acceptodds
Under review as a conference paper at ICLR 2027

Diffusion Drafts, AR Verifies: Accelerating Document OCR with Self-Speculative Decoding

Abstract

Autoregressive OCR vision–language models accurately convert document images into text and structured markup, but require one sequential decoding step per output token, limiting inference speed. Unlike open-ended text generation, OCR outputs are strongly grounded in the input image, making diffusion-based parallel generation promising. However, when several tokens are predicted in one diffusion step, each is predicted before the others are known. Committing them directly can therefore introduce errors. We therefore introduce GravityOCR, a parameter-shared AR–block-diffusion model jointly trained for parallel drafting and causal AR verification. Verifying drafts before commitment lets the model commit multiple output tokens per round without a separate drafting network. The causal AR path also enables GRPO with sequence- and structure-level OCR rewards, avoiding diffusion-trajectory likelihood estimation while updating the shared drafter parameters. On OmniDocBench v1.6, AR-path GRPO improves the Overall score from 94.92 to 95.16 without reducing diffusion drafting efficiency, while the final model remains close to the original GLM-OCR score of 95.48. In an SGLang serving deployment, GravityOCR commits an average of 9.7 output tokens per forward pass and achieves a 3.94× decode-only speedup on region crops and a 1.32× end-to-end page-processing speedup over AR decoding.

Primary area:
generative models
Keywords:
optical character recognition, diffusion language models, speculative decoding, vision-language models, efficient inference

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_PGVpOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses faster decoding for document OCR VLMs. AR OCR models are slow because each token needs one forward pass. Committing several tokens at once with diffusion can break reading order and markup structure. The paper therefore proposes to separate drafting from commitment. The proposed method, GravityOCR, adds only a mask token embedding to a pretrained AR OCR model and trains it jointly as a block-diffusion drafter and a causal AR verifier. At inference, the diffusion path drafts a whole block in one pass, and the AR path commits only the prefix that matches its next-token predictions. The authors claim that this keeps the output identical to greedy AR while greatly speeding up decoding, and that GRPO recovers accuracy close to the base model.

Critical strengths
  1. It achieves the fastest page throughput among the systems measured while keeping top-tier accuracy. It presents a wide range of experimental comparisons and analyses, such as the comparison with direct block diffusion and breakdowns by batch size, output length, and content type. It also does not hide the results that the gain shrinks with larger batches and is small for short outputs.

  2. The paper is well written and clean. The figures make the problem setting and the method clear at a glance, and the decoding procedure and measurement settings are fully written out in the appendix, so it was easy to follow.

Critical weaknesses
  1. It is not clear what this paper contributes. The design where diffusion drafts and AR verifies on one backbone already exists in TiDAR and other work the authors cite as closest, and the authors themselves write that they applied it to OCR. At inference the whole block is predicted at once and then verified sequentially by AR, so the advantage of diffusion, unmasking high-confidence tokens first regardless of position, is not used, and the method looks in effect like MTP with an adaptive window. Meanwhile the only element aimed at OCR is the post-training reward, so it is hard to judge clearly what contribution the authors established to solve the problems of the OCR task.

  2. The results themselves are not very impactful. The OmniDocBench Overall score is actually lower than the base model. On speed, the page-level end-to-end gain is only 1.32×, and it shrinks to 1.13× at larger batch sizes, so the benefit is limited in high-load serving. To judge whether this trade-off is good, the base model should be made to emit more tokens at once and compared on accuracy at a speed similar to GravityOCR. The MTP in the paper was measured on a different serving engine, and its TPF excludes the drafting cost, so it is not a comparison on the same basis. I cannot tell whether the method is better than alternatives that lose no accuracy, so it is hard to assess its value.

(Minor details) There is no limitations section and no failure case analysis. Inputs where the verifier often rejects the draft and conditions where the speed gain shrinks are not summarized, and without this analysis users cannot judge how far the method applies.

Questions for authors
  1. What are the speeds in Table 1 if GLM-OCR's MTP is measured on SGLang, the same engine as GravityOCR? The GLM-OCR technical report reports 5.2 tokens per step and about 50% higher throughput with MTP, so why is it 3.7 TPF and slower than AR in this paper? Can it be compared on the same basis with MTP using more draft tokens?

  2. In one-shot drafting the whole block is masked and in the end it is resolved sequentially, so what does bidirectional attention actually contribute in the proposed method? If the draft is switched to causal attention under the same training, how does the acceptance length change?

  3. TiDAR handles drafting and verification in one forward pass, so why split them into two? Please state clearly how this paper's contribution differs from such similar prior work.

  4. Are there any other elements or analyses designed for OCR besides the reward? Before the OCR-targeted post-training, does the same method give the same gain on a general (M)LLM? After RL with OCR rewards, does the model lose its original general ability?

6 comments
AuthorsOpenReview ↗

Response to Reviewer PGVp (Part 1/2)

Thank you for the detailed review and for recognizing our page throughput, recognition accuracy, and breadth of experimental analysis. We also appreciate your positive assessment of the presentation, figures, and decoding and measurement procedures.


1. [W1, Q3, Q4] Contribution to OCR and Other Domains

GravityOCR builds on existing joint AR–diffusion self-speculation; we do not propose a new framework. We analyze OCR's parallel-generation potential and limitations, demonstrate practical acceleration with near-AR quality, and show that AR-path OCR-reward post-training improves recognition while preserving drafting efficiency.

Our additional Diffusion Gemma experiment found 3.6× lower mean draft entropy and 2.17× higher TPF on OCR than image description (23.57 vs. 10.88). This supports OCR as a favorable setting for parallel diffusion generation. However, parallel predictability alone does not ensure accurate transcription: GravityOCR's direct-diffusion outputs retain errors (Fig. 2), and AR verification improves quality at the same checkpoint (Sec. 4.3).

Resource constraints prevented applying our training recipe to a general (M)LLM. We instead compared OCR and image description using the general-purpose self-speculative VLMs Fast-dVLM and Nemotron. Neither achieved OCR-level TPF on image description: 1.62 vs. 2.65 for Fast-dVLM and 1.75 vs. 5.54 for Nemotron. Thus, even with the same model and decoding method, the evaluated non-OCR task did not reach OCR's drafting efficiency. This task dependence reinforces the value of analyzing and validating self-speculation for OCR. We will add these analyses.


2. [W2, Q1] Recognition Quality and Speed

The practical value is faster document processing while retaining recognition quality close to GLM-OCR. GravityOCR achieves 1.73× the page throughput of GLM-OCR AR while maintaining an Overall score of 95.16 versus 95.48 for GLM-OCR. Post-submission serving optimizations increased page throughput for both AR and GravityOCR; this comparison uses the improved AR baseline. At batch sizes 64 and 256, GravityOCR achieves 6,157.6 and 6,655.1 tok/s, respectively, versus 3,580.2 and 4,302.2 tok/s for AR decoding of the same checkpoint. The resulting 1.72× and 1.55× speedups show that the throughput advantage persists at the evaluated larger batch sizes.

In response, we evaluated GLM-OCR MTP and a 165M DFlash drafter trained for original GLM-OCR on our data, both in SGLang. To match the data used for DFlash training, we also fine-tuned GLM-OCR’s existing MTP head on GravityOCR’s training data and labels. We initialized it from the released MTP weights and trained for 22,000 steps with the rest of GLM-OCR frozen.

MethodOverallTokens/roundCrop tok/sPages/s
GLM-OCR AR95.481775 (1.00×)0.843 (1.00×)
MTP95.464.63984 (1.27×)0.970 (1.15×)
MTP + fine-tuned head95.469.601,849 (2.39×)1.153 (1.37×)
DFlash95.3910.281,953 (2.52×)1.143 (1.36×)
GravityOCR95.1619.364,975 (6.42×)1.455 (1.73×)

Tokens/round counts all committed tokens; crop throughput is decode-only.

Timings use the paper's SGLang protocols and GLM-OCR AR baseline. Default MTP was slower than AR (0.67× crop, 0.84× pages/s). For this comparison, we optimized MTP serving by enabling draft-extend CUDA graphs and reducing CPU-side and synchronization overhead. These optimizations increased crop decoding and page throughput from 0.67× and 0.84× to 1.27× and 1.15× relative to AR. The fine-tuned-head configuration achieved 2.39× and 1.37×, respectively.

To address your question about the reported token efficiency and longer drafts, we additionally varied the MTP draft length. We obtained 5.33 tokens per step at length 7, close to the reported 5.2, and 6.65 at length 20. The table's MTP row uses length 7 on 8,922 crops.

GravityOCR improves quality over direct diffusion (Sec. 4.3) and MinerU-Diffusion (Table 1). It achieves 1.455 pages/s versus 1.153 for MTP with a fine-tuned head and 1.143 for DFlash (1.26–1.27× higher), with 2.55× crop throughput over DFlash. We will add these results.

AuthorsOpenReview ↗

Response to Reviewer PGVp (Part 2/2)


3. [W1, Q3] Relation to TiDAR

TiDAR and GravityOCR both use a single shared backbone for diffusion drafting and AR verification. As discussed above, our contribution lies in the OCR-focused analysis and validation, rather than this shared self-speculation principle. TiDAR’s fusion verifies the current draft and prepares possible next drafts in one forward pass. This execution strategy is complementary to our contributions and can also be applied to GravityOCR without retraining.

In our 32-token-draft configuration, applying TiDAR-style fusion processes 1,089 token positions per forward, versus 33 per forward in two-pass execution. At batch 1 in Hugging Face, fusion achieves 1.24× end-to-end throughput over two-pass execution. However, fusion trades fewer sequential forwards for additional speculative computation. With a fixed draft length, this additional work can outweigh the savings from fewer forwards as larger batches become compute-bound. We therefore treat fusion as an optional, workload-dependent optimization rather than an unconditional replacement for two-pass execution.


4. [W1, Q2] Architectural Choices

The within-block attention pattern is an architectural choice within GravityOCR, rather than the central contribution of this work. We adopted bidirectional attention used in prior block-diffusion models, allowing draft positions to exchange contextual representations during parallel prediction. To examine this design choice, we additionally trained a causal-attention variant under the same data and loss recipe, following the reviewer’s suggestion. The comparison at 10k training steps is shown below:

Draft attentionTPFAccepted draft tokens/roundOverall
Bidirectional6.7511.5095.02
Causal7.8813.7794.96

At 10k steps, the causal variant yields more accepted draft tokens per round and higher TPF, while the Overall scores differ by 0.06 points. This provides a useful design insight and a promising direction for further improving drafting efficiency within the same OCR self-speculation approach. We will include this ablation in the revised paper and examine whether the trend persists with longer training.


5. [Q4] Performance after OCR-Reward RL

GravityOCR starts from GLM-OCR, which was already specialized for document OCR before our adaptation and GRPO. All three evaluated OCR tasks improve after GRPO: the Overall gain does not come at the expense of text, table, or formula recognition. Changes in non-OCR capabilities were not evaluated.

TaskBeforeAfter
Text (1−Edit)95.9695.98
Table (TEDS)92.2392.80
Formula (CDM)96.5696.71

6. [Minor] Limitations and Failure Cases

Short outputs offer limited speedups because decoding accounts for little total runtime. We will document these applicability limits and draft-rejection cases at word boundaries, punctuation, and LaTeX commands in the revised paper's limitations and failure analysis.

Thank you again for your detailed feedback. We have uploaded a revised manuscript, with revisions marked in blue. The updated paper clarifies the OCR-focused contribution and incorporates the same-engine MTP and DFlash comparisons, the TiDAR-style fusion experiment, and the causal-versus-bidirectional draft-attention ablation. It also includes the OCR-versus-image-description analysis, draft-rejection patterns, and a dedicated limitations section.

Following an additional draft-length sweep for the fine-tuned MTP head, we selected draft length 15, which achieved the highest throughput among the tested settings. We updated Table 2 accordingly: decode-only throughput increases from 1,849 to 1,917 tokens/s, and page throughput from 1.153 to 1.191 pages/s. GravityOCR achieves 1.455 pages/s, corresponding to 1.22× the page throughput of this updated MTP baseline.

Please let us know if you have any remaining concerns or would find further clarification helpful. If our responses and revisions have addressed your concerns, we would be grateful if you could consider revising your score upward.

Reviewer_iyAKOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents GravityOCR, which adapts GLM-OCR through joint autoregressive and block-diffusion training. At inference, the shared model drafts multiple tokens in parallel and uses its causal AR path to verify and accept the longest matching prefix. The authors further apply GRPO through the AR path using OCR-specific rewards. Recognition quality is evaluated on OmniDocBench, PubTabNet, and UniMER, while inference speed is measured in SGLang for both token decoding and full-page processing.

Critical strengths
  1. The method offers a useful speed-quality tradeoff in a practical OCR pipeline. Compared with AR decoding of the same checkpoint in SGLang, GravityOCR achieves a 3.94× decode-only speedup and a 1.32× end-to-end page-processing speedup. Its final OmniDocBench Overall score is 95.16, compared with 95.48 for the original GLM-OCR. These results support the practical value of shared-model self-speculative decoding for OCR.

  2. The experiments provide useful evidence for the main design choices. Comparing direct diffusion and self-speculative decoding with the same checkpoint helps isolate the benefit of causal verification. Ablations of the AR objective, drafting strategy, and GRPO, together with throughput measurements at different concurrency levels, help explain where the method is effective and where its gains become smaller.

Critical weaknesses
  1. The contribution beyond the closest prior work remains limited. Fast-dVLM and Nemotron-Labs-Diffusion already study joint AR-diffusion training and parallel drafting with causal verification within a shared model, including vision-language settings. The paper acknowledges these connections, but leaves unclear which additional challenges arise in OCR and which new design choices are necessary to address them. The AR-path GRPO and masking-strategy experiments provide useful empirical findings, although the current contribution remains primarily an adaptation and validation of this framework for OCR.

  2. The comparisons with other speculative decoding approaches do not establish the relative benefit of using the full shared backbone as the drafter. The GLM-OCR MTP baseline runs on vLLM, whereas GravityOCR runs on SGLang, so their speed difference includes both decoding-method and backend effects. The DFlash comparison uses a different backbone, HunyuanOCR-1.5. These system-level comparisons are informative, and the matched AR baseline supports the reported acceleration. However, the evidence for choosing the shared-model design over alternative drafting strategies remains incomplete.

Questions for authors
  1. Relative to directly adapting the Fast-dVLM or Nemotron-Labs-Diffusion recipe to GLM-OCR, which training or decoding changes are specific to this work, and which existing ablations establish their contribution? A comparison matched for backbone, data, and training budget would help clarify the added value.

  2. Could the authors include at least one comparison with MTP or a lightweight diffusion drafter using the same backbone and comparable serving settings, reporting recognition quality and wall-clock latency? This would help establish the relative merits of the shared-model approach for OCR.

3 comments

Thank you for your thoughtful review and for recognizing the practical speed–quality trade-off of shared-model self-speculation for OCR. We also appreciate your assessment that the same-checkpoint comparison and ablations provide useful evidence for our design choices. Below, we clarify our contribution and present additional comparisons under matched serving conditions.


1. [W1, Q1] OCR-Specific Contributions

GravityOCR builds on existing joint AR–diffusion self-speculation; we do not propose a new self-speculation framework. Our contribution is to analyze the potential and limitations of parallel generation in document OCR and, on this basis, demonstrate practical acceleration with near-AR recognition quality.

As an additional analysis for this response, we evaluated Diffusion Gemma on OCR and image description to examine OCR's suitability for parallel generation:

TaskMean draft entropy* (nats)TPF
OCR0.7423.57
Image description2.6710.88

*Entropy is averaged over positions k=1–32.

Compared with image description, OCR has 3.6× lower mean draft entropy and 2.17× higher TPF, supporting its suitability for parallel generation. However, directly committing diffusion predictions can still introduce omissions and repetitions, as shown by GravityOCR's failure cases (Fig. 2). Our same-checkpoint comparison shows that AR verification improves recognition quality over direct diffusion decoding (Sec. 4.3). We therefore use parallel predictions as drafts and let the causal AR path determine which tokens are committed.

We further show that OCR-reward post-training through the AR path improves recognition while preserving drafting efficiency. GRPO increases Overall from 94.92 to 95.16, with TPF nearly unchanged at 9.61 and 9.68 (Sec. 4.5). Continued SFT from the same starting checkpoint and RL data, with matched data exposure, reaches 94.98. Thus, GRPO provides a further 0.18-point improvement over continued SFT.

Separately, the AR-loss ablation (Sec. 4.4) and partial-versus-full masking comparison (Appendix D) quantify the effects of the joint-training recipe on verifier accuracy and drafting efficiency. These experiments support the adopted training configuration independently of the GRPO comparison. We will clarify these contributions and the roles of the experiments in the revised paper.


2. [W2, Q2] Comparison with MTP and DFlash

Thank you for the suggestion. In response, we evaluated GLM-OCR MTP in SGLang and trained a 165M-parameter DFlash drafter for original GLM-OCR on the same training data as GravityOCR, using the same engine. To match the data used for DFlash training, we also fine-tuned GLM-OCR’s existing MTP head on GravityOCR’s training data and labels. We initialized it from the released MTP weights and trained for 22,000 steps with the rest of GLM-OCR frozen. Since submission, our serving optimizations have increased page throughput for both AR and GravityOCR.

MethodOverallTokens/roundCrop tok/sPages/s
GLM-OCR AR95.481775 (1.00×)0.843 (1.00×)
MTP95.464.63984 (1.27×)0.970 (1.15×)
MTP + fine-tuned head95.469.601,849 (2.39×)1.153 (1.37×)
DFlash95.3910.281,953 (2.52×)1.143 (1.36×)
GravityOCR95.1619.364,975 (6.42×)1.455 (1.73×)

Tokens/round counts all committed tokens; crop throughput is decode-only.

Timing follows the paper's protocols, with all methods served in SGLang. SGLang's default MTP path was slower than AR (0.67× crop decoding, 0.84× pages/s). For this comparison, we optimized MTP serving by enabling draft-extend CUDA graphs and reducing CPU-side and synchronization overhead. These optimizations increased crop decoding and page throughput from 0.67× and 0.84× to 1.27× and 1.15× relative to AR. The fine-tuned-head configuration achieved 2.39× and 1.37×, respectively. All table speedups are relative to the GLM-OCR AR row.

Despite using the full backbone for drafting rather than a smaller auxiliary module, GravityOCR achieves 6.42× crop decoding throughput and 1.73× page throughput over GLM-OCR AR, and 2.55× crop decoding and 1.27× page throughput over DFlash. We will add these comparisons to the paper.

Thank you again for your detailed feedback. We have uploaded a revised manuscript, with revisions marked in blue. The updated paper clarifies our OCR-focused contribution relative to prior shared AR–diffusion methods and incorporates the OCR-versus-image-description analysis and the MTP and DFlash comparisons using GLM-OCR and SGLang.

Following an additional draft-length sweep for the fine-tuned MTP head, we selected draft length 15, which achieved the highest throughput among the tested settings. We updated Table 2 accordingly: decode-only throughput increases from 1,849 to 1,917 tokens/s, and page throughput from 1.153 to 1.191 pages/s.

Please let us know if you have any remaining concerns or would find further clarification helpful. If our responses and revisions have addressed your concerns, we would be grateful if you could consider revising your score upward.

Reviewer_iyAKOpenReview ↗

Response

Thank you for the detailed response and the additional same-engine comparisons with MTP and DFlash, which address my second question. The clarified scope of the contribution is also helpful. I have updated my review and raised my rating to 3.

Reviewer_TX1aOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper uses a shared model for block-diffusion drafting and autoregressive verification, then improves OCR quality through GRPO on the autoregressive path. It aims to reduce sequential decoding steps without a separate drafting network while largely preserving the original model's recognition quality.

Critical strengths
  1. The engineering benefit is clear. With the same checkpoint and serving stack, decoding is approximately 4x faster and full-page processing approximately 1.3x faster, while recognition quality remains close to the original model. These timing scopes must be distinguished, but the full-page measurement itself provides evidence of practical deployment value.

  2. Evaluation covers different benchmarks, output lengths, batch sizes, and content types, with documented timing scopes and serving stacks. Experiments also support compatibility between AR-path optimization and the shared drafter, making the contribution more than an isolated inference trick.

Critical weaknesses
  1. Novelty lies primarily in OCR adaptation and training/serving integration. Fast-dVLM and Nemotron already explore shared AR-diffusion models and self-speculation. Although cross-backbone comparisons such as HunyuanOCR-DFlash are included, a closely related approach with the same backbone, serving stack, and a similar budget would better establish the quality-speed-extra-parameter trade-off.

  2. The GRPO quality gain is small, without uncertainty analysis or an equal-budget continued-SFT control, so its independent benefit has not been sufficiently isolated. The authors disclose finite-precision output discrepancies; evaluating quality on that subset would strengthen the evidence, but differing strings alone do not establish an overall quality decline.

Established algorithmic foundations do not negate the engineering contribution. OCR adaptation, same-checkpoint acceleration, and serving-level evaluation provide credible practical value. The uncertain independent benefit of GRPO is a secondary limitation, not grounds to overturn the main acceleration result.

Questions for authors

1.Can you provide a matched comparison with a closely related approach and show whether the GRPO gain exceeds evaluation uncertainty or the benefit of continued supervised training? 2.Can you also report recognition quality on the numerically divergent subset?

2 comments

Thank you for the careful review and for recognizing the practical value of our same-checkpoint acceleration and serving evaluation. We also appreciate your positive assessment of the breadth of our experiments and the compatibility between AR-path optimization and the shared drafter. Below, we address your suggestions with additional comparisons and analyses.


1. [W1] Contribution to Document OCR

GravityOCR's contribution is to analyze the potential and accuracy limitations of parallel generation in document OCR and demonstrate near-AR recognition quality with practical end-to-end acceleration using existing AR–diffusion self-speculation.

OCR outputs are strongly constrained by the input image, making parallel generation promising. In an additional experiment for this response, Diffusion Gemma showed 3.6× lower mean draft entropy (k=1–32) and 2.17× higher TPF on OCR than on image description (23.57 vs. 10.88), supporting this potential. However, GravityOCR's direct-diffusion outputs still show omissions and repetitions (Fig. 2) and lower recognition quality than self-speculation at the same checkpoint (Sec. 4.3).

These additional results support combining parallel drafting with AR verification. We further show that AR-path post-training improves recognition while retaining drafting efficiency. We will clarify this contribution and add the new results to the paper.


2. [W1, Q1] Comparison with MTP and DFlash

Thank you for the suggestion. In response, we evaluated GLM-OCR MTP in SGLang and trained a 165M-parameter DFlash drafter for original GLM-OCR on the same training data as GravityOCR, using the same engine. To match the data used for DFlash training, we also fine-tuned GLM-OCR’s existing MTP head on GravityOCR’s training data and labels. We initialized it from the released MTP weights and trained for 22,000 steps with the rest of GLM-OCR frozen. Since submission, our serving optimizations have increased page throughput for both AR and GravityOCR.

MethodOverallTokens/roundCrop tok/sPages/s
GLM-OCR AR95.481775 (1.00×)0.843 (1.00×)
MTP95.464.63984 (1.27×)0.970 (1.15×)
MTP + fine-tuned head95.469.601,849 (2.39×)1.153 (1.37×)
DFlash95.3910.281,953 (2.52×)1.143 (1.36×)
GravityOCR95.1619.364,975 (6.42×)1.455 (1.73×)

Tokens/round counts all committed tokens; crop throughput is decode-only.

Timing follows the paper's protocols, with all methods served in SGLang. SGLang's default MTP path was slower than AR (0.67× crop decoding, 0.84× pages/s). For this comparison, we optimized MTP serving by enabling draft-extend CUDA graphs and reducing CPU-side and synchronization overhead. These optimizations increased crop decoding and page throughput from 0.67× and 0.84× to 1.27× and 1.15× relative to AR. The fine-tuned-head configuration achieved 2.39× and 1.37×, respectively. All table speedups are relative to the GLM-OCR AR row.

Despite using the full backbone for drafting rather than a smaller auxiliary module, GravityOCR achieves 6.42× crop decoding throughput and 1.73× page throughput over GLM-OCR AR, and 2.55× crop decoding and 1.27× page throughput over DFlash. We will add these comparisons to the paper.


3. [W2, Q1] GRPO Gains and the Continued-SFT Control

Following your suggestion, we compared continued SFT and GRPO from the same starting checkpoint, using the RL data and reference labels with matched data exposure. OmniDocBench Overall is 94.92 before RL, 94.98 after SFT, and 95.16 after GRPO. GRPO achieves an additional 0.18-point gain over continued SFT, while TPF remains nearly unchanged at 9.61, 9.64, and 9.68, respectively. Thus, GRPO provides an additional recognition improvement while preserving drafting efficiency.

To address whether the GRPO gain exceeds evaluation uncertainty, we performed a paired page-level randomization test (sign-flip) [1] on 1,641 scored pages, using 100,000 randomizations and the benchmark's Overall aggregation. The +0.24-point gain over the pre-GRPO checkpoint was statistically significant (p = 0.0022).


4. [W2, Q2] Quality of Numerically Divergent Outputs

Among the 562 pages with differing output strings, 390 receive tied recognition scores. Applying the benchmark's Overall aggregation to this subset gives 94.09 for AR and 94.15 for self-speculation. AR is better on 90 pages and self-speculation on 82. Thus, the observed string-level discrepancies did not translate into a degradation in aggregate recognition quality.


References

[1] Stefan Riezler and John T. Maxwell. 2005. On Some Pitfalls in Automatic Evaluation and Significance Testing for MT.

Thank you again for your thoughtful feedback. We have uploaded a revised manuscript incorporating the experiments and clarifications discussed in our response, with revisions marked in blue. The updated paper clarifies the OCR-focused contribution and includes the matched MTP and DFlash comparisons, the continued-SFT control, the statistical analysis of the pre-/post-GRPO improvement, and the recognition-quality analysis of numerically divergent outputs.

Following an additional draft-length sweep for the fine-tuned MTP head, we selected draft length 15, which achieved the highest throughput among the tested settings. We updated Table 2 accordingly: decode-only throughput increases from 1,849 to 1,917 tokens/s, and page throughput from 1.153 to 1.191 pages/s.

Please let us know if you have any remaining questions or concerns. Thank you again for helping us improve the paper.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.