lengths, decoding steps, and measurement methodology) for each baseline.
minor comments (6)
  1. [Figure 5] Typo: 'Reloop-UME' should be 'ReLoop-UME'.
  2. [Section 3.2] The notation s^+_{i,l} and s^-_{i,l} in Eq. (2) is used as if defined; please define them explicitly as the positive and the Q_0.8 negative similarity values for sample i at layer l.
  3. [Section 3.3] The sentence listing stage boundaries is ambiguous: 'This yields layers 0–16, 17–26, and 27 for Qwen2-VL-2B, Qwen2-VL-7B, and Qwen3-VL-2B' could be read as all three using identical boundaries. Make explicit that Qwen2-VL-7B uses the same partition as Qwen2-VL-2B, as stated later.
  4. [Section 4.5 / Supplement F] The MRMR aggregates in Table 6 differ from the official per-subtask numbers reproduced in Supplement Table 10. The supplementary explains evaluation-snapshot alignment, but the main text should note this explicitly so readers do not attempt to average the per-subtask cells from Table 10 to recover Table 6.
  5. [Section 4.2] The paper clearly acknowledges the video limitation (ReLoop-UME scores below PLUME and UME-R1 on video). This is a strength, but the abstract and conclusion might mention this limitation alongside the overall gains to avoid overgeneralization.
  6. [Section H.1 / Table 15] The full-decoder recurrence uses 112 decoder applications vs 58 for layers 17–26, so the comparison is not iso-compute. The conclusion that 'additional computation alone is insufficient' is reasonable, but the phrase 'more computation' should be qualified by the compute difference.

Circularity Check

0 steps flagged

No significant circularity: the central claims rest on external benchmark evaluation and controlled ablations, not on definitions or self-citations.

full rationale

The paper's stage localization (Eq. 2) is an empirical measurement on an independently trained single-forward UME model; the recurrent architecture uses the resulting interval as a fixed hyperparameter chosen before recurrent training. The claimed improvement is then evaluated on the external MMEB-V2 and MRMR benchmarks after full training, and the interval ablation in Table 3 compares fixed layer intervals on the test benchmark. No equation in the paper reduces the final All score to the S_l curve, and no fitted parameter is renamed as a prediction: the terminal InfoNCE objective is applied only to the recurrently produced embedding, and the gains are measured rather than derived. The efficiency formula D_loop(T)=Lp+T Lr+Ls is definitional, but the paper supplements it with independently measured latency, so it is not used as a substitute for empirical evaluation. Prior works by overlapping authors (PLUME, TRACE) appear only as baselines and related work; no load-bearing self-citation, uniqueness theorem, or ansatz-smuggling citation is used to justify the central claim. The fixed stage interval is an untested transfer assumption from single-forward analysis to recurrent training, but that is a modeling risk, not circularity. The paper is therefore self-contained against external benchmarks, and no circular step is identified.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The central method rests on a small number of hand-chosen hyperparameters (T, M, tau) and on the stage interval derived from the single-forward diagnostic. The most consequential choice is the interval itself: it is selected before recurrent training and never re-validated after the model becomes recurrent. The registers are an invented trainable component, but with modest parameter cost and clear ablative support.

free parameters (4)
  • Recurrent depth T = 4
    Selected by ablating T in {1,2,4,8} on MMEB-V2 All (Table 5); directly controls the accuracy-efficiency tradeoff.
  • Register count M = 5
    Selected by ablating M in {0,5,8,10} on MMEB-V2 All (Table 4); adds only 7,680 parameters at 2B scale.
  • Stage boundaries (a,b) = 17-26 for 28-layer Qwen backbones; 12-22 for Qwen3.5-2B
    Fit by a pooled piecewise-linear change-point algorithm to training-set layer-wise separability S_l of a single-forward model (Eq. 2). Not tuned on final retrieval accuracy, but not re-derived after recurrent training.
  • Contrastive temperature tau = 0.05
    Hand-chosen temperature for InfoNCE; affects gradient scale and absolute retrieval scores.
axioms (4)
  • domain assumption Layer-wise normalized readout discriminability S_l on the training set of a single-forward UME identifies the functional stage that should be recurrently reused, and this choice remains valid after recurrent training.
    The whole localized-recurrence design depends on Section 3.2's change-point localization; no causal intervention or post-training re-localization is provided.
  • domain assumption Terminal InfoNCE with in-batch negatives is sufficient to shape all unrolled recurrent states; no per-loop supervision or auxiliary loss is needed.
    Section 3.4 applies supervision only after the final loop; this is a training assumption, supported only indirectly by the ablations.
  • domain assumption Causal attention with registers appended after the input allows the final register to read all content and earlier registers, while content tokens never see registers.
    Section 3.3 relies on this masking scheme; alternative register placements or bidirectional access are not tested.
  • domain assumption MMEB-V2 and MRMR official evaluation protocols and candidate pools measure retrieval without test-set leakage or contamination.
    Used throughout Section 4; no contamination check is reported, and some MMEB-V2 evaluation tasks share dataset names with training sources.
invented entities (1)
  • Learnable Retrieval Registers (R = [r_1 ... r_M]) no independent evidence
    purpose: Persistent fixed-size workspace appended to multimodal tokens; accumulates and exchanges evidence across loops, with the final register serving as the embedding readout.
    Supported only by in-paper ablations (Table 4 and the factorial Table 17); no external falsifiable prediction is offered. Conceptually similar to register/memory tokens in prior literature.

pith-pipeline@v1.3.0-alltime-deepseek · 26881 in / 16387 out tokens · 179656 ms · 2026-08-03T00:29:38.124588+00:00 · methodology

0 comments
read the original abstract

Universal multimodal embedding (UME) maps heterogeneous multimodal inputs into a shared embedding space. Existing UME models either form embeddings through single forward encoding or add computation through explicit rationale tokens and latent autoregressive states. Although token expansion can improve complex matching, serial generation increases retrieval latency and makes the final embedding depend on generated intermediate states. This raises a different question: can useful computation be expanded along model depth while keeping the token workspace fixed? We analyze positive-negative similarity separation at every layer of independently trained UME models and observe a shared progression: early layers contextualize multimodal inputs, a contiguous middle-to-late stage forms retrieval-discriminative features, and the final layers map them into the embedding space. Based on this finding, we propose ReLoop-UME, which executes the early layers once, recurrently reuses a parameter-shared retrieval-forming block, and applies the final mapping layers after the last loop. Learnable Retrieval Registers provide persistent retrieval-specific states that accumulate and exchange evidence across loops, with the final register serving as the embedding readout. On MMEB-V2 and MRMR, ReLoop-UME consistently improves retrieval across different backbones while running 44.9x faster than UME-R1 and 1.5x faster than PLUME.

Figures

Figures reproduced from arXiv: 2607.28751 by Guangyu Cao, Haiyun Guo, Shijie Wang, Xiangzhao Hao, Xinyu Tang, Yueti Li.

Figure 1
Figure 1. Figure 1: Abstract layer-wise retrieval-discriminability trend [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of UME computation paths. Existing methods allocate computation through either single forward [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the recurrent retrieval encoder. Multimodal tokens and learnable retrieval registers are processed once by [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Task-group comparison among 2B models on [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: ReLoop-UME achieves a favorable accuracy– [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 1
Figure 1. Figure 1: Layer-wise positive–negative separation for Qwen2-VL-2B and Qwen3-VL-2B. Both 28-layer backbones exhibit a [PITH_FULL_IMAGE:figures/full_fig_p014_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Layer-wise separation for the 24-layer Qwen3.5-2B hybrid stack. All four quantiles select prefix layers 0–11, retrieval [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Explicit-CoT failures caused by irrelevant-information hijacking and query-intent over-compression. In (a), the generated rationale follows visually salient company-overview cues and retrieves an “About” page rather than the shareholder message. In (b), a long dialogue containing multiple positive and negative constraints is reduced to the generic concept “a lone skier in snow,” removing the details that s… view at source ↗
Figure 4
Figure 4. Figure 4: Explicit-CoT failures in which a plausible linguistic prior overrides the discriminative visual evidence. The correct [PITH_FULL_IMAGE:figures/full_fig_p019_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Four PLUME cases selected to expose fine-grained category collapse, local OCR loss, temporal-state confusion, and [PITH_FULL_IMAGE:figures/full_fig_p020_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Three representative ReLoop-UME failures. The green item is the positive candidate and the red item is the retrieved [PITH_FULL_IMAGE:figures/full_fig_p021_6.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

39 extracted references · 12 linked inside Pith

  1. [1]

    Bai, S.; et al. 2025. Qwen3-VL Technical Report. arXiv:2511.21631

    Pith/arXiv arXiv 2025

  2. [2]

    N.; Awasthi, A.; Pan, X.; Ahuja, C.; Mishra, S

    Cui, X.; Cheng, J.; Chen, H.-y.; Shukla, S. N.; Awasthi, A.; Pan, X.; Ahuja, C.; Mishra, S. K.; Guo, Q.; Lim, S.-N.; Singh, A.; and Fan, X. 2025. Think Then Embed: Generative Context Improves Multimodal Embedding. arXiv:2510.05014

    arXiv 2025

  3. [3]

    Dehghani, M.; Gouws, S.; Vinyals, O.; Uszkoreit, J.; and Kaiser, L. 2019. Universal Transformers. In International Conference on Learning Representations

    2019

  4. [4]

    Fan, Y.; Du, Y.; Ramchandran, K.; and Lee, K. 2025. Looped Transformers for Length Generalization. In International Conference on Learning Representations

    2025

  5. [5]

    Faysse, M.; Sibille, H.; Wu, T.; Omrani, B.; Viaud, G.; Hudelot, C.; and Colombo, P. 2025. ColPali : Efficient Document Retrieval with Vision Language Models. In International Conference on Learning Representations

    2025

  6. [6]

    R.; Kailkhura, B.; Bhatele, A.; and Goldstein, T

    Geiping, J.; McLeish, S.; Jain, N.; Kirchenbauer, J.; Singh, S.; Bartoldson, B. R.; Kailkhura, B.; Bhatele, A.; and Goldstein, T. 2025. Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach. In Advances in Neural Information Processing Systems, volume 38

    2025

  7. [7]

    Gu, T.; Yang, K.; Feng, Z.; Wang, X.; Zhang, Y.; Long, D.; Chen, Y.; Cai, W.; and Deng, J. 2025. Breaking the Modality Barrier: Universal Embedding Learning with Multimodal LLM s. arXiv:2504.17432

    arXiv 2025

  8. [8]

    Hao, X.; Wang, S.; Yang, T.; Wang, T.; Guo, H.; and Wang, J. 2026. TRACE : Task-Adaptive Reasoning and Representation Learning for Universal Multimodal Retrieval. arXiv:2603.02929

    arXiv 2026

  9. [9]

    He, C.; Hao, X.; Yang, T.; Ma, Y.; Jia, Y.; Wu, L.; Zhao, C.; Guo, H.; and Wang, J. 2026. PLUME : Latent Reasoning Based Universal Multimodal Embedding. arXiv:2604.02073

    Pith/arXiv arXiv 2026

  10. [10]

    Jia, C.; Yang, Y.; Xia, Y.; Chen, Y.-T.; Parekh, Z.; Pham, H.; Le, Q.; Sung, Y.-H.; Li, Z.; and Duerig, T. 2021. Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision. In Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, 4904--4916. PMLR

    2021

  11. [11]

    Jiang, H.; Wang, Y.; Zhu, Y.; Lu, X.; Qin, W.; Wang, M.; Wan, P.; and Tang, Y. 2026. Embed-RL : Reinforcement Learning for Reasoning-Driven Multimodal Embeddings. arXiv:2602.13823

    arXiv 2026

  12. [12]

    Jiang, T.; Song, M.; Zhang, Z.; Huang, H.; Deng, W.; Sun, F.; Zhang, Q.; Wang, D.; and Zhuang, F. 2024. E5-V : Universal Embeddings with Multimodal Large Language Models. arXiv:2407.12580

    Pith/arXiv arXiv 2024

  13. [13]

    Jiang, Z.; Meng, R.; Yang, X.; Yavuz, S.; Zhou, Y.; and Chen, W. 2025. VLM2Vec : Training Vision-Language Models for Massive Multimodal Embedding Tasks. In The Thirteenth International Conference on Learning Representations

    2025

  14. [14]

    Koishekenov, Y.; Lipani, A.; and Cancedda, N. 2025. Encode, Think, Decode: Scaling Test-Time Reasoning with Recursive Latent Thoughts. arXiv preprint arXiv:2510.07358

    arXiv 2025

  15. [15]

    Lan, Z.; Niu, L.; Meng, F.; Zhou, J.; and Su, J. 2026. UME-R1 : Exploring Reasoning-Driven Generative Multimodal Embeddings. In The Fourteenth International Conference on Learning Representations

    2026

  16. [16]

    Li, M.; Zhang, Y.; Long, D.; Chen, K.; Song, S.; Bai, S.; Yang, Z.; Xie, P.; Yang, A.; Liu, D.; Zhou, J.; and Lin, J. 2026. Qwen3-VL-Embedding and Qwen3-VL-Reranker : A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking. arXiv:2601.04720

    Pith/arXiv arXiv 2026

  17. [17]

    Lin, S.-C.; Lee, C.; Shoeybi, M.; Lin, J.; Catanzaro, B.; and Ping, W. 2024. MM-Embed : Universal Multimodal Retrieval with Multimodal LLM s. arXiv:2411.02571

    Pith/arXiv arXiv 2024

  18. [18]

    Liu, C.; Yang, J.; Gao, R.; Zhu, Y.; Zhu, F.; Zhao, R.; and Wang, L. 2025 a . Reasoning Guided Embeddings: Leveraging MLLM Reasoning for Improved Multimodal Retrieval. arXiv:2511.16150

    arXiv 2025

  19. [19]

    Liu, Y.; Zhang, Y.; Cai, J.; Jiang, X.; Hu, Y.; Yao, J.; Wang, Y.; and Xie, W. 2025 b . LamRA : Large Multimodal Model as Your Advanced Retrieval Assistant. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 4015--4025

    2025

  20. [20]

    Meng, R.; Jiang, Z.; Liu, Y.; Su, M.; Yang, X.; Fu, Y.; Qin, C.; Thirukovalluru, R.; Zhang, X.; Chen, Z.; Xu, R.; Xiong, C.; Zhou, Y.; Chen, W.; and Yavuz, S. 2026. VLM 2Vec-V2 : Advancing Multimodal Embedding for Videos, Images, and Visual Documents. Transactions on Machine Learning Research

    2026

  21. [21]

    Park, T.; Lee, Y.; Kim, D.; and Bae, H. 2026. LoopUS : Recasting Pretrained LLM s into Looped Latent Refinement Models. arXiv:2605.11011

    Pith/arXiv arXiv 2026

  22. [22]

    Qwen Team . 2026. Qwen3.5 : Towards Native Multimodal Agents

    2026

  23. [23]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; Krueger, G.; and Sutskever, I

    Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; Krueger, G.; and Sutskever, I. 2021. Learning Transferable Visual Models From Natural Language Supervision. In Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, ...

    2021

  24. [24]

    Saunshi, N.; Dikkala, N.; Li, Z.; Kumar, S.; and Reddi, S. J. 2025. Reasoning with Latent Thoughts: On the Power of Looped Transformers. In International Conference on Learning Representations

    2025

  25. [25]

    Sun, S.; Ren, J.; Liao, Z.; Mao, D.; Ren, X.; Zhang, Y.; Zhao, H.; Lin, W.; Jiang, S.; Zhang, L.; and Zheng, Y. 2026. Bottleneck Tokens for Unified Multimodal Retrieval. arXiv:2604.11095

    Pith/arXiv arXiv 2026

  26. [26]

    Wang, P.; Bai, S.; Tan, S.; Wang, S.; Fan, Z.; Bai, J.; Chen, K.; Liu, X.; Wang, J.; Ge, W.; Fan, Y.; Dang, K.; Du, M.; Ren, X.; Men, R.; Liu, D.; Zhou, C.; Zhou, J.; and Lin, J. 2024. Qwen2-VL : Enhancing Vision-Language Model's Perception of the World at Any Resolution. arXiv:2409.12191

    Pith/arXiv arXiv 2024

  27. [27]

    Wei, C.; Chen, Y.; Chen, H.; Hu, H.; Zhang, G.; Fu, J.; Ritter, A.; and Chen, W. 2024. UniIR : Training and Benchmarking Universal Multimodal Information Retrievers. In European Conference on Computer Vision, 387--404. Springer

    2024

  28. [28]

    Xiao, Z.; Ma, Q.; Gu, M.; cheng Jason Chen, C.; Chen, X.; Ordonez, V.; and Mohan, V. 2026. MetaEmbed: Scaling Multimodal Retrieval at Test-Time with Flexible Late Interaction. arXiv:2509.18095

    Pith/arXiv arXiv 2026

  29. [29]

    Zhai, X.; Mustafa, B.; Kolesnikov, A.; and Beyer, L. 2023. Sigmoid Loss for Language Image Pre-Training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 11975--11986

    2023

  30. [30]

    Zhang, K.; Luan, Y.; Hu, H.; Lee, K.; Qiao, S.; Chen, W.; Su, Y.; and Chang, M.-W. 2024. MagicLens : Self-Supervised Image Retrieval with Open-Ended Instructions. arXiv:2403.19651

    Pith/arXiv arXiv 2024

  31. [31]

    T.; and Zhao, C

    Zhang, S.; Gao, Y.; Zhou, X.; Zhao, Y.; Song, T.; Cohan, A.; Luu, A. T.; and Zhao, C. 2026. MRMR : A Realistic and Expert-Level Multidisciplinary Benchmark for Reasoning-Intensive Multimodal Retrieval. In The Fourteenth International Conference on Learning Representations

    2026

  32. [32]

    Zhang, X.; Zhang, Y.; Xie, W.; Li, M.; Dai, Z.; Long, D.; Xie, P.; Zhang, M.; Li, W.; and Zhang, M. 2025. Bridging Modalities: Improving Universal Multimodal Retrieval by Multimodal Large Language Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 9274--9285

    2025

  33. [33]

    J.; and Lian, D

    Zhou, J.; Xiong, Y.; Liu, Z.; Liu, Z.; Xiao, S.; Wang, Y.; Zhao, B.; Zhang, C. J.; and Lian, D. 2025. MegaPairs : Massive Data Synthesis for Universal Multimodal Retrieval. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 19076--19095

    2025

  34. [34]

    Zhu, R.-J.; Wang, Z.; Hua, K.; Zhang, T.; Li, Z.; Que, H.; Wei, B.; Wen, Z.; Yin, F.; Xing, H.; Li, L.; Shi, J.; Ma, K.; Li, S.; Kergan, T.; Smith, A.; Qu, X.; Hui, M.; Wu, B.; Min, Q.; Huang, H.; Zhou, X.; Ye, W.; Liu, J.; Yang, J.; Shi, Y.; Lin, C.; Zhao, E.; Cai, T.; Zhang, G.; Huang, W.; Bengio, Y.; and Eshraghian, J. 2025. Scaling Latent Reasoning vi...

    Pith/arXiv arXiv 2025

  35. [35]

    Qwen Team . 2024. Qwen2-VL-2B-Instruct : Official Model Configuration. Hugging Face

    2024

  36. [36]

    Qwen Team . 2025 a . Qwen3 Embedding : Advancing Text Embedding and Reranking Through Foundation Models. Qwen technical blog

    2025

  37. [37]

    Qwen Team . 2025 b . Qwen3-VL-2B-Instruct : Official Model Configuration. Hugging Face

    2025

  38. [38]

    Qwen Team . 2026. Qwen3.5-2B : Official Model Configuration. Hugging Face

    2026

  39. [39]

    Zhang, Y.; Li, M.; Long, D.; Zhang, X.; Lin, H.; Yang, B.; Xie, P.; Yang, A.; Liu, D.; Lin, J.; Huang, F.; and Zhou, J. 2025. Qwen3 Embedding : Advancing Text Embedding and Reranking Through Foundation Models. arXiv:2506.05176

    Pith/arXiv arXiv 2025