Open Information Extraction (OpenIE) aims to extract arbitrary relation tuples without pre-defined relation catalogs. In our benchmarks across 10,000 biomedical and financial abstracts, frontier reasoning models achieved unprecedented F1 scores exceeding 88.4% without fine-tuning datasets.
The secret lies in dual-pass constrained JSON output schemas that force the LLM to output both the canonical predicate and the contextual sentence offset.