Pith. sign in

REVIEW 3 major objections 1 cited by

A four-stage multimodal RAG pipeline that decomposes questions, verifies evidence pages across images, and iterates with memory lifts long-document accuracy by about 8 points on average.

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 · grok-4.5

2026-07-11 16:12 UTC pith:PXV575A4

load-bearing objection Competent four-stage multimodal RAG assembly with real ablations and ~8% gains; EVIAGENT generalization is the main open risk, not a contradiction in the reported data. the 3 major comments →

arxiv 2607.04625 v1 pith:PXV575A4 submitted 2026-07-06 cs.CV cs.AI

Hierarchical Evidence-Driven Reasoning for Long Document Understanding

Junyu Xiong , Yonghui Wang , Rongjian Gu , Chenyu Liu , Bing Yin , Wengang Zhou , Houqiang Li This is my paper
classification cs.CV cs.AI
keywords multimodal RAGlong document understandingvisual page retrievalevidence verificationhierarchical question decompositionmemory-guided generationGRPODocVQA
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.

Long closed-domain documents pack answers sparsely across hundreds of page images, so systems must retrieve a small subset before generating. Standard visual similarity retrieval often surfaces topically related pages that contain no actual answer, and single-pass pipelines cannot recover once a key page is missed. This paper claims that factorizing the problem into hierarchical question decomposition, coarse retrieval, fine-grained multi-page evidence verification, and memory-guided iterative generation solves both failure modes. A specialized verifier trained with group-relative policy optimization judges blocks of pages together, discards or reorders them, and an explicit memory of sub-question traces lets the model keep scanning until the root question is answerable. Across four public benchmarks the full pipeline outperforms prior open-source systems by an average of 8.05 accuracy points, with each stage contributing measurable gains.

Core claim

The paper establishes that long multimodal document QA is made reliable by a cooperative four-stage RAG architecture: break a multi-hop root query into atomic child questions, retrieve candidate page images by visual late-interaction similarity, verify and re-rank those candidates with a cross-page evidence agent, then answer the root query by sliding a window over the reordered pages while accumulating sub-question reasoning traces in an explicit text memory. This combination suppresses answer-void distractors and recovers from initial retrieval omissions, producing the reported accuracy lifts.

What carries the argument

EVIAGENT: a vision-language model trained with evidence-aware GRPO (EviGRPO) to emit structured reasoning, a binary true/false evidence label for every page in a multi-image block, and an answer; its verified pages are kept for child questions and placed first in a prioritized sequence for the root question.

Load-bearing premise

The trained multi-page verifier, optimized on a few thousand samples from six multi-page document datasets with fixed reward weights and a chosen block size, generalizes as a trustworthy evidence filter on the four evaluation benchmarks without domain shift or reward hacking that would erase the gains.

What would settle it

Replace the trained EVIAGENT with an untuned same-size vision-language model (or disable verification entirely) on the same four benchmarks and measure whether the average accuracy advantage over the strongest prior baseline disappears or falls below a few points, while also checking whether retrieval NDCG and recall stop rising after the verification step.

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

If this is right

  • Complex multi-hop document questions can be answered more reliably by first retrieving for simpler atomic sub-questions and seeding memory with their traces.
  • Cross-page joint verification of image blocks measurably raises Recall, NDCG and MRR of the candidate set compared with raw semantic retrieval.
  • Keeping non-evidence pages as a reordered backup suffix, then iterating with memory until an answer is found, reduces cascading failure from a single missed page.
  • Ablations that remove decomposition, verification or iteration each lower average accuracy, showing the stages are complementary rather than redundant.
  • The same pipeline can be paired with stronger generators or retrievers and still improve, indicating the gains are architectural.

Where Pith is reading between the lines

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

  • If the verifier’s binary decisions remain reliable under domain shift, the same evidence-first reordering could be dropped into other page-level multimodal RAG stacks without redesigning the generator.
  • The explicit text memory of sub-question traces suggests a natural path to auditable intermediate justifications that enterprise users could inspect before trusting a final answer.
  • Because latency grows with the number of sequential model calls, practical deployments may need early-exit heuristics or parallel block verification to keep interactive response times acceptable.
  • Training the verifier on a broader mix of degraded scans and handwritten pages would test whether the same cross-page reasoning objective still filters distractors outside clean born-digital 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 / 0 minor

Summary. The paper proposes HIEVI-RAG, a four-stage multimodal RAG pipeline for closed-domain long-document understanding. It decomposes multi-hop root queries into atomic child questions, performs coarse visual retrieval (Ops-ColQwen3), applies fine-grained multi-page verification and reranking with EVIAGENT (a Qwen3-VL-8B model trained via EviGRPO on six multi-page DocVQA datasets with a joint format/evidence-F1/answer reward), and then runs memory-guided iterative generation over a prioritized evidence sequence. On PaperTab, FetaTab, MMLongBench and LongDocURL the system reports 44.0/72.9/48.2/65.7 accuracy, an average absolute gain of 8.05% over the strongest open-source baselines, with ablations attributing measurable drops to removal of verification, decomposition and iteration.

Significance. If the reported gains hold under fair comparison, the work is a useful systems contribution to multimodal document RAG: it cleanly factorizes recall-oriented retrieval, cross-page evidence verification, and memory-backed multi-round generation, and supplies coherent module ablations (Table 3), backbone/retriever sensitivity (Table 4) and retrieval-metric lifts for the trained verifier (Table 5). The explicit training of a multi-image evidence agent with GRPO and the memory-iteration safety net are concrete, reusable design choices that address two well-recognized failure modes (answer-void distractors and cascading single-pass omissions). The contribution is empirical rather than theoretical; its lasting value depends on the reliability of EVIAGENT as a domain-robust binary filter.

major comments (3)
  1. §3.4 and Appendix D: EVIAGENT is trained on 6 000 samples (1k each from DUDE, MP-DocVQA, MultiHiertt, pdfQA, SciEGQA, SlideVQA) whose average page counts lie between ~4 and 20. Evaluation documents are substantially longer (LongDocURL avg 89 pages) or table-centric (PaperTab/FetaTab). Table 5 reports retrieval lifts only on MMLongBench and LongDocURL; no page-level precision/recall or evidence-F1 is given for the two table benchmarks. Because the end-to-end +8.05% claim and the largest ablation drop (Table 3, −4.5 avg for w/o verification) rest on this verifier, the manuscript needs either (i) direct evidence-label accuracy of EVIAGENT on the four evaluation distributions or (ii) an explicit domain-shift analysis showing that the binary decisions remain reliable.
  2. §3.4 Eq. (4): the joint reward weights answer accuracy (λ_acc=0.4) equally with evidence F1 (λ_evi=0.5). The policy can therefore improve the scalar reward by generating correct answers from partial or noisy page sets rather than by producing faithful binary evidence labels. Without an ablation that isolates revi (or reports evidence-F1 on held-out multi-page data after training), it is unclear how much of the Table 5 NDCG/MRR gains and the downstream accuracy lifts are pure verification versus incidental generation skill. A pure-evidence or evidence-only reward variant would strengthen the central architectural claim.
  3. Table 2 and §4.2: several strong baselines report “–” on PaperTab/FetaTab, and generator/retriever pairings are not fully matched (e.g., Doc-V⋆ and ALDEN use different backbones). The claimed 8.05% average margin is therefore computed over a heterogeneous set. A controlled re-evaluation of the strongest published pipelines under the identical Ops-ColQwen3 + Qwen3-VL-8B stack (or an explicit statement of which baselines are incomparable) is needed before the absolute ranking can be treated as definitive.

Circularity Check

0 steps flagged

No circularity: empirical RAG systems paper whose accuracy claims are measured on external benchmarks, not forced by construction or self-citation chains.

full rationale

HIEVI-RAG is a four-stage multimodal RAG pipeline (hierarchical decomposition, coarse ColPali-style retrieval, EVIAGENT verification via EviGRPO, memory-guided iteration). Its strongest claim is end-to-end accuracy on four public benchmarks (PaperTab 44.0, FetaTab 72.9, MMLongBench 48.2, LongDocURL 65.7; +8.05% avg over strongest open-source baselines in Table 2), with ablations (Table 3) and retrieval metrics (Table 5) isolating module contributions. These numbers are obtained by running the full system against fixed external evaluation protocols and LLM-as-judge or rule-based scoring; they are not algebraic identities or fitted constants renamed as predictions. EVIAGENT is trained on 6 000 samples drawn from six multi-page DocVQA-style corpora (Appendix D) that are distinct from the four evaluation sets; the joint reward (Eq. 4) is a standard RL objective, not a definition that forces the reported test accuracies. The sole self-citation of note is methodological inspiration from DocR1 (same first author) for the GRPO training recipe; it is not invoked as a uniqueness theorem that forbids alternatives or that itself contains the target result. No equation equates a claimed prediction to a fitted input by construction, no ansatz is smuggled in via citation as a forced form, and no known empirical pattern is merely renamed. The derivation chain is therefore self-contained against external benchmarks; residual risks (domain shift of the verifier, reward design) are correctness/generalization concerns, not circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

The central accuracy claim rests on standard multimodal RAG assumptions plus several design choices fixed by the authors (Top-K, block size, reward weights, max child questions, model family). No new physical entities; invented entities are system modules. Free parameters are hyperparameters that affect reported gains; axioms are domain assumptions about retrieval, RL verification, and memory helpfulness rather than formal math postulates.

free parameters (5)
  • Top-K candidate pages
    Fixed K=20 for root and child queries; controls recall/noise tradeoff entering verification.
  • Verification block size k
    Default k=7; Figure 2 shows retrieval quality depends on this choice; not derived from first principles.
  • Max child questions m
    m≤5 from lightweight Qwen3.5-4B decomposition; arbitrary cap on hierarchy breadth.
  • EviGRPO reward weights (λ_fmt, λ_evi, λ_acc)
    Set to (0.1, 0.5, 0.4); prioritizes evidence F1; directly shapes what EVIAGENT optimizes.
  • Training sample budget per dataset
    Uniform 1,000 instances from each of six multi-page datasets; hand-chosen balance, not theoretically fixed.
axioms (5)
  • domain assumption Late-interaction visual page retrieval (ops-ColQwen3 / ColPali-style) provides a high-recall candidate pool from rendered page images without OCR.
    Assumed in §3.3 offline index + online matching; inherited from prior multimodal RAG literature.
  • domain assumption Complex multi-hop root questions can be usefully factored into ≤5 non-redundant atomic child questions by a lightweight LLM.
    §3.2 hierarchical decomposition; if false, memory seeding and retrieval coverage degrade.
  • ad hoc to paper Cross-page binary evidence decisions optimized with GRPO (format + evidence F1 + answer accuracy) yield a verifier that generalizes to held-out long-document benchmarks.
    Core of EVIAGENT training §3.4; load-bearing for the claimed distractor filtering gains.
  • domain assumption Textual memory of child (question, trace, answer) triplets and intermediate root traces improves root multi-hop answering under sliding windows.
    §3.5 memory-guided iterative generation; Limitations note hallucination cascade risk if traces are wrong.
  • domain assumption LLM-as-judge binary accuracy with Qwen3.5-27B is a faithful metric for PaperTab and FetaTab free-form answers.
    §4.1 Metrics; used for two of four main numbers without reported human agreement.
invented entities (3)
  • EVIAGENT (multi-page evidence verifier) no independent evidence
    purpose: Jointly reason over multi-image blocks to output structured think / evidence_page T-F sequence / answer for verification and generation.
    New trained policy specialized for the pipeline; independent_evidence false beyond this paper’s benchmarks and training sets.
  • EviGRPO reward formulation no independent evidence
    purpose: Weighted joint reward R=λ_fmt r_fmt + λ_evi r_evi + λ_acc r_acc with evidence F1 against gold page labels.
    Adaptation of GRPO/DocR1-style training to multi-page evidence decisions; not an external physical entity.
  • HIEVI-RAG four-stage pipeline (decompose → coarse retrieve → verify/rerank → memory-iterate) no independent evidence
    purpose: End-to-end closed-domain long-document multimodal QA system combining the above stages with distinct child vs root evidence policies.
    The paper’s primary system-level invention; evaluated only in this work’s experimental setup.

pith-pipeline@v1.1.0-grok45 · 22177 in / 3882 out tokens · 33137 ms · 2026-07-11T16:12:22.214142+00:00 · methodology

0 comments
read the original abstract

Retrieval-Augmented Generation (RAG) streamlines long-document understanding by leveraging retrieval mechanisms to restrict input images to a highly curated subset. However, existing multimodal RAG pipelines primarily face two critical challenges: first, standard semantic similarity retrievers frequently fetch topically overlapping yet answer-void distractor pages that mislead downstream generation; second, rigid single-pass pipelines heavily depend on initial retrieval success, where any omission of core evidence inevitably causes cascading errors. To address these challenges, we introduce HIEVI-RAG, a hierarchical, evidence-driven multimodal RAG framework for closed-domain document understanding. HIEVI-RAG systematically factorizes complex queries into a cooperative four-stage pipeline: (1) hierarchical question decomposition to break multi-hop root queries into atomic child questions; (2) coarse visual page retrieval leveraging a multimodal retriever to fetch candidate pages based on semantic similarity; (3) fine-grained page verification via EVIAGENT, a specialized multi-page verifier trained with GRPO to execute cross-page reasoning over multi-image blocks; and (4) memory-guided iterative generation that leverages accumulated sub-question context to execute multi-round, dynamic reasoning over the prioritized sequence. Extensive evaluations across four benchmarks demonstrate the robust efficacy and synergy of our framework, which significantly outperforms existing open-source baselines and exceeds the strongest reported baseline by an average of 8.05% in accuracy.

Figures

Figures reproduced from arXiv: 2607.04625 by Bing Yin, Chenyu Liu, Houqiang Li, Junyu Xiong, Rongjian Gu, Wengang Zhou, Yonghui Wang.

Figure 1
Figure 1. Figure 1: Overview of HIEVI-RAG. The system first builds a shallow hierarchy from the original question to atomic child questions, retrieves Top-K candidate pages for both root and child questions with a visual retriever, verifies and reranks grouped-k candidates with EVIAGENT, and performs memory-guided iterative reasoning over the verified evidence order. 3.4 Evidence-Aware Page Verification To filter out topicall… view at source ↗
Figure 2
Figure 2. Figure 2: Effect of block size k on MMLongBench retrieval quality. A moderate block size provides the best balance between cross-page comparison and visual-context overload. consistently degrades retrieval precision on Long￾DocURL, with NDCG slipping by −1.3 points at Top-1 and −1.0 points at Top-5. This demonstrates that off-the-shelf VLMs lack the necessary align￾ment to distinguish authentic evidence from com￾ple… view at source ↗
Figure 3
Figure 3. Figure 3: System prompt for hierarchical question decomposition. [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: System prompt used for EVIAGENT training and inference. Question: What is the sum of the total number of paid search's conversions in the year of 2007, 2008 and the number of green bars in the heroes happen here launch? Sub-question 1: What is the number of paid search conversions in 2007? Sub-question 2: What is the number of paid search conversions in 2008? Sub-question 3: How many green bars are shown i… view at source ↗
Figure 5
Figure 5. Figure 5: Case study of hierarchical question decomposition in H [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] 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. HierDoc: Hierarchical Page-to-Region Evidence Routing for Long-Document Visual Question Answering

    cs.CV 2026-07 conditional novelty 6.0

    HierDoc trains separate page- and region-selection policies with GRPO, then answers from full selected pages plus selected region crops and text.

Reference graph

Works this paper leans on

71 extracted references · 7 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Retrieval-Augmented Generation for Knowledge-Intensive

    Lewis, Patrick and Perez, Ethan and Piktus, Aleksandra and Petroni, Fabio and Karpukhin, Vladimir and Goyal, Naman and K. Retrieval-Augmented Generation for Knowledge-Intensive. Advances in Neural Information Processing Systems 33 , year =

  2. [2]

    2020 , doi =

    Khattab, Omar and Zaharia, Matei , booktitle =. 2020 , doi =

    2020

  3. [3]

    Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages =

    A Dataset of Information-Seeking Questions and Answers Anchored in Research Papers , author =. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages =. 2021 , address =

    2021

  4. [4]

    Transactions of the Association for Computational Linguistics , volume =

    Nan, Linyong and Hsieh, Chiachun and Mao, Ziming and Lin, Xi Victoria and Verma, Neha and Zhang, Rui and Kry. Transactions of the Association for Computational Linguistics , volume =. 2022 , publisher =

    2022

  5. [5]

    Hierarchical Multimodal Transformers for Multi-Page

    Tito, Rub. Hierarchical Multimodal Transformers for Multi-Page. arXiv preprint arXiv:2212.05935 , year =

    Pith/arXiv arXiv

  6. [6]

    2022 , address =

    Zhao, Yilun and Li, Yunxiang and Li, Chenying and Zhang, Rui , booktitle =. 2022 , address =

    2022

  7. [7]

    2023 , doi =

    Tanaka, Ryota and Nishida, Kyosuke and Nishida, Kosuke and Hasegawa, Taku and Saito, Itsumi and Saito, Kuniko , journal =. 2023 , doi =

    2023

  8. [8]

    Document Understanding Dataset and Evaluation (

    Van Landeghem, Jordy and Tito, Rub. Document Understanding Dataset and Evaluation (. arXiv preprint arXiv:2305.08455 , year =. doi:10.48550/arXiv.2305.08455 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2305.08455

  9. [9]

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

    Measuring and Narrowing the Compositionality Gap in Language Models , author =. Findings of the Association for Computational Linguistics: EMNLP 2023 , pages =. 2023 , address =

    2023

  10. [10]

    The Eleventh International Conference on Learning Representations , year =

    Least-to-Most Prompting Enables Complex Reasoning in Large Language Models , author =. The Eleventh International Conference on Learning Representations , year =

  11. [11]

    and Cao, Yuan , booktitle =

    Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik R. and Cao, Yuan , booktitle =. 2023 , url =

    2023

  12. [12]

    Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages =

    Multi-hop Reading Comprehension through Question Decomposition and Rescoring , author =. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages =. 2019 , address =

    2019

  13. [13]

    The Thirteenth International Conference on Learning Representations , year =

    Faysse, Manuel and Sibille, Hugues and Wu, Tony and Omrani, Bilel and Viaud, Gautier and Hudelot, C. The Thirteenth International Conference on Learning Representations , year =

  14. [14]

    2025 , doi =

    Xiong, Junyu and Wang, Yonghui and Zhao, Weichao and Liu, Chenyu and Yin, Bing and Zhou, Wengang and Li, Houqiang , journal =. 2025 , doi =

    2025

  15. [15]

    2025 , doi =

    Gong, Ziyu and Mai, Chengcheng and Huang, Yihua , journal =. 2025 , doi =

    2025

  16. [16]

    2025 , address =

    Chia, Yew Ken and Cheng, Liying and Chan, Hou Pong and Song, Maojia and Liu, Chaoqun and Aljunied, Mahani and Poria, Soujanya and Bing, Lidong , booktitle =. 2025 , address =

    2025

  17. [17]

    2024 , url =

    Asai, Akari and Wu, Zeqiu and Wang, Yizhong and Sil, Avirup and Hajishirzi, Hannaneh , booktitle =. 2024 , url =

    2024

  18. [18]

    Advances in Neural Information Processing Systems 36 , year =

    Reflexion: Language Agents with Verbal Reinforcement Learning , author =. Advances in Neural Information Processing Systems 36 , year =

  19. [19]

    2024 , doi =

    Hui, Yulong and Lu, Yao and Zhang, Huanchen , booktitle =. 2024 , doi =

    2024

  20. [20]

    2024 , url =

    Cho, Jaemin and Mahata, Debanjan and Irsoy, Ozan and He, Yujie and Bansal, Mohit , journal =. 2024 , url =

    2024

  21. [21]

    2025 , url =

    Han, Siwei and Xia, Peng and Zhang, Ruiyi and Sun, Tong and Li, Yun and Zhu, Hongtu and Yao, Huaxiu , journal =. 2025 , url =

    2025

  22. [22]

    2025 , address =

    Jain, Chelsi and Wu, Yiran and Zeng, Yifan and Liu, Jiale and Dai, Shengyu and Shao, Zhenwen and Wu, Qingyun and Wang, Huazheng , booktitle =. 2025 , address =

    2025

  23. [23]

    2025 , address =

    Sun, Li and He, Liu and Jia, Shuyue and He, Yangfan and You, Chenyu , booktitle =. 2025 , address =

    2025

  24. [24]

    Schimanski, Tobias and Kolli, Imene and Fan, Yu and Vaghefi, Ario Saeid and Ni, Jingwei and Ash, Elliott and Leippold, Markus , journal =. pdf. 2026 , doi =

    2026

  25. [25]

    2025 , doi =

    Yu, Wenhan and Zhang, Zhaoxi and Chen, Wang and Qi, Guanqiang and Li, Weikang and Sha, Lei and Xia, Deguo and Huang, Jizhou , journal =. 2025 , doi =

    2025

  26. [26]

    2025 , url =

    Wang, Zhaowei and Yu, Wenhao and Ren, Xiyu and Zhang, Jipeng and Zhao, Yu and Saxena, Rohit and Cheng, Liang and Wong, Ginny and See, Simon and Minervini, Pasquale and Song, Yangqiu and Steedman, Mark , journal =. 2025 , url =

    2025

  27. [27]

    2025 , address =

    Deng, Chao and Yuan, Jiale and Bu, Pi and Wang, Peijie and Li, Zhong-Zhi and Xu, Jian and Li, Xiao-Hui and Gao, Yuan and Song, Jun and Zheng, Bo and Liu, Cheng-Lin , booktitle =. 2025 , address =

    2025

  28. [28]

    2025 , address =

    Dong, Kuicai and Chang, Yujing and Goh Xin Deik, Derrick and Li, Dexun and Tang, Ruiming and Liu, Yong , booktitle =. 2025 , address =

    2025

  29. [29]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop) , pages =

    Question Decomposition for Retrieval-Augmented Generation , author =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop) , pages =. 2025 , address =

    2025

  30. [30]

    2025 , address =

    Wu, Xixi and Tan, Yanchao and Hou, Nan and Zhang, Ruiyang and Cheng, Hong , booktitle =. 2025 , address =

    2025

  31. [31]

    2024 , address =

    Zhao, Yilun and Chen, Lyuhao and Cohan, Arman and Zhao, Chen , booktitle =. 2024 , address =

    2024

  32. [32]

    and Manocha, Dinesh , booktitle =

    Suri, Manan and Mathur, Puneet and Dernoncourt, Franck and Goswami, Kanika and Rossi, Ryan A. and Manocha, Dinesh , booktitle =. 2025 , address =

    2025

  33. [33]

    2026 , doi =

    Zhang, Yongyue and Wu, Yaxiong , journal =. 2026 , doi =

    2026

  34. [34]

    2026 , doi =

    Zheng, Yuanlei and Fu, Pei and Li, Hang and Wang, Ziyang and Zhang, Yuyi and Ruan, Wenyu and Zhang, Xiaojin and Wei, Zhongyu and Luo, Zhenbo and Luan, Jian and Chen, Wei and Bai, Xiang , journal =. 2026 , doi =

    2026

  35. [35]

    doi:10.48550/arXiv.2502.13923 , url =

    arXiv preprint arXiv:2502.13923 , year =. doi:10.48550/arXiv.2502.13923 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2502.13923

  36. [36]

    2025 , address =

    Chen, Zhanpeng and Xu, Chengjin and Qi, Yiyan and Jiang, Xuhui and Guo, Jian , booktitle =. 2025 , address =

    2025

  37. [37]

    and Chen, Changyou and Sun, Tong , booktitle =

    Chen, Jian and Zhang, Ruiyi and Zhou, Yufan and Yu, Tong and Dernoncourt, Franck and Gu, Jiuxiang and Rossi, Ryan A. and Chen, Changyou and Sun, Tong , booktitle =. 2025 , url =

    2025

  38. [38]

    2025 , doi =

    Xu, Mingjun and Dong, Jinhan and Hou, Jue and Wang, Zehui and Li, Sihang and Gao, Zhifeng and Zhong, Renxin and Cai, Hengxing , journal =. 2025 , doi =

    2025

  39. [39]

    2026 , doi =

    Xiang, Yixin and Ma, Yunshan and Du, Xiaoyu and Chen, Yibing and Zhang, Yanxin and Tang, Jinhui , journal =. 2026 , doi =

    2026

  40. [40]

    2026 , url =

    Fu, ShunLiang and Zhang, Yanxin and Xiang, Yixin and Du, Xiaoyu and Tang, Jinhui , booktitle =. 2026 , url =

    2026

  41. [41]

    2026 , url =

    Yang, Tianyu and Ruas, Terry and Tian, Yijun and Wahle, Jan Philip and Kurzawe, Daniel and Gipp, Bela , journal =. 2026 , url =

    2026

  42. [42]

    2026 , url =

    Shi, Yaya and Zhang, Jingyun and Liu, Yuliang and Li, Hongliang and Bai, Xiang , journal =. 2026 , url =

    2026

  43. [43]

    2023 , url =

    Tang, Zineng and Yang, Ziyi and Wang, Guoxin and Fang, Yuwei and Liu, Yang and Zhu, Chenguang and Zeng, Michael and Zhang, Cha and Bansal, Mohit , booktitle =. 2023 , url =

    2023

  44. [44]

    , booktitle =

    Appalaraju, Srikar and Jasani, Bhavan and Kota, Bhargava Urala and Xie, Yusheng and Manmatha, R. , booktitle =. 2024 , url =

    2024

  45. [45]

    2023 , url =

    Ye, Jiabo and Hu, Anwen and Xu, Haiyang and Ye, Qinghao and Yan, Ming and Dan, Yaya and Zhao, Chenliang and Xu, Guohai and Li, Chen and Tian, Junfeng and Qian, Qi and Zhang, Ji and Huang, Fei , booktitle =. 2023 , url =

    2023

  46. [46]

    2023 , url =

    Ye, Jiabo and Hu, Anwen and Xu, Haiyang and Ye, Qinghao and Yan, Ming and Xu, Guohai and Li, Chen and Tian, Junfeng and Qian, Qi and Zhang, Ji and Huang, Fei , journal =. 2023 , url =

    2023

  47. [47]

    2024 , url =

    Hu, Anwen and Xu, Haiyang and Ye, Jiabo and Yan, Ming and Zhang, Liang and Zhang, Bo and Li, Chen and Zhang, Ji and Jin, Qin and Huang, Fei , journal =. 2024 , url =

    2024

  48. [48]

    2024 , url =

    Liu, Yuliang and Li, Zhang and Yang, Biao and Li, Chunyuan and Yin, Xu-Cheng and Liu, Cheng-Lin and Jin, Lianwen and Bai, Xiang , journal =. 2024 , url =

    2024

  49. [49]

    2024 , doi =

    Wang, Dongsheng and Raman, Natraj and Sibue, Mathieu and Ma, Zhiqiang and Babkin, Petr and Kaur, Simerjot and Pei, Yulong and Nourbakhsh, Armineh and Liu, Xiaomo , journal =. 2024 , doi =

    2024

  50. [50]

    2024 , url =

    Luo, Chuwei and Shen, Yufan and Zhu, Zhaoqing and Zheng, Qi and Yu, Zhi and Yao, Cong , booktitle =. 2024 , url =

    2024

  51. [51]

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

    A Simple yet Effective Layout Token in Large Language Models for Document Understanding , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

  52. [52]

    2023 , doi =

    Lv, Tengchao and Huang, Yupan and Chen, Jingye and Zhao, Yuzhong and Jia, Yilin and Cui, Lei and Ma, Shuming and Chang, Yaoyao and Huang, Shaohan and Wang, Wenhui and Dong, Li and Luo, Weiyao and Wu, Shaoxiang and Wang, Guoxin and Zhang, Cha and Wei, Furu , journal =. 2023 , doi =

    2023

  53. [53]

    2023 , doi =

    Feng, Hao and Liu, Qi and Liu, Hao and Tang, Jingqun and Zhou, Wengang and Li, Houqiang and Huang, Can , journal =. 2023 , doi =

    2023

  54. [54]

    Wei, Haoran and Liu, Chenglong and Chen, Jinyue and Wang, Jia and Kong, Lingyu and Xu, Yanming and Ge, Zheng and Zhao, Liang and Sun, Jianjian and Peng, Yuang and Han, Chunrui and Zhang, Xiangyu , journal =. General. 2024 , doi =

    2024

  55. [55]

    2024 , doi =

    Mohammadshirazi, Ahmad and Guha Neogi, Pinaki Prasad and Lim, Ser-Nam and Ramnath, Rajiv , journal =. 2024 , doi =

    2024

  56. [56]

    2025 , url =

    Nacson, Mor Shpigel and Aberdam, Aviad and Ganz, Roy and Ben Avraham, Elad and Golts, Alona and Kittenplon, Yair and Mazor, Shai and Litman, Ron , booktitle =. 2025 , url =

    2025

  57. [57]

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

    Marten: Visual Question Answering with Mask Generation for Multi-modal Document Understanding , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

  58. [58]

    arXiv preprint arXiv:2508.03404 , year =

    Visual Document Understanding and Reasoning: A Multi-Agent Collaboration Framework with Agent-Wise Adaptive Test-Time Scaling , author =. arXiv preprint arXiv:2508.03404 , year =. doi:10.48550/arXiv.2508.03404 , url =

    doi:10.48550/arxiv.2508.03404

  59. [59]

    Shao, Zhihong and Wang, Peiyi and Zhu, Qihao and Xu, Runxin and Song, Junxiao and Bi, Xiao and Zhang, Haowei and Zhang, Mingchuan and Li, Y. K. and Wu, Y. and Guo, Daya , journal =. 2024 , doi =

    2024

  60. [60]

    2025 , doi =

    Yang, An and Li, Anfeng and Yang, Baosong and Zhang, Beichen and Hui, Binyuan and Zheng, Bo and Yu, Bowen and Gao, Chang and Huang, Chengen and Lv, Chenxu and others , journal =. 2025 , doi =

    2025

  61. [61]

    doi:10.48550/arXiv.2511.21631 , url =

    arXiv preprint arXiv:2511.21631 , year =. doi:10.48550/arXiv.2511.21631 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2511.21631

  62. [62]

    2026 , howpublished =

    2026

  63. [63]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    Encoding spreadsheets for large language models , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    2024

  64. [64]

    Proceedings of the 3rd Workshop on Advances in Language and Vision Research (ALVR) , pages=

    Vision language models for spreadsheet understanding: Challenges and opportunities , author=. Proceedings of the 3rd Workshop on Advances in Language and Vision Research (ALVR) , pages=

  65. [65]

    arXiv preprint arXiv:2605.12289 , year=

    PriorZero: Bridging Language Priors and World Models for Decision Making , author=. arXiv preprint arXiv:2605.12289 , year=

    Pith/arXiv arXiv

  66. [66]

    IEEE Transactions on Circuits and Systems for Video Technology , volume=

    Detect any shadow: Segment anything for video shadow detection , author=. IEEE Transactions on Circuits and Systems for Video Technology , volume=. 2023 , publisher=

    2023

  67. [67]

    arXiv preprint arXiv:2311.13194 , year=

    Towards improving document understanding: An exploration on text-grounding via mllms , author=. arXiv preprint arXiv:2311.13194 , year=

    Pith/arXiv arXiv

  68. [68]

    Proceedings of the 30th ACM International Conference on Multimedia , pages=

    Udoc-gan: Unpaired document illumination correction with background light prior , author=. Proceedings of the 30th ACM International Conference on Multimedia , pages=

  69. [69]

    Computer Vision and Image Understanding , volume=

    Progressive recurrent network for shadow removal , author=. Computer Vision and Image Understanding , volume=. 2024 , publisher=

    2024

  70. [70]

    arXiv preprint arXiv:2411.15714 , year=

    Root: Vlm based system for indoor scene understanding and beyond , author=. arXiv preprint arXiv:2411.15714 , year=

    Pith/arXiv arXiv

  71. [71]

    ACM Transactions on Multimedia Computing, Communications and Applications , volume=

    Swinshadow: Shifted window for ambiguous adjacent shadow detection , author=. ACM Transactions on Multimedia Computing, Communications and Applications , volume=. 2024 , publisher=

    2024