Diffusion Drafts, AR Verifies: Accelerating Document OCR with Self-Speculative Decoding
Dohyun Kim, Sungjun Han, Hyungguk Kim, Yusik Kim, Jamin Shin, Paul Hongsuck Seo, Hongjoon Ahn
cs.CL · cs.CV
Quick verdict
GravityOCR turns GLM-OCR into a shared diffusion-drafter/AR-verifier: 3.94× decode speedup at −0.32 Overall; rigorous ablations, one base model, no CIs.
Summary
The paper introduces GravityOCR, which converts the pretrained autoregressive OCR model GLM-OCR into a single parameter-shared model supporting both block-diffusion drafting and causal AR verification. Each training example passes through one clean AR stream plus two complementary corrupted diffusion streams, jointly optimizing next-token and denoising objectives. At inference, one forward pass produces a causal boundary token and a 32-token parallel draft; a causal verification pass commits only the longest AR-consistent draft prefix, so output matches greedy AR decoding in exact arithmetic — verified at 96.6% string identity under bf16 across 8,922 crops, with divergences traced to near-tie argmax flips. GRPO runs through the causal path with task-aware rewards (edit similarity, TEDS-based table scores, canonicalized LaTeX), avoiding diffusion-trajectory likelihood estimation. On OmniDocBench v1.6, Overall improves from 94.92 to 95.16 after GRPO, 0.32 below the original GLM-OCR (95.48). In SGLang the model commits 9.7 tokens per forward, giving 3.94× decode-only and 1.32× page-level speedups, narrowing to 1.13× at batch 64; direct block-diffusion commitment reaches only 92.53 Overall at comparable parallelism. Limitations: a single base model, predominantly English training with speed measured on the English subset only, no significance testing of small deltas, and mixed inference stacks across compared systems.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A pretrained causal AR OCR model can be adapted into a parameter-shared block-diffusion drafter plus causal AR verifier, adding only a mask-token embedding and no separate drafting network.
- Self-speculative decoding is output-equivalent to greedy AR decoding in exact arithmetic, with 96.6% string identity across 8,922 crops under bf16 serving kernels.
- GRPO applied only through the causal AR path improves OmniDocBench v1.6 Overall from 94.92 to 95.16 while tokens-per-forward stays ~9.7.
- In SGLang serving, GravityOCR commits 9.7 tokens per forward and delivers a 3.94× decode-only and 1.32× page-level speedup, whereas direct confidence-based diffusion commitment drops to 92.53 Overall at comparable parallelism.
Strengths
Exemplary AR-equivalence auditing
Appendix B gives a sound induction argument for exact equivalence under greedy decoding, then honestly quantifies bf16 divergence: 96.6% string identity on 8,922 crops, 305 first divergences with median top-two logit margin of 0.14 nats, and an fp32 re-check attributing flips to near-ties (65 below 0.05 nats).
Parameter-shared design with honest accounting
Only a mask-token embedding is added — no drafter network or auxiliary head (§3.1); unlike the † rows of Table 1 (MTP, DFlash), GravityOCR's TPF of 9.7 counts both draft and verification forwards, a conservative convention.
Unusually complete ablations
RL without diffusion likelihoods
Transparent speed protocols
Weaknesses
No uncertainty quantification
PubTabNet gain is confounded
Cross-stack speed comparisons
English-centric evaluation
Single base model
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
GravityOCR targets decode-time acceleration of generative document OCR. The authors convert the released GLM-OCR checkpoint into a single parameter set serving two attention regimes: block diffusion (bidirectional within a 32-token block, causal across blocks) for parallel drafting, and token-level causal attention for verification. Four claims anchor the paper (§1, §4): (1) the adaptation needs only a mask-token embedding — no drafter network or auxiliary head; (2) self-speculative decoding is output-equivalent to greedy AR decoding, exact in exact arithmetic and 96.6% string-identical across 8,922 English-subset crops under bf16 (Appendix B.3); (3) GRPO applied solely through the causal path lifts OmniDocBench v1.6 Overall from 94.92 to 95.16 while TPF stays ~9.7 (Table 6); (4) SGLang serving yields 9.7 tokens per forward, a 3.94× decode-only speedup and a 1.32× page-level speedup (Tables 1, 3).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is GravityOCR?
GravityOCR is a document-OCR model created by adapting the pretrained GLM-OCR model so that one shared set of parameters supports two decoding modes: block diffusion, which drafts many tokens in parallel, and causal autoregressive decoding, which verifies each draft before any token is committed — speeding up inference without training a separate drafting model.
How much faster is GravityOCR than standard autoregressive OCR?
In an SGLang deployment on one H100 it commits about 9.7 output tokens per forward pass, yielding a 3.94× decode-only speedup on region crops and a 1.32× end-to-end page-processing speedup. The advantage shrinks to 1.13× at batch size 64, so it helps most in low-concurrency settings.
Does parallel drafting hurt OCR accuracy?
By design, no: the causal verifier only commits draft tokens that match its own next-token predictions, so the output equals greedy autoregressive decoding in exact arithmetic (96.6% of crops are string-identical under bf16). The final model scores 95.16 on OmniDocBench v1.6 versus 95.48 for the original GLM-OCR — a small drop of 0.32 points.
Why is GRPO applied on the AR path instead of the diffusion path?
Sequence likelihood in masked diffusion requires marginalizing over unmasking trajectories, which is hard to estimate. Because GravityOCR's verifier has an exact causal factorization, standard GRPO with exact token log-probabilities can be used, and since parameters are shared, the diffusion drafter is updated too — no diffusion-specific RL objective is needed.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–7 and the qualitative Figures 8–10 are available only as extracted captions and axis labels; curve shapes and example renderings could not be independently inspected.
- Code and weights appear as link badges in the PDF; the actual URLs and repository contents were not extractable from the text, so availability is asserted, not verified.
- Competitor speed numbers depend on each system's official stack and tuning level (Appendix C); their optimization quality cannot be audited from the paper alone.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.