Pith. sign in

REVIEW 4 major objections 7 minor 30 references

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

T0 review · 4 major / 7 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read MinerU.Chem recognizes molecular structures in real literature at 93% SMILES accuracy, ahead of seven comparison systems.

desk verdict A genuinely useful system report whose headline accuracy figure is unproven because the evaluation benchmark and training protocol likely overlap. read the letter →

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

classification cs.CV
keywords opticalchemicalstructurerecognitionOCSRCARBONmolecularreactionschemeparsingdocumentMolRecBench-WildMinerU.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

MinerU.Chem is a document-parsing system that turns organic chemistry papers and patents into machine-readable chemical records: a Molecule Summary List and a Reaction Summary List with source locations, identifiers, and structures. The paper's central claim is that its molecular structure recognition module, built on the CARBON graph representation, reaches 93.02% SMILES exact-match accuracy on the 2,392-image SMILES-evaluable subset of MolRecBench-Wild, and 79.66% graph accuracy on all 5,024 graph-annotated images, outperforming the best comparison system by 18.15 percentage points on SMILES and 43.25 points on graph accuracy. If the numbers hold, chemists and AI-for-chemistry pipelines could extract molecule and reaction data from real literature at near-human reliability, with traceability back to the source image. The system is deployed in the MinerU online platform.

What carries the argument

The load-bearing object is CARBON (Complex Atomic Representation and Bonding Object Notation), an atom-centric graph format for molecular structure depictions. It stores each atom's element or superatom label, coordinates, charge, valence, isotope, and radical, plus bonds with a vocabulary covering single, double, aromatic, wedge, dative, and 'any' types, preserving the visual layout and non-SMILES-compatible semantics such as Markush fragments and repeating units. It carries the argument by making the model's predictions align with the source image, and by serving as the native prediction format whose derived MolFile/SMILES exports are what the paper evaluates; the paper extends CARBON from an evaluation target into a production prediction format.

What would settle it

Re-run the evaluation on the 2,392-sample SMILES subset with the comparison systems configured by their own vendors and with an explicit train/test overlap audit: if the overlap between MolRecBench-Wild images and MinerU.Chem's training components is nonzero, or if baseline accuracy rises above 93.02% under vendor-recommended settings, the claimed margin fails.

Watch

Extended reading notes

Core claim

The paper claims that treating molecular structure recognition as image-to-graph generation rather than SMILES generation is the key to high accuracy on real-world literature. The recognition model GTR-VL-1.4.13 predicts CARBON, an atom-centric graph with image-aligned coordinates, superatoms for abbreviations, and a rich bond-type vocabulary, then exports MolFile and SMILES for downstream tools. Trained on curated open data, human-annotated real literature, and synthetic data covering the 37 MOSAIC difficulty labels, the model scores 93.02% exact SMILES match on the 2,392 SMILES-annotated MolRecBench-Wild images and 79.66% exact graph match on all 5,024 images, with the gap to the best comparison system widening on the hardest difficulty subset C (70.13% vs 60.07% SMILES, 55.42% vs 19.04% graph).

Load-bearing premise

The margin over the comparison systems presumes that the MolRecBench-Wild test images were not used in training GTR-VL-1.4.13 and that the seven baselines were evaluated with fair, representative prompting and decoding conditions; the paper does not state either explicitly.

Editorial extensions

If this is right

  • If the stated accuracy holds, MinerU.Chem can convert molecule and reaction images from papers into SMILES, MolFile, and structured reaction records automatically, reducing manual data curation for chemistry knowledge bases.
  • Because recognition outputs retain CARBON's image-aligned coordinates and internal superatoms, failures can be traced back to the exact atom or bond in the source image, which is a prerequisite for human verification of extracted data.
  • On difficulty-stratified subsets, SMILES accuracy falls from 98.28% (subset A) to 70.13% (subset C), so residual errors concentrate where chemical-semantic complexity such as stereochemistry, Markush structures, and non-standard bonds is present.
  • Reaction scheme parsing currently covers single-step and a subset of multi-step schemes, and the authors state that Markush expansion, electron-pushing arrows, and table-based condition or substrate extraction are not yet supported.
  • The integrated system produces Molecule Summary Lists and Reaction Summary Lists linked to page numbers and bounding boxes, enabling each extracted reaction component to be traced back to its source document evidence.

Reading between the lines

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

  • If the reported margin survives a clean train/test split, it suggests that an image-to-graph objective with a layout-preserving representation is the dominant factor in real-world OCSR, and that SMILES-first captioning in the strongest baselines is a bottleneck; a testable extension would be to ablate CARBON versus SMILES supervision on the same model architecture.
  • The CARBON-as-native-format decision implies that evaluation should eventually include a layout-coordinate metric, since graph accuracy ignores 2D coordinates; the paper's own commentary notes that CARBON preserves image-aligned spatial information which the current graph metric does not measure.
  • A practical consequence not developed in the paper is that the same pipeline could be applied to patents, where reaction schemes and Markush structures are even more prevalent; the authors' stated Markush and table limitations point to this as the next frontier.
  • The synthetic data component designed around the 37 MOSAIC difficulty labels could be reused as a targeted data-augmentation recipe for other chemistry vision models, separate from its role in this system.
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.

Referee Report

4 major / 7 minor

Summary. MinerU.Chem is presented as a chemistry-aware post-processing layer on top of the MinerU document parser, comprising five modules: chemistry relevance filtering, molecular structure detection, molecule identifier extraction, molecular structure recognition, and reaction scheme parsing. The recognition module adopts CARBON — the graph representation introduced by the authors' own MolRecBench-Wild benchmark [28] — as its native output format, while exporting MolFile and SMILES for downstream toolchains. On the latest revised MolRecBench-Wild (5,024 graph-annotated images, of which 2,392 have valid SMILES), the deployed model GTR-VL-1.4.13 is reported to achieve 93.02% SMILES exact-match accuracy and 79.66% graph accuracy, against best comparison values of 74.87% (GPT-5.6-Sol) and 36.41% (Gemini-3.5-flash-thinking), respectively. Difficulty-stratified results are given for subsets A, B, and C defined by the MOSAIC framework. The paper also describes the reaction scheme parsing module and cites prior-work detection results (MolYOLO on MolDet-33k-test).

Significance. If the headline result holds, this would be a substantial practical advance: a deployed, integrated system that converts real-world organic chemistry documents into traceable Molecule and Reaction Summary Lists, with a reported margin of more than 18 percentage points in SMILES accuracy over the best evaluated comparison system on a wild benchmark. The manuscript has genuine internal strengths: the subset arithmetic is consistent (1,219+875+298 = 2,392; the weighted SMILES and graph accuracies reproduce the reported full-set values exactly); the SMILES-evaluable versus graph-annotated distinction is stated clearly; the A/B/C stratification gives a useful picture of performance degradation under visual and chemical-semantic difficulty; and the system is available as an online service. The significance is conditional, however, on two conditions that the paper does not yet establish: (i) that the revised MolRecBench-Wild test images do not overlap with GTR-VL-1.4.13's training mixture, including its human-annotated and MOSAIC-synthetic components, and (ii) that the compared systems were evaluated at parity in prompting, decoding, and output parsing.

major comments (4)
  1. [§5.3, §5.4] The headline claim of 93.02% versus 74.87% presupposes that the 5,024 images of the revised MolRecBench-Wild test set, their human annotations, and any near-duplicate synthetic variants generated from them were not used in training GTR-VL-1.4.13, but the paper never states this. Section 5.3 describes the training mixture as (a) curated open-source data, (b) human-annotated real-world literature 'using the annotation protocol of MolRecBench-Wild [28]', and (c) synthetic data following the two-dimensional MOSAIC difficulty framework that defines the benchmark's A/B/C subsets; reference [28] is self-authored by the same group as this paper. Because test images could plausibly enter components (b) or (c), the reported 18.15-point margin could be a contamination artifact rather than a capability gain. The authors must state explicitly that the revised test images and annotations were excluded from all three training components and should provide supporting evidence, such as document-level disjointness or an image-similarity deduplication pass between training and test data. This is the load-bearing requirement for the paper's central claim.
  2. [§5.4, Table 4] Comparison-parity is not established for the seven LLM/VLM baselines (GPT-5.6-Sol, Gemini-3.5-flash-thinking, Claude-opus-4-8, InternVL3.5, GLM-4.5V, Intern-S1, Seed1.6-Thinking) or the commercial services. The paper reports no system prompts, decoding settings, temperature, number of runs, or output-parsing protocol, yet extracting a valid SMILES string or graph from a free-form VLM response requires a specific parsing and canonicalization procedure. Without this documentation, the comparison may understate baseline performance. Please provide the exact prompts, decoding parameters, the response-to-SMILES conversion procedure, and per-system run counts, ideally with raw prediction files as supplementary material.
  3. [§5.4, §5.2] Graph accuracy is computed over CARBON, which is simultaneously (i) the native prediction representation of MinerU.Chem and (ii) the annotation target of the self-authored benchmark [28]. The paper does not document how baseline systems' outputs were mapped into this graph space, which atom- and bond-level attributes (e.g., dative or 'any' bonds, superatoms, charge/valence/radical fields) were scored as exact matches, or whether some attributes are unrepresentable in the baselines' output formats. A concrete test would be to convert all systems' outputs — including MinerU.Chem's own exported MolFile — into a neutral canonical graph (e.g., an RDKit molecule graph) and report graph accuracy on that common ground, or at minimum to publish the baseline-to-CARBON alignment protocol. Without this, the 43.25-point graph margin may partly measure representation compatibility rather than recognition quality.
  4. [§6] The title, abstract, and contribution list advertise reaction recognition, and reaction scheme parsing is one of the five claimed modules, but Section 6 provides no quantitative evaluation of this module: no benchmark, no accuracy figures, and no error analysis. The only reaction-related numbers in the paper (Table 2) evaluate MolYOLO detection and are explicitly credited to prior work, RxnCaption, rather than to this system. The authors should either include an evaluation of the reaction scheme parsing module or explicitly restrict the paper's accuracy claims to molecular structure recognition in the title and abstract.
minor comments (7)
  1. [Abstract, §1] The arXiv abstract uses 'MinerU-Chem' while the manuscript body uses 'MinerU.Chem'; the manuscript abstract also contains a spacing error ('introducesMinerU.Chem'). Please standardize the system name and fix the spacing.
  2. [Tables 3 and 4] Uni-Parser MolParser 1.5 [7] and MolMole/ViMore [5] are discussed as representative OCSR systems in Table 3 but are absent from the Table 4 evaluation; given the state-of-the-art claim, the absence of at least Uni-Parser should be justified or remedied.
  3. [§5.4] The SMILES exact-match metric is not fully specified: the paper should state whether canonical SMILES were compared, which toolkit and version were used (e.g., RDKit), and how tautomers, salts, and stereochemistry were handled.
  4. [§5.3] The claimed complementary roles of the three training components (open-source, human-annotated, synthetic) are not supported by any ablation study; a component-wise ablation would substantiate the design choices and help isolate the source of the reported gains.
  5. [§5.4] The 'latest revised version' of MolRecBench-Wild is not version-pinned; report the exact release identifier, URL, and revision date so that results can be reproduced against a fixed test set.
  6. [Figure 4] The CARBON JSON examples contain inconsistent field naming (atom/atom1/atom2) and formatting artifacts (spacing inside 'symbols'); normalize them to a single schema so the format is unambiguous.
  7. [§5.4] No error bars or multi-run results are reported. This is secondary for the headline margin (a binomial standard error at N=2,392 and p=0.93 is roughly 0.5 percentage points), but confidence intervals would still be useful for the smaller subset-C strata, where the standard error is about 2.7 percentage points.

Circularity Check

1 steps flagged · score 4.0 of 10

Self-authored evaluation benchmark and training data share the same annotation protocol and MOSAIC difficulty framework, so the reported SOTA margins are not independent of the construction of the model's training set.

  1. self citation load bearing [Section 5.3 (Training Data Composition) and Section 5.4 (Evaluation Method and Results)]
    "The second component comprises molecular structure depictions manually annotated from real-world organic chemistry literature using the annotation protocol of MolRecBench-Wild [28]. ... The third component comprises synthetic data targeting challenging scenarios. ... Following the two-dimensional MOSAIC difficulty framework introduced by MolRecBench-Wild [28], the synthesis process covers 18 categories of visual complexity ... and 19 categories of chemical complexity ... This report evaluates the online model on MolRecBench-Wild [28] following its multi-track evaluation protocol"

    MolRecBench-Wild [28] is authored by the same group, with Haote Yang as first author of both this report and the benchmark. Section 5.3 states that GTR-VL-1.4.13 was trained on human annotations created with the MolRecBench-Wild annotation protocol and on synthetic data generated according to the MOSAIC difficulty framework that defines the benchmark's A/B/C subsets. Section 5.4 then reports the headline 93.02% vs. 74.87% SMILES and 79.66% vs. 36.41% graph margins on the latest revised MolRecBench-Wild. The paper never states that the benchmark images were excluded from this training mixture.

full rationale

The paper is a system report rather than a mathematical derivation, so most derivation-level circularity patterns do not apply. The evaluation numbers in Table 4 are empirical measurements, not consequences of an equation. However, the load-bearing evidence for the SOTA claim is MolRecBench-Wild [28], a benchmark authored by the same group, and Section 5.3 states that the training mixture includes human annotations created with the MolRecBench-Wild annotation protocol and synthetic data generated under the MOSAIC difficulty framework that defines the benchmark's A/B/C subsets. The paper nowhere states that the revised benchmark images are disjoint from this training mixture, so the benchmark is not independent for the system under test. The graph-accuracy metric also uses the CARBON representation that is the system's native output, further favoring the proposed system in that comparison. The SMILES exact-match comparison against external systems such as GPT-5.6-Sol, Gemini, and Claude provides some independent content, so the paper is not wholly circular. Score 4 reflects substantial self-citation in the evaluation chain without an equation-level reduction of the central claim.

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

No fitted constants or hand-chosen numerical parameters are central to the claim; the claims rest on training and evaluation choices, benchmark provenance, and unstated train/test separation. CARBON and MOSAIC are adopted from the team's prior benchmark paper rather than introduced here.

assumptions (4)
  • domain assumption MolRecBench-Wild ground-truth annotations and SMILES/graph conversions are correct.
    All accuracy claims in Table 4 depend on the benchmark labels being accurate; the paper does not audit the benchmark.
  • domain assumption The revised test set is disjoint from all training data, including human-annotated and synthetic components.
    Section 5.3 describes training data but never states that the evaluation images were excluded.
  • domain assumption The comparison systems were evaluated under fair, representative conditions.
    Section 5.4 lists model names only, with no prompts, decoding parameters, or attempt counts.
  • domain assumption SMILES exact-match and graph accuracy are the appropriate primary metrics for comparing OCSR systems.
    The choice of these metrics shapes the ranking; the paper argues for them in Section 5.1.

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 .

Figures

Figures reproduced from arXiv: 2608.03525 by the authors.

Figure 1
Figure 1. The overall workflow of MinerU.Chem. 2 System Overview 2.1 Overall Workflow MinerU.Chem operates as a chemistry-aware post-processing layer within the MinerU pipeline. MinerU first converts PDFs or page images into document-level Markdown and layout outputs; MinerU.Chem then parses molecule- and reaction-related regions to produce structured chemical records [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. An example of the Molecule Summary List (left) and the Reaction Summary List (right). [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. An example of the molecule-identifier pairing. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Comparison of molecular representation methods: (a) representative molecular representa [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Cleaning, normalization, and CARBON representation conversion workflow for open-source [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Example of fine-grained human annotation of molecular structure depictions from real-world [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Examples of synthetic training data generated according to the two-dimensional MOSAIC [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Comparison of SMILES predictions across the evaluated systems. [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Comparison of molecular graph predictions across the evaluated systems. [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: Comparison of molecular graph predictions across the evaluated systems. [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]
Figure 11
Figure 11. Figure 11: Comparison of molecular graph predictions across the evaluated systems. [PITH_FULL_IMAGE:figures/full_fig_p019_11.png]
Figure 12
Figure 12. Figure 12: Comparison of molecular graph predictions across the evaluated systems. [PITH_FULL_IMAGE:figures/full_fig_p019_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 7 linked inside Pith

  1. [28]

    Molrecbench-wild: A real-world benchmark for optical chemical structure recognition

    Haote Yang, Hui Wang, Chen Zhu, Jingchao Wang, Linye Li, Hongbin Lai, Huijie Ao, Yongxuan Lyu, Jiang Wu, Jiaxing Sun, et al. Molrecbench-wild: A real-world benchmark for optical chemical structure recognition. arXiv preprint arXiv:2605.05832, 2026

  2. [1]

    Intern-s1: A scientific multimodal foundation model.arXiv preprint arXiv:2508.15763, 2025

    Lei Bai, Zhongrui Cai, Yuhang Cao, Maosong Cao, Weihan Cao, Chiyu Chen, Haojiong Chen, Kai Chen, Pengcheng Chen, Ying Chen, et al. Intern-s1: A scientific multimodal foundation model.arXiv preprint arXiv:2508.15763, 2025

  3. [2]

    Logics-parsing technical report.arXiv preprint arXiv:2509.19760, 2025

    Xiangyang Chen, Shuzhao Li, Xiuwen Zhu, Yongfan Chen, Fan Yang, Cheng Fang, Lin Qu, Xiaoxiao Xu, Hu Wei, and Minggang Wu. Logics-parsing technical report.arXiv preprint arXiv:2509.19760, 2025

  4. [3]

    Molnextr: a generalized deep learning model for molecular image recognition.Journal of Cheminformatics, 16(1):141, 2024

    Yufan Chen, Ching Ting Leung, Yong Huang, Jianwei Sun, Hao Chen, and Hanyu Gao. Molnextr: a generalized deep learning model for molecular image recognition.Journal of Cheminformatics, 16(1):141, 2024

  5. [4]

    Towards large-scale chemical reaction image parsing via a multimodal large language model.Chemical Science, 16(45): 21464–21474, 2025

    Yufan Chen, Ching Ting Leung, Jianwei Sun, Yong Huang, Linyan Li, Hao Chen, and Hanyu Gao. Towards large-scale chemical reaction image parsing via a multimodal large language model.Chemical Science, 16(45): 21464–21474, 2025

  6. [5]

    Molmole: Molecule mining from scientific literature.arXiv e-prints, pages arXiv–2505, 2025

    Sehyun Chun, Jiye Kim, Ahra Jo, Yeonsik Jo, Seungyul Oh, Seungjun Lee, Kwangrok Ryoo, Jongmin Lee, Seung Hwan Kim, Byung Jun Kang, et al. Molmole: Molecule mining from scientific literature.arXiv e-prints, pages arXiv–2505, 2025

  7. [6]

    Ocsu: Optical chemical structure understanding for molecule-centric scientific discovery.arXiv preprint arXiv:2501.15415, 2025

    Siqi Fan, Yuguang Xie, Bowen Cai, Ailin Xie, Gaochao Liu, Mu Qiao, Jie Xing, and Zaiqing Nie. Ocsu: Optical chemical structure understanding for molecule-centric scientific discovery.arXiv preprint arXiv:2501.15415, 2025

  8. [7]

    Uni-parser technical report.arXiv preprint arXiv:2512.15098, 2025

    Xi Fang, Haoyi Tao, Shuwen Yang, Chaozheng Huang, Suyang Zhong, Haocheng Lu, Han Lyu, Junjie Wang, Xinyu Li, Linfeng Zhang, et al. Uni-parser technical report.arXiv preprint arXiv:2512.15098, 2025

Show all 30 references
  1. [8]

    Molparser: End-to-end visual recognition of molecule structures in the wild

    Xi Fang, Jiankun Wang, Xiaochen Cai, Shangqian Chen, Shuwen Yang, Haoyi Tao, Nan Wang, Lin Yao, Linfeng Zhang, and Guolin Ke. Molparser: End-to-end visual recognition of molecule structures in the wild. InProceedings of the IEEE/CVF International Conference on Computer Vision,...

  2. [9]

    Csmiles: A compact, human-readable smiles extension for conformations.Journal of Chemical Information and Modeling, 65(19):10289–10310, 2025

    James W Furness, Kevin B Moore III, and Art Bochevarov. Csmiles: A compact, human-readable smiles extension for conformations.Journal of Chemical Information and Modeling, 65(19):10289–10310, 2025

  3. [10]

    Glm-4.5 v and glm-4.1 v-thinking: Towards versatile multimodal reasoning with scalable reinforcement learning.arXiv preprint arXiv:2507.01006, 2025

    Wenyi Hong, Wenmeng Yu, Xiaotao Gu, Guo Wang, Guobing Gan, Haomiao Tang, Jiale Cheng, Ji Qi, Junhui Ji, Lihang Pan, et al. Glm-4.5 v and glm-4.1 v-thinking: Towards versatile multimodal reasoning with scalable reinforcement learning.arXiv preprint arXiv:2507.01006, 2025

  4. [11]

    Self- referencing embedded strings (selfies): A 100% robust molecular string representation.Machine Learning: Science and Technology, 1(4):045024, 2020

    Mario Krenn, Florian H ¨ase, AkshatKumar Nigam, Pascal Friederich, and Alan Aspuru-Guzik. Self- referencing embedded strings (selfies): A 100% robust molecular string representation.Machine Learning: Science and Technology, 1(4):045024, 2020

  5. [12]

    Chemvlm: Exploring the power of multimodal large language models in chemistry area

    Junxian Li, Di Zhang, Xunzhi Wang, Zeying Hao, Jingdi Lei, Qian Tan, Cai Zhou, Wei Liu, Yaotian Yang, Xinrui Xiong, et al. Chemvlm: Exploring the power of multimodal large language models in chemistry area. InProceedings of the AAAI Conference on Artificial Intelligence, volum...

  6. [13]

    Molgrapher: graph-based visual recognition of chemical structures

    Lucas Morin, Martin Danelljan, Maria Isabel Agea, Ahmed Nassar, Valery Weber, Ingmar Meijer, Peter Staar, and Fisher Yu. Molgrapher: graph-based visual recognition of chemical structures. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 19552–19...

  7. [14]

    Junbo Niu, Zheng Liu, Zhuangcheng Gu, Bin Wang, Linke Ouyang, Zhiyuan Zhao, Tao Chu, Tianyao He, Fan Wu, Qintong Zhang, et al. Mineru2. 5: A decoupled vision-language model for efficient high-resolution document parsing.arXiv preprint arXiv:2509.22186, 1(2):7, 2025

  8. [15]

    Rxnscribe: a sequence generation model for reaction diagram parsing.Journal of chemical information and modeling, 63(13):4030–4041, 2023

    Yujie Qian, Jiang Guo, Zhengkai Tu, Connor W Coley, and Regina Barzilay. Rxnscribe: a sequence generation model for reaction diagram parsing.Journal of chemical information and modeling, 63(13):4030–4041, 2023

  9. [16]

    Molscribe: robust molecular structure recognition with image-to-graph generation.Journal of chemical information and modeling, 63(7):1925–1934, 2023

    Yujie Qian, Jiang Guo, Zhengkai Tu, Zhening Li, Connor W Coley, and Regina Barzilay. Molscribe: robust molecular structure recognition with image-to-graph generation.Journal of chemical information and modeling, 63(7):1925–1934, 2023

  10. [17]

    Decimer: towards deep learning for chemical image recognition.Journal of Cheminformatics, 12(1):65, 2020

    Kohulan Rajan, Achim Zielesny, and Christoph Steinbeck. Decimer: towards deep learning for chemical image recognition.Journal of Cheminformatics, 12(1):65, 2020. 14 MinerU.Chem: A High-Precision System for Optical Chemical Structure and Reaction Recognition

  11. [18]

    Decimer- segmentation: Automated extraction of chemical structure depictions from scientific literature.Journal of cheminformatics, 13(1):20, 2021

    Kohulan Rajan, Henning Otto Brinkhaus, Maria Sorokina, Achim Zielesny, and Christoph Steinbeck. Decimer- segmentation: Automated extraction of chemical structure depictions from scientific literature.Journal of cheminformatics, 13(1):20, 2021

  12. [19]

    Kohulan Rajan, Henning Otto Brinkhaus, M Isabel Agea, Achim Zielesny, and Christoph Steinbeck. Decimer. ai: an open platform for automated optical chemical structure identification, segmentation and recognition in scientific publications.Nature communications, 14(1):5045, 2023

  13. [20]

    Rxncaption: Reformulating reaction diagram parsing as visual prompt guided captioning.arXiv preprint arXiv:2511.02384, 2025

    Jiahe Song, Chuang Wang, Bowen Jiang, Yinfan Wang, Hao Zheng, Xingjian Wei, Chengjin Liu, Rui Nie, Junyuan Gao, Jiaxing Sun, et al. Rxncaption: Reformulating reaction diagram parsing as visual prompt guided captioning.arXiv preprint arXiv:2511.02384, 2025

  14. [21]

    Molecular identifier visual prompt and verifiable reinforcement learning for chemical reaction diagram parsing.arXiv preprint arXiv:2603.15011, 2026

    Jiahe Song, Chuang Wang, Yinfan Wang, Hao Zheng, Rui Nie, Bowen Jiang, Xingjian Wei, Junyuan Gao, Yubin Wang, Bin Wang, et al. Molecular identifier visual prompt and verifiable reinforcement learning for chemical reaction diagram parsing.arXiv preprint arXiv:2603.15011, 2026

  15. [22]

    Mineru: An open-source solution for precise document content extraction.arXiv preprint arXiv:2409.18839, 2024

    Bin Wang, Chao Xu, Xiaomeng Zhao, Linke Ouyang, Fan Wu, Zhiyuan Zhao, Rui Xu, Kaiwen Liu, Yuan Qu, Fukai Shang, et al. Mineru: An open-source solution for precise document content extraction.arXiv preprint arXiv:2409.18839, 2024

  16. [23]

    Bin Wang, Tianyao He, Linke Ouyang, Fan Wu, Zhiyuan Zhao, Tao Chu, Yuan Qu, Zhenjiang Jin, Weijun Zeng, Ziyang Miao, et al. Mineru2. 5-pro: Pushing the limits of data-centric document parsing at scale.arXiv preprint arXiv:2604.04771, 2026

  17. [24]

    Gtr-cot: Graph traversal as visual chain of thought for molecular structure recognition

    Jingchao Wang, Yifan He, Haote Yang, Jiang Wu, Lingli Ge, Xingjian Wei, Yinfan Wang, Linye Li, Huijie Ao, Chengjin Liu, et al. Gtr-cot: Graph traversal as visual chain of thought for molecular structure recognition. arXiv preprint arXiv:2506.07553, 2025

  18. [25]

    Internvl3

    Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, et al. Internvl3. 5: Advancing open-source multimodal models in versatility, reasoning, and efficiency.arXiv preprint arXiv:2508.18265, 2025

  19. [26]

    Smiles, a chemical language and information system

    David Weininger. Smiles, a chemical language and information system. 1. introduction to methodology and encoding rules.Journal of chemical information and computer sciences, 28(1):31–36, 1988

  20. [27]

    αextractor: a system for automatic extraction of chemical information from biomedical literature.Sci

    Jiacheng Xiong, Xiaohong Liu, Zhaojun Li, Hongzhong Xiao, Guangchao Wang, Zhenjiang Niu, Chaoyuan Fei, Feisheng Zhong, Gang Wang, Wei Zhang, et al. αextractor: a system for automatic extraction of chemical information from biomedical literature.Sci. China: Life Sci., 67:618–621, 2023

  21. [29]

    Chemdfm-x: towards large multimodal model for chemistry.Science China Information Sciences, 67(12):220109, 2024

    Zihan Zhao, Bo Chen, Jingpiao Li, Lu Chen, Liyang Wen, Pengyu Wang, Zichen Zhu, Danyang Zhang, Yansi Li, Zhongyang Dai, et al. Chemdfm-x: towards large multimodal model for chemistry.Science China Information Sciences, 67(12):220109, 2024

  22. [30]

    Yode-segmentation: automated noise-free retrieval of molecular structures from scientific publications.Journal of Cheminformatics, 15(1):111, 2023

    Chong Zhou, Wei Liu, Xiyue Song, Mengling Yang, and Xiaowang Peng. Yode-segmentation: automated noise-free retrieval of molecular structures from scientific publications.Journal of Cheminformatics, 15(1):111, 2023. 15 MinerU.Chem: A High-Precision System for Optical Chemical S...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.