arxiv: 2605.10230 · v1 · submitted 2026-05-11 · 💻 cs.LG

Recognition: no theorem link

FORGE: Fragment-Oriented Ranking and Generation for Context-Aware Molecular Optimization

Qingchuan Zhang , He Cao , Hao Li , Yanjun Shao , Zhiyuan Liu , Shihang Wang , Shufang Xie , Shenghua Gao

show 1 more author

Xinwu Ye

Authors on Pith no claims yet

Pith reviewed 2026-05-12 05:21 UTC · model grok-4.3

classification 💻 cs.LG

keywords molecular optimizationfragment editingcontext-aware generationlanguage modelschemical prioredit pair miningblack-box optimizationlocal structural edits

0 comments

The pith

FORGE reformulates molecular optimization as context-aware fragment editing with mined pairs, letting a 0.6B model outperform larger language models and graph methods.

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

The paper claims that treating molecular optimization as prompt-based sequence generation creates scaling limits, hallucinations, and misses how fragment changes depend on the surrounding molecule. FORGE instead mines verified low-to-high edit pairs automatically and uses them in a two-stage process on a compact language model. The first stage ranks candidate fragments according to the property gain they would produce inside the full current molecule. The second stage then produces the actual replacement. If this holds, molecular improvement becomes more reliable and scalable without relying on human text descriptions or ever-larger models.

Core claim

FORGE is a two-stage framework that reformulates molecular optimization as context-aware local editing. Stage 1 ranks candidate fragments by their property contribution under the full molecular context using automatically mined and verified low-to-high edit pairs; Stage 2 generates the explicit fragment replacements. Built on a 0.6B language model that adapts to unseen black-box objectives through in-context demonstrations, the method outperforms prior approaches including substantially larger language models and graph methods on Prompt-MolOpt, PMO-1k, and ChemCoTBench.

What carries the argument

Two-stage ranking-then-generation on mined low-to-high edit pairs that evaluates each fragment's effect inside the complete molecule rather than through language prompts.

If this is right

Molecular optimization can proceed without expensive human text annotations while still preserving structural similarity to the starting compound.
A compact model suffices for black-box objectives once fragment contributions are ranked in full context.
Explicit fragment replacements reduce chemical hallucinations compared with free-form sequence generation.
Performance gains on standard benchmarks arise from local edits whose effects are measured against the surrounding molecule rather than global text conditioning.

Where Pith is reading between the lines

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

The same mining and ranking approach could be tested on other structured objects such as proteins or materials where local edits must be evaluated inside global context.
If fragment effects prove strongly context-dependent, the ranked candidates themselves might surface previously hidden structure-property patterns for human inspection.
Extending the edit-pair mining to multi-step trajectories could turn the method into a general planner for longer optimization sequences.

Load-bearing premise

Automatically mined and verified low-to-high edit pairs supply a sufficient, unbiased chemical prior that captures the strong context dependence of fragment effects better than natural-language supervision.

What would settle it

A new benchmark or test set of molecules where FORGE produces more invalid structures or lower property gains than larger language-model baselines would falsify the claim that the mined pairs provide a superior prior.

Figures

Figures reproduced from arXiv: 2605.10230 by Hao Li, He Cao, Qingchuan Zhang, Shenghua Gao, Shihang Wang, Shufang Xie, Xinwu Ye, Yanjun Shao, Zhiyuan Liu.

**Figure 1.** Figure 1: Why prompt-only generation is insufficient for molecular optimization. (a) Grouping fragment attributions by ECFP neighborhoods reduces attribution variance across two distinct property predictors, indicating that fragment effects depend on chemical context. (b) The same fragment can have different effects across host molecules. (c) On PMO, replacing the true target property in the prompt with an unrelat… view at source ↗

**Figure 2.** Figure 2: Overview of the FORGE framework. (a) Left: Empirical observations reveal that fragment contributions are highly context-dependent. (b) Middle: Natural language is a suboptimal interface for optimization due to weak prompt control, lack of oracle semantics, and distillation hallucinations. (c) Right: The FORGE pipeline decomposes the task into context-aware fragment ranking (Stage 1) and explicit modificati… view at source ↗

**Figure 3.** Figure 3: Marginal gains of Stage 1 and Stage 2 across representative benchmarks. Stage 2 is the larger increment throughout. This is expected: Stage 2 supervision is in the same output space used at inference, so the gradient signal directly trains the inference behavior. The increment is largest on the real-target tasks in ChemCoTBench (JNK3 SR: 67 → 83; Solubility: 82 → 95). Stage 1 adds a smaller but consistent… view at source ↗

**Figure 4.** Figure 4: Fragment decoupling of ChemBERTa on lipop B. More experiment details [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗

read the original abstract

Molecular optimization seeks to improve a molecule through small structural edits while preserving similarity to the starting compound. Recent language-model approaches typically treat this task as prompt-conditioned sequence generation. However, relying on natural language introduces an inherent data-scaling bottleneck, often leads to chemical hallucinations, and ignores the strong context dependence of fragment effects. We present FORGE, a two-stage framework that reformulates molecular optimization as context-aware local editing. By utilizing automatically mined, verified low-to-high edit pairs instead of expensive human text annotations, Stage 1 ranks candidate fragments by their property contribution under the full molecular context to inject chemical prior, and Stage 2 generates explicit fragment replacements. Built on a compact 0.6B language model, FORGE further adapts to unseen black-box objectives through in-context demonstrations. Across Prompt-MolOpt, PMO-1k and ChemCoTBench, FORGE consistently outperforms prior methods, including substantially larger language models and graph methods. These results highlight the value of explicit fragment-level supervision as a more easily obtainable, scalable, and hallucination-less alternative to natural language training.

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.

Desk Editor's Note private letter to a colleague

FORGE's two-stage fragment ranking plus generation on mined edit pairs is a clean technical shift from prompt-based LMs, but the outperformance claims stand or fall on whether the mining process introduces bias.

read the letter

The main point is that FORGE replaces text-prompt training with a two-stage process: rank candidate fragments by their measured property contribution inside the full molecular context, then generate the replacement. It trains on automatically mined low-to-high edit pairs instead of human annotations and uses a 0.6B model that adapts to new objectives via in-context examples. That setup directly targets the context dependence of fragments and the data-scaling problem the abstract flags in prior LM work.

Referee Report

2 major / 2 minor

Summary. The paper introduces FORGE, a two-stage framework for molecular optimization that reformulates the task as context-aware local editing. Stage 1 ranks candidate fragments by their property contribution under full molecular context using automatically mined and verified low-to-high edit pairs; Stage 2 generates explicit fragment replacements. The approach is built on a compact 0.6B language model that adapts to unseen black-box objectives via in-context demonstrations, and it reports consistent outperformance over prior methods (including substantially larger LMs and graph methods) on the Prompt-MolOpt, PMO-1k, and ChemCoTBench benchmarks.

Significance. If the empirical results prove robust, the work demonstrates that explicit fragment-level supervision derived from mined edit pairs can serve as a scalable, chemically grounded, and hallucination-resistant alternative to natural-language supervision for molecular optimization, potentially lowering data requirements while preserving context dependence of fragment effects.

major comments (2)

[Methods (likely §3–4)] The central claim that automatically mined low-to-high edit pairs supply an unbiased chemical prior capturing strong context dependence of fragment effects (stronger than natural-language supervision) rests on the mining/verification pipeline. The methods section provides only a high-level description of this pipeline and does not specify the mining algorithm, the verification oracle, or controls against selection bias or leakage from the property predictors later used in benchmarking; without these details the reported gains on Prompt-MolOpt, PMO-1k and ChemCoTBench cannot be evaluated for generalizability.
[Experimental results (likely §5)] The experimental results section claims consistent outperformance but supplies no information on data splits, baseline re-implementations, statistical significance tests, or whether the same property predictors appear in both the mining stage and the evaluation; these omissions are load-bearing for the claim that FORGE outperforms larger LMs and graph methods.

minor comments (2)

[Abstract] The abstract refers to “substantially larger language models” without naming the models or reporting parameter counts; this comparison should be made explicit.
[Introduction and Methods] Notation for fragment ranking scores and in-context demonstration formatting is introduced without a dedicated notation table or early definition, which reduces readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. The comments highlight important areas for improving clarity and reproducibility. We will revise the manuscript to provide the requested details on the mining pipeline and experimental setup while preserving the core contributions.

read point-by-point responses

Referee: [Methods (likely §3–4)] The central claim that automatically mined low-to-high edit pairs supply an unbiased chemical prior capturing strong context dependence of fragment effects (stronger than natural-language supervision) rests on the mining/verification pipeline. The methods section provides only a high-level description of this pipeline and does not specify the mining algorithm, the verification oracle, or controls against selection bias or leakage from the property predictors later used in benchmarking; without these details the reported gains on Prompt-MolOpt, PMO-1k and ChemCoTBench cannot be evaluated for generalizability.

Authors: We agree that the original submission described the mining/verification pipeline at a high level. In the revised manuscript we will expand Section 3 with the exact mining algorithm (including fragmentation rules, property-delta thresholds, and pair extraction logic), the verification oracle (chemical validity filters plus property-improvement checks), and explicit controls for selection bias and leakage. We will also document the data partitioning that keeps mining data disjoint from the evaluation benchmarks and confirm that property predictors used during mining are independent of those used in benchmarking. revision: yes
Referee: [Experimental results (likely §5)] The experimental results section claims consistent outperformance but supplies no information on data splits, baseline re-implementations, statistical significance tests, or whether the same property predictors appear in both the mining stage and the evaluation; these omissions are load-bearing for the claim that FORGE outperforms larger LMs and graph methods.

Authors: We acknowledge these omissions in the experimental reporting. The revised Section 5 will specify the data splits for each benchmark, provide implementation details and hyperparameter settings for all re-implemented baselines, report statistical significance tests (e.g., paired t-tests or Wilcoxon signed-rank tests with p-values), and explicitly state that the property predictors employed in the mining stage are distinct from those used in evaluation, thereby eliminating leakage. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical benchmark claims independent of self-referential steps

full rationale

The paper describes a two-stage framework (ranking then generation) that relies on automatically mined edit pairs as chemical prior and evaluates via external benchmarks (Prompt-MolOpt, PMO-1k, ChemCoTBench). No equations, derivations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided text. The central outperformance claim rests on comparative results rather than any reduction of outputs to inputs by construction. The mining/verification process is noted as a potential empirical weakness but does not constitute circularity under the defined patterns, as no self-definition or ansatz smuggling is exhibited.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Based on abstract only; the central claim rests on the domain assumption that fragment effects are strongly context-dependent and that automatically mined pairs can be verified and used as reliable supervision without introducing selection bias.

axioms (2)

domain assumption Fragment effects on molecular properties are strongly context-dependent
Invoked to justify moving away from natural-language prompting.
domain assumption Automatically mined low-to-high edit pairs can be verified and provide unbiased chemical prior
Stated as the data source replacing human text annotations.

pith-pipeline@v0.9.0 · 5518 in / 1297 out tokens · 43576 ms · 2026-05-12T05:21:15.195358+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

58 extracted references · 58 canonical work pages · 3 internal anchors

[1]

Advances and 42 challenges in deep generative models for de novo molecule generation.Wiley Interdis- ciplinary Reviews: Computational Molecular Science2019,9, e1395

Genmol: A drug discovery generalist with discrete diffusion , author=. arXiv preprint arXiv:2501.06158 , year=

work page arXiv
[2]

Advances in Neural Information Processing Systems , volume=

Molecule generation with fragment retrieval augmentation , author=. Advances in Neural Information Processing Systems , volume=

work page
[3]

Nature communications , volume=

Chemistry-intuitive explanation of graph neural networks for molecular property prediction with substructure masking , author=. Nature communications , volume=. 2023 , publisher=

work page 2023
[4]

Nature Machine Intelligence , volume=

Leveraging language model for advanced multiproperty molecular optimization via prompt engineering , author=. Nature Machine Intelligence , volume=. 2024 , publisher=

work page 2024
[5]

Advances in neural information processing systems , volume=

Sample efficiency matters: a benchmark for practical molecular optimization , author=. Advances in neural information processing systems , volume=

work page
[6]

Beyond chemical qa: Evaluating llm’s chemical reasoning with modular chemical operations.arXiv preprint arXiv:2505.21318, 2025

Beyond Chemical QA: Evaluating LLM's Chemical Reasoning with Modular Chemical Operations , author=. arXiv preprint arXiv:2505.21318 , year=

work page arXiv
[7]

Lico: Large language models for in-context molecular optimization.arXiv preprint arXiv:2406.18851, 2024

Lico: Large language models for in-context molecular optimization , author=. arXiv preprint arXiv:2406.18851 , year=

work page arXiv
[8]

arXiv preprint arXiv:2602.07075 , year=

Latentchem: From textual cot to latent thinking in chemical reasoning , author=. arXiv preprint arXiv:2602.07075 , year=

work page arXiv
[9]

Qwen3 Technical Report

Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[10]

mmpdb: An Open-Source Matched Molecular Pair Platform for Large Multiproperty Data Sets , journal =

Dalke, Andrew and Hert, J. mmpdb: An Open-Source Matched Molecular Pair Platform for Large Multiproperty Data Sets , journal =. 2018 , doi =

work page 2018
[11]

and Bento, A

Gaulton, Anna and Bellis, Louisa J. and Bento, A. Patricia and Chambers, Jon and Davies, Mark and Hersey, Anne and Light, Yvonne and McGlinchey, Shaun and Michalovich, David and Al-Lazikani, Bissan and Overington, John P. , title =. Nucleic Acids Research , volume =. 2012 , month =. doi:10.1093/nar/gkr777 , url =

work page doi:10.1093/nar/gkr777 2012
[12]

arXiv preprint arXiv:2510.08744 , year=

Graph diffusion transformers are in-context molecular designers , author=. arXiv preprint arXiv:2510.08744 , year=

work page arXiv
[13]

Nature machine intelligence , volume=

A deep generative model for molecule optimization via one fragment modification , author=. Nature machine intelligence , volume=. 2021 , publisher=

work page 2021
[14]

Journal of chemical information and modeling , volume=

Exposing the limitations of molecular machine learning with activity cliffs , author=. Journal of chemical information and modeling , volume=. 2022 , publisher=

work page 2022
[15]

International conference on machine learning , pages=

Junction tree variational autoencoder for molecular graph generation , author=. International conference on machine learning , pages=. 2018 , organization=

work page 2018
[16]

Advances in neural information processing systems , volume=

Graph convolutional policy network for goal-directed molecular graph generation , author=. Advances in neural information processing systems , volume=

work page
[17]

Artificial intelligence review , volume=

Machine learning in drug discovery: a review , author=. Artificial intelligence review , volume=. 2022 , publisher=

work page 2022
[18]

IEEE Transactions on Neural Networks and Learning Systems , volume=

Graph polish: A novel graph generation paradigm for molecular optimization , author=. IEEE Transactions on Neural Networks and Learning Systems , volume=. 2021 , publisher=

work page 2021
[19]

ChemRxiv , volume =

Jannik Philipp Roth , title =. ChemRxiv , volume =. 2026 , doi =. https://chemrxiv.org/doi/pdf/10.26434/chemrxiv.15002302/v1 , abstract =

work page doi:10.26434/chemrxiv.15002302/v1 2026
[20]

arXiv preprint arXiv:2505.22252 , year=

B-XAIC Dataset: Benchmarking Explainable AI for Graph Neural Networks Using Chemical Data , author=. arXiv preprint arXiv:2505.22252 , year=

work page arXiv
[21]

arXiv preprint arXiv:2508.15015 , year=

Fragment-Wise Interpretability in Graph Neural Networks via Molecule Decomposition and Contribution Analysis , author=. arXiv preprint arXiv:2508.15015 , year=

work page arXiv
[22]

Seyone Chithrananda, Gabriel Grand, Bharath Ramsun- dar, et al

ChemBERTa: large-scale self-supervised pretraining for molecular property prediction , author=. arXiv preprint arXiv:2010.09885 , year=

work page arXiv 2010
[23]

Digital Discovery , volume=

Gp-molformer: A foundation model for molecular generation , author=. Digital Discovery , volume=. 2025 , publisher=

work page 2025
[24]

arXiv preprint arXiv:2601.15279 , year=

MolecularIQ: Characterizing Chemical Reasoning Capabilities Through Symbolic Verification on Molecular Graphs , author=. arXiv preprint arXiv:2601.15279 , year=

work page arXiv
[25]

arXiv preprint arXiv:2407.18897 , year=

Small molecule optimization with large language models , author=. arXiv preprint arXiv:2407.18897 , year=

work page arXiv
[26]

Communications Chemistry , year=

ChemFM as a scaling law guided foundation model pre-trained on informative chemicals , author=. Communications Chemistry , year=

work page
[27]

arXiv preprint arXiv:2402.09391 (2024)

Llasmol: Advancing large language models for chemistry with a large-scale, comprehensive, high-quality instruction tuning dataset , author=. arXiv preprint arXiv:2402.09391 , year=

work page arXiv
[28]

TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second

Tabpfn: A transformer that solves small tabular classification problems in a second , author=. arXiv preprint arXiv:2207.01848 , year=

work page internal anchor Pith review arXiv
[29]

TabICL: A tabular foundation model for in-context learning on large data.arXiv preprint arXiv:2502.05564, 2025

Tabicl: A tabular foundation model for in-context learning on large data , author=. arXiv preprint arXiv:2502.05564 , year=

work page arXiv
[30]

Journal of Chemical Information and Modeling , volume=

Test-time training scaling laws for chemical exploration in drug design , author=. Journal of Chemical Information and Modeling , volume=. 2025 , publisher=

work page 2025
[31]

BOOM: Benchmarking out-of-distribution molecular property predictions of machine learning models.arXiv preprint arXiv:2505.01912, 2025

Boom: benchmarking out-of-distribution molecular property predictions of machine learning models , author=. arXiv preprint arXiv:2505.01912 , year=

work page arXiv
[32]

Forty-second International Conference on Machine Learning , year=

Drug-tta: Test-time adaptation for drug virtual screening via multi-task meta-auxiliary learning , author=. Forty-second International Conference on Machine Learning , year=

work page
[33]

GPT-4o System Card

Gpt-4o system card , author=. arXiv preprint arXiv:2410.21276 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[34]

The Eleventh International Conference on Learning Representations , year=

Uni-Mol: A Universal 3D Molecular Representation Learning Framework , author=. The Eleventh International Conference on Learning Representations , year=

work page
[35]

Chemical science , volume=

MoleculeNet: a benchmark for molecular machine learning , author=. Chemical science , volume=. 2018 , publisher=

work page 2018
[36]

International conference on machine learning , pages=

Hierarchical generation of molecular graphs using structural motifs , author=. International conference on machine learning , pages=. 2020 , organization=

work page 2020
[37]

arXiv preprint arXiv:2508.13408 , year=

NovoMolGen: Rethinking Molecular Language Model Pretraining , author=. arXiv preprint arXiv:2508.13408 , year=

work page arXiv
[38]

arXiv e-prints , keywords =

Molecular De Novo Design through Deep Reinforcement Learning. arXiv e-prints , keywords =. doi:10.48550/arXiv.1704.07555 , archivePrefix =. 1704.07555 , primaryClass =

work page doi:10.48550/arxiv.1704.07555
[39]

Advances in Neural Information Processing Systems , volume=

Genetic-guided GFlowNets for sample efficient molecular optimization , author=. Advances in Neural Information Processing Systems , volume=

work page
[40]

and Irwin, J

Sterling, Teague and Irwin, John J. , title =. Journal of Chemical Information and Modeling , year =. doi:10.1021/acs.jcim.5b00559 , url =

work page doi:10.1021/acs.jcim.5b00559
[41]

, title =

Jensen, Jan H. , title =. Chemical Science , year =. doi:10.1039/c8sc05372c , publisher =

work page doi:10.1039/c8sc05372c
[42]

Gaussian process optimization in the bandit setting: No regret and experimental design,

Gaussian Process Optimization in the Bandit Setting: No Regret and Experimental Design. arXiv e-prints , keywords =. doi:10.48550/arXiv.0912.3995 , archivePrefix =. 0912.3995 , primaryClass =

work page doi:10.48550/arxiv.0912.3995
[43]

Chemllm: A chemical large language model.arXiv preprint arXiv:2402.06852, 2024

ChemLLM: A Chemical Large Language Model. arXiv e-prints , keywords =. doi:10.48550/arXiv.2402.06852 , archivePrefix =. 2402.06852 , primaryClass =

work page doi:10.48550/arxiv.2402.06852
[44]

Training a scientific reasoning model for chemistry.arXiv preprint arXiv:2506.17238, 2025

Training a Scientific Reasoning Model for Chemistry. arXiv e-prints , keywords =. doi:10.48550/arXiv.2506.17238 , archivePrefix =. 2506.17238 , primaryClass =

work page doi:10.48550/arxiv.2506.17238
[45]

Efficient evolutionary search over chemical space with large language models, 2025

Efficient Evolutionary Search Over Chemical Space with Large Language Models. arXiv e-prints , keywords =. doi:10.48550/arXiv.2406.16976 , archivePrefix =. 2406.16976 , primaryClass =

work page doi:10.48550/arxiv.2406.16976
[46]

arXiv e-prints , keywords =

Augmented Memory: Capitalizing on Experience Replay to Accelerate De Novo Molecular Design. arXiv e-prints , keywords =. doi:10.48550/arXiv.2305.16160 , archivePrefix =. 2305.16160 , primaryClass =

work page doi:10.48550/arxiv.2305.16160
[47]

Mol-instructions: A large-scale biomolecular instruction dataset for large language models.arXiv preprint arXiv:2306.08018, 2023

Mol-Instructions: A Large-Scale Biomolecular Instruction Dataset for Large Language Models. arXiv e-prints , keywords =. doi:10.48550/arXiv.2306.08018 , archivePrefix =. 2306.08018 , primaryClass =

work page doi:10.48550/arxiv.2306.08018
[48]

arXiv preprint arXiv:1812.01070 , year=

Learning multimodal graph-to-graph translation for molecular optimization , author=. arXiv preprint arXiv:1812.01070 , year=

work page arXiv
[49]

arXiv preprint arXiv:1907.11223 , year=

Hierarchical graph-to-graph translation for molecules , author=. arXiv preprint arXiv:1907.11223 , year=

work page arXiv 1907
[50]

Journal of chemical information and modeling , volume=

GuacaMol: benchmarking models for de novo molecular design , author=. Journal of chemical information and modeling , volume=. 2019 , publisher=

work page 2019
[51]

Journal of the American Chemical Society , volume=

Deep lead optimization: leveraging generative AI for structural modification , author=. Journal of the American Chemical Society , volume=. 2024 , publisher=

work page 2024
[52]

ACS omega , volume=

AI-driven drug discovery: a comprehensive review , author=. ACS omega , volume=. 2025 , publisher=

work page 2025
[53]

Journal of Cheminformatics , volume=

Reinvent 4: Modern AI--driven generative molecule design , author=. Journal of Cheminformatics , volume=. 2024 , publisher=

work page 2024
[54]

Journal of Cheminformatics , volume=

kGCN: a graph-based deep learning framework for chemical structures , author=. Journal of Cheminformatics , volume=. 2020 , publisher=

work page 2020
[55]

Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing

BioT5: Enriching Cross-modal Integration in Biology with Chemical Knowledge and Natural Language Associations , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023

work page 2023
[56]

Feinberg, Joseph Gomes, Caleb Geniesse, Aneesh S

MoleculeNet: A Benchmark for Molecular Machine Learning. arXiv e-prints , keywords =. doi:10.48550/arXiv.1703.00564 , archivePrefix =. 1703.00564 , primaryClass =

work page doi:10.48550/arxiv.1703.00564
[57]

arXiv e-prints , keywords =

Retrieval-based Controllable Molecule Generation. arXiv e-prints , keywords =. doi:10.48550/arXiv.2208.11126 , archivePrefix =. 2208.11126 , primaryClass =

work page doi:10.48550/arxiv.2208.11126
[58]

arXiv e-prints , keywords =

Exploring Chemical Space with Score-based Out-of-distribution Generation. arXiv e-prints , keywords =. doi:10.48550/arXiv.2206.07632 , archivePrefix =. 2206.07632 , primaryClass =

work page doi:10.48550/arxiv.2206.07632