Structured information extraction from geoscientific literature with GeoExtractor

Earth science research increasingly depends on the integration of large volumes of heterogeneous data distributed across publications. A substantial amount of quantitative information remains embedded in unstructured literature, making large-scale data compilation labor-intensive and difficult to maintain. Here, we introduce GeoExtractor, a large language model (LLM)-powered multi-agent framework designed to automatically extract structured information from geoscience literature. GeoExtractor employs a hierarchical extraction strategy that combines an iterative decision–retrieval–generation workflow with a two-stage retrieval mechanism, enabling the recovery of complex multi-instance scientific records from heterogeneous document sources. We evaluate the framework on four benchmark datasets spanning diverse geoscience sub-disciplines, using three LLMs and repeated runs to quantify variability. GeoExtractor generally performs better than the comparison methods on multi-record and multi-field tasks, while its limitations on single-record datasets are also characterized. To demonstrate its practical utility, we compiled a literature-derived database containing 2,259 zircon analyses from 179 publications of the Central Asian Orogenic Belt (CAOB), and the resulting dataset reproduces previously recognized tectonic patterns in the region. These results demonstrate that LLM-based multi-agent systems can facilitate the transformation of unstructured geoscientific literature into structured knowledge resources, substantially reducing the effort required for database construction and accelerating data-driven research across Earth science disciplines.