Spatially Grounded Explanations in Vision Language Models for Document Visual Question Answering

Dimosthenis Karatzas; H\'ector Cerezo-Costas; Maximiliano Hormaz\'abal Lagos

REVIEW 3 major objections 5 minor 22 references

Spatially Grounded Explanations in Vision Language Models for Document Visual Question Answering

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read By masking a document image down to the regions singled out by the model's own explanation, EaGERS improves DocVQA exact match from 71.17% to 74.50% with no training.

desk verdict A tidy, honest training-free DocVQA grounding pipeline; the headline accuracy gain isn't yet attributable to the explanation, because no control isolates it. read the letter →

arxiv 2507.12490 v1 pith:OD63PH6V submitted 2025-07-15 cs.CV cs.AIcs.CLcs.LG

Maximiliano Hormazábal Lagos , Héctor Cerezo-Costas , Dimosthenis Karatzas This is my paper

classification cs.CVcs.AIcs.CLcs.LG

keywords DocumentVisualQuestionAnsweringVisionLanguageModelsExplainabilitySpatialGroundingMultimodalEmbeddingsTraining-freePipelineDocVQA

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

EaGERS is a training-free, model-agnostic pipeline that makes a vision-language model answer document questions only from regions the model itself can verbally justify. It first asks the model to explain where the answer is, converts the image into an $m \times n$ grid, compares embedding vectors of the explanation and each grid cell, keeps the top 30% of cells by majority voting across three embedding models, masks everything else, and re-asks the question on the masked image. On the DocVQA validation split, the best configuration (5 columns, 10 rows, 15% margin) raises exact match from 71.17% (base model) to 74.50% and ANLS from 82.90 to 83.31. The intended payoff is that answers become spatially traceable to document sub-regions without any fine-tuning or prompt engineering.

What carries the argument

EaGERS (Explanation-Guided Region Selection), a three-stage pipeline. Stage A generates a natural-language spatial explanation with the VLM; Stage B divides the image into grid cells, embeds each cell and the explanation with BLIP, CLIP, and ALIGN, computes cosine similarities, and selects the top 30% of cells by majority vote (ties broken by mean similarity); Stage C blacks out all unselected cells and re-queries the same VLM on the masked image. The mask is the load-bearing object: it forces the final answer to be reconstructed only from regions that scored highly against the explanation.

What would settle it

Run EaGERS on the DocVQA validation split with the kept cells chosen by random selection (same fraction of cells and same 15% margin) instead of by explanation-to-cell similarity; if random masking matches or exceeds the 74.50 EM of the reported configuration, then explanation-guided selection is not responsible for the gain.

Watch

Extended reading notes

Core claim

The paper's central claim is that explanation-guided masking is a viable mechanism for spatial grounding in document VQA. The authors report that, using Qwen2.5-VL-3B as the frozen backbone, a $5 \times 10$ grid with a 15% expansion of each kept region yields EM 74.50 and ANLS 83.31, both above the 71.17 EM and 82.90 ANLS of the same model answering directly. They attribute the gain to restricting the model's view to the sub-regions most similar to its own explanation, so the answer must be reconstructed from explicitly grounded evidence. The paper frames the improvement as modest but emphasizes that the method adds explainability and reproducibility at no training cost.

Load-bearing premise

The whole method works only if the model's verbal explanation points at roughly the right part of the document often enough that the embedding-based ranking keeps the true answer region visible; when the explanation is spatially wrong, the mask can hide the answer and hurt accuracy.

Editorial extensions

If this is right

If the reported numbers reproduce, other instruction-following VLMs could gain the same kind of spatial grounding and a small accuracy lift by adding the EaGERS mask step, without updating weights.
The 15% margin is essential to the gain: the 5x10 grid without margin gives EM 66.67 and ANLS 77.94, while the same grid with a 15% margin gives EM 74.50 and ANLS 83.31.
Because the retained regions are recorded, each answer comes with an explicit visible area, making outputs easier to audit than attention-only explanations.
The method trades compute for transparency: mean inference time rises from 7.21 s per question to 17.48 s, which is part of the current cost of training-free grounding.

Reading between the lines

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

The paper does not test transfer to other backbones, but if explanation-guided selection is the causal driver, swapping Qwen2.5-VL-3B for another open VLM on the same DocVQA split should preserve most of the gain.
The reported margin effect suggests an interior optimum: too little margin cuts off answers that straddle grid borders, while too much margin weakens the grounding; a margin sweep beyond 15% would test that.
A stronger version of the paper's assumption is that masking acts as an attentional prior; one way to check this is to measure whether the kept cells actually overlap the document regions named in the explanation.

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.

Desk Editor's Note

read the letter

Two things to know. First, the paper is a clean, honest incremental contribution: a training-free pipeline that generates a VLM rationale, maps it onto grid regions via CLIP/BLIP/ALIGN cosine similarity with majority voting, masks everything else, and re-queries the same VLM. Code is provided. Second, the stress-test note is correct: there is no control that isolates explanation-based selection, so the reported EM/ANLS gains cannot currently be attributed to the spatial grounding mechanism.

What is actually new: the specific combination, plus the finding that explanation-guided masking with a 15% margin improves over the base model on DocVQA validation (EM 74.50 vs 71.17; ANLS 83.31 vs 82.90 for Qwen2.5-VL-3B). The paper does not claim SOTA, and it candidly lists limitations, including the assumption that VLM explanations are spatially accurate.

The soft spots are real and mostly about experimental design. Table 1 compares EaGERS only against a single base-model pass. There is no ablation with random or fixed region selection, no margin-only mask, and no second forward pass on the full image. The 0%-margin configs perform below base; the 15%-margin configs beat it, which means the margin and the re-query might be doing most of the work. That is the load-bearing gap. Beyond that, there are no error bars or significance tests; the grid and top-k were tuned on the same validation split; only one VLM is tested despite the model-agnostic claim; and the explanations themselves are never quantitatively evaluated for localization fidelity (e.g., IoU with ground-truth answer regions or text-coherence metrics). The inference time roughly doubles, which matters for deployment.

Who should read it: people working on explainable document VQA will want this as a baseline for training-free grounding, but with caution. It deserves a serious referee because the idea is plausible, the writing is clear, and the missing control is a tractable fix. I'd send it to review and request the ablation, a second VLM, and error bars before accepting. Otherwise it's a solid workshop paper.

Referee Report

3 major / 5 minor

Summary. The paper proposes EaGERS, a training-free and model-agnostic pipeline for Document Visual Question Answering. The method first asks a vision-language model (Qwen2.5-VL-3B) to generate a natural-language explanation, then divides the document image into an m×n grid, embeds the explanation and each grid region with BLIP, CLIP, and ALIGN, selects the top 30% of regions by cosine similarity and majority voting, masks all other regions, and re-queries the same VLM on the masked image to produce the final answer. Experiments on the DocVQA validation split report that the best configuration (5 columns, 10 rows, 15% margin) achieves 74.50 exact match and 83.31 ANLS, compared with 71.17 EM and 82.90 ANLS for the base model run directly. The paper also reports inference-time overhead and discusses limitations including fixed-grid partitioning and dependence on the spatial accuracy of VLM explanations.

Significance. If the claimed effect is real, the paper demonstrates a useful and inexpensive way to add spatial grounding to off-the-shelf VLMs: no fine-tuning, model-agnostic, and with a released code repository. The main strengths are the simplicity of the pipeline, the inclusion of three complementary embedding models, and the measurement of both accuracy and inference time. The reported best-configuration numbers do support the headline improvement over the base model, but the paper does not establish that the improvement comes from explanation-guided region selection rather than from the additional forward pass, the masking procedure, or the margin expansion. The explanatory/transparency contribution is also asserted rather than directly measured, since no fidelity metric such as IoU is reported.

major comments (3)

[§4.2, Table 1] The experimental design does not include a control that isolates explanation-based selection. The only comparisons are EaGERS configurations versus the base model, so the reported gains (EM +3.33, ANLS +0.41 in the best configuration) could arise from the second forward pass, the 15% margin that partially restores visibility around selected cells, the masking operation itself, or general attention focusing, rather than from the spatial accuracy of the generated explanation. To support the central claim that explanation-guided grounding is what improves DocVQA accuracy, the authors should add an ablation that keeps the masking/re-query pipeline unchanged but replaces explanation-based ranking with random selection, fixed/central selection, or selection based on a trivial heuristic. This is load-bearing because it concerns whether the proposed mechanism, not merely the pipeline, is responsible for the result.
[§4.1 and §4.2] The hyperparameters that define the best configuration — top-k ratio k=30%, grid dimensions (5×5 and 5×10), and margin size (15%) — appear to be selected on the same validation split that is used to report the final numbers, with no error bars, confidence intervals, or significance tests. Without a held-out test split or statistical uncertainty quantification, the advantage of EaGERS 50|15 over the base model may be inflated by selection bias. The authors should either report results on a separate test split, or provide bootstrap confidence intervals over the validation set, and clearly state whether the reported configuration was chosen before or after seeing these metrics.
[§5, Limitations and Future Work] The paper explicitly acknowledges that the pipeline assumes spatial accuracy of VLM explanations and that inaccurate rationales can degrade fidelity, but it does not quantify how often this occurs or measure the spatial fidelity of the selected regions (e.g., IoU with answer bounding boxes or visual-text coherence scores). Since one of the stated contributions is traceability and explainability, the absence of any direct fidelity measurement leaves the 'spatially grounded' aspect of the claim supported only indirectly through downstream accuracy. Adding a region-fidelity evaluation on a subset with localization annotations, or at least a manual error analysis of selected regions, would substantially strengthen the paper.

minor comments (5)

[§2.2] The section heading 'Explainability en DocVQA' should be 'Explainability in DocVQA'.
[Table 1] The column header 'A vg Time' contains a stray space, and the acronym CV is not defined in the caption or in the text where it first appears.
[§3.2] The choice of k=30% is said to be 'based on preliminary experiments,' but no details of those experiments, such as the range of k values tried or the metric used for selection, are provided; this is needed for reproducibility.
[§4.1] The resizing preprocessing is mentioned but no target resolution or resizing rule is given; please specify the exact preprocessing steps so that the reported timing and accuracy numbers can be reproduced.
[Abstract and §6] The paper states that the pipeline 'enhances transparency and reproducibility,' but no user study, quantitative explainability metric, or region-fidelity measure is reported; this claim should be softened or empirically supported.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: EaGERS is an empirical pipeline whose final answers are scored against ground truth, with no fitted parameter or definitional equivalence between inputs and outputs.

full rationale

The paper's claimed chain — generate an explanation, embed the explanation and grid cells with BLIP/CLIP/ALIGN, select top cells by cosine similarity and majority voting, mask the remaining cells, re-query the same VLM, and measure EM/ANLS on DocVQA — contains no step in which a predicted quantity is defined in terms of the target quantity or in which a fitted parameter is renamed as a prediction. The k=30% region fraction and the grid/margin choices are reported as configuration choices, with Table 1 comparing several settings against the base model; none of these values is derived from the final EM/ANLS numbers by construction. The same VLM supplies both the explanation and the final answer, but the final answer is forced to be read from the masked image and is scored against ground-truth answers, so the explanation is an input to the selection mechanism rather than the object being predicted. The acknowledged assumption in Section 5 — that explanation spatial accuracy may fail and degrade fidelity — is a failure-mode caveat, not a circular reduction. The absence of a random-selection ablation is a valid attribution concern about whether explanation-based selection causes the observed gain, but that is an experimental control issue, not a circularity of the derivation. No load-bearing self-citations or imported uniqueness claims appear in the methodology.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central result rests on three domain assumptions: the VLM explanation localizes the answer, embedding cosine similarity captures relevance, and a fixed grid isolates answer regions. The paper explicitly acknowledges the first and third assumptions in Section 5. No new entities are introduced, and the free parameters are the top-k fraction, grid size, and margin, which are chosen on the validation split.

free parameters (2)

top_k selection ratio (k = 30%) = 30% of grid cells
Chosen "based on preliminary experiments" in Section 3.2 rather than fixed a priori; affects which regions are kept for the final answer.
grid dimensions and margin = 5x10 grid with 15% margin (best)
Several grid/margin configurations are evaluated on the DocVQA validation split in Section 4.2 and the best is selected, so these are tuned on the evaluation set.

assumptions (3)

domain assumption VLM-generated spatial explanations are accurate enough to localize answer regions
Stated explicitly in Section 5: "Our pipeline assumes spatial accuracy of VLM explanations; however, when these do not match ground truth, the final fidelity may degrade."
domain assumption Cosine similarity in BLIP/CLIP/ALIGN embedding space reflects semantic relevance of document sub-regions to the explanation
Used in Section 3.2 to rank grid cells; no validation is provided that this similarity aligns with true answer location.
domain assumption A fixed grid partitions documents in a way that captures answer-relevant regions
Acknowledged as a limitation in Sections 3.2 and 5: fixed grids may not generalize to irregular layouts or split relevant zones across cell borders.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Spatially Grounded Explanations in Vision Language Models for Document Visual Question Answering." pith.science (2026). https://pith.science/paper/OD63PH6V

@misc{pith2026250712490,
  author       = {Pith},
  title        = {Pith review of: Spatially Grounded Explanations in Vision Language Models for Document Visual Question Answering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OD63PH6V}},
  note         = {Machine review of arXiv:2507.12490}
}

read the original abstract

We introduce EaGERS, a fully training-free and model-agnostic pipeline that (1) generates natural language rationales via a vision language model, (2) grounds these rationales to spatial sub-regions by computing multimodal embedding similarities over a configurable grid with majority voting, and (3) restricts the generation of responses only from the relevant regions selected in the masked image. Experiments on the DocVQA dataset demonstrate that our best configuration not only outperforms the base model on exact match accuracy and Average Normalized Levenshtein Similarity metrics but also enhances transparency and reproducibility in DocVQA without additional model fine-tuning.

Figures

Figures reproduced from arXiv: 2507.12490 by the authors.

**Figure 1.** EaGERS Document VQA pipeline: (1) the multimodal model generates a spatial natural language explanation from the image and the question; (2) the image is segmented into an m × n grid; (3) embeddings of the explanation and each sub-region are obtained using BLIP, CLIP, and ALIGN; (4) majority voting selects the most relevant regions; (5) the image is masked to retain only those regions, and the model is re-queried … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 14 canonical work pages

[1]

In: Proceedin gs of the IEEE/CVF international conference on computer vision

Appalaraju, S., Jasani, B., Kota, B.U., Xie, Y., Manmatha , R.: Docformer: End-to- end transformer for document understanding. In: Proceedin gs of the IEEE/CVF international conference on computer vision. pp. 993–1003 (2021)

work page 2021
[2]

4290–4300

Biten, A., Tito, R., Maﬂa, A., Gomez, L., Rusinol, M., Jawa har, C., Val- veny, E., Karatzas, D.: Scene text visual question answerin g, pp. 4290–4300. Proceedings of the IEEE International Conference on Comput er Vision, Insti- tute of Electrical and Electronics Engineers Inc., United S tates (Oct 2019). https://doi.org/10.1109/ICCV.2019.00439, funding I...

arXiv 2019
[3]

Pattern Recognition Letters 150, 242–249 (2021)

Gómez, L., Biten, A.F., Tito, R., Maﬂa, A., Rusiñol, M., Va lveny, E., Karatzas, D.: Multimodal grid features and cell pointers fo r scene text visual question answering. Pattern Recognition Letters 150, 242–249 (2021). https://doi.org/https://doi.org/10.1016/j.patrec.2021.06.026, https://www.sciencedirect.com/science/article/pii/S0167865521002336

work page doi:10.1016/j.patrec.2021.06.026 2021
[4]

, Huang, F., Zhou, J.: mplug-docowl2: High-resolution compressing for ocr-f ree multi-page document understanding (2024), https://arxiv.org/abs/2409.03420

Hu, A., Xu, H., Zhang, L., Ye, J., Yan, M., Zhang, J., Jin, Q. , Huang, F., Zhou, J.: mplug-docowl2: High-resolution compressing for ocr-f ree multi-page document understanding (2024), https://arxiv.org/abs/2409.03420

arXiv 2024
[5]

In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR)

Hu, R., Singh, A., Darrell, T., Rohrbach, M.: Iterative An swer Pre- diction With Pointer-Augmented Multimodal Transformers f or TextVQA . In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR). pp. 9989–9999. IEEE Computer Society, Los Al ami- tos, CA, USA (Jun 2020). https://doi.org/10.1109/CVPR42600.2020.01001, https://doi.iee...

arXiv 2020
[6]

, Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-langu age representation learning with noisy text supervision

Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H. , Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-langu age representation learning with noisy text supervision. In: Meila, M., Zhang, T. (eds.) Proceed- ings of the 38th International Conference on Machine Learni ng. Proceedings of Machine Learning Research, vol. 139, pp. 4...

work page 2021
[7]

In: European Conference on Computer Vision

Kim, G., Hong, T., Yim, M., Nam, J., Park, J., Yim, J., Hwang , W., Yun, S., Han, D., Park, S.: Ocr-free document understanding transfo rmer. In: European Conference on Computer Vision. pp. 498–517. Springer (2022 )

work page 2022
[8]

In: Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., Scarlet t, J

Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: Bootstrappi ng language-image pre- training with frozen image encoders and large language mode ls. In: Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., Scarlet t, J. (eds.) Proceed- ings of the 40th International Conference on Machine Learni ng. Proceedings of Machine Learning Research, vol. 20...

work page 2023

Show all 22 references

[9]

In: Chaudhuri, K., Jegelka, S., Song, L., Szepesvari, C., Niu, G., Sabato, S

Li, J., Li, D., Xiong, C., Hoi, S.: BLIP: Bootstrapping lan guage-image pre- training for uniﬁed vision-language understanding and gen eration. In: Chaudhuri, K., Jegelka, S., Song, L., Szepesvari, C., Niu, G., Sabato, S . (eds.) Proceedings of the 39th International Conferenc...

2022
[10]

Li, K., Vosselman, G., Yang, M.Y.: Convincing rationale s for visual question an- swering reasoning (2025), https://arxiv.org/abs/2402.03896

2025 arXiv
[11]

In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M

Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tun ing. In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M. , Levine, S. (eds.) Advances in Neural Information Processin g Sys- tems. vol. 36, pp. 34892–34916. Curran Associates, Inc. (20 23), https://proceeding...

2023
[12]

In: Proceedings of the IEEE/CVF winter conference o n applications of computer vision

Mathew, M., Karatzas, D., Jawahar, C.: Docvqa: A dataset for vqa on document images. In: Proceedings of the IEEE/CVF winter conference o n applications of computer vision. pp. 2200–2209 (2021)

2021
[13]

Mohammadshirazi, A., Neogi, P.P.G., Lim, S.N., Ramnath , R.: Dlava: Document language and vision assistant for answer localization with enhanced interpretability and trustworthiness (2024), https://arxiv.org/abs/2412.00151

2024 arXiv
[14]

, Pietruszka, M., Pałka, G.: Going full-tilt boogie on document understanding with t ext-image-layout trans- former

Powalski, R., Borchmann, Ł., Jurkiewicz, D., Dwojak, T. , Pietruszka, M., Pałka, G.: Going full-tilt boogie on document understanding with t ext-image-layout trans- former. In: Document Analysis and Recognition–ICDAR 2021: 16th International Conference, Lausanne, Switzerland, ...

2021
[15]

In: Meila , M., Zhang, T

Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila , M., Zhang, T. (eds.) Proceedings of the 38th In...

2021
[16]

pp. 732–747. Springer (2021)

2021
[17]

In: 2019 IEEE/CVF Conference on Computer Vi sion and Pattern Recognition (CVPR)

Rezatoﬁghi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I., Savarese, S.: Generalized Intersection Over Union: A Metric and a Loss for Bounding Box Regression . In: 2019 IEEE/CVF Conference on Computer Vi sion and Pattern Recognition (CVPR). pp. 658–666. IEEE Computer Soc iet...

2019
[18]

I n: International Confer- ence on Document Analysis and Recognition

Saifullah, S., Agne, S., Dengel, A., Ahmed, S.: Docxplai n: A novel model-agnostic explainability method for document image classiﬁcation. I n: International Confer- ence on Document Analysis and Recognition. pp. 103–123. Spr inger (2024)

2024
[19]

In: 2017 IEEE International Conference on Computer Vision (ICC V)

Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Pa rikh, D., Batra, D.: Grad- cam: Visual explanations from deep networks via gradient-b ased localization. In: 2017 IEEE International Conference on Computer Vision (ICC V). pp. 618–626 (2017). https://doi.org/10.1109/ICCV.2017.74

2017 doi
[20]

In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recogni tion (CVPR)

Singh, A., Hu, R., Goswami, V., Couairon, G., Galuba, W., Rohrbach, M., Kiela, D.: Flava: A foundational language and vision alignm ent model. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recogni tion (CVPR). pp. 15617–15629 (2022). https://doi.org/10.1109/CVPR5...

2022
[21]

In: Proceedings of the 26th ACM SIGKDD International Confer - ence on Knowledge Discovery & Data Mining

Xu, Y., Li, M., Cui, L., Huang, S., Wei, F., Zhou, M.: Layou tlm: Pre-training of text and layout for document image understa nding. In: Proceedings of the 26th ACM SIGKDD International Confer - ence on Knowledge Discovery & Data Mining. p. 1192–1200. KDD ’20, ACM (Aug 2020). ...

2020
[22]

In: The E leventh International Conference on Learning Representations (2022)

Zhou, Y., Muresanu, A.I., Han, Z., Paster, K., Pitis, S., Chan, H., Ba, J.: Large language models are human-level prompt engineers. In: The E leventh International Conference on Learning Representations (2022)

2022

Pith tools

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

[1] [1]

In: Proceedin gs of the IEEE/CVF international conference on computer vision

Appalaraju, S., Jasani, B., Kota, B.U., Xie, Y., Manmatha , R.: Docformer: End-to- end transformer for document understanding. In: Proceedin gs of the IEEE/CVF international conference on computer vision. pp. 993–1003 (2021)

work page 2021

[2] [2]

4290–4300

Biten, A., Tito, R., Maﬂa, A., Gomez, L., Rusinol, M., Jawa har, C., Val- veny, E., Karatzas, D.: Scene text visual question answerin g, pp. 4290–4300. Proceedings of the IEEE International Conference on Comput er Vision, Insti- tute of Electrical and Electronics Engineers Inc., United S tates (Oct 2019). https://doi.org/10.1109/ICCV.2019.00439, funding I...

arXiv 2019

[3] [3]

Pattern Recognition Letters 150, 242–249 (2021)

Gómez, L., Biten, A.F., Tito, R., Maﬂa, A., Rusiñol, M., Va lveny, E., Karatzas, D.: Multimodal grid features and cell pointers fo r scene text visual question answering. Pattern Recognition Letters 150, 242–249 (2021). https://doi.org/https://doi.org/10.1016/j.patrec.2021.06.026, https://www.sciencedirect.com/science/article/pii/S0167865521002336

work page doi:10.1016/j.patrec.2021.06.026 2021

[4] [4]

, Huang, F., Zhou, J.: mplug-docowl2: High-resolution compressing for ocr-f ree multi-page document understanding (2024), https://arxiv.org/abs/2409.03420

Hu, A., Xu, H., Zhang, L., Ye, J., Yan, M., Zhang, J., Jin, Q. , Huang, F., Zhou, J.: mplug-docowl2: High-resolution compressing for ocr-f ree multi-page document understanding (2024), https://arxiv.org/abs/2409.03420

arXiv 2024

[5] [5]

In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR)

Hu, R., Singh, A., Darrell, T., Rohrbach, M.: Iterative An swer Pre- diction With Pointer-Augmented Multimodal Transformers f or TextVQA . In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR). pp. 9989–9999. IEEE Computer Society, Los Al ami- tos, CA, USA (Jun 2020). https://doi.org/10.1109/CVPR42600.2020.01001, https://doi.iee...

arXiv 2020

[6] [6]

, Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-langu age representation learning with noisy text supervision

Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H. , Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-langu age representation learning with noisy text supervision. In: Meila, M., Zhang, T. (eds.) Proceed- ings of the 38th International Conference on Machine Learni ng. Proceedings of Machine Learning Research, vol. 139, pp. 4...

work page 2021

[7] [7]

In: European Conference on Computer Vision

Kim, G., Hong, T., Yim, M., Nam, J., Park, J., Yim, J., Hwang , W., Yun, S., Han, D., Park, S.: Ocr-free document understanding transfo rmer. In: European Conference on Computer Vision. pp. 498–517. Springer (2022 )

work page 2022

[8] [8]

In: Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., Scarlet t, J

Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: Bootstrappi ng language-image pre- training with frozen image encoders and large language mode ls. In: Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., Scarlet t, J. (eds.) Proceed- ings of the 40th International Conference on Machine Learni ng. Proceedings of Machine Learning Research, vol. 20...

work page 2023