MinerU.Chem: A High-Precision System for Optical Chemical Structure and Reaction Recognition

Bin Wang; Bowen Jiang; Bowen Zhou; Chengjin Liu; Chen Zhu; Chuang Wang; Conghui He; Dahua Lin; Dechen Lin; Guangyu Wang

MinerU.Chem: A High-Precision System for Optical Chemical Structure and Reaction Recognition

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A chemistry-aware document parser converts molecular drawings to exact SMILES strings 93.02% of the time on a 2,392-image benchmark, beating the best general model by 18 points, while preserving layout and chemistry the strings lose.

desk verdict MinerU.Chem is a credible deployed system with strong internal numbers, but the same-team benchmark and training protocol leave the headline accuracy gains unanchored until artifacts are released. read the letter →

arxiv 2608.03525 v2 pith:OBKHLTSF submitted 2026-08-04 cs.CV

Haote Yang , Jiang Wu , Jingchao Wang , Xingjian Wei , Lixin Ma , Linye Li , Chen Zhu , Xiaolong Wu

show 37 more authors

Yuheng Lu Ziran Zhu Junyuan Gao Lingli Ge Yuan Xu Huijie Ao QianQian Wu Dechen Lin Huaiyu Gu Lu Chen Shengxin Lu ShaSha Wang Yuanyuan Cao Zhejia Yu Ruijie Zhang Zimai Tian Jiaxing Sun Yinfan Wang Jiahe Song Chuang Wang Yubin Wang Rui Nie Hao Zheng Bowen Jiang Hongbin Lai Yifan He Chengjin Liu Tingting Zhang Liqun Wei Lijun Wu Bin Wang Yuqiang Li Guangyu Wang Wei Li Bowen Zhou Dahua Lin Conghui He

This is my paper

classification cs.CV

keywords opticalchemicalstructurerecognitionOCSRCARBONrepresentationmoleculargraphreactionschemeparsingdocumentSMILESMinerU.Chem

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that a document parsing system specialized for organic chemistry literature can turn molecular structure drawings and reaction schemes in papers and patents into structured, traceable, machine-readable records, at an accuracy that general-purpose parsers and large multimodal models do not reach. The load-bearing capability is molecular structure recognition: instead of generating a SMILES string from an image, the model predicts an atom-centred graph (CARBON) that keeps each atom's coordinates and models abbreviations, non-standard bonds, and repeating units as first-class objects, then derives MolFile and SMILES for downstream tools. On the revised MolRecBench-Wild benchmark, the module reports 93.02% SMILES exact-match accuracy on 2,392 SMILES-evaluable images and 79.66% graph accuracy on all 5,024 images, leading the best evaluated comparison systems by 18.15 and 43.25 percentage points. If these numbers hold, chemists and data builders gain a working automatic step for converting literature images into verifiable structured data for reaction prediction, retrosynthesis, property prediction, and drug design, with each record traceable to its source page and bounding box.

What carries the argument

The central object is CARBON (Complex Atomic Representation and Bonding Object Notation), an atom-centric graph representation for molecular structure depictions: each atom—or superatom, for abbreviated groups such as CF3 or Rα—carries its two-dimensional coordinates from the source image, and bonds carry rich types including dative, wavy, wedge, and 'any' bonds. Its job is to serve as the native prediction format of the recognition module, so recognition preserves the visual layout and the chemical semantics that SMILES cannot express; a post-processing step expands superatoms and derives MolFile and SMILES. A second machinery element is the MOSAIC difficulty framework — two dimensions of 1

What would settle it

Run the deployed model on the original, unrevised MolRecBench-Wild images and on a freshly annotated set of molecular structure images from papers the team has not touched; if SMILES exact-match and graph accuracy fall well below the reported 93.02% and 79.66% — or the gap to GPT-5.6-Sol and Gemini-3.5-flash-thinking shrinks substantially — the claimed margins do not generalize. A cheaper audit: have an independent group re-annotate a random sample of the 5,024 images and measure agreement with the released ground truth.

Watch

Extended reading notes

Core claim

The central claim is that faithful structure reconstruction — recovering the graph, atom attributes, and two-dimensional layout of a molecular depiction — is achievable, and that a representation built for that purpose (CARBON) is why the recognition module beats string-first systems. The module is image-to-graph: it outputs atoms, bonds, superatoms, and image-aligned coordinates natively, then derives MolFile and SMILES without discarding the internal graph. On the latest revised MolRecBench-Wild, it reports 93.02% SMILES exact-match on 2,392 SMILES-evaluable images and 79.66% graph accuracy on all 5,024 images, versus 74.87% and 36.41% for best comparison systems. The same graph core, embe

Load-bearing premise

The reported lead over other systems depends on the latest revised version of MolRecBench-Wild being an independent, unbiased testbed; the benchmark, the CARBON representation, the annotation protocol, and the model itself come from the same team, and the paper neither releases the revised benchmark nor documents the train/test separation, so the size of the advantage is not externally auditable.

Editorial extensions

If this is right

If the reported accuracies hold, literature-scale data production for AI-for-chemistry becomes practical: molecule and reaction records come with source-page and bounding-box traceability, not just bare strings.
The representation-first result implies that string-first recognizers lose information that a coordinate-preserving graph keeps; downstream formats can be derived from the graph rather than predicted directly.
The difficulty-stratified results (98.28%/92.15% on the easy subset A, 70.13%/55.42% on the hard subset C) locate the remaining bottleneck in chemical-semantic complexity — Markush fragments, non-standard bonds, stereochemistry — rather than in basic visual recognition.
Reaction scheme parsing produces reactants/products/conditions records linked to molecule records, an enabling step toward reconstructing multi-step synthetic routes from natural-product and medicinal-chemistry papers.
The paper itself limits reaction scheme parsing to common single-step and a subset of multi-step schemes, and leaves Markush/R-group expansion and mechanism diagrams to future work, so the Molecule and Reaction Summary Lists cover a defined scope rather than all of chemistry.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A reader should treat the 18-and-43-point margins as upper bounds until an independent evaluation: the benchmark, the CARBON representation, the annotation protocol, and the model come from the same team, and the paper does not release the 'latest revised version' or document how its revision was informed by the model's errors.
The design suggests a testable extension beyond chemistry: predicting structured, coordinate-preserving graphs instead of flat strings may be a generally stronger recipe for multimodal models on any visual task where spatial layout carries meaning, such as diagrams, schematics, and maps.
The subset-C results point to the next concrete capability to build: automated expansion of Markush structures and R-group substitution tables, explicitly left as future work in the paper, which is the main blocker for property-dataset construction from medicinal-chemistry series.
If the graph-recognition advantage replicates on independently annotated data, downstream tasks sensitive to stereochemistry and layout, such as retrosynthesis and synthesis-route mapping, should measurably improve when trained on these coordinate-preserving records rather than on SMILES alone.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

Read MinerU.Chem. It's a systems paper, and judged on that, it's solid. The real contribution is integration: five chemistry-specific modules bolted onto MinerU, with CARBON as the native prediction format and MolFile/SMILES for export. That's genuinely useful for chemistry knowledge-base construction, and the online deployment means people can actually run it. The paper doesn't oversell novelty—it credits MolYOLO, Mid-Mapper, GTR-VL, and the CARBON representation to prior work. The evaluation is careful for what it is: stratified by MOSAIC difficulty, with visualizations and a separate reaction-parsing module, and the limitations (no Markush expansion, no mechanism interpretation) are stated plainly.

The soft spot is exactly where the stress-test puts it: benchmark independence. Training data includes human annotations following the MolRecBench-Wild protocol and synthetic data generated from the same MOSAIC framework used to label that benchmark. The evaluation runs on the 'latest revised version' of MolRecBench-Wild. All of that is same-team. The paper doesn't release the revised benchmark, the training set, or an overlap analysis, so we can't rule out that the revision or annotation protocol was shaped by the model's errors. If that happened, the 18-point SMILES gap and the 43-point graph gap would shrink. I'm not saying it did—but the paper doesn't show it didn't.

Two smaller points. The SMILES metric is computed on 2,392 of 5,024 images, and the paper explains the exclusion (Markush, non-standard bonds). That's fair, but the headline number applies to a subset. And the graph accuracy is defined on CARBON, which only MinerU.Chem emits; the comparison systems get dashed out. So the 79.66% vs. 36.41% is not apples-to-apples. That's disclosed, but readers should keep it in mind.

On balance, the system is real and the engineering is sound. The paper deserves a serious referee and likely publication as a systems report, but the review should make the authors release the benchmark, training data, and overlap analysis. Without that, the accuracy claims stay unverified externally.

Recommendation: send to peer review, condition acceptance on artifact release and an analysis of train/test overlap. I'd bring this to the reading group once the artifacts land.

Referee Report

3 major / 5 minor

Summary. The paper presents MinerU.Chem, a chemistry-aware document-parsing layer integrated into the MinerU platform. It consists of five modules: chemistry relevance filtering, molecular structure detection, molecule identifier extraction, molecular structure recognition, and reaction scheme parsing. The central contribution is the molecular structure recognition module, which predicts molecules in a CARBON graph format and exports SMILES/MolFile. The module is evaluated on MolRecBench-Wild: the paper reports 93.02% SMILES exact-match accuracy on the 2,392-image SMILES-evaluable subset and 79.66% graph accuracy on all 5,024 images, outperforming comparison systems by large margins. The paper also describes the system architecture, training data composition, and an online deployment.

Significance. If the benchmark evaluation is trustworthy, the reported margins over general-purpose multimodal models (e.g., 18.15 points over GPT-5.6-Sol on SMILES, 43.25 points over Gemini-3.5-flash-thinking on graph accuracy) are substantial and would make MinerU.Chem a practically useful tool for converting chemistry literature into structured data. The system design is pragmatic and the integration into a public platform is a strength. However, the evaluation rests on a benchmark, a representation, and a difficulty taxonomy all originating from the same research group as the model, with no released split, overlap analysis, or independent audit. In addition, the graph metric is defined on the home-team's CARBON format, which disadvantages systems that do not emit CARBON. The paper is also transparent about some limitations (e.g., SMILES-evaluable subset, limited Markush handling), but the central numerical claims are not yet externally anchored. The work is a useful systems contribution, but its main evaluation claims require substantial additional evidence.

major comments (3)

[§5.3 and §5.4] The training data include human annotations 'using the annotation protocol of MolRecBench-Wild [28]' and synthetic data generated according to the MOSAIC difficulty framework introduced in the same reference. Evaluation is then performed on the 'latest revised version of MolRecBench-Wild [28]'. Reference [28] shares authors with the current paper. The manuscript provides no test/training overlap analysis, no released split, and no independent audit. If the benchmark revision or the shared annotation protocol was influenced by the model's failure modes, the reported gaps over GPT-5.6-Sol (74.87% SMILES) and Gemini-3.5-flash-thinking (36.41% graph) could shrink substantially. This is load-bearing because the benchmark is the only external validation of the system. Please provide an overlap analysis, a frozen and versioned release of the benchmark with a documented train/test split, or an e
[§5.4, Table 4] The graph-accuracy comparison is not format-agnostic. Graph accuracy is defined on CARBON graph annotations, and specialized OCSR systems that do not emit CARBON are simply dashed out. The 43.25-point graph gap over Gemini-3.5-flash-thinking therefore conflates recognition quality with the ability to emit the home-team's representation. To make the comparison fair, convert all systems' outputs into a common standard graph format (e.g., MolFile) and recompute graph agreement, or at least report the graph metric on a comparable subset and discuss the effect of the format mismatch.
[§5.4, Table 4 and Abstract] The headline SMILES exact-match accuracy of 93.02% is computed on only 2,392 of 5,024 images. The paper states that the excluded 2,632 images cannot be converted to valid SMILES because they contain Markush structures, unconventional abbreviations, illegal direction keys, or valence-violating bonds—precisely the hard cases that motivated the wild benchmark. The abstract highlights 93.02% as the system's headline number, while the full-benchmark graph accuracy is 79.66%. The paper should report a full-set metric (e.g., treating unparseable outputs as incorrect) so that the headline cannot be read as performance on the full benchmark, and should justify the representativeness of the SMILES-evaluable subset.

minor comments (5)

[Figure 4] In the CARBON example, the JSON keys for bonds are inconsistent: some entries use "atom" and "atom2" while others use "atom1" and "atom2". This should be standardized to avoid ambiguity.
[Table 4] No confidence intervals or statistical significance tests are reported. A McNemar test on the headline SMILES comparison between MinerU.Chem and GPT-5.6-Sol would strengthen the claim that the difference is meaningful.
[§5.3] The relative sizes of the three training-data components (open-source, human-annotated, synthetic) are not given. Reporting these proportions would help readers assess how much of the performance gain comes from synthetic data and whether the synthetic component closely mirrors the MOSAIC taxonomy used in the benchmark.
[Reproducibility] The paper does not release the model checkpoint, inference code, or evaluation scripts. The online service URL is useful, but the specific model version (GTR-VL-1.4.13) and benchmark version should be paired with a reproducible evaluation protocol, ideally including the exact prompt/parsing code used for comparison systems.
[Support Information, Table 5] The bond-type visualization relies on colors that may be difficult to distinguish in grayscale printing. Consider adding texture or line-style distinctions for color-blind and grayscale readers.

Circularity Check

1 steps flagged · score 4.0 of 10

No formal equation-level circularity; headline accuracy rests on a same-team, unreleased benchmark (MolRecBench-Wild), making the central evaluation load-bearing on self-citation.

self citation load bearing [Section 5.4, Evaluation Method and Results; Section 5.3, Training Data Composition]
"We use the latest revised version of MolRecBench-Wild [28], which contains 5,024 molecular structure images, all with graph annotations."

MolRecBench-Wild [28] is a same-team preprint (overlapping authors with the present paper), and it is the sole benchmark for the paper's headline claims. Section 5.3 states that the training data include human annotations 'using the annotation protocol of MolRecBench-Wild [28]' and synthetic data generated 'following the two-dimensional MOSAIC difficulty framework introduced by MolRecBench-Wild [28]'. The evaluation is performed on an unreleased 'latest revised version' of this benchmark, and the paper reports no train/test overlap analysis and does not release the revised benchmark. The central performance numbers (93.02% SMILES, 79.66% graph) are therefore anchored to an internally controlled, self-cited benchmark rather than an independent external test. This is load-bearing self-citati

full rationale

No step in the paper reduces to its input by construction in the sense of Eq. X = Eq. Y or a fitted parameter being renamed as a prediction. The model genuinely emits SMILES and CARBON graph predictions, and the reported accuracies are measured, not derived from the training labels. The main concern is external validity: the benchmark, the CARBON representation, the MOSAIC difficulty framework, and the annotation protocol all originate from the same research group, and the evaluation uses a revised, unreleased version of that benchmark while the training data are generated using the same protocol. This does not make the results definitionally forced—comparison systems are scored on the same benchmark and the model could fail—but it does mean the headline result is not independently anchored. The graph-accuracy comparison is further weakened because systems that cannot emit CARBON are not scored on that metric. These are correctness/contamination risks rather than strict derivation-level circularity, so the score is 4 rather than higher.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

All central accuracy claims rest on the MolRecBench-Wild benchmark being correct, stable, and independent of the model. The paper does not document test/training separation or provide the revised benchmark, and the framework used to create training data is the same one used to label the test set.

assumptions (4)

domain assumption MolRecBench-Wild annotations (SMILES and graph) are ground truth.
Section 5.4 uses the benchmark scores as the basis for the 93.02% and 79.66% claims; any errors in the annotations would change the numbers.
domain assumption The 'latest revised version' of MolRecBench-Wild is stable and its test split does not overlap with training data.
Section 5.4 and 5.3 imply the model was trained separately from the benchmark; the paper does not document split or version control.
domain assumption CARBON graph representation is a faithful and comparable target for all graph-evaluated systems.
Section 5.2 uses CARBON as native format and graph accuracy equates predicted graphs to benchmark graphs; if CARBON omits or distorts some depiction semantics, graph scores may be biased.
domain assumption All baselines were evaluated under identical, fair prompt and post-processing conditions.
Section 5.4 compares against GPT-5.6-Sol and others without specifying prompt templates or output normalization, so fairness is assumed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of MinerU.Chem: A High-Precision System for Optical Chemical Structure and Reaction Recognition." pith.science (2026). https://pith.science/paper/OBKHLTSF

@misc{pith2026260803525,
  author       = {Pith},
  title        = {Pith review of: MinerU.Chem: A High-Precision System for Optical Chemical Structure and Reaction Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OBKHLTSF}},
  note         = {Machine review of arXiv:2608.03525}
}

read the original abstract

In organic chemistry papers and patents, molecular structures, reaction schemes, and experimental conditions are often presented as molecular structure depictions, reaction diagrams, and complex tables or figures. Such information is difficult for general-purpose document parsing systems to directly convert into machine-readable data. This limits data production for organic chemistry knowledge base construction and for AI for Chemistry tasks such as reaction prediction, retrosynthesis, condition recommendation, molecular property prediction, and drug molecule design. This report introduces MinerU-Chem, a document parsing system for organic chemistry literature integrated into the MinerU online platform. Built on top of MinerU's general document parsing pipeline, MinerU-Chem adds five chemistry-specific modules: chemistry relevance filtering, molecular structure detection, molecule identifier extraction, molecular structure recognition, and reaction scheme parsing. Together, these modules convert organic-chemistry-related image regions in documents into a Molecule Summary List and a Reaction Summary List. For molecular structure recognition, MinerU-Chem uses CARBON (Complex Atomic Representation and Bonding Object Notation) as its core representation. CARBON enables recognition results to preserve both the visual layout of the original image and complex chemical semantics, while supporting the export of standard downstream formats such as MolFile and SMILES. On the SMILES-evaluable subset of MolRecBench-Wild (N=2,392), MinerU-Chem's molecular structure recognition module achieves a SMILES exact-match accuracy of 93.02%, outperforming the best evaluated comparison system, GPT-5.6-Sol (74.87%), by 18.15 percentage points. The system has been integrated into the MinerU online platform and is available at https://mineru.net/OpenSourceTools/Extractor .