Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

This paper claims that retrieval-augmented generation—retrieving relevant chunks or image patches before answering—consistently beats concatenating the whole document, by up to +22.5 ANLS, and +5.0 without OCR.

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 ↗

Retrieval-augmented generation improves multi-page document VQA accuracy for small and medium models, with text-based retrieval up to +22.5 ANLS and visual retrieval up to +5.0 ANLS.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Useful empirical study, but the headline +22.5 gain rests on an unverified baseline—Qwen2.5-VL-baseline is never described, and RAG-Qwen gets LoRA fine-tuning that may not be matched. the 3 major comments →

arxiv 2508.18984 v2 pith:6MSFQPEM submitted 2025-08-26 cs.CV

Enhancing Document VQA Models via Retrieval-Augmented Generation

Eric L\'opez , Artemis Llabr\'es , Ernest Valveny This is my paper
classification cs.CV
keywords Document Visual Question AnsweringRetrieval-Augmented Generationmulti-page documentstext retrievalvisual retrievallate interactionOCR-freeANLS
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 reading

The paper sets out to establish that careful evidence selection, not longer context or larger models, is the practical route to accurate multi-page Document VQA. It wraps three existing generators—VT5, Qwen2.5-VL-7B, and Pix2Struct—in a retrieval-augmented pipeline: chunk the document, retrieve candidate chunks by embedding similarity, rerank with a cross-encoder, then generate from only the top-k chunks. On MP-DocVQA, DUDE, and InfographicVQA, the text-based RAG variants beat the concatenate-all-pages baseline by up to +22.5 ANLS, and the fully visual OCR-free variant gains +5.0 ANLS. An ablation shows the reranker and contrastive fine-tuning of the embedder carry the gain, while layout-guided chunking hurts. If the claim holds, retrieval-augmented generation is a memory-cheap way to bring multi-page document question answering to small and medium models.

Core claim

On the paper's own terms, the central discovery is that retrieval-augmented evidence selection consistently improves multi-page Document VQA across model sizes and benchmarks, and that the mechanism is retrieval plus reranking rather than smarter chunking. The text-centric variant improves the concatenate-all-pages baseline by up to +22.5 ANLS on MP-DocVQA (RAG-Qwen, 51.2 to 73.7), while the visual variant improves RAG-Pix2Struct by +5.0 ANLS without requiring OCR. The same pattern holds on DUDE and InfographicVQA, with one outlier: visual RAG drops 0.8 ANLS on DUDE, attributed to its higher reasoning complexity. The ablation attributes most of the gain to the cross-encoder reranker and cont

What carries the argument

The pipeline is the machinery. Offline, the document is split into small OCR-token chunks (size 60, overlap 10) or overlapping visual patches (512px, 50% overlap); each chunk is embedded by a bi-encoder (bge-en-small-v1.5) for text, or by the Pix2Struct vision encoder for images. At query time, the question is embedded the same way, the top k′=20 chunks are retrieved by cosine similarity (or by ColBERT-style late interaction for images), and a cross-encoder reranker (bge-reranker-v2-m3) narrows this to the top k=10 (or k=5 visual) chunks. The generator consumes only these chunks, with their bounding boxes and image crops for VT5, or as a prompt plus crops for Qwen, or as tiled image patches

Load-bearing premise

The head-to-head numbers assume the baseline and RAG models get the same task-specific fine-tuning; the paper does not state whether the Qwen2.5-VL baseline was LoRA-tuned like RAG-Qwen, so part of the +22.5 ANLS gap could be training rather than retrieval.

What would settle it

Give the Qwen2.5-VL baseline the same LoRA fine-tuning on the same training split and prompt, then re-run MP-DocVQA: if the RAG-vs-baseline gap collapses, the retrieval benefit is mostly an artifact of unequal training. Also verify that the contrastive positive chunks for embedder training are drawn only from training queries.

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

If this is right

  • Small and medium Document VQA models can reach multi-page accuracies close to or above published state of the art by adding a retrieval stack instead of scaling up context.
  • Text-based RAG can be bolted onto an existing OCR pipeline with a +601M-parameter retrieval stack, making the approach compatible with current OCR-heavy systems.
  • OCR-free visual RAG works without text extraction, gaining +5.0 ANLS on MP-DocVQA and +3.3 on InfographicVQA while cutting GPU cost relative to self-attention retrieval.
  • Retrieval and reranking, not chunking strategy, carry the accuracy gain; layout-guided chunking over-segments pages and can hurt.
  • The one reported drop (-0.8 ANLS for RAG-Pix2Struct on DUDE) indicates reasoning-heavy datasets may resist retrieval-based pruning, so gains are not universal.

Where Pith is reading between the lines

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

  • The evidence index is the reusable asset: since chunks are embedded offline, the same retriever and reranker can serve any future generator, so the marginal cost of upgrading the answer model is small.
  • A text-plus-image hybrid scorer could fix failures like the paper's Example 2, where text retrieval found the right table but omitted the column headers needed to disambiguate the answer.
  • The newly introduced Chunk Score@k metric—did the answer appear in the retrieved text—could double as a training reward for the retriever, since it measures evidence utility rather than page hit rate.
Share X Bluesky LinkedIn Reddit HN

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 / 5 minor

Summary. The paper proposes retrieval-augmented generation (RAG) for multi-page Document VQA. It introduces a textual RAG pipeline for VT5 and Qwen2.5-VL-7B, using OCR-token chunking with learned embeddings and cross-encoder reranking, and a visual RAG pipeline for Pix2Struct using patch-level late-interaction retrieval without OCR. Experiments on MP-DocVQA, DUDE, and InfographicVQA report consistent ANLS gains over concatenate-all-pages or per-page MaxConf baselines, with the largest headline gain +22.5 ANLS for RAG-Qwen on MP-DocVQA. An ablation on RAG-VT5 attributes most gains to reranking and embedder fine-tuning, while layout-guided chunking is reported not to help. The paper also reports retrieval-precision and a chunk-level score.

Significance. If the empirical claims hold, the paper provides a practical and cheaper alternative to full-context multi-page processing: evidence selection improves accuracy for small and medium document-VQA models on several benchmarks, with a fully OCR-free variant as a bonus. The study is broad: two textual generators and one visual generator, three benchmarks, retrieval-quality metrics, and an ablation isolating retrieval components. The paper also honestly reports a negative result for layout-guided chunking and an outlier failure on DUDE for the visual variant. These are useful and credible empirical contributions. The main caveat is that the comparison between RAG variants and baselines is not clearly controlled for equal task-specific training, which is essential to attribute the measured gains to retrieval rather than fine-tuning.

major comments (3)
  1. [§4.2, §4.3, Table 2] The central claim that retrieval itself yields large gains is not yet supported because the baselines' training condition is unspecified. Section 4.2 describes the multi-page extension for VT5 ('concatenate') and Pix2Struct ('MaxConf') but gives no description of the Qwen2.5-VL baseline. Section 4.3 states that RAG-VT5 is fully fine-tuned and RAG-Qwen is LoRA-fine-tuned on each target dataset, but does not state whether the VT5/Qwen baselines receive the same task-specific fine-tuning. If the Qwen2.5-VL baseline is used zero-shot or in instruction mode only, the +22.5 ANLS difference in Table 2 conflates LoRA adaptation with retrieval. The same issue applies, though less severely, to RAG-VT5 versus VT5-baseline. Please specify the baseline training protocol explicitly and, ideally, add a baseline with identical LoRA/full fine-tuning to the RAG generator.
  2. [§4.3] The contrastive fine-tuning of the bi-encoder is described as creating positives by processing 'each document VQA dataset' through the RAG pipeline, selecting chunks whose generated answer has ANLS > 0.8 with ground truth. The manuscript does not state whether this procedure is restricted to training queries or whether it also uses validation/test queries. If evaluation queries are included in embedder fine-tuning, the reported retrieval metrics (Retrieval Precision@k, Chunk Score@k) and the downstream ANLS gains become optimistic, and the ablation in Table 3 would be affected. Please clarify the split and restrict positive mining to the training split.
  3. [§5.1 / Table 2, §6] The ablation study in Table 3 is performed only for RAG-VT5, but the headline cross-model generalization claim rests on Qwen2.5-VL and Pix2Struct. The reader cannot tell whether the +22.5 and +5.0 gains are driven by the same retrieval components or by different interaction effects in those models. Adding at least an ablation of reranking/embedding fine-tuning for RAG-Qwen, or a clear argument that the RAG-VT5 ablation transfers, would strengthen the manuscript. Without this, the claim 'careful evidence selection consistently boosts accuracy across multiple model sizes' is only partially verified.
minor comments (5)
  1. [§4.2 / Table 1] Hyperparameters are reported as 'the ones that yield better results', but the selection method, search range, and validation split used for tuning are not described. This is a reproducibility concern: adding a sentence on the tuning protocol would be helpful.
  2. [§4.1, Eq. (3)] Retrieval Precision@k is page-level, not chunk-level: it counts a positive if any retrieved chunk is on the ground-truth page. Since the baseline chunking can produce many chunks from the same page, this metric may overestimate the retrieval of the actual answer-bearing content. Consider reporting a stricter span-level or chunk-answer overlap metric as well.
  3. [§3.2] The visual retrieval renders the query as an image, but no details are given about font, size, width/height, or rendering resolution. These details affect reproducibility of the late-interaction retrieval, especially for typed queries with numbers or mathematical symbols.
  4. [§5.1] The conclusion claims computational and memory efficiency, but the paper does not report wall-clock time, GPU memory, or latency for any method. Since the 'RAG total' includes 601M additional retrieval parameters (Table 2 footnote), an explicit runtime/memory comparison with concatenate-all-pages would make the efficiency claim concrete.
  5. [Table 3] The row headers in Table 3 are abbreviated ('Rerank', 'Train Emb.', 'Layout', 'Cluster', 'Train Layout'). It would be clearer to spell them out and to state in the caption which configuration corresponds to the default RAG-VT5 pipeline.

Circularity Check

0 steps flagged

No circularity: the empirical RAG comparisons are evaluated against external benchmarks; the fairness/split concerns are experimental confounds, not construction-level circularity.

full rationale

The paper is an empirical study, not a formal derivation. The claimed improvements (RAG-VT5 +13.2/+7.3/+10.6 ANLS, RAG-Qwen +22.5/+4.0/+2.0, RAG-Pix2Struct +5.0/+3.3/-0.8) are measured by standard accuracy and ANLS against ground-truth answers on the external MP-DocVQA, DUDE, and InfographicVQA benchmarks. No equation in the paper defines a predicted quantity in terms of the input that is then reported as a discovery; the retrieval metrics (Retrieval Precision@k, Chunk Score@k) are explicitly defined as evaluation metrics, not as fitted predictions. The bi-encoder positive-pair construction in Section 4.3 uses ground-truth answers to label chunks as positives (ANLS > 0.8), which is standard supervised training when confined to the training split; the paper does not state that the evaluation queries are excluded, so this is a data-contamination risk, but it is not a circularity of reasoning because the final benchmark numbers do not reduce to those labels by construction. Similarly, Section 4.3 describes LoRA fine-tuning for RAG-Qwen without stating whether the Qwen baseline receives equivalent fine-tuning, so the +22.5 ANLS gain may conflate fine-tuning with retrieval; this is an experimental-fairness confound, not a case where a fitted parameter is renamed a prediction. Self-citations to Hi-VT5, DUDE, InfographicVQA, and Pix2Struct-SA-retrieval are used as prior baselines or dataset sources, not as an unverified uniqueness theorem or as the sole justification for a central claim. The ablation in Table 3 is an empirical decomposition whose outcome (reranking and embedder fine-tuning help; layout chunking does not) is contingent and could have been otherwise. Therefore, no load-bearing step in the paper reduces by definition to its own inputs, and the appropriate circularity score is 0.

Axiom & Free-Parameter Ledger

11 free parameters · 6 axioms · 0 invented entities

The method uses existing model components and off-the-shelf retrievers; no new entities are introduced. The free parameters are hyperparameters tuned by experimentation, and the axioms are standard assumptions about metric validity, embedding relevance, and dataset suitability.

free parameters (11)
  • chunk_size_L = 60 tokens
    Sliding window length for text chunks, selected by experimentation to yield better ANLS.
  • chunk_overlap_O = 10 tokens
    Shared tokens between neighboring chunks, tuned manually.
  • chunk_tolerance_tau = 0.2
    Allows last chunk on a page to expand by 20%, avoiding small isolated remainders.
  • initial_candidates_k_prime = 20
    Number of chunks retrieved by bi-encoder before reranking.
  • final_chunks_k = 10
    Number of chunks kept after reranker and fed to generator.
  • visual_patch_size_P = 512 pixels
    Vertical patch height for visual RAG.
  • visual_patch_overlap_Opix = 256 pixels
    Half-patch overlap to preserve information near boundaries.
  • visual_retrieved_k = 5
    Number of visual patches selected by late interaction.
  • ANLS_positive_threshold_t = 0.8
    Threshold for a query-chunk pair to be a positive training example for the bi-encoder.
  • LoRA_alpha_rank_dropout = alpha=16, rank=8, dropout=0.05
    PEFT hyperparameters for Qwen2.5-VL fine-tuning.
  • VT5_learning_rate = 2e-4
    Initial learning rate for VT5 fine-tuning.
axioms (6)
  • domain assumption The ANLS metric is a valid proxy for answer correctness in Document VQA.
    Used as primary evaluation metric across all benchmarks; if ANLS mis-ranks answers, the headline gains lose meaning.
  • domain assumption Cosine similarity between bi-encoder embeddings reflects relevance of a chunk to a question.
    The retrieval pipeline assumes dense embeddings capture semantic relevance, an empirical premise validated only indirectly by downstream QA accuracy.
  • domain assumption Ground-truth answer page and answer text can serve as labels for retrieval quality and retriever training.
    Retrieval Precision@k and Chunk Score@k use the ground-truth answer page; the training positives are selected by comparing generated answers to ground truth.
  • domain assumption The publicly available fine-tuned Pix2Struct checkpoints are appropriate generators for the visual RAG evaluation.
    The paper relies on existing dataset-specific fine-tuned Pix2Struct versions without further training.
  • domain assumption OCR token sequences are available for all evaluated text-based datasets.
    Text RAG operates on OCR tokens from MP-DocVQA, DUDE, and InfographicVQA; if OCR quality varies, results may not transfer.
  • ad hoc to paper The hyperparameters reported in Table 1 generalize to all datasets and models.
    Chunk size, overlap, and k values are fixed across datasets even though optimal values likely differ; this is a modeling choice that could hide dataset-specific behavior.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Document VQA Models via Retrieval-Augmented Generation." pith.science (2026). https://pith.science/paper/6MSFQPEM

@misc{pith2026250818984,
  author       = {Pith},
  title        = {Pith review of: Enhancing Document VQA Models via Retrieval-Augmented Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6MSFQPEM}},
  note         = {Machine review of arXiv:2508.18984}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Document Visual Question Answering (Document VQA) must cope with documents that span dozens of pages, yet leading systems still concatenate every page or rely on very large vision-language models, both of which are memory-hungry. Retrieval-Augmented Generation (RAG) offers an attractive alternative, first retrieving a concise set of relevant segments before generating answers from this selected evidence. In this paper, we systematically evaluate the impact of incorporating RAG into Document VQA through different retrieval variants - text-based retrieval using OCR tokens and purely visual retrieval without OCR - across multiple models and benchmarks. Evaluated on the multi-page datasets MP-DocVQA, DUDE, and InfographicVQA, the text-centric variant improves the "concatenate-all-pages" baseline by up to +22.5 ANLS, while the visual variant achieves +5.0 ANLS improvement without requiring any text extraction. An ablation confirms that retrieval and reranking components drive most of the gain, whereas the layout-guided chunking strategy - proposed in several recent works to leverage page structure - fails to help on these datasets. Our experiments demonstrate that careful evidence selection consistently boosts accuracy across multiple model sizes and multi-page benchmarks, underscoring its practical value for real-world Document VQA.

Figures

Figures reproduced from arXiv: 2508.18984 by Artemis LLabr\'es, Eric L\'opez, Ernest Valveny.

Figure 1
Figure 1. Figure 1: Overview of the Textual RAG pipeline. Offline (pink panel): A multi-page document is segmented into chunks consisting of OCR (text and boxes) and the image crop of the chunk. A Bi-Encoder converts the chunk text into a dense embedding for later retrieval. Online (green panel): the user question is encoded in the same way, and cosine-similarity is used to select the top-k ′ chunks, whose text is then passed… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the Visual RAG pipeline. Offline (pink panel): A multi-page document is segmented into image patches, each of which is passed to a Visual Encoder to produce multi-vector embeddings for later retrieval. Online (green panel): the user question is rendered as an image, encoded in the exact same way and matched to the patch embeddings through late interaction. The top-k image patches are concatenat… view at source ↗
Figure 3
Figure 3. Figure 3: Different concatenation and generation strategies for the proposed RAG meth￾ods. For RAG-VT5, the information of different modalities is separately concatenated and fed to a specialized encoder, before aggregating and passing it to the decoder. For RAG-Qwen, the text of the chunks is concatenated along with the question and an instruction prompt, and is passed together with the separate chunk image crops t… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative examples of Question Answering on long-context multi-page sam￾ples of MP-DocVQA. Textual RAG uses VT5 as generator, while Visual RAG uses Pix2Struct. Example 1 shows a correct retrieval and correct model generation, while Example 2 shows a correct retrieval but a wrong generation. † While 5 image patches are retrieved, some may be overlapping so they are merged into a bigger one [PITH_FULL_IMA… view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Scaling Beyond Context: A Survey of Multimodal Retrieval-Augmented Generation for Document Understanding

    cs.CL 2025-10 accept novelty 4.0

    A systematic survey of Multimodal RAG for document understanding proposing a taxonomy based on domain, retrieval modality, and granularity while reviewing graph structures, agentic frameworks, datasets, benchmarks, ap...

Reference graph

Works this paper leans on

29 extracted references · 17 canonical work pages · cited by 1 Pith paper

  1. [1]

    In: Proc

    Appalaraju, S., Jasani, B., Kota, B.U., Xie, Y., Manmatha, R.: Docformer: End- to-end transformer for document understanding. In: Proc. of the IEEE/CVF In- ternational Conference on Computer Vision (ICCV). pp. 993–1003 (2021)

    work page 2021

  2. [2]

    Proceedings of the AAAI ConferenceonArtificialIntelligence 38(2),709–718(Mar2024)

    Appalaraju, S., Tang, P., Dong, Q., Sankaran, N., Zhou, Y., Manmatha, R.: Doc- formerv2: Local features for document understanding. Proceedings of the AAAI ConferenceonArtificialIntelligence 38(2),709–718(Mar2024). https://doi.org/ 10.1609/aaai.v38i2.27828, https://ojs.aaai.org/index.php/AAAI/article/ view/27828

    work page doi:10.1609/aaai.v38i2.27828

  3. [3]

    CoRRabs/2502.13923 (2025), arXiv:2502.13923

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., Zhong, H., Zhu, Y., Yang, M., Li, Z., Wan, J., Wang, P., Ding, W., Fu, Z., Xu, Y., Ye, J., Zhang, X., Xie, T., Cheng, Z., Zhang, H., Yang, Z., Xu, H., Lin, J.: Qwen2.5-vl technical report. CoRRabs/2502.13923 (2025), arXiv:2502.13923

    Pith/arXiv arXiv 2025

  4. [4]

    CoRRabs/2407.07726 (2024), arXiv:2407.07726

    Beyer, L., Steiner, A., Pinto, A.S., Kolesnikov, A., Wang, X., Salz, D., Neumann, M., Alabdulmohsin, I., Tschannen, M., Bugliarello, E., Unterthiner, T., Keysers, D., Koppula, S., Liu, F., Grycner, A., Gritsenko, A., Houlsby, N., Kumar, M., Rong, K., Eisenschlos, J., Kabra, R., Bauer, M., Bošnjak, M., Chen, X., Minderer, M., Voigtlaender, P., Bica, I., Ba...

    Pith/arXiv arXiv 2024

  5. [5]

    In: Proc

    Blau, T., Fogel, S., Ronen, R., Golts, A., Tsiper, S., Ben-Avraham, E., Aber- dam, A., Bronstein, I., Litman, R., Mazor, S., Appalaraju, S., Manmatha, R.: Gram: Global reasoning for multi-page vqa. In: Proc. of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR). pp. 15598–15607 (2024). https://doi.org/10.1109/CVPR52733.2024.01477

    arXiv 2024

  6. [6]

    Chen, J., Xiao, S., Zhang, P., Luo, K., Lian, D., Liu, Z.: Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self- knowledge distillation (2024)

    work page 2024

  7. [7]

    Delestre, C.: (2024), https://huggingface.co/cmarkea/ dit-base-layout-detection

    work page 2024

  8. [8]

    In: Proc

    Devlin,J.,Chang,M.,Lee,K.,Toutanova,K.:BERT:Pre-trainingofDeepBidirec- tional Transformers for Language Understanding. In: Proc. of NAACL-HLT (Vol. 1: Long and Short Papers). pp. 4171–4186 (2019)

    work page 2019

  9. [9]

    In: Proc

    Faysse, M., Sibille, H., Wu, T., Omrani, B., Viaud, G., Hudelot, C., Colombo, P.: Colpali: Efficient document retrieval with vision language models. In: Proc. of the Int. Conf. on Learning Representations (ICLR) (2025)

    work page 2025

  10. [10]

    In: Proc

    He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770–778 (2016).https://doi.org/10.1109/CVPR.2016.90

    doi:10.1109/cvpr.2016.90 2016

  11. [11]

    CoRRabs/1705.00652 (2017), arXiv:1705.00652

    Henderson, M., Al-Rfou, R., Strope, B., Sung, Y., Lukacs, L., Guo, R., Kumar, S., Miklos, B., Kurzweil, R.: Efficient natural language response suggestion for smart reply. CoRRabs/1705.00652 (2017), arXiv:1705.00652

    Pith/arXiv arXiv 2017

  12. [12]

    In: Proc

    Huang, Y., Lv, T., Cui, L., Lu, Y., Wei, F.: Layoutlmv3: Pre-training for document ai with unified text and image masking. In: Proc. of the 30th ACM Int. Conf. on Multimedia (ACM-MM) (2022) Enhancing Document VQA Models via Retrieval-Augmented Generation 17

    work page 2022

  13. [13]

    In: Proc

    Kang, L., Tito, R., Valveny, E., Karatzas, D.: Multi-page document visual question answering using self-attention scoring mechanism. In: Proc. of the 17th Int. Conf. on Document Analysis and Recognition (ICDAR) (2024)

    work page 2024

  14. [14]

    In: Proc

    Khattab, O., Zaharia, M.: Colbert: Efficient and effective passage search via con- textualized late interaction over bert. In: Proc. of the 43rd Int. ACM SIGIR Conf. on Research and Development in Information Retrieval (SIGIR). pp. 39–48 (2020). https://doi.org/10.1145/3397271.3401075

    arXiv 2020

  15. [15]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Landeghem, J.V., Tito, R., Borchmann, Ł., Pietruszka, M., Józiak, P., Powalski, R., Jurkiewicz, D., Coustaty, M., Anckaert, B., Valveny, E., Blaschko, M., Moens, S., Stanisławek, T.: Document Understanding Dataset and Evaluation (DUDE). In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 19528–19540 (2023)

    work page 2023

  16. [16]

    In: Proc

    Lee, K., Joshi, M., Turc, I., Hu, H., Liu, F., Eisenschlos, J., Khandelwal, U., Shaw, P., Chang, M., Toutanova, K.: Pix2struct: Screenshot parsing as pretraining for visual language understanding. In: Proc. of the 40th International Conference on Machine Learning (ICML) (2023)

    work page 2023

  17. [17]

    Li,C.,Liu,Z.,Xiao,S.,Shao,Y.:Makinglargelanguagemodelsabetterfoundation for dense retrieval (2023)

    work page 2023

  18. [18]

    In: Globerson, A., Mackey, L., Belgrave, D., Fan, A., Paquet, U., Tomczak, J., Zhang, C

    Ma,Y.,Zang,Y.,Chen,L.,Chen,M.,Jiao,Y.,Li,X.,Lu,X.,Liu,Z.,Ma,Y.,Dong, X., Zhang, P., Pan, L., Jiang, Y.G., Wang, J., Cao, Y., Sun, A.: Mmlongbench- doc: Benchmarking long-context document understanding with visualizations. In: Globerson, A., Mackey, L., Belgrave, D., Fan, A., Paquet, U., Tomczak, J., Zhang, C. (eds.) Advances in Neural Information Processi...

    work page 2024

  19. [19]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)

    Mathew, M., Bagal, V., Tito, R., Karatzas, D., Valveny, E., Jawahar, C.V.: Info- graphicvqa. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1697–1706 (2022)

    work page 2022

  20. [20]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)

    Mathew, M., Karatzas, D., Jawahar, C.V.: Docvqa: A dataset for vqa on document images. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 2199–2208 (2021).https://doi.org/10.1109/ WACV48630.2021.00225

    arXiv 2021

  21. [21]

    Journal of Machine Learning Research21(140), 1–67 (2020)

    Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research21(140), 1–67 (2020)

    2020

  22. [22]

    In: Proc

    Tito, R., Karatzas, D., Valveny, E.: Document collection visual question answer- ing. In: Proc. 16th Int. Conf. on Document Analysis and Recognition (ICDAR). LNCS, vol. 12822, pp. 778–792. Springer, Cham (2021). https://doi.org/10. 1007/978-3-030-86331-9_50

    work page 2021

  23. [23]

    Pattern Recognition144, 109834 (2023)

    Tito, R., Karatzas, D., Valveny, E.: Hierarchical multimodal transformers for multi-page docvqa. Pattern Recognition144, 109834 (2023). https://doi.org/ 10.1016/j.patcog.2023.109834

    work page arXiv 2023

  24. [24]

    CoRRabs/2501.05485 (2025), arXiv:2501.05485

    Verma, P.: S2 chunking: A hybrid framework for document segmentation through integrated spatial and semantic analysis. CoRRabs/2501.05485 (2025), arXiv:2501.05485

    Pith/arXiv arXiv 2025

  25. [25]

    CoRRabs/2401.00908 (2023), arXiv:2401.00908

    Wang,D.,Raman,N.,Sibue,M.,Ma,Z.,Babkin,P.,Kaur,S.,Pei,Y.,Nourbakhsh, A., Liu, X.: Docllm: A layout-aware generative language model for multimodal document understanding. CoRRabs/2401.00908 (2023), arXiv:2401.00908

    Pith/arXiv arXiv 2023

  26. [26]

    López et al

    Xiao, S., Liu, Z., Zhang, P., Muennighoff, N.: C-pack: Packaged resources to ad- vance general chinese embedding (2023) 18 E. López et al

    work page 2023

  27. [27]

    In: Proc

    Xu, Y., Xu, Y., Lv, T., Cui, L., Wei, F., Wang, G., Lu, Y., Florencio, D., Zhang, C., Che, W., Zhang, M., Zhou, L.: Layoutlmv2: Multi-modal pre-training for visually- rich document understanding. In: Proc. of ACL/IJCNLP (Volume 1: Long Papers). pp. 2579–2591 (2021)

    work page 2021

  28. [28]

    In: Proc

    Xu,Y.,Li,M.,Cui,L.,Huang,S.,Wei,F.,Zhou,M.:Layoutlm:Pre-trainingoftext and layout for document image understanding. In: Proc. of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD). pp. 1192–1200 (2020). https://doi.org/10.1145/3394486.3403172

    arXiv 2020

  29. [29]

    CoRR abs/2410.12628 (2024), arXiv:2410.12628

    Zhao, Z., Kang, H., Wang, B., He, C.: Doclayout-yolo: Enhancing document layout analysis through diverse synthetic data and global-to-local adaptive perception. CoRR abs/2410.12628 (2024), arXiv:2410.12628

    Pith/arXiv arXiv 2024

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.