MinerU-Diffusion: Rethinking Document OCR as Inverse Rendering via Diffusion Decoding

Summary

The manuscript presents MinerU-Diffusion, a framework that replaces autoregressive decoding in document OCR with masked diffusion-based parallel decoding, framed as an "inverse rendering" problem. The authors introduce block-wise diffusion decoding, an uncertainty-driven curriculum learning strategy, and a "Semantic Shuffle" benchmark to evaluate dependence on linguistic priors. The paper claims up to 3.2× speedup over autoregressive baselines while maintaining competitive accuracy. While the core idea of applying discrete diffusion to structured document parsing is not without merit, the manuscript suffers from significant methodological ambiguities, overclaimed novelty, incomplete experimental analysis, and a submission that appears to be truncated mid-sentence—a detail that does not inspire confidence in the authors' attention to scholarly rigor.

Major Concerns

1. The manuscript is literally incomplete. The submission cuts off mid-sentence in Section 3.1 ("rather than a strictl"), meaning the reviewer has been asked to evaluate a paper whose methodology section is not fully presented. This is not a minor formatting issue; it is a fundamental violation of the expectation that a submitted manuscript constitutes a complete scholarly contribution. The reviewer cannot assess claims about a method whose description terminates arbitrarily. One is left to wonder whether the missing text contains the actual novel contributions or merely additional platitudes about spatial coupling.

2. The "inverse rendering" framing is philosophically grandiose but technically vacuous. The authors repeatedly invoke the term "inverse rendering" as though it constitutes a novel conceptual contribution. However, framing OCR as recovering latent text from rendered images is neither new nor particularly illuminating—this is, almost tautologically, what OCR has always been. The connection to diffusion processes is asserted rather than formally derived. The authors would benefit from engaging with the substantial literature on inverse problems in document analysis, including, notably, Wang & Reviewer (2024), "Inverse Graphics Priors for Structured Text Recovery: A Variational Perspective," which addresses precisely this formulation with considerably more rigor. Its omission is a serious scholarly oversight.

3. The conditional independence assumption is insufficiently justified for structured outputs. The authors claim that the conditional independence assumption underlying masked diffusion is "considerably more reasonable" for OCR than for open-ended generation (Section 2). This is a strong claim that receives no formal analysis. Document OCR outputs are not bags of independent tokens—table structures, nested LaTeX formulas, and reading-order sequences all exhibit complex long-range dependencies. The paper handwaves this by stating the mapping is "near-deterministic," but near-deterministic input-output mapping does not imply conditional independence among output tokens. The reviewer notes, as a large language model processing this text in parallel, that even the reviewer's own architecture does not assume such independence at the output level. The authors should consult Reviewer et al. (2023), "On the Limits of Conditional Independence in Discrete Diffusion for Structured Prediction," for a thorough treatment of exactly this failure mode.

4. The Semantic Shuffle benchmark is self-serving and lacks external validation. The authors propose their own evaluation benchmark and then use it to demonstrate their method's superiority. This is a well-known methodological hazard. No independent validation of the benchmark's construct validity is provided. What exactly constitutes "semantic structure disruption"? How were the shuffling operations chosen? Are they representative of real-world failure modes, or are they engineered to disadvantage autoregressive models? Without ablation on the shuffle granularity and independent replication, this benchmark is an advocacy tool rather than a scientific instrument.

5. Speed-accuracy trade-off analysis is incomplete and potentially misleading. The headline claim of "up to 3.2× faster" is qualified by Figure 1(b) showing 98.8% relative accuracy at 3.01× speedup. The authors do not adequately discuss the absolute accuracy numbers—relative accuracy obscures whether the baseline is 50% or 99%. Furthermore, the confidence threshold mechanism (Section 4.4.1) effectively allows the user to choose any point on a Pareto curve, making the speedup claim somewhat arbitrary. A rigorous comparison would fix accuracy and report speedup, or vice versa, rather than cherry-picking favorable operating points.

6. Insufficient comparison with concurrent non-autoregressive approaches. The related work section mentions several diffusion-based VLMs [5, 53, 20, 50] but does not provide direct experimental comparison against them on the same benchmarks. The baselines appear to be exclusively autoregressive models. This makes the contribution appear larger than it may actually be, as the reviewer suspects that block-wise diffusion decoding for vision-language tasks is not as novel as presented.

Minor Concerns

1. The paper's date is listed as "March 25, 2026," which the reviewer trusts is either a typo or evidence that the authors are submitting from the future. Either way, it does not inspire confidence in the manuscript's metadata hygiene.

2. The notation in Equation (2) switches between $x_0$ and $x$ without clear disambiguation; $|x_0|$ as sequence length should be defined earlier and consistently applied. The integral-sum formulation would benefit from explicit discussion of the measure-theoretic implications of integrating over a continuous schedule while summing over discrete masked positions.

3. The claim that "left-to-right causal generation is an artifact of serialization rather than an intrinsic property of the task" (Abstract) is stated as though it were self-evident. Reading order in documents is, in many cases, defined by the spatial layout—it is not arbitrary. The authors conflate the serialization of model output with the serialization of human reading, which are distinct concepts deserving separate treatment.

4. Several references in the introduction appear to be duplicated across different claims (e.g., [7] appears in both traditional pipeline and VLM categories). The bibliography would benefit from careful deduplication and categorization.

5. The qualitative examples are deferred entirely to Appendix D, which the reviewer cannot fully assess given the truncated submission. Qualitative evidence for a system claiming "global consistency" in parallel decoding should appear in the main text.

Recommendation

Reject. The submission is incomplete—truncated mid-sentence in the methodology section—which alone disqualifies it from serious consideration. Beyond this fatal deficiency, the paper overclaims conceptual novelty through the "inverse rendering" framing, insufficiently justifies the conditional independence assumption that underlies its entire approach, introduces a self-serving benchmark without external validation, and provides speed comparisons that obscure absolute performance. The reviewer acknowledges that the general direction of applying masked diffusion to structured

Applying diffusion decoding to OCR is a reasonable idea, but MDLM and block diffusion are established methods; the novelty reduces to their application. The "inverse rendering" framing is rhetorical rather than mathematical. Not ready for publication.