{"id":"2549407e-70da-447a-92f9-1714cf9780f5","arxiv_id":"2607.28532","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"OCSRGlyph reaches 93.8% stereo-exact USPTO OCSR and MarkushGlyph beats multi-stage Markush parsers, with a stricter graph-equality metric that catches prior scorer failures.","lead":"Two models turn chemical drawings into machine-readable text: OCSRGlyph for single molecules and MarkushGlyph for patent Markush families. They beat prior systems on standard benchmarks by stereo-aware training and a single-stage vision-language read of the whole image.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Markush SOTA rests on a thin, incompletely scored margin and an unreproduced baseline.","rationale":"The reader already isolated the right load-bearing premise: benchmark labels + CXSMILES/VF2 scoring as sufficient ground truth, plus the failed MG2 USPTO-M reproduction and ignored substituent text. Stress-testing does not surface a deeper internal inconsistency in the method (single-stage VLM, stereo oversampling, strict scorer validation in App. C, and released code/weights are real support). It does sharpen the same concern quantitatively: under strict equality the flagship IP5-M gain is ~1pp, majority-vote helps but does not remove dependence on the legacy score for a clean sweep narrative, and M2S structure-only grading omits half of the stated record. That justifies keeping CONDITIONAL rather than upgrading to ACCEPT or downgrading to REJECT—the OCSR result and the stricter metric are genuine contributions; the Markush SOTA packaging needs tighter baseline reproduction, uncertainty on small deltas, and explicit limits on what the metrics grade. No change to the reader’s verdict.","tokens_in":17757,"tokens_out":685,"duration_ms":38585,"concrete_test":"Using the released predictions/weights: (1) bootstrap 95% CIs on MarkushGlyph−MG2 strict deltas on the fixed public IP5-M 878 and M2S 103 splits; (2) re-infer MarkushGrapher-2 from its public checkpoint on USPTO-M 74 under both scorers and record whether 55% is recoverable; (3) add exact/normalized <stable> match on M2S. If the IP5-M strict CI includes 0, or USPTO-M MG2 remains ~42% while legacy tables still cite 55%, the cross-benchmark Markush SOTA claim as written does not hold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The bundled SOTA claim is only as strong as the Markush half. Under the authors’ preferred strict parsed-graph equality (Table 3), MarkushGlyph’s edge on the largest set IP5-M is 52.1% vs MarkushGrapher-2’s 51.1% (≈9 examples on n=878). That comparison already uses the authors’ own MG2 run, and on USPTO-Markush they could not reproduce the published MG2 score (41.9% vs 55%; Table 3 footnote)—so the “exceeds prior systems on all three” headline leans on the looser legacy cxsmi_equality numbers and an unstable baseline. Independently, §4.4 and App. A state that neither scorer checks substituent-definition text in <stable>, even though that text is part of the emitted record and of the M2S task. If the IP5-M strict delta is noise, the USPTO-M baseline is mis-measured, or “correct Markush translation” is defined without the definition prose the model is trained to emit, the Markush SOTA margin is not what Tables 2–3 imply. OCSRGlyph’s 93.8% claim is on firmer ground (matched stereo conventions, data ablations); the load-bearing soft spot is the Markush evaluation premise.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper frames OCSR and Markush structure recognition as image-to-text problems and presents two models: OCSRGlyph, a compact Swin encoder–Transformer decoder trained with stereochemistry-enriched data, and MarkushGlyph, a LoRA-tuned Qwen3.5-2B VLM that reads the full patent crop (structure plus printed text) in one early-fusion stage and emits a unified cxsmiles_opt record. On the 5,719-image USPTO OCSR benchmark, OCSRGlyph reports 93.8% canonical (full-stereo) exact match, above prior published numbers under matched stereo conventions. MarkushGlyph is reported to exceed MarkushGrapher-2 and earlier systems on IP5-M, M2S, and USPTO-Markush under both the legacy cxsmi_equality score and a new strict parsed-graph equality metric based on feature-constrained VF2 isomorphism. The authors also ablate stereo oversampling and the ordinary-molecule mixture share, release code/weights/manifests, and document scorer validation in the appendix.","tokens_in":18124,"tokens_out":1792,"duration_ms":52755,"significance":"If the results hold under tightened evaluation, the work is a clear incremental advance for chemical document AI: a simple stereo-curation recipe that moves full-stereo OCSR past the previous best reported USPTO numbers, and a single-stage VLM that matches or beats multi-stage Markush pipelines without a separate OCR/layout front-end. The strict parsed-graph equality metric is a genuine methodological contribution—it exposes a real failure mode of the prior recall-style scorer (extra/transposed annotations on symmetric scaffolds) and is backed by VF2/RDKit differential checks, fuzz, and determinism tests. Public weights, training/eval code, and prediction dumps materially raise the bar for reproducibility in this subfield. The remaining gap is that Markush accuracy still lags OCSR by a large margin and that the Markush SOTA margin under the stricter metric is thin on the largest set.","major_comments":[{"comment":"Table 3, strict parsed-graph equality, IP5-M column: MarkushGlyph greedy is 52.1% vs MarkushGrapher-2 at 51.1% (≈9/878 examples). That delta is within ordinary binomial noise and is the main support for “exceeds prior systems … under strict equality” on the largest public set. Please report Wilson/bootstrap CIs (or a paired McNemar test on the same 878 IDs) for all three benchmarks under both scorers, and qualify the abstract/§5.1 SOTA wording where the strict IP5-M edge is not significant. Majority-vote (54.7%) may still clear the bar; the single-greedy claim currently does not on IP5-M alone.","section":"Table 3, §5.1"},{"comment":"Table 3 footnote and §5.1: the authors’ MarkushGrapher-2 reproduction on USPTO-Markush reaches only 41.9% under both metrics versus the published 55.0% MarkushGrapher-2 score. The paper still headlines “exceeds … on all three” while mixing published MG2 numbers (cxsmi_equality) with a self-run strict baseline that could not match the published score on one of three sets. This is load-bearing for the comparative claim. Please (i) diagnose the gap (checkpoint, preprocessing, example filter, CXSMILES canonicalization, or denominator), (ii) report a single consistent MG2 baseline under both scorers from one public inference recipe, and (iii) mark any unreproduced cell as such in the main table rather than only in a footnote.","section":"Table 3, §5.1"},{"comment":"§4.4 and Appendix A.3 state that neither Markush metric scores the <stable> substituent-definition text, even though MarkushGlyph is trained to emit it and M2S includes that text in the image and target. For M2S in particular, “correct Markush translation” as scored can ignore errors or omissions in the definition prose the model is asked to produce. Either add a simple exact/normalized match (or field-level F1) on <stable> for M2S and report it beside Table 3, or explicitly scope all Markush accuracy claims to “structure+annotations only, substituent text unscored” in the abstract, contributions list, and §5. Leaving the omission only in the methods appendix overstates what the tables measure.","section":"§4.4, Appendix A.3, Table 3"},{"comment":"Table 2 vs Table 3 / Figure 5: MarkushGlyph’s USPTO OCSR canonical exact match is 69.0% (5% ordinary-molecule mix) against OCSRGlyph’s 93.8%, and the conclusion correctly notes the unification gap. The multi-task claim in §1 and §6 (“without a priori knowledge of what type of chemical structure is being parsed”) is only weakly supported at current OCSR accuracy. Please either (i) show a practical router/ensemble result (MarkushGlyph for Markush pages, OCSRGlyph for molecule-only crops) with end-to-end PDF-style numbers, or (ii) soften the unified-parser framing to match the measured OCSR gap and the 0–15% mixture ablation.","section":"Table 2, Figure 5, §1, §6"}],"minor_comments":[{"comment":"Figure 3 omits strict parsed-graph bars; pointing readers only to Table 3 makes the looser metric visually dominant. Consider a paired bar plot or dual-axis figure for both scorers.","section":"Figure 3"},{"comment":"§3.1 / Appendix A.2: cxsmiles_opt is described clearly, but a short side-by-side of one structure in standard CXSMILES vs cxsmiles_opt in the main text (not only the Rv example) would help non-ChemAxon readers.","section":"§3.1"},{"comment":"Table 2 baseline cells mix “own reported” and MolSight-re-reported values (mark m). A one-column note listing canonicalizer version (RDKit) and stereo flags per baseline would reduce residual protocol doubt.","section":"Table 2, Appendix A.1"},{"comment":"Appendix C Table 4 is strong validation; citing the 541 vs 25 disagreement asymmetry once in §5.3 main text (not only the figure) would better motivate adopting the strict metric as primary.","section":"§5.3, Appendix C"},{"comment":"Minor wording: abstract “state-of-the-art OCSR model” is fine under canonical match; ensure the same phrase is not applied unqualified to Markush under strict IP5-M until CIs are shown. Also fix “the a strict” typo in Appendix A.3.","section":"Abstract, Appendix A.3"},{"comment":"References include 2026 system cards and arXiv preprints (MarkushGrapher-2, MolSight, etc.); keep versions/dates pinned in the camera-ready for longevity.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"Solid applied CV/cheminformatics paper with real artifacts (weights, strict grader, manifests). I would not reject: OCSRGlyph’s stereo result and the strict Markush metric are publishable even if Markush SOTA rhetoric is toned down. The skeptic note is right that the Markush half is thinner than the abstract implies; requiring CIs, a clean MG2 re-run, and an explicit substituent-text policy is enough. Fit is appropriate for a methods-focused CV or cheminformatics venue; less so for a pure theory venue. No integrity red flags beyond the unreproduced USPTO-M baseline, which looks like an evaluation-hygiene issue rather than selective reporting—the authors already footnoted the 41.9% figure."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is twofold. First, OCSRGlyph’s 93.8% full-stereo exact match on USPTO is on firm ground: same public split, three stereo conventions, and a simple curriculum ablation (stereo/adjacent-ring oversampling) that moves the residual from stereo-only errors without tanking graph accuracy. Second, MarkushGlyph shows a single early-fusion VLM plus LoRA can beat multi-stage Markush pipelines, and they ship a stricter parsed-graph equality metric that actually catches the failure modes of the prior cxsmi_equality scorer (extra labels, transposed R-groups on symmetric scaffolds, hallucinated Sg brackets). Code, weights, and manifests are released. That is real work.\n\nWhat is new is mostly engineering and evaluation discipline, not a new paradigm. Image-to-SMILES, Swin decoders, VLM+LoRA, and cxsmiles_opt are established. The distinctive pieces are the stereo data recipe, the unified OCSR-as-empty-Markush training mix (Fig. 5 is honest about the tradeoff), and the validated strict scorer (App. C: VF2/RDKit differentials, fuzz, determinism).\n\nSoft spots, in proportion. The Markush headline is softer than OCSR. Under strict equality on IP5-M the edge is 52.1 vs 51.1—about nine examples on n=878. On USPTO-Markush they could not reproduce MarkushGrapher-2’s published 55% (they get 41.9%) and they footnote it; the “beats all three” claim leans harder on the looser legacy score and majority vote. Neither scorer checks <stable> substituent prose, which they state plainly—so “correct Markush” still means scaffold+annotations, not the full claim text the model emits. MarkushGlyph also remains well below OCSRGlyph on plain molecules, which matters if you want one bulk PDF parser. None of that is hidden or circular; it just means the Markush SOTA margin is thinner than Tables 2–3 first suggest.\n\nWho it is for: people building patent indexing, OCSR training sets, or Markush eval. Citation pattern is appropriate; math is standard isomorphism/canonicalization, not load-bearing theory. I would send it to peer review. Engage if you care about stereo OCSR or Markush metrics; skim the Markush SOTA claim and read the scorer appendix and the USPTO-M footnote before you lean on the leaderboard numbers.","headline":"Solid infra paper: stereo-curated OCSR is the clean win; Markush SOTA is real but thinner than the abstract sells, and they mostly own the gaps.","tokens_in":18779,"tokens_out":621,"would_cite":true,"duration_ms":21804,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A single image-to-text model can read both ordinary molecules and Markush patent drawings, with stereo-aware training lifting full-stereo OCSR to 93.8% exact match.","keywords":["optical chemical structure recognition","Markush structures","CXSMILES","vision-language model","stereochemistry","image-to-sequence","patent informatics","parsed-graph equality"],"falsifier":"Re-annotate a fixed slice of IP5-M/M2S/USPTO-Markush with independent chemist-verified CXSMILES (including stereo and full substituent text), then re-score MarkushGlyph and the prior best system under strict parsed-graph equality; if the ranking or the reported margins reverse, the central Markush claim fails.","tokens_in":18600,"feed_emoji":"🧪","tokens_out":1012,"duration_ms":27208,"temperature":0.7,"pith_summary":"Chemistry is still mostly published as pictures. Turning those pictures into line notations is how databases get indexed and how training sets for chemistry models get built. This paper treats both ordinary molecule drawings and Markush structures (drawings that describe a whole family of related compounds) as one image-to-text problem. For single molecules it shows that carefully oversampling chiral examples closes most of the remaining stereochemistry errors and reaches 93.8% full-stereo exact match on the standard USPTO set. For Markush drawings it shows that one early-fusion vision-language model reading the whole crop—scaffold, R-labels, brackets, and printed text together—beats multi-stage pipelines that split vision and text. It also replaces a looser prior score with a strict atom-mapping test that rejects extra or swapped variable groups the older metric missed.","feed_headline":"One model reads molecules and Markush drawings from images","feed_subtitle":"Stereo-aware training hits 93.8% full-stereo OCSR; a single VLM beats multi-stage Markush parsers","key_machinery":"cxsmiles_opt plus strict parsed-graph equality: an optimized CXSMILES record with inline R-labels, positional-variation and repeat-unit annotations, used as the shared target for both tasks; equality is decided by finding one atom bijection (VF2-style) that preserves the molecular graph, R-group assignments, attachment sets, repeat units, and fragments—not by per-feature recall that can credit extras or label swaps.","core_discovery":"OCSR and Markush translation can be solved as a single image-to-sequence task. Stereo-enriched training data makes a compact encoder-decoder state of the art on full-stereo OCSR (93.8% canonical exact match on 5,719 USPTO images). An early-fusion vision-language model that ingests the entire Markush image in one pass, without a separate OCR or layout stage, exceeds prior multi-stage systems on IP5-M, M2S, and USPTO-Markush under both the legacy score and a new strict parsed-graph equality metric; mixing a small share of ordinary molecules into training lets the same model handle both input types.","pith_inferences":["If substituent-definition text were folded into the strict metric, measured Markush accuracy would likely drop and expose a second error mode the current tables do not show.","The remaining OCSR gap inside the unified Markush model suggests a natural next experiment: scale the ordinary-molecule mixture or add stereo-rich Markush synthetics until one checkpoint matches specialist OCSR.","A strict scorer that also demands stereo agreement would tighten the Markush leaderboard the same way canonical exact match tightened OCSR comparisons."],"forward_implications":["Patent and literature bulk parsers can run one model without first deciding whether a figure is a single molecule or a Markush family.","Stereo-aware curriculum alone is enough to push full-stereo OCSR past prior reported numbers on the shared USPTO benchmark.","Strict atom-mapping equality becomes the fairer public yardstick, because it fails predictions that add or transpose variable groups the older recall-style score accepted.","Majority vote over a handful of samples further lifts Markush accuracy without changing the architecture.","Unifying both tasks in one record format makes ordinary molecules the empty-variable special case of Markush output."],"fun_headline_variants":["Single model reads molecules and Markush drawings as images","Stereo-aware OCSR hits 93.8% full-stereo match on USPTO images","One-pass VLM tops multi-stage parsers on Markush structures","Image-to-sequence solves both OCSR and Markush translation","Early-fusion VLM handles full Markush images without OCR stages"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The published Markush benchmark labels and the graph-level scorers are treated as ground truth for a correct translation, even though neither scorer checks the free-text substituent definitions and the older score ignores stereochemistry.","fun_headline_variants_meta":{"raw":{"variants":["Single model reads molecules and Markush drawings as images","Stereo-aware OCSR hits 93.8% full-stereo match on USPTO images","One-pass VLM tops multi-stage parsers on Markush structures","Image-to-sequence solves both OCSR and Markush translation","Early-fusion VLM handles full Markush images without OCR stages"]},"model":"grok-4.5","effort":"low","cost_usd":0.004889,"raw_usage":{"total_tokens":1424,"prompt_tokens":805,"num_sources_used":0,"completion_tokens":77,"cost_in_usd_ticks":48888000,"prompt_tokens_details":{"text_tokens":805,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":542,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":805,"tokens_out":77,"duration_ms":9291,"temperature":1.0,"reasoning_tokens":542,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T08:38:33.524832+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-annotate a fixed slice of IP5-M/M2S/USPTO-Markush with independent chemist-verified CXSMILES (including stereo and full substituent text), then re-score MarkushGlyph and the prior best system under strict parsed-graph equality; if the ranking or the reported margins reverse, the central Markush claim fails.","supporting_citations":[],"review_version":2}