RxnOptBench: Benchmarking LLMs for Reaction-Condition Optimization in Organic Methodology
Abstract
Chemical reaction-condition optimization -- choosing the catalyst, ligand, solvent, reagent, temperature, time, and atmosphere that jointly maximize yield and stereoselectivity -- is a central, judgement-laden subtask of organic methodology research that large language models are increasingly expected to support. Yet existing chemistry benchmarks evaluate reaction-class labelling, retrosynthesis, or SMILES manipulation, and do not ask models to read a real condition-screening table and pick the best set. We introduce RxnOptBench, a benchmark whose every option and precedent is a real wet-lab entry mined from the optimization tables of organic-methodology papers published in 2025, graded by a continuous relative score derived from a declared headline utility that combines reported yield with enantiomeric excess (ee), diastereomeric ratio (dr), and regioisomeric ratio (rr), and equipped with a paired precedents-vs-no-precedents design that isolates in-context use of literature evidence from parametric memorization. Across nine frontier LLMs and three Chemistry LLMs, even the best models leave substantial headroom: chemistry-specialized models fall to the random-baseline floor on multi-axis selection, while open-weight models have closed most of the gap to proprietary frontier models. We release the final human-reviewed benchmark test set and evaluation code.
Related
- Reaction-Level Machine Learning for Condition Optimization in Suzuki–Miyaura Cross-Coupling
- A practical one-electrode electrochemical methodology for tracking catalyst reconstruction in NiFe-Prussian blue analogues for the oxygen evolution reaction.
- Two-Stage Optimization of Oleic Acid Esterification via Response Surface Methodology and Green Chemistry Metrics Selection for Biodiesel from WCO
- OPRD-100: An Expert-Curated Dataset of Organic Reactions for Benchmarking Reaction Data Mining Tools
- RxnScout: Deploying a Vision-Language Agent for Nickel-Catalyzed C-Glycosylation Condition Optimization
BibTeX
@misc{ge2026rxnoptbench,
title = {{RxnOptBench: Benchmarking LLMs for Reaction-Condition Optimization in Organic Methodology}},
author = {Lingli Ge and Yubin Wang and Junyuan Gao and Jiahe Song and Jiaxing Sun and Boyu Zhu and Haote Yang and Jingchao Wang and Lixin Ma and Jiang Wu and Yuqiang Li and Conghui He},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.02242},
eprint = {2610.02242},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.02242},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - RxnOptBench: Benchmarking LLMs for Reaction-Condition Optimization in Organic Methodology
AU - Lingli Ge
AU - Yubin Wang
AU - Junyuan Gao
AU - Jiahe Song
AU - Jiaxing Sun
AU - Boyu Zhu
AU - Haote Yang
AU - Jingchao Wang
AU - Lixin Ma
AU - Jiang Wu
AU - Yuqiang Li
AU - Conghui He
PY - 2026
DA - 2026/10/01
PB - arXiv
DO - 10.48550/arxiv.2610.02242
UR - https://arxiv.org/abs/2610.02242
AB - Chemical reaction-condition optimization -- choosing the catalyst, ligand, solvent, reagent, temperature, time, and atmosphere that jointly maximize yield and stereoselectivity -- is a central, judgement-laden subtask of organic methodology research that large language models are increasingly expected to support. Yet existing chemistry benchmarks evaluate reaction-class labelling, retrosynthesis, or SMILES manipulation, and do not ask models to read a real condition-screening table and pick the best set. We introduce RxnOptBench, a benchmark whose every option and precedent is a real wet-lab entry mined from the optimization tables of organic-methodology papers published in 2025, graded by a continuous relative score derived from a declared headline utility that combines reported yield with enantiomeric excess (ee), diastereomeric ratio (dr), and regioisomeric ratio (rr), and equipped with a paired precedents-vs-no-precedents design that isolates in-context use of literature evidence from parametric memorization. Across nine frontier LLMs and three Chemistry LLMs, even the best models leave substantial headroom: chemistry-specialized models fall to the random-baseline floor on multi-axis selection, while open-weight models have closed most of the gap to proprietary frontier models. We release the final human-reviewed benchmark test set and evaluation code.
ER -