{"id":"0e9f0419-65b8-47c7-9886-2a9dd35a0288","arxiv_id":"2608.12898","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"NaviDC-OCR is a 1.2B parameter document parser that combines deformation-aware training, adaptive layout sampling, and content-structure decoupled learning to reach state-of-the-art scores on OmniDocBench v1.6, Wild-OmniDocBench, and PureDocBench.","lead":"This paper presents NaviDC-OCR, a vision-language model that parses digital and camera-captured documents into structured text, tables, and formulas. It reports top scores on three public document parsing benchmarks by adding geometry-aware training and a two-step table and formula learning strategy.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PureDocBench's headline 78.41 hinges on post hoc removal of six invalid prediction files; the appendix's claim that missing files score zero is unverified and could mask inflated scoring if the evaluator drops unmatched pages.","rationale":"The central claim is the set of SOTA benchmark scores. Of these, the PureDocBench number is the most vulnerable because the appendix itself discloses post hoc removal of prediction files. The fairness of that removal depends on an unverified claim about evaluator behavior: if missing files are dropped from the denominator rather than scored as zero, the 78.41 is inflated. The transfer assumption identified by the reader is important for explaining why the method works, but it is downstream of score validity; even a perfect transfer mechanism cannot rescue an unreliable headline number. I therefore focus on the scoring protocol rather than the synthetic-to-real transfer. The reader's weakest_assumption differs, though the reader's rationale did mention the post hoc exclusion as a supporting concern, hence partial agreement. The verdict remains unchanged as conditional because a concrete rerun with the six files restored can settle the issue; if the score shifts materially, the PureDocBench SOTA claim would need to be withdrawn or substantially qualified.","tokens_in":22674,"tokens_out":6481,"duration_ms":66611,"concrete_test":"Rerun the official PureDocBench evaluator on the NaviDC-OCR prediction directory with the six removed files restored exactly as originally generated, and separately with those files replaced by empty prediction files. Recompute the three track scores and the 78.41 aggregate. Also inspect the evaluator source, or insert a probe, to determine whether an absent prediction file contributes a per-page zero or drops the page from the denominator. If the restored-file run changes the reported overall score by more than 0.1 points, the 78.41 headline is not robust; if the score is unchanged, the post hoc exclusion is benign and the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim rests on three headline scores, and the PureDocBench score of 78.41 is the least secure. Appendix D (Table 5) discloses that six NaviDC-OCR Markdown prediction files were removed from PureDocBench scoring because of 'severe repetitive generation.' The paper asserts this is fair because 'missing predictions are assigned a score of zero under the evaluation protocol.' That assertion is the load-bearing link between the raw model outputs and the reported 78.41. If the official evaluator instead treats a missing prediction file as an unmatched page excluded from the denominator, or if the malformed files would have earned partial credit, removing them improves the aggregate score. The appendix provides no source-level verification of the zero-missing behavior, and the fact that removal was necessary suggests the invalid outputs could not pass through the normal scoring path. Separately, Section 5.2's claim of SOTA on the 'Degraded track' is contradicted by Table 3, where OvisOCR2 scores 77.77 versus NaviDC-OCR's 77.47 on Digital Degraded; if 'Degraded' includes that track, the overclaim compounds the issue. This concern directly targets the reliability of a headline number, before even assessing the synthetic-to-real transfer assumption.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents NaviDC-OCR, a 1.2B-parameter decoupled vision-language model for unified document parsing of digital and camera-captured documents. The method combines three claimed contributions: deformation-aware learning that replaces rectangular layout detection with polygon segmentation and adds point-level geometric supervision; a Curvature-Guided Douglas–Peucker (CGDP) sampling mechanism for adaptive polygon point selection; and a content-structure decoupled learning strategy that explicitly models formula grammars and table structures. Training is organized in four progressive stages, with multi-node consensus voting and a self-judgement VLM for pseudo-label filtering. The authors report state-of-the-art overall scores of 96.87 on OmniDocBench v1.6, 88.53 on Wild-OmniDocBench, 78.41 on PureDocBench, and first place in the ICDAR 2026 Sci-ImageMiner Challenge. The evaluation uses public benchmarks with a fixed OmniDocBench-style protocol.","tokens_in":23102,"tokens_out":5698,"duration_ms":52526,"significance":"If the reported results hold, NaviDC-OCR would be a meaningful advance in practical document parsing, particularly for camera-captured documents where decoupled pipelines traditionally struggle. The paper offers a detailed data-engineering pipeline, a unified prompt interface for eight parsing tasks, and competitive results against much larger models (e.g., 235B general VLMs and 30B Ovis2.6). The authors explicitly disclose the removal of six invalid PureDocBench prediction files in Appendix D, which is a degree of transparency that should be credited. However, the significance is conditional: the central claim of state-of-the-art performance rests on benchmark tables without ablations that isolate the three proposed contributions, without error bars or significance tests, and with an unresolved issue concerning the handling of missing prediction files in the PureDocBench scoring. The synthetic-to-real transfer assumption behind the deformation-aware learning is also not directly validated.","major_comments":[{"comment":"The paper attributes the performance gains to deformation-aware learning, CGDP sampling, and content-structure decoupled learning, but no ablation experiment isolates any of these components. For example, there is no comparison of Stage 2 training with and without the point-level deformation supervision, nor a comparison of polygon boundaries produced by CGDP versus uniform sampling, nor a removal of the content-structure decoupled objective in Stage 3. Without such ablations, the headline improvements over OvisOCR2, PaddleOCR-VL-1.6, and MinerU2.5-Pro cannot be attributed to the proposed mechanisms rather than to the larger or better-curated training data, the multi-model pseudo-labeling, or the four-stage training schedule. I request a dedicated ablation table, at minimum on OmniDocBench v1.6 and Wild-OmniDocBench, that turns off each contribution individually.","section":"§4, Tables 1–3"},{"comment":"The PureDocBench overall score of 78.41 is load-bearing for the abstract's and conclusion's SOTA claims, yet it depends on the post hoc removal of six invalid Markdown prediction files. The paper asserts that 'missing predictions are assigned a score of zero under the evaluation protocol,' but this behavior is not verified against the evaluator source code, and the appendix does not demonstrate what the official OmniDocBench quick_match evaluator does with an absent prediction file. If the evaluator silently drops unmatched pages from the denominator rather than scoring them as zero, or if the malformed files would have received partial credit, the reported 78.41 would be inflated. The authors must either (a) provide source-level verification of the zero-missing behavior, or (b) re-score with the six files replaced by empty predictions and report both the original and corrected numbers. The claim in the current form is not verifiable from the manuscript.","section":"Appendix D, Table 5"},{"comment":"The sentence 'NaviDC-OCR achieves state-of-the-art performance on the Degraded track in Table 3, outperforming existing end-to-end document parsing methods' is contradicted by the 'Digital Degraded' column, where OvisOCR2 scores 77.77 overall versus NaviDC-OCR's 77.47. If the intended claim is limited to the 'Real Degraded' track, the text should say so explicitly; otherwise, the statement overstates the results. Similar caution is needed for the PureDocBench-Clean discussion in §5.1, where the claim of surpassing OvisOCR2 is supported (86.90 vs. 82.14) and should be kept distinct from the degraded-track claim.","section":"§5.2, Table 3"},{"comment":"The central innovation, deformation-aware learning, relies on a transfer assumption that is not tested. Synthetic warps are generated by applying Doc3D/ForCenNet forward mappings to undistorted pages and training the model to predict the warped control points and region boundaries. There is no quantitative analysis of the distribution gap between these synthetic deformations and the real capture conditions in Wild-OmniDocBench, and no ablation that removes the deformation-aware supervision while keeping the additional synthetic data volume fixed. Consequently, the improvement on Wild-OmniDocBench (Table 2) could in principle come from the sheer increase in training data rather than from the geometric awareness the paper claims. I would like to see a direct comparison on Wild-OmniDocBench between (i) the full Stage 2 pipeline, (ii) the same data volume without the control-point and polygon supervision, and (iii) a version trained only on digital documents, to establish that the deformation-specific losses are what drive the gain.","section":"§3.2.1, §4.2, Table 2"}],"minor_comments":[{"comment":"In the OvisOCR2 row, the numeric entries '0.12990.3785.13' appear to be missing separators and are inconsistent with the column structure; this looks like a formatting error that should be corrected.","section":"Table 2"},{"comment":"The HunyuanOCR-1.5 row lists the exact same Formula CDM, Table TEDS, Table TEDS-S, and Read Order Edit values as the PaddleOCR-VL-1.6 row; please verify whether this is a copy-and-paste error or genuinely identical measurements.","section":"Table 1"},{"comment":"The paper states that 'we release the complete implementation and model configurations,' but no repository URL, model checkpoint link, or contact for code release is given anywhere in the text; please add the release location.","section":"§4"},{"comment":"The notation for the number of deformation control points is inconsistent: §3.2.1 introduces 'M×M control points P' while §4.2 uses 'N^2 = 1,024 control points'; unify the notation and clarify whether M and N refer to the same quantity.","section":"§3.2.1 and §4.2"},{"comment":"The abstract refers to 'Wild-OmniDocBench' without a version, while §5.2 and Table 2 specify 'v1.5'; harmonize the version information across the paper.","section":"Abstract and §5.2"},{"comment":"The approximation 'd ≈ κL^2/8' for the Douglas–Peucker distance is stated without derivation or a citation; please provide a reference or a brief derivation, and define κ and L precisely.","section":"§3.2.2"},{"comment":"In the upper-left panel of Figure 2, the label 'NaviO R' appears garbled and should be corrected to a readable model name.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a strong systems-style submission with broad benchmark coverage, but the central SOTA claim is not yet fully supported. The absence of ablations and the unverified zero-scoring behavior in Appendix D are fixable with additional experiments and verification; the claim about the 'Degraded track' is a wording/scope issue. I do not see a fatal flaw in the methodology itself, so I recommend major revision rather than rejection. One additional editorial note: the paper cites many 2026 arXiv preprints, some of which may not be peer-reviewed; this is common in fast-moving areas, but the authors should confirm that all cited works are publicly accessible at the time of publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: the paper is a serious engineering effort with a clear system story, but the headline SOTA claim is not yet solid. The PureDocBench 78.41 depends on a post hoc exclusion of six prediction files, and the claim that missing predictions score zero is unverified; if the evaluator drops unmatched pages instead, the number is inflated. Also the 'Degraded track SOTA' in Section 5.2 conflicts with Table 3, where OvisOCR2's Digital Degraded overall (77.77) beats NaviDC-OCR's 77.47.\n\nWhat's actually new: the combination of deformation-aware learning with point-level control point prediction, CGDP adaptive sampling, and content-structure decoupled learning for formulas and tables. The data engineering pipeline (multi-node consensus voting, self-judgement VLM) is a credible component, and the four-stage training is described in enough detail to be reproduced. The ICDAR Sci-ImageMiner win gives independent evidence that the decoupled structure idea works. The evaluation is more transparent than most: the appendix discloses the removed files instead of hiding them.\n\nSoft spots in proportion: (1) No ablations isolating the three contributions, and no error bars or significance tests; several wins are within a point or two, so variance matters. (2) The missing-prediction scoring assumption is load-bearing and not source-verified. (3) The transfer assumption that synthetic Doc3D/ForCenNet deformations match Wild-OmniDocBench is plausible but not validated by test-set analysis. (4) On OmniDocBench, formula CDM is below OvisOCR2 (96.36 vs 97.53), so the overall win is driven by text and tables, not uniform superiority. (5) No code, data, or weights, despite a promise to 'release the complete implementation.'\n\nBottom line: worth a serious referee, because if the benchmark numbers reproduce, this is a practically useful system. But the referee should demand ablations, a clarification of the PureDocBench exclusion against the official evaluator, and code release before the SOTA claims are accepted. I would not cite it yet; I'd wait for the revised version.","headline":"A credible engineering system with a plausible design story, but the headline SOTA claim is not yet solid: the PureDocBench number rests on an unverified post hoc exclusion, and the 'Degraded track' claim conflicts with the paper's own Table 3.","tokens_in":23494,"tokens_out":1818,"would_cite":false,"duration_ms":18870,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"NaviDC-OCR claims that a single 1.2B vision-language model parses both digital and camera-captured documents at state-of-the-art accuracy, without a separate dewarping module.","keywords":["document parsing","vision-language models","camera-captured documents","geometric distortion","layout segmentation","table structure recognition","formula recognition","scientific figure-to-table"],"falsifier":"Train NaviDC-OCR twice, once with and once without the Stage 2 synthetic-deformation samples, keeping everything else identical, and compare Wild-OmniDocBench scores on creased and folded pages; if the gap is within noise, the deformation transfer claim fails. As a sharper check, hold out real photos with warp types absent from the Doc3D/ForCenNet synthesis, such as page curl from bound books or water-ripple distortion, and see whether accuracy drops.","tokens_in":22517,"feed_emoji":"📄","tokens_out":10653,"duration_ms":95248,"temperature":0.7,"pith_summary":"NaviDC-OCR is a unified document-parsing framework that tries to establish that one model can handle both born-digital and camera-captured documents, including tables, formulas, and scientific figures. The paper's core claim is that a decoupled VLM can beat end-to-end rivals on distorted photos if geometric deformation is learned internally rather than removed by a preprocessing dewarping step. To do this, NaviDC-OCR predicts polygonal region boundaries and a coarse grid of deformation control points, samples boundary points with a curvature-aware simplification rule, and trains tables and formulas by predicting structure before content. On three public benchmarks and one ICDAR competition it reports state-of-the-art overall scores: 96.87 on OmniDocBench v1.6, 88.53 on Wild-OmniDocBench, 78.41 on PureDocBench, and first place in the Sci-ImageMiner figure-to-table task. If these numbers hold, the practical payoff is a single small model that replaces a rectangular-detection-plus-dewarping pipeline with geometry-aware parsing.","feed_headline":"One 1.2B model tops three document-parsing benchmarks","feed_subtitle":"NaviDC-OCR learns page warps instead of rectifying them, reporting scores of 96.87, 88.53, and 78.41.","key_machinery":"The load-bearing machinery is a set of three training-side mechanisms. Deformation-aware learning turns synthetic warps into two supervision signals: region-level polygon boundaries (replacing rectangular boxes) and a downsampled $32\\times 32$ grid of control points that the VLM must predict, with the warps generated through forward mappings from Doc3D and ForCenNet. Curvature-Guided Douglas–Peucker Sampling (CGDP) assigns each candidate boundary point an importance $S_i = d_i(1 + \\lambda\\hat{\\kappa}_i)$, mixing global simplification distance with local curvature so creases and sharp corners keep enough sample points. Content-structure decoupled learning makes the model emit a structural skeleton first—OTSL table topology without cell contents, or LaTeX syntax tokens—and then fill in content, which the paper argues lowers structural uncertainty. Around these sit a data engine (heterogeneous-model consensus voting plus a self-judgement VLM that compares rendered predictions to the original image) and a four-stage training schedule ending in GRPO reinforcement learning with normalized edit distance, TEDS, and CDM rewards.","core_discovery":"The central claim is that document geometry can be absorbed into a vision-language model as an intermediate prediction target, so that camera-captured documents no longer need explicit rectification. The authors report that applying a dewarping model to Wild-OmniDocBench substantially improves a two-stage parser, and that structure-heavy tasks such as tables and formulas show high prediction entropy; NaviDC-OCR is their answer to both. Layout analysis is reformulated as polygon boundary prediction with an adaptive number of points, supplemented by a coarse global control-point field that teaches the VLM the page's warp, and structured tasks are taught in two passes: first the OTSL table skeleton or LaTeX syntax, then the cell contents. The trained 1.2B model reports 96.87 overall on OmniDocBench v1.6, 88.53 on Wild-OmniDocBench v1.5, 78.41 on PureDocBench, and the best TEDS score in the ICDAR 2026 Sci-ImageMiner Challenge. Taken at face value, the discovery is that deformation-aware, structure-decoupled training makes a decoupled VLM competitive across clean, degraded, and physically captured documents.","pith_inferences":["An implication the authors leave implicit is that the deformation-aware recipe (adaptive polygons plus a coarse warp field) could transfer to other distorted-perception tasks such as camera-captured forms, receipts, or whiteboards, where rectangular detectors fail the same way.","A testable extension is to isolate CGDP: swapping the curvature modulation $\\lambda\\hat{\\kappa}_i$ off to plain Douglas–Peucker should specifically hurt crease and corner regions in layout segmentation, which would confirm the sampling rule as the active ingredient.","The paper's own admission of repetitive generation on several PureDocBench pages points to a follow-up: page-level structure conditioning or constrained decoding for long structured documents, rather than deleting invalid outputs.","Because the model already predicts a coarse $32\\times 32$ control-point grid, the same supervision could be extended to recover a backward mapping and produce rectified document images as a byproduct, something NaviDC-OCR does not currently output."],"forward_implications":["If the reported scores replicate, decoupled VLM parsers can drop the dedicated dewarping preprocessing module; geometric robustness comes from training on synthetic warps.","Polygon-based layout prediction with adaptive sampling becomes a viable replacement for axis-aligned boxes on curved, folded, or creased pages, improving crops for downstream content recognition.","The structure-then-content recipe generalizes beyond tables and formulas to any task with a strong grammar, such as code blocks and scientific figure-to-table conversion, the paper suggests.","The data pipeline's consensus voting plus rendered-image self-judgement offers a way to build large document-parsing training sets with fewer human annotations.","Task-specific verifiable rewards in GRPO provide a template for aligning parsing models to text, table, and formula metrics simultaneously."],"supporting_citations":[{"why":"Supplies the forward-mapping deformation generation used to synthesize warped training samples for deformation-aware learning.","marker":"Cai et al. (2025)"},{"why":"Provides the backward mappings from which the forward deformation mapping is derived, defining the synthetic distortion distribution.","marker":"Das et al. (2019)"},{"why":"Defines Wild-OmniDocBench, the camera-captured benchmark used both to motivate the approach and to report the 88.53 overall score.","marker":"Li et al. (2026a)"},{"why":"Defines the OmniDocBench evaluator and quick-match protocol used to score the benchmark comparisons reported in the paper.","marker":"Ouyang et al. (2024)"},{"why":"Defines PureDocBench's clean, degraded, and real tracks against which the 78.41 overall score is reported.","marker":"Li et al. (2026c)"},{"why":"Supplies OTSL, the table structure tokenization used as the structural skeleton in content-structure decoupled learning.","marker":"Lysak et al. (2023)"},{"why":"Contributes the GRPO reinforcement-learning algorithm used in Stage 4 to optimize parsing rewards.","marker":"Shao et al. (2024)"},{"why":"Defines the CDM formula metric used as a benchmark score and as a reward in reinforcement learning.","marker":"Wang et al. (2025a)"},{"why":"Defines the TEDS table similarity metric used to score table outputs and as a reinforcement reward.","marker":"Zhong et al. (2020)"},{"why":"Organizes the ICDAR 2026 Sci-ImageMiner Challenge whose scientific figure-to-table task the paper reports winning.","marker":"Ahmed et al. (2026)"}],"fun_headline_variants":["NaviDC-OCR: learns page warp, skips dewarp, tops 3 benchmarks","One 1.2B model parses flat, warped, and captured docs","NaviDC-OCR: two-pass decoding for tables and formulas","ICDAR 2026 winner: NaviDC-OCR sets document parsing record","NaviDC-OCR: deformation-aware parsing, no rectification needed"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The main assumption is that the computer-generated wrinkles, curves, and folds used in training stand in for the distortions in real camera-captured test photos; if that synthetic-to-real transfer fails, the Wild-OmniDocBench advantage and the deformation-aware design are not established.","fun_headline_variants_meta":{"raw":{"variants":["NaviDC-OCR: learns page warp, skips dewarp, tops 3 benchmarks","One 1.2B model parses flat, warped, and captured docs","NaviDC-OCR: two-pass decoding for tables and formulas","ICDAR 2026 winner: NaviDC-OCR sets document parsing record","NaviDC-OCR: deformation-aware parsing, no rectification needed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000836,"raw_usage":{"total_tokens":3705,"prompt_tokens":1060,"completion_tokens":2645,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":676,"completion_tokens_details":{"reasoning_tokens":2540}},"tokens_in":676,"tokens_out":2645,"duration_ms":18698,"temperature":1.0,"reasoning_tokens":2540,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:59:42.762341+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train NaviDC-OCR twice, once with and once without the Stage 2 synthetic-deformation samples, keeping everything else identical, and compare Wild-OmniDocBench scores on creased and folded pages; if the gap is within noise, the deformation transfer claim fails. As a sharper check, hold out real photos with warp types absent from the Doc3D/ForCenNet synthesis, such as page curl from bound books or water-ripple distortion, and see whether accuracy drops.","supporting_citations":[{"cited_title":"Forcennet: Foreground-centric network for document image rectification","cited_arxiv_id":null,"evidence_quote":"Supplies the forward-mapping deformation generation used to synthesize warped training samples for deformation-aware learning."},{"cited_title":"Dewarpnet: Single-image document unwarping with stacked 3d and 2d regression networks","cited_arxiv_id":null,"evidence_quote":"Provides the backward mappings from which the forward deformation mapping is derived, defining the synthetic distortion distribution."},{"cited_title":"Optimized table tokenization for table structure recognition","cited_arxiv_id":null,"evidence_quote":"Supplies OTSL, the table structure tokenization used as the structural skeleton in content-structure decoupled learning."},{"cited_title":"Image-based table recognition: data, model, and evaluation","cited_arxiv_id":null,"evidence_quote":"Defines the TEDS table similarity metric used to score table outputs and as a reinforcement reward."}],"review_version":1}