pith. sign in

arxiv: 2606.03057 · v1 · pith:GYSDJ3TAnew · submitted 2026-06-02 · 💻 cs.LG · cs.AI

Rethinking Molecular Text Representations for LLMs: An Empirical Study

Arun Raja , Garrett M. Morris , Kian Ming A. Chai This is my paper

Pith reviewed 2026-06-28 11:14 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords molecular representationsLLMschemical benchmarksSMILESInChIIUPACCMLMolJSON
0
0 comments X

The pith

LLM performance on molecular tasks varies sharply with the choice of text representation for molecules.

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

This paper conducts a large-scale benchmark of how different ways to write molecules as text affect large language model results on chemical problems. Nine representations including SMILES strings, InChI, IUPAC names, and structured formats like CML and MolJSON are tested on eight tasks with sixteen models spanning general, reasoning, and chemistry-specialized types. Results show that no format is best for everything, though CML leads overall and explicit structures handle structural questions better while IUPAC handles meaning-based ones. The study also reveals that models trained mainly on SMILES struggle when given other formats, and internal model behaviors like attention differ by representation. These patterns indicate that assuming one representation fits all uses is not reliable for chemistry applications of LLMs.

Core claim

The authors establish through systematic evaluation that molecular text representation choice is a dominant factor in LLM competence on chemistry tasks. Explicit structured representations such as CML and MolJSON perform best on structural tasks, IUPAC names lead on semantic tasks including molecule retrieval for every model tested, and common SMILES variants are seldom the top choice. Chemistry-specialized models excel with SMILES but suffer notable drops with structured text, while tokenization and attention analyses indicate that different representations engage distinct internal mechanisms within the models.

What carries the argument

The benchmark comparing nine molecular text representations on eight tasks across sixteen LLMs, combined with mechanistic probes of tokenization, linear probes, and attention patterns.

Load-bearing premise

The selected eight tasks and nine representations adequately cover the variety of real-world molecular applications for LLMs.

What would settle it

Finding a molecular task or set of models where one representation, such as canonical SMILES, achieves the highest performance across all evaluated cases would challenge the dependence on representation.

Figures

Figures reproduced from arXiv: 2606.03057 by Arun Raja, Garrett M. Morris, Kian Ming A. Chai.

Figure 2
Figure 2. Figure 2: Gemini 3 Flash’s scoring (from 1 to 5) across [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Reasoning models reduce hallucination. ning, 1997), as a chemical weapon and refuses to proceed with the response. Given that hallucination errors are diverse, using LLM-as-a-judge in addi￾tion to rigid metrics helps to analyse the qualitative nature of those errors. Finally, reasoning LLMs reduce hallucination ( [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Token count distribution across 8 representa [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Linear probing performance of Qwen3-4B across all representations for molecular weight, log P, and TPSA prediction Using the same 1,000 molecules from Section 6.1, we interrogate the hidden embeddings of each rep￾resentation in Qwen3-4B to see if they are rich enough to predict chemical properties. Such linear probing can elicit if the representation already pro￾vides sufficient signals in the intermediate… view at source ↗
Figure 6
Figure 6. Figure 6: Attention between the last token and the [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Attention within the molecule within-molecule attention compared to MolJSON and CML, suggesting that compact representations encourage richer token-to-token interaction within the molecular string. 7 Conclusion We introduced MolRepBench, a systematic eval￾uation of nine molecular text representations across eight tasks and 16 LLMs spanning open￾weight, reasoning-tuned, domain-specialized, and closed fronti… view at source ↗
Figure 9
Figure 9. Figure 9: Attention between the last token and molecule [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Attention within the molecule tokens. We perform this analysis from layers 17 to 24. We chose this range of layers as they form the intermediate layers and exhibit reasonable perfor￾mance during linear probing. CML and MolJSON have the largest last token to molecule attention, nearly 5 to 10 times that of canonical SMILES. This is consistent with the long length of the explicit graph representations: the … view at source ↗
Figure 8
Figure 8. Figure 8: Reasoning token lengths for atom counting [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗
Figure 11
Figure 11. Figure 11: Linear CKA similarity between representation pairs of chemistry-specialized models, ChemDFM-v2.0 [PITH_FULL_IMAGE:figures/full_fig_p025_11.png] view at source ↗
read the original abstract

Large language models (LLMs) are increasingly used for molecular tasks, but it remains unclear which molecular representation to use. We present a systematic benchmark evaluating LLM molecular competence across nine representations and eight chemical tasks. We benchmark 16 LLMs across five model families, including reasoning and non-reasoning variants, chemistry-specialized LLMs, and closed frontier models. Performance is strongly representation-dependent and no single representation wins across tasks, though CML is the best, followed by MolJSON, InChI, and then canonical SMILES. Explicit structured text representations (CML and MolJSON) dominate structural tasks; IUPAC dominates semantic tasks, winning molecule retrieval for all 16 LLMs; and SMILES variants are rarely optimal despite their prevalence in pretraining. Chemistry-specialized models perform well with SMILES at the cost of large degradations with structured text representations, suggesting SMILES-only evaluation rewards specialization that does not generalize. Using LLM-as-a-judge, we find that IUPAC produces the highest fraction of correct molecule generations. A mechanistic study via tokenization audits, linear probes and attention shows that representations are encoded differently inside the model; for example, structured representations require higher attention across the molecular span. Our results argue against representation-invariant evaluation and motivate task-aware representation routing for LLM-based chemistry.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 3 minor

Summary. The manuscript reports a systematic empirical benchmark of nine molecular text representations (CML, MolJSON, InChI, canonical SMILES, and variants, plus IUPAC) across eight chemical tasks using sixteen LLMs spanning five families. It claims performance is strongly representation-dependent with no universal winner; CML ranks highest overall, followed by MolJSON, InChI, and canonical SMILES; explicit structured representations dominate structural tasks while IUPAC dominates semantic tasks (including perfect retrieval wins); SMILES variants are rarely optimal; chemistry-specialized models excel with SMILES but degrade sharply on structured text; LLM-as-a-judge favors IUPAC for generation correctness; and mechanistic probes (tokenization audits, linear probes, attention) reveal representation-specific internal encodings, e.g., higher attention span for structured reps. The work concludes against representation-invariant evaluation and for task-aware routing.

Significance. If the benchmark results and mechanistic findings hold under proper statistical controls, the paper makes a substantive contribution by demonstrating that representation choice is a first-order factor in molecular LLM performance and by providing concrete evidence for task-dependent routing. Credit is due for the breadth (16 models including closed frontier and specialized variants), the inclusion of both structural and semantic tasks, the LLM-as-a-judge analysis, and the mechanistic component that links surface performance to internal model behavior. The study supplies a useful reference point for future work on representation-aware chemistry LLMs.

major comments (2)
  1. [§4, Tables 1–3] §4 (Results) and Tables 1–3: Performance differences and rankings (e.g., CML best overall, structured reps dominating structural tasks) are reported without error bars, number of evaluation runs, random seeds, or statistical significance tests. This is load-bearing for the central claim that performance is “strongly representation-dependent” and that specific orderings hold.
  2. [§3] §3 (Benchmark Design): The selection of exactly eight tasks and nine representations is presented without an explicit argument or coverage analysis showing that these adequately sample the space of practical molecular LLM applications (e.g., multi-step synthesis planning, conformer generation, or large-molecule regimes). This directly affects the generalizability of the “no single winner” and “task-aware routing” conclusions.
minor comments (3)
  1. [Abstract] Abstract: “CML” is used without expansion on first occurrence.
  2. [Figures 4–6] Figure captions and legends: Attention and probe figures would benefit from explicit scale bars and clearer indication of which model layers are shown.
  3. [§2] Related work section: Prior molecular representation benchmarks for LLMs are cited but could more explicitly contrast the current task set with those earlier studies.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments, which help improve the rigor and clarity of our work. We address each major comment point by point below, indicating the revisions we will make to the manuscript.

read point-by-point responses

  1. Referee: [§4, Tables 1–3] §4 (Results) and Tables 1–3: Performance differences and rankings (e.g., CML best overall, structured reps dominating structural tasks) are reported without error bars, number of evaluation runs, random seeds, or statistical significance tests. This is load-bearing for the central claim that performance is “strongly representation-dependent” and that specific orderings hold.

    Authors: We agree that the lack of error bars, run counts, seeds, and statistical tests weakens the support for our central claims on representation dependence and specific rankings. In the revised manuscript we will add these elements: results will be averaged over at least three independent runs with distinct random seeds where feasible, error bars will be included in Tables 1–3, and we will report statistical significance (e.g., paired Wilcoxon tests or bootstrap confidence intervals) for key comparisons. For closed-API models we will transparently note any practical constraints on repeated sampling while still providing variance estimates where multiple queries were performed. revision: yes

  2. Referee: [§3] §3 (Benchmark Design): The selection of exactly eight tasks and nine representations is presented without an explicit argument or coverage analysis showing that these adequately sample the space of practical molecular LLM applications (e.g., multi-step synthesis planning, conformer generation, or large-molecule regimes). This directly affects the generalizability of the “no single winner” and “task-aware routing” conclusions.

    Authors: We accept that an explicit coverage argument is needed to substantiate the generalizability of the “no single winner” and task-aware routing conclusions. In the revision we will expand §3 with a new subsection that (i) justifies the eight tasks by mapping them to core molecular LLM use-cases (structure, property, retrieval, generation), (ii) explains the choice of nine representations as spanning string, structured, and nomenclature families, and (iii) provides a coverage analysis together with an explicit limitations paragraph addressing multi-step synthesis, conformer generation, and large-molecule regimes. We will also qualify how the task-routing recommendation may or may not extend to those regimes. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical benchmark with direct task measurements

full rationale

This is a pure empirical benchmark study evaluating nine representations on eight chemical tasks across 16 LLMs. No derivations, equations, fitted parameters renamed as predictions, or self-citations appear as load-bearing steps. All reported patterns (CML/MolJSON dominance on structural tasks, IUPAC on semantic tasks, etc.) are measured directly against external task performance and are falsifiable by independent replication. The central claims do not reduce to any input by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests entirely on the outcomes of the described benchmark experiments; no free parameters, axioms, or invented entities are introduced or required.

pith-pipeline@v0.9.1-grok · 5763 in / 1173 out tokens · 26537 ms · 2026-06-28T11:14:51.964090+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

89 extracted references · 19 canonical work pages · 1 internal anchor

  1. [1]

    Proceedings of the 40th International Conference on Machine Learning , pages =

    Unifying Molecular and Textual Representations via Multi-task Language Modelling , author =. Proceedings of the 40th International Conference on Machine Learning , pages =. 2023 , volume =

    2023

  2. [2]

    Nature Machine Intelligence , year=

    Leveraging large language models for predictive chemistry , author=. Nature Machine Intelligence , year=

  3. [3]

    Journal of Cheminformatics , year=

    Randomized SMILES strings improve the quality of molecular generative models , author=. Journal of Cheminformatics , year=

  4. [4]

    Nature Machine Intelligence , year=

    Invalid SMILES are beneficial rather than detrimental to chemical language models , author=. Nature Machine Intelligence , year=

  5. [5]

    and Friederich, Pascal and Gaudin, Théophile and Gayle, Alberto Alexander and Jablonka, Kevin Maik and Lameiro, Rafael F

    Krenn, Mario and Ai, Qianxiang and Barthel, Senja and Carson, Nessa and Frei, Angelo and Frey, Nathan C. and Friederich, Pascal and Gaudin, Théophile and Gayle, Alberto Alexander and Jablonka, Kevin Maik and Lameiro, Rafael F. and Lemm, Dominik and Lo, Alston and Moosavi, Seyed Mohamad and Nápoles-Duarte, José Manuel and Nigam, AkshatKumar and Pollice, Ro...

    work page doi:10.1016/j.patter.2022.100588 2022

  6. [6]

    Empowering Molecule Discovery for Molecule-Caption Translation With Large Language Models: A ChatGPT Perspective , volume=

    Li, Jiatong and Liu, Yunqing and Fan, Wenqi and Wei, Xiao-Yong and Liu, Hui and Tang, Jiliang and Li, Qing , year=. Empowering Molecule Discovery for Molecule-Caption Translation With Large Language Models: A ChatGPT Perspective , volume=. IEEE Transactions on Knowledge and Data Engineering , publisher=. doi:10.1109/tkde.2024.3393356 , number=

    work page doi:10.1109/tkde.2024.3393356 2024

  7. [7]

    What can Large Language Models do in chemistry? A comprehensive benchmark on eight tasks , url =

    Guo, Taicheng and Guo, kehan and Nan, Bozhao and Liang, Zhenwen and Guo, Zhichun and Chawla, Nitesh and Wiest, Olaf and Zhang, Xiangliang , booktitle =. What can Large Language Models do in chemistry? A comprehensive benchmark on eight tasks , url =

  8. [8]

    nach0: multimodal natural and chemical languages foundation model , volume=

    Livne, Micha and Miftahutdinov, Zulfat and Tutubalina, Elena and Kuznetsov, Maksim and Polykovskiy, Daniil and Brundyn, Annika and Jhunjhunwala, Aastha and Costa, Anthony and Aliper, Alex and Aspuru-Guzik, Alán and Zhavoronkov, Alex , year=. nach0: multimodal natural and chemical languages foundation model , volume=. Chemical Science , publisher=. doi:10....

    work page doi:10.1039/d4sc00966e

  9. [9]

    2022 , eprint=

    Galactica: A Large Language Model for Science , author=. 2022 , eprint=

    2022

  10. [10]

    The Troubling Emergence of Hallucination in Large Language Models - An Extensive Definition, Quantification, and Prescriptive Remediations

    Rawte, Vipula and Chakraborty, Swagata and Pathak, Agnibh and Sarkar, Anubhav and Tonmoy, S.M Towhidul Islam and Chadha, Aman and Sheth, Amit and Das, Amitava. The Troubling Emergence of Hallucination in Large Language Models - An Extensive Definition, Quantification, and Prescriptive Remediations. Proceedings of the 2023 Conference on Empirical Methods i...

    work page doi:10.18653/v1/2023.emnlp-main.155 2023

  11. [11]

    Translation be- tween Molecules and Natural Language

    Edwards, Carl and Lai, Tuan and Ros, Kevin and Honke, Garrett and Cho, Kyunghyun and Ji, Heng. Translation between Molecules and Natural Language. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022. doi:10.18653/v1/2022.emnlp-main.26

    work page doi:10.18653/v1/2022.emnlp-main.26 2022

  12. [12]

    SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules , author=. J. Chem. Inf. Comput. Sci. , year=

  13. [13]

    Deva , title =

    Raghunathan, Shampa and Priyakumar, U. Deva , title =. International Journal of Quantum Chemistry , volume =. doi:https://doi.org/10.1002/qua.26870 , url =. https://onlinelibrary.wiley.com/doi/pdf/10.1002/qua.26870 , abstract =

    work page doi:10.1002/qua.26870

  14. [14]

    Molecular representations in AI-driven drug discovery: a review and practical guide , volume =

    David, Laurianne and Thakkar, Amol and Mercado, Rocío and Engkvist, Ola , year =. Molecular representations in AI-driven drug discovery: a review and practical guide , volume =. Journal of Cheminformatics , doi =

  15. [15]

    and Goodman, Jonathan M

    Wigh, Daniel S. and Goodman, Jonathan M. and Lapkin, Alexei A. , title =. WIREs Computational Molecular Science , volume =. doi:https://doi.org/10.1002/wcms.1603 , url =. https://wires.onlinelibrary.wiley.com/doi/pdf/10.1002/wcms.1603 , abstract =

    work page doi:10.1002/wcms.1603

  16. [16]

    Self-referencing embedded strings (SELFIES): A 100\ volume=

    Krenn, Mario and Häse, Florian and Nigam, AkshatKumar and Friederich, Pascal and Aspuru-Guzik, Alan , year=. Self-referencing embedded strings (SELFIES): A 100\ volume=. Machine Learning: Science and Technology , publisher=. doi:10.1088/2632-2153/aba947 , number=

    work page doi:10.1088/2632-2153/aba947

  17. [17]

    Chemistry international , year=

    The IUPAC international chemical identifier: InChl-A new standard for molecular informatics , author=. Chemistry international , year=

  18. [18]

    ChemDoodle Web Components: HTML5 toolkit for chemical graphics, interfaces, and informatics , volume =

    Burger, Melanie , year =. ChemDoodle Web Components: HTML5 toolkit for chemical graphics, interfaces, and informatics , volume =. Journal of cheminformatics , doi =

  19. [19]

    2026 , eprint=

    In-Context Molecular Property Prediction with LLMs: A Blinding Study on Memorization and Knowledge Conflicts , author=. 2026 , eprint=

    2026

  20. [20]

    2025 , eprint=

    A Survey on Data Contamination for Large Language Models , author=. 2025 , eprint=

    2025

  21. [21]

    NLP evaluation in trouble: On the need to measure LLM data contamination for each benchmark

    Sainz, Oscar and Campos, Jon and Garc \'i a-Ferrero, Iker and Etxaniz, Julen and de Lacalle, Oier Lopez and Agirre, Eneko. NLP Evaluation in trouble: On the Need to Measure LLM Data Contamination for each Benchmark. Findings of the Association for Computational Linguistics: EMNLP 2023. 2023. doi:10.18653/v1/2023.findings-emnlp.722

    work page doi:10.18653/v1/2023.findings-emnlp.722 2023

  22. [22]

    Conference on Empirical Methods in Natural Language Processing , year=

    Molecular String Representation Preferences in Pretrained LLMs: A Comparative Study in Zero- & Few-Shot Molecular Property Prediction , author=. Conference on Empirical Methods in Natural Language Processing , year=

  23. [23]

    ArXiv , year=

    Benchmarking Large Language Models for Molecule Prediction Tasks , author=. ArXiv , year=

  24. [24]

    Benchmarking Large Language Models for Polymer Property Predictions , ISSN=

    Gupta, Sonakshi and Mahmood, Akhlak and Shukla, Shivank and Ramprasad, Rampi , year=. Benchmarking Large Language Models for Polymer Property Predictions , ISSN=. doi:10.1002/marc.202500388 , journal=

    work page doi:10.1002/marc.202500388

  25. [25]

    Comparing Text Representations: A Theory-Driven Approach

    Yauney, Gregory and Mimno, David. Comparing Text Representations: A Theory-Driven Approach. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 2021. doi:10.18653/v1/2021.emnlp-main.449

    work page doi:10.18653/v1/2021.emnlp-main.449 2021

  26. [26]

    2025 , eprint=

    RE-IMAGINE: Symbolic Benchmark Synthesis for Reasoning Evaluation , author=. 2025 , eprint=

    2025

  27. [27]

    2018 , eprint=

    MoleculeNet: A Benchmark for Molecular Machine Learning , author=. 2018 , eprint=

    2018

  28. [28]

    and Ahmad, Irfan

    Al-shaibani, Maged S. and Ahmad, Irfan. Consonant is all you need: a compact representation of E nglish text for efficient NLP. Findings of the Association for Computational Linguistics: EMNLP 2023. 2023. doi:10.18653/v1/2023.findings-emnlp.775

    work page doi:10.18653/v1/2023.findings-emnlp.775 2023

  29. [29]

    2026 , eprint=

    Regression with Large Language Models for Materials and Molecular Property Prediction , author=. 2026 , eprint=

    2026

  30. [30]

    ChemRxiv , year=

    DeepSMILES: An Adaptation of SMILES for Use in Machine-Learning of Chemical Structures , author=. ChemRxiv , year=

  31. [31]

    Mol-Instructions: A Large-Scale Biomolecular Instruction Dataset for Large Language Models

    Fang, Yin and Liang, Xiaozhuan and Zhang, Ningyu and Liu, Kangwei and Huang, Rui and Chen, Zhuo and Fan, Xiaohui and Chen, Huajun. Mol-Instructions: A Large-Scale Biomolecular Instruction Dataset for Large Language Models. arXiv preprint arXiv:2306.08018. 2023

    arXiv 2023

  32. [32]

    2024 , eprint=

    Are large language models superhuman chemists? , author=. 2024 , eprint=

    2024

  33. [33]

    Multi-modal Molecule Structure-text Model for Text-based Retrieval and Editing

    Liu, Shengchao and Nie, Weili and Wang, Chengpeng and Lu, Jiarui and Qiao, Zhuoran and Liu, Ling and Tang, Jian and Xiao, Chaowei and Anandkumar, Anima. Multi-modal Molecule Structure-text Model for Text-based Retrieval and Editing. Nature Machine Intelligence. 2023

    2023

  34. [34]

    GIMLET : A Unified Graph-Text Model for Instruction-Based Molecule Zero-Shot Learning

    Zhao, Haiteng and Liu, Shengchao and Ma, Chang and Xu, Hannan and Fu, Jie and Deng, Zhihong and Kong, Lingpeng and Liu, Qi. GIMLET : A Unified Graph-Text Model for Instruction-Based Molecule Zero-Shot Learning. Advances in Neural Information Processing Systems. 2023

    2023

  35. [35]

    , author=

    Nomenclature of organic chemistry: sections A, B, C, D, E, F and H - 1979 ed. , author=. 1979 , url=

    1979

  36. [36]

    2013 , publisher=

    Nomenclature of Organic Chemistry: IUPAC Recommendations and Preferred Names 2013 , author=. 2013 , publisher=

    2013

  37. [37]

    A Deep-learning System Bridging Molecule Structure and Biomedical Text with Comprehension Comparable to Human Professionals

    Zeng, Zheni and Yao, Yuan and Liu, Zhiyuan and Sun, Maosong. A Deep-learning System Bridging Molecule Structure and Biomedical Text with Comprehension Comparable to Human Professionals. Nature Communications. 2022

    2022

  38. [38]

    Nature Machine Intelligence , year=

    Chemical language models enable navigation in sparsely populated chemical space , author=. Nature Machine Intelligence , year=

  39. [39]

    2023 , eprint=

    Language models can generate molecules, materials, and protein binding sites directly in three dimensions as XYZ, CIF, and PDB files , author=. 2023 , eprint=

    2023

  40. [40]

    Aho and Jeffrey D

    Alfred V. Aho and Jeffrey D. Ullman , title =. 1972

    1972

  41. [41]

    Publications Manual , year = "1983", publisher =

    1983

  42. [42]

    , title =

    Davies, Mark and Nowotka, Michał and Papadatos, George and Dedman, Nathan and Gaulton, Anna and Atkinson, Francis and Bellis, Louisa and Overington, John P. , title =. Nucleic Acids Research , volume =. 2015 , month =. doi:10.1093/nar/gkv352 , url =

    work page doi:10.1093/nar/gkv352 2015

  43. [43]

    2018 , eprint=

    Fr\'echet ChemNet Distance: A metric for generative models for molecules in drug discovery , author=. 2018 , eprint=

    2018

  44. [44]

    Chandra and Dexter C

    Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243

    work page doi:10.1145/322234.322243 1981

  45. [45]

    Scalable training of

    Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable training of. 2007 , url=

    2007

  46. [46]

    Tetreault , title =

    Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url =

    2015

  47. [47]

    A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

    Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =. 2005 , url=

    2005

  48. [48]

    and Tukey, John W

    Cooley, James W. and Tukey, John W. , journal=. An algorithm for the machine calculation of complex. 1965 , url=

    1965

  49. [49]

    2026 , eprint=

    Molecular Representations for Large Language Models , author=. 2026 , eprint=

    2026

  50. [50]

    Table Meets

    Sui, Yuan and Zhou, Mengyu and Zhou, Mingjie and Han, Shi and Zhang, Dongmei , booktitle=. Table Meets

  51. [51]

    Tabular Representation, Noisy Operators, and Impacts on Table Structure Understanding Tasks in

    Singha, Ananya and Cambronero, Jos. Tabular Representation, Noisy Operators, and Impacts on Table Structure Understanding Tasks in. NeurIPS 2023 Second Table Representation Learning Workshop , year=

    2023

  52. [52]

    Journal of Cheminformatics , year=

    Open chemistry: RESTful web APIs, JSON, NWChem and the modern web application , author=. Journal of Cheminformatics , year=

  53. [53]

    Matthew Swain , title =

  54. [54]

    2025 , howpublished =

    2025

  55. [55]

    2025 , eprint=

    Training a Scientific Reasoning Model for Chemistry , author=. 2025 , eprint=

    2025

  56. [56]

    Yu, Zhaoning and Xu, Xiangyang and Gao, Hongyang , journal=

  57. [57]

    2025 , month =

    Anthropic , title =. 2025 , month =

    2025

  58. [58]

    2025 , month = dec, url =

    Gemini 3 Flash Model Card , institution =. 2025 , month = dec, url =

    2025

  59. [59]

    The properties of known drugs. 1. Molecular frameworks. , author=. Journal of medicinal chemistry , year=

  60. [60]

    Journal of cheminformatics , volume=

    Open Babel: An open chemical toolbox , author=. Journal of cheminformatics , volume=. 2011 , publisher=

    2011

  61. [61]

    Chemical Markup, XML, and the Worldwide Web

    Murray-Rust, Peter and Rzepa, Henry , year =. Chemical Markup, XML, and the Worldwide Web. 1. Basic Principles , volume =. Journal of Chemical Information and Computer Sciences , doi =

  62. [62]

    Chemical markup, XML, and the world wide web

    Murray-Rust, Peter and Rzepa, Henry , year =. Chemical markup, XML, and the world wide web. 4. CML schema , volume =. Journal of chemical information and computer sciences , doi =

  63. [63]

    Nov , volume=

    IBM internal report , author=. Nov , volume=

  64. [64]

    Exploring Forgetting in Large Language Model Pre-Training

    Liao, Chonghua and Xie, Ruobing and Sun, Xingwu and Sun, Haowen and Kang, Zhanhui. Exploring Forgetting in Large Language Model Pre-Training. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025. doi:10.18653/v1/2025.acl-long.105

    work page doi:10.18653/v1/2025.acl-long.105 2025

  65. [65]

    Belinkov

    Belinkov, Yonatan , title =. Computational Linguistics , volume =. 2022 , month =. doi:10.1162/coli_a_00422 , url =

    work page internal anchor Pith review doi:10.1162/coli_a_00422 2022

  66. [66]

    Insights into LLM Long-Context Failures: When Transformers Know but Don ' t Tell

    Gao, Muhan and Lu, TaiMing and Yu, Kuai and Byerly, Adam and Khashabi, Daniel. Insights into LLM Long-Context Failures: When Transformers Know but Don ' t Tell. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024. doi:10.18653/v1/2024.findings-emnlp.447

    work page doi:10.18653/v1/2024.findings-emnlp.447 2024

  67. [67]

    A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., Hassabis, D., Clopath, C., Kumaran, D., and Hadsell, R

    Kirkpatrick, James and Pascanu, Razvan and Rabinowitz, Neil and Veness, Joel and Desjardins, Guillaume and Rusu, Andrei A. and Milan, Kieran and Quan, John and Ramalho, Tiago and Grabska-Barwinska, Agnieszka and Hassabis, Demis and Clopath, Claudia and Kumaran, Dharshan and Hadsell, Raia , year=. Overcoming catastrophic forgetting in neural networks , vol...

    work page doi:10.1073/pnas.1611835114

  68. [68]

    Journal of chemical information and computer sciences , year=

    Reoptimization of MDL Keys for Use in Drug Discovery , author=. Journal of chemical information and computer sciences , year=

  69. [69]

    Journal of chemical information and modeling , year=

    Extended-Connectivity Fingerprints , author=. Journal of chemical information and modeling , year=

  70. [70]

    Statistical Significance Tests for Machine Translation Evaluation

    Koehn, Philipp. Statistical Significance Tests for Machine Translation Evaluation. Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing. 2004

    2004

  71. [71]

    2024 , eprint=

    LlaSMol: Advancing Large Language Models for Chemistry with a Large-Scale, Comprehensive, High-Quality Instruction Tuning Dataset , author=. 2024 , eprint=

    2024

  72. [72]

    RDKit: Open-Source Cheminformatics Software , url =

    Landrum, Greg , biburl =. RDKit: Open-Source Cheminformatics Software , url =

  73. [73]

    Chemical Name to Structure: OPSIN, an Open Source Solution , volume =

    Lowe, Daniel and Corbett, Peter and Murray-Rust, Peter and Glen, Robert , year =. Chemical Name to Structure: OPSIN, an Open Source Solution , volume =. Journal of chemical information and modeling , doi =

  74. [74]

    Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year=

    Efficient Memory Management for Large Language Model Serving with PagedAttention , author=. Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year=

  75. [75]

    arXiv preprint arXiv:2505.09388 , year =

    Qwen3 Technical Report , author =. arXiv preprint arXiv:2505.09388 , year =

    Pith/arXiv arXiv

  76. [76]

    arXiv preprint arXiv:2412.08905 , year =

    Abdin, Marah and Aneja, Jyoti and Behl, Harkirat and Bubeck, S. arXiv preprint arXiv:2412.08905 , year =

    Pith/arXiv arXiv

  77. [77]

    arXiv preprint arXiv:2504.21318 , year =

    Abdin, Marah and Agarwal, Sahaj and Awadallah, Ahmed and Balachandran, Vidhisha and Behl, Harkirat and Chen, Lingjiao and de Rosa, Gustavo and Gunasekar, Suriya and Javaheripi, Mojan and Joshi, Neel and Kauffmann, Piero and Lara, Yash and Mendes, Caio C. arXiv preprint arXiv:2504.21318 , year =

    Pith/arXiv arXiv

  78. [78]

    2026 , eprint=

    Olmo 3 , author=. 2026 , eprint=

    2026

  79. [79]

    Developing

    Zhao, Zihan and Ma, Da and Chen, Lu and Sun, Liangtai and Li, Zihao and Xia, Yi and Chen, Bo and Xu, Hongshen and Zhu, Zichen and Zhu, Su and Fan, Shuai and Shen, Guodong and Yu, Kai and Chen, Xin , journal=. Developing. 2024 , note =

    2024

  80. [80]

    Zhao, Zihan and others , journal=

Showing first 80 references.