Beyond patent citations: a benchmark for detecting implicit science–technology linkages using open large language models

Science–technology linkage (STL) analysis relies predominantly on patent citations to scientific literature, an indicator known to underestimate actual knowledge interactions. Computational alternatives exist but have not been systematically compared, and whether links detected beyond citations represent genuine intellectual relationships remains unestablished. This study addresses both questions through a multi-paradigm benchmark in wind energy (3,611 patents, 43,582 publications, 350 citation-verified ground-truth links). Five retrieval methods are evaluated: lexical (TF-IDF, BM25), BERT-era embeddings (MiniLM), and modern dense embeddings (mxbai-embed, nomic-embed). A validation stage employing two open-weight generative LLMs as independent judges assesses the top embedding-detected implicit links. Results show that methods with comparable aggregate recall identify largely non-overlapping link subsets (pairwise Jaccard < 6%), that a widely used general-purpose sentence encoder significantly underperforms all alternatives including a lexical baseline, and that 81% of top implicit links receive dual-model validation as meaningful patent–publication relationships, substantially exceeding random same-domain controls. All models are open-weight, offering a reproducible two-stage pipeline for detecting citation-invisible science–technology interactions.