A Survey of Large Language Models for Text-Guided Molecular Discovery: from Molecule Generation to Optimization

Abhishek Pandey; Han Liu; Kaize Ding; Kexin Zhang; Yibo Wen; Zihan Zhao; Ziqing Wang

REVIEW 4 major objections 4 minor 1 cited by

A Survey of Large Language Models for Text-Guided Molecular Discovery: from Molecule Generation to Optimization

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This survey positions itself as the first systematic review of large language models as the text-guided engine for both molecule generation and molecule optimization, organized by whether the model is used frozen or fine-tuned.

desk verdict A useful organizational survey, but the 'first review' claim is unproven and the TOMG-Bench dataset size is inconsistent. read the letter →

arxiv 2505.16094 v1 pith:IFPRI4MW submitted 2025-05-22 cs.LG cs.CL

Ziqing Wang , Kexin Zhang , Zihan Zhao , Yibo Wen , Abhishek Pandey , Han Liu , Kaize Ding This is my paper · ORCID

classification cs.LGcs.CL

keywords LargelanguagemodelsMoleculegenerationoptimizationText-guidedmoleculardiscoveryLearningparadigmsSupervisedfine-tuningPreferencetuningIn-context

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

This survey aims to be the first systematic review of large language models used as the core engine for text-guided molecular discovery, covering the two central tasks of generating new molecules and optimizing existing ones. It argues that existing reviews either cover general deep generative models or treat LLMs only as part of broader chemical-AI surveys, leaving this specific intersection unmapped. The authors propose a taxonomy organized by learning paradigm: models used without any parameter update (zero-shot prompting and in-context learning) versus models adapted through supervised fine-tuning or preference tuning. They also catalogue the datasets, benchmarks, and metrics used to train and evaluate these systems. A reader would care because the taxonomy gives a common vocabulary for comparing approaches and identifies where LLMs actually carry the generative workload rather than playing auxiliary roles.

What carries the argument

The load-bearing object is the taxonomy itself, defined by task (generation versus optimization) crossed with learning paradigm (without tuning: zero-shot prompting and in-context learning; with tuning: supervised fine-tuning and preference tuning). The taxonomy does the argumentative work: it is what makes the survey first rather than merely another collection of papers, and every method reviewed is assigned a slot in it. Secondary machinery includes the decomposition of every prompt into instruction, few-shot examples, and property constraints, and the split of evaluation resources into structure-based metrics (validity, similarity, diversity) and property-based metrics (single- and multi-property success).

What would settle it

A reader could settle the central claim by checking whether the cited prior surveys, particularly Liao et al. (2024) and Zhang et al. (2025), contain sections treating both text-guided molecule generation and molecule optimization with a comparable learning-paradigm taxonomy; if either does, the claim of being the first systematic review fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that the emerging field of LLM-centric molecular discovery can be systematically organized by two tasks and four learning strategies. For molecule generation, the LLM is the de novo designer that converts instructions, optional few-shot examples, and property constraints into valid molecular strings. For molecule optimization, the LLM edits a given starting molecule to improve specified properties while preserving structural similarity. Within each task, methods divide into those that leave the pretrained model frozen and guide it through prompts, and those that update the model's weights, either by supervised fine-tuning on instruction datasets or by preference tuning with rewards or chosen-versus-rejected pairs. Grouping methods this way is meant to reveal how different strategies exploit LLM capabilities and what each costs in data and compute.

Load-bearing premise

The contribution rests on the claim that no earlier survey already covers LLM-centric molecule generation and optimization at comparable depth; if such a survey exists, the novelty is incremental rather than first.

Editorial extensions

If this is right

Researchers can position any new LLM method by answering two questions: whether it generates fresh molecules or edits existing ones, and whether it tunes the model or leaves it frozen.
The survey's dataset and metric catalogue gives a common reference for comparing methods, so reported gains in generation or optimization can be checked against shared benchmarks.
Preference tuning is identified as a growing route for aligning LLM outputs with chemical constraints by learning from chosen/rejected molecule pairs rather than absolute labels.
Multi-modal input, such as graphs, 3D coordinates, and fingerprints, is an emerging direction that addresses the limitations of text-only SMILES representations.
The authors identify open problems, including distinguishing harmful hallucination from useful creative leaps and building LLM agents that close the loop between prediction and experimental validation.

Reading between the lines

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

If the taxonomy is adopted, method comparisons are likely to shift from which architecture is used to which learning paradigm and training data are used, changing how contributions in this area are framed.
The learning-paradigm split predicts a trade-off: frozen-model approaches are cheaper and flexible but depend heavily on prompt quality, while tuned methods gain control at the cost of curated instruction data; this could be tested by re-benchmarking representative methods on the same dataset.
A testable extension is to build a shared leaderboard where every method in a given taxonomy cell runs on the same benchmark with the same validity and property metrics, exposing whether the taxonomy cells actually predict performance.
The paper's suggestion that controlled hallucination may serve as a creativity mechanism implies a quantitative test: measure whether models trained to suppress invalid outputs generate fewer novel scaffolds than models allowed modest hallucination rates.

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.

Desk Editor's Note

read the letter

Thanks for the report. I mostly agree with the conditional verdict. The value here is organizational: a clean taxonomy of zero-shot prompting, in-context learning, SFT, and preference tuning applied to molecule generation and optimization, plus curated tables of methods, datasets, and metrics. For someone entering the field, that is a real time-saver. The paper is readable and the citations look appropriate. As a survey, it does what it promises.

The soft spots are mostly specific. The TOMG-Bench scale mismatch is concrete: Table 1 lists 5k, Appendix B.2 says 45,000. That undermines confidence in a contribution that is explicitly a dataset summary. The 'first systematic review' claim is asserted but not demonstrated; the authors cite Liao et al. and Zhang et al. as general or small-model-focused, but never check whether those surveys already cover text-guided generation and optimization at comparable depth. Since the taxonomy is the main novelty, this priority claim is load-bearing. It may be true, but it needs evidence. Table 2 also has minor input-type misclassifications (FrontierX shows 'Instruction' only, though the text describes few-shot examples). Minor, but in a survey, summary accuracy is the product.

None of this is fatal. The paper is a reasonable map of the field in mid-2025. It is not the first and only resource, but it is a decent entry point. I would send it to peer review with a request to fix the dataset error and either substantiate or soften the 'first' framing. The taxonomy can stand without the priority claim.

Referee Report

4 major / 4 minor

Summary. This manuscript is a survey of large language models applied to text-guided molecular discovery, focusing on two tasks: molecule generation and molecule optimization. It proposes a taxonomy that separates methods operating without LLM tuning (zero-shot prompting and in-context learning) from those with tuning (supervised fine-tuning and preference tuning), and it maps representative methods to these categories. It also catalogs common datasets, benchmarks, and evaluation metrics, and it discusses future challenges such as hallucination mitigation, LLM agents, and multi-modal modeling. The paper claims to be the first systematic review focused on the intersection of LLMs and text-guided molecule generation and optimization, and it provides a continuously updated GitHub reading list.

Significance. If the priority claim is accepted, this survey would provide a useful organizing framework for a fast-growing area. The learning-paradigm taxonomy is simple and applicable, the dataset and metric catalog in the appendices is detailed, and the discussion of future directions is informed. The paper does not introduce new algorithms or results, but a well-scoped survey with a clear taxonomy is a legitimate contribution. The authors also provide a public reading list, which is a practical resource. However, the contribution depends on the accuracy of the taxonomy and dataset summaries; the priority claim and several table inconsistencies need to be verified before the survey can be relied upon.

major comments (4)

[Section 1] The paper's central novelty claim is not yet substantiated. Section 1 states that existing reviews 'fail to provide a comprehensive analysis of this specific intersection' and dismisses Liao et al. (2024) and Zhang et al. (2025) as too general or built around smaller language models, but it never demonstrates that those surveys lack LLM-centric molecule generation and optimization sections at comparable depth. Because 'first systematic review' is the stated contribution, please add a concrete scope comparison with the cited prior surveys (for example, a small table listing whether each covers generation, optimization, and the learning-paradigm taxonomy) or revise the claim to avoid asserting priority.
[Table 1 and Appendix B] The dataset summary contains factual inconsistencies that affect a stated contribution. Table 1 gives TOMG-Bench Scale=5k, while Appendix B.2 states that TOMG-Bench 'contains 45,000 samples'; Table 1 gives PubChem Scale=119M, while Appendix B.1 says PubChem 'holds over 750 million records'; Table 1 gives ZINC Scale >980M, while Appendix B.1 says ZINC contains 'over 20 million commercially available molecules.' Please define what 'Scale' counts (samples, compounds, records, or purchasable molecules) and harmonize every entry with the appendix text.
[Table 2] Table 2 contains classification and citation errors that undermine confidence in the survey's systematization. FrontierX is placed under 'In-Context Learning' with Input Type 'Instruction', but Section 3.1 states that FrontierX supplies 'detailed instructions alongside few-shot examples' within the prompt; the Input Type should include few-shot examples or the method should be moved to zero-shot prompting. Also, Mol-Instructions is listed as 'ICLR' in Table 2, while the reference list gives arXiv:2306.08018 and no ICLR publication; please verify and correct the venue.
[Figure 1 and Section 2.3] Figure 1's taxonomy is internally inconsistent with the text. Section 2.3 defines 'Without LLM Tuning' as encompassing both Zero-Shot Prompting and In-Context Learning, but Figure 1 shows only an 'In-Context Learning' branch under Generation w/o Tuning, while Optimization w/o Tuning has both branches. Section 3.1 also discusses zero-shot prompting for generation as 'challenging' but does not say it is excluded. Since the learning-paradigm taxonomy is a central contribution, please either include the (possibly empty) Zero-Shot Prompting branch for generation or explain explicitly why no zero-shot generation methods are categorized.

minor comments (4)

[Section 2.2] The itemized list contains the typo 'F ew-Shot Examples' with a stray space after 'F'; please remove the space.
[Appendix B.1 and Table 1] For PubChem, the appendix says 'over 750 million records' while Table 1 reports Scale=119M; please clarify whether the table counts compounds, substances, or records, and apply the same clarification to all large databases.
[Table 2] The 'Venue' column is inconsistent in formatting and completeness, with entries such as 'Arxiv' rather than 'arXiv' and missing preprint numbers; please standardize the venue information throughout.
[Section 5.1] MuMOInstruct and ChemData are categorized as 'Pretraining-Only' datasets, but they are instruction-tuning datasets rather than pretraining corpora in the standard LLM sense; please rename the category or clarify that 'training' here includes instruction tuning.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: this is a survey with an externally imposed taxonomy, not a derivation, fit, or prediction.

full rationale

This manuscript is a literature survey and taxonomy paper. It contains no equations that map inputs to predicted outputs, no fitted parameters later reported as predictions, and no derivation chain in which a claimed result is equivalent by construction to its inputs. The proposed taxonomy (without vs. with LLM tuning; zero-shot prompting, in-context learning, supervised fine-tuning, preference tuning) is an organizational framework imposed on the cited literature, not a consequence of any formula or fitted quantity, so none of the seven circularity patterns applies. The priority claim that this is 'the first systematic review focused on LLMs for text-guided molecular discovery for both generation and optimization tasks' is a factual claim about the literature that would need verification against earlier surveys such as Liao et al. (2024) and Zhang et al. (2025); an unverified priority claim is a correctness or novelty risk, not a circularity reduction. Similarly, the inconsistency between Appendix B.2's '45,000 samples' and Table 1's '5k' for TOMG-Bench is an internal consistency or accuracy issue, not evidence that any result is defined in terms of itself. The limitation section explicitly narrows the scope to generation and optimization and defers other areas to future work, which is transparent rather than circular. No self-citation is used to justify a load-bearing premise, and the paper does not invoke any uniqueness theorem or prior author-specific ansatz. Therefore the appropriate finding is no significant circularity with score 0.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

No free parameters or invented entities. The survey's conclusions rest on the adequacy of the chosen taxonomy, the representativeness of selected papers, and the accuracy of the authors' reading of those papers.

assumptions (3)

domain assumption LLMs are operationalized as foundation models with at least 1 billion parameters.
Section 2.1 declares this scope, which affects which methods are included or excluded.
domain assumption The taxonomy of four learning paradigms (zero-shot prompting, in-context learning, supervised fine-tuning, preference tuning) is sufficient and appropriate for categorizing the field.
The entire survey is organized around this framework; if a fifth paradigm exists, the claim of comprehensive coverage weakens.
domain assumption The selected methods are representative of the literature and are described accurately.
The survey does not use a formal systematic search protocol; coverage depends on the authors' selection and reading of primary sources.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Survey of Large Language Models for Text-Guided Molecular Discovery: from Molecule Generation to Optimization." pith.science (2026). https://pith.science/paper/IFPRI4MW

@misc{pith2026250516094,
  author       = {Pith},
  title        = {Pith review of: A Survey of Large Language Models for Text-Guided Molecular Discovery: from Molecule Generation to Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IFPRI4MW}},
  note         = {Machine review of arXiv:2505.16094}
}

read the original abstract

Large language models (LLMs) are introducing a paradigm shift in molecular discovery by enabling text-guided interaction with chemical spaces through natural language, symbolic notations, with emerging extensions to incorporate multi-modal inputs. To advance the new field of LLM for molecular discovery, this survey provides an up-to-date and forward-looking review of the emerging use of LLMs for two central tasks: molecule generation and molecule optimization. Based on our proposed taxonomy for both problems, we analyze representative techniques in each category, highlighting how LLM capabilities are leveraged across different learning settings. In addition, we include the commonly used datasets and evaluation protocols. We conclude by discussing key challenges and future directions, positioning this survey as a resource for researchers working at the intersection of LLMs and molecular science. A continuously updated reading list is available at https://github.com/REAL-Lab-NU/Awesome-LLM-Centric-Molecular-Discovery.

Figures

Figures reproduced from arXiv: 2505.16094 by the authors.

**Figure 2.** Overview of LLM-Centric Molecular Discovery. Left: Typical input components (Instruction, Few-Shot Examples, Property Constraints) for molecule generation and optimization. Right: Core learning paradigms for applying LLMs to Zero-Shot Prompting & In-Context Learning, Supervised Fine-Tuning and Preference Tuning. pairs to serve as more effective few-shot context, thereby boosting ICL performance. 3.2 Molecule Generat… view at source ↗

**Figure 3.** A Taxonomy of Benchmarking & Evaluation in Molecule Discovery. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Illustration of an example molecule and its representation in different data modalities. From left to right following the 2D chemical structure diagram: its 1D SMILES string representation, a simplified 2D graph view, and its 3D ball-and-stick model. fine-tuning chemic…

**Figure 5.** Figure 5: Visualization of the Instruction dataset of [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

A Systematic Survey on Large Language Models for Evolutionary Optimization: From Modeling to Solving
cs.NE 2025-09 conditional novelty 4.0 of 10

A literature survey that classifies LLM-based optimization research into modeling and solving, with solving divided into LLMs as optimizers, low-level components, and high-level managers.

Reference graph

Works this paper leans on

14 extracted references · 11 canonical work pages · cited by 1 Pith paper

[2]

ChatMol: A Versatile Molecule Designer Based on the Numerically Enhanced Large Language Model

Chatmol: A versatile molecule designer based on the numerically enhanced large language model.arXiv preprint arXiv:2502.19794. Yin Fang, Xiaozhuan Liang, Ningyu Zhang, Kangwei Liu, Rui Huang, Zhuo Chen, Xiaohui Fan, and Huajun Chen. 2023. Mol-instructions: A large-scale biomolecu- lar instruction dataset for large language models.arXiv preprint arXiv:2306...

work page Pith review arXiv 2023
[5]

Sunghwan Kim, Paul A Thiessen, Evan E Bolton, Jie Chen, Gang Fu, Asta Gindulyte, Lianyi Han, Jane He, Siqian He, Benjamin A Shoemaker, and 1 others

Pubchem 2025 update.Nucleic Acids Research, 53(D1):D1516–D1525. Sunghwan Kim, Paul A Thiessen, Evan E Bolton, Jie Chen, Gang Fu, Asta Gindulyte, Lianyi Han, Jane He, Siqian He, Benjamin A Shoemaker, and 1 others. 2016. Pubchem substance and compound databases.Nucleic acids research, 44(D1):D1202–D1213. Mario Krenn, Florian Häse, AkshatKumar Nigam, Pas- ca...

arXiv 2025
[8]

serves as a vast public chemical informa- tion repository, holding over 750 million records. It covers a wide array of data, including chemi- cal structures, identifiers, bioactivity outcomes, genes, proteins, and patents, and is organized into three interlinked databases: Substance (con- tributed chemical information), Compound (stan- dardized unique str...

work page
[10]

ready-to-dock

is presented as the first high-quality instruction-tuning dataset focused on complex, multi-property molecular optimization tasks. Un- like datasets such as MolOpt-Instruction (Ye et al., 2025) that primarily target single- or dual- property tasks, MuMOInstruct emphasizes tasks involving at least three properties, facilitating the evaluation of LLMs in bo...

work page 2025
[12]

• Novelty Rate:The novelty rate (Brown et al.,

and Unique@10k (within 10,000 valid samples) (Bagal et al., 2021), to assess short- range and broader diversity, respectively. • Novelty Rate:The novelty rate (Brown et al.,

work page 2021
[14]

Cate- gory

measures the size of the largest subset of generated molecules where no two molecules have a Tanimoto similarity exceeding a prede- fined threshold. A higher NCircles value indi- cates greater structural dissimilarity within the set. C.2 Property-Based Metrics Property-based metrics evaluate whether a de- signed or modified molecule satisfies specific phy...

work page 1968
[1968]

The journal of organic chemistry, 33(1):347–350

Linear free-energy relationship between partition coefficients and the aqueous solubility of organic liquids. The journal of organic chemistry, 33(1):347–350. Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, and 1 others. 2021. Lora: Low-rank adaptation of large language models. InICLR. Weihua Hu, Matthias Fey, ...

arXiv 2021
[1988]

These repre- sentations are well-suited for LLMs due to their compatibility with token-based language model- ing

andSELFIES(Self-Referencing Embed- ded Strings) (Krenn et al., 2020). These repre- sentations are well-suited for LLMs due to their compatibility with token-based language model- ing. Another format used in certain settings is the IUPAC nomenclature (Favre and Powell, 2014), which provides systematic names for molecules and is employed as an alternative o...

work page 2020

Show all 14 references

[2019]

It serves as an indicator of the model’s generalization ability and its potential to discover previously unseen chemical entities

measures the fraction of valid and unique generated molecules that are not present in the training dataset. It serves as an indicator of the model’s generalization ability and its potential to discover previously unseen chemical entities. A low novelty rate may suggest overfit...

2017
[2020]

InInternational Conference on Machine Learning, pages 4849–4859

Multi-objective molecule generation using inter- pretable substructures. InInternational Conference on Machine Learning, pages 4849–4859. PMLR. Sunghwan Kim, Jie Chen, Tiejun Cheng, Asta Gin- dulyte, Jia He, Siqian He, Qingliang Li, Benjamin A Shoemaker, Paul A Thiessen, Bo Yu...

2019
[2022]

provides a large-scale, well-curated dataset for molecular docking. It encompasses an exten- sive collection of docking scores and poses for more than 260,000 ligands against 58 medically relevant targets, and includes pharmaceutically relevant benchmark tasks such as virtual ...

2020
[2023]

is a large-scale, diverse, and high- quality dataset designed for the biomolecular do- main, featuring over 2 million carefully curated biomolecular instructions. It is structured around three core components: molecule-oriented in- structions (148.4K across six tasks focusing ...
[2024]

InAnnual Conference on Neural Information Processing Systems

Learning to optimize molecules with a chemi- cal language model. InAnnual Conference on Neural Information Processing Systems. 12 Sakhinana Sagar Srinivas and Venkataramana Runk- ana. 2024. Crossing new frontiers: Knowledge- augmented large language model prompting for zero- s...

2024 arXiv
[2025]

Nicola De Cao and Thomas Kipf

A zero-shot single-point molecule optimization model: Mimicking medicinal chemists’ expertise. Nicola De Cao and Thomas Kipf. 2018. Molgan: An implicit generative model for small molecular graphs. arXiv preprint arXiv:1805.11973. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and K...

2018 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [2]

ChatMol: A Versatile Molecule Designer Based on the Numerically Enhanced Large Language Model

Chatmol: A versatile molecule designer based on the numerically enhanced large language model.arXiv preprint arXiv:2502.19794. Yin Fang, Xiaozhuan Liang, Ningyu Zhang, Kangwei Liu, Rui Huang, Zhuo Chen, Xiaohui Fan, and Huajun Chen. 2023. Mol-instructions: A large-scale biomolecu- lar instruction dataset for large language models.arXiv preprint arXiv:2306...

work page Pith review arXiv 2023

[2] [5]

Sunghwan Kim, Paul A Thiessen, Evan E Bolton, Jie Chen, Gang Fu, Asta Gindulyte, Lianyi Han, Jane He, Siqian He, Benjamin A Shoemaker, and 1 others

Pubchem 2025 update.Nucleic Acids Research, 53(D1):D1516–D1525. Sunghwan Kim, Paul A Thiessen, Evan E Bolton, Jie Chen, Gang Fu, Asta Gindulyte, Lianyi Han, Jane He, Siqian He, Benjamin A Shoemaker, and 1 others. 2016. Pubchem substance and compound databases.Nucleic acids research, 44(D1):D1202–D1213. Mario Krenn, Florian Häse, AkshatKumar Nigam, Pas- ca...

arXiv 2025

[3] [8]

serves as a vast public chemical informa- tion repository, holding over 750 million records. It covers a wide array of data, including chemi- cal structures, identifiers, bioactivity outcomes, genes, proteins, and patents, and is organized into three interlinked databases: Substance (con- tributed chemical information), Compound (stan- dardized unique str...

work page

[4] [10]

ready-to-dock

is presented as the first high-quality instruction-tuning dataset focused on complex, multi-property molecular optimization tasks. Un- like datasets such as MolOpt-Instruction (Ye et al., 2025) that primarily target single- or dual- property tasks, MuMOInstruct emphasizes tasks involving at least three properties, facilitating the evaluation of LLMs in bo...

work page 2025

[5] [12]

• Novelty Rate:The novelty rate (Brown et al.,

and Unique@10k (within 10,000 valid samples) (Bagal et al., 2021), to assess short- range and broader diversity, respectively. • Novelty Rate:The novelty rate (Brown et al.,

work page 2021

[6] [14]

Cate- gory

measures the size of the largest subset of generated molecules where no two molecules have a Tanimoto similarity exceeding a prede- fined threshold. A higher NCircles value indi- cates greater structural dissimilarity within the set. C.2 Property-Based Metrics Property-based metrics evaluate whether a de- signed or modified molecule satisfies specific phy...

work page 1968

[7] [1968]

The journal of organic chemistry, 33(1):347–350

Linear free-energy relationship between partition coefficients and the aqueous solubility of organic liquids. The journal of organic chemistry, 33(1):347–350. Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, and 1 others. 2021. Lora: Low-rank adaptation of large language models. InICLR. Weihua Hu, Matthias Fey, ...

arXiv 2021

[8] [1988]

These repre- sentations are well-suited for LLMs due to their compatibility with token-based language model- ing

andSELFIES(Self-Referencing Embed- ded Strings) (Krenn et al., 2020). These repre- sentations are well-suited for LLMs due to their compatibility with token-based language model- ing. Another format used in certain settings is the IUPAC nomenclature (Favre and Powell, 2014), which provides systematic names for molecules and is employed as an alternative o...

work page 2020