PolyReader: an agentic LLM pipeline for extracting structured data from unstructured polymer scientific literature
6 affiliations
- 1Department of Chemistry and Materials Science
- 2Xi’an Jiaotong-Liverpool University
- 3Department of Chemistry
- 4University of Liverpool
- 5College of Carbon Neutrality Future Technology, State Key Laboratory of Heavy Oil Processing
- 6China University of Petroleum (Beijing
Abstract
Machine learning for polymer design requires structured, sample-level records linking polymer identity, synthesis, composition, molecular-weight characteristics, processing conditions, and measured properties. Yet decades of polymer data remain dispersed across the texts, tables, figures, and supporting information of the literature, where individual values are often separated from the samples and experimental context they describe. Here we introduce PolyReader, an LLM agent for provenance-aware extraction of polymer data from scientific papers. Rather than treating a paper as a single block of text, PolyReader represents the manuscript and supporting information as navigable hierarchical section trees. The agent surveys this structure, selectively reads relevant pages, follows cross-references between documents, and integrates dispersed evidence into schema-validated, per-sample records. The resulting extracted fields are thus linked to its source document and page. On a benchmark of 42 polymer papers evaluated by trained annotators, PolyReader achieved F1 scores of 0.92–0.96 across 11 property categories, with comparable performance across three LLM models. Holding the LLM model, parsed document representation, and output schema fixed, PolyReader improved F1 by 13–17 points over single-pass prompting. These results show that reliable polymer-data extraction depends not only on recognizing individual values, but on reconstructing the sample-level relationships distributed throughout a paper. PolyReader produces records that are directly auditable against the primary literature, providing a scalable foundation for constructing traceable polymer datasets for machine learning. The extraction pipeline and benchmark annotations are released for reuse in the community.
Related
- From unstructured to structured: opportunities, risks and engineering practices for LLM-based data extraction in chemistry and materials
- Neural Constrained Charge Equilibration for Extracting Charge Transport Parameters from Machine Learned Interatomic Potentials
- vibe-qc: expert-directed agentic engineering for scientific software development
- Scientific Symbiosis: Bridging Expertise and Fostering Effective Collaborations with Mass Spectrometrists
- Agentic AI for Scientific Reasoning in Autonomous Quantum Sensing Experiments
BibTeX
@misc{shi2026polyreader,
title = {{PolyReader: an agentic LLM pipeline for extracting structured data from unstructured polymer scientific literature}},
author = {Yuwei Shi and Xinyan Zhong and Yuqi Wei and Heechae Choi and Tianhang Zhou and Zhenghao Wu},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008614/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008614/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - PolyReader: an agentic LLM pipeline for extracting structured data from unstructured polymer scientific literature
AU - Yuwei Shi
AU - Xinyan Zhong
AU - Yuqi Wei
AU - Heechae Choi
AU - Tianhang Zhou
AU - Zhenghao Wu
PY - 2026
DA - 2026/09/10
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008614/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008614/v1
AB - Machine learning for polymer design requires structured, sample-level records linking polymer identity, synthesis, composition, molecular-weight characteristics, processing conditions, and measured properties. Yet decades of polymer data remain dispersed across the texts, tables, figures, and supporting information of the literature, where individual values are often separated from the samples and experimental context they describe. Here we introduce PolyReader, an LLM agent for provenance-aware extraction of polymer data from scientific papers. Rather than treating a paper as a single block of text, PolyReader represents the manuscript and supporting information as navigable hierarchical section trees. The agent surveys this structure, selectively reads relevant pages, follows cross-references between documents, and integrates dispersed evidence into schema-validated, per-sample records. The resulting extracted fields are thus linked to its source document and page. On a benchmark of 42 polymer papers evaluated by trained annotators, PolyReader achieved F1 scores of 0.92–0.96 across 11 property categories, with comparable performance across three LLM models. Holding the LLM model, parsed document representation, and output schema fixed, PolyReader improved F1 by 13–17 points over single-pass prompting. These results show that reliable polymer-data extraction depends not only on recognizing individual values, but on reconstructing the sample-level relationships distributed throughout a paper. PolyReader produces records that are directly auditable against the primary literature, providing a scalable foundation for constructing traceable polymer datasets for machine learning. The extraction pipeline and benchmark annotations are released for reuse in the community.
ER -