Pith. sign in

REVIEW 2 major objections 6 minor 1 cited by

Two-stage speculative decoding speeds up document-parsing AI by 2.78x without retraining, and up to 7.04x on long documents.

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 23:38 UTC pith:JC46EOP2

load-bearing objection HSD delivers a real, broad, training-free speedup for document-parsing VLMs, but the 'near-lossless' label needs the missing tau=1.0 accuracy control before I'd fully trust it. the 2 major comments →

arxiv 2602.12957 v3 pith:JC46EOP2 submitted 2026-02-13 cs.CV

HSD: Training-Free Acceleration for Document Parsing Vision-Language Models with Hierarchical Speculative Decoding

Wenhui Liao , Hongliang Li , Pengyu Xie , Xinyu Cai , Yufan Shen , Yi Xin , Qi Qin , Shenglong Ye
show 10 more authors
Tianbin Li Ming Hu Junjun He Yihao Liu Wenhai Wang Min Dou Bin Fu Botian Shi Yu Qiao Lianwen Jin
This is my paper
classification cs.CV
keywords speculative decodingdocument parsingvision-language modelsinference accelerationhierarchical verificationtraining-free accelerationlayout analysislatency reduction
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 the main bottleneck of end-to-end vision-language document parsers—slow autoregressive generation of long, structured output—can be removed without modifying the model or retraining. It proposes Hierarchical Speculative Decoding: a fast pipeline parser first drafts coarse region-level outputs, then the target parser verifies those drafts in parallel per region, and finally verifies the assembled page once more to restore global coherence. The paper reports near-lossless acceleration across multiple parsers and benchmarks, with the target model remaining the final arbiter that corrects draft errors. If correct, this gives end-to-end parsers hybrid-level efficiency while preserving their accuracy.

Core claim

The central claim is that region-level parallelism and page-level global coherence are not in conflict. A lightweight pipeline parser generates a region partition and coarse token drafts once per page. Stage 1 verifies each region draft in parallel on cropped images using speculative decoding; Stage 2 aggregates the refined regions into a page draft and runs one more verification pass on the full page to fix cross-region inconsistencies. Because the drafts are generated a single time and never refreshed, a decoupled speculative verification scheme aligns them to the target's current prefix via a short reference window, and a prefix-tree batching mechanism verifies multiple candidate continua

What carries the argument

The central mechanism is the two-stage hierarchical verification loop combined with decoupled draft-target alignment. A lightweight pipeline drafter runs once per page, producing semantic region partitions and coarse per-region token sequences. Stage 1 runs the target parser on region crops in parallel, verifying drafts; Stage 2 runs the target parser on the full page with the aggregated refined regions as the draft, correcting structural errors and restoring reading order. Decoupled Speculative Verification handles the resulting prefix-draft misalignment: a reference window of accepted tokens is matched against each draft to extract candidate suffixes, which are organized into a prefix tree

Load-bearing premise

The speedup and near-lossless accuracy both rest on the pipeline drafter producing region partitions and coarse drafts whose token sequences align with what the target parser would generate; on document types where the drafter's output is poor, the speedup degrades—the paper's own qualitative analysis shows a handwritten page with end-to-end speedup of only 1.22x.

What would settle it

Run HSD on a set of documents deliberately chosen to fool the pipeline drafter—cursive handwriting, unusual reading orders, mixed-language layouts, or heavy noise—and measure both end-to-end speedup and parsing accuracy relative to the unmodified target parser. If speedup stays high while accuracy deviates noticeably from the baseline, the near-lossless claim would be falsified.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • End-to-end document parsers can be accelerated to roughly the efficiency of hybrid systems while keeping near-lossless accuracy, without retraining or architectural changes.
  • Speedup grows with output length and region parallelism: long, multi-block documents such as newspapers and academic papers benefit most, while short or sparse pages are limited by fixed vision/prefill cost.
  • Draft quality is the primary factor in acceptance length; the method remains beneficial but weaker on handwritten or degraded scans where the pipeline produces poor drafts.
  • The method is plug-and-play and stackable with other acceleration techniques such as visual token compression, yielding additional speedup on top.

Where Pith is reading between the lines

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

  • The method's ceiling is set by the pipeline drafter's competence: on document types outside the drafter's ability, the speedup can approach 1x. A natural extension is to adapt or replace the drafter per domain, or to use a small vision-language model as the drafter.
  • The decoupled, hierarchical verification idea could transfer to other structured generation tasks (table extraction, form filling, code generation) where a fast rule-based or heuristic system can produce coarse structured drafts.
  • The two-stage local-then-global verification principle might generalize beyond decoding—for example, to multi-stage agentic pipelines where cheap local hypotheses are verified in parallel before a global consistency pass.

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

2 major / 6 minor

Summary. The paper proposes Hierarchical Speculative Decoding (HSD), a training-free inference-time acceleration method for VLM-based end-to-end document parsers. A lightweight pipeline drafter (PP-StructureV3) first generates region-level layout partitions and coarse drafts. Stage 1 verifies these drafts independently and in parallel on cropped regions; Stage 2 aggregates the refined region outputs and performs a full-page verification to restore global coherence. To handle the mismatch between fixed drafts and the target model's evolving prefix, the paper introduces Decoupled Speculative Verification (DSV), which matches a reference window of accepted tokens against the drafts, builds a prefix tree of candidate continuations, and verifies them in parallel with a tree-ancestry attention mask. A soft acceptance threshold τ=0.75 accepts tokens whose probability is within 75% of the greedy token. Experiments on OmniDocBench v1.5, olmOCR-Bench, and Ocean-OCR-Bench with six parsers (dots.ocr, HunyuanOCR, Qwen2.5-VL-3B/7B, Qwen3-VL-2B/8B) report end-to-end speedups of 1.35×–7.04× and claim near-lossless accuracy, with a headline 2.78× speedup for HunyuanOCR on OmniDocBench v1.5.

Significance. If the results hold, HSD is a practically valuable contribution: it is plug-and-play, requires no training of the target model, and combines the efficiency of pipeline drafters with the accuracy of end-to-end VLMs. The paper includes a broad empirical study across six models and three benchmarks, ablations of hierarchical design and DSV, a drafter-robustness analysis, and a Roofline-style FLOPs/latency analysis. The speedup numbers are consistent and substantial. However, the central 'near-lossless' claim rests on a tolerance-based acceptance rule that abandons the distribution-preservation guarantee of standard speculative decoding, and the empirical evidence for quality preservation is currently incomplete. These issues are addressable but require additional experiments and analysis.

major comments (2)
  1. [§3.2, Eq. (11); §4.5, Tables 7–8] The acceptance rule in Eq. (11) with τ=0.75 accepts a candidate u* even when pθ(u*) is smaller than the greedy token pθ(û), as long as the log-probability gap is within log τ. This deliberately produces outputs that differ from the target model's greedy decoding, breaking the distribution-preservation guarantee of standard speculative decoding. The paper calls the result "near-lossless" but never isolates the accuracy cost of this tolerance. Table 8 compares τ=1.0 and τ=0.75 only on AAL and speedup; no accuracy is reported for τ=1.0. Since τ=1.0 is exact (it accepts only the greedy argmax), a direct accuracy comparison between τ=1.0 and τ=0.75 on the same benchmarks would quantify the trade-off that underlies the near-lossless claim. Without it, the central claim is not substantiated. I also note that no sensitivity analysis over τ is provided, so the choice 0.75 is unvalidated.
  2. [§4.4, Tables 5 and 7] The accuracy evidence supporting "near-lossless across models, document types, and languages" is incomplete. Table 7 reports accuracy only for dots.ocr, and the deltas are point estimates without error bars or significance tests: +0.40 on OmniDocBench v1.5, −0.50 on olmOCR-Bench, and +1.11 on Ocean-OCR-Bench. For HunyuanOCR, the model used for the headline 2.78× speedup, accuracy is given only for OmniDocBench v1.5 (Table 5: 94.10→94.02). No accuracy numbers are reported for HunyuanOCR on olmOCR-Bench or Ocean-OCR-Bench, nor for any of the Qwen-family parsers used in the acceleration tables. The claim that HSD is near-lossless across all evaluated settings therefore rests on a small and partially selected subset. Please report accuracy for all model-benchmark pairs in the main tables, ideally with bootstrap confidence intervals or repeated-run variance.
minor comments (6)
  1. [§4.5, Table 8 text] The text states that τ=1.0 yields AAL 3.87 and τ=0.75 yields 4.98, but Table 8 lists 2.87 and 3.98. Please correct the text or the table so they are consistent.
  2. [§3.1] The term "training-free" is used, but the method requires a trained pipeline drafter (PP-StructureV3). Clarify that the target VLM is not modified or trained; the external drafter is pretrained.
  3. [§4.5, Fig. 3] The x-axis "Accuracy of Drafter" is not clearly defined; specify which metric from OmniDocBench is used and how the noise injections affect it.
  4. [§3.2, Eq. (10)] The notation û (hat u) can be confused with the update notation in Eq. (12); consider using a different symbol for the global argmax, e.g., v*.
  5. [Appendix A, Fig. A3] The authors are transparent about the 1.22× worst-case speedup on a handwritten page; I suggest mentioning this draft-limited behavior in the main text as a caveat to the headline speedups.
  6. [Table 2] The column header "Old ScansMathTables" appears to concatenate two categories ("Old Scans" and "Math Tables"). Please fix the formatting.

Circularity Check

0 steps flagged

No circularity: HSD's speedup and accuracy claims are measured empirically against external baselines; the method's equations define an inference procedure, not a derivation from its conclusions.

full rationale

The paper's central claims (2.78x end-to-end speedup with HunyuanOCR, up to 7.04x on long documents, near-lossless accuracy) are empirical measurements, not derived quantities. Eqs. (1)-(12) define the verification procedure, but no predicted result is obtained by fitting a parameter to the benchmark and then re-reporting it. The acceptance threshold tau=0.75 and reference window n=3 are fixed hyperparameters; Table 8 compares tau=1.0 vs tau=0.75 for AAL and speedup, and Table 7 reports accuracy against baseline parsers on OmniDocBench, olmOCR-Bench, and Ocean-OCR-Bench. Even though some benchmarks/drafters (OmniDocBench v1.5/MinerU2.5, HunyuanOCR) share authors/ecosystem with the present paper, those citations are not load-bearing theoretical premises: the speedups are wall-clock ratios versus the same target model, and accuracy is scored by external protocols. No self-citation chain is used to justify correctness, no uniqueness theorem is imported, and no known result is merely renamed. The only notable gap (tau=1.0 accuracy not reported) is an evaluation-reporting concern, not a circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

No new physical or architectural entity is introduced. The method relies on two hand-set hyperparameters (n and tau) and on the external pipeline drafter as a strong prior. The near-lossless claim depends on an unproven but empirically checked tolerance assumption.

free parameters (2)
  • reference-window length n = 3
    Set in Sec 4.3. Controls how draft-target matching finds candidate continuations. No sensitivity analysis is reported.
  • acceptance threshold tau = 0.75
    Set in Sec 4.3. Tolerates draft tokens within 75% of the top-1 probability, trading exactness for speed. Only tau=1.0 vs 0.75 is ablated in Table 8.
axioms (3)
  • domain assumption A pipeline parser (PP-StructureV3) supplies region partitions and drafts whose token sequences align with the target VLM's output often enough.
    HSD's speedup and near-lossless accuracy depend on high draft acceptance. Fig. A3 shows that when drafts are poor (handwriting), AAL drops to 2.09 and speedup to 1.22x.
  • standard math Tree-ancestry attention masking lets the target model score all candidate suffixes in one forward pass without changing the conditional distribution of each path.
    Invoked in Sec 3.2 (Eq 8-11). This is standard for causal transformer tree verification, assuming the model is a causal transformer.
  • ad hoc to paper Accepting a token whose probability is at least 0.75 of the greedy top token preserves document parsing quality.
    Eq. 11 defines this acceptance rule, but no proof or bound shows that it preserves the target output distribution. It is supported only by aggregate accuracy in Tables 5 and 7.

pith-pipeline@v1.3.0-alltime-deepseek · 17207 in / 11834 out tokens · 108589 ms · 2026-08-02T23:38:46.271379+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of HSD: Training-Free Acceleration for Document Parsing Vision-Language Models with Hierarchical Speculative Decoding." pith.science (2026). https://pith.science/paper/JC46EOP2

@misc{pith2026260212957,
  author       = {Pith},
  title        = {Pith review of: HSD: Training-Free Acceleration for Document Parsing Vision-Language Models with Hierarchical Speculative Decoding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JC46EOP2}},
  note         = {Machine review of arXiv:2602.12957}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Document parsing is a fundamental task in multimodal understanding, supporting a wide range of downstream applications such as information extraction and intelligent document analysis. Benefiting from strong semantic modeling and robust generalization, VLM-based end-to-end approaches have emerged as the mainstream paradigm in recent years. However, these models often suffer from substantial inference latency, as they must autoregressively generate long, full-page sequences when processing long-form documents. While recent hybrid methods mitigate this issue via region-level parallel decoding with VLMs, independent region decoding loses full-page context and might weaken global coherence. To address this issue, we propose Hierarchical Speculative Decoding (HSD), a two-stage local-to-global framework for document parsing. HSD first employs a lightweight pipeline drafter to predict region partitions and generate coarse drafts for each region. The first stage verifies the generated region-level drafts in parallel for efficiency, while the second stage further performs page-level verification on these refined outputs to preserve full-page coherence. Experimental results show that HSD achieves a near-lossless 2.7x speedup with HunyuanOCR on OmniDocBench v1.5 and up to 7.04x speedup on long-document parsing tasks, demonstrating the effectiveness of the proposed method. The code is available at https://github.com/whlscut/HSD.

Figures

Figures reproduced from arXiv: 2602.12957 by Bin Fu, Botian Shi, Hongliang Li, Junjun He, Lianwen Jin, Min Dou, Ming Hu, Pengyu Xie, Qi Qin, Shenglong Ye, Tianbin Li, Wenhai Wang, Wenhui Liao, Xinyu Cai, Yihao Liu, Yi Xin, Yufan Shen, Yu Qiao.

Figure 3
Figure 3. Figure 3: Evaluation metrics of different methods evaluated on a test set [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 1
Figure 1. Figure 1: Overview of the proposed hierarchical speculative decoding paradigm. A lightweight pipeline generates coarse region drafts via layout analysis and element recognition. The end-to-end parser verifies these drafts in two stages: Stage 1 performs region-level verification on cropped regions in parallel to produce refined drafts; Stage 2 aggregates them into a page-level draft and performs global, page-level v… view at source ↗
Figure 2
Figure 2. Figure 2: Visualization of decoupled speculative verification. (a) Draft–target matching aligns a reference window from the accepted tokens with multiple drafts to extract candidate continuations. (b) The prefix tree organizes candidates by merging common prefixes. (c) The tree-ancestry attention mask enables parallel verification, where to￾kens attend only to the accepted tokens and their ancestors in the prefix tr… view at source ↗
Figure 3
Figure 3. Figure 3: Impacts of using different pipelines as drafters. Accuracy of Drafter is the Om￾niDocBench v1.5 score; “add . . . ” denotes noise injected into PP-StructureV3 drafts. benchmarks. These results show that HSD achieves substantial speedups while maintaining parsing quality. Impact of Framework Designs. Tab. 8 presents an ablation study to an￾alyze the influence of different designs. Comparing page-level-only … view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. HunyuanOCR-1.5: Making Lightweight OCR VLMs Faster and Better

    cs.CV 2026-07 conditional novelty 5.0

    HunyuanOCR-1.5 makes a lightweight end-to-end OCR VLM the fastest among peers via DFlash speculative decoding while expanding long-tail capabilities through agent-driven data construction.

Reference graph

Works this paper leans on

49 extracted references · 18 linked inside Pith · cited by 1 Pith paper

  1. [1]

    NeurIPS (2022)

    Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. NeurIPS (2022)

    2022

  2. [2]

    arXiv preprint arXiv:2511.21631 (2025)

    Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., Ge, W., Guo, Z., Huang, Q., Huang, J., Huang, F., Hui, B., Jiang, S., Li, Z., Li, M., Li, M., Li, K., Lin, Z., Lin, J., Liu, X., Liu, J., Liu, C., Liu, Y., Liu, D., Liu, S., Lu, D., Luo, R., Lv, C., Men, R., Meng, L., Ren, X., Ren, X., Song, S., Sun, Y., Tang, ...

    Pith/arXiv arXiv 2025

  3. [3]

    arXiv preprint arXiv:2502.13923 (2025)

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., Zhong, H., Zhu, Y., Yang, M., Li, Z., Wan, J., Wang, P., Ding, W., Fu, Z., Xu, Y., Ye, J., Zhang, X., Xie, T., Cheng, Z., Zhang, H., Yang, Z., Xu, H., Lin, J.: Qwen2.5-VL Technical Report. arXiv preprint arXiv:2502.13923 (2025)

    Pith/arXiv arXiv 2025

  4. [4]

    arXiv preprint arXiv:2308.13418 (2023)

    Blecher, L., Cucurull, G., Scialom, T., Stojnic, R.: Nougat: Neural optical under- standing for academic documents. arXiv preprint arXiv:2308.13418 (2023)

    Pith/arXiv arXiv 2023

  5. [5]

    In: ICML

    Cai, T., Li, Y., Geng, Z., Peng, H., Lee, J.D., Chen, D., Dao, T.: Medusa: Simple LLM inference acceleration framework with multiple decoding heads. In: ICML. pp. 5209–5235 (2024)

    2024

  6. [6]

    arXiv preprint arXiv:2302.01318 (2023)

    Chen, C., Borgeaud, S., Irving, G., Lespiau, J.B., Sifre, L., Jumper, J.: Acceler- ating large language model decoding with speculative sampling. arXiv preprint arXiv:2302.01318 (2023)

    Pith/arXiv arXiv 2023

  7. [7]

    arXiv preprint arXiv:2501.15558 (2025)

    Chen, S., Guo, X., Li, Y., Zhang, T., Lin, M., Kuang, D., Zhang, Y., Ming, L., Zhang, F., Wang, Y., et al.: Ocean-OCR: Towards general OCR application via a vision-language model. arXiv preprint arXiv:2501.15558 (2025)

    Pith/arXiv arXiv 2025

  8. [8]

    In: ICLR (2022)

    Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., et al.: PaLI: A jointly-scaled multilingual language-image model. In: ICLR (2022)

    2022

  9. [9]

    com / chatdoc - com / OCRFlux(2025), accessed:2025-11-10

    chatdoc com: Ocrflux.https : / / github . com / chatdoc - com / OCRFlux(2025), accessed:2025-11-10

    2025

  10. [10]

    arXiv preprint arXiv:2510.14528 (2025) 16 W

    Cui, C., Sun, T., Liang, S., Gao, T., Zhang, Z., Liu, J., Wang, X., Zhou, C., Liu, H., Lin, M., Zhang, Y., Zhang, Y., Zheng, H., Zhang, J., Zhang, J., Liu, Y., Yu, D., Ma, Y.: PaddleOCR-VL: Boosting multilingual document parsing via a 0.9B ultra-compact vision-language model. arXiv preprint arXiv:2510.14528 (2025) 16 W. Liao and H. Li et al

    arXiv 2025

  11. [11]

    arXiv preprint arXiv:2507.05595 (2025)

    Cui, C., Sun, T., Lin, M., Gao, T., Zhang, Y., Liu, J., Wang, X., Zhang, Z., Zhou, C., Liu, H., et al.: PaddleOCR 3.0 technical report. arXiv preprint arXiv:2507.05595 (2025)

    Pith/arXiv arXiv 2025

  12. [12]

    In: NeurIPS (2022)

    Dao,T.,Fu,D.Y.,Ermon,S.,Rudra,A.,Ré,C.:FlashAttention:Fastandmemory- efficient exact attention with IO-awareness. In: NeurIPS (2022)

    2022

  13. [13]

    Elhoushi, M., Shrivastava, A., Liskovich, D., Hosmer, B., Wasti, B., Lai, L., Mah- moud, A., Acun, B., Agarwal, S., Roman, A., Aly, A., Chen, B., Wu, C.J.: Lay- erSkip: Enabling early exit inference and self-speculative decoding. In: ACL. pp. 12622–12642 (2024)

    2024

  14. [14]

    arXiv preprint arXiv:2505.14059 (2025)

    Feng, H., Wei, S., Fei, X., Shi, W., Han, Y., Liao, L., Lu, J., Wu, B., Liu, Q., Lin, C., et al.: Dolphin: Document image parsing via heterogeneous anchor prompting. arXiv preprint arXiv:2505.14059 (2025)

    Pith/arXiv arXiv 2025

  15. [15]

    In: CVPR Workshops

    Gagrani, M., Goel, R., Jeon, W., Park, J., Lee, M., Lott, C.: On speculative decod- ing for multimodal large language models. In: CVPR Workshops. pp. 8285–8289 (2024)

    2024

  16. [16]

    In: NeurIPS (2025)

    Kang, J., Shu, H., Li, W., Zhai, Y., Chen, X.: ViSpec: Accelerating vision-language models with vision-aware speculative decoding. In: NeurIPS (2025)

    2025

  17. [17]

    In: ICML

    Leviathan, Y., Kalman, M., Matias, Y.: Fast inference from transformers via spec- ulative decoding. In: ICML. pp. 19274–19286 (2023)

    2023

  18. [18]

    In: ICML (2022)

    Li, J., Li, D., Xiong, C., Hoi, S.: BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML (2022)

    2022

  19. [19]

    In: EMNLP

    Li, Y., Wei, F., Zhang, C., Zhang, H.: EAGLE-2: Faster inference of language models with dynamic draft trees. In: EMNLP. pp. 7421–7432 (2024)

    2024

  20. [20]

    In: ICML

    Li, Y., Wei, F., Zhang, C., Zhang, H.: EAGLE: Speculative sampling requires rethinking feature uncertainty. In: ICML. vol. 235, pp. 28935–28948 (2024)

    2024

  21. [21]

    arXiv preprint arXiv:2506.05218 (2025)

    Li, Z., Liu, Y., Liu, Q., Ma, Z., Zhang, Z., Zhang, S., Guo, Z., Zhang, J., Wang, X., Bai, X.: MonkeyOCR: Document parsing with a structure-recognition-relation triplet paradigm. arXiv preprint arXiv:2506.05218 (2025)

    arXiv 2025

  22. [22]

    arXiv preprint arXiv:2505.14260 (2025)

    Lin, L., Lin, Z., Zeng, Z., Ji, R.: Speculative decoding reimagined for multimodal large language models. arXiv preprint arXiv:2505.14260 (2025)

    Pith/arXiv arXiv 2025

  23. [23]

    arXiv:2310.03744 (2023)

    Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv:2310.03744 (2023)

    Pith/arXiv arXiv 2023

  24. [24]

    arXiv preprint arXiv:2509.01215 (2025)

    Liu, Y., Zhao, Z., Tian, L., Wang, H., Ye, X., You, Y., Yu, Z., Wu, C., Zhou, X., Yu, Y., et al.: POINTS-Reader: Distillation-free adaptation of vision-language models for document conversion. arXiv preprint arXiv:2509.01215 (2025)

    Pith/arXiv arXiv 2025

  25. [25]

    arXiv preprint arXiv:2501.17887 (2025)

    Livathinos, N., Auer, C., Lysak, M., Nassar, A., Dolfi, M., Vagenas, P., Ramis, C.B., Omenetti, M., Dinkla, K., Kim, Y., et al.: Docling: An efficient open-source toolkit for AI-driven document conversion. arXiv preprint arXiv:2501.17887 (2025)

    Pith/arXiv arXiv 2025

  26. [26]

    Mandalm, S.: Nanonets-ocr-s.https://nanonets.com/research/nanonets-ocr- s/(2025), accessed:2025-11-10

    2025

  27. [27]

    In: ASPLOS

    Miao, X., Oliaro, G., Zhang, Z., Cheng, X., Wang, Z., Zhang, Z., Wong, R.Y.Y., Zhu, A., Yang, L., Shi, X., Shi, C., Chen, Z., Arfeen, D., Abhyankar, R., Jia, Z.: SpecInfer: Accelerating large language model serving with tree-based speculative inference and verification. In: ASPLOS. p. 932–949 (2024)

    2024

  28. [28]

    arXiv preprint arXiv:2509.22186 (2025)

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

    Pith/arXiv arXiv 2025

  29. [29]

    NVIDIA: Nvidia a100 tensor core gpu datasheet.https : / / www . nvidia . com / content / dam / en - zz / Solutions / Data - Center / a100 / pdf / nvidia - a100 - datasheet-nvidia-us-2188504-web.pdf(2021), accessed: 2026-03-09

    2021

  30. [30]

    OpenDataLab: MinerU.https://github.com/opendatalab/MinerU(2025), ac- cessed 2026-03-04

    2025

  31. [31]

    com / datalab - to / marker(2025), accessed:2025-11-10

    Paruchuri, V.: Marker.https : / / github . com / datalab - to / marker(2025), accessed:2025-11-10

    2025

  32. [32]

    arXiv preprint arXiv:2502.18443 (2025)

    Poznanski, J., Rangapur, A., Borchardt, J., Dunkelberger, J., Huff, R., Lin, D., Wilhelm, C., Lo, K., Soldaini, L.: olmOCR: Unlocking trillions of tokens in PDFs with vision language models. arXiv preprint arXiv:2502.18443 (2025)

    arXiv 2025

  33. [33]

    arXiv preprint arXiv:2510.19817 (2025)

    Poznanski, J., Soldaini, L., Lo, K.: olmOCR 2: Unit test rewards for document OCR. arXiv preprint arXiv:2510.19817 (2025)

    arXiv 2025

  34. [34]

    PyTorch, T.: FlexAttention: The flexibility of pytorch with the perfor- mance of flashattention.https://pytorch.org/blog/flexattention/(2025), accessed:2025-11-10

    2025

  35. [35]

    rednote: dots.ocr: Multilingual document layout parsing in a single vision-language model.https://github.com/rednote-hilab/dots.ocr(2025), accessed:2025-11- 10

    2025

  36. [36]

    arXiv preprint arXiv:2106.04970 (2021)

    Sun, X., Ge, T., Wei, F., Wang, H.: Instantaneous grammatical error correction with shallow aggressive decoding. arXiv preprint arXiv:2106.04970 (2021)

    Pith/arXiv arXiv 2021

  37. [37]

    arXiv preprint arXiv:2511.19575 (2025)

    Team, H.V., Lyu, P., Wan, X., Li, G., Peng, S., Wang, W., Wu, L., Shen, H., Zhou, Y., Tang, C., Yang, Q., Peng, Q., Luo, B., Yang, H., Zhang, X., Zhang, J., Peng, H., Yang, H., Xie, S., Zhou, L., Pei, G., Wu, B., Yan, R., Wu, K., Yang, J., Wang, B., Liu, K., Zhu, J., Jiang, J., Linus, Hu, H., Zhang, C.: HunyuanOCR technical report. arXiv preprint arXiv:25...

    arXiv 2025

  38. [38]

    Team, M.A.: Mistral-ocr.https://mistral.ai/news/mistral-ocr?utm_source= ai-bot.cn(2025), accessed:2025-11-10

    2025

  39. [39]

    arXiv preprint arXiv:2506.03197 (2025)

    Wang, B., Wu, B., Li, W., Fang, M., Huang, Z., Huang, J., Wang, H., Liang, Y., Chen, L., Chu, W., Qi, Y.: Infinity Parser: Layout aware reinforcement learning for scanned document parsing. arXiv preprint arXiv:2506.03197 (2025)

    arXiv 2025

  40. [40]

    arXiv preprint arXiv:2409.18839 (2024)

    Wang, B., Xu, C., Zhao, X., Ouyang, L., Wu, F., Zhao, Z., Xu, R., Liu, K., Qu, Y., Shang, F., et al.: MinerU: An open-source solution for precise document content extraction. arXiv preprint arXiv:2409.18839 (2024)

    Pith/arXiv arXiv 2024

  41. [41]

    In: Proceedings of the 5th Deep Learning on Supercomputers Workshop (DLS) (2020)

    Wang, Y., Yang, C., Farrell, S., Zhang, Y., Kurth, T., Williams, S.: Time-based roofline for deep learning performance analysis. In: Proceedings of the 5th Deep Learning on Supercomputers Workshop (DLS) (2020)

    2020

  42. [42]

    arXiv preprint arXiv:2409.01704 (2024)

    Wei, H., Liu, C., Chen, J., Wang, J., Kong, L., Xu, Y., Ge, Z., Zhao, L., Sun, J., Peng, Y., et al.: General OCR Theory: Towards OCR-2.0 via a unified end-to-end model. arXiv preprint arXiv:2409.01704 (2024)

    Pith/arXiv arXiv 2024

  43. [43]

    arXiv preprint arXiv:2510.18234 (2025)

    Wei, H., Sun, Y., Li, Y.: DeepSeek-OCR: Contexts optical compression. arXiv preprint arXiv:2510.18234 (2025)

    Pith/arXiv arXiv 2025

  44. [44]

    Communications of the ACM52(4), 65–76 (2009)

    Williams, S., Waterman, A., Patterson, D.: Roofline: An insightful visual per- formance model for multicore architectures. Communications of the ACM52(4), 65–76 (2009)

    2009

  45. [45]

    Liao and H

    Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., Cistac, P., Rault, T., Louf, R., Funtowicz, M., Davison, J., Shleifer, S., von Platen, P., Ma, C., Jernite, Y., Plu, J., Xu, C., Le Scao, T., Gugger, S., Drame, M., Lhoest, Q., 18 W. Liao and H. Li et al. Rush, A.: Transformers: State-of-the-art natural language processing. In: ACL. pp. 3...

    2020

  46. [46]

    arXiv preprint arXiv:2509.23928 (2025)

    Xie, Z., Wang, P., Cheng, J.: HiViS: Hiding visual tokens from the drafter for speculative decoding in vision-language models. arXiv preprint arXiv:2509.23928 (2025)

    arXiv 2025

  47. [47]

    arXiv preprint arXiv:2602.15318 (2026)

    Zhang, L., Zhang, Z., Hong, W., Qiao, P., Li, D.: Sparrow: Text-anchored window attention with visual-semantic glimpsing for speculative decoding in video llms. arXiv preprint arXiv:2602.15318 (2026)

    arXiv 2026

  48. [48]

    arXiv preprint arXiv:2410.21169 (2024)

    Zhang, Q., Wang, B., Huang, V.S.J., Zhang, J., Wang, Z., Liang, H., He, C., Zhang, W.: Document parsing unveiled: Techniques, challenges, and prospects for structured information extraction. arXiv preprint arXiv:2410.21169 (2024)

    Pith/arXiv arXiv 2024

  49. [49]

    Zhao, W., Huang, Y., Han, X., Xiao, C., Liu, Z., Sun, M.: Ouroboros: Speculative decoding with large model enhanced drafting. arXiv preprint arXiv:2402.13720 (2024) HSD 19 HSD: Training-Free Acceleration for Document Parsing Vision-Language Models with Hierarchical Speculative Decoding Supplementary Material A Qualitative Analysis We qualitatively analyze...

    Pith/arXiv arXiv 2024