Pith. sign in

REVIEW 4 major objections 7 minor 36 references

HPD-Parsing claims that replacing full-page autoregressive decoding with a layout-coordinated main branch and forked parallel content branches — plus progressive multi-token prediction — reaches 4,752 tokens per second (2.62× the fastest pr

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

2026-08-01 14:09 UTC pith:NXSHD7XV

load-bearing objection A genuinely new decoding paradigm for document parsing with credible speedups, but the headline number is overstated and the cross-model comparisons need tighter controls. the 4 major comments →

arxiv 2607.18839 v1 pith:NXSHD7XV submitted 2026-07-21 cs.CL

HPD-Parsing: Hierarchical Parallel Document Parsing

Shu Wei , Jingjing Wu , Lingshu Zhang , Qunyi Xie , Hao Zou , Le Xiang , Xu Fan , Yangliu Xu
show 5 more authors
Manhui Lin Xiaolong Ma Cheng Cui Tengyu Du YY
This is my paper
classification cs.CL
keywords document parsinghierarchical parallel decodingvision-language modelmulti-token predictioninference efficiencylayout analysisreading orderOCR
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 argues that full-page document parsing should not be generated token by token in one autoregressive stream, because layout is global but each block's content is local. It proposes Hierarchical Parallel Decoding: a main layout branch emits forking tokens that spawn concurrent content branches, each decoding one region, with shared-prefix KV cache reuse and progressive multi-token prediction inside every branch. On public benchmarks the resulting 1B model reaches 4,752 tokens per second, about 2.62× the fastest existing parser and 3.06× its own autoregressive baseline, while staying at competitive accuracy. A fair reader should take away that document parsing can be globally coordinated and locally parallelized instead of sequentially decoded.

Core claim

HPD-Parsing claims that a single full-page autoregressive trajectory is an unnecessarily sequential organization for document parsing. The model instead maintains one layout branch that predicts reading order and region boxes, and whenever it emits a <FORK> token the scheduler forks a content branch that decodes only that region, reusing visual and structural KV cache and attending only to shared context plus its own history. P-MTP is integrated into every branch to draft and verify multiple future tokens per step (average accepted length 6.6), shortening the intra-branch path. The paper reports 4,752 TPS at batch 512 on OmniDocBench v1.6, 3.06× its autoregressive baseline, with an overall a

What carries the argument

Hierarchical Parallel Decoding (HPD): the main decoding trajectory is split into a global layout branch and dynamically forked local content branches, joined by <FORK>/<CHILD> routing tokens; content branches share the visual and structural prefix KV cache but isolate their attention from other regions' text. P-MTP, a progressive multi-token prediction module with a lightweight residual MLP and depth-weighted loss, runs inside every branch to accept several tokens per step. Together they shorten the effective sequential decoding path along two axes: across branches through concurrency and within branches through multi-token acceptance.

Load-bearing premise

The load-bearing premise is that a block's content — text, tables, formulas — is grounded in its own visual region and barely depends on distant regions; if that locality fails, forked content branches lose needed context and accuracy drops.

What would settle it

Build or select a test set of documents where content genuinely crosses block boundaries — paragraphs continuing across columns, tables split across pages, headings whose text disambiguates later blocks — and compare HPD-Parsing against its own autoregressive baseline on text edit distance and reading-order edit distance. If the parallel model loses accuracy on such pages, the locality premise is the bottleneck.

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

If this is right

  • Full-page autoregressive generation is no longer required: globally coordinated, locally parallel decoding is a viable paradigm for unified document parsers.
  • The throughput advantage widens with document length, so text-dense pages benefit most — up to 18× fewer decoding steps in the longest output bucket.
  • A 1B unified parser can match or beat much larger general VLMs and remain competitive with pipeline parsers while being 2.6–3× faster than the fastest prior specialized parser.
  • Repetition and error propagation are confined to a single block, since content branches do not attend to other regions' decoded text.
  • Hierarchical parallel decoding extends beyond single-page parsing to other structured generation tasks such as key information extraction and multi-page document understanding.

Where Pith is reading between the lines

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

  • If the locality premise holds as a general property of documents, the same skeleton-plus-branches decomposition could be applied to any structured generation task with a global plan and localizable leaves — for example, long-document summarization by section, or structured extraction from multi-field forms.
  • Because HPD-Parsing consumes about 4,800 visual tokens per page, pairing it with visual-token compression should compound the gain; the paper already beats a compressed-input competitor despite a four-times larger input budget.
  • The accuracy of reading order and region assignment now rests almost entirely on the layout branch; a stress test that artificially corrupts layout predictions would reveal how much of the final accuracy depends on that single branch.
  • Comparing against a standard VLM on single-block pages would under-estimate the paradigm's risk; the interesting test is pages with cross-block text flow.

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 / 7 minor

Summary. The paper proposes HPD-Parsing, a 1B-parameter unified VLM document parser that replaces full-page autoregressive decoding with a hierarchical parallel decoding scheme: a main layout branch generates layout and routing tokens, and dynamically forks concurrent content branches that decode localized regions with shared-prefix KV-cache reuse. Progressive Multi-Token Prediction (P-MTP) is integrated within each branch to further reduce decoding steps, and training uses a three-stage adaptation with a difficulty-aware data curation pipeline. On OmniDocBench v1.6, the model reports 94.91 overall accuracy and 4,752 TPS, claiming 3.06× throughput over its own autoregressive baseline and 1.62× over DeepSeek-OCR-2; the abstract additionally claims 2.62× over the fastest existing model, a figure that is inconsistent with Table 2.

Significance. If the efficiency and accuracy results hold under controlled comparison, the paper makes a meaningful contribution: it restructures the decoding trajectory itself rather than only compressing visual tokens or predicting multiple tokens, and it is potentially orthogonal to those acceleration strategies. The scheduler algorithm, shared-prefix KV reuse, and staged adaptation recipe are concrete and reproducible in principle, and the authors provide model URLs. The qualitative analyses are informative. However, the central quantitative claim is not currently supported as stated: the headline ratio is misreported, the speed comparison is not controlled across systems, and the accuracy argument rests on a locality assumption that is not directly tested. The significance is therefore conditional.

major comments (4)
  1. [Abstract; §6.2, Table 2] The abstract and contribution list claim '2.62× the throughput of the fastest existing document parsing model'. Table 2 gives the relevant comparison: 4752.1/2932.1 = 1.62× TPS and 2.68/2.05 = 1.31× PPS vs DeepSeek-OCR-2. The 2.62× value is actually the PPS ratio vs the authors' own autoregressive baseline (2.68/1.02). §6.2 correctly reports 1.62×/1.31× vs DeepSeek-OCR-2. The headline must be corrected and all occurrences made consistent.
  2. [Table 2; §6.2] The efficiency comparison is not controlled. TPS/PPS numbers for competing systems are taken from external papers with no common hardware, vLLM version, batching policy, input/output length distribution, tokenizer, or sampling parameters, and no error bars or repeated-run variance are reported. Because 'fastest existing document parsing model' is a central claim, the authors should benchmark at least DeepSeek-OCR-2 and Unlimited OCR (the closest TPS competitors) on the same A800/vLLM setup with the same OmniDocBench requests and report mean ± std.
  3. [§3.3; §6.2] The accuracy argument rests on the locality premise from §3.1, but the architecture hard-wires it: each content branch attends only to shared visual context, its structural prefix, and its own generation history. The §6.2 statement that each region is interpreted 'with awareness of the entire page' is therefore not supported. No experiment checks pages with cross-block dependencies (paragraph continuation across regions, table cells referencing footnotes, or reading order that is content-dependent). The authors should add a targeted benchmark, or at least a stratified error analysis on OmniDocBench, to test whether missing cross-block textual context degrades accuracy.
  4. [Table 2; §3.4; §4] The reported improvement is for the full system (HPD + P-MTP + staged training), so the individual contribution of hierarchical forking to throughput is not isolated. Since P-MTP is prior work [22] and accounts for an average acceptance length of 6.6 tokens/step, readers cannot tell how much of the 3.06×/2.62× gain comes from branch concurrency versus multi-token prediction. An ablation with (a) full-page AR, (b) full-page AR + P-MTP, (c) HPD without P-MTP, and (d) HPD + P-MTP would substantiate the 'hierarchical parallel decoding' claim.
minor comments (7)
  1. [§1] 'HPD-Parsing achieves best throughout' should be 'best throughput'.
  2. [Figure 4 caption] 'Checkpointed HDP-Parsing' is a typo for 'HPD-Parsing'.
  3. [Algorithm 1, line 24] The token 'FO R KID' appears with a stray space; use '<FORK>' consistently.
  4. [§6.2] The sentence 'corresponding to improvements 2.62× and 3.06×, respectively' is correct but easy to misread; explicitly state that 2.62× is PPS vs the baseline and 1.62× is TPS vs DeepSeek-OCR-2.
  5. [Table 2] Several Avg Input Tokens entries are missing; report or justify the omissions.
  6. [§3.4] The average accepted length of 6.6 tokens/step is given without measurement details or variance; provide the protocol used to measure it.
  7. [References] References [5] and [30] are duplicates and should be merged.

Circularity Check

0 steps flagged

No constructed circularity; the throughput and accuracy claims are empirical measurements on external benchmarks. The only notable self-citation is P-MTP [22], used as a component without an in-paper ablation, but no equation or fitted value reduces to the claimed result.

full rationale

Walking the claimed derivation chain, the central claims are measurements rather than derived predictions. Throughput (4,752 TPS, 2.62x/3.06x) is an empirical result on OmniDocBench v1.6, and accuracy (94.91 overall) is benchmarked against external methods; neither is obtained by fitting a parameter and then 'predicting' the same quantity. The hierarchical parallel decoding speedup follows from the architecture (main layout branch forking concurrent content branches, shared-prefix KV cache reuse, and localized attention horizons) and is not forced by an equation. Equation (2) is a standard multi-token training objective with masks and weights; it is not used to produce the reported test-set numbers. The only self-citation of substance is P-MTP [22], described as 'our previous work' and adopted for intra-branch multi-token prediction, with the paper reporting an average accepted length of 6.6 tokens/step and referring to [22] for loss-weighting details. This is a verification/support concern — no independent artifact or ablation isolating the HPD contribution from P-MTP is provided — but it is not definitional circularity: the paper does not derive P-MTP's effectiveness from HPD-Parsing's own outputs. The §3.1 locality premise ('limited dependence on the detailed content of distant regions') is a strong modelling assumption with no targeted cross-block dependency test, but an unvalidated assumption is a correctness risk, not a circular step. No equation or fitted value in this paper is equivalent by construction to the claimed result.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

The central claims rest on one domain assumption (locality of block content), one self-referential component (P-MTP from ref [22]), and an unverified comparison protocol for baseline speeds. These are the entries a reader must accept before the 3× throughput conclusion bites.

free parameters (3)
  • P-MTP weighting coefficients (e.g., depth decay 0.8^k and consistency weights) = 0.8^k shown in Figure 3(b); other coefficients from ref [22]
    These control the progressive multi-token objective; the resulting acceptance length (6.6 tokens/step) drives the speedup, and the coefficients are inherited from the authors' prior work without independent verification.
  • Scheduler thresholds (KV occupancy τ, draft window K, concurrency cap N_max) = not reported
    Algorithm 1's throughput depends on these hand-chosen serving parameters; without values, replication of the 4,752 TPS number is impossible.
  • Difficulty-aware curation proportions (Easy/Medium/Hard mix, attribute balance, max refinement attempts N) = not reported
    The 100K Stage-2 and 600 Stage-3 datasets are selected by hand-set balancing ratios and refinement limits that affect the reported accuracy.
axioms (3)
  • domain assumption Document block content is local: regions show limited dependence on distant regions' content (§3.1, 'Overcoming this bottleneck...')
    Licenses parallel content branches; if false, accuracy collapses.
  • domain assumption P-MTP from the authors' prior work [22] transfers to this model with ~6.6 accepted tokens/step (§3.4)
    The efficiency claim depends on this unpublished, unverified component; no artifact is provided.
  • domain assumption Baseline efficiency numbers in Table 2 are comparable across papers
    The paper does not describe a common measurement protocol for rival systems.

pith-pipeline@v1.3.0-alltime-deepseek · 15742 in / 16248 out tokens · 140047 ms · 2026-08-01T14:09:39.756058+00:00 · methodology

0 comments
read the original abstract

Efficient teamwork typically combines global coordination with parallel execution, a principle not yet fully reflected in unified Vision-Language Model (VLM)-based document parsers. Existing unified parsers process an entire page jointly but generate its output through a single token-by-token autoregressive trajectory, creating a sequential bottleneck that grows with document length. Such full-page sequential generation overlooks a key property of document parsing: layout must be analyzed globally, whereas block content can be parsed in parallel. Based on this observation, we introduce HPD-Parsing, which replaces full-page autoregressive generation with a Hierarchical Parallel Decoding paradigm. A main layout branch organizes the overall document structure and dynamically assigns block-level content decoding to concurrent branches, while progressive multi-token prediction (P-MTP) further reduces the decoding steps within each branch. Experiments on public benchmarks show that HPD-Parsing achieves 4,752 tokens per second, delivering $2.62\times$ the throughput of the fastest existing document parsing model and $3.06\times$ that of the vanilla autoregressive baseline, while maintaining competitive parsing accuracy. These results establish hierarchical parallel decoding as an effective alternative to full-page autoregressive generation, opening a new direction for efficient unified document parsing.

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

36 extracted references · 18 linked inside Pith

  1. [1]

    Paddleocr-vl: Boosting multilingual document parsing via a 0.9 b ultra-compact vision-language model

    Cheng Cui, Ting Sun, Suyin Liang, Tingquan Gao, Zelun Zhang, Jiaxuan Liu, Xueqing Wang, Changda Zhou, Hongen Liu, Manhui Lin, et al. Paddleocr-vl: Boosting multilingual document parsing via a 0.9 b ultra-compact vision-language model. arXiv preprint arXiv:2510.14528, 2025

    arXiv 2025

  2. [2]

    Paddleocr-vl-1.5: Towards a multi-task 0.9 b vlm for robust in-the-wild document parsing

    Cheng Cui, Ting Sun, Suyin Liang, Tingquan Gao, Zelun Zhang, Jiaxuan Liu, Xueqing Wang, Changda Zhou, Hongen Liu, Manhui Lin, et al. Paddleocr-vl-1.5: Towards a multi-task 0.9 b vlm for robust in-the-wild document parsing. arXiv preprint arXiv:2601.21957, 2026

    Pith/arXiv arXiv 2026

  3. [3]

    Mineru: An open-source solution for precise document content extraction

    Bin Wang, Chao Xu, Xiaomeng Zhao, Linke Ouyang, Fan Wu, Zhiyuan Zhao, Rui Xu, Kaiwen Liu, Yuan Qu, Fukai Shang, et al. Mineru: An open-source solution for precise document content extraction. arXiv preprint arXiv:2409.18839, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    Bin Wang, Tianyao He, Linke Ouyang, Fan Wu, Zhiyuan Zhao, Tao Chu, Yuan Qu, Zhenjiang Jin, Weijun Zeng, Ziyang Miao, et al. Mineru2. 5-pro: Pushing the limits of data-centric document parsing at scale. arXiv preprint arXiv:2604.04771, 2026

    Pith/arXiv arXiv 2026

  5. [6]

    Youtu-parsing: Perception, structuring and recognition via high-parallelism decoding

    Kun Yin, Yunfei Wu, Bing Liu, Zhongpeng Cai, Xiaotian Li, Huang Chen, Xin Li, Haoyu Cao, Yinsong Liu, Deqiang Jiang, et al. Youtu-parsing: Perception, structuring and recognition via high-parallelism decoding. arXiv preprint arXiv:2601.20430, 2026

    Pith/arXiv arXiv 2026

  6. [7]

    Dolphin: Document image parsing via heterogeneous anchor prompting

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

    2025

  7. [8]

    Dolphin-v2: Universal document parsing via scalable anchor prompt- ing

    Hao Feng, Wei Shi, Ke Zhang, Xiang Fei, Lei Liao, Dingkang Yang, Yongkun Du, Xuecheng Wu, Jingqun Tang, Yang Liu, et al. Dolphin-v2: Universal document parsing via scalable anchor prompt- ing. arXiv preprint arXiv:2602.05384, 2026

    arXiv 2026

  8. [9]

    Glm-ocr technical report

    Shuaiqi Duan, Yadong Xue, Weihan Wang, Zhe Su, Huan Liu, Sheng Yang, Guobing Gan, Guo Wang, Zihan Wang, Shengdong Yan, et al. Glm-ocr technical report. arXiv preprint arXiv:2603.10910, 2026

    arXiv 2026

  9. [10]

    Firered-ocr technical report

    Hao Wu, Haoran Lou, Xinyue Li, Zuodong Zhong, Zhaojun Sun, Phellon Chen, Xuanhe Zhou, Kai Zuo, Yibo Chen, Xu Tang, et al. Firered-ocr technical report. arXiv preprint arXiv:2603.01840, 2026

    arXiv 2026

  10. [11]

    Qianfan-ocr: A unified end-to-end model for document intelligence

    Daxiang Dong, Mingming Zheng, Dong Xu, Chunhua Luo, Bairong Zhuang, Yuxuan Li, Ruoyun He, Haoran Wang, Wenyu Zhang, Wenbo Wang, et al. Qianfan-ocr: A unified end-to-end model for document intelligence. arXiv preprint arXiv:2603.13398, 2026

    arXiv 2026

  11. [12]

    Hunyuanocr technical report.arXiv preprint arXiv:2511.19575, 2025

    Hunyuan Vision Team, Pengyuan Lyu, Xingyu Wan, Gengluo Li, Shangpin Peng, Weinong Wang, Liang Wu, Huawen Shen, Yu Zhou, Canhui Tang, et al. Hunyuanocr technical report.arXiv preprint arXiv:2511.19575, 2025

    arXiv 2025

  12. [13]

    Yumeng Li, Guang Yang, Hao Liu, Bowen Wang, and Colin Zhang. dots. ocr: Multilingual document layout parsing in a single vision-language model. arXiv preprint arXiv:2512.02498, 2025

    arXiv 2025

  13. [14]

    Ocean-ocr: Towards general ocr application via a vision-language model

    Song Chen, Xinyu Guo, Yadong Li, Tao Zhang, Mingan Lin, Dongdong Kuang, Youwei Zhang, Lingfeng Ming, Fengyu Zhang, Yuran Wang, et al. Ocean-ocr: Towards general ocr application via a vision-language model. arXiv preprint arXiv:2501.15558, 2025

    Pith/arXiv arXiv 2025

  14. [15]

    Nougat: Neu- ral optical understanding for academic documents

    Lukas Blecher, Guillem Cucurull Preixens, Thomas Scialom, and Robert Stojnic. Nougat: Neu- ral optical understanding for academic documents. In International Conference on Learning Representations, volume 2024, pages 37646–37663, 2024

    2024

  15. [16]

    Lightonocr: A 1b end-to-end multilin- gual vision-language model for state-of-the-art ocr

    Said Taghadouini, Adrien Cavaillès, and Baptiste Aubertin. Lightonocr: A 1b end-to-end multilin- gual vision-language model for state-of-the-art ocr. arXiv preprint arXiv:2601.14251, 2026. 19

    Pith/arXiv arXiv 2026

  16. [17]

    Deepseek-ocr: Contexts optical compression

    Haoran Wei, Yaofeng Sun, and Yukun Li. Deepseek-ocr: Contexts optical compression. arXiv preprint arXiv:2510.18234, 2025

    Pith/arXiv arXiv 2025

  17. [18]

    Deepseek-ocr 2: Visual causal flow

    Haoran Wei, Yaofeng Sun, and Yukun Li. Deepseek-ocr 2: Visual causal flow. arXiv preprint arXiv:2601.20552, 2026

    arXiv 2026

  18. [19]

    Logics-parsing technical report

    Xiangyang Chen, Shuzhao Li, Xiuwen Zhu, Yongfan Chen, Fan Yang, Cheng Fang, Lin Qu, Xiaoxiao Xu, Hu Wei, and Minggang Wu. Logics-parsing technical report. arXiv preprint arXiv:2509.19760, 2025

    arXiv 2025

  19. [20]

    Unlimited ocr works

    Youyang Yin, Huanhuan Liu, Qunyi Xie, Chaorun Liu, Shiqi Yang, Shaohua Wang, Zhanlong Liu, Hao Zou, Jinyue Chen, Shu Wei, et al. Unlimited ocr works. arXiv preprint arXiv:2606.23050, 2026

    Pith/arXiv arXiv 2026

  20. [21]

    Hunyuanocr-1.5: Making lightweight ocr vlms faster and better

    Gengluo Li, Xingyu Wan, Shangpin Peng, Weinong Wang, Hao Feng, Yongkun Du, Binghong Wu, Zheng Ruan, Zhiqiong Lu, Liang Wu, et al. Hunyuanocr-1.5: Making lightweight ocr vlms faster and better. arXiv preprint arXiv:2607.04884, 2026

    Pith/arXiv arXiv 2026

  21. [22]

    P-mtp: Efficient document parsing via multi-token prediction with progressive depth scaling

    Le Xiang, Chenxi Zhai, Shu Wei, Jingjing Wu, Qunyi Xie, Xiao Tan, Kunbin Chen, and Wei He. P-mtp: Efficient document parsing via multi-token prediction with progressive depth scaling. arXiv preprint arXiv:2606.24447, 2026

    Pith/arXiv arXiv 2026

  22. [23]

    Structextv3: An efficient vision-language model for text-rich image perception, comprehension, and beyond

    Pengyuan Lyu, Yulin Li, Hao Zhou, Weihong Ma, Xingyu Wan, Qunyi Xie, Liang Wu, Chengquan Zhang, Kun Yao, Errui Ding, et al. Structextv3: An efficient vision-language model for text-rich image perception, comprehension, and beyond. arXiv preprint arXiv:2405.21013, 2024

    Pith/arXiv arXiv 2024

  23. [24]

    Internvl3

    Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, et al. Internvl3. 5: Advancing open-source multimodal models in versatility, reasoning, and efficiency.arXiv preprint arXiv:2508.18265, 2025

    Pith/arXiv arXiv 2025

  24. [25]

    Kimi Team, Tongtong Bai, Yifan Bai, Yiping Bao, SH Cai, Yuan Cao, Y Charles, HS Che, Cheng Chen, Guanduo Chen, et al. Kimi k2. 5: Visual agentic intelligence. arXiv preprint arXiv:2602.02276, 2026

    Pith/arXiv arXiv 2026

  25. [26]

    Openai gpt-5 system card

    Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaughlin, Aiden Low, AJ Ostrow, Akhila Ananthram, et al. Openai gpt-5 system card. arXiv preprint arXiv:2601.03267, 2025

    Pith/arXiv arXiv 2025

  26. [27]

    Qwen3-vl technical report

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report. arXiv preprint arXiv:2511.21631, 2025

    Pith/arXiv arXiv 2025

  27. [28]

    Gemini 3 Pro Model Card

    Google. Gemini 3 Pro Model Card. https://storage.googleapis.com/deepmind-media/M odel-Cards/Gemini-3-Pro-Model-Card.pdf, 2026

    2026

  28. [29]

    Ovis2.5 Technical Report

    Shiyin Lu, Yang Li, Yu Xia, Yuwei Hu, Shanshan Zhao, Yanqing Ma, Zhichao Wei, Yinglun Li, Lunhao Duan, Jianshan Zhao, Yuxuan Han, Haijun Li, et al. Ovis2.5 Technical Report. arXiv:2508.11737, 2025

    Pith/arXiv arXiv 2025

  29. [30]

    MonkeyOCR: Document parsing with a structure-recognition-relation triplet paradigm

    Zhang Li, Yuliang Liu, Qiang Liu, Zhiyin Ma, Ziyang Zhang, Shuo Zhang, Zidun Guo, Jiarui Zhang, Xinyu Wang, and Xiang Bai. MonkeyOCR: Document parsing with a structure-recognition-relation triplet paradigm. arXiv preprint arXiv:2506.05218, 2025

    arXiv 2025

  30. [31]

    Unirec-0.1b: Unified text and formula recognition with 0.1b parameters

    Yongkun Du, Zhineng Chen, Yazhen Xie, Weikang Bai, Hao Feng, Wei Shi, Yuchen Su, Can Huang, and Yu-Gang Jiang. Unirec-0.1b: Unified text and formula recognition with 0.1b parameters. arXiv preprint arXiv:2512.21095, 2025

    Pith/arXiv arXiv 2025

  31. [32]

    MinerU2.5: A decoupled vision-language model for efficient high-resolution document parsing

    Junbo Niu, Zheng Liu, Zhuangcheng Gu, Bin Wang, Linke Ouyang, Zhiyuan Zhao, Tao Chu, Tianyao He, Fan Wu, Qintong Zhang, et al. MinerU2.5: A decoupled vision-language model for efficient high-resolution document parsing. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics–Industry Track, 2025. 20

    2025

  32. [33]

    Paddleocr-vl-1.6: Expanding the frontier of document parsing with under-optimized region refinement and progressive post-training

    Zelun Zhang, Hongen Liu, Suyin Liang, Yubo Zhang, Yiqing Xiang, Jiaxuan Liu, Ting Sun, Manhui Lin, Yue Zhang, Changda Zhou, et al. Paddleocr-vl-1.6: Expanding the frontier of document parsing with under-optimized region refinement and progressive post-training. arXiv preprint arXiv:2606.03264, 2026

    Pith/arXiv arXiv 2026

  33. [34]

    POINTS-Reader: Distillation-free adaptation of vision-language models for document conversion

    Yuan Liu, Zhongyin Zhao, Le Tian, Haicheng Wang, Xubing Ye, Yangxiu You, Zilin Yu, Chuhan Wu, Zhou Xiao, Yang Yu, et al. POINTS-Reader: Distillation-free adaptation of vision-language models for document conversion. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2025

    2025

  34. [35]

    Nanonets-OCR-S: A model for transforming documents into structured markdown with intelligent content recognition and semantic tagging, 2025

    Souvik Mandal, Ashish Talewar, Paras Ahuja, and Prathamesh Juvatkar. Nanonets-OCR-S: A model for transforming documents into structured markdown with intelligent content recognition and semantic tagging, 2025. URLhttps://huggingface.co/nanonets/Nanonets-OCR-s

    2025

  35. [36]

    olmocr: Unlocking trillions of tokens in pdfs with vision language models

    Jake Poznanski, Aman Rangapur, Jon Borchardt, Jason Dunkelberger, Regan Huff, Daniel Lin, Christopher Wilhelm, Kyle Lo, and Luca Soldaini. olmocr: Unlocking trillions of tokens in pdfs with vision language models. arXiv preprint arXiv:2502.18443, 2025

    arXiv 2025

  36. [37]

    OCRVerse: Towards holistic OCR in end-to-end vision- language models

    Yufeng Zhong, Lei Chen, Xuanle Zhao, Wenkang Han, Liming Zheng, Jing Huang, Deyang Jiang, Yilin Cao, Lin Ma, and Zhixiong Zeng. OCRVerse: Towards holistic OCR in end-to-end vision- language models. arXiv preprint arXiv:2601.21639, 2026. 21

    arXiv 2026