Pith. sign in

REVIEW 3 major objections 6 minor 46 references

KoVRE: Training an Efficient Embedding Model for Korean Visual Document Retrieval

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper claims that a compact 2-billion-parameter single-vector retriever, KoVRE, can beat both an 8B single-vector model and a 4B multi-vector baseline on Korean visual document retrieval.

desk verdict The recipe is real and the 2B-vs-8B result holds up; the multi-vector 'outperforms' claim is a 0.0034 margin with no variance, and train/eval disjointness is never documented. read the letter →

arxiv 2608.01389 v1 pith:QQAPIQD7 submitted 2026-08-02 cs.AI

classification cs.AI
keywords Koreanvisualdocumentretrievalsingle-vectorembeddinghard-negativeminingknowledgedistillationMatryoshkarepresentationlearningbilingualtrainingQwen3-VL-EmbeddingnDCG@10
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 claims that a compact 2-billion-parameter single-vector retriever, KoVRE, can beat both an 8B single-vector model and a 4B multi-vector baseline on Korean visual document retrieval. It trains the model in two stages: contrastive learning on 708,729 Korean and English query–page pairs with positive-aware hard-negative mining, then knowledge distillation from a reranker teacher on Korean data. Across the KoViDoRe and SDS KoPub VDR benchmarks, KoVRE reaches an average nDCG@10 of 0.5919, versus 0.5489 for the 8B model and 0.5885 for the multi-vector baseline. The broader claim is that targeted bilingual supervision and careful training design can substitute for a larger backbone or storage-heavy multi-vector representations.

What carries the argument

The load-bearing mechanism is the two-stage training recipe: Stage 1 contrastive learning with positive-aware hard-negative mining (selecting hard negatives while filtering out candidates whose similarity exceeds 95% of the positive, plus stop-gradient self-guide filtering and hardness weighting), and Stage 2 reranker-based knowledge distillation with per-query min–max score normalization. Matryoshka Representation Learning is applied throughout, training truncated embeddings jointly so that the retriever works at 128–2,048 dimensions.

What would settle it

Remove from the training corpus every page that appears in the KoViDoRe evaluation split, retrain KoVRE under identical settings, and compare average nDCG@10 with the reported 0.5919; a material drop would show the gain depends on train/eval overlap.

Watch

Extended reading notes

Core claim

KoVRE is a single-vector Korean visual document retriever initialized from Qwen3-VL-Embedding-2B and trained in two stages. Stage 1 uses an InfoNCE contrastive objective over paired positives, seven mined hard negatives per query, and in-batch negatives, with self-guide filtering that masks potential false negatives and hardness weighting that amplifies the gradient from hard negatives; training is done jointly across Matryoshka dimensions from 128 to 2,048. Stage 2 distills Qwen3-VL-Reranker-8B scores into the student by minimizing KL divergence over a per-query candidate set of one positive and eight teacher-selected negatives, with per-query min–max normalization applied to both teacher a

Load-bearing premise

The headline numbers assume the KoViDoRe evaluation set shares no pages with the ko-vdr-train-public training set; the paper never states that the evaluation split excludes training pages.

Editorial extensions

If this is right

  • If KoVRE's results hold, Korean document retrieval can be deployed as a single 2B embedding model, avoiding the storage and latency costs of multi-vector indexes.
  • The bilingual contrastive stage protects the backbone's existing multilingual ability, while adding an extra Korean text dataset hurts, suggesting vision-language alignment is the binding constraint.
  • Per-query min–max normalization turns reranker distillation from harmful to beneficial, giving a recipe for other compact retrievers.
  • Matryoshka training lets collections use 256-dimensional vectors with little loss, cutting embedding storage to one-eighth of the full dimension while beating the 8B baseline at its full dimension.
  • No single model dominates every document domain; KoVRE offers the most balanced performance across the evaluated Korean benchmark domains.

Reading between the lines

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

  • The recipe likely generalizes to other under-resourced languages: the same bilingual-rehearsal and distillation steps could be applied to any language with a small visual document retrieval corpus.
  • If the KoViDoRe evaluation split shares pages with the training set (the paper does not state they are disjoint), the reported gains could be partly memorization; a clean split test would settle this.
  • The finding that extra Korean text data hurt suggests that mixing text-only triplets can distract a visual retriever from page-level cues; this is testable on other benchmarks.
  • The 256-dim result hints that large-scale Korean VDR deployments could cut storage costs by 87.5% with a minor quality trade-off, which is worth benchmarking against multi-vector systems.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper introduces KoVRE/KOVRE, a 2B-parameter single-vector retriever for Korean visual document retrieval, initialized from Qwen3-VL-Embedding-2B and trained in two stages: (1) contrastive learning over 708,729 Korean and English query–page pairs with mined hard negatives, self-guide filtering, and hardness weighting, and (2) reranker-based knowledge distillation on Korean data using a Qwen3-VL-Reranker-8B teacher. The authors report nDCG@10 on KoViDoRe and SDS KoPub VDR, claiming that the final model substantially improves over the 2B backbone and outperforms both the 8B single-vector model and a strong multi-vector baseline, while also providing ablations on training-data composition, negative treatment, score normalization in distillation, and Matryoshka embedding dimensions.

Significance. If the results hold, this is a useful empirical contribution: a compact 2B model can deliver competitive Korean VDR with a single vector, avoiding the storage overhead of multi-vector representations. The paper also provides systematic ablations of several training techniques (hard-negative filtering, hardness weighting, min–max normalized distillation, MRL) that are valuable for practitioners. The strengths are the breadth of controlled experiments and the use of public benchmarks. However, the two most load-bearing claims — generalization beyond the training corpus and outperformance of the multi-vector baseline — are not rigorously supported by the evidence as reported: the train/evaluation disjointness is not documented, and the multi-vector margin is tiny with no variance estimates.

major comments (3)
  1. [§3.2, §4.1] The Korean Stage-1 training data includes 'ko-vdr-train-public released with KoViDoRe' (Table 1), and the evaluation is on the KoViDoRe benchmark (§4.1). Nowhere do the authors state that the evaluation queries/pages are disjoint from the training pages. If the benchmark does not have an explicit train/test split, or if overlapping pages were not removed, the reported gains (e.g., 0.5919 vs 0.3591 over the base backbone) could reflect memorization rather than generalization. This is load-bearing for the abstract and conclusion claims. Please document the exact split/dedup procedure and, if overlap exists, re-report numbers on a strictly disjoint subset. The Limitations section acknowledges the limited Korean page diversity but does not address this potential overlap.
  2. [§5.1, Table 2] The headline claim that KOVRE 'outperforms' the multi-vector jina-embeddings-v4 rests on OVR nDCG@10 of 0.5919 vs 0.5885, a difference of 0.0034. No standard deviations, number of runs, or significance tests are reported for any row of Table 2, so under standard assumptions this margin is indistinguishable from noise. Moreover, the macro-average hides sizable losses on specific domains: KOVRE trails jina-embeddings-v4 (multi-vector) on Cybersecurity (0.7627 vs 0.7714), Energy (0.6576 vs 0.6752), and especially SDS KoPub VDR (0.7324 vs 0.7802, a 4.78-point deficit). A reweighting of domains or a different seed could flip the aggregate ordering. The claims against the 2B backbone (0.3591) and the 8B single-vector model (0.5489) are robust, but the multi-vector outperformance claim is not supported by the evidence as presented.
  3. [§5.3, §5.4] The recipe choices that lead to the final configuration (α=2.0, self-guide threshold -0.1, and min–max normalization in Stage 2) are based on single-run comparisons. For example, adding self-guide improves OVR by only 0.0031 (0.5761 → 0.5792, Table 3), and Figure 2 contrasts individual checkpoints. These deltas are as small as the headline multi-vector margin, and without repeated runs or error bars, the conclusions that these components are 'critical' or 'support using' are not statistically grounded, even though the ablations are internally consistent.
minor comments (6)
  1. [Title and throughout] The model name is inconsistently styled: 'KoVRE' in the title and Abstract, but 'KOVRE' in most of the body and table captions. Please standardize.
  2. [Table 1] Formatting issue: 'After hard-negative selection416,304' and 'After positive refinement406,945' are missing spaces. The same problem appears in the caption text for the sub-total row.
  3. [Eq. (1)–(3)] The self-guide condition is written as sg(s(q_i,d)) > sg(s(q_i,d+)) + 0.1, where sg denotes stop-gradient. Since stop-gradient normally applies to tensors/parameters rather than scores, please clarify the intended semantics (e.g., stop-gradient on the embeddings before computing the score, or a detached score).
  4. [§5.2] The text mentions 'Qwen3-Embedding-8B' for selecting hard negatives in the ko-triplet experiment; this should be 'Qwen3-VL-Embedding-8B' for consistency with the rest of the paper.
  5. [Figure 2] The y-axis label 'Avg nDCG@10' is ambiguous: is this the OVR macro-average over both benchmarks, or an average over KoViDoRe domains only? Please state this explicitly in the caption.
  6. [References] Several references are dated 2026 and may be preprints or project theses in press; please verify that all citations have stable identifiers or published versions, and update the reference list accordingly.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the headline result is an external benchmark measurement; self-citations are present but not load-bearing.

full rationale

The central claim is an empirical comparison on KoViDoRe and SDS KoPub VDR. The model is trained with a contrastive objective and KL distillation, then evaluated with MTEB; no training objective is defined in terms of the reported nDCG@10, and no parameter is fitted to the evaluation scores and then reported as a prediction. The teacher Qwen3-VL-Reranker-8B belongs to the same model family as the Qwen3-VL-Embedding baselines, but the student is a different model and the comparison is not forced by construction. The KoViDoRe benchmark and ko-triplet dataset are self-citations, and the Korean training split uses ko-vdr-train-public released with KoViDoRe; however, the paper also evaluates on the externally authored SDS KoPub VDR benchmark, and the self-cited benchmark functions as an evaluation instrument rather than as a proof step. One reporting gap is that the paper never states that the KoViDoRe evaluation split is disjoint from ko-vdr-train-public pages; if they overlapped, the numbers would be inflated. That would be data leakage or overfitting, not a definitional circularity. Likewise, the paper's ablations are used to select hyperparameters on the same benchmarks, which is a test-set-tuning concern rather than a circular-derivation concern. The 0.0034 aggregate margin over the multi-vector baseline lacks significance testing, but absence of error bars is a statistical robustness issue, not circularity. Overall, no load-bearing step reduces to its input by construction.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new theoretical entities. The central claim rests on a handful of hand-set training hyperparameters (alpha, self-guide threshold, filtering thresholds, temperature) and on assumptions about backbone transferability, teacher reliability, and data quality. These are normal for an empirical ML systems paper, but the free-parameter count is nontrivial and several are set via ablations or manual inspection.

free parameters (6)
  • Hardness weight alpha = 2.0
    Set after an ablation comparing alpha=2.0 and alpha=5.0 (Table 3); it controls the stop-gradient hardness term applied to mined negatives.
  • Self-guide threshold = -0.1
    Chosen as the default cutoff for excluding negatives whose similarity exceeds the positive by 0.1, after an ablation with no self-guide filtering (Table 3).
  • Positive-score filter threshold = 0.3
    Queries are retained only if at least one positive scores above 0.3, a threshold set 'based on separate inspection of the Korean and English score distributions and manually reviewed samples' (Section 3.2).
  • False-negative exclusion ratio = 95% of positive similarity
    Mined candidates whose similarity exceeds 95% of the annotated positive similarity are excluded to reduce false negatives; adopted from NV-Retriever but still a free choice affecting the training set.
  • Number of mined hard negatives per query = 7
    Kept constant across training; pairs with fewer than seven candidates are discarded (Section 3.2).
  • Temperature tau = not reported
    The contrastive temperature tau appears in Eq. (1) and in the hardness-weighted logit, but its value is never specified; presumably inherited from the Qwen3 recipe.
assumptions (4)
  • domain assumption Qwen3-VL-Embedding-2B provides a suitable backbone whose visual-text alignment transfers to Korean after adaptation.
    Invoked throughout Section 3.1; if the backbone cannot represent Korean visual documents, the training recipe cannot succeed.
  • domain assumption Qwen3-VL-Reranker-8B scores are reliable teacher relevance labels for distillation.
    The whole Stage 2 (Section 3.4) treats these scores as ground-truth-like training targets; incorrect teacher scores would train the student toward wrong rankings.
  • domain assumption The source datasets (VisRAG, ColPali, TAT-QA, vdr-multilingual-train, etc.) contain correct positive query-page annotations.
    Training data quality is taken for granted in Section 3.2; noisy positives would undermine the contrastive objective.
  • standard math The MTEB framework computes nDCG@10 faithfully for these benchmarks.
    Evaluation in Section 4.1 relies on the MTEB implementation; no manual verification is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of KoVRE: Training an Efficient Embedding Model for Korean Visual Document Retrieval." pith.science (2026). https://pith.science/paper/QQAPIQD7

@misc{pith2026260801389,
  author       = {Pith},
  title        = {Pith review of: KoVRE: Training an Efficient Embedding Model for Korean Visual Document Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QQAPIQD7}},
  note         = {Machine review of arXiv:2608.01389}
}
read the original abstract

Visual Document Retrieval (VDR) directly matches text queries against document images, preserving visual and structural information that may be lost during text extraction. However, existing VDR models and training resources remain predominantly English-centric, while many high-performing systems rely on massive backbones or storage-intensive multi-vector representations. To address these limitations, we introduce KoVRE: Korean Visual Document Retrieval Embedding, a single-vector retriever for Korean visual documents, alongside a comprehensive training recipe. We train the model on 708,729 Korean and English query-page pairs using positive-aware hard-negative mining and conduct controlled analyses of training-data composition, hard-negative treatment, and reranker-based knowledge distillation. Across Korean visual document retrieval benchmarks, our 2B model substantially improves over the base backbone model, outperforming both its 8B single-vector counterpart and a strong multi-vector baseline. These results demonstrate that targeted bilingual supervision and our carefully designed training strategies can produce a highly effective Korean VDR model across diverse document domains, without requiring a scaled-up backbone or multi-vector representations.

Figures

Figures reproduced from arXiv: 2608.01389 by the authors.

Figure 1
Figure 1. Effect of training-data composition on Korean [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Effect of per-row min–max score normal￾ization on Stage 2 distillation, measured by average nDCG@10. 5.4 Score Normalization for Knowledge Distillation During Stage 2, we distill the reranker teacher into KOVRE by matching, for each query, the student’s distribution over the positive and eight negatives to that of the teacher. Because the reranker assigns scores on an unbounded, query-dependent scale, ap￾plying a so… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 14 canonical work pages

  1. [1]

    2020 , eprint=

    Approximate Nearest Neighbor Negative Contrastive Learning for Dense Text Retrieval , author=. 2020 , eprint=

  2. [2]

    2021 , eprint=

    RocketQA: An Optimized Training Approach to Dense Passage Retrieval for Open-Domain Question Answering , author=. 2021 , eprint=

  3. [3]

    arXiv preprint arXiv:2010.02666 , year=

    Improving efficient neural ranking models with cross-architecture knowledge distillation , author=. arXiv preprint arXiv:2010.02666 , year=

  4. [4]

    2023 , eprint=

    RocketQAv2: A Joint Training Method for Dense Passage Retrieval and Passage Re-ranking , author=. 2023 , eprint=

  5. [5]

    2022 , eprint=

    ColBERTv2: Effective and Efficient Retrieval via Lightweight Late Interaction , author=. 2022 , eprint=

  6. [6]

    2024 , eprint=

    GISTEmbed: Guided In-sample Selection of Training Negatives for Text Embedding Fine-tuning , author=. 2024 , eprint=

  7. [7]

    Advances in neural information processing systems , volume=

    Retrieval-augmented generation for knowledge-intensive nlp tasks , author=. Advances in neural information processing systems , volume=

  8. [8]

    International Conference on Learning Representations , volume=

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

Show all 46 references
  1. [9]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Vdocrag: Retrieval-augmented generation over visually-rich documents , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  2. [10]

    V i D o RAG : Visual Document Retrieval-Augmented Generation via Dynamic Iterative Reasoning Agents

    Wang, Qiuchen and Ding, Ruixue and Chen, Zehui and Wu, Weiqi and Wang, Shihang and Xie, Pengjun and Zhao, Feng. V i D o RAG : Visual Document Retrieval-Augmented Generation via Dynamic Iterative Reasoning Agents. Proceedings of the 2025 Conference on Empirical Methods in Natur...

  3. [11]

    arXiv preprint arXiv:2407.01449 , year=

    Colpali: Efficient document retrieval with vision language models , author=. arXiv preprint arXiv:2407.01449 , year=

  4. [12]

    Unifying Multimodal Retrieval via Document Screenshot Embedding

    Ma, Xueguang and Lin, Sheng-Chieh and Li, Minghan and Chen, Wenhu and Lin, Jimmy. Unifying Multimodal Retrieval via Document Screenshot Embedding. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024. doi:10.18653/v1/2024.emnlp-main.373

  5. [13]

    arXiv preprint arXiv:2505.17166 , year=

    Vidore benchmark v2: Raising the bar for visual retrieval , author=. arXiv preprint arXiv:2505.17166 , year=

  6. [14]

    and Liu, Bo and Faysse, Manuel and Hudelot, Celine and Viaud, Gautier

    Loison, Ant \'o nio and Mac \'e , Quentin and Edy, Antoine and Xing, Victor and Balough, Tom and Moreira, Gabriel de Souza P. and Liu, Bo and Faysse, Manuel and Hudelot, Celine and Viaud, Gautier. V i D o R e V3: A Comprehensive Evaluation of Retrieval Augmented Generation in ...

  7. [15]

    arXiv preprint arXiv:2510.03663 , year=

    Unidoc-bench: A unified benchmark for document-centric multimodal rag , author=. arXiv preprint arXiv:2510.03663 , year=

  8. [16]

    arXiv preprint arXiv:2602.17687 , year=

    Irpapers: A visual document benchmark for scientific retrieval and question answering , author=. arXiv preprint arXiv:2602.17687 , year=

  9. [17]

    REAL - MM - RAG : A Real-World Multi-Modal Retrieval Benchmark

    Wasserman, Navve and Pony, Roi and Naparstek, Oshri and Goldfarb, Adi Raz and Schwartz, Eli and Barzelay, Udi and Karlinsky, Leonid. REAL - MM - RAG : A Real-World Multi-Modal Retrieval Benchmark. Proceedings of the 63rd Annual Meeting of the Association for Computational Ling...

  10. [18]

    2025 , publisher=

    Nomic Embed Multimodal: Interleaved Text, Image, and Screenshots for Visual Document Retrieval , author=. 2025 , publisher=

  11. [19]

    arXiv preprint arXiv:2507.05513 , year=

    Llama nemoretriever colembed: Top-performing text-image retrieval model , author=. arXiv preprint arXiv:2507.05513 , year=

  12. [20]

    arXiv preprint arXiv:2602.03992 , year=

    Nemotron ColEmbed V2: Top-Performing Late Interaction embedding models for Visual Document Retrieval , author=. arXiv preprint arXiv:2602.03992 , year=

  13. [21]

    jina-embeddings-v4: Universal Embeddings for Multimodal Multilingual Retrieval

    G. jina-embeddings-v4: Universal Embeddings for Multimodal Multilingual Retrieval. Proceedings of the 5th Workshop on Multilingual Representation Learning (MRL 2025). 2025. doi:10.18653/v1/2025.mrl-main.36

  14. [22]

    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=

  15. [23]

    2025 , url =

    Huang, Xin and Tan, Kye Min , title =. 2025 , url =

  16. [24]

    arXiv preprint arXiv:2509.18095 , year=

    Metaembed: Scaling multimodal retrieval at test-time with flexible late interaction , author=. arXiv preprint arXiv:2509.18095 , year=

  17. [25]

    arXiv preprint arXiv:2505.11651 , year=

    Miracl-vision: A large, multilingual, visual document retrieval benchmark , author=. arXiv preprint arXiv:2505.11651 , year=

  18. [26]

    MIRACL : A Multilingual Retrieval Dataset Covering 18 Diverse Languages

    Zhang, Xinyu and Thakur, Nandan and Ogundepo, Odunayo and Kamalloo, Ehsan and Alfonso-Hermelo, David and Li, Xiaoguang and Liu, Qun and Rezagholizadeh, Mehdi and Lin, Jimmy. MIRACL : A Multilingual Retrieval Dataset Covering 18 Diverse Languages. Transactions of the Associatio...

  19. [27]

    arXiv preprint arXiv:2511.04910 , year=

    SDS KoPub VDR: A Benchmark Dataset for Visual Document Retrieval in Korean Public Documents , author=. arXiv preprint arXiv:2511.04910 , year=

  20. [28]

    K o V i D o R e: A Benchmark for K orean Visual Document Retrieval

    Choi, Yongbin and Song, Yongwoo and Sung, Mujeen. K o V i D o R e: A Benchmark for K orean Visual Document Retrieval. Proceedings of the 2nd Workshop on Multimodal Augmented Generation via Multimodal Retrieval ( MAGM a R 2026). 2026. doi:10.18653/v1/2026.magmar-main.11

  21. [29]

    TAT - QA : A Question Answering Benchmark on a Hybrid of Tabular and Textual Content in Finance

    Zhu, Fengbin and Lei, Wenqiang and Huang, Youcheng and Wang, Chao and Zhang, Shuo and Lv, Jiancheng and Feng, Fuli and Chua, Tat-Seng. TAT - QA : A Question Answering Benchmark on a Hybrid of Tabular and Textual Content in Finance. Proceedings of the 59th Annual Meeting of the...

  22. [30]

    2025 , publisher =

    LlamaIndex , title =. 2025 , publisher =

  23. [31]

    arXiv preprint arXiv:2407.15831 , year=

    NV-Retriever: Improving text embedding models with effective hard-negative mining , author=. arXiv preprint arXiv:2407.15831 , year=

  24. [32]

    Advances in Neural Information Processing Systems , volume=

    Matryoshka representation learning , author=. Advances in Neural Information Processing Systems , volume=

  25. [33]

    MTEB : Massive Text Embedding Benchmark

    Muennighoff, Niklas and Tazi, Nouamane and Magne, Loic and Reimers, Nils. MTEB : Massive Text Embedding Benchmark. Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics. 2023. doi:10.18653/v1/2023.eacl-main.148

  26. [34]

    arXiv preprint arXiv:2503.04812 , year=

    Llave: Large language and vision embedding models with hardness-weighted contrastive learning , author=. arXiv preprint arXiv:2503.04812 , year=

  27. [35]

    arXiv preprint arXiv:2509.20354 , year=

    Embeddinggemma: Powerful and lightweight text representations , author=. arXiv preprint arXiv:2509.20354 , year=

  28. [36]

    Proceedings of the 6th Workshop on Representation Learning for NLP (RepL4NLP-2021) , pages=

    Scaling deep contrastive learning batch size under memory limited setup , author=. Proceedings of the 6th Workshop on Representation Learning for NLP (RepL4NLP-2021) , pages=

  29. [37]

    arXiv preprint arXiv:2502.13595 , year=

    Mmteb: Massive multilingual text embedding benchmark , author=. arXiv preprint arXiv:2502.13595 , year=

  30. [38]

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

    Ma, Yubo and Li, Jinsong and Zang, Yuhang and Wu, Xiaobao and Dong, Xiaoyi and Zhang, Pan and Cao, Yuhang and Duan, Haodong and Wang, Jiaqi and Cao, Yixin and Sun, Aixin. Towards Storage-Efficient Visual Document Retrieval: An Empirical Study on Reducing Patch-Level Embeddings...

  31. [39]

    MMD oc IR : Benchmarking Multimodal Retrieval for Long Documents

    Dong, Kuicai and Chang, Yujing and Goh Xin Deik, Derrick and Li, Dexun and Tang, Ruiming and Liu, Yong. MMD oc IR : Benchmarking Multimodal Retrieval for Long Documents. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025. doi:10.18653/...

  32. [40]

    arXiv preprint arXiv:2411.04952 , year=

    M3docrag: Multi-modal retrieval is what you need for multi-page multi-document understanding , author=. arXiv preprint arXiv:2411.04952 , year=

  33. [41]

    arXiv preprint arXiv:2510.01149 , year=

    ModernVBERT: Towards Smaller Visual Document Retrievers , author=. arXiv preprint arXiv:2510.01149 , year=

  34. [42]

    arXiv preprint arXiv:2412.08802 , year=

    jina-clip-v2: Multilingual multimodal embeddings for text and images , author=. arXiv preprint arXiv:2412.08802 , year=

  35. [43]

    arXiv preprint arXiv:2605.08384 , year=

    jina-embeddings-v5-omni: Geometry-preserving Embeddings via Locked Aligned Towers , author=. arXiv preprint arXiv:2605.08384 , year=

  36. [44]

    Annual Conference on Human and Language Technology , pages=

    KoE5: A New Dataset and Model for Improving Korean Embedding Performance , author=. Annual Conference on Human and Language Technology , pages=. 2024 , organization=

  37. [45]

    Leitner-Guided Memory Replay for Cross-lingual Continual Learning

    M ' hamdi, Meryem and May, Jonathan. Leitner-Guided Memory Replay for Cross-lingual Continual Learning. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024....

  38. [46]

    Mitigating Catastrophic Forgetting in Large Language Models with Self-Synthesized Rehearsal

    Huang, Jianheng and Cui, Leyang and Wang, Ante and Yang, Chengyi and Liao, Xinting and Song, Linfeng and Yao, Junfeng and Su, Jinsong. Mitigating Catastrophic Forgetting in Large Language Models with Self-Synthesized Rehearsal. Proceedings of the 62nd Annual Meeting of the Ass...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.