Pith. sign in

REVIEW 5 major objections 5 minor 51 references

ChemActor: Enhancing Automated Extraction of Chemical Synthesis Actions with LLM-Generated Data

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read ChemActor, a fully fine-tuned 7B LLM boosted by 50,000 distribution-selected LLM-generated training samples, achieves state-of-the-art extraction of chemical synthesis actions from reaction descriptions.

desk verdict Real gains on a real task, but the paper's central attribution of those gains to chemically meaningful synthetic labels is not established, and a few internal contradictions need fixing before the SOTA claim is fully credible. read the letter →

arxiv 2506.23520 v2 pith:7QBAE2IW submitted 2025-06-30 cs.AI

Yu Zhang , Ruijie Yu , Jidong Tian , Feng Zhu , Jiapeng Liu , Xiaokang Yang , Yaohui Jin , Yanyan Xu This is my paper
classification cs.AI
keywords chemicalsynthesisactionextractionLLM-generateddataselectionsupervisedfine-tuningdescription-to-actionreaction-to-descriptionLLMevaluationdebateretrosynthesisvalidation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes ChemActor, a fully fine-tuned 7B-scale LLM that turns unstructured chemical reaction descriptions into structured, machine-executable action sequences such as add, reflux, and concentrate. To overcome scarce and unevenly distributed annotations, the authors build a sequential data-generation pipeline: starting from a single target molecule, they generate plausible reactants, validate the reaction with a forward model, have GPT-4o write reaction descriptions, and then use an existing action-extraction model plus GPT-4o formatting to produce action labels. A data-selection module keeps only generated samples whose embedding distribution diverges from real data by a KL-based threshold. On the CHEM TRANS benchmark, ChemActor trained with 50,000 such samples reaches state-of-the-art scores, improving BLEU-2 by about 9 points and Exact Match by 4.5 points over the same model without generated data. The paper also introduces a multi-round LLM debate metric that ranks ChemActor above baselines in a way consistent with human evaluation.

What carries the argument

The central mechanism is the sequential LLM-generated data framework. It starts from a target molecule, uses a single-step retrosynthesis model to propose reactants, a forward prediction model to check that the proposed reaction is plausible, GPT-4o with in-context examples to write a reaction description, the Paragraph2Actions model to predict procedural actions from that description, and GPT-4o once more to reformat the actions. A data-selection module embeds both real and generated (reaction, description, action) triplets with the D2A encoder, projects them with UMAP, and keeps only samples whose KL divergence from real data exceeds threshold $\tau=0.7$. That filtered synthetic set is mixed alternately with real data during supervised fine-tuning of LLaMA-2-7B. The multi-round circle review metric, based on iterative LLM debate, carries the paper's semantic-evaluation claim.

What would settle it

Take a random sample of the 50,000 generated training triplets and have expert chemists judge whether each predicted action sequence is chemically sensible and consistent with both the reaction description and the proposed reactants; if a substantial share are judged invalid or inconsistent, then the reported performance gains on CHEM TRANS cannot be attributed to learning genuine chemical procedure semantics.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that scaling up LLM-generated, distribution-selected synthetic data materially improves a fine-tuned LLM's ability to transcribe reaction descriptions into structured action sequences. ChemActor with the generated 50,000 samples achieves SM-A 76.88, BLEU-2 84.74, BLEU-4 76.93, and Exact Match 36.40 on CHEM TRANS D2A, surpassing both T5-ChemTrans and ChemActor without generated data; on OPEN EXP, ChemActor without generated data already outperforms all molecule-pretrained and description-based baselines on most metrics. The gains are attributed to the data selection module broadening chemical-space coverage rather than to raw data volume, since randomly adding 50,000 OPEN EXP samples yields only a 2.7% EM gain while selected generated samples yield a 14.1% EM gain. A secondary claim is that the multi-round LLM circle review, in which four LLMs debate semantic similarity scores, reflects advanced chemical understanding and agrees with human ranking.

Load-bearing premise

The load-bearing premise is that the LLM-generated data pipeline produces chemically valid action labels, because GPT-4o writes the description, Paragraph2Actions predicts the actions, GPT-4o only reformats them, and the forward-reaction model checks only the reactant-to-product transformation rather than the procedure text or action sequence.

Editorial extensions

If this is right

  • ChemActor with generated data can be used as a drop-in chemical executor for robotic synthesis platforms, translating literature descriptions without per-dataset rule engineering.
  • The KL-divergence data selection criterion should transfer to other low-annotation scientific domains, reducing the need for expensive human labels.
  • The reported gains imply that a 7B open-weight model can outperform much larger or molecule-pretrained models when given enough selected synthetic procedure data.
  • The LLM circle review metric offers a human-aligned way to judge semantic correctness of generated action sequences, complementing exact-match metrics.
  • The type-matching and low-frequency action results suggest synthetic data helps rare actions such as DISTILL and REFLUX, improving coverage of the action taxonomy.

Reading between the lines

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

  • The paper leaves implicit that the same selection pipeline could be run in reverse: sampling molecules from underrepresented regions of chemical space to create targeted training data for rare action types, a testable extension.
  • If the generated labels are even partially noisy, the gains might come from distributional smoothing rather than correct chemistry; a direct check would compare against a model trained on the same volume of synthetic descriptions with independently human-verified actions.
  • The circle review's agreement with human ranking is demonstrated on 279 bad cases; a broader validation would run the debate metric on the full test set against multiple expert annotators.
  • A practical consequence the authors do not develop is that the same framework could generate synthetic reaction descriptions for the action-to-description direction, which Table 1 also shows improves.
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

5 major / 5 minor

Summary. This paper describes ChemActor, a fully fine-tuned LLaMA-2-7B model for converting unstructured chemical reaction descriptions into structured action sequences (D2A) and the reverse direction (A2D). To address data scarcity, the authors propose a sequential LLM-generated data framework: a retrosynthesis model proposes reactants, a forward model validates the reaction, GPT-4o writes a reaction description, Paragraph2Actions predicts procedural actions, and GPT-4o refines the formatting. A data selection module filters generated triplets based on the KL divergence between UMAP projections of real and generated description embeddings. The authors also introduce a multi-round LLM 'circle review' metric for semantic evaluation. On CHEM TRANS and OPEN EXP, ChemActor trained with 50,000 generated samples is reported to outperform baselines, with gains of roughly 10% over ChemActor without generated data (e.g., BLEU-2 75.63 to 84.74 and EM 31.9 to 36.4 on CHEM TRANS D2A). The paper attributes these gains to the LLM-generated data and the proposed selection module, and claims that the new circle-review metric reflects advanced chemical understanding.

Significance. If the central claim holds, the paper makes a meaningful contribution: it demonstrates that procedurally generated synthetic data can improve structured action extraction in chemistry, and it provides an open implementation (code is available at the cited repository). The numerical results are real measurements, and the paper includes extensive ablations on training proportions, data size, and type-matching errors, which are useful for the community. The type-matching analysis in Appendix C.2 is a particular strength because it goes beyond token-overlap metrics. However, the contribution is conditional on the validity of the synthetic action labels. Because the action labels are produced by Paragraph2Actions, which achieves EM=0 on the target CHEM TRANS task, and are never independently validated, the observed gains could plausibly reflect increased coverage of description styles or schema-format exposure rather than chemically meaningful action supervision. The inconsistent statement of the selection criterion and the absence of a same-pool random control for the selection module further weaken the attribution.

major comments (5)
  1. [Section 3.3 / Algorithm 1] The data selection criterion is stated inconsistently. Algorithm 1 appends a triplet set when delta >= tau, while Section 3.3 states that data are retained when tau > delta and resampled otherwise. These are opposite rules. The experimental behavior in Table 4, where larger tau increases the time spent searching for qualifying data, supports the Algorithm 1 version (delta >= tau), because a higher threshold makes the condition harder to satisfy. The text must be corrected and all reported results tied unambiguously to one rule.
  2. [Section 3.2 / Figure 1B / Table 7] The paper provides no evidence that the 50,000 LLM-generated (description, action) pairs are chemically valid. The action labels are produced by Paragraph2Actions, which scores EM=0 and SM-A=21.57 on CHEM TRANS (Table 1), followed only by LLM reformatting; the forward-reaction model validates reactants/products, not the procedure text or the action sequence. The selection module operates on description embeddings and is therefore blind to action-label correctness. Because Table 7 attributes the main gain to these generated samples, the claim that the improvement reflects enriched understanding of chemical procedures is not established. The authors should add a validity check, for example a human expert evaluation of a random sample of generated pairs, or an ablation that pairs generated descriptions with existing gold action sequences.
  3. [Table 7] The random-OPENEXP control does not isolate the effect of the data selection module. Randomly sampled OPEN EXP data differ from the LLM-generated pool in source, label provenance, and schema coverage; the comparison therefore conflates the selection mechanism with dataset differences. A proper control would randomly sample the same LLM-generated pool without selection (and ideally also apply the selection module to OPEN EXP). Without this control, the conclusion that the data selection module generates more valuable data is not supported.
  4. [Table 2] The row 'ChemActor*, wo/g' is placed inside the block 'With molecule pre-training, reaction equation yields experimental actions', but ChemActor is a description-to-action model without molecule pre-training. The asterisk is never defined, and the input modality for this row is unspecified. As written, the row appears to compare ChemActor against reaction-equation models under the wrong input setting, which affects the claimed superiority on OPEN EXP. The row should be moved to the description-input block or clearly labeled with its actual input.
  5. [Section 4.4 / Appendix C.4] The description of the bad-case filtering is contradictory: the text says cases are retained when difflib similarity is greater than 0.4 and then calls the remaining set 'bad cases', but bad cases should be those with low similarity. In addition, GPT-4o participates both as the generator of synthetic reaction descriptions and as one of the evaluators in the circle review, so the metric may systematically reward outputs that resemble LLM-generated text. The paper should report per-item agreement or correlation with human scores, not only the ranking trend, and should justify or mitigate the evaluator/generator overlap.
minor comments (5)
  1. [Section 4.5.1] The sentence 'Increasing tau from 0.5 to 0.8 consistently improves all evaluation metrics' is not supported by Table 4: BLEU-2 drops from 76.17 at tau=0.6 to 76.01 at tau=0.7, and BLEU-4 drops from 68.74 to 68.71. Please rephrase to reflect the metric-wise changes.
  2. [Sections 1 and 3.2] The text contains typos: 'repspectively' in Section 3.2 and 'nature language' in the introduction; these should be corrected.
  3. [Figure 1] The box labeled 'E.x. Submission' contains garbled text that appears to be a rendering artifact; please replace it with a clean figure.
  4. [Table 3] Only 'ChemActor (wo/g)' is evaluated under the circle review, not the final 'ChemActor, w/generated' model. Since the abstract's advanced-understanding claim refers to the augmented model, this should be stated explicitly or the final model should be included.
  5. [Section 4.5.3] The phrase 'Method criterion refers to strategies for data augmentation' is an incomplete definition; please clarify what the 'criterion' column means in Table 7.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central SOTA claim is an empirical result on held-out test sets using standard metrics, not an artifact of the data-generation or evaluation loops.

full rationale

The paper's main claim, that ChemActor with 50,000 LLM-generated samples outperforms the baseline by roughly 10%, is supported by Table 1 and Table 7 on the CHEM TRANS test set using standard BLEU, EM, and SM-A metrics. These scores are measured against held-out human annotations and are not defined in terms of the data-selection criterion or the LLM review scores. The data selection module (Algorithm 1) uses the D2A encoder's embeddings to filter generated triplets by distribution divergence; this is a training-data filtering heuristic, not a construction of the test-set predictions. Even if the selector and final model share an encoder lineage, the evaluation is independent of the selection rule. The self-citation to Ualign (Zeng et al., 2024) is a component dependency in the generation pipeline, and that model is externally validated on retrosynthesis benchmarks; it is not used to forbid alternatives or to define the outcome. The multi-round LLM circle review is an auxiliary metric and is not load-bearing for the abstract's performance claim; any potential evaluator overlap with the data generator is an evaluation-quality concern, not a by-construction equivalence. No equation in the paper defines a predicted quantity in terms of the fitted parameters or the training signal, and no 'prediction' is statistically forced by a subset fit. The derivation chain from data generation to training to evaluation is therefore self-contained with respect to circularity, and the reported gains, whatever their chemical validity may be, are genuine empirical measurements rather than artifacts of circular reasoning.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

No new physical or chemical entities are introduced. The 'multi-round LLMs circle review' is a new evaluation procedure, not an invented entity. The central claim rests mainly on the quality and representativeness of synthetic training data and on a tuned selection threshold.

free parameters (2)
  • Data selection threshold tau = 0.7
    Tuned in Table 4 from {0.5, 0.6, 0.7, 0.8} on CHEM TRANS; all main results use this value, and performance varies from 30.94 to 34.33 EM as tau changes.
  • Number of generated training samples n = 50,000
    The paper reports 300, 600, 2,100 and 50,000 generated samples (Table 7) and adopts 50,000 for the headline result; larger n is not tested beyond this, so the gain may depend on this specific quantity.
assumptions (5)
  • domain assumption GPT-4o-generated reaction descriptions, conditioned on retrosynthesis products, are chemically plausible and representative of real experimental procedures.
    Invoked in Section 3.2 and Figure 1B; no chemical validation of the generated descriptions is provided beyond a forward-reaction correctness check on the reactant-product pair.
  • domain assumption Action labels produced by Paragraph2Actions and refined by GPT-4o are accurate enough to serve as training targets for 50,000 generated examples.
    Section 3.2 and Table 7; Paragraph2Actions has EM=0 on CHEM TRANS (Table 1), so the supervision signal for generated data is noisy at best.
  • domain assumption KL divergence between UMAP projections of embedding sets is a valid measure of distribution difference for data selection.
    Algorithm 1 and Section 3.3 use KL(e) and KL(e*) without defining how distributions are estimated from 2D UMAP points; the validity of the criterion is assumed.
  • domain assumption The forward prediction model's self-correctness check ensures validity of the generated reaction.
    Section 3.2 uses Schwaller et al. forward model to validate reactants-to-product, but this does not validate the procedure description or action sequence.
  • domain assumption LLM circle-review scores reflect semantic correctness of chemical actions.
    Section 4.4 and Appendix C.4; human evaluation is limited to six students each scoring a subset, and the reported consistency is contradicted by Table 3 for Mistral vs T5-ChemTrans.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ChemActor: Enhancing Automated Extraction of Chemical Synthesis Actions with LLM-Generated Data." pith.science (2026). https://pith.science/paper/7QBAE2IW

@misc{pith2026250623520,
  author       = {Pith},
  title        = {Pith review of: ChemActor: Enhancing Automated Extraction of Chemical Synthesis Actions with LLM-Generated Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7QBAE2IW}},
  note         = {Machine review of arXiv:2506.23520}
}
read the original abstract

With the increasing interest in robotic synthesis in the context of organic chemistry, the automated extraction of chemical procedures from literature is critical. However, this task remains challenging due to the inherent ambiguity of chemical language and the high cost of human annotation required for developing reliable computer-aided extraction protocols. Here, we present ChemActor, a fully fine-tuned large language model (LLM), as a chemical executor to convert between unstructured experimental procedures and structured action sequences. We propose a sequential LLM-generated data framework to address the challenges of insufficient and low-quality annotated data. This framework integrates a data selection module that selects data based on distribution divergence, with a general-purpose LLM, to generate machine-executable actions from a single molecule input. Additionally, we introduce a novel multi-round LLMs circle review metric, which reflects the model's advanced understanding of chemical experimental procedures. Extensive experiments on reaction-to-description (R2D) and description-to-action (D2A) tasks demonstrate that ChemActor, augmented by LLM-generated data, achieves state-of-the-art performance, outperforming the baseline model by 10%. The code is available at: https://github.com/Zhanghahah/ChemActor.

Figures

Figures reproduced from arXiv: 2506.23520 by the authors.

Figure 1
Figure 1. The overview of ChemActor. (A) Task definition; (B) Illustration of the framework of LLM-generated data; (C) BERTScore and multi-round LLMs circle review metrics for evalution. generated and real data in this low-dimensional space. Specifically, we compute the KL divergence for the distributions of LLM-generated data and original data, denoted as KL(e ∗ ), and KL(e), re￾spectively. We then define a threshold τ for t… view at source ↗
Figure 2
Figure 2. Results for multi-round LLMs circle reviews. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Comparison of the action types on the CHEMTRANS and OPENEXP datasets. pairs. • T5 (Raffel et al., 2020). T5 is a method based on an encoder-decoder architecture, which has achieved competitive performance on a broad range of natural language processing tasks. For the T5-Base model, it consists of 12 en￾coder layers and 12 decoder layers, and is scaled up to 220M parameters. • T5-ChemTrans (Zeng et al., 2023). T5- Ch… view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Examples of constructed Q&A datasets on C [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: Data distribution of the OPENEXP and CHEMTRANS datasets. the codes released by (Zhang et al., 2024). During the training period, we set the batch size as eight, the maximum number of epochs as five and the learning rate as 1e-5. • Mistral-7B-Instruct-v0.2. We follow (Z…
Figure 6
Figure 6. Figure 6: Heatmap of type matching results of ChemActor and T5-ChemTrans. [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: The performance varying different training proportions on C [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: The performance of different data curriculum [PITH_FULL_IMAGE:figures/full_fig_p020_8.png]
Figure 9
Figure 9. Figure 9: The pipeline of the multi-round LLMs circle review. [PITH_FULL_IMAGE:figures/full_fig_p021_9.png]
Figure 10
Figure 10. Figure 10: An example of the performance of multi-round LLMs circle review (bad case 1). [PITH_FULL_IMAGE:figures/full_fig_p022_10.png]
Figure 11
Figure 11. Figure 11: An example of the performance of multi-round LLMs circle review (bad case 2). [PITH_FULL_IMAGE:figures/full_fig_p023_11.png]
Figure 12
Figure 12. Figure 12: Experimental results of ChemActor on CHEMTRANS and OPENEXP datasets [PITH_FULL_IMAGE:figures/full_fig_p024_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 34 canonical work pages

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. GPT-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  2. [2]

    Zachary J Baum, Xiang Yu, Philippe Y Ayala, Yanan Zhao, Steven P Watkins, and Qiongqiong Zhou. 2021. Artificial Intelligence in Chemistry: Current Trends and Future Directions . Journal of Chemical Information and Modeling , 61(7):3197--3212

    work page 2021

  3. [3]

    Daniil A Boiko, Robert MacKnight, Ben Kline, and Gabe Gomes. 2023. Autonomous Chemical Research with Large Language Models . Nature, 624(7992):570--578

    work page 2023

  4. [4]

    Xunxin Cai, Meng Xiao, Zhiyuan Ning, and Yuanchun Zhou. 2023. Resolving the imbalance issue in hierarchical disciplinary topic inference via llm-based data augmentation. In 2023 IEEE International Conference on Data Mining Workshops (ICDMW), pages 1424--1429. IEEE

    work page 2023

  5. [5]

    Richard B Canty and Milad Abolhasani. 2024. Reproducibility in automated chemistry laboratories using computer science abstractions. Nature Synthesis, 3(11):1327--1339

    work page 2024

  6. [6]

    Jiaao Chen, Derek Tam, Colin Raffel, Mohit Bansal, and Diyi Yang. 2023. An empirical survey of data augmentation for limited data learning in nlp. Transactions of the Association for Computational Linguistics, 11:191--211

    work page 2023

  7. [7]

    Steffi Chern, Ethan Chern, Graham Neubig, and Pengfei Liu. 2024. Can large language models be trusted for evaluation? scalable meta-evaluation of llms as evaluators via agent debate. arXiv preprint arXiv:2401.16788

    arXiv 2024

  8. [8]

    Dimitrios Christofidellis, Giorgio Giannone, Jannis Born, Ole Winther, Teodoro Laino, and Matteo Manica. 2023. Unifying molecular and textual representations via multi-task language modelling. In International Conference on Machine Learning, pages 6140--6157. PMLR

    work page 2023

Show all 51 references
  1. [9]

    Zhumin Chu, Qingyao Ai, Yiteng Tu, Haitao Li, and Yiqun Liu. 2024. Pre: A peer review based large language model evaluator. arXiv preprint arXiv:2401.15641

    2024 arXiv

  2. [10]

    Connor W Coley, Dale A Thomas III, Justin AM Lummiss, Jonathan N Jaworski, Christopher P Breen, Victor Schultz, Travis Hart, Joshua S Fishman, Luke Rogers, Hanyu Gao, et al. 2019. A robotic platform for flow synthesis of organic compounds informed by AI planning. Science, (645...

    2019

  3. [11]

    Ian W Davies. 2019. The digitization of organic synthesis. Nature, 570(7760):175--181

    2019

  4. [12]

    Linnea Evanson, Yair Lakretz, and Jean-R \'e mi King. 2023. Language acquisition: Do children and language models follow similar learning stages? arXiv preprint arXiv:2306.03586

    2023 arXiv

  5. [13]

    Vincent Fan, Yujie Qian, Alex Wang, Amber Wang, Connor W Coley, and Regina Barzilay. 2024. Openchemie: An information extraction toolkit for chemistry literature. Journal of Chemical Information and Modeling, 64(14):5521--5534

    2024

  6. [14]

    Cong Gao, Benjamin D Killeen, Yicheng Hu, Robert B Grupp, Russell H Taylor, Mehran Armand, and Mathias Unberath. 2023. Synthetic data accelerates the development of generalizable learning-based algorithms for X-ray image analysis. Nature Machine Intelligence , 5(3):294--308

    2023

  7. [15]

    Alexander G Godfrey, Thierry Masquelin, and Horst Hemmerle. 2013. A remote-controlled adaptive medchem lab: An innovative approach to enable drug discovery in the 21st century. Drug Discovery Today, 18(17-18):795--802

    2013

  8. [16]

    Jiang Guo, A Santiago Ibanez-Lopez, Hanyu Gao, Victor Quach, Connor W Coley, Klavs F Jensen, and Regina Barzilay. 2021. Automated chemical reaction extraction from scientific literature. Journal of chemical information and modeling, 62(9):2035--2045

    2021

  9. [17]

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12):1--38

    2023

  10. [18]

    Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825

    2023 arXiv

  11. [19]

    Steven M Kearnes, Michael R Maser, Michael Wleklinski, Anton Kast, Abigail G Doyle, Spencer D Dreher, Joel M Hawkins, Klavs F Jensen, and Connor W Coley. 2021. The open reaction database. Journal of the American Chemical Society , 143(45):18820--18826

    2021

  12. [20]

    Vladimir I Levenshtein et al. 1966. Binary codes capable of correcting deletions, insertions, and reversals. In Soviet physics doklady, volume 10, pages 707--710. Soviet Union

    1966

  13. [21]

    Haoran Li, Qingxiu Dong, Zhengyang Tang, Chaojun Wang, Xingxing Zhang, Haoyang Huang, Shaohan Huang, Xiaolong Huang, Zeqiang Huang, Dongdong Zhang, et al. 2024. Synthetic data (almost) from scratch: Generalized instruction tuning for language models. arXiv preprint arXiv:2402.13064

    2024 arXiv

  14. [22]

    Zhuoyan Li, Hangxiao Zhu, Zhuoran Lu, and Ming Yin. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.647 Synthetic data generation with large language models for text classification: Potential and limitations . In Proceedings of the 2023 Conference on Empirical Methods in Nat...

    2023 doi

  15. [23]

    Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74--81

    2004

  16. [24]

    reaction-ship

    Zhiyuan Liu, Yaorui Shi, An Zhang, Sihang Li, Enzhi Zhang, Xiang Wang, Kenji Kawaguchi, and Tat-Seng Chua. 2024. Reactxt: Understanding molecular" reaction-ship" via reaction-contextualized molecule-text pretraining. arXiv preprint arXiv:2405.14225

    2024 arXiv

  17. [25]

    Daniel Lowe. 1976. Chemical reactions from us patents (1976-sep2016), 2017. DOI, 10:m9

    1976

  18. [26]

    Bran, Sam Cox, Oliver Schilter, Carlo Baldassari, Andrew D White, and Philippe Schwaller

    Andres M. Bran, Sam Cox, Oliver Schilter, Carlo Baldassari, Andrew D White, and Philippe Schwaller. 2024. Augmenting large language models with chemistry tools. Nature Machine Intelligence, pages 1--11

    2024

  19. [27]

    S Hessam M Mehr, Matthew Craven, Artem I Leonov, Graham Keenan, and Leroy Cronin. 2020. A universal system for digitization and automatic execution of the chemical synthesis literature. Science, 370(6512):101--108

    2020

  20. [28]

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: A method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311--318

    2002

  21. [29]

    Mark Peplow. 2014. the pobo-chemist. Nature, 512(7512):20

    2014

  22. [30]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1--67

    2020

  23. [31]

    Laria Reynolds and Kyle McDonell. 2021. Prompt Programming for Large Language Models: Beyond the Few-Shot Paradigm . In Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems , pages 1--7

    2021

  24. [32]

    Philippe Schwaller, Riccardo Petraglia, Valerio Zullo, Vishnu H Nair, Rico Andreas Haeuselmann, Riccardo Pisoni, Costas Bekas, Anna Iuliano, and Teodoro Laino. 2020. Predicting retrosynthetic pathways using transformer-based models and a hyper-graph exploration strategy. Chemi...

    2020

  25. [33]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open Foundation and Fine-Tuned Chat Models . arXiv preprint arXiv:2307.09288

    2023 arXiv

  26. [34]

    Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. 2018. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In Proceedings of the IEEE c...

    2018

  27. [35]

    Trieu H Trinh, Yuhuai Wu, Quoc V Le, He He, and Thang Luong. 2024. Solving olympiad geometry without human demonstrations. Nature, 625(7995):476--482

    2024

  28. [36]

    Alain C Vaucher, Philippe Schwaller, Joppe Geluykens, Vishnu H Nair, Anna Iuliano, and Teodoro Laino. 2021. Inferring experimental procedures from text-based representations of chemical reactions. Nature communications, 12(1):2573

    2021

  29. [37]

    Alain C Vaucher, Federico Zipoli, Joppe Geluykens, Vishnu H Nair, Philippe Schwaller, and Teodoro Laino. 2020. Automated extraction of chemical synthesis actions from experimental procedures. Nature communications, 11(1):3601

    2020

  30. [38]

    Eric Walker, Joshua Kammeraad, Jonathan Goetz, Michael T Robo, Ambuj Tewari, and Paul M Zimmerman. 2019. Learning to predict reaction conditions: Relationships between solvent, molecular structure, and catalyst. Journal of chemical information and modeling, 59(9):3645--3654

    2019

  31. [39]

    Smith, Daniel Khashabi, and Hannaneh Hajishirzi

    Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2023. https://doi.org/10.18653/v1/2023.acl-long.754 Self-Instruct: Aligning Language Models with Self-Generated Instructions . In Proceedings of the 61st Annual Mee...

    2023 doi

  32. [40]

    Zhenhua Wang, Guang Xu, and Ming Ren. 2024. Llm-generated natural language meets scaling laws: New explorations and data augmentation methods. arXiv preprint arXiv:2407.00322

    2024 arXiv

  33. [41]

    Chenxi Whitehouse, Monojit Choudhury, and Alham Fikri Aji. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.44 LLM -powered data augmentation for enhanced cross-lingual performance . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pa...

    2023 doi

  34. [42]

    Jiayi Yuan, Ruixiang Tang, Xiaoqian Jiang, and Xia Hu. 2023. Large language models for healthcare data augmentation: An example on patient-trial matching. In AMIA Annual Symposium Proceedings, volume 2023, page 1324. American Medical Informatics Association

    2023

  35. [43]

    Kaipeng Zeng, Bo Yang, Xin Zhao, Yu Zhang, Fan Nie, Xiaokang Yang, Yaohui Jin, and Yanyan Xu. 2024. Ualign: pushing the limit of template-free retrosynthesis prediction with unsupervised smiles alignment. Journal of Cheminformatics, 16(1):80

    2024

  36. [44]

    Zheni Zeng, Yi-Chen Nie, Ning Ding, Qian-Jun Ding, Wei-Ting Ye, Cheng Yang, Maosong Sun, E Weinan, Rong Zhu, and Zhiyuan Liu. 2023. Transcription between human-readable synthetic descriptions and machine-executable instructions: An application of the latest pre-training techno...

    2023

  37. [45]

    Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2019. Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations

    2019

  38. [46]

    Wei Zhang, Qinggong Wang, Xiangtai Kong, Jiacheng Xiong, Shengkun Ni, Duanhua Cao, Buying Niu, Mingan Chen, Yameng Li, Runze Zhang, et al. 2024. Fine-tuning large language models for chemical text mining. Chemical Science

    2024

  39. [47]

    Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, et al. 2023. Siren's song in the ai ocean: A survey on hallucination in large language models. arXiv preprint arXiv:2309.01219

    2023 arXiv

  40. [48]

    Ruochen Zhao, Wenxuan Zhang, Yew Ken Chia, Deli Zhao, and Lidong Bing. 2024. Auto arena of llms: Automating llm evaluations with agent peer-battles and committee discussions. arXiv preprint arXiv:2405.20267

    2024 arXiv

  41. [49]

    Ming Zhong, Siru Ouyang, Minhao Jiang, Vivian Hu, Yizhu Jiao, Xuan Wang, and Jiawei Han. 2023. https://doi.org/10.18653/v1/2023.findings-acl.767 R eact IE : Enhancing chemical reaction extraction with weak supervision . In Findings of the Association for Computational Linguist...

    2023 doi

  42. [50]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  43. [51]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.