Local Open-Weight LLMs Extract Domain Models From Source Code Without Ever Seeing the Whole Repo

arXiv 2608.12228·low signalXBlueskyLinkedInCopy link

Proprietary frontier LLMs are off-limits in many industrial reverse-engineering contexts for privacy and confidentiality reasons, while compact local models can't fit large codebases in their context window. The proposed approach combines structural and semantic heuristics with iterative LLM reasoning over progressively ranked subsets of code elements, identifying domain concepts and refining domain boundaries without full-system context. It reports high F1-scores across a dataset of ten projects, each pairing a curated domain model with its implementation, while remaining fully executable on locally deployable models.