acceptodds
Under review as a conference paper at ICLR 2027

ViCoR: Reliable Molecular Structure Extraction via Spatially Aligned Verification and Executable Revision

Abstract

Reliable optical chemical structure recognition (OCSR) is essential for building high-quality chemical data from scientific literature, yet even small recognition errors can propagate into chemical databases and downstream models. In practice, recognized structures often require manual inspection and correction before use, making large-scale data curation costly and difficult to scale. We therefore study Selective Structure Recognition (SSR), a post-recognition setting that automatically produces reliable structured outputs while rejecting unresolved cases. Selection-only approaches can improve reliability by rejection, but cannot create additional correct outputs beyond those produced by the base recognizer. We propose ViCoR, a repair-before-rejection framework for iterative VerIfiCatiOn and Revision. Its key idea is to make observation–prediction correspondence explicit: coordinate-preserving rendering establishes spatial correspondence between the source image and predicted structure, while index anchoring maps localized visual discrepancies to executable graph edits without full-structure regeneration. A shared VLM is progressively trained from verification to revision. On two real-world OCSR benchmarks, ViCoR improves overall accuracy from 73.53% to 88.26% and from 61.83% to 84.32%, while achieving over 97% accepted accuracy at 85–89% coverage. The resulting molecular data further improve reaction-extraction F1 by 15.5 points and literature-sourced reaction prediction accuracy by 7.7 and 5.8 points, demonstrating the value of automated reliability control for scientific data curation and downstream chemical learning. Code will be released.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
reliable OCSR for downstream tasks, selective structure prediction, Vision-Language Models

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_CiepOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper focuses on chemical structure recognition and extraction from images. Instead of directly predicting the final chemical structure, it proposes ViCoR, a pipeline of iterative verification and revision. This is motivated by the observation that initial predictions often contain only minor errors compared with the ground truth.

The pipeline works as follows. The initial molecular graph prediction Y^0 is first rendered into an image. The coordinates also appear to be specified during rendering to align it with the input image. A verifier, implemented as a VLM, is then asked whether the two images represent the same structure. If they match, the prediction is returned. Otherwise, a reviser proposes local corrections using six types of node- and edge-level edits based on the input and rendered images (relates to Question 1 later). Although this is not explained clearly, my understanding from the prompt on page 15 is that the reviser produces editing instructions in a language or JSON format. An executor must therefore apply these instructions (related to Question 2 later). The process repeats until verification succeeds or the maximum number of revisions, T, is reached.

The verifier and reviser share a single VLM, which appears to be trained by the authors, although the training details are largely unclear in the main text (relates to Question 3 later). Finally, the authors conduct experiments and compare their approach with several existing methods.

Critical strengths
  1. The topic of the chemical structure extraction is well motivated and has been studied for a long time.
  2. Compared with the text content of the paper, the figures are more informative and help me a lot to understand the paper. However, some figures lack references from the main text, as noted later.
Critical weaknesses
  1. To be honest, I feel this manuscript is poorly prepared, particularly the method section. Several easy-found problems of mathematical writing, cross-referencing, and grammar are listed below. Setting aside the academic contributions of the paper, I believe that, in terms of the writing and structure of the manuscript itself, it is not yet ready for submission, let alone be accepted.
  2. Continuing the previous point, some important methodological details remain vague in the main content. After reading the method section, I found it difficult to follow the method and evaluate its soundness. I may have missed information in the appendix, but the main text does not provide sufficient context or clear references to appendix how these issues are addressed. I do not mean that every detail should be included in the main text. However, the authors should at least indicate that these questions have been considered and direct readers to the relevant appendix sections for further explanation. Meanwhile, repeated and redundant statements waste space that could be used to explain more important methodological details.
  3. I also have concerns that the method is technically unsound. Several important concepts or details are not clearly described, leaving many unresolved questions. Please see the questions below.
  4. Although code and checkpoints are not strictly required for submission, but I feel these are important, especially when it is hard to directly evaluate the results by only reading the paper. I believe it should be straightforward to provide code and checkpoints through an anonymous repository, and I hope the authors will do so. I would reconsider my assessment once the results are verified by myself; otherwise, I remain unconvinced by the reported results.
  5. Commercial tools such as Mathpix support optical chemical structure recognition and extraction as well. I believe comparisons with these tools are also useful information for providing readers more context of the positions of the methods with the commercial software.

Some issues of the academic writing:

  1. Some of the mathematical writing is unclear, and the notation sometimes appears inconsistent or incorrect. The authors appear to formulate nearly every component mathematically, but these formulations do not seem to have been carefully considered and checked for clarity and consistency, making the method unnecessarily difficult to follow.

    (a). In Eq. (10), is there a specific reason to introduce P? Why not directly write V = (I || I_tilde)? Similarly, why is P_idx needed in Eq. (11)? These additional symbols seem to complicate the presentation without a clear benefit.

    (b). In Eq. (12), how many arguments does the reviser take, and what are they? The notation appears inconsistent even within the equation. Does the reviser receive only the prediction? My understanding is that it should receive both the prediction and the original input image.

    (c). In Eq. (15), what are p_ver and p_rev? They are not explicitly defined. The reviser’s arguments also appear different from those used in the earlier equations. Please clarify these definitions and make the notation consistent.

  2. Figures should be referenced in the main text. Figures 1 and 2 are not referenced in the main text, and Figure 2 is referenced only in the appendix. Explicitly referencing and discussing figures is a basic requirement of academic writing.

  3. Displayed equations that are not referenced in the text are commonly left unnumbered. Although this is not a strict rule, it is a standard writing convention.


Typos and grammatical issues:

  1. Line 169: "an base" should be "a base."
  2. Line 169: "recognizer first the predicts" should be "recognizer first predicts."
  3. The sentence around Eq. (5) is incomplete.
Questions for authors
  1. Some edits may require the coordinates of the location to be changed, such as when adding an atom. How does the VLM obtain these coordinates? Does it predict them directly from the image, or does it use additional tools? The prompt on page 15 does not appear to provide explicit coordinate information to the VLM.

  2. How are the JSON-formatted editing instructions applied as actual modifications to the molecular graph? I assume that a rule-based executor performs these operations, but its operation is not clear from the main text.

  3. The authors use only lines 291–300 to explain VLM training in the method section. How exactly is the model trained, and how are the training data constructed? After reading the method section, I still could not determine how training is performed. The experiments section briefly discusses synthetic data construction, but the explanation remains limited.

For example, how are the reviser's target edit sequences obtained? Are the corrections to erroneous predictions annotated by humans, or generated automatically (I think the synthetic data construction is targeting for this, right)? I think this is one of the most important aspects of the paper. If the appendix explains these details, the main text should explicitly indicate that these questions are addressed and direct readers to the relevant sections. The current organization and presentation make the training procedure difficult and struggling to understand.

Reviewer_HudMOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper introduces ViCoR, a method for correcting mistakes made by an OCSR system (a system that converts the diagram of a molecule to a structured form like a SMILES string). ViCoR renders the output and asks a VLM to compare it with the original image. The VLM is used to determine correctness and make localized edits to the graph to correct for any mistakes. The VLM is a small model the authors trained on synthetic data. Evaluation results validate the method, and show that it improves downstream task performance.

Critical strengths
  1. The evaluation is thorough*, including 10 benchmarks and ~10 methods, and further evaluation of two downstream tasks and ablation studies further justify the method.
  2. The evaluation results show that the method does meaningfully provide performance and coverage boost over existing methods.

*With one exception – see Weakness 1 below.

Critical weaknesses
  1. (major) The evaluations on the benchmarks made by the authors had more baselines (Table 1) than the established benchmarks (Table 2). What are the numbers for the missing (method, benchmark) pairs? For example, GPT-5.6-sol on the established benchmarks or MolVec on the new benchmarks? Update: This is largely addressed, though minor discrepancies (e.g. GPT-4o) remain.
  2. (minor) The conceptual contribution is somewhat incremental, given that the idea of making corrections through a given action interface is well-known in the space of LLM agents. The contribution appears more like a combination of largely existing ideas. Update: I maintain that the methodological contribution itself is incremental, though I now appreciate the gains despite the simplicity of the approach.

Update: I raised my soundness score from 2 to 3 and overall score from 2 (weak reject) to 3 (weak accept).

Questions for authors
  1. Does ViCoR assume that the OCSR system outputs the graph structure with all atom coordinates, so that there’s no additional special alignment step needed in the graph generation step? If so, it would be helpful to explicitly state it when describing the method (e.g., in the introduction and Section 3.1).

Suggestions:

  1. The downstream task methods are described only very briefly in Section 4.2. For example, it is unclear how ViCoR was integrated into OpenChemIE. A more elaborate description in the main text would be helpful.
3 comments

I thank the authors for their responses. W1 (inconsistent of baselines in Tables 1 vs 2) is largely addressed. I recommend including GPT-4o in Table 2 as well, and more generally make the rows of Table 1 identical to the rows of Table 2 whenever possible.

Regarding W2 (the contribution is incremental), the main challenge in coordinate-preserving rendering is obtaining the coordinates and edges, after which one simply renders the atoms at the given coordinates and draws in the edges. This hard part is already provided by MolScribe. Similarly, the interface for index-anchored rendering is standard graph edit operations. Nevertheless, we appreciate the ablation studies showing that they do matter, so the observation that such a simple change is all that's needed for significantly improving verification could be considered interesting.

Overall, I raise my soundness score to 3 and overall score from 2 (weak rejection) to 3 (weak acceptance).

Thank you very much for your thoughtful and encouraging follow-up. We sincerely appreciate the time you have taken to reconsider our work and your recognition of the value of our approach and supporting ablation studies.

Following your additional suggestion, we have included GPT-4o in Table 2 and made the baseline coverage across both tables as consistent as possible.

Your constructive feedback has helped us strengthen the experimental evaluation and clarify our contributions. We are truly grateful for your careful consideration and continued support.

I see the revised Table 2 in the paper -- thanks for addressing my feedback.

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

The paper introduces ViCoR, a selective structure recognition framework to reduce errors in OCSR. The key idea is “repair-before-rejection.” First, ViCoR renders the predicted molecular graph while preserving its spatial layout. A VLM compares this rendering with the original image: if it detects a discrepancy, it uses an indexed rendering to propose localized graph edits. ViCoR applies the edits, renders the revised graph, and verifies it again. A verified graph is accepted, and one that remains unverified after reaching the revision budget is rejected. The authors progressively train a VLM for verification and revision using synthetic data. Experiments across standard OCSR benchmarks and two real-world datasets show improvements in recognition accuracy, as well as downstream utility for chemical reaction extraction and literature-based reaction prediction.

Critical strengths
  • The iterative “repair-before-rejection” approach is intuitive and well motivated. It can correct localized errors while retaining the rest of a mostly correct structure, and the experiments validates the approach.

  • The evaluation is comprehensive, covering 10 OCSR benchmarks, multiple model configurations, and downstream chemistry tasks (e.g., reaction extraction)

  • The paper is well organized and explains the method clearly.

Critical weaknesses
  • While the framework handles various structural corner cases, the paper would benefit from a deeper qualitative analysis of where index anchoring breaks down (e.g. in dense, highly fused ring systems or complex stereochemical depictions) and how these failures led to rejection.

  • Since ViCoR relies on an iterative verify-render-revise loop, a detailed runtime profiling or latency breakdown across varying molecule complexities would be helpful for users to gauge the computational tradeoff.

Questions for authors
  1. Is there a fine-grained analysis showing how the error rate and the required number of revision iterations scale with molecule size (e.g., atom and bond counts)?
Reviewer_DyHZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ViCoR, a post-recognition framework for reliable Optical Chemical Structure Recognition (OCSR). Instead of simply rejecting uncertain molecular structures, ViCoR follows a "repair-before-rejection" paradigm, attempting to correct recognition errors before rejecting unresolved predictions.

The proposed framework consists of three main components. First, coordinate-preserving rendering is used to establish spatial correspondence between the original molecular image and the predicted molecular graph. Second, atom indexing provides explicit anchors for identifying erroneous graph entities, enabling a vision-language model (VLM) to generate executable node- and edge-level editing operations. Third, an iterative verification-and-revision mechanism repeatedly checks the corrected structure and performs additional edits when necessary.

The authors further introduce a progressive training strategy, first training a shared VLM on verification tasks and subsequently fine-tuning it for graph revision. The model is trained using 1.3 million synthetic samples.

The proposed method is evaluated on eight existing OCSR benchmarks and two newly curated real-world datasets, OCIE and JIE. With MolScribe as the base recognizer, ViCoR improves overall accuracy from 73.53% to 88.26% on OCIE and from 61.83% to 84.32% on JIE, while maintaining accepted accuracy above 97% and coverage above 85%. The authors also demonstrate improvements in chemical reaction extraction and literature-based reaction prediction, suggesting that more reliable molecular structures can benefit downstream chemical learning tasks.

Critical strengths

The repair-before-rejection paradigm addresses an important limitation of conventional confidence-based filtering. Instead of discarding potentially recoverable predictions, ViCoR attempts to correct localized structural errors while preserving already-correct parts of the molecular graph. The substantial improvements in overall accuracy, accepted accuracy, and coverage on OCIE and JIE demonstrate the potential value of this approach.

Critical weaknesses

Insufficient discussion of closely related molecular structure correction methods.

The related work section does not sufficiently discuss prior work on molecular structure correction. In particular, AtomLenz (CVPR 2024) has already explored graph edit-correction for molecular structure recognition, while MolNexTR (2024) incorporates an abbreviation expansion and self-correction mechanism. Although these methods differ from ViCoR in their objectives and correction settings, they are closely related to the proposed approach. The current manuscript only mentions MolNexTR as an OCSR baseline, without discussing its self-correction mechanism, and does not discuss AtomLenz. The authors should clarify the technical differences between these approaches and ViCoR, particularly regarding inference-time image-conditioned correction, spatial correspondence, and executable graph editing. This is important for establishing the novelty of the proposed framework.

The chemical validity and reliability of executable graph revisions require stronger validation.

The appendix describes sequential graph updates, cascading edge deletion, obsolete-operation handling, and index-collision resolution. These mechanisms help maintain graph-topological consistency. However, topological validity does not necessarily imply chemical validity.

For example, the paper does not sufficiently clarify whether the revision executor explicitly enforces chemical constraints such as atomic valence, formal charges, aromaticity, and stereochemical consistency. This is particularly important because the method relies on VLM-generated graph-edit instructions, which may contain incorrect or chemically implausible operations.

Furthermore, since verification determines whether a prediction is accepted, false acceptance errors are particularly consequential. A more detailed analysis of chemically invalid revisions and incorrect verification decisions would strengthen the reliability claims.

Questions for authors

Relationship to previous work: How does ViCoR fundamentally differ from the graph edit-correction mechanism in AtomLenz and the self-correction mechanism in MolNexTR? Could the authors clarify which aspects of the proposed framework represent fundamentally new contributions beyond these previous approaches?

Chemical validity: Does the graph update operator explicitly enforce chemical constraints, such as atomic valence, formal charges, aromaticity, and stereochemistry? If so, how are these constraints implemented? If not, how do the authors ensure that the executed graph edits produce chemically valid molecular structures?

Contribution of spatial alignment: Can the authors provide a more controlled comparison using the same VLM backbone, training data, and inference budget, but replacing coordinate-preserving rendering with conventional molecular rendering or direct graph/SMILES revision? Such a comparison would help isolate the contribution of spatial alignment from the effects of task-specific training and structured graph-edit outputs.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.