acceptodds
Desk rejected at ICLR 2027

MinerU-Diffusion v2: Bridging Autoregressive and Diffusion Decoding for Document OCR

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Diffusion decoding offers a new path to parallel generation for document OCR, but training from scratch leaves the capabilities of existing autoregressive parsers underused. We introduce MinerU-Diffusion v2, a 1.2B-parameter block diffusion document parser initialized from MinerU2.5-Pro. The design balances decoding parallelism and recognition quality while accounting for the downstream effects of layout decisions. Autoregressive-to-Diffusion Continued Training (ADCT) reuses existing document supervision through progressive block-size training and smaller-block replay, enabling a single checkpoint to support multiple inference block sizes. Hierarchical Joint Policy Optimization (HJPO) uses downstream rewards and conditional credit assignment to optimize final parsing quality, alternately updating a shared policy for layout generation and content recognition. On OmniDocBench v1.5, the model achieves Overall scores of 92.63 with predicted layouts and 94.82 with ground-truth layouts, outperforming standard MinerU2.5 in both settings. Compared with MinerU-Diffusion, it uses fewer than half the parameters while improving the respective scores by 3.69 and 1.45 points. Effective output throughput increases from 131.22 to 152.36 tokens per second, with flexible speed–accuracy trade-offs supported by configurable inference block sizes.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Document Parsing, Optical Character Recognition, Block Diffusion Language Model, Reinforcement Learning, Vision Language Model, OCR

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_BBxBOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MinerU-Diffusion v2, a 1.2B-parameter block-diffusion document parser converted from an autoregressive MinerU2.5-Pro checkpoint. Its two components are (i) Autoregressive-to-Diffusion Continued Training (ADCT), which progressively increases the denoising block size, replays smaller blocks, and then contracts the main block size; and (ii) Hierarchical Joint Policy Optimization (HJPO), which uses a page-level parsing reward and nested layout/recognition rollouts to train the same policy for two stages. The authors report 92.63 overall on OmniDocBench v1.5 with predicted layouts, 94.82 with ground-truth layouts, and 152.36 effective output tokens/s at the quality-oriented operating point

Critical strengths

Relevant efficiency problem and useful reuse of an existing model. AR-to-diffusion conversion avoids building a document parser from scratch and enables a tunable parallelism/accuracy trade-off. The paper reports results on text, formulas, tables, and reading order rather than a single recognition task.

Meaningful ablation of the downstream-layout objective. HJPO's idea of judging layout choices by how they affect final recognition is well motivated. Table 6 compares independent GRPO, a shared downstream reward, and conditional comparisons with matched nested rollout counts for the last two conditions. These are more informative than relying only on final benchmark scores.

Strong practical efficiency signal. The reported model is smaller than the prior 2.5B MinerU-Diffusion model and improves both its benchmark score and reported token throughput. The element-aware decoding results demonstrate an adjustable quality/parallelism frontier

Critical weaknesses
  1. The most relevant autoregressive baseline is the model used to initialize the submission, but v2 does not match it in quality. The abstract and main comparison emphasize improvements over standard MinerU2.5 and MinerU-Diffusion. However, Table 6 reports the MinerU2.5-Pro initializer at 93.07 predicted-layout / 95.39 GT-layout, compared with 92.63 / 94.82 for the final v2 model. The converted and RL-trained model therefore remains 0.44 / 0.57 points below its own initializer, despite its throughput gain. On predicted layouts it exceeds PaddleOCR-VL (92.56) by only 0.07 points (Table 1). This does not negate the efficiency contribution, but it fundamentally changes its interpretation: the main value proposition is a quality–latency trade-off relative to Pro, not an unqualified quality improvement. The manuscript should foreground a directly matched Pro-versus-v2 accuracy/latency/compute Pareto comparison and avoid using weaker historical checkpoints as the primary quality claim.

  2. The speed evaluation does not sufficiently establish end-to-end document throughput. Table 3 reports TPF and effective output tokens/s, but document OCR also incurs page layout generation, image encoding, cropping, visual-token prefill, assembly, and possibly separate recognition calls. The supplementary evaluation description states that the authors' evaluations use H200s, but the table contains starred entries using official numbers when consistent decoding could not be run. A speedup in output-token decoding need not translate proportionally into faster complete-document parsing. Please report matched end-to-end pages/s (or pages/min), median/p95 page latency, peak GPU memory, image resolutions, batch/concurrency settings, and cost per document for MinerU2.5-Pro and v2 on the same hardware/runtime. Quality should be compared at matched document-level latency, not only at one selected decoding setting. These measurements are central rather than peripheral for an efficiency paper.

  3. The proposed ADCT curriculum is not adequately isolated. ADCT combines conversion of a pretrained AR model, progressive expansion from block size 1 to 64, replay of smaller block sizes, and contraction to size 16. Table 4 varies data duration/recipe, initialization, and inference block length, but does not compare fixed-block conversion vs. progressive expansion, with vs. without smaller-block replay, or expansion-only vs. expansion-and-contraction under the same training exposures. Consequently, the evidence supports the feasibility of conversion but does not establish which claimed design component enables multi-block-length competence or retains recognition quality. In addition, Appendix E.2 describes a 'small fraction' of replay samples without an explicit per-stage sampling probability/schedule, complicating reproduction. These are the highest-value missing ablations for the first contribution.

  4. HJPO's incremental benefit is modest, and the conditional-credit contribution needs a more direct attribution. In Table 6, the final HJPO checkpoint improves over ADCT from 91.81 to 92.63 on predicted layouts (+0.82) and from 94.32 to 94.82 with GT layouts (+0.50). The GRPO-only configuration actually drops below ADCT. The downstream-coupled and HJPO variants do use the same four layouts/eight OCR trajectories per layout (Appendix E.3), which is good. However, comparison to independent GRPO simultaneously changes its input crops and its rewards; the effect of page-level reward is not cleanly separated from training on predicted layouts. Also, nested rollouts can be costly: four layouts times eight recognition trajectories per page is substantial for multi-region documents. Please report training GPU-hours/number of generated recognition sequences and test a simple shared-page-reward baseline with predicted-layout exposure but without nested conditional normalization. The theory in Appendix A analyzes ideal on-policy score-function/unnormalized leave-one-out estimators; the actual implementation adds group normalization, clipping, KL, and sequential shared-parameter updates. These are reasonable practices, but the idealized variance arguments should not be presented as a proof for the implemented optimizer.

  5. The adaptive decoding comparison changes two knobs simultaneously. In Table 5, uniform block-size configurations use threshold 0.95, whereas element-aware configurations use 0.92/0.96/0.92 by element. Hence the observed gain cannot be attributed uniquely to adaptive block sizing. A uniform-block baseline using the same per-element thresholds, or a threshold-only ablation, would resolve this issue. Likewise, the claimed 11.8% conversion-exposure comparison against training MinerU-Diffusion from scratch excludes the extensive pretraining already invested in the Pro initializer. Please describe this specifically as additional conversion samples rather than evidence of lower total training expense.

Questions for authors

What are fully matched end-to-end page latency/throughput and quality for the Pro initializer versus the final v2 model, including both stages and identical hardware/runtime?

What do the three ADCT component ablations (fixed versus progressive block size; replay versus none; contraction versus none) show at the same training exposure?

How much does each HJPO modification help when the use of predicted-layout crops, downstream reward, rollout counts, and KL/training budget are individually controlled?

Does the small predicted-layout advantage over strong alternatives persist across document categories, seeds, and a matched operating point? Please give uncertainty and per-category trade-offs.

Reviewer_TaThOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents MinerU-Diffusion v2, a 1.2B document parser converted from MinerU2.5-Pro. ADCT gradually changes the training block size and replays smaller blocks to support different decoding configurations within one checkpoint. HJPO jointly optimizes layout generation and content recognition using page-level rewards, with recognition trajectories compared under the same layout. The model improves over MinerU-Diffusion v1 in both parsing accuracy and reported throughput, although its accuracy remains below the AR initializer.

Critical strengths
  1. The conversion approach is practically useful. Reusing an existing OCR model and its supervision reduces the additional training needed to obtain a diffusion parser. The resulting model is also substantially smaller than v1, while improving predicted-layout Overall from 88.94 to 92.63. This is a meaningful improvement over the previous system.

  2. HJPO addresses a real issue in two-stage document parsing. A layout with better geometric overlap does not necessarily produce crops that are easier to recognize. Evaluating layouts through downstream parsing quality makes sense, and comparing recognition outputs under the same layout is a reasonable way to reduce interference from layout differences. The improvement over ADCT alone, from 91.81 to 92.63 with predicted layouts, supports the usefulness of this design.

Critical weaknesses
  1. The advantage over an accelerated AR parser is not established. MinerU2.5-Pro remains more accurate with both predicted layouts (93.07 versus 92.63) and GT layouts (95.39 versus 94.82). This does not establish a lower ceiling for diffusion OCR, but it leaves the accuracy cost of conversion unresolved. Since speed is a main reason for accepting this cost, the relevant comparison should include the same AR model with MTP-based or speculative decoding. Appendix B discusses these alternatives, but does not provide a matched experiment. Without such a comparison, it is difficult to tell whether direct diffusion offers a better practical trade-off.

  2. The evidence for robustness is narrower than the intended application. Semantic Shuffle measures sensitivity to disrupted word order, but it does not directly measure how often a model replaces visible characters with plausible text. It also does not establish whether the model retains useful contextual reasoning for difficult documents. Results on realistic acquisition conditions are missing. In addition, the gains from the conversion recipe are not fully attributed: the paper does not separately test smaller-block replay or the contraction stage, and the element-aware decoding comparison changes both block sizes and confidence thresholds.

Questions for authors
  1. What is the speed–accuracy trade-off against an accelerated version of the AR initializer?

Could the authors compare against MinerU2.5-Pro with MTP-based or speculative decoding under the same hardware, serving engine, batch size, and input conditions? Page-level latency and throughput would be more informative than decoding TPS alone. If an accelerated AR model preserves the initializer’s accuracy at a similar speed, what practical advantage remains for direct diffusion? I would also like to understand which errors account for the remaining accuracy gap after conversion.

  1. Can the reduction in hallucination be measured more directly?

The Semantic Shuffle results are useful, but low sensitivity to word-order changes is not the same as a low hallucination rate. Could the authors evaluate on CHAOS-Bench, released with HunyuanOCR-1.5, or a comparable character-corruption benchmark? In particular, please report errors at altered character positions separately from ordinary transcription errors, using the AR initializer and v1 as controls.

  1. Does the change in decoding affect useful contextual understanding?

Block decoding still conditions on preceding blocks, so I do not assume that it removes context. However, reduced reliance on linguistic context can have both benefits and costs. Could the authors examine ambiguous characters, long formulas, table consistency, and relationships across regions as block size changes? The current shuffle experiment tests resistance to misleading context, but not the ability to use context when it is needed.

  1. Do the gains hold on documents captured under realistic conditions?

Results on Real5-OmniDocBench and Wild-OmniDocBench, compared with MinerU2.5-Pro, would help answer this. Please include predicted-layout results, since recognition on GT crops would miss failures introduced by layout detection. A breakdown by acquisition condition would also show whether the conversion introduces particular weaknesses.

  1. Which parts of ADCT are responsible for the improvement?

Could the authors provide matched comparisons with a fixed training block size, without smaller-block replay, and without contraction from 64 to 16? For element-aware decoding, please vary block allocation and confidence thresholds separately. Table 4 establishes the effect of training duration and initialization, but does not isolate these proposed design choices.

  1. How much of HJPO’s benefit depends on diffusion decoding?

The conditional layout–recognition reward construction appears applicable to an AR parser as well. An AR + HJPO comparison would clarify whether this is a general improvement to document parsing or whether it particularly benefits diffusion. For the existing HJPO comparison, repeated runs or uncertainty estimates would also help assess the additional 0.59-point gain over downstream-coupled GRPO.

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

This paper presents MinerU-Diffusion v2, a 1.2B-parameter document OCR parser obtained by converting the 1.2B MinerU2.5-Pro autoregressive parser into a block-diffusion decoder. The proposed Autoregressive-to-Diffusion Continued Training (ADCT) uses progressive block-size expansion/contraction and replay of smaller block sizes. The paper additionally proposes Hierarchical Joint Policy Optimization (HJPO), which uses page-level parsing rewards and conditional leave-one-out credit assignment to jointly optimize layout generation and recognition through a shared policy. On OmniDocBench v1.5, the model reports Overall scores of 92.63 with predicted layouts and 94.82 with ground-truth layouts, together with 152.36 effective output tokens/s at one operating point. It also reports results on CC-OCR, OCRBench v2, UniMER-Test, and a semantic-shuffle benchmark. Ablations study training duration, initialization, block allocation, and the two HJPO components.

Critical strengths
  1. Strong practical result at a smaller model size. The proposed 1.2B model improves over the reported 2.5B MinerU-Diffusion baseline from 88.94 to 92.63 with predicted layouts and from 93.37 to 94.82 with ground-truth layouts (Table 1), while also reporting higher throughput. This is a consequential result for document parsing systems where memory and decoding latency matter.

  2. The paper targets an important coupling in document parsing. HJPO explicitly evaluates recognition under candidate layouts and uses final page-level utility rather than only layout IoU. The cumulative ablation in Table 6 gives a useful decomposition: downstream coupling improves over independent GRPO, and conditional credit assignment gives an additional improvement. The configurable block lengths and element-aware allocation also expose a useful quality–parallelism trade-off.

Critical weaknesses
  1. The contribution of ADCT is not sufficiently isolated. The conversion strategy shares several ingredients with prior AR-to-diffusion methods, particularly Fast-dLLM v2 and Fast-dVLM. While the results are promising, Table 4 mainly varies training duration and initialization, without ablating smaller-block replay or the expansion–contraction schedule. It is therefore unclear how much of the capability retention comes from the proposed curriculum rather than the strong MinerU2.5-Pro initialization. A simpler conversion baseline under a matched budget would help clarify the contribution.

  2. The gains from HJPO and the decoding strategy need more careful attribution. Table 6 provides encouraging results, but the leave-one-out reward in Eq. (4) is a rescaled group-mean-centered reward, which becomes nearly equivalent after group-wise standardization. This suggests that the improvement may mainly come from layout-conditioned grouping rather than the leave-one-out construction. Moreover, Table 5 changes both block sizes and confidence thresholds, making it difficult to isolate the effect of element-aware block allocation. More controlled comparisons would strengthen these claims.

Questions for authors
  1. Since the leave-one-out reward in Eq. (4) is a rescaled group-mean-centered reward, would layout-conditioned grouping with ordinary GRPO-style normalization yield similar results? This comparison would clarify which part of HJPO accounts for the gain in Table 6.
  2. Could the authors independently ablate smaller-block replay and the final contraction stage in ADCT? A simpler conversion baseline under matched training conditions would help establish whether the proposed curriculum is necessary.
  3. In Table 5, block sizes and confidence thresholds change together. Could the authors report a controlled comparison that varies only block sizes, and clarify whether the inference settings were selected on a validation set independent of OmniDocBench test labels?

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

Related papers

Discussion (0)

Sign in to comment.