Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

This paper argues that single-step retrosynthesis should be judged by chemical plausibility — whether the proposed reaction's reaction-center and functional-group context has precedent in known chemistry — rather than by whether the predict

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 04:54 UTC pith:B2DP64MX

load-bearing objection ChemCensor is a useful idea, but the paper's central claim that it measures chemical plausibility is unvalidated, and the C3LM results are confounded by using the same scorer for filtering, reward, and evaluation. the 3 major comments →

arxiv 2602.03554 v2 pith:B2DP64MX submitted 2026-02-03 cs.LG cs.AIcs.CEcs.CL

When Single Answer Is Not Enough: Rethinking Single-Step Retrosynthesis Benchmarks for LLMs

Bogdan Zagribelnyy , Ivan Ilin , Maksim Kuznetsov , Nikita Bondarev , Mathieu Reymond , Roman Schutski , Thomas MacDougall , Rim Shayakhmetov
show 5 more authors
Zulfat Miftakhutdinov Mikolaj Mizera Vladimir Aladinskiy Alex Aliper Alex Zhavoronkov
This is my paper
classification cs.LG cs.AIcs.CEcs.CL
keywords retrosynthesischemical plausibilityreaction centerfunctional groupsLLM evaluationbenchmark datasetsingle-step retrosynthesissynthesis planning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that the standard way of scoring single-step retrosynthesis models — checking whether the predicted reactant set exactly matches the one recorded answer — is the wrong yardstick. It introduces ChemCensor, a scorer that rates a proposed reaction by whether its reaction center and surrounding functional groups have precedents in a large database of known reactions, giving a 0–5 plausibility score. Under this metric, the paper builds a large training set (CREED) of plausibility-verified reactions and shows that a model fine-tuned on it outperforms both general-purpose and chemistry-specialized LLMs. The broader claim is that evaluation should reward chemically reasonable alternatives, not a single ground truth, and that plausibility-based training data is what makes language models useful for real synthesis planning.

Core claim

The paper's central claim is that chemical plausibility, operationalized as precedent support for a reaction's reaction center and functional-group context, is a better measure of single-step retrosynthesis quality than exact-match Top-K accuracy. It embodies this claim in ChemCensor, which decomposes a proposed reaction into a hierarchical reaction center (five context levels) and a 515-functional-group signature, then looks up both against a library built from a large public patent reaction database; reactions without precedent get zero, and higher scores reflect broader matched context. On an out-of-domain set of 100 novel targets, the paper shows that most leading LLMs degrade sharply on

What carries the argument

ChemCensor: a precedent-based scorer that splits a reaction into a reaction center (RC, the atoms/bonds that change, represented at five increasing context sizes RC1–RC5) and a static functional-group (FG) signature (from a curated list of 515 FG patterns). It matches these against a relational database of reaction centers and FG co-occurrences built from a large public patent reaction set, and assigns a confidence score 0–5 based on how much substructural context is supported. This score replaces exact-match accuracy in the benchmark, is used as a filter to build CREED, and serves as a reward signal in reinforcement learning for C3LM.

Load-bearing premise

A reaction is scored plausible only if its reaction center and functional-group context have a recorded precedent in the patent-based reference database; absence of a precedent is treated as zero plausibility, so reactions that are genuinely valid but undocumented get mislabeled as implausible.

What would settle it

Take a collection of reactions published after the reference database's cutoff (or sourced from non-patent literature) that expert chemists confirm as working, and run ChemCensor on them; if a substantial fraction score zero solely due to missing precedents, the metric's central equivalence between 'undocumented' and 'implausible' is refuted. A cheaper check: use the paper's own example of a chloro-Stille coupling that ChemCensor scores low despite being found in the same patent family as the ground-truth reaction, and see whether expert chemists judge it as a real synthetic route — if they do

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Model rankings change substantially when plausibility replaces exact-match: several frontier LLMs that score well on the standard test collapse on the novel out-of-domain set, suggesting memorization inflates their Top-K numbers.
  • Training on plausibility-verified multi-answer data (CREED) improves performance on novel targets without needing a larger base model; the gains come from data quality and coverage of diverse disconnections.
  • Using ChemCensor as a reward in reinforcement learning reduces functional-group compatibility errors, indicating that the metric encodes chemoselectivity and regioselectivity knowledge transferable to the model.
  • The benchmark exposes that even strong proprietary models fail at functional-group compatibility on specific groups like aryl halides and amines, giving a concrete diagnostic for where synthesis models need improvement.
  • Exact-match ground truth is shown to be misleading: alternative disconnections that chemists would accept receive zero under Top-K but pass plausibility scoring.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because ChemCensor equates 'no precedent' with 'implausible,' it will systematically downgrade genuinely new or recently invented chemistry absent from the patent corpus; as reaction space evolves, the reference library would need continuous updating, and future benchmarks should test whether this creates a false-negative bias.
  • The CREED dataset is generated from a few thousand expert templates plus drug-scheme scaffolds, so its diversity is bounded by template coverage; a model trained on it might underperform on reactions outside that template space even when ChemCensor deems them plausible.
  • A natural testable extension: combine ChemCensor with a forward-reaction validator (round-trip check) to catch false positives where precedent exists but the predicted reactants would not actually form the product; the paper's own experiments suggest round-trip rewards degrade out-of-domain, hinting that these two signals capture different failure modes.
  • The 100-target expert-validated set is small but leakage-resistant; scaling it to a few thousand novel targets would sharpen the ranking between top models and reduce variance in the reported averages.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes ChemCensor, a precedent-based metric that scores single-step retrosynthesis predictions by matching their reaction centers and functional-group contexts against USPTO-full. It introduces CREED (~6.4M reactions, filtered by ChemCensor), URSA-expert-2026 (100 novel expert-validated targets), and C3LM, a Qwen3-8B model fine-tuned on CREED (and, for some variants, with ChemCensor or other rewards). The central claim is that ChemCensor provides a more meaningful plausibility-based evaluation than exact-match Top-K accuracy, and that C3LM trained on CREED outperforms a broad set of proprietary and open-weight LLMs on this benchmark.

Significance. If ChemCensor is a faithful proxy for chemical plausibility, the paper contributes a useful alternative to exact-match benchmarking, a large training dataset, and a new out-of-domain test set. The release of CREED and URSA-expert-2026 would be valuable resources for the community, and the C3LM results would suggest that large curated reaction sets improve SSRS performance beyond current baselines. However, because ChemCensor is used both as the training-filter/reward and as the evaluation metric without independent validation, the significance of these claims is currently uncertain.

major comments (3)
  1. [Sec. 3.1, 3.4, 4.3, 4.4] ChemCensor is used at three points in the pipeline: (i) to filter CREED (Sec. 3.4: 'only those with CC Score >0 were retained'), (ii) as the RFT reward (Sec. 4.3: 'ChemCensor Reward'), and (iii) as the evaluation metric (Sec. 4.4). This creates a closed loop: the training data, the optimization target, and the test score all derive from the same scorer. The reported superiority of C3LM (especially the ChemCensor-reward RFT variant) may reflect optimization of the metric rather than improved chemical plausibility. To support the central claim, the authors must validate ChemCensor against an external reference (e.g., human expert plausibility ratings, laboratory feasibility, or an independent forward-synthesis metric) and provide results on such a reference for the benchmarked models.
  2. [Sec. 3.1] The zero-score rule — 'if the algorithm cannot find supporting precedents for the RC and/or its FGs environment, the score is zero' — operationally defines chemical plausibility as 'has a precedent in USPTO-full (1976–2016 US patents)'. This conflates database coverage with feasibility and will penalize valid modern or niche chemistry outside that corpus. Since CREED is built by keeping only CC>0 reactions and the benchmark scores all predictions with ChemCensor, this assumption is load-bearing. A concrete test: take expert-validated reactions from the URSA-expert-2026 schemes (or other chemistry outside USPTO-full) and show their ChemCensor scores; if many are zero, this confirms the bias. The paper does not report such a test.
  3. [Sec. J, Table 4] The curation of USPTO-50K-test is described inconsistently. The main text (Sec. 3.2) says 'removing a small subset', Appendix J states 'We removed 28 entries', but Table 4 lists 20 + 4 + 5 + 3 = 32 removed entries. This discrepancy directly affects the benchmark surface and reproducibility. Please correct the count and the table, and state the exact version of the curated test set used for all reported results.
minor comments (5)
  1. [Table 1] In the RetroDFM-R row, '4.361.56' appears to be a formatting error; the value should likely be '4.36' or '1.56'.
  2. [Fig. 1 / Appendix F] The caption in Fig. 1 contains 'accriss' (likely 'across'). Similar typo in Appendix F caption.
  3. [Sec. 4.4] The evaluation says 15 independent responses per product; however, Table 1 reports percentages for 'Unique' but the exact definition of uniqueness (after deduplication) is only given in Sec. 3.3. Clarify whether 'Unique' refers to the fraction of targets with at least one valid unique reaction or the fraction of all generated samples.
  4. [Appendix K] Only one expert-validated synthesis scheme is provided as an example. Since URSA-expert-2026 is claimed to be expert-annotated, please release all 100 schemes or provide a public link; otherwise the 'expert assessment' claim is not verifiable.
  5. [Sec. 2.2 / References] Some references are incomplete or informal (e.g., 'Vl´eduts, 1963' spelling, and several ArXiv-only preprints). Ensure all citations have full bibliographic details.

Circularity Check

2 steps flagged

C3LM is filtered, rewarded, and scored by the same ChemCensor function; its reported benchmark superiority is largely self-confirming.

specific steps
  1. fitted input called prediction [Section 4.3 (ChemCensor Reward) vs. Section 3.3 and Table 1]
    "ChemCensor Reward: Applies the confidence level of ChemCensor as a reward. For a generated reaction, the reward is the chemical plausibility between 0−5 , scaled to (0,+1) . … We used ChemCensor Score (CC) as an evaluation metric for a primary reaction."

    During RFT, GRPO maximizes the ChemCensor score of generated reactants; Table 1 reports exactly this score as Av. PT-Max CC and Av. PT-Top-K CC. The paper's finding that 'RFT with the ChemCensor reward increases PT-Max CC' is therefore the definition of reward maximization, not an empirical discovery about chemical plausibility. The same function appears on both sides of the pipeline: as the reward (input) and as the headline evaluation metric (output). This is the fitted-input-called-prediction pattern: the model is optimized on the metric that is then reported as independent evidence of improvement.

  2. self definitional [Section 3.1 (ChemCensor Metric) and Section 3.4 (CREED)]
    "if the algorithm cannot find supporting precedents for the RC and/or its FGs environment, the score is zero. … All candidate reactions were subsequently verified with CC, and only those with CC Score >0 were retained."

    ChemCensor operationalizes 'chemical plausibility' as 'has a USPTO-full precedent for RC and FG context.' CREED is then filtered to retain only CC>0 reactions, i.e., only reactions that satisfy this definition. C3LM is trained on CREED and subsequently evaluated with the same CC function. Thus C3LM's advantage over baselines on the ChemCensor benchmark partly reduces to: 'a model trained on CC-positive reactions scores higher on CC.' No external criterion (laboratory validation, expert ratings, or an independent plausibility metric) is used to break the loop; the expert schemes in URSA-expert-2026 are provided but are not converted into the reported evaluation scores.

full rationale

ChemCensor is simultaneously the proposed scientific measure of chemical plausibility, the filter that selects CREED, the reward used in RFT, and the metric reported in Table 1. The clearest reduction is in Section 4.3: the ChemCensor Reward is the exact score later reported as Av. PT-Max/PT-Top-K CC. Thus the RFT-CC model's improvement over baselines is the GRPO objective itself — a reward maximizer will increase the metric it is trained on, without providing new information about chemical plausibility. Similarly, CREED was constructed by keeping only CC>0 reactions, so a model fine-tuned on CREED has been trained on the evaluator's positive class. The paper presents no external validation linking CC scores to laboratory feasibility or to the expert-validated schemes in URSA-expert-2026; those schemes are not used as quantitative evaluation labels. Because the training data, reward, and evaluation all reduce to the same unvalidated function, the central performance claim is substantially self-confirming. I set 7 rather than 10 because the paper does contain independent components — the CREED resource, the expert target set, and the empirical observation that baselines fail FG-compatibility checks — and the loop is an optimization/evaluation loop rather than a formal logical tautology.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The paper introduces algorithmic artifacts (ChemCensor, CREED, C3LM) rather than physical entities. Its main free parameters are the chosen RC context levels, the 515-FG pattern set, and the positive-score retention threshold for CREED. All of these are hand-set, none are calibrated against independent human assessments, and no sensitivity analysis is reported.

free parameters (3)
  • RC context levels RC1-RC5 = topological distances 1, 2, 3, 4, and an unspecified RC5
    The ChemCensor score is defined by hand-chosen substructural context sizes. No calibration against external plausibility judgments or sensitivity analysis is provided, and RC5 is not explicitly defined.
  • 515 functional-group SMARTS patterns = 515 curated patterns
    The FG-compatibility signature is built from a hand-curated list of 515 patterns. The count and composition are arbitrary design choices that determine which incompatibilities are detected.
  • CREED curation threshold = ChemCensor score > 0
    Reactions in CREED are retained only if ChemCensor gives a positive score. This threshold controls the training-data distribution, and no analysis shows how varying it changes downstream model behavior.
axioms (4)
  • domain assumption USPTO-full (1976-2016 US patents) is a sufficient reference database for chemical plausibility.
    Used in Section 3.1 as the grounding knowledge base of ChemCensor. It assumes absence from USPTO-full is informative, ignoring modern literature, theses, and post-2016 patents.
  • domain assumption Absence of a precedent implies the reaction is chemically implausible.
    Stated in Section 3.1: if no RC precedent is found, the score is zero. This is the core semantic of ChemCensor and is never validated against known counterexamples.
  • domain assumption Reaction centers and functional-group compatibility are fully captured by SMARTS-like pattern matching at topological distances 1-4.
    Used throughout Appendix D to define RC1-RC4. This assumes that all chemically relevant context can be encoded in substructure patterns of limited size, which is a strong simplification of regioselectivity and stereoelectronic effects.
  • domain assumption The ~3K expert-coded templates and drug-derived synthetic scheme templates generate chemically valid reactions.
    Used in Section 3.4 to build CREED. The validity of the enumerated reactions is checked by ChemCensor, but the templates themselves are not independently audited in the paper.

pith-pipeline@v1.3.0-alltime-deepseek · 39506 in / 10933 out tokens · 107188 ms · 2026-08-03T04:54:18.024376+00:00 · methodology

0 comments
read the original abstract

Recent progress has expanded the use of large language models (LLMs) in drug discovery, including synthesis planning. However, objective evaluation of retrosynthesis performance remains limited. Existing benchmarks and metrics typically rely on published synthetic procedures and Top-K accuracy based on single ground-truth, which does not capture the open-ended nature of real-world synthesis planning. We propose a new benchmarking framework for single-step retrosynthesis that evaluates both general-purpose and chemistry-specialized LLMs using ChemCensor, a novel metric for chemical plausibility. By emphasizing plausibility over exact match, this approach better aligns with human synthesis planning practices. We also introduce CREED, a novel dataset comprising millions of ChemCensor-validated reaction records for LLM training, and use it to train a model that improves over the LLM baselines under this benchmark.

Figures

Figures reproduced from arXiv: 2602.03554 by Alex Aliper, Alex Zhavoronkov, Bogdan Zagribelnyy, Ivan Ilin, Maksim Kuznetsov, Mathieu Reymond, Mikolaj Mizera, Nikita Bondarev, Rim Shayakhmetov, Roman Schutski, Thomas MacDougall, Vladimir Aladinskiy, Zulfat Miftakhutdinov.

Figure 1
Figure 1. Figure 1: Assessment of generated reactants from benchmarked models by ChemCensor; each answer is assigned to pass or specific non-pass categories, and models are ordered by the fraction of ChemCensor passed reactions. models are emerging (Shee et al., 2025), many CASP tools still decompose the problem into (i) single-step retrosynthe￾sis (SSRS), which proposes candidate disconnections, and (ii) multistep retrosynth… view at source ↗
Figure 2
Figure 2. Figure 2: ChemCensor part of the URSA benchmark. A. The procedure of retrieval and chemistry-aware knowledge aggregation from USPTO-full into the relational DB of RCs and FGs. B. The pipeline of LLMs benchmarking for the SSRS task. C. The example of RC and FGs intercompatibility assessment during single reaction scoring within ChemCensor. match only a topologically limited substructural context of RC, the score is l… view at source ↗
Figure 3
Figure 3. Figure 3: CREED (Comprehensive Reactant Exhaustive Enumeration Dataset) design process. 3.3. Aggregated Metrics We used ChemCensor Score (CC) as an evaluation metric for a primary reaction. As integral metrics for this benchmarking, we implemented additional metrics: Av. PT-Max CC and Av. PT-Top-K: Av. PT-Max CC is calculated as the average of the Max CC values over all target molecules t in the benchmark set of siz… view at source ↗
Figure 4
Figure 4. Figure 4: LLMs’ performance on URSA-expert-2026 set. language models (Livne et al., 2024; Pei et al., 2023), we ex￾tend the base vocabulary with SMILES format (Weininger, 1988) specific tokens; this is aimed at isolating chemical tokens from natural-language tokens and providing a con￾sistent representation of SMILES entities. During the train￾ing, we tokenize SMILES into specialised tokens always in model outputs, … view at source ↗
Figure 5
Figure 5. Figure 5: FG–related error profiles across models on test sets. The heat map shows the count of CC FG incompatibility errors across the best 10 models by Av. PT-Top-10 CC (see [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: LLMs’ performance in URSA benchmark on URSA-expert-2026 (left) and USPTO-50K-test (right) sets. Higher values along both axes indicate better performance (top-right is best). B. Glossary Target molecule The target molecule is the desired chemical compound that represents the ultimate goal of synthesis planning. It is the molecule for which the system generates or evaluates synthetic routes, and it serves a… view at source ↗
Figure 7
Figure 7. Figure 7: The ground truth for the example #1 from the USPTO-50K-test set. Besides the reaction shown in [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Scoring three different outcomes predicted by RetroKNN for example #1. reference document ID and if one checks other reactions present in the patent with the ground-truth reaction, the chloro-Stille coupling very similar to the ground truth with a fluorine atom replaced with a methoxy group is found ( [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: The additional reaction found in the patent application US20070249579A1, where the ground truth for example #1 was extracted from. To additionally illustrate the difference between our scoring approach and commonly used metrics, another target molecule from the USPTO-50K-test, example #2, was used to predict single-step retrosynthetic output within the LocalRetro model (Chen & Jung, 2021). The target molec… view at source ↗
Figure 10
Figure 10. Figure 10: The ground truth for example #2. N N F NH N N NH N O O S N N F NH N N Cl O S NH2 N + O N N N N NH N O I F NH O S B O O N N F NH N N NH N O O + S N N F NH2 N N NH N O S Cl O N N F NH N N NH N O O + S max LC = 4 Synthetic precedent found in US20110098296A1 max LC = 1 Synthetic precedent found in US20040192767A1 max LC = 5 Synthetic precedent found in US20090149456A1 Ground truth [PITH_FULL_IMAGE:figures/fu… view at source ↗
Figure 11
Figure 11. Figure 11: Scoring three different outcomes predicted by LocalRetro for example #2. The examples provided illustrate the advantages of the present ChemCensor as a tool to assess SSRS models’ output. It is not biased to the ground truth concept and provides the user with a powerful engine for single reaction evaluation. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: ChemCensor operands. A. Input reaction. B. RCNs (highlighted in blue). C. FG signatures (FGs are highlighted). D.1. Reaction center (RC) In this work, we distinguish five levels of reaction center annotation, denoted RC1, RC2, RC3, RC4 and RC5 cor- responding to confidence levels and ChemCensor Scores 1, 2, 3, 4 and 5, respectively. The RCN is the representation of an RC, where N shows the size of the sub… view at source ↗
Figure 13
Figure 13. Figure 13: Example reactions predicted by the C3LM SFT model which passed ChemCensor. If the input reaction is not reflected by the reference dataset of synthetic precedents in terms of its RC and FG context, ChemCensor provides a zero score and the reaction can be considered as chemically implausible. In that case, the module prints the exact reason of failure: (1) there are no precedents of RC representations or (… view at source ↗
Figure 14
Figure 14. Figure 14: Example reactions which didn’t pass ChemCensor due to functional group incompatibility. Reactions for which reaction center representations are absent from the reference dataset are annotated with extracted mapped SMARTS only and a message of absence of synthetic precedents ( [PITH_FULL_IMAGE:figures/full_fig_p022_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: A reaction example for which RC representations are absent from the reference dataset. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Visualized SMARTS patterns for FGs with the highest FG incompatibility error count accriss the Top-10 models by means of Av. PT-Top-10 CC. Individual figure preparation was performed in (Ehrt et al., 2020). G. CREED Details CREED splitting The splitting was performed randomly while preserving the distribution of the number of reactions per product, ensuring that identical products appear in only one split… view at source ↗
Figure 17
Figure 17. Figure 17: CREED dataset distributions of reactions per product: (left) distribution by category; (right) overall distribution. Split Unique reactions Unique products Train 5,107,681 522,865 Validation 635,977 65,285 Test 638,184 65,350 [PITH_FULL_IMAGE:figures/full_fig_p024_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Reaction examples for particular product molecule from USPTO-full (left) and CREED (right). O O F F F F N NH N O O N N N F F F + O O Cl F F F N NH N O O N N N F F F + O O Br F F F N NH N O O N N N F F F + O O I F F F N NH N O O N N N F F F + O O O S O F O F F F F F N NH N O O N N N F F F + O O N I F F F Cl N N O O N N N F F F + O O N N N F F F O O N N N F F F O O N N N O F F F O O N N N F F F OH O OH N N … view at source ↗
Figure 19
Figure 19. Figure 19: Examples of reactions from CREED for the product molecule, which is not included in USPTO-full. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Theoretical synthesis scheme for the compound X404-135-0036 from URSA-expert-2026 set. The URSA-expert-2026 was used as a holdout set. The CREED and USPTO-full were decontaminated within URSA￾expert-2026, so the training datasets for our models don’t contain reactions and product molecules from the holdout at all. L. Benchmark Models Completions Examples Most of the general purpose models we tested suppor… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. URSA: Chemistry-Aware Benchmark for Utilitarian Retrosynthesis Assessment

    cs.LG 2026-07 accept novelty 6.0

    Specialized retrosynthesis models outperform LLMs on chemically plausible multi-step routes when scored by the new URSA Solv-2 protocol using ChemCensor.

Reference graph

Works this paper leans on

26 extracted references · 3 canonical work pages · cited by 1 Pith paper

  1. [1]

    tBuNC, K2CO3, air, Pd(PPh3)4, toluene

  2. [2]

    TFA Org. Biomol. Chem. (2020), 18(26), 4936-4940 10.1039/D0OB01092H K2CO3, MeOH NH2 N O NN N tBuONa, Pd(dpa)2, toluene Cl N HN O Cl 932-22-9 25637-16-5 O Br Mg, THF Cl NN H O O Bioorganic Med. Chem. Lett. (2015), 25(5), 1030-1035 10.1016/j.bmcl.2015.01.024 O Cl Cl NN O O O 727985-37-7 Br HO TBDMSCl, DMAP Br TBDMSO TBAF, MeOH N H N O NN N OH N H N O NN N O...

    work page doi:10.1039/d0ob01092h 2020

  3. [8]

    URL https://chemistry-europe

    doi: https://doi.org/10.1002/chem.201605499. URL https://chemistry-europe. onlinelibrary.wiley.com/doi/abs/10. 1002/chem.201605499. Segler, M. H. S., Preuss, M., and Waller, M. P. Planning chemical syntheses with deep neural networks and sym- bolic AI.Nature, 555(7698):604–610, March 2018. doi: 10.1038/nature25978. URL https://www.nature. com/articles/nat...

    Pith/arXiv arXiv 2018

  4. [9]

    URL https://data.x.ai/ 2025-11-17-grok-4-1-model-card.pdf. Xia, Y ., Jin, P., Xie, S., He, L., Cao, C., Luo, R., Liu, G., Wang, Y ., Liu, Z., Chen, Y .-J., Guo, Z., Bai, Y ., Deng, P., Min, Y ., Lu, Z., Hao, H., Yang, H., Li, J., Liu, C., Zhang, J., Zhu, J., Bi, R., Wu, K., Zhang, W., Gao, K., Pei, Q., Wang, Q., Liu, X., Li, Y ., Zhu, H., Lu, Y ., Ma, M.,...

    Pith/arXiv arXiv 2025

  5. [14]

    The URSA-expert-2026 was used as a holdout set

    AcOH, H2O Figure 20.Theoretical synthesis scheme for the compound X404-135-0036 from URSA-expert-2026 set. The URSA-expert-2026 was used as a holdout set. The CREED and USPTO-full were decontaminated within URSA- expert-2026, so the training datasets for our models don’t contain reactions and product molecules from the holdout at all. L. Benchmark Models ...

    2026

  6. [15]

    **Left Arm (N-aryl-oxadiazole- heterocycle):** ‘Nc2noc(-c3cnn4ccccc34)n2‘ 3

    **Central Phenyl Core:** ‘Cc1ccc(N...)c(CO...)c1‘ 2. **Left Arm (N-aryl-oxadiazole- heterocycle):** ‘Nc2noc(-c3cnn4ccccc34)n2‘ 3. **Right Arm (Ether-linked pyrazolone derivative):** ‘COc2cnn(C(=O)C3CC3)c(=O)c2C2CCOCC2‘ Based on this, we can propose the following disconnections: **Disconnection 1: N-aryl bond formation** The N-aryl bond (‘Ar-NH-Ar’‘) is co...

  7. [16]

    **Hydrazine derivative:** ‘H2N-NH-C(=O)C2CC2‘ (N-cyclopropylcarbonyl hydrazine) 2. **Beta-keto ester:** ‘R-CO-CH2-COOEt‘ where ‘R‘ is ‘C1CCOCC1‘ (ethyl 3-oxo-3-(tetrahydropyran-4-yl)propanoate) <skipped reasoning> **Overall Proposed Reactants:** The most direct disconnection leads to three main fragments that are then coupled. * **Fragment 1 (central core...

  8. [17]

    This would yield ‘<smiles>Cc1ccc(Nc2noc(-c3cnn4ccccc34)n2)c(CO)c1</smiles >‘

    Buchwald-Hartwig coupling of ‘Cc1ccc(Br)c(CO)c1‘ with ‘Nc1noc(-c2cnn3ccccc23)n1‘ to form the N- aryl bond. This would yield ‘<smiles>Cc1ccc(Nc2noc(-c3cnn4ccccc34)n2)c(CO)c1</smiles >‘. 2. Then, a Mitsunobu reaction or a Williamson ether synthesis between the benzylic alcohol (‘-CH2OH‘) from the product of step 1 and the phenolic pyrazolone (‘Oc1cnn(C(=O)C...

  9. [18]

    The CH2O ether linkage could be formed via Williamson ether synthesis 3

    The NH-oxadiazole bond could be formed via nucleophilic aromatic substitution 2. The CH2O ether linkage could be formed via Williamson ether synthesis 3. The cyclopropylcarbonyl on the nitrogen could be an acylation Key fragments I can identify: - The oxadiazole with indazole: -c1noc(-c2cnn3ccccc23)n1 - The aromatic amine with CH2OH: Cc1ccc(N)c(CO)c1 or s...

  10. [19]

    A nucleophilic aromatic substitution between an aniline and a chloro-oxadiazole 2. The aniline already contains the CH2O-pyridazinone substituent **Plausible reactants:** <smiles>Clc1noc(-c2cnn3ccccc23)n1</smiles> <smiles>Cc1ccc(N)c(COc2cnn(C(=O)C3CC3)c(=O)c2C2CCOCC2)c1</smiles> The reaction would proceed via nucleophilic aromatic substitution where the a...

  11. [20]

    There’s a methylphenyl group (toluene derivative) with: - A -Nc2noc(-c3cnn4ccccc34)n2 substituent (an oxadiazole ring with an indazole) - A -COc2cnn(C(=O)C3CC3)c(=O)c2C2CCOCC2 substituent (a pyridazinone with cyclopropyl amide and tetrahydropyran) Looking at this structure, I can see several key components:

  12. [21]

    Jin, W., Coley, C., Barzilay, R., and Jaakkola, T

    URL https://www.sciencedirect.com/ science/article/pii/S2095809922005665. Jin, W., Coley, C., Barzilay, R., and Jaakkola, T. Predicting organic reaction outcomes with weisfeiler-lehman 10 When Single Answer Is Not Enough: Rethinking Single-Step Retrosynthesis Benchmarks for LLMs network. In Guyon, I., Luxburg, U. V ., Bengio, S., Wallach, H., Fergus, R., ...

  13. [22]

    An oxadiazole ring (-noc-n-) connected to: - An aniline nitrogen - An indazole (-c3cnn4ccccc34)

  14. [23]

    The ether bond (C-O-C) between the benzyl and pyridazinone 2

    A pyridazinone system with: - An ether linkage (COc) - A cyclopropyl carbonyl group C(=O)C3CC3 - A tetrahydropyran ring (C2CCOCC2) Possible disconnections: 1. The ether bond (C-O-C) between the benzyl and pyridazinone 2. The C-N bond between the aniline and oxadiazole For the synthesis, plausible reactants could be:

  15. [24]

    Guo, T., Guo, K., Nan, B., Liang, Z., Guo, Z., Chawla, N

    URL https://www.sciencedirect.com/ science/article/pii/S2451929418300858. Guo, T., Guo, K., Nan, B., Liang, Z., Guo, Z., Chawla, N. V ., Wiest, O., and Zhang, X. What can large language models do in chemistry? a comprehensive benchmark on eight tasks. InProceedings of the 37th International Conference on Neural Information Processing Systems, NIPS ’23, Re...

    Pith/arXiv arXiv 2023

  16. [25]

    A halogenated pyridazinone (like a chloropyridazinone): Clc1cnn(C(=O)C2CC2)c(=O)c1C1CCOCC1

  17. [26]

    A nucleophilic aromatic substitution or Williamson ether synthesis could couple these fragments together

    A hydroxymethyl-substituted aniline bearing the oxadiazole group: HOCc1cc(C)ccc1Nc2noc(- c3cnn4ccccc34)n2 The key disconnections appear to be the ether linkage between the benzyl carbon and pyridazinone oxygen, and potentially the C-N bond connecting the aniline to the oxadiazole. A nucleophilic aromatic substitution or Williamson ether synthesis could co...

  18. [27]

    - Multiple electron-withdrawing groups activate the ring and allow substitution

    **SNAr (substituent on aromatic rings)**: - Benzene substituents can undergo coupling reactions, like substitution of an existing leave group with an amine subsequent from another carbonyl or bromide source. - Multiple electron-withdrawing groups activate the ring and allow substitution

  19. [28]

    - Perhaps a Buchwald–Hartwig amination? Possible precursors: The first part is the source molecule, and the second component is the amine

    **Benzimidazole derivatives can be formed via a C-N bond formation:** - The amine and aryl chloride or halide could attach to form a substituting amine. - Perhaps a Buchwald–Hartwig amination? Possible precursors: The first part is the source molecule, and the second component is the amine. Given the complexity, the reaction likely links these parts. **Re...

  20. [70]

    emnlp-main.70

    URL https://aclanthology.org/2023. emnlp-main.70. Rasley, J., Rajbhandari, S., Ruwase, O., and He, Y . Deep- Speed: System optimizations enable training deep learn- ing models with over 100 billion parameters. InPro- ceedings of the 26th ACM SIGKDD International Con- ference on Knowledge Discovery & Data Mining, KDD ’20, pp. 3505–3506, New York, NY , USA,...

    arXiv 2023

  21. [827]

    Zhang, C., Lin, Q., Zhu, B., Yang, H., Lian, X., Deng, H., Zheng, J., and Liao, K

    URL https://spj.science.org/doi/ abs/10.34133/research.0827. Zhang, C., Lin, Q., Zhu, B., Yang, H., Lian, X., Deng, H., Zheng, J., and Liao, K. SynAsk: unleashing the power of large language models in organic synthesis.Chem. Sci., 16:43–56, 2025a. doi: 10.1039/D4SC04757E. URL http://dx.doi.org/10.1039/D4SC04757E. Zhang, D., Liu, W., Tan, Q., Chen, J., Yan...

    Pith/arXiv arXiv 2024

  22. [840]

    2025/840

    URL https://doi.org/10.24963/ijcai. 2025/840. Liu, B., Ramsundar, B., Kawthekar, P., Shi, J., Gomes, J., Luu Nguyen, Q., Ho, S., Sloane, J., Wender, P., and Pande, V . Retrosynthetic reaction prediction us- ing neural sequence-to-sequence models.ACS Cen- tral Science, 3(10):1103–1113, 2017. doi: 10.1021/ acscentsci.7b00303. URL https://doi.org/10. 1021/ac...

    doi:10.24963/ijcai 2025

  23. [1694]

    wiley.com/doi/abs/10.1002/wcms.1694

    URL https://wires.onlinelibrary. wiley.com/doi/abs/10.1002/wcms.1694. 14 When Single Answer Is Not Enough: Rethinking Single-Step Retrosynthesis Benchmarks for LLMs A. Models’ Performance no reas, CREED reas, CREED+USPTO-fullCC rewCREED MAexact matching no reas, CREED+USPTO-fullGrok-4.1 Gemini 3 Flash prev MT rew NatureLM Claude 4.5 SonnetClaude 4.5 Opus ...

    work page doi:10.1002/wcms.1694 2026

  24. [2017]

    cc/paper_files/paper/2017/file/ ced556cd9f9c0c8315cfbe0744a3baf0-Paper

    URL https://proceedings.neurips. cc/paper_files/paper/2017/file/ ced556cd9f9c0c8315cfbe0744a3baf0-Paper. pdf. Kang, C., Liu, X., and Guo, F. Retrointext: A multimodal large language model enhanced framework for retrosyn- thetic planning via in-context representation learning. In The Thirteenth International Conference on Learning Representations, 2025. UR...

    arXiv 2017

  25. [2025]

    Livne, M., Miftahutdinov, Z., Tutubalina, E., Kuznetsov, M., Polykovskiy, D., Brundyn, A., Jhunjhunwala, A., Costa, A., Aliper, A., Aspuru-Guzik, A., and Zhavoronkov, A

    URL https://openreview.net/forum? id=30iBKSQMXn. Livne, M., Miftahutdinov, Z., Tutubalina, E., Kuznetsov, M., Polykovskiy, D., Brundyn, A., Jhunjhunwala, A., Costa, A., Aliper, A., Aspuru-Guzik, A., and Zhavoronkov, A. nach0: multimodal natural and chemical languages foun- dation model.Chem. Sci., 15:8380–8389, 2024. doi: 10.1039/D4SC00966E. URL http://dx...

    arXiv 2024

  26. [9294]

    doi: https://doi.org/10.1016/j.chempr.2018.02

    work page doi:10.1016/j.chempr.2018.02 2018