Pith. sign in

REVIEW 4 major objections 5 minor 71 references

HierDoc proposes that treating page selection and region selection as two successive, independently trained set-prediction policies improves long-document VQA, reporting best open-weight scores on four of five benchmarks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 02:48 UTC pith:V37PI4QY

load-bearing objection HierDoc is a credible two-stage evidence-routing system; the central claim that trained region selection beats random crops is plausible but not yet statistically nailed down. the 4 major comments →

arxiv 2607.29638 v1 pith:V37PI4QY submitted 2026-07-31 cs.CV

HierDoc: Hierarchical Page-to-Region Evidence Routing for Long-Document Visual Question Answering

Rongjian Gu , Wengang Zhou , Junyu Xiong , Yonghui Wang , Bing Yin , Bei Wang , Houqiang Li This is my paper
classification cs.CV
keywords long-document visual question answeringhierarchical evidence routingpage selectionregion selectionGRPOstructured-set predictionmultimodal document understandingparser-native action space
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that long-document visual question answering improves when evidence acquisition is factored into two explicit, answer-agnostic stages: first choose which pages contain evidence, then choose which semantic regions within those pages do. Prior systems either retrieve whole pages without fine-grained region selection or assume the relevant pages are already given, leaving page and region decisions disconnected. HierDoc trains each stage separately with a reinforcement-learning objective (GRPO) and a structured-set reward, then feeds the selected full pages, region crops, and OCR text to an answer model. Reported results include top open-weight accuracy on four of five benchmarks, with a 16.87% relative gain on LongDocURL, and ablations show the region stage adds 5.51% accuracy over page-only answering. A sympathetic reader would take the paper's core claim to be that the page-to-region factorization itself, not merely a stronger answer backbone, is what drives the gains.

Core claim

The central claim is that decomposing long-document evidence acquisition into two successive structured-set policies—a page policy that returns a set of evidence pages and a region policy that returns a set of parser-native semantic regions drawn from those pages—yields better end-to-end VQA than page-centric or region-centric alternatives. The two policies are optimized independently with GRPO, using granularity-specific set rewards: the page reward favors recall and compactness, the region reward balances precision and recall. The paper demonstrates that HierDoc with a Qwen2.5-VL answer backbone already outperforms all compared methods on MMLongBench-Doc and LongDocURL, and that the full r

What carries the argument

The central mechanism is hierarchical evidence routing: a windowed page policy trained with GRPO and a recall-sensitive structured-set reward selects a compact set of pages; a layout parser converts those pages into typed semantic regions (paragraphs, tables, figures, captions) with bounding boxes and OCR/table text; a second GRPO-trained region policy selects a subset of these parser-native aliases using a precision-inclusive reward; and a grounded answering stage fuses the selected full pages, region crops, and OCR. The parser-native action space and stage-wise structured-set rewards are what carry the argument: they turn evidence acquisition into discrete set decisions that can be indepen

Load-bearing premise

The whole pipeline assumes the page policy never skips a page that contains answer evidence, and that the layout parser segments that evidence into candidate regions; a miss at either step cannot be recovered later.

What would settle it

Run HierDoc on a long-document benchmark while artificially withholding a random 10% of gold evidence pages from the page policy's candidate windows (or corrupting the parser to drop 10% of gold evidence regions), then measure the drop in end-to-end QA accuracy: if the drop is small, the routing policies are not actually carrying the reported gains, but if it is large, the system's success is contingent on the very recall assumptions the paper acknowledges.

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

If this is right

  • Long-document VQA systems can be built modularly: page routing and region routing can be improved, swapped, or debugged independently without retraining the answer model.
  • The two-stage evidence-set formulation transfers across answer backbones, as shown by the Qwen2.5-VL variant, so the gains are not tied to a single model's strength.
  • Because page routing is recall-oriented and region routing is precision-oriented, practitioners can tune each stage's reward separately to manage the trade-off between evidence coverage and context volume.
  • Bounded reflection at both stages produces more compact page and region sets, which should lower inference cost while preserving or improving downstream accuracy.
  • Page-level retrieval quality becomes directly measurable and comparable to existing retrievers, making evidence routing more auditable in deployed systems.

Where Pith is reading between the lines

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

  • An unstated consequence is that the two-stage page-to-region factorization could serve as a general recipe for retrieval-augmented generation over any multi-part visual or textual document—slide decks, video frames, or multi-chapter reports—by learning a coarse selector and a fine selector with set rewards at each granularity.
  • Because the region policy acts on parser-native regions, the system's ceiling is partly set by the parser's recall; improving layout parsing or adding a fallback that re-submits omitted pages could lift performance even if the routing policies stay unchanged.
  • The reward asymmetry (page recall-weighted, region precision-weighted) encodes an assumption about error cost that could be tested further: on documents where evidence is highly concentrated, an adaptive threshold for page selection might outperform the fixed reflection thresholds.
  • A natural extension the paper leaves implicit is to feed the answer model's outcome (e.g., 'Not answerable') back into the routers, creating a closed loop that recovers from page-omission errors—something the paper's limitations section explicitly identifies as future work.

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

4 major / 5 minor

Summary. HierDoc proposes a hierarchical evidence-routing framework for long-document VQA in which page selection and region selection are cast as two successive, answer-agnostic structured-set policies, each trained independently with GRPO and granularity-specific rewards. A windowed page selector first retrieves evidence pages; MinerU2.5 parses these pages into semantic regions; a region selector then chooses a subset of parser-native regions; and a downstream answer model receives the selected full pages, region crops, and OCR/table text. The paper reports best or competitive open-weight results on five benchmarks (e.g., 53.62 accuracy on MMLongBench-Doc, 65.80 on LongDocURL), with a Qwen2.5-VL variant used to show that gains are not solely from the stronger Qwen3-VL answer model. Controlled evidence-composition ablations report that trained selected regions improve page-only answering by 5.51% accuracy and 4.82% F1, exceeding matched random-region controls. The conclusion explicitly acknowledges irreversible page-to-region error propagation and dependence on parser quality and gold annotations.

Significance. If the empirical claims hold, the paper makes a useful design contribution: it treats page and region acquisition as separate, optimizable evidence-set decisions with a parser-native discrete action space, rather than as navigation aids or post-retrieval operations. The strengths are real: the page-retrieval evaluation in Table 2 directly measures the evidence pool available to later stages; the random-region controls in Figure 2 are a sound way to separate learned selection from mere input augmentation; the supplement provides dataset accounting, complete prompts, and reward details; and the Qwen2.5-VL variant supports backbone transfer. The main weakness is statistical: the central trained-vs-random and incremental-ablation comparisons rest on single-run points, and the SOTA claim omits several directly relevant open-weight baselines discussed in the paper itself. With multi-seed variance estimates and either additional baseline comparisons or a narrowed claim, the contribution would be solid and publishable.

major comments (4)
  1. [Figure 2; Section "Effects of Evidence Composition"] The caption states that random-region controls report mean and sample standard deviation over four seeds, while "all other points are single runs." The text's key conclusion — that gains arise primarily from question-conditioned region routing rather than from adding more inputs — rests on the trained-vs-random differences (3.18% accuracy / 4.46% F1 for selected crops+OCR). Since the trained side has no variance estimate or significance test, and the random-control standard deviations are explicitly shown, the claimed advantage may be within run-to-run noise. Please report multiple seeds for the trained configuration (or a paired bootstrap over questions) with confidence intervals and a significance test.
  2. [Table 3; "Effects of Selector Training and Reflection"] The incremental ablation is the main evidence that GRPO training and bounded reflection improve both selector quality and downstream QA, but every number in the table appears to be a single run. For example, the page block attributes a 5.26% accuracy gain to GRPO training, and the region block shows Region F1 unchanged at 46.9 while QA accuracy changes by 1.54%; without error bars or a significance test these changes are hard to separate from noise. Provide multiple seeds, or at least a paired test over questions, for the key rows that support the factorization claim.
  3. [Table 1; "Compared Methods"] The Related Work names DocLens, HIEVI-RAG, and VRAG-RL as directly relevant open-weight hierarchical/coarse-to-fine document systems, but none appears in Table 1 or in the "Compared Methods" list. The abstract and main text claim "state-of-the-art or competitive performance among open-weight systems" and a 16.87% relative improvement over the "strongest reported open-weight baseline" on LongDocURL. Since those omitted systems are open-weight and designed for the same task, the SOTA claim is not currently established. Either add these baselines (directly or with clearly matched reported scores) or narrow the claim to "among the evaluated baselines."
  4. [Supplementary "Semantic Evaluation for PaperTab and FetaTab"] PaperTab and FetaTab results rely on LLM-as-a-judge scores. HierDoc uses GPT-5.5 with a rubric, while Table 1 imports reported scores from prior work. If the prior systems were judged with different judge models, prompts, or answer-format instructions, the relative gains (e.g., 37.06% on PaperTab) may reflect judge mismatch rather than method quality. Report judge agreement/human validation, or re-score the baselines with the same judge protocol.
minor comments (5)
  1. [Figure 2] Add error bars or variance shading for the trained-selection points; the caption currently highlights that they are single runs, which is precisely the configuration whose effect size matters most.
  2. [Table 3] Clarify what "Selector F1" measures in the region block. It is surprising that GRPO training leaves Region F1 at 46.9 while QA accuracy improves; if this is due to the metric being defined on a different candidate set, say so explicitly.
  3. [Implementation Details] The page-window capacity Mp=16 and reflection thresholds τp=τr=8 are fixed without sensitivity analysis. A short sensitivity study or a justification for these values would strengthen the paper.
  4. [Method; Figure 1] The parser is called "MinerU2.5" in the method text and "MinerU2.5-Pro" in Figure 1 and the abstract. Unify the nomenclature.
  5. [Supplementary Material] No code or checkpoint release is mentioned. Given the complexity of the pipeline, releasing code and trained selector checkpoints would materially aid reproducibility.

Circularity Check

0 steps flagged

No significant circularity: routing policies are trained on external gold sets and evaluated on held-out benchmarks; self-citations are not load-bearing.

full rationale

The paper's derivation chain is not circular. Page and region selectors are trained with GRPO against external gold page/region annotations (MMDocIR subset and ViDoRe-v3 corpus) using explicit rule-based set rewards, and end-to-end accuracy is measured on separate held-out benchmarks (MMLongBench-Doc, LongDocURL, SlideVQA, PaperTab, FetaTab). The claim that region routing contributes is supported by a controlled ablation (Figure 2) that compares trained selection to four-seed random-region controls on the same answer model and page set; the random-control advantage (2.26%/0.35%) is weaker than trained selection (5.51%/4.82%), so the gain is not an artifact of merely adding inputs. The Qwen2.5-VL variant and page-retrieval evaluation further separate the routing contribution from the stronger Qwen3-VL answer backbone. Self-citations (DocR1, HIEVI-RAG) are related-work precedents for GRPO-style evidence selection and are not used as load-bearing justification for HierDoc's claims. The paper explicitly acknowledges parser dependence and irreversible page-routing error propagation in the Conclusion, which is consistent with an empirical pipeline rather than a result forced by definition. Caveats such as possible reward-balance selection on MMLongBench-Doc (Supplementary Table 7) and single-run ablation points (Figure 2 caption) are statistical/validity concerns, not circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The method's contribution is empirical. Its central claim rests on: (i) gold evidence annotations being complete and correct; (ii) MinerU parsing producing a sufficiently complete candidate region set; (iii) LLM-as-a-judge scores being faithful and comparable across papers; and (iv) several hand-chosen hyperparameters (reward weights, window capacity, reflection thresholds, box-mapping thresholds, GRPO settings). No new theoretical entities are introduced.

free parameters (5)
  • Page-policy reward weights (λp_rec, λp_f1, λp_fmt) = (0.40, 0.45, 0.15)
    Hand-chosen; Table 7 ablates the recall/F1 balance and selects 'Balanced' based on downstream QA accuracy, indicating these values are fitted to performance rather than derived.
  • Region-policy reward weights (λr_rec, λr_f1, λr_prec, λr_fmt) = (0.20, 0.50, 0.20, 0.10)
    Selected by authors; no sensitivity analysis for these values is reported.
  • Page-window capacity M_p = 16
    Inference hyperparameter; no ablation or justification beyond 'fixed-size windows'.
  • Reflection thresholds τp, τr = 8
    Set to 8 at inference; no sensitivity analysis reported.
  • Gold-box to parser-region matching thresholds = IoU ≥ 0.10 or containment ≥ 0.80
    Hand-chosen mapping from evidence boxes to parser aliases; changes in these thresholds would change the region-policy training target and Region F1.
axioms (6)
  • domain assumption Benchmark gold pages G_p and gold evidence boxes are complete and correct annotations for each question.
    Page and region rewards are computed against these sets (Sec. Page Routing / Region Routing); if annotations are incomplete, the learned selectors are trained to incomplete targets.
  • domain assumption MinerU2.5-Pro's parsed candidate regions C(P̂) contain all semantic units needed to answer.
    Region policy can only select from parser-native candidates; the authors list parser dependence as a limitation in the Conclusion. This is load-bearing because downstream answering cannot recover a missing region.
  • domain assumption LLM-as-a-judge (GPT-5.5) with a binary rubric gives a faithful correctness signal for PaperTab/FetaTab free-form answers.
    Used for two of five benchmarks; no human agreement or judge-version sensitivity is reported in the paper.
  • standard math GRPO set-reward optimization (DeepSeekMath formulation) is applicable as specified to vision-language selectors.
    Equations (1)-(2) in the supplement adopt the standard GRPO objective from Shao et al.; the paper adds no new RL theory.
  • domain assumption Prior-paper LLM-judge scores for PaperTab/FetaTab are directly comparable despite potential differences in judge model, prompt, and rubric.
    The paper says 'we use their reported scores directly' (Supplementary Material), so comparability is assumed rather than demonstrated.
  • domain assumption Training on MMDocIR and ViDoRe-v3 subsets transfers to the five evaluated benchmarks.
    Page/region policies are trained on these corpora and evaluated on different datasets; no overlap analysis is provided.

pith-pipeline@v1.3.0-daily-deepseek · 16170 in / 14223 out tokens · 126727 ms · 2026-08-03T02:48:15.188155+00:00 · methodology

0 comments
read the original abstract

Multi-page document visual question answering requires locating sparse evidence at both the page and region levels. Existing approaches typically emphasize one level over the other: page-centric methods focus on page acquisition, with region operations serving mainly as navigation aids, whereas region-centric methods assume that the relevant pages have already been supplied. Consequently, page and region selection remain disconnected rather than forming successive evidence decisions. We propose HierDoc, a hierarchical evidence-routing framework that formulates long-document evidence acquisition as two-stage set prediction from pages to regions. A page policy selects evidence pages from the full document; these pages are then parsed for semantic elements, after which a region policy selects the elements passed to a downstream answer model. Both answer-agnostic policies are optimized with stage-wise GRPO using granularity-specific structured-set rewards. The answer model receives selected full pages together with selected region crops and OCR or table text, preserving global context while emphasizing fine-grained evidence. Across the evaluated benchmarks, HierDoc achieves state-of-the-art or competitive performance among open-weight systems, improving LongDocURL by 16.87% relative to the strongest reported open-weight baseline. Controlled ablations further show that selected regional evidence improves the page-only system in accuracy and F1 by 5.51% and 4.82%, respectively. These results demonstrate the benefit of organizing coarse page routing and fine-grained region routing as successive, separately optimized stages of a unified evidence-acquisition process.

Figures

Figures reproduced from arXiv: 2607.29638 by Bei Wang, Bing Yin, Houqiang Li, Junyu Xiong, Rongjian Gu, Wengang Zhou, Yonghui Wang.

Figure 1
Figure 1. Figure 1: Overview of HierDoc inference. A trainable page policy selects evidence pages from fixed-size windows and routes [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Evidence-composition ablation for the final answer model. Random-region controls report the mean and sample [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Single-page evidence routing on MMLongBench-Doc. For an AP [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Cross-page compositional routing on MMLongBench-Doc. From a 43-page document, the page selector retains [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

71 extracted references · 5 canonical work pages

  1. [3]

    A.; Chen, C.; and Sun, T

    Chen, J.; Zhang, R.; Zhou, Y.; Yu, T.; Dernoncourt, F.; Gu, J.; Rossi, R. A.; Chen, C.; and Sun, T. 2025. SV-RAG : LoRA -Contextualizing Adaptation of MLLMs for Long Document Understanding. In The Thirteenth International Conference on Learning Representations

    2025

  2. [5]

    Deng, C.; Yuan, J.; Bu, P.; Wang, P.; Li, Z.-Z.; Xu, J.; Li, X.-H.; Gao, Y.; Song, J.; Zheng, B.; and Liu, C.-L. 2025. LongDocURL : A Comprehensive Multimodal Long Document Benchmark Integrating Understanding, Reasoning, and Locating. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 1135--1159

    2025

  3. [6]

    Dong, K.; Chang, Y.; Goh Xin Deik, D.; Li, D.; Tang, R.; and Liu, Y. 2025. MMDocIR : Benchmarking Multimodal Retrieval for Long Documents. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, 30971--31005

    2025

  4. [7]

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

    2025

  5. [8]

    GLM-V Team . 2025. GLM-4.5V and GLM-4.1V-Thinking : Towards Versatile Multimodal Reasoning with Scalable Reinforcement Learning. arXiv:2507.01006

    Pith/arXiv arXiv 2025

  6. [10]

    Hu, A.; Xu, H.; Zhang, L.; Ye, J.; Yan, M.; Zhang, J.; Jin, Q.; Huang, F.; and Zhou, J. 2025. mPLUG-DocOwl2 : High-resolution Compressing for OCR -free Multi-page Document Understanding. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 5817--5834

    2025

  7. [11]

    Hui, Y.; Lu, Y.; and Zhang, H. 2024. UDA : A Benchmark Suite for Retrieval Augmented Generation in Real-World Document Analysis. In Advances in Neural Information Processing Systems, volume 37, 67200--67217

    2024

  8. [12]

    Li, Y.; Lu, Z.; Liu, Z.; Sun, Y.; Liu, C.; and Xie, H. 2026. RegionRAG : Region-level Retrieval-Augmented Generation for Visual Document Understanding. Proceedings of the AAAI Conference on Artificial Intelligence, 40(8): 6662--6670

    2026

  9. [13]

    Loison, A.; Mac \'e , Q.; Edy, A.; Xing, V.; Balough, T.; Moreira, G.; Liu, B.; Faysse, M.; Hudelot, C.; and Viaud, G. 2026. ViDoRe V3 : A Comprehensive Evaluation of Retrieval Augmented Generation in Complex Real-World Scenarios. arXiv:2601.08620

    Pith/arXiv arXiv 2026

  10. [14]

    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.; and Sun, A. 2024. MMLongBench-Doc : Benchmarking Long-context Document Understanding with Visualizations. In Advances in Neural Information Processing Systems, volume 37, 95963--96010

    2024

  11. [15]

    X.; Tan, J

    Masry, A.; Long, D. X.; Tan, J. Q.; Joty, S.; and Hoque, E. 2022. ChartQA : A Benchmark for Question Answering about Charts with Visual and Logical Reasoning. In Findings of the Association for Computational Linguistics: ACL 2022, 2263--2279

    2022

  12. [16]

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

    2022

  13. [17]

    Mathew, M.; Karatzas, D.; and Jawahar, C. 2021. DocVQA: A Dataset for VQA on Document Images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2200--2209

    2021

  14. [18]

    Niu, J.; Liu, Z.; Gu, Z.; Wang, B.; Ouyang, L.; Zhao, Z.; Chu, T.; He, T.; Wu, F.; Zhang, Q.; Jin, Z.; Liang, G.; Zhang, R.; Zhang, W.; Qu, Y.; Ren, Z.; Sun, Y.; Tang, Z.; Niu, B.; Zheng, Y.; Ma, D.; Miao, Z.; Dong, H.; Qian, S.; Zhang, J.; Wang, F.; Chen, J.; Zhao, X.; Wei, L.; Li, W.; Wang, S.; Xu, R.; Cao, Y.; Chen, L.; Wu, Q.; Gu, H.; Lu, L.; Lin, D.;...

    2026

  15. [20]

    Tanaka, R.; Nishida, K.; Nishida, K.; Hasegawa, T.; Saito, I.; and Saito, K. 2023. SlideVQA : A Dataset for Document Visual Question Answering on Multiple Images. Proceedings of the AAAI Conference on Artificial Intelligence, 37(11): 13636--13645

    2023

  16. [21]

    Tito, R.; Karatzas, D.; and Valveny, E. 2023. Hierarchical Multimodal Transformers for Multi-Page DocVQA. Pattern Recognition, 144: 109834

    2023

  17. [22]

    Van Landeghem, J.; Tito, R.; Borchmann, .; Pietruszka, M.; Joziak, P.; Powalski, R.; Jurkiewicz, D.; Coustaty, M.; Anckaert, B.; Valveny, E.; Blaschko, M.; Moens, S.; and Stanislawek, T. 2023. Document Understanding Dataset and Evaluation (DUDE). In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 19528--19540

    2023

  18. [23]

    Wang, Q.; Ding, R.; Chen, Z.; Wu, W.; Wang, S.; Xie, P.; and Zhao, F. 2025 a . ViDoRAG : Visual Document Retrieval-Augmented Generation via Dynamic Iterative Reasoning Agents. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, 9113--9134

    2025

  19. [28]

    Wu, X.; Tan, Y.; Hou, N.; Zhang, R.; and Cheng, H. 2025. MoLoRAG : Bootstrapping Document Understanding via Multi-modal Logic-aware Retrieval. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, 14024--14045

    2025

  20. [29]

    Xiong, J.; Wang, Y.; Gu, R.; Liu, C.; Yin, B.; Zhou, W.; and Li, H. 2026 a . Hierarchical Evidence-Driven Reasoning for Long Document Understanding. arXiv:2607.04625

    Pith/arXiv arXiv 2026

  21. [30]

    Xiong, J.; Wang, Y.; Zhao, W.; Liu, C.; Yin, B.; Zhou, W.; and Li, H. 2026 b . DocR1 : Evidence Page-Guided GRPO for Multi-Page Document Understanding. Proceedings of the AAAI Conference on Artificial Intelligence, 40(13): 11178--11186

    2026

  22. [32]

    Yu, S.; Tang, C.; Xu, B.; Cui, J.; Ran, J.; Yan, Y.; Liu, Z.; Wang, S.; Han, X.; Liu, Z.; and Sun, M. 2025. VisRAG : Vision-based Retrieval-augmented Generation on Multi-modality Documents. In The Thirteenth International Conference on Learning Representations

    2025

  23. [33]

    P.; Zhang, H.; Gonzalez, J

    Zheng, L.; Chiang, W.-L.; Sheng, Y.; Zhuang, S.; Wu, Z.; Zhuang, Y.; Lin, Z.; Li, Z.; Li, D.; Xing, E. P.; Zhang, H.; Gonzalez, J. E.; and Stoica, I. 2023. Judging LLM -as-a-Judge with MT-Bench and Chatbot Arena . In Advances in Neural Information Processing Systems, volume 36, 46595--46623

    2023

  24. [34]

    Zheng, Y.; Fu, P.; Li, H.; Wang, Z.; Zhang, Y.; Ruan, W.; Zhang, X.; Wei, Z.; Luo, Z.; Luan, J.; Chen, W.; and Bai, X. 2026. Doc- V^* : Coarse-to-Fine Interactive Visual Reasoning for Multi-Page Document VQA. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 45901--45923

    2026

  25. [35]

    Zhu, D.; Meng, R.; Chen, J.; Li, S.; Pfister, T.; and Yoon, J. 2026. DocLens : A Tool-Augmented Multi-Agent Framework for Long Visual Document Understanding. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 26804--26829

    2026

  26. [36]

    Zhu, F.; Lei, W.; Feng, F.; Wang, C.; Zhang, H.; and Chua, T.-S. 2022. Towards Complex Document Understanding By Discrete Reasoning. In Proceedings of the 30th ACM International Conference on Multimedia, 4857--4866

    2022

  27. [38]

    , title =

    Mathew, Minesh and Karatzas, Dimosthenis and Jawahar, C.V. , title =. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , month = jan, year =

  28. [39]

    , title =

    Mathew, Minesh and Bagal, Viraj and Tito, Rub\`en and Karatzas, Dimosthenis and Valveny, Ernest and Jawahar, C.V. , title =. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , month = jan, year =. doi:10.1109/WACV51458.2022.00264 , url =

    arXiv 2022

  29. [40]

    2022 , pages =

    Masry, Ahmed and Long, Do Xuan and Tan, Jia Qing and Joty, Shafiq and Hoque, Enamul , booktitle =. 2022 , pages =. doi:10.18653/v1/2022.findings-acl.177 , url =

    doi:10.18653/v1/2022.findings-acl.177 2022

  30. [41]

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

    Towards Complex Document Understanding By Discrete Reasoning , author =. Proceedings of the 30th ACM International Conference on Multimedia , year =. doi:10.1145/3503161.3548422 , url =

    arXiv

  31. [42]

    Pattern Recognition , volume =

    Tito, Rub\`en and Karatzas, Dimosthenis and Valveny, Ernest , title =. Pattern Recognition , volume =. 2023 , doi =

    2023

  32. [43]

    Document Understanding Dataset and Evaluation (DUDE) , booktitle =

    Van Landeghem, Jordy and Tito, Rub\`en and Borchmann,. Document Understanding Dataset and Evaluation (DUDE) , booktitle =. 2023 , pages =

    2023

  33. [44]

    2024 , doi =

    Ma, Yubo and Zang, Yuhang and Chen, Liangyu and Chen, Meiqi and Jiao, Yizhu and Li, Xinze and Lu, Xinyuan and Liu, Ziyu and Ma, Yan and Dong, Xiaoyi and Zhang, Pan and Pan, Liangming and Jiang, Yu-Gang and Wang, Jiaqi and Cao, Yixin and Sun, Aixin , booktitle =. 2024 , doi =

    2024

  34. [45]

    2024 , doi =

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

    2024

  35. [46]

    2025 , pages =

    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 , pages =. doi:10.18653/v1/2025.acl-long.57 , url =

    doi:10.18653/v1/2025.acl-long.57 2025

  36. [47]

    2025 , pages =

    Dong, Kuicai and Chang, Yujing and Goh Xin Deik, Derrick and Li, Dexun and Tang, Ruiming and Liu, Yong , booktitle =. 2025 , pages =. doi:10.18653/v1/2025.emnlp-main.1576 , url =

    doi:10.18653/v1/2025.emnlp-main.1576 2025

  37. [48]

    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 =

  38. [49]

    2025 , url =

    Yu, Shi and Tang, Chaoyue and Xu, Bokai and Cui, Junbo and Ran, Junhao and Yan, Yukun and Liu, Zhenghao and Wang, Shuo and Han, Xu and Liu, Zhiyuan and Sun, Maosong , booktitle =. 2025 , url =

    2025

  39. [50]

    2026 , doi =

    Li, Yinglu and Lu, Zhiying and Liu, Zhihang and Sun, Yiwei and Liu, Chuanbin and Xie, Hongtao , journal =. 2026 , doi =

    2026

  40. [51]

    2025 , pages =

    Wang, Qiuchen and Ding, Ruixue and Chen, Zehui and Wu, Weiqi and Wang, Shihang and Xie, Pengjun and Zhao, Feng , booktitle =. 2025 , pages =. doi:10.18653/v1/2025.emnlp-main.464 , url =

    work page doi:10.18653/v1/2025.emnlp-main.464 2025

  41. [52]

    and Sheng, Tao and Ravi, Sujith and Dehghani, Morteza and Roth, Dan , booktitle =

    Sourati, Zhivar and Wang, Zheng and Liu, Marianne Menglin and Hu, Yazhe and Guo, Mengqing and Bharadwaj, Sujeeth and Han, Kyu J. and Sheng, Tao and Ravi, Sujith and Dehghani, Morteza and Roth, Dan , booktitle =. 2026 , pages =. doi:10.18653/v1/2026.acl-long.724 , url =

    work page doi:10.18653/v1/2026.acl-long.724 2026

  42. [53]

    Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =

    Doc- V^* : Coarse-to-Fine Interactive Visual Reasoning for Multi-Page Document VQA , author =. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =. doi:10.18653/v1/2026.acl-long.2129 , url =

    doi:10.18653/v1/2026.acl-long.2129 2026

  43. [54]

    2503.13964 , archivePrefix =

    Han, Siwei and Xia, Peng and Zhang, Ruiyi and Sun, Tong and Li, Yun and Zhu, Hongtu and Yao, Huaxiu , year =. 2503.13964 , archivePrefix =

    Pith/arXiv arXiv

  44. [55]

    2023 , url =

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

    2023

  45. [56]

    2026 , pages =

    Niu, Junbo and Liu, Zheng and Gu, Zhuangcheng and Wang, Bin and Ouyang, Linke and Zhao, Zhiyuan and Chu, Tao and He, Tianyao and Wu, Fan and Zhang, Qintong and Jin, Zhenjiang and Liang, Guang and Zhang, Rui and Zhang, Wenzheng and Qu, Yuan and Ren, Zhifei and Sun, Yuefeng and Tang, Zirui and Niu, Boyu and Zheng, Yuanhong and Ma, Dongsheng and Miao, Ziyang...

    2026

  46. [57]

    2511.21631 , archivePrefix =

    Bai, Shuai and Cai, Yuxuan and Chen, Ruizhe and Chen, Keqin and Chen, Xionghui and Cheng, Zesen and Deng, Lianghao and Ding, Wei and Gao, Chang and Ge, Chunjiang and Ge, Wenbin and Guo, Zhifang and Huang, Qidong and Huang, Jie and Huang, Fei and Hui, Binyuan and Jiang, Shutong and Li, Zhaohai and Li, Mingsheng and Li, Mei and Li, Kaixin and Lin, Zicheng a...

    Pith/arXiv arXiv

  47. [58]

    2502.13923 , archivePrefix =

    Bai, Shuai and Chen, Keqin and Liu, Xuejing and Wang, Jialin and Ge, Wenbin and Song, Sibo and Dang, Kai and Wang, Peng and Wang, Shijie and Tang, Jun and Zhong, Humen and Zhu, Yuanzhi and Yang, Mingkun and Li, Zhaohai and Wan, Jianqiang and Wang, Pengfei and Ding, Wei and Fu, Zheren and Xu, Yiheng and Ye, Jiabo and Zhang, Xi and Xie, Tianbao and Cheng, Z...

    Pith/arXiv arXiv

  48. [59]

    2508.18265 , archivePrefix =

    Wang, Weiyun and Gao, Zhangwei and Gu, Lixin and Pu, Hengjun and Cui, Long and Wei, Xingguang and Liu, Zhaoyang and Jing, Linglin and Ye, Shenglong and Shao, Jie and Wang, Zhaokai and Chen, Zhe and Zhang, Hongjie and Yang, Ganlin and Wang, Haomin and Wei, Qi and Yin, Jinhui and Li, Wenhao and Cui, Erfei and Chen, Guanzhou and Ding, Zichen and Tian, Changy...

    Pith/arXiv arXiv

  49. [60]

    2025 , pages =

    Hu, Anwen and Xu, Haiyang and Zhang, Liang and Ye, Jiabo and Yan, Ming and Zhang, Ji and Jin, Qin and Huang, Fei and Zhou, Jingren , booktitle =. 2025 , pages =. doi:10.18653/v1/2025.acl-long.291 , url =

    doi:10.18653/v1/2025.acl-long.291 2025

  50. [61]

    2510.18234 , archivePrefix =

    Wei, Haoran and Sun, Yaofeng and Li, Yukun , year =. 2510.18234 , archivePrefix =

    Pith/arXiv arXiv

  51. [62]

    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 , year =. 2402.03300 , archivePrefix =

    Pith/arXiv arXiv

  52. [63]

    2026 , doi =

    Xiong, Junyu and Wang, Yonghui and Zhao, Weichao and Liu, Chenyu and Yin, Bing and Zhou, Wengang and Li, Houqiang , journal =. 2026 , doi =

    2026

  53. [64]

    2026 , pages =

    Zhu, Dawei and Meng, Rui and Chen, Jiefeng and Li, Sujian and Pfister, Tomas and Yoon, Jinsung , booktitle =. 2026 , pages =. doi:10.18653/v1/2026.acl-long.1234 , url =

    work page doi:10.18653/v1/2026.acl-long.1234 2026

  54. [65]

    2602.24134 , archivePrefix =

    Wang, Zhengren and Ma, Dongsheng and Zhong, Huaping and Li, Jiayu and Zhang, Wentao and Wang, Bin and He, Conghui , year =. 2602.24134 , archivePrefix =

    arXiv

  55. [66]

    2026 , pages =

    Lin, Jiahang and Hu, Kai and Wang, Binghai and Zhou, Yuhao and Xi, Zhiheng and Guo, Honglin and Liu, Shichun and Wang, Junzhe and Dou, Shihan and Zhou, Enyu and Yan, Hang and Han, Zhenhua and Gui, Tao and Zhang, Qi and Huang, Xuanjing , booktitle =. 2026 , pages =. doi:10.18653/v1/2026.findings-acl.1488 , url =

    work page doi:10.18653/v1/2026.findings-acl.1488 2026

  56. [67]

    Look as You Think: Unifying Reasoning and Visual Evidence Attribution for Verifiable Document

    Liu, Shuochen and Luo, Pengfei and Zhang, Chao and Chen, Yuhao and Zhang, Haotian and Liu, Qi and Kou, Xin and Xu, Tong and Chen, Enhong , journal =. Look as You Think: Unifying Reasoning and Visual Evidence Attribution for Verifiable Document. 2026 , doi =

    2026

  57. [68]

    2023 , doi =

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

    2023

  58. [69]

    2020 , pages =

    Khattab, Omar and Zaharia, Matei , booktitle =. 2020 , pages =. doi:10.1145/3397271.3401075 , url =

    arXiv 2020

  59. [70]

    2024 , pages =

    Chen, Jianlyu and Xiao, Shitao and Zhang, Peitian and Luo, Kun and Lian, Defu and Liu, Zheng , booktitle =. 2024 , pages =. doi:10.18653/v1/2024.findings-acl.137 , url =

    doi:10.18653/v1/2024.findings-acl.137 2024

  60. [71]

    2411.04952 , archivePrefix =

    Cho, Jaemin and Mahata, Debanjan and Irsoy, Ozan and He, Yujie and Bansal, Mohit , year =. 2411.04952 , archivePrefix =

    Pith/arXiv arXiv

  61. [72]

    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

  62. [73]

    2025 , pages =

    Wu, Xixi and Tan, Yanchao and Hou, Nan and Zhang, Ruiyang and Cheng, Hong , booktitle =. 2025 , pages =. doi:10.18653/v1/2025.emnlp-main.708 , url =

    work page doi:10.18653/v1/2025.emnlp-main.708 2025

  63. [74]

    Advances in Neural Information Processing Systems , volume =

    Chain of Agents: Large Language Models Collaborating on Long-Context Tasks , author =. Advances in Neural Information Processing Systems , volume =. 2024 , doi =

    2024

  64. [75]

    2025 , eprint =

    Visual Document Understanding and Reasoning: A Multi-Agent Collaboration Framework with Agent-Wise Adaptive Test-Time Scaling , author =. 2025 , eprint =

    2025

  65. [76]

    2505.22019 , archivePrefix =

    Wang, Qiuchen and Ding, Ruixue and Zeng, Yu and Chen, Zehui and Chen, Lin and Wang, Shihang and Xie, Pengjun and Huang, Fei and Zhao, Feng , year =. 2505.22019 , archivePrefix =

    Pith/arXiv arXiv

  66. [77]

    Look Back to Reason Forward: Revisitable Memory for Long-Context

    Shi, Yaorui and Chen, Yuxin and Wang, Siyuan and Li, Sihang and Cai, Hengxing and Gu, Qi and Wang, Xiang and Zhang, An , booktitle =. Look Back to Reason Forward: Revisitable Memory for Long-Context. 2026 , url =

    2026

  67. [78]

    and Zhang, Hao and Gonzalez, Joseph E

    Zheng, Lianmin and Chiang, Wei-Lin and Sheng, Ying and Zhuang, Siyuan and Wu, Zhanghao and Zhuang, Yonghao and Lin, Zi and Li, Zhuohan and Li, Dacheng and Xing, Eric P. and Zhang, Hao and Gonzalez, Joseph E. and Stoica, Ion , booktitle =. Judging. 2023 , doi =

    2023

  68. [79]

    2026 , eprint =

    Hierarchical Evidence-Driven Reasoning for Long Document Understanding , author =. 2026 , eprint =

    2026

  69. [80]

    2026 , eprint =

    Loison, Ant. 2026 , eprint =

    2026

  70. [81]

    2512.12658 , archivePrefix =

    Xu, Qixin and Wang, Haozhe and Liu, Che and Lin, Fangzhen and Chen, Wenhu , year =. 2512.12658 , archivePrefix =

    arXiv

  71. [82]

    2504.10479 , archivePrefix =

    Zhu, Jinguo and Wang, Weiyun and Chen, Zhe and Liu, Zhaoyang and Ye, Shenglong and Gu, Lixin and Tian, Hao and Duan, Yuchen and Su, Weijie and Shao, Jie and Gao, Zhangwei and Cui, Erfei and Wang, Xuehui and Cao, Yue and Liu, Yangzhou and Wei, Xingguang and Zhang, Hongjie and Wang, Haomin and Xu, Weiye and Li, Hao and Wang, Jiahao and Deng, Nianchen and Li...

    Pith/arXiv arXiv