DRIFT: Decompose, Retrieve, Illustrate, then Formalize Theorems

arXiv:2510.10815v4 Announce Type: replace Abstract: Automating the formalization of mathematical statements for theorem proving remains a major challenge for Large Language Models (LLMs). LLMs struggle to identify and utilize the prerequisite mathematical knowledge and its corresponding formal representation in languages like Lean. Current retrieval-augmented autoformalization methods query external libraries using the informal statement directly, but overlook a fundamental limitation: informal