- 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
-
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.
-
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.
-
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.
-
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.
-
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.