Pith. sign in

REVIEW 3 major objections 5 minor 40 references

The paper claims that progressively sharpening normalized reference-based rewards during RL training—Step-Aware Annealing—makes edit-distance-style rewards discriminative enough for document parsing to reach state-of-the-art element-level s

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 ↗

A reward-annealing trick that progressively sharpens edit-distance rewards improves GRPO-style RL for document parsing across text, tables, and formulas.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Solid empirical paper on reward annealing for document parsing, but the mechanism story oversells the near-tie sharpening and the single-run results need error bars. the 3 major comments →

arxiv 2608.00536 v1 pith:GG7EXLYF submitted 2026-08-01 cs.CV

DocPO: Advancing Document Policy Optimization via Tailored Step-Aware Rewards

Yunhao Wang , Binghong Wu , Zhenyu Huang , Jiacheng Shi , Shuo Huang , Tinghao Yu , Feng Zhang This is my paper
classification cs.CV
keywords Document ParsingReinforcement LearningReward EngineeringStep-Aware AnnealingGRPOEdit DistanceTable RecognitionFormula Recognition
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

DocPO tackles the problem that edit-distance rewards become nearly flat for near-perfect document parsing outputs, giving reinforcement learning little signal to improve hard cases. Its proposed mechanism, Step-Aware Annealing (SAA), applies a power-law transform M^gamma to any normalized reference reward and gradually increases gamma during training, amplifying tiny quality differences among high-scoring candidates while preserving their original ranking. On top of this, DocPO pairs SAA with element-specific rewards: normalized edit distance for text, weighted tree edit distance for tables, and a syntax-gated rubric-plus-edit reward for formulas. Across OmniDocBench and the harder DocElemHard benchmark, SAA consistently beats non-annealed rewards and static sharpening alternatives, and the resulting 3B model outperforms specialized systems with custom vision encoders and post-processing. If correct, this means reward-level optimization alone, on a general-purpose backbone, can substitute for much heavier architectural investment in document parsing.

Core claim

DocPO's central claim is that the bottleneck in RL-based document parsing is not the choice of base reward family but the curvature of the reward in the high-accuracy regime. Step-Aware Annealing transforms any normalized base reward M into M^gamma, with gamma rising from 1 toward a maximum of 9 on an exponential schedule whose pace is modulated by a task-wise Dynamic Dispersion Controller. This sharpens relative differences between near-correct candidates—the relative gap (1+delta/m)^gamma - 1 grows monotonically with gamma—while leaving the base reward's ranking untouched. Applied to NED for text, weighted tree edit distance for tables, and syntax-gated Rubric+edit for formulas, SAA improv

What carries the argument

The load-bearing object is Step-Aware Annealing (SAA): a power-law reward transform f_gamma(M)=M^gamma applied to normalized base rewards, with gamma annealed from 1 to 1+Delta_gamma (default Delta_gamma=8) over training steps. The Dynamic Dispersion Controller (DDC) modulates the annealing time constant using a rolling coefficient of variation of recent rewards computed separately per task. The power law preserves the original reward ranking while monotonically amplifying relative gaps between nearby high scores, so the same mechanism can sharpen any reference-based reward without changing what correctness means.

Load-bearing premise

The whole gain depends on the assumption that making small raw-reward gaps larger actually improves the learning signal after GRPO re-centers and re-scales rewards inside each group; the paper proves the raw-gap amplification but acknowledges it does not guarantee the normalized advantage improves.

What would settle it

Run a controlled comparison of GRPO with SAA versus a fixed high exponent (gamma=9) across multiple random seeds on the same RL dataset. If fixed-gamma matches or beats the annealed schedule, the claimed benefit of progressive annealing is not the cause. Alternatively, log the group-normalized advantages with and without SAA; if the separation among top-ranked normalized advantages does not increase while scores rise, the stated mechanism—sharper raw rewards improving optimization—is refuted.

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

If this is right

  • Document parsing RL can work with simple reference-based edit-distance rewards plus a curvature schedule, rather than requiring learned reward models or handcrafted composite scoring.
  • Architectural specialization—dedicated high-resolution encoders, aggressive vision-token compression, or additional multimodal pretraining—may not be necessary to reach top element-level parsing results on a general-purpose backbone.
  • SAA transfers across element types: one M^gamma schedule improves text, table, and formula rewards, making it a plug-in mechanism for any normalized reference reward.
  • SAA accelerates training as well as improving the final result, reaching a fixed TEDS threshold about 1.8x faster and maintaining a higher plateau than the non-annealed baseline.
  • The gains concentrate on structurally hard cases—vertical layouts, spanning cells, fuzzy scans, watermarks, and mixed-script text—where flat edit-distance rewards otherwise provide little learning signal.

Where Pith is reading between the lines

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

  • A natural extension the paper leaves implicit: if raw reward curvature is the active ingredient, SAA should transfer to other high-accuracy structured-generation tasks with saturating similarity metrics, such as code repair, math proof-step generation, or chart understanding, by applying the same M^gamma schedule to their existing normalized rewards.
  • Appendix B.10 shows the sharpening effect on raw rewards but stops short of proving that GRPO's group-normalized advantages improve monotonically; a direct follow-up is to instrument the normalized advantages and check whether SAA increases top-sample separation after normalization, which would confirm or refute the stated mechanism.
  • The DDC's dispersion-based schedule suggests a broader adaptive-curriculum principle: annealing speed can be set from reward dispersion, and one could test per-sample or per-batch curvature instead of per-task curvature to make the schedule even more responsive to difficulty.
  • Because SAA leaves the base reward unchanged, it composes with future improved base rewards—any new metric can be plugged in with the same annealing, meaning the contribution is orthogonal to reward design and likely to persist as better metrics appear.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes Step-Aware Annealing (SAA), a reward-sharpening mechanism for GRPO-style RL in document parsing. SAA applies a power-law transform f_gamma(M)=M^gamma to element-specific base rewards (NED for text, weighted TEDS for tables, a syntax-gated Rubric+edit reward for formulas) and anneals gamma from 1 to a maximum of 9 during training, with an additional Dynamic Dispersion Controller (DDC) modulating the annealing pace from recent reward statistics. The authors build DocPO on an unmodified Qwen2.5-VL-3B backbone and report state-of-the-art element-level results on OmniDocBench and a newly constructed DocElemHard benchmark, with ablations claiming SAA consistently outperforms non-annealed and fixed-curvature rewards across all three element types. The appendix provides formal propositions about the raw reward profile and acknowledges that they do not guarantee improvement in the normalized GRPO advantage.

Significance. If the empirical claims hold, the paper demonstrates a practically important result: reward-level optimization alone, without vision-encoder modifications, additional pretraining, or post-processing, can outperform specialized document VLMs with dedicated high-resolution encoders. The paper includes external validation on WikiTableSet and UniMERNet and reports training-stability diagnostics, which strengthen the empirical case. However, the central mechanism is not theoretically established, and the key ablations are single runs without error bars, so the significance of the specific contribution (SAA vs. a fixed nonlinearity) remains uncertain.

major comments (3)
  1. [3.4, App. B.10] The formal analysis does not support the stated mechanism for GRPO. Prop. B.1 and B.2 concern the raw reward profile, and App. B.10 explicitly concedes they 'do not by themselves guarantee monotonic growth in the normalized advantage.' The concern is stronger: for rewards near m=1, f_gamma(m+delta) = m^gamma + gamma m^(gamma-1) delta + O(delta^2), which is approximately affine in delta. GRPO advantages are invariant to affine transformations of group rewards, so A_i(gamma) ≈ A_i(1) in the near-tie high-score regime. This is exactly the regime SAA is motivated by. The reported gains therefore currently rest on empirical coincidence rather than the stated mechanism. Please provide either a corrected theoretical statement, or direct evidence that within-group advantage distributions actually change under SAA (e.g., histograms of advantage values or group reward spreads at different training
  2. [Tables 2, 4, 5, 7] All ablations are single runs with no error bars or seed variation. Key differences are small: Table 4 shows SAA at 93.1 vs. fixed gamma=8 at 92.3; Table 5 shows DDC on/off at 93.10 vs. 92.73; Table 2 formula OmniDocBench is 93.93 vs. 94.70. These magnitudes are within typical seed noise for RL training. Without multiple seeds and variance reporting, the central claim that dynamic annealing (rather than fixed curvature, reward choice, or seed noise) is the operative ingredient is not established. Please report mean and standard deviation over at least 3 seeds for the main SAA-vs-baseline comparisons, or provide learning-curve overlays from multiple runs.
  3. [Sec. 4.1, App. B.7] The main hyperparameters of SAA appear to be selected using the same benchmark on which the final results are reported. Delta_gamma=8 is justified by Table 14 (OmniDocBench Overall) and the formula weights (alpha=0.8, beta=0.2) are selected from Table 15 (OmniDocBench CDM). If OmniDocBench is the test set, this selection procedure risks overfitting to that benchmark, especially given that the final claims are measured against it. Please clarify whether a validation split was used for these choices, or report sensitivity of the chosen hyperparameters on held-out benchmarks such as WikiTableSet and UniMERNet.
minor comments (5)
  1. [Eq. (7)] The definition of Overall is ambiguous: it should be written as Overall = ((1-NED)*100 + TEDS + CDM)/3, with explicit parentheses around the numerator.
  2. [Limitations] The Limitations section mentions RL cost, proxy rewards, and element coverage, but does not mention that all experiments are single runs or that App. B.10 concedes the normalized-advantage mechanism is unproven. These are material caveats for readers.
  3. [Sec. 3.3, App. B.1-B.2] The formula rubric is instantiated with Qwen2.5-7B-Instruct, an external pretrained model. The claim 'without additional human supervision' is technically true, but the reward construction is not purely reference-based. Please clarify explicitly that the rubric is a frozen auxiliary model, and report its computational overhead in the training-cost comparison.
  4. [Figure 7] The '1.8x faster' claim is based on a single threshold crossing (91.0 TEDS) in one run. Please provide confidence intervals or multiple runs for this convergence comparison, as the variance in step count is likely substantial.
  5. [Sec. 4.2] The phrase 'no KL divergence constraint' is unusual for GRPO and may affect exploration stability. A one-sentence justification or citation for omitting KL control would help reproducibility.

Circularity Check

0 steps flagged

No significant circularity: the formal analysis is transparently limited and the reported gains are empirical measurements, not derivations from the reward transform.

full rationale

The paper's only formal derivation is in Appendix B.10, where Propositions B.1 and B.2 prove properties of the power-law transform f_gamma(M)=M^gamma: relative margin amplification and exponential concentration of reward mass. These are direct, correct consequences of the definition of the transform, and the paper does not claim they by themselves prove the central empirical claim. On the contrary, the appendix explicitly states: 'GRPO subsequently centers and scales rewards within each group through A_i=(r_i-mu_G)/sigma_G, so they do not by themselves guarantee monotonic growth in the normalized advantage.' This is an honest limitation, not a circular step. The main claims—that Step-Aware Annealing improves GRPO-style RL and that DocPO reaches strong OmniDocBench/DocElemHard scores—are supported by training runs and benchmark measurements, not derived from the reward definition. The base rewards (NED, TEDS, Rubric+edit) are defined before training and are not fitted to the reported scores; they align with evaluation metrics but that is a standard reward-design choice, not circularity. The self-citations (HunyuanOCR, StrucTab, MORE, progressive hard-case mining) are contextual or related-work references and are not load-bearing for the SAA mechanism. Some hyperparameters (Delta_gamma, alpha, beta) were selected using OmniDocBench outcomes (Tables 14-15), which is a benchmark-selection concern, but this is not a case of a fitted parameter being renamed as a prediction: the paper reports experimental results rather than claiming to predict them from the transform. Overall, no load-bearing step in the derivation chain reduces by construction to its own inputs.

Axiom & Free-Parameter Ledger

8 free parameters · 5 axioms · 1 invented entities

The method introduces several hand-set hyperparameters (gamma_init, Delta_gamma, tau_base, alpha/beta, TEDS costs) and a data pre-filtering rule; these are not derived from theory. The key domain assumptions are that the base rewards and the frozen LLM rubric are reliable, and that raw-reward sharpening transfers to GRPO advantages. No new physical entities are introduced; the SAA/DDC mechanism is an algorithmic construct validated only internally.

free parameters (8)
  • gamma_init = 1
    Default baseline curvature; identity at start of training.
  • Delta_gamma = 8
    Maximum curvature increase; chosen via sensitivity on OmniDocBench (Appendix B.7, Table 14).
  • tau_base = roughly half of total training steps
    Time constant of the annealing exponential in Eq. (5); set by hand, no sensitivity reported.
  • s_window (DDC threshold/window) = 3
    Early-phase cutoff and rolling window size in Eq. (6); ambiguous specification.
  • alpha (formula semantic weight) = 0.8
    Selected via sensitivity on OmniDocBench (Table 15) as the best of the tested grid.
  • beta (formula structural weight) = 0.2
    Selected together with alpha via sensitivity on OmniDocBench (Table 15).
  • Weighted TEDS operation costs = structural ops cost 2, content mismatch cost 1
    Hand-set in Figure 4; no sensitivity analysis.
  • Rollout pre-filter thresholds = remove all-perfect (8/8) and all-zero (0/8) samples
    Data pre-filtering rule in Appendix B.9; no ablation on these thresholds.
axioms (5)
  • domain assumption The base rewards (NED, weighted TEDS, Rubric+edit) are valid proxies for document parsing quality as measured by the evaluation metrics.
    The entire RL signal is these rewards; if they misalign with quality, SAA amplifies the wrong signal. Stated in Sections 3.1-3.3.
  • domain assumption The frozen Qwen2.5-7B-Instruct rubric provides reliable semantic-equivalence judgments for formulas during training.
    Validated on only 200 formula pairs (97.5% accuracy, Appendix B.2) and used as a reward component with caching.
  • domain assumption Sharpening the raw reward improves the GRPO normalized advantage in practice.
    Propositions B.1 and B.2 only characterize raw rewards; Appendix B.10 explicitly states they do not guarantee advantage improvement; the claim is empirically assumed.
  • standard math APTED computes the weighted tree edit distance correctly and efficiently enough for training.
    The paper cites APTED [22,23] and uses it as a reward; no independent verification in the paper.
  • standard math The power-law transform M^gamma preserves reward ordering and is differentiable for policy optimization.
    M^gamma is monotone for M in [0,1] with gamma>=1; standard calculus, used in Eq. (4).
invented entities (1)
  • Step-Aware Annealing (SAA) with Dynamic Dispersion Controller (DDC) no independent evidence
    purpose: Progressively increase reward curvature during RL to amplify near-tie differences among high-scoring samples; DDC adapts the annealing pace based on task-wise reward dispersion.
    Supported only by internal ablations (Tables 4, 5); no external falsifiable handle beyond the reported benchmarks. It is an algorithmic construct, not a physical entity.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of DocPO: Advancing Document Policy Optimization via Tailored Step-Aware Rewards." pith.science (2026). https://pith.science/paper/GG7EXLYF

@misc{pith2026260800536,
  author       = {Pith},
  title        = {Pith review of: DocPO: Advancing Document Policy Optimization via Tailored Step-Aware Rewards},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GG7EXLYF}},
  note         = {Machine review of arXiv:2608.00536}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Reinforcement learning (RL) for document parsing often relies on reference-based rewards rooted in edit distance (e.g., tree edit distance), yet it remains hard to optimize in the high-accuracy regime because such rewards become weakly discriminative: near-correct outputs receive very similar scores, providing limited learning signal for hard cases. We propose Step-Aware Annealing (SAA), a plug-and-play reward sharpening mechanism that progressively increases reward curvature during training, amplifying subtle quality differences among high-scoring samples while preserving stability in early learning. Built on SAA, we introduce DocPO, a document policy optimization framework with element-specific, reference-based rewards anchored by edit-distance signals: normalized string edit distance (NED) for text, tree edit distance similarity (TEDS) for tables, and a hybrid Rubric+edit reward for formulas. Experiments on OmniDocBench and DocElemHard show that SAA consistently improves GRPO-style RL across document elements over non-annealed rewards, without requiring additional human supervision for reward construction.

Figures

Figures reproduced from arXiv: 2608.00536 by Binghong Wu, Feng Zhang, Jiacheng Shi, Shuo Huang, Tinghao Yu, Yunhao Wang, Zhenyu Huang.

Figure 1
Figure 1. Figure 1: Two complementary views motivating DocPO. Left: Step-Aware Annealing improves optimization in the high-accuracy [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: DocPO pipeline. Element-specific rewards are normalized to [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Examples of LaTeX evaluation challenges. The [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of reward rescaling functions with [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of weighted tree edit distance-based [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of advantage distributions for [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Evaluation TEDS score during training (table recog [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Optimization-stability diagnostics for DocPO and [PITH_FULL_IMAGE:figures/full_fig_p011_8.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

40 extracted references · 15 canonical work pages · 4 internal anchors

  1. [1]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. 2025. Qwen2.5-VL Technical Report.arXiv preprint arXiv:2502.13923(2025)

    Pith/arXiv arXiv 2025

  2. [2]

    Lukas Blecher, Guillem Cucurull, Thomas Scialom, and Robert Stojnic. 2023. Nougat: Neural optical understanding for academic documents.arXiv preprint arXiv:2308.13418(2023)

    Pith/arXiv arXiv 2023

  3. [3]

    Cheng Cui, Ting Sun, Suyin Liang, Tingquan Gao, Zelun Zhang, Jiaxuan Liu, Xueqing Wang, Changda Zhou, Hongen Liu, Manhui Lin, et al. 2025. PaddleOCR- VL: Boosting Multilingual Document Parsing via a 0.9B Ultra-Compact Vision- Language Model.arXiv preprint arXiv:2510.14528(2025)

    arXiv 2025

  4. [4]

    Cheng Cui, Ting Sun, Manhui Lin, Tingquan Gao, Yubo Zhang, Jiaxuan Liu, Xueqing Wang, Zelun Zhang, Changda Zhou, Hongen Liu, et al. 2025. PaddleOCR 3.0 Technical Report.arXiv preprint arXiv:2507.05595(2025)

    Pith/arXiv arXiv 2025

  5. [5]

    Lei Cui, Yiheng Xu, Tengchao Lv, and Furu Wei. 2021. Document AI: Benchmarks, Models and Applications.arXiv preprint arXiv:2111.08609(2021)

    Pith/arXiv arXiv 2021

  6. [6]

    Hao Feng, Shu Wei, Xiang Fei, Wei Shi, Yingdong Han, Lei Liao, Jinghui Lu, Binghong Wu, Qi Liu, Chunhui Lin, et al . 2025. Dolphin: Document image parsing via heterogeneous anchor prompting. InFindings of the Association for Computational Linguistics: ACL 2025. 21919–21936

    work page 2025

  7. [7]

    Ling Fu, Zhebin Kuang, Jiajun Song, Mingxin Huang, Biao Yang, Yuzhe Li, Linghao Zhu, Qidi Luo, Xinyu Wang, Hao Lu, et al. 2025. OCRBench v2: An Improved Benchmark for Evaluating Large Multimodal Models on Visual Text Localization and Reasoning. InAdvances in Neural Information Processing Systems, Vol. 38

    work page 2025

  8. [8]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al . 2025. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning.arXiv preprint arXiv:2501.12948(2025)

    Pith/arXiv arXiv 2025

  9. [9]

    Jaekyu Ha, Robert M Haralick, and Ihsin T Phillips. 1995. Recursive XY cut using bounding boxes of connected components. InProceedings of 3rd International Conference on Document Analysis and Recognition, Vol. 2. IEEE, 952–955

    work page 1995

  10. [10]

    Hunyuan Vision Team, Pengyuan Lyu, Xingyu Wan, Gengluo Li, Shangpin Peng, Weinong Wang, Liang Wu, Huawen Shen, Yu Zhou, Canhui Tang, et al. 2025. HunyuanOCR Technical Report.arXiv preprint arXiv:2511.19575(2025)

    arXiv 2025

  11. [11]

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. 2024. GPT-4o System Card.arXiv preprint arXiv:2410.21276(2024)

    Pith/arXiv arXiv 2024

  12. [12]

    Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. 2022. OCR-Free Document Understanding Transformer. InComputer Vision– ECCV 2022. 498–517. doi:10.1007/978-3-031-19815-1_29

    doi:10.1007/978-3-031-19815-1_29 2022

  13. [13]

    Aviral Kumar, Vincent Zhuang, Rishabh Agarwal, Yi Su, John D Co-Reyes, Avi Singh, Kate Baumli, Shariq Iqbal, Colton Bishop, Rebecca Roelofs, et al . 2024. Training language models to self-correct via reinforcement learning.arXiv preprint arXiv:2409.12917(2024)

    Pith/arXiv arXiv 2024

  14. [14]

    Gengluo Li, Shangpin Peng, Chengquan Zhang, Binghong Wu, Hao Feng, Weinong Wang, Pengyuan Lyu, Huawen Shen, Xingyu Wan, Zhuotao Tian, et al. 2026. StrucTab: A Structured Optimization Framework for Table Parsing. arXiv preprint arXiv:2606.29905(2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  15. [15]

    Gengluo Li, Xingyu Wan, Shangpin Peng, Weinong Wang, Hao Feng, Yongkun Du, Binghong Wu, Zheng Ruan, Zhiqiong Lu, Liang Wu, et al. 2026. HunyuanOCR-1.5: Making Lightweight OCR VLMs Faster and Better.arXiv preprint arXiv:2607.04884 (2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  16. [16]

    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 preprint arXiv:2512.02498(2025)

    arXiv 2025

  17. [17]

    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 preprint arXiv:2506.05218(2025)

    arXiv 2025

  18. [18]

    Jinghui Lu, Haiyang Yu, Yanjie Wang, Yongjie Ye, Jingqun Tang, Ziwei Yang, Binghong Wu, Qi Liu, Hao Feng, Han Wang, et al . 2025. A Bounding Box is Worth One Token: Interleaving Layout and Text in a Large Language Model for Document Understanding. InFindings of the Association for Computational Linguistics: ACL 2025. 7252–7273

    work page 2025

  19. [19]

    Tengchao Lv, Yupan Huang, Jingye Chen, Yuzhong Zhao, Yilin Jia, Lei Cui, Shuming Ma, Yaoyao Chang, Shaohan Huang, Wenhui Wang, et al. 2023. Kosmos- 2.5: A multimodal literate model.arXiv preprint arXiv:2309.11419(2023)

    Pith/arXiv arXiv 2023

  20. [20]

    Junbo Niu, Zheng Liu, Zhuangcheng Gu, Bin Wang, Linke Ouyang, Zhiyuan Zhao, Tao Chu, Tianyao He, Fan Wu, Qintong Zhang, et al . 2025. MinerU2.5: A Decoupled Vision-Language Model for Efficient High-Resolution Document Parsing.arXiv preprint arXiv:2509.22186(2025)

    Pith/arXiv arXiv 2025

  21. [21]

    Linke Ouyang, Yuan Qu, Hongbin Zhou, Jiawei Zhu, Rui Zhang, Qunshu Lin, Bin Wang, Zhiyuan Zhao, Man Jiang, Xiaomeng Zhao, et al. 2025. OmniDocBench: Benchmarking Diverse PDF Document Parsing with Comprehensive Annotations. InProceedings of the Computer Vision and Pattern Recognition Conference. 24838– 24848

    work page 2025

  22. [22]

    Mateusz Pawlik and Nikolaus Augsten. 2015. Efficient computation of the tree edit distance.ACM Transactions on Database Systems (TODS)40, 1 (2015), 1–40

    work page 2015

  23. [23]

    Mateusz Pawlik and Nikolaus Augsten. 2016. Tree edit distance: Robust and memory-efficient.Information Systems56 (2016), 157–173

    work page 2016

  24. [24]

    Jake Poznanski, Aman Rangapur, Jon Borchardt, Jason Dunkelberger, Regan Huff, Daniel Lin, Christopher Wilhelm, Kyle Lo, and Luca Soldaini. 2025. olmOCR: Unlocking Trillions of Tokens in PDFs with Vision Language Models.arXiv preprint arXiv:2502.18443(2025)

    arXiv 2025

  25. [25]

    Jake Poznanski, Luca Soldaini, and Kyle Lo. 2025. olmOCR 2: Unit Test Rewards for Document OCR.arXiv preprint arXiv:2510.19817(2025)

    arXiv 2025

  26. [26]

    Jingqun Tang, Chunhui Lin, Zhen Zhao, Shu Wei, Binghong Wu, Qi Liu, Hao Feng, Yang Li, Siqi Wang, Lei Liao, et al. 2024. TextSquare: Scaling up text-centric visual instruction tuning.arXiv preprint arXiv:2404.12803(2024)

    Pith/arXiv arXiv 2024

  27. [27]

    Baode Wang, Biao Wu, Weizhen Li, Meng Fang, Zuming Huang, Jun Huang, Haozhe Wang, Yanjie Liang, Ling Chen, Wei Chu, and Yuan Qi. 2025. Infinity Parser: Layout Aware Reinforcement Learning for Scanned Document Parsing. arXiv preprint arXiv:2506.03197(2025)

    arXiv 2025

  28. [28]

    Bin Wang, Chao Xu, Xiaomeng Zhao, Linke Ouyang, Fan Wu, Zhiyuan Zhao, Rui Xu, Kaiwen Liu, Yuan Qu, Fukai Shang, et al. 2024. MinerU: An Open-Source So- lution for Precise Document Content Extraction.arXiv preprint arXiv:2409.18839 (2024)

    Pith/arXiv arXiv 2024

  29. [29]

    Zilong Wang, Yiheng Xu, Lei Cui, Jingbo Shang, and Furu Wei. 2021. Lay- outReader: Pre-training of Text and Layout for Reading Order Detection.arXiv preprint arXiv:2108.11591(2021)

    Pith/arXiv arXiv 2021

  30. [30]

    Haoran Wei, Yaofeng Sun, and Yukun Li. 2025. DeepSeek-OCR: Contexts Optical Compression.arXiv preprint arXiv:2510.18234(2025)

    Pith/arXiv arXiv 2025

  31. [31]

    Binghong Wu, Yehui Yang, Dalu Yang, Junde Wu, Xiaorong Wang, Haifeng Huang, Lei Wang, and Yanwu Xu. 2021. Progressive hard-case mining across pyramid levels for object detection.arXiv preprint arXiv:2109.07217(2021)

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  32. [32]

    Long Xu, Binghong Wu, Tinghao Yu, Hao Feng, Zhenyu Huang, Haoqing Jiang, Yunhao Wang, Shuo Huang, and Feng Zhang. 2026. MORE: A Multilingual Document Parsing Benchmark and Evaluation.arXiv preprint arXiv:2607.02956 (2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  33. [33]

    Biao Yang, Bin Wen, Boyang Ding, et al. 2025. Kwai Keye-VL 1.5 Technical Report. arXiv preprint arXiv:2509.01563(2025)

    Pith/arXiv arXiv 2025

  34. [34]

    Jiarui Zhang, Yuliang Liu, Zijun Wu, Guosheng Pang, Zhili Ye, Yupei Zhong, Junteng Ma, Tao Wei, Haiyang Xu, Weikai Chen, et al. 2025. MonkeyOCR v1.5 Technical Report: Unlocking Robust Document Parsing for Complex Patterns. arXiv preprint arXiv:2511.10390(2025)

    arXiv 2025

  35. [35]

    Weichao Zhao, Hao Feng, Qi Liu, Jingqun Tang, Binghong Wu, Lei Liao, Shu Wei, Yongjie Ye, Hao Liu, Wengang Zhou, et al. 2024. TabPedia: Towards compre- hensive visual table understanding with concept synergy. InAdvances in Neural Information Processing Systems, Vol. 37. 7185–7212

    work page 2024

  36. [36]

    Zhen Zhao, Jingqun Tang, Chunhui Lin, Binghong Wu, Can Huang, Hao Liu, Xin Tan, Zhizhong Zhang, and Yuan Xie. 2024. Multi-modal In-Context Learn- ing Makes an Ego-evolving Scene Text Recognizer. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 15567–15576

    work page 2024

  37. [37]

    ""{gt}""

    Zhen Zhao, Jingqun Tang, Binghong Wu, Chunhui Lin, Shu Wei, Hao Liu, Xin Tan, Zhizhong Zhang, Can Huang, and Yuan Xie. 2024. Harmonizing Visual Text Comprehension and Generation. InAdvances in Neural Information Processing Systems, Vol. 37. 97499–97522. MM ’26, November 10–14, 2026, Rio de Janeiro, Brazil Wang et al. Supplementary Material A Prompts for D...

    work page arXiv 2024

  38. [38]

    GRPO subsequently centers and scales rewards within each group through𝐴𝑖 =(𝑟 𝑖− 𝜇G)/𝜎G, so they do not by themselves guarantee monotonic growth in the normalized advantage

    For the concentration result, note that for any 𝑗≠𝑖 ∗ where 𝑖∗ =arg max𝑖𝑀𝑖, we have𝑞𝑗/𝑞𝑖∗ =(𝑀 𝑗/𝑀𝑖∗)𝛾→ 0as 𝛾→∞ , so 𝑞𝑖∗→1.□ Connection to GRPO Advantage.The propositions above de- scribe the pre-normalization reward profile. GRPO subsequently centers and scales rewards within each group through𝐴𝑖 =(𝑟 𝑖− 𝜇G)/𝜎G, so they do not by themselves guarantee monot...

    work page 2026

  39. [39]

    Nam Ly, Atsuhiro Takasu, Phuc Nguyen, and Hideaki Takeda. 2023. Rethinking Image-Based Table Recognition Using Weakly Supervised Methods. InProceedings of the 12th International Conference on Pattern Recognition Applications and Methods (ICPRAM). 872–880. https://doi.org/10.5220/0011682600003411

    doi:10.5220/0011682600003411 2023

  40. [40]

    Bin Wang, Zhuangcheng Gu, Guang Liang, Chao Xu, Bo Zhang, Botian Shi, and Conghui He. 2024. UniMERNet: A Universal Network for Real-World Mathematical Expression Recognition.arXiv preprint arXiv:2404.15254. https://arxiv.org/abs/2404. 15254

    Pith/arXiv arXiv 2024

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