arxiv: 2604.02692 · v1 · submitted 2026-04-03 · 💻 cs.CV

Recognition: no theorem link

Parser-Oriented Structural Refinement for a Stable Layout Interface in Document Parsing

Fuyuan Liu , Dianyu Yu , He Ren , Nayu Liu , Xiaomian Kang , Delai Qiu , Fa Zhang , Genpeng Zhen

show 5 more authors

Shengping Liu Jiaen Liang Wei Huang Yining Wang Junnan Zhu

Authors on Pith no claims yet

Pith reviewed 2026-05-13 21:04 UTC · model grok-4.3

classification 💻 cs.CV

keywords document parsinglayout analysisstructural refinementDETR detectorparser interfacereading orderinstance retentiondocument layout analysis

0 comments

The pith

A lightweight structural refinement module between a DETR-style detector and parser stabilizes the layout interface by jointly deciding instance retention, refining boxes, and predicting input order.

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

The paper introduces a lightweight structural refinement stage placed between a DETR-style detector and the downstream parser in document layout analysis pipelines. This module treats raw detector outputs as a hypothesis pool and performs set-level reasoning over query features, semantic cues, box geometry, and visual evidence to select retained instances, refine their localization, and determine the serialization order handed to the parser. The goal is to prevent inconsistent retained sets and mismatched orders on dense pages with overlaps and ambiguous boundaries, which otherwise produce severe downstream parsing errors. Retention-oriented supervision and a difficulty-aware ordering objective are added to better align outputs with parser expectations on complex pages. When integrated into end-to-end pipelines, the approach improves page-level layout quality and reduces sequence mismatch.

Core claim

Treating raw detector outputs as a compact hypothesis pool, the proposed module performs set-level reasoning over query features, semantic cues, box geometry, and visual evidence. From a shared refined structural state, it jointly determines instance retention, refines box localization, and predicts parser input order before handoff, with retention-oriented supervision and a difficulty-aware ordering objective to align the retained set and order with final parser input.

What carries the argument

The lightweight structural refinement module that performs set-level reasoning over query features, semantic cues, box geometry, and visual evidence to produce a retained instance set and parser-compatible order.

If this is right

Consistently improves page-level layout quality across public benchmarks.
Substantially reduces sequence mismatch when integrated into standard end-to-end parsing pipelines.
Achieves a Reading Order Edit of 0.024 on OmniDocBench.
Delivers stronger results on structurally complex pages through the difficulty-aware ordering objective.

Where Pith is reading between the lines

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

The same interface-stabilization idea could apply to other modular detection pipelines where downstream components require consistent ordered instance sets.
Tighter coupling between the refinement module and detector training might further reduce retained-set inconsistencies.
Set-level reasoning over geometry and semantics may help similar interface problems in tasks like multi-object tracking or scene parsing.

Load-bearing premise

The lightweight module can reliably perform set-level reasoning over query features, semantic cues, box geometry, and visual evidence to produce a retained instance set and order that matches what the parser expects, without access to the full detector output.

What would settle it

Measuring whether the Reading Order Edit score stays near 0.024 or rises on a new test set of pages with unseen dense overlaps and ambiguous boundaries would show if the stabilization holds.

Figures

Figures reproduced from arXiv: 2604.02692 by Delai Qiu, Dianyu Yu, Fa Zhang, Fuyuan Liu, Genpeng Zhen, He Ren, Jiaen Liang, Junnan Zhu, Nayu Liu, Shengping Liu, Wei Huang, Xiaomian Kang, Yining Wang.

**Figure 1.** Figure 1: Illustration of detector-to-parser hando! failures endtoend parsers have advanced rapidly, pipeline systems remain tttiitil dlt bf thi#i biguous boundaries, and visually similar text blocks. Around a true [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

**Figure 3.** Figure 3: Qualitative comparison on two challenging layouts (Case A: equations, Case B: chemical schemes). Our method avoids [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

read the original abstract

Accurate document parsing requires both robust content recognition and a stable parser interface. In explicit Document Layout Analysis (DLA) pipelines, downstream parsers do not consume the full detector output. Instead, they operate on a retained and serialized set of layout instances. However, on dense pages with overlapping regions and ambiguous boundaries, unstable layout hypotheses can make the retained instance set inconsistent with its parser input order, leading to severe downstream parsing errors. To address this issue, we introduce a lightweight structural refinement stage between a DETR-style detector and the parser to stabilize the parser interface. Treating raw detector outputs as a compact hypothesis pool, the proposed module performs set-level reasoning over query features, semantic cues, box geometry, and visual evidence. From a shared refined structural state, it jointly determines instance retention, refines box localization, and predicts parser input order before handoff. We further introduce retention-oriented supervision and a difficulty-aware ordering objective to better align the retained instance set and its order with the final parser input, especially on structurally complex pages. Extensive experiments on public benchmarks show that our method consistently improves page-level layout quality. When integrated into a standard end-to-end parsing pipeline, the stabilized parser interface also substantially reduces sequence mismatch, achieving a Reading Order Edit of 0.024 on OmniDocBench.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper adds a lightweight refinement module after a DETR detector to stabilize retained instances and reading order for downstream parsers, but the abstract leaves the actual gains and failure modes untested.

read the letter

The core idea is inserting a small set-level module between detector and parser. It takes the hypothesis pool, reasons over features, semantics, geometry and visuals, then decides what to keep, tweaks the boxes, and sets the order before handing off. They add retention supervision and a difficulty-aware ordering loss to push the output closer to what the parser actually needs on messy pages. That targets a real pipeline pain point where unstable detections cause sequence errors downstream, and the abstract claims it drops Reading Order Edit to 0.024 on OmniDocBench while lifting page-level layout quality overall. The design is straightforward and the supervision signals are tied directly to parser needs rather than generic detection metrics, which is a reasonable way to close the interface gap. The main weakness is that nothing in the abstract shows baselines, ablations, or how the module behaves when the initial hypotheses are incomplete or heavily overlapping. The stress-test concern holds: if the detector misses instances or assigns wrong semantics, a lightweight refiner working only on the pool cannot recover them. Without those controls or error analysis, the reported numbers are hard to trust as general improvement rather than dataset-specific tuning. This is useful reading for anyone running DETR-based document parsers who sees order or retention errors on dense layouts. It is not a broad advance but a practical fix worth checking once the experiments are visible. I would send it for peer review to get the full details and see whether the module actually delivers under the conditions the abstract describes.

Referee Report

2 major / 2 minor

Summary. The paper introduces a lightweight structural refinement module inserted between a DETR-style detector and a downstream parser in document layout analysis pipelines. Treating detector outputs as a hypothesis pool, the module performs set-level reasoning over query features, semantic cues, box geometry, and visual evidence to jointly decide instance retention, refine bounding boxes, and predict parser input order. Retention-oriented supervision and a difficulty-aware ordering objective are proposed to align the retained set with parser expectations, particularly on complex pages. The authors claim consistent improvements in page-level layout quality and, when integrated into end-to-end parsing, a reduction in sequence mismatch to a Reading Order Edit of 0.024 on OmniDocBench.

Significance. If substantiated, the work provides a practical, parser-aligned stabilization layer that could reduce downstream parsing errors arising from unstable layout hypotheses on dense or ambiguous pages. The emphasis on retention-oriented and difficulty-aware supervision tailored to the parser interface is a targeted strength that may generalize across DETR-based DLA systems without altering the detector or parser.

major comments (2)

[Abstract] Abstract: The central claim of a Reading Order Edit of 0.024 on OmniDocBench is presented without any baseline comparisons, ablation studies, statistical significance tests, or experimental setup details; this directly undermines assessment of whether the refinement module delivers the reported reduction in sequence mismatch.
[Abstract] Abstract: The load-bearing assumption that set-level reasoning from query features, semantic cues, box geometry, and visual evidence alone (without full detector output or global context) can reliably recover correct retention and order on pages with overlapping regions is not supported by failure-mode analysis or targeted ablations; if this assumption fails, the claimed ROE improvement would not hold.

minor comments (2)

[Abstract] The abstract references 'public benchmarks' and 'OmniDocBench' but does not name the full set of datasets or provide any quantitative layout-quality metrics beyond the single ROE value.
[Abstract] The term 'refined structural state' is used without an accompanying equation or diagram reference in the summary, making the joint prediction mechanism harder to follow.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We agree that the abstract would benefit from additional context to better substantiate the reported results and assumptions. We will revise the abstract and add supporting analysis in the main text as detailed below.

read point-by-point responses

Referee: [Abstract] Abstract: The central claim of a Reading Order Edit of 0.024 on OmniDocBench is presented without any baseline comparisons, ablation studies, statistical significance tests, or experimental setup details; this directly undermines assessment of whether the refinement module delivers the reported reduction in sequence mismatch.

Authors: We acknowledge that the abstract as currently written presents the key quantitative result without accompanying context. The full manuscript (Sections 4 and 5) contains the requested baseline comparisons against standard DETR-style detectors, ablation studies on the refinement components, and details of the OmniDocBench evaluation protocol. In the revised version we will expand the abstract to include a concise statement of the baseline ROE, the magnitude of improvement, and a brief note on the experimental setup, while preserving the abstract's length constraints. revision: yes
Referee: [Abstract] Abstract: The load-bearing assumption that set-level reasoning from query features, semantic cues, box geometry, and visual evidence alone (without full detector output or global context) can reliably recover correct retention and order on pages with overlapping regions is not supported by failure-mode analysis or targeted ablations; if this assumption fails, the claimed ROE improvement would not hold.

Authors: We agree that explicit failure-mode analysis and targeted ablations isolating the set-level reasoning would strengthen the support for this assumption. Our current experiments already evaluate performance on dense pages containing overlapping regions (reported in Table 3 and the qualitative analysis), but we did not include a dedicated failure-case study or component-wise ablation on overlapping subsets. In the revision we will add both a targeted ablation on the contribution of visual evidence and box geometry for overlapping instances and a short failure-mode section discussing remaining error cases on such pages. revision: yes

Circularity Check

0 steps flagged

No circularity: new supervision signals and module are independent of target metrics

full rationale

The paper introduces a lightweight structural refinement module that performs set-level reasoning and adds retention-oriented supervision plus a difficulty-aware ordering objective. These elements are explicitly new additions tied to parser input needs rather than being defined in terms of the downstream Reading Order Edit metric or any fitted parameter. No equations or steps in the described method reduce a prediction to its own inputs by construction, and no load-bearing self-citations or ansatzes are invoked to force the reported 0.024 ROE result. The claims rest on experimental integration into an end-to-end pipeline and benchmark measurements, which remain falsifiable and non-circular.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides insufficient technical detail to enumerate free parameters, axioms, or invented entities; review is limited to the abstract.

pith-pipeline@v0.9.0 · 5567 in / 1069 out tokens · 35204 ms · 2026-05-13T21:04:48.340635+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages · 3 internal anchors

[1]

Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexan- der Kirillov, and Sergey Zagoruyko. 2020. End-to-End Object Detection with Transformers. InProceedings of the 16th European Conference on Computer Vision (ECCV). 213–229. doi:10.1007/978-3-030-58452-8_13

work page doi:10.1007/978-3-030-58452-8_13 2020
[2]

Yufan Chen, Ruiping Liu, Junwei Zheng, Di Wen, Kunyu Peng, Jiaming Zhang, and Rainer Stiefelhagen. 2025. Graph-based Document Structure Analysis. InThe Thirteenth International Conference on Learning Representations (ICLR). https: //openreview.net/forum?id=Fu0aggezN9

work page 2025
[3]

Cheng Cui, Ting Sun, Suyin Liang, Tingquan Gao, and others. 2025. PaddleOCR- VL: Boosting Multilingual Document Parsing via a 0.9B Ultra-Compact Vision- Language Model. arXiv:2510.14528 https://arxiv.org/abs/2510.14528

work page arXiv 2025
[4]

Cheng Cui, Ting Sun, Suyin Liang, Tingquan Gao, and others. 2026. PaddleOCR- VL-1.5: Towards a Multi-Task 0.9B VLM for Robust In-the-Wild Document Parsing. arXiv:2601.21957 https://arxiv.org/abs/2601.21957

work page internal anchor Pith review Pith/arXiv arXiv 2026
[5]

Cheng Cui, Ting Sun, Manhui Lin, Tingquan Gao, and others. 2025. PaddleOCR 3.0 Technical Report. arXiv:2507.05595 https://arxiv.org/abs/2507.05595

work page internal anchor Pith review arXiv 2025
[6]

Cheng Da, Chuwei Luo, Qi Zheng, and Cong Yao. 2023. Vision Grid Transformer for Document Layout Analysis. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 19405–19415. doi:10.1109/ICCV51070. 2023.01783

work page doi:10.1109/iccv51070 2023
[7]

Shuaiqi Duan, Yadong Xue, Weihan Wang, Zhe Su, and others. 2026. GLM-OCR Technical Report. arXiv:2603.10910 https://arxiv.org/abs/2603.10910

work page arXiv 2026
[8]

Hao Feng, Shu Wei, Xiang Fei, Wei Shi, Yingdong Han, Lei Liao, Jinghui Lu, Binghong Wu, Qi Liu, Chunhui Lin, Jingqun Tang, Hao Liu, and Can Huang. 2025. Dolphin: Document Image Parsing via Heterogeneous Anchor Prompting. In Findings of the Association for Computational Linguistics: ACL 2025. 21919–21936. https://aclanthology.org/2025.findings-acl.1130/

work page 2025
[9]

Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. 2022. Lay- outLMv3: Pre-training for Document AI with Unified Text and Image Masking. In Proceedings of the 30th ACM International Conference on Multimedia. 4083–4091. https://doi.org/10.1145/3503161.3548112

work page doi:10.1145/3503161.3548112 2022
[10]

Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, and Furu Wei. 2022. DiT: Self-supervised Pre-training for Document Image Transformer. InProceedings of the 30th ACM International Conference on Multimedia. 3530–3539. doi:10.1145/ 3503161.3547911

work page arXiv 2022
[11]

Qiwei Li, Zuchao Li, Xiantao Cai, Bo Du, and Hai Zhao. 2023. Enhancing Visually- Rich Document Understanding via Layout Structure Modeling. InProceedings of the 31st ACM International Conference on Multimedia. 4513–4523. doi:10.1145/ 3581783.3612327

work page arXiv 2023
[12]

Xin Li, Mingming Gong, Yunfei Wu, Jianxin Dai, Antai Guo, Xinghua Jiang, Haoyu Cao, Yinsong Liu, Deqiang Jiang, and Xing Sun. 2025. DREAM: Document Reconstruction via End-to-end Autoregressive Model. InProceedings of the 33rd ACM International Conference on Multimedia. 2949–2957. doi:10.1145/3746027. 3754906

work page doi:10.1145/3746027 2025
[13]

Yumeng Li, Guang Yang, Hao Liu, Bowen Wang, and Colin Zhang. 2025. dots.ocr: Multilingual Document Layout Parsing in a Single Vision-Language Model. arXiv:2512.02498 https://arxiv.org/abs/2512.02498

work page arXiv 2025
[14]

Zhang Li, Yuliang Liu, Qiang Liu, Zhiyin Ma, Ziyang Zhang, Shuo Zhang, Zidun Guo, Jiarui Zhang, Xinyu Wang, and Xiang Bai. 2025. MonkeyOCR: Document Parsing with a Structure-Recognition-Relation Triplet Paradigm. arXiv:2506.05218 https://arxiv.org/abs/2506.05218

work page arXiv 2025
[15]

Fuyuan Liu, Dianyu Yu, He Ren, Nayu Liu, Xiaomian Kang, Delai Qiu, Fa Zhang, Genpeng Zhen, Shengping Liu, Jiaen Liang, Wei Huang, Yining Wang, and Jun- nan Zhu. 2026. FocalOrder: Focal Preference Optimization for Reading Order Detection. arXiv:2601.07483 https://arxiv.org/abs/2601.07483

work page arXiv 2026
[16]

Fuyuan Liu, Dianyu Yu, He Ren, Nayu Liu, Xiaomian Kang, Delai Qiu, Fa Zhang, Genpeng Zhen, Shengping Liu, Jiaen Liang, Wei Huang, Yining Wang, and Junnan Zhu. 2026. PARL: Position-Aware Relation Learning Network for Document Layout Analysis. arXiv:2601.07620 https://arxiv.org/abs/2601.07620

work page arXiv 2026
[17]

Yuan Liu, Zhongyin Zhao, Le Tian, Haicheng Wang, Xubing Ye, Yangxiu You, Zilin Yu, Chuhan Wu, Zhou Xiao, Yang Yu, and Jie Zhou. 2025. POINTS-Reader: Distillation-Free Adaptation of Vision-Language Models for Document Con- version. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP). 1576–1601. https://aclanthol...

work page 2025
[18]

Said Gurbuz, and Peter W

Nikolaos Livathinos, Christoph Auer, Ahmed Nassar, Rafael Teixeira de Lima, Maksym Lysak, Brown Ebouky, Cesar Berrospi, Michele Dolfi, Panagiotis Vagenas, Matteo Omenetti, Kasper Dinkla, Yusik Kim, Valery Weber, Lucas Morin, Ingmar Meijer, Viktor Kuropiatnyk, Tim Strohmeyer, A. Said Gurbuz, and Peter W. J. Staar. 2025. Advanced Layout Analysis Models for ...

work page arXiv 2025
[19]

Junbo Niu, Zheng Liu, Zhuangcheng Gu, Bin Wang, and others. 2025. MinerU2.5: A Decoupled Vision-Language Model for Efficient High-Resolution Document Parsing. arXiv:2509.22186 https://arxiv.org/abs/2509.22186

work page arXiv 2025
[20]

Linke Ouyang, Yuan Qu, Hongbin Zhou, Jiawei Zhu, Rui Zhang, Qunshu Lin, Bin Wang, Zhiyuan Zhao, Man Jiang, Xiaomeng Zhao, Jin Shi, Fan Wu, Pei Chu, Minghao Liu, Zhenxiang Li, Chao Xu, Bo Zhang, Botian Shi, Zhongying Tu, and Conghui He. 2025. OmniDocBench: Benchmarking Diverse PDF Docu- ment Parsing with Comprehensive Annotations. InProceedings of the IEEE...

work page 2025
[21]

Yansong Peng, Hebei Li, Peixi Wu, Yueyi Zhang, Xiaoyan Sun, and Feng Wu. 2025. D-FINE: Redefine Regression Task of DETRs as Fine-grained Distribution Re- finement. InThe Thirteenth International Conference on Learning Representations (ICLR). https://openreview.net/forum?id=MFZjrTFE7h

work page 2025
[22]

Nassar, and Peter Staar

Birgit Pfitzmann, Christoph Auer, Michele Dolfi, Ahmed S. Nassar, and Peter Staar

work page
[23]

and Staar, Peter , title =

DocLayNet: A Large Human-Annotated Dataset for Document-Layout Segmentation. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 3743–3751. doi:10.1145/3534678.3539043

work page doi:10.1145/3534678.3539043
[24]

Jake Poznanski, Luca Soldaini, and Kyle Lo. 2025. olmOCR 2: Unit Test Rewards for Document OCR. arXiv:2510.19817 https://arxiv.org/abs/2510.19817

work page arXiv 2025
[25]

Reid, and Silvio Savarese

Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian D. Reid, and Silvio Savarese. 2019. Generalized Intersection over Union: A Metric and a Loss for Bounding Box Regression. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 658–666. doi:10.1109/CVPR. 2019.00075

work page doi:10.1109/cvpr 2019
[26]

Ting Sun, Cheng Cui, Yuning Du, and Yi Liu. 2025. PP-DocLayout: A Unified Document Layout Detection Model to Accelerate Large-Scale Data Construction. arXiv:2503.17213 https://arxiv.org/abs/2503.17213

work page arXiv 2025
[27]

Hunyuan Vision Team, Pengyuan Lyu, Xingyu Wan, Gengluo Li, and others

work page
[28]

Hunyuanocr Technical Report.arXiv preprint arXiv:2511.19575, 2025

HunyuanOCR Technical Report. arXiv:2511.19575 https://arxiv.org/abs/ 2511.19575

work page arXiv
[29]

Jiawei Wang, Kai Hu, and Qiang Huo. 2024. DLAFormer: An End-to-End Transformer For Document Layout Analysis. InProceedings of the 18th In- ternational Conference on Document Analysis and Recognition (ICDAR). 40–57. https://doi.org/10.1007/978-3-031-70546-5_3

work page doi:10.1007/978-3-031-70546-5_3 2024
[30]

Wenjin Wang, Zhengjie Huang, Bin Luo, Qianglong Chen, Qiming Peng, Yinxu Pan, Weichong Yin, Shikun Feng, Yu Sun, Dianhai Yu, and Yin Zhang. 2022. mmLayout: Multi-grained MultiModal Transformer for Document Understanding. InProceedings of the 30th ACM International Conference on Multimedia. 4877–4886. doi:10.1145/3503161.3548406

work page doi:10.1145/3503161.3548406 2022
[31]

Zilong Wang, Yiheng Xu, Lei Cui, Jingbo Shang, and Furu Wei. 2021. Lay- outReader: Pre-training of Text and Layout for Reading Order Detection. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Pro- cessing (EMNLP). 4735–4744. https://doi.org/10.18653/v1/2021.emnlp-main.389

work page doi:10.18653/v1/2021.emnlp-main.389 2021
[32]

Haoran Wei, Yaofeng Sun, and Yukun Li. 2025. DeepSeek-OCR: Contexts Optical Compression. arXiv:2510.18234 https://arxiv.org/abs/2510.18234

work page internal anchor Pith review Pith/arXiv arXiv 2025
[33]

Hao Wu, Haoran Lou, Xinyue Li, Zuodong Zhong, and others. 2026. FireRed-OCR Technical Report. arXiv:2603.01840 https://arxiv.org/abs/2603.01840 Parser-Oriented Structural Refinement for a Stable Layout Interface in Document Parsing Conference acronym ’XX, June 03–05, 2018, Woodstock, NY

work page arXiv 2026
[34]

Kun Yin, Yunfei Wu, Bing Liu, Zhongpeng Cai, and others. 2026. Youtu- Parsing: Perception, Structuring and Recognition via High-Parallelism Decoding. arXiv:2601.20430 https://arxiv.org/abs/2601.20430

work page arXiv 2026
[35]

Jiarui Zhang, Yuliang Liu, Zijun Wu, Guosheng Pang, and others. 2025. Monkey- OCR v1.5 Technical Report: Unlocking Robust Document Parsing for Complex Patterns. arXiv:2511.10390 https://arxiv.org/abs/2511.10390

work page arXiv 2025
[36]

Peng Zhang, Yunlu Xu, Zhanzhan Cheng, Shiliang Pu, Jing Lu, Liang Qiao, Yi Niu, and Fei Wu. 2020. TRIE: End-to-End Text Reading and Information Extraction for Document Understanding. InProceedings of the 28th ACM International Conference on Multimedia. 1413–1422. doi:10.1145/3394171.3413900

work page doi:10.1145/3394171.3413900 2020
[37]

Zhiyuan Zhao, Hengrui Kang, Bin Wang, and Conghui He. 2024. DocLayout- YOLO: Enhancing Document Layout Analysis through Diverse Synthetic Data and Global-to-Local Adaptive Perception. arXiv:2410.12628 https://arxiv.org/ abs/2410.12628

work page arXiv 2024
[38]

Changda Zhou, Ziyue Gao, Xueqing Wang, Tingquan Gao, Cheng Cui, Jing Tang, and Yi Liu. 2026. Real5-OmniDocBench: A Full-Scale Physical Reconstruction Benchmark for Robust Document Parsing in the Wild. arXiv:2603.04205 https: //arxiv.org/abs/2603.04205

work page arXiv 2026