Falconer is a collaborative framework that combines LLM's agent-like inference with lightweight proxy models for knowledge mining tasks that extract structured information from large-scale unstructured text based on user instructions. LLM acts as a planner, decomposing user instructions into executable pipelines and generating supervision for training small proxies. Falconer integrates classification and extraction into two atomic operations (get label and get span), allowing a single instruction-following model to replace multiple task-specific components. New benchmarks evaluate the consistency between the proxy model and annotations provided by humans and larger models. Falconer approaches the state-of-the-art LLM in instruction-following accuracy, while reducing inference costs by up to 90% and accelerating large-scale knowledge mining by more than 20x, providing an efficient and scalable foundation for deep research.
