acceptodds
Withdrawn from ICLR 2027

MinerU-Popo: Universal Post-Processing Model for Structured Document Parsing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

VLM-based OCR models have become the de facto choice for document parsing, as they can accurately extract page-level elements (e.g., paragraphs within individual pages) together with their bounding boxes and textual content. However, downstream applications such as RAG require coherent document-level information, whereas these models often break cross-page continuity and fail to recover disrupted structures, such as paragraphs and tables truncated by page boundaries. Such relationships are not confined to a single page; instead, they require joint analysis of titles, paragraphs, tables, and images spanning multiple pages. A natural solution is therefore to reuse existing OCR outputs and reconstruct document-level logical structures through post-processing. To this end, we propose MinerU-Popo, a lightweight and universal framework for POst-Processing OCR outputs, which converts page-level results from diverse parsers into coherent document-level structures. MinerU-Popo decomposes the problem into four focused subtasks: text truncation recovery, table truncation recovery, title hierarchy reconstruction, and image-text association. To address these effectively, we build a task-oriented data engine with task-specific input filtering, and use the generated data (30K) to fine-tune a lightweight post-processing model (Qwen3-VL-4B). To support long documents, we introduce dynamic chunking with overlap-based synchronization, which aligns chunk-level outputs from the fine-tuned model and preserves global consistency. Finally, we assemble the aligned outputs into a tree-structured document representation, further enriched with node chunking and summaries for downstream retrieval and analysis. Empirical results show MinerU-Popo improves title-hierarchy TEDS by at least 20% across all five tested OCR models, improves RAG accuracy on most subsets, and reduces per-query latency by up to 70% compared with raw OCR baselines.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
VLM, Post OCR Processing, Cross-page Relation, Document Layouts and Semantics

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ryN6OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper provides four new tasks in OCR, including text truncation, table truncation, title hierarchy and image-text association. To evaluate the performance of their pipeline, they also construct the PostDocBench including 30K instances and propose their model, MinerU-Popo.

Critical strengths

1: This paper revisited the key failure mode of current OCR paradigm and proposed a new pipeline and model called MinerU-Popo. 2: This paper provided a new benchmark called PostDocBench with 30K training instances which is helpful to the community.

Critical weaknesses

1: The baseline comparision is not fair. MinerU-Popo is finetuned on 30K training datasets, however, the rest of models in table 2 is not finetuned. Also, the author mentioned several baselines in related work including DocHieNet, HRDoc, ZenDB but didn't include them in comparision.

2: The author only provides results regarding title hierarchy task, what is the performance on text truncation, table truncation and image-text association? Current results only provide the absolute value of MinerU-Popo without any other baselines.

Questions for authors

Please see above.

Reviewer_HQ4jOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes MinerU-PoPo, a post-processing framework that reconstructs document-level structures from existing OCR outputs without retraining or replacing OCR models. The authors address structural discontinuities and logical relationships across pages by dividing the problem into four tasks: text continuation, table continuation, heading hierarchy analysis, and image-text association. To efficiently handle these tasks, they fine-tune Qwen3-VL-4B with task-specific input filtering and introduce Dynamic Chunking and Synchronization for processing long documents. Finally, the reconstructed structures are organized into a hierarchical document tree, which the authors claim improves the performance and efficiency of downstream RAG and QA tasks.

Critical strengths
  1. The paper addresses a well-motivated research problem with well-defined practical relevance.

  2. The authors propose a post-processing framework that can be applied to diverse OCR models without relying on a specific architecture.

Critical weaknesses
  1. The abstract is overly long, and I suggest consolidating it into a single paragraph rather than two. For example, the following sentence could be removed, as it adds limited information: “A natural solution is therefore to reuse existing OCR outputs and reconstruct document-level logical structures through post-processing.” In addition, the description beginning with “To address these effectively, we build a task-oriented data engine...” is too detailed for an abstract and should be substantially condensed to focus on the key methodological contributions rather than specific implementation details.

  2. The most significant weakness of this paper is the lack of discussion of two closely related recent studies, M3DocDep (CVPR 2026) [1] and HiKEY (ACL 2026) [2]. MinerU-Popo appears to share substantial similarities with these works in both problem formulation and methodology, including (1) leveraging page-level OCR outputs, (2) reconstructing document-level hierarchical structures, (3) inferring cross-page structural relationships, (4) linking images, tables, and captions, (5) building tree-based document representations, (6) structure-aware chunking, and (7) evaluating downstream RAG/QA performance. Given that the main contributions of this paper lie primarily in leveraging existing models, decomposing the problem into specific tasks, input filtering, cross-chunk synchronization, and pipeline design rather than introducing a new model architecture, the omission of these closely related works raises concerns about the novelty and distinct contributions of the proposed components. A more thorough comparison with these studies is necessary to clarify the technical differences and justify the paper's contributions.

[1] Shin, J., Park, J., Seo, J., & Lim, H. (2026). M3DocDep: Multi-modal, Multi-page, Multi-document Dependency Chunking with Large Vision-Language Models. arXiv preprint arXiv:2605.18774.

[2] Shin, J., Shim, G., Park, J., Seo, J., & Lim, H. S. (2026, July). HiKEY: Hierarchical Multimodal Retrieval for Open-Domain Document Question Answering. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 17967-17987).

  1. The paper does not evaluate MinerU-PoPo on established benchmarks that would allow direct comparisons with prior work on document hierarchy reconstruction. For example, although the authors cite DocHiNet, they evaluate hierarchy reconstruction performance only on their own constructed dataset. The lack of evaluation on existing benchmarks limits the comparability of the results and makes it difficult to assess whether the proposed approach offers meaningful improvements over prior methods.

  2. The ablation study does not provide sufficient evidence to verify the individual contributions of Dynamic Chunking and Synchronization. Furthermore, it remains unclear whether the improvements in downstream RAG performance actually come from reconstructing the logical document structure or from other components of the proposed framework.

  3. The baseline comparisons in Table 2 raise concerns about the fairness of the experimental evaluation. MinerU-PoPo uses a fine-tuned Qwen3-VL-4B model, while the baseline Qwen3-VL models are evaluated without additional fine-tuning. Therefore, it is unclear whether the reported improvements actually come from the proposed methodology or simply from the additional training. A fair comparison with fine-tuned baselines under the same training and input conditions is necessary to clearly demonstrate the effectiveness of the proposed approach.

Questions for authors

Given the substantial similarities with closely related studies in both problem formulation and pipeline design, could the authors clarify the novelty and distinct technical contributions of MinerU-PoPo? Why are these studies neither cited nor discussed?

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

MinerU-Popo reconstructs document-level structures from page-level OCR output. The fine-tuned Qwen3-VL-4B model processes text continuation, table continuation and boundary-cell merging, title hierarchy, and image/table associations. Task-specific filtering reduces input complexity and synchronizes overlapping chunks before building document trees rich with summaries and child nodes. The authors introduce PostDocBench, report hierarchical improvements in five OCR trunks, and evaluate downstream retrieval and question answering on ViDoRe V3 and MMDA.

Critical strengths
  1. A well-motivated and practically relevant problem. The reuse of typed OCR blocks to fix document relationships is a reasonable way to bridge page-level parser and document-level applications. The four tasks are clearly decomposed and mapped to specific structural failures (Sections 3-4).
  2. Substantial structural improvements were reported. Figure 1 shows hierarchical structural improvements for five parsers (including MinerU, ranging from 53.7 to 90.6 TEDS). Table 2 reports 90.6 and 78.0 TEDS for Qwen3-VL-32B, with higher document throughput under the common deployment conditions described. These are all encouraging small-model results.
  3. A useful assessment beyond a single structural metric. The paper reports text and association precision/recall, cell-merge accuracy, external downstream evaluations, and ablations. Table 6 shows the benefits of chunking/synchronizing packages, while Table 7 reveals the trade-offs of summary retrieval. The discussion also acknowledges the failure of detailed industry queries, as well as stronger visualization of RAG financial results.
Critical weaknesses
  1. The independence and reference quality of the main benchmarks need clarification. Appendix A describes Gemini-generated annotations, manual checks, then rapid refinement, further 100 instance checks, achieving more than 95% consistency, and dual model element filtering. It also specifies 30K training and 165 test instances per subtask. However, random partitioning units are not adequately defined and cannot establish complete source documentation, overlapping chunks, versions, and near duplicates. It is also unclear whether the final test targets will accept independent human judgment, or whether teachers-generated targets will be retained in the training annotation pipeline. Because PostDocBench supports major structural claims, document/source-level separation and clearly defined test label validation are crucial. These are unsolved issues, not evidence of leakage or incorrect test labels (Appendix A, pp. 12-13).
  2. The reported scope has not established a common end-to-end recovery across all four tasks. Section 4.1 applies heuristic candidate filtering prior to task prediction, including punctuation-based exclusions and table-boundary/width/column checks. Whether the evaluation denominator contains these rule exclusions is not clear. 92.7% of table results measure the accuracy of unit-level merge methods, rather than separately reporting table-continuation precision/recall or complete-table reconstruction; class balance and simple always-alone references are also not specified. Figure 1 does provide five-parser hierarchy transfers, but the equivalent transfer results for the other three tasks and the reference alignment protocol are not reported. These differences are important because the strong performance on selected candidates is narrower than the scope for restoring all relevant document relationships (Section 5.3; Figure 1; Appendix A).
  3. Downstream improvements cannot be adequately attributed to structural repairs, and latency calculations are incomplete. The proposed searchers embed header paths and generated summaries, while the original OCR embeds the full text. Table 7 helps eliminate summaries and child nodes, but does not provide the original OCR baseline with comparable summaries, chunking, retrieval budgets, and answer-context budgets. Table 5 also changes the representation and whether structural repairs exist, so it cannot isolate serialization quality. Furthermore, each query timing does not specify the processing of caching or postprocessing, summarization, embedding, and index construction. Matching representative controls and independent offline/online costs will make utility and efficiency claims easier to explain (Sections 4.3 and 5.3-5.4; Tables 3, 5, and 7).
Questions for authors
  1. What exactly constitutes the training/test instance? Are the complete PDFs, overlapping chunks, document versions, and near duplicates grouped before splitting? How are the final test labels independently checked, and what are the manually validated annotation units, denominators, and task-per-task agreement rates for the report?
  2. Do structural metrics calculate the relationship of pre-filtered omissions? Please clarify whether cell-merge class distribution and table-continuation assessment, as well as whether the other three tasks are transmitted across the same five OCR parsers under a shared model and a documented reference-alignment procedure.
  3. How does downstream comparison change when the raw OCR receives comparable chunking and summaries and retrieval/context budget matches? Separate one-time preprocessing/indexing costs from query-time retrieval and answering, and clarify the cache policy. Existing measurements or a single centralized control will be particularly helpful

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

Related papers

Discussion (0)

Sign in to comment.