Pith. sign in

REVIEW 3 major objections 5 minor 53 references

OvisOCR2 is a 0.8B end-to-end document parser that the paper reports as scoring 96.58 on OmniDocBench v1.6 and 75.06 Avg3 on PureDocBench, ahead of pipeline-based systems on both public 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

2026-08-02 04:35 UTC pith:7EXYNIU6

load-bearing objection A real system paper with plausible design and strong numbers, but the headline SOTA claim rests on missing decontamination evidence and metric-aligned RL rewards. the 3 major comments →

arxiv 2607.13639 v1 pith:7EXYNIU6 submitted 2026-07-15 cs.CV cs.AI

OvisOCR2 Technical Report

Shiyin Lu , Yinglun Li , Yu Xia , Yuhui Chen , An-Yang Ji , Jun-Peng Jiang , Qing-Guo Chen , Jianshan Zhao
show 5 more authors
En Lin Haijun Li Cheng Qin Zhao Xu Weihua Luo
This is my paper
classification cs.CV cs.AI
keywords document parsingend-to-end OCRMarkdown generationsynthetic datareinforcement learningon-policy distillationtable 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 paper claims that a compact 0.8B model, reading a page image and emitting Markdown in one pass, can outperform multi-stage pipeline parsers on mainstream document parsing leaderboards. It credits three ingredients: a data engine that merges filtered real-document annotations with synthetic pages whose images and Markdown labels derive from the same HTML source; reinforcement learning on a larger 4B branch using rewards for text fidelity, formula rendering, and table structure; and on-policy distillation of that aligned policy into the 0.8B student. If the numbers hold, end-to-end parsing becomes the leading approach at a deployment-friendly scale, potentially simplifying systems and cutting serving costs while removing cascading layout errors.

Core claim

The paper introduces OvisOCR2, a 0.8B end-to-end model that converts a document page image into a Markdown representation in natural reading order. The central claim is that this single-pass model achieves state-of-the-art results on two public benchmarks — an overall score of 96.58 on OmniDocBench v1.6 and an Avg3 score of 75.06 on PureDocBench — placing an end-to-end model at the top of leaderboards previously dominated by pipeline approaches. The report argues that the result comes from a source-of-truth synthetic data pipeline, a multi-component RL reward that targets structural errors (tables, formulas, long pages), and a distillation scheme that transfers a 4B teacher's reward-aligned

What carries the argument

The central mechanism is the source-of-truth synthetic data pipeline: one HTML source renders both the document image and the Markdown target, making labels deterministic and bounding boxes exact, while preserving realistic typography and reading order. Training is carried by GRPO reinforcement learning with a composite reward — normalized edit distance for text, CDM for formulas, TEDS for tables — followed by on-policy distillation using student top-k reverse KL, which transfers the teacher's preferences on tables, formulas, and dense layouts rather than just its output tokens.

Load-bearing premise

The benchmark pages and their annotations never leaked into the model's training data through the large-scale real-image ingest or the synthetic template library, since no de-duplication against the public benchmarks is described — if they leaked, the state-of-the-art scores are inflated.

What would settle it

Compare document-image hashes between the training corpus (real and synthetic) and the pages of OmniDocBench v1.6 and PureDocBench, then re-score OvisOCR2 on the non-overlapping subset; if the scores drop materially, the leaderboard claim is contaminated. Alternatively, run the model on a withheld, independently sourced set of camera-captured and heavily degraded pages.

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

If this is right

  • If the reported scores hold, end-to-end parsing at 0.8B parameters can displace pipeline systems on mainstream parsing benchmarks, simplifying deployment and removing error accumulation across layout, recognition, and merging stages.
  • State-of-the-art-level parsing at 0.8B would enable high-quality document understanding on resource-constrained devices and at substantially lower serving cost than large general-purpose vision-language models.
  • The source-of-truth synthetic principle — generating image and label from the same HTML — offers a transferable recipe for producing clean, controllable training data for other structured-output tasks like forms understanding or code-from-layout.
  • On heavily degraded real-world images, the paper's own results show the model trails large general vision-language models on the Real track of PureDocBench, so the lead is strongest on clean-to-moderately degraded documents and handwriting and complex-table subsets.
  • The two-branch training pattern — RL on a larger teacher, then on-policy distillation into a small student — may extend to other long-output structured tasks where direct RL on compact models is unstable.

Where Pith is reading between the lines

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

  • Because the real-data pipeline ingests large-scale document images with no stated de-duplication against the public benchmarks, the absolute leaderboard numbers are best read as plausible but unverified; an independent overlap check would settle it.
  • The synthetic pipeline's HTML-to-image rendering resembles the generation style of some public benchmarks, so the clean-track scores may overestimate generalization to entirely unseen document styles; camera-captured, low-quality real-world pages are the stronger test.
  • A testable extension is ablating each reward component (text, formula, table) to measure its marginal contribution to the leaderboard gaps, which would say whether the gains are driven mainly by table and formula rewards.
  • The top-k reverse-KL distillation mechanism suggests a general recipe for small-model structured generation: do RL at larger scale, then transfer via mode-seeking distillation rather than direct RL on the compact model.

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. This technical report presents OvisOCR2, a 0.8B parameter end-to-end document parser trained from Qwen3.5-0.8B. The training data are built by a two-pipeline data engine: real document pages are parsed by specialized OCR systems and normalized/filtered into Markdown, while synthetic pages are generated from HTML sources so that image and Markdown target are derived from the same source. The training recipe proceeds from supervised fine-tuning through GRPO reinforcement learning on a 4B branch, on-policy distillation into the 0.8B student, and weighted model fusion. The paper reports state-of-the-art results on OmniDocBench v1.6 (overall 96.58), PureDocBench (Avg3 75.06), and an author-built in-house benchmark of more than 1,000 pages, with additional breakdowns for handwriting and complex-table subsets. The model is released on HuggingFace.

Significance. If the reported scores are uncontaminated, the paper's central result is significant: a compact 0.8B single-pass model outperforming pipeline-based systems and much larger general VLMs on two public document-parsing benchmarks would be a meaningful advance for end-to-end document parsing. The report is also strong on system description: the source-aligned synthetic pipeline, the rule-based normalization of real annotations, the hard-sample mining procedure, and the concrete OPD/RL infrastructure choices (top-k reverse KL, object-store tensor referencing, common-prefix masking) are detailed and plausible. The main weakness is external validity: the report does not state that public benchmark pages or near-duplicates were excluded from training, the RL rewards are the same metrics as the OmniDocBench overall score, and the in-house benchmark is unreleased. These issues must be addressed before the SOTA claim can be taken as evidence of generalization rather than benchmark fitting.

major comments (3)
  1. [§2.1.1, §2.2.1, §4.1] The headline SOTA claim assumes OmniDocBench and PureDocBench pages are outside the training distribution. The real-world pipeline filters duplicated source images only within the candidate set, and the synthetic pipeline mines hard samples from 'failure cases identified through multi-faceted assessments' without specifying whether those failure cases include benchmark pages, benchmark HTML, or derived templates. No decontamination step is described anywhere in §2 or §4. Please add an explicit statement and evidence that benchmark pages, their rendered HTML, and near-duplicates were excluded from SFT, RL, and synthetic template mining; for PureDocBench, which is source-traceable, report the overlap check directly.
  2. [§3.3.2, Table 1] The RL reward components are normalized edit distance, CDM, and TEDS — exactly the metrics that compose the OmniDocBench overall score. This is not circular because evaluation is external, but it raises a metric-fitting risk, especially if any training pages resemble benchmark pages. Please report the distributional relationship between the RL training set and the public benchmarks, include an out-of-distribution holdout, and provide an SFT-only vs. RL comparison on OmniDocBench and PureDocBench so readers can see whether the gains generalize or are localized to the reward metrics.
  3. [§4.3, Tables 4–7] The in-house benchmark is author-built, follows the OmniDocBench protocol, and is not released; the synthetic template families are also not released. Consequently, readers cannot verify that the more than 1,000 in-house pages do not overlap with the synthetic templates generated from mined hard samples. Since this benchmark is presented as evidence of generalization beyond the public benchmarks, please release the benchmark and annotations, or provide an overlap analysis between the in-house pages and the synthetic training template families, together with a public-benchmark decontamination report.
minor comments (5)
  1. [§4.1] The overall score is said to be computed by averaging 'the text score converted from normalized edit distance, formula CDM, and table TEDS', but the exact conversion from normalized edit distance to a text score is not given. Please state the formula.
  2. [Figure 4] The validation-set table TEDS is plotted, but the validation set is not described. State its composition, size, and whether it was used for any training decisions.
  3. [§3.2–§3.5] Several training hyperparameters are omitted, including the real:synthetic data mixture ratio, the OPD top-k support size, the on-policy hard-case filtering thresholds, and the model-fusion weights. If these are intentionally withheld, say so; otherwise list them in an appendix.
  4. [Tables 2 and 3] All reported scores are single-run with no variance or confidence intervals. Given the small margins, e.g., the TEDS tie at 94.76 in Table 2, providing repeated evaluations or releasing the evaluation harness would substantially strengthen the comparison.
  5. [Appendix A] The qualitative examples do not state how pages were selected. Please describe the selection criteria to avoid the impression of cherry-picking, especially for the Unlimited-OCR outputs in Figures A.1 and A.2.

Circularity Check

0 steps flagged

No significant circularity: the SOTA claims rest on external benchmarks and are not reduced to the paper's own inputs.

full rationale

The central claims are empirical scores on public benchmarks (OmniDocBench v1.6, PureDocBench) that this paper does not construct or annotate, so the headline results have independent grounding. The training recipe (SFT, RL, OPD, fusion) is described operationally; no equation defines a reported benchmark score in terms of the model's own fitted parameters or training labels. The RL reward components in Sec. 3.3.2 (normalized edit distance, CDM, TEDS) do match the composition of the OmniDocBench overall score in Sec. 4.1, which is a genuine caveat about metric-alignment and potential overfitting, but it is a training-objective choice rather than a derivation of the evaluation result: the benchmark pages are external and the paper does not fit a parameter to a subset of benchmark data and then rename that fit as a prediction. No load-bearing argument rests on self-citation; the same-team references [10] and [27] appear only as examples/baselines. The absence of an explicit decontamination statement (no deduplication of SFT/RL data against OmniDocBench/PureDocBench, and the unreleased in-house benchmark in Sec. 4.3) is a data-hygiene and external-validity risk, not a circularity in the derivation chain. Accordingly, no enumerated circular step is present.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 0 invented entities

The paper's load-bearing infrastructure is not a mathematical derivation but an empirical training pipeline. It leans on numerous undisclosed hand-chosen hyperparameters (data mixture, fusion weights, OPD k, filtering thresholds) and on domain assumptions about label quality, synthetic realism, benchmark integrity, and metric validity. No new physical or mathematical entities are introduced.

free parameters (8)
  • Real:synthetic data mixture ratio = not disclosed
    The mix of real-world filtered annotations and synthetic HTML pages in SFT/RL is central to performance and is never quantified (Sec. 3.2).
  • RL reward component weighting = equal average over available text/table/formula components
    The page reward is a designed aggregation; any other weighting would change the optimization target (Sec. 3.3.2).
  • OPD top-k support size k = not disclosed
    The top-k reverse-KL distillation restricts teacher supervision to the student's top-k tokens; k controls the distillation bias/efficiency trade-off (Sec. 3.4.2).
  • Training epochs / fraction = 0.8B SFT: 2 epochs; 4B SFT: 0.2 epoch
    Chosen to balance cost and quality, with no ablation shown (Sec. 3.2).
  • Max sequence length and image-resolution budget = 16K tokens; dynamic resolution budget unspecified
    Controls which long/tall pages the model can parse at all; the resolution schedule is not detailed (Sec. 3.2).
  • Model fusion weights = not disclosed
    Weighted parameter averaging of candidate variants; the weights and number of variants are not reported (Sec. 3.5).
  • On-policy hard-case filtering thresholds = not disclosed
    RL training keeps only pages where the current policy occasionally beats its average; operational thresholds are omitted (Sec. 3.3.1).
  • Spot-checking subset removal criteria = manual/conservative, no quantitative threshold
    Real-data subsets are kept or removed by manual inspection; this is a human-chosen selection rule that shapes the corpus (Sec. 2.1.2).
axioms (6)
  • domain assumption Filtered outputs of PaddleOCR-VL-1.5 and MinerU2.5-Pro are accurate enough to serve as document-to-Markdown training labels after deterministic normalization and spot-checking.
    Central to the real-world data pipeline; if downstream recognizer errors survive filtering, the model inherits them (Sec. 2.1).
  • domain assumption HTML-rendered pages via Playwright are a valid proxy for real document images for training.
    The synthetic pipeline uses the same HTML source for image and label; realism is only visually inspected (Sec. 2.2.4).
  • domain assumption The OmniDocBench/PureDocBench public scores transcribed from leaderboards and technical reports are accurate and comparable at the time of submission.
    The headline SOTA claim rests on these external numbers (Sec. 4.1, 4.2).
  • domain assumption The evaluation benchmark pages are disjoint from the training corpus.
    No contamination check is reported; the data engine ingests large-scale real pages (Sec. 2.1, 4.1).
  • domain assumption The in-house benchmark of over 1,000 pages is representative, fairly scored, and disjoint from the synthetic training templates.
    It is author-constructed and not released, so this cannot be independently checked (Sec. 4.3).
  • domain assumption CDM, TEDS, and text-edit-based rewards measure the parsing qualities that matter for downstream use.
    These metrics are also the benchmark metrics, so optimizing them may not improve other qualities such as semantic fidelity or authenticity (Sec. 3.3.2, 4.1).

pith-pipeline@v1.3.0-alltime-deepseek · 28924 in / 14714 out tokens · 139705 ms · 2026-08-02T04:35:25.940566+00:00 · methodology

0 comments
read the original abstract

We introduce OvisOCR2, a 0.8B document parsing model. OvisOCR2 is designed as an end-to-end parser: given a document page image, it generates a Markdown representation in natural reading order, covering text, formulas, tables, and visual regions. We build a data engine that combines filtered real-document annotations with synthetic pages whose rendered images and Markdown targets are derived from the same HTML source. The training recipe includes supervised fine-tuning, reinforcement learning on a 4B branch with a multi-component reward design, on-policy distillation into the 0.8B model, and model fusion. On OmniDocBench v1.6, OvisOCR2 achieves a state-of-the-art overall score of 96.58, placing an end-to-end model at the top of this leaderboard previously dominated by pipeline methods and highlighting the potential of end-to-end document parsing. On PureDocBench, OvisOCR2 also achieves the highest Avg3 score of 75.06. Beyond these two public benchmarks, we evaluate OvisOCR2 on an in-house benchmark designed to cover a broader set of long-tail and challenging scenarios. OvisOCR2 obtains the best overall performance among the compared methods, providing further evidence of its generalization and robustness. OvisOCR2 is available at https://huggingface.co/ATH-MaaS/OvisOCR2.

Figures

Figures reproduced from arXiv: 2607.13639 by An-Yang Ji, Cheng Qin, En Lin, Haijun Li, Jianshan Zhao, Jun-Peng Jiang, Qing-Guo Chen, Shiyin Lu, Weihua Luo, Yinglun Li, Yuhui Chen, Yu Xia, Zhao Xu.

Figure 1
Figure 1. Figure 1: Performance of OvisOCR2 on OmniDocBench v1.6 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Architecture of the data engine itself a key part of the system design. No single data source is sufficient: real-world documents provide natural layouts and varied visual conditions, but their annotations are noisy and difficult to standardize; synthetic documents offer controllable training targets and scalable coverage, but may suffer from limited realism if synthesized naively. To address this, we buil… view at source ↗
Figure 3
Figure 3. Figure 3: Two-branch training of OvisOCR2. The 4B branch produces an RL-aligned teacher, [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Training stability comparison between 4B RL and 0.8B RL. Panel (a) reports actor KL [PITH_FULL_IMAGE:figures/full_fig_p011_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

53 extracted references · 18 linked inside Pith

  1. [1]

    OmniDocBench: Benchmarking diverse PDF document parsing with comprehensive annotations

    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, Zhenxi- ang Li, Chao Xu, Bo Zhang, Botian Shi, Zhongying Tu, and Conghui He. OmniDocBench: Benchmarking diverse PDF document parsing with comprehensive annotations. InProceed- ings of the IEEE/CVF...

    2025

  2. [2]

    How far is document parsing from solved? PureDocBench: A source-traceable benchmark across clean, degraded, and real-world settings.arXiv preprint arXiv:2605.07492, 2026

    Zhiheng Li, Zongyang Ma, Jiaxian Chen, Jianing Zhang, Zhaolong Su, Yutong Zhang, Zhiyin Yu, Ruiqi Liu, Xiaolei Lv, Bo Li, Jun Gao, Ziqi Zhang, Chunfeng Yuan, Bing Li, and Weiming Hu. How far is document parsing from solved? PureDocBench: A source-traceable benchmark across clean, degraded, and real-world settings.arXiv preprint arXiv:2605.07492, 2026

    Pith/arXiv arXiv 2026

  3. [3]

    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

    Zelun Zhang, Hongen Liu, Suyin Liang, Yubo Zhang, Yiqing Xiang, Jiaxuan Liu, Ting Sun, Manhui Lin, Yue Zhang, Changda Zhou, Tingquan Gao, Cheng Cui, Yi Liu, Dianhai Yu, and Yanjun Ma. 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

  4. [4]

    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, Zhenjiang Jin, Guang Liang, Rui Zhang, Wenzheng Zhang, Yuan Qu, Zhifei Ren, Yuefeng Sun, Zirui Tang, Boyu Niu, Yuanhong Zheng, Dongsheng Ma, Ziyang Miao, Hejun Dong, Siyi Qian, Junyuan Zhang, Fangdong 16 Wang, Jingzhou Chen, Xiaomeng Zha...

    2026

  5. [5]

    GLM-OCR technical report.arXiv preprint arXiv:2603.10910, 2026

    Shuaiqi Duan, Yadong Xue, Weihan Wang, Zhe Su, Huan Liu, Sheng Yang, Guobing Gan, Guo Wang, Zihan Wang, Shengdong Yan, Dexin Jin, Yuxuan Zhang, Guohong Wen, Yanfeng Wang, Yutao Zhang, Xiaohan Zhang, Wenyi Hong, Yukuo Cen, Da Yin, Bin Chen, Wenmeng Yu, Xiaotao Gu, and Jie Tang. GLM-OCR technical report.arXiv preprint arXiv:2603.10910, 2026

    arXiv 2026

  6. [6]

    MinerU2.5-Pro: Pushing the limits of data-centric document parsing at scale.arXiv preprint arXiv:2604.04771, 2026

    Bin Wang, Tianyao He, Linke Ouyang, Fan Wu, Zhiyuan Zhao, Tao Chu, Yuan Qu, Zhen- jiang Jin, Weijun Zeng, Ziyang Miao, Bangrui Xu, Junbo Niu, Mengzhang Cai, Jiantao Qiu, Qintong Zhang, Dongsheng Ma, Yuefeng Sun, Hejun Dong, Wenzheng Zhang, Jutao Xiao, Jiayong Shi, Pengyu Liao, Xiaomeng Zhao, Huaping Zhong, Liqun Wei, Jing Yu, Jie Yang, Wei Li, Shasha Wang...

    Pith/arXiv arXiv 2026

  7. [7]

    OmniDocBench public leaderboard

    OpenDataLab. OmniDocBench public leaderboard. https://opendatalab.com/ omnidocbench, 2026. Accessed June 29, 2026

    2026

  8. [8]

    HunyuanOCR-1.5: Making lightweight OCR VLMs faster and better.arXiv preprint arXiv:2607.04884, 2026

    Gengluo Li, Xingyu Wan, Shangpin Peng, Weinong Wang, Hao Feng, Yongkun Du, Binghong Wu, Zheng Ruan, Zhiqiong Lu, Liang Wu, Pengyuan Lyu, Huawen Shen, Zibin Lin, Shijing Hu, Jieneng Yang, Hongbing Wen, Guanghua Yu, Hong Liu, Bochao Wang, Can Ma, Han Hu, Chengquan Zhang, and Yu Zhou. HunyuanOCR-1.5: Making lightweight OCR VLMs faster and better.arXiv prepri...

    Pith/arXiv arXiv 2026

  9. [9]

    Unlimited OCR works.arXiv preprint arXiv:2606.23050, 2026

    Youyang Yin, Huanhuan Liu, YY, Qunyi Xie, Chaorun Liu, Shiqi Yang, Shaohua Wang, Zhanlong Liu, Hao Zou, Jinyue Chen, Shu Wei, Jingjing Wu, Mingxin Huang, Zhen Wu, Guibin Wang, Tengyu Du, and Lei Jia. Unlimited OCR works.arXiv preprint arXiv:2606.23050, 2026

    Pith/arXiv arXiv 2026

  10. [10]

    OvisOCR: End-to-end document parsing via aligning specialized perception with general reasoning

    Jun-Peng Jiang, Shiyin Lu, An-Yang Ji, Yinglun Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, De-Chuan Zhan, and Han-Jia Ye. OvisOCR: End-to-end document parsing via aligning specialized perception with general reasoning. InProceedings of the 43rd International Conference on Machine Learning, volume 306 ofProceedings of Machine Learning Research, Se...

    2026

  11. [11]

    Qwen3.5: Towards native multimodal agents

    Qwen Team. Qwen3.5: Towards native multimodal agents. https://qwen.ai/blog?id= qwen3.5, 2026

    2026

  12. [12]

    PaddleOCR-VL-1.5: Towards a multi-task 0.9B VLM for robust in-the- wild document parsing.arXiv preprint arXiv:2601.21957, 2026

    Cheng Cui, Ting Sun, Suyin Liang, Tingquan Gao, Zelun Zhang, Jiaxuan Liu, Xueqing Wang, Changda Zhou, Hongen Liu, Manhui Lin, Yue Zhang, Yubo Zhang, Yi Liu, Dianhai Yu, and Yanjun Ma. PaddleOCR-VL-1.5: Towards a multi-task 0.9B VLM for robust in-the- wild document parsing.arXiv preprint arXiv:2601.21957, 2026

    Pith/arXiv arXiv 2026

  13. [13]

    Playwright: Fast and reliable end-to-end testing for modern web apps

    Microsoft. Playwright: Fast and reliable end-to-end testing for modern web apps. https: //playwright.dev/, 2026. 17

    2026

  14. [14]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. DeepSeekMath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

    Pith/arXiv arXiv 2024

  15. [15]

    Image over text: Transforming formula recognition evaluation with character detection matching

    Bin Wang, Fan Wu, Linke Ouyang, Zhuangcheng Gu, Rui Zhang, Renqiu Xia, Botian Shi, Bo Zhang, and Conghui He. Image over text: Transforming formula recognition evaluation with character detection matching. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    2025

  16. [16]

    Image-based table recognition: Data, model, and evaluation

    Xu Zhong, Elaheh ShafieiBavani, and Antonio Jimeno Yepes. Image-based table recognition: Data, model, and evaluation. InProceedings of the European Conference on Computer Vision (ECCV), 2020

    2020

  17. [17]

    Distilling the knowledge in a neural network

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015

    Pith/arXiv arXiv 2015

  18. [18]

    On-policy distillation of language models: Learning from self-generated mistakes

    Rishabh Agarwal, Nino Vieillard, Yongchao Zhou, Piotr Stanczyk, Sabela Ramos, Matthieu Geist, and Olivier Bachem. On-policy distillation of language models: Learning from self-generated mistakes. InInternational Conference on Learning Representations, 2024

    2024

  19. [19]

    Rethinking on-policy distillation of large language models: Phenomenology, mechanism, and recipe.arXiv preprint arXiv:2604.13016, 2026

    Yaxuan Li, Yuxin Zuo, Bingxiang He, Jinqian Zhang, Chaojun Xiao, Cheng Qian, Tianyu Yu, Huan-ang Gao, Wenkai Yang, Zhiyuan Liu, and Ning Ding. Rethinking on-policy distillation of large language models: Phenomenology, mechanism, and recipe.arXiv preprint arXiv:2604.13016, 2026

    Pith/arXiv arXiv 2026

  20. [20]

    Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and Ludwig Schmidt

    Mitchell Wortsman, Gabriel Ilharco, Samir Yitzhak Gadre, Rebecca Roelofs, Raphael Gontijo- Lopes, Ari S. Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and Ludwig Schmidt. Model Soups: Averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. InProceedings of the 39th International Conferen...

    2022

  21. [21]

    InternVL3.5: Advancing open-source multimodal models in versatility, reasoning, and efficiency.arXiv preprint arXiv:2508.18265, 2025

    Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, Zhaokai Wang, Zhe Chen, Hongjie Zhang, Ganlin Yang, Haomin Wang, Qi Wei, Jinhui Yin, Wenhao Li, Erfei Cui, Guanzhou Chen, Zichen Ding, Changyao Tian, Zhenyu Wu, Jingjing Xie, Zehao Li, Bowen Yang, Yuchen Duan, Xuehui Wang, Zhi Hou,...

    Pith/arXiv arXiv 2025

  22. [22]

    Kimi K2.5: Visual agentic intelligence.arXiv preprint arXiv:2602.02276, 2026

    Kimi Team. Kimi K2.5: Visual agentic intelligence.arXiv preprint arXiv:2602.02276, 2026

    Pith/arXiv arXiv 2026

  23. [23]

    Update to GPT-5 system card: GPT-5.2

    OpenAI. Update to GPT-5 system card: GPT-5.2. https://cdn.openai.com/pdf/ 3a4153c8-c748-4b71-8e31-aecbde944f8d/oai_5_2_system-card.pdf, 2025

    2025

  24. [24]

    Qwen3-VL technical report.arXiv preprint arXiv:2511.21631, 2025

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, Mei Li, Kaixin Li, Zicheng Lin, Junyang Lin, Xuejing Liu, Jiawei Liu, Chenglong Liu, Yang Liu, Dayiheng Liu, Shixuan ...

    Pith/arXiv arXiv 2025

  25. [25]

    Gemini 3 Flash model card

    Google DeepMind. Gemini 3 Flash model card. https://deepmind.google/models/ model-cards/gemini-3-flash/, 2025

    2025

  26. [26]

    Gemini 3 Pro model card

    Google DeepMind. Gemini 3 Pro model card. https://deepmind.google/models/ model-cards/gemini-3-pro/, 2025

    2025

  27. [27]

    Ovis2.5 Technical Report.arXiv:2508.11737, 2025

    Shiyin Lu, Yang Li, Yu Xia, Yuwei Hu, Shanshan Zhao, Yanqing Ma, Zhichao Wei, Yinglun Li, Lunhao Duan, Jianshan Zhao, Yuxuan Han, Haijun Li, Wanying Chen, Junke Tang, Chengkun Hou, Zhixing Du, Tianli Zhou, Wenjie Zhang, Huping Ding, Jiahe Li, Wen Li, Gui Hu, Yiliang Gu, Siran Yang, Jiamang Wang, Hailong Sun, Yibo Wang, Hui Sun, Jinlong Huang, Yuping He, S...

    Pith/arXiv arXiv 2025

  28. [28]

    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, Jingqun Tang, Hao Liu, and Can Huang. Dolphin: Document image parsing via heterogeneous anchor prompting. InFindings of the Association for Computational Linguistics: ACL 2025, pages 21919–21936, Vienna, Austria, July 2025. Association for Computat...

    2025

  29. [29]

    MonkeyOCR: Document parsing with a structure-recognition-relation triplet paradigm.arXiv preprint arXiv:2506.05218, 2025

    Zhang Li, Yuliang Liu, Qiang Liu, Zhiyin Ma, Ziyang Zhang, Shuo Zhang, Biao Yang, 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. [30]

    Youtu-Parsing: Perception, structuring and recognition via high-parallelism decoding.arXiv preprint arXiv:2601.20430, 2026

    Kun Yin, Yunfei Wu, Bing Liu, Zhongpeng Cai, Xiaotian Li, Huang Chen, Xin Li, Haoyu Cao, Yinsong Liu, Deqiang Jiang, Xing Sun, Yunsheng Wu, Qianyu Li, Antai Guo, Yanzhen Liao, Yanqiu Qu, Haodong Lin, Chengxu He, and Shuangyin Liu. Youtu-Parsing: Perception, structuring and recognition via high-parallelism decoding.arXiv preprint arXiv:2601.20430, 2026

    Pith/arXiv arXiv 2026

  31. [31]

    PaddleOCR-VL: Boosting multilingual document parsing via a 0.9B ultra-compact vision-language model.arXiv preprint arXiv:2510.14528, 2025

    Cheng Cui, Ting Sun, Suyin Liang, Tingquan Gao, Zelun Zhang, Jiaxuan Liu, Xueqing Wang, Changda Zhou, Hongen Liu, Manhui Lin, Yue Zhang, Yubo Zhang, Handong Zheng, Jing Zhang, Jun Zhang, Yi Liu, Dianhai Yu, and Yanjun Ma. PaddleOCR-VL: Boosting multilingual document parsing via a 0.9B ultra-compact vision-language model.arXiv preprint arXiv:2510.14528, 2025

    arXiv 2025

  32. [32]

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

    Yuan Liu, Zhongyin Zhao, Le Tian, Haicheng Wang, Xubing Ye, Yangxiu You, Zilin Yu, Chuhan Wu, Xiao Zhou, Yang Yu, and Jie Zhou. POINTS-Reader: Distillation-free adapta- tion of vision-language models for document conversion. InProceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2025

    2025

  33. [33]

    Nanonets-OCR-S: A model for transforming documents into structured Markdown with intelligent con- tent recognition and semantic tagging

    Souvik Mandal, Ashish Talewar, Paras Ahuja, and Prathamesh Juvatkar. Nanonets-OCR-S: A model for transforming documents into structured Markdown with intelligent con- tent recognition and semantic tagging. https://nanonets.com/research/nanonets-ocr-s, 2025

    2025

  34. [34]

    olmOCR: Unlocking trillions of tokens in PDFs with vision language models.arXiv preprint arXiv:2502.18443, 2025

    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. 19

    arXiv 2025

  35. [35]

    OCRVerse: Towards holistic OCR in end-to- end vision-language models.arXiv preprint arXiv:2601.21639, 2026

    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

    arXiv 2026

  36. [36]

    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, Qi Yang, Qiming Peng, Bin Luo, Hower Yang, Xinsong Zhang, Jinnian Zhang, Houwen Peng, Hongming Yang, Senhao Xie, Longsha Zhou, Ge Pei, Binghong Wu, Rui Yan, Kan Wu, Jieneng Yang, Bochao Wang, Kai Liu, Jianchen Zhu, Jie Jiang...

    arXiv 2025

  37. [37]

    DeepSeek-OCR 2: Visual causal flow.arXiv preprint arXiv:2601.20552, 2026

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

    arXiv 2026

  38. [38]

    UniRec-0.1B: Unified text and formula recognition with 0.1B parameters.arXiv preprint arXiv:2512.21095, 2025

    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

  39. [39]

    dots.ocr: Multilingual document layout parsing in a single vision-language model.arXiv preprint arXiv:2512.02498, 2025

    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

  40. [40]

    FireRed-OCR technical report.arXiv preprint arXiv:2603.01840, 2026

    Hao Wu, Haoran Lou, Xinyue Li, Zuodong Zhong, Zhaojun Sun, Phellon Chen, Xuanhe Zhou, Kai Zuo, Yibo Chen, Xu Tang, Yao Hu, Boxiang Zhou, Jian Wu, Yongji Wu, Wenxin Yu, Yingmiao Liu, Yuhao Huang, Manjie Xu, Gang Liu, Yidong Ma, Zhichao Sun, and Changhao Qiao. FireRed-OCR technical report.arXiv preprint arXiv:2603.01840, 2026

    arXiv 2026

  41. [41]

    ABot-OCR technical report.arXiv preprint arXiv:2605.27978, 2026

    Kaitao Jiang, Ruiyan Gong, Xiaolong Cheng, Kangning Niu, Tianlun Li, and Mu Xu. ABot-OCR technical report.arXiv preprint arXiv:2605.27978, 2026

    Pith/arXiv arXiv 2026

  42. [42]

    Logics-Parsing technical report.arXiv preprint arXiv:2509.19760, 2025

    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

  43. [43]

    Qianfan-OCR: A unified end-to-end model for document intelligence.arXiv preprint arXiv:2603.13398, 2026

    Daxiang Dong, Mingming Zheng, Dong Xu, Chunhua Luo, Bairong Zhuang, Yuxuan Li, Ruoyun He, Haoran Wang, Wenyu Zhang, Wenbo Wang, Yicheng Wang, Xue Xiong, Ayong Zheng, Xiaoying Zuo, Ziwei Ou, Jingnan Gu, Quanhao Guo, Jianmin Wu, Dawei Yin, and Dou Shen. Qianfan-OCR: A unified end-to-end model for document intelligence.arXiv preprint arXiv:2603.13398, 2026

    arXiv 2026

  44. [44]

    MiniCPM-V 4.5: Cooking efficient MLLMs via architecture, data, and training recipe.arXiv preprint arXiv:2509.18154, 2025

    Tianyu Yu, Zefan Wang, Chongyi Wang, Fuwei Huang, Wenshuo Ma, Zhihui He, Tianchi Cai, Weize Chen, Yuxiang Huang, Yuanqian Zhao, Bokai Xu, Junbo Cui, Yingjing Xu, Liqing Ruan, Luoyuan Zhang, Hanyu Liu, Jingkun Tang, Hongyuan Liu, Qining Guo, Wenhao Hu, Bingxiang He, Jie Zhou, Jie Cai, Ji Qi, Zonghao Guo, Chi Chen, Guoyang Zeng, Yuxuan Li, Ganqu Cui, Ning D...

    Pith/arXiv arXiv 2025

  45. [45]

    STEP3-VL-10B technical report.arXiv preprint arXiv:2601.09668, 2026

    Ailin Huang, Chengyuan Yao, Chunrui Han, Fanqi Wan, Hangyu Guo, Haoran Lv, Hongyu Zhou, Jia Wang, Jian Zhou, Jianjian Sun, Jingcheng Hu, Kangheng Lin, Liang Zhao, Mitt Huang, Song Yuan, Wenwen Qu, Xiangfeng Wang, Yanlin Lai, Yingxiu Zhao, Yinmin Zhang, Yukang Shi, Yuyang Chen, Zejia Weng, Ziyang Meng, Ang Li, Aobo Kong, Bo Dong, Changyi Wan, David Wang, D...

    arXiv 2026

  46. [46]

    Gemini 3.1 Pro model card

    Google DeepMind. Gemini 3.1 Pro model card. https://deepmind.google/models/ model-cards/gemini-3-1-pro/, 2026

    2026

  47. [47]

    SVTRv2: CTC beats encoder-decoder models in scene text recognition

    Yongkun Du, Zhineng Chen, Hongtao Xie, Caiyan Jia, and Yu-Gang Jiang. SVTRv2: CTC beats encoder-decoder models in scene text recognition. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20147–20156, 2025

    2025

  48. [48]

    Multimodal OCR: Parse anything from documents

    Handong Zheng, Yumeng Li, Kaile Zhang, Liang Xin, Guangwei Zhao, Hao Liu, Jiayu Chen, Jie Lou, Qi Fu, Rui Yang, Shuo Jiang, Weijian Luo, Weijie Su, Weijun Zhang, Xingyu Zhu, Yabin Li, Yiwei Ma, Yu Chen, Yuqiu Ji, Zhaohui Yu, Guang Yang, Colin Zhang, Lei Zhang, Yuliang Liu, and Xiang Bai. Multimodal OCR: Parse anything from documents. arXiv preprint arXiv:...

    arXiv 2026

  49. [49]

    OCRFlux: A multimodal toolkit for converting documents into Markdown

    ChatDOC. OCRFlux: A multimodal toolkit for converting documents into Markdown. https://github.com/chatdoc-com/OCRFlux, 2025

    2025

  50. [50]

    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

  51. [51]

    Nanonets-OCR2: A model for transforming documents into structured Markdown with intelligent content recognition and semantic tagging

    Souvik Mandal, Ashish Talewar, Siddhant Thakuria, Paras Ahuja, and Prathamesh Juvatkar. Nanonets-OCR2: A model for transforming documents into structured Markdown with intelligent content recognition and semantic tagging. https://nanonets.com/research/ nanonets-ocr-2, 2025

    2025

  52. [52]

    olmOCR 2: Unit test rewards for document OCR.arXiv preprint arXiv:2510.19817, 2025

    Jake Poznanski, Luca Soldaini, and Kyle Lo. olmOCR 2: Unit test rewards for document OCR.arXiv preprint arXiv:2510.19817, 2025

    arXiv 2025

  53. [53]

    Reading or reasoning? format decoupled reinforcement learning for document OCR

    Yufeng Zhong, Lei Chen, Zhixiong Zeng, Xuanle Zhao, Deyang Jiang, Liming Zheng, Jing Huang, Haibo Qiu, Peng Shi, Siqi Yang, and Lin Ma. Reading or reasoning? format decoupled reinforcement learning for document OCR. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026. 21 A Qualitative examples This appendix pre...

    2026