LLMs as Idiomatic Decompilers: Small Specialized Models Recover High-Level Dart Code From x86-64 Assembly

arXiv·medium signal

First study of LLM-based decompilation targeting modern languages beyond C — specifically Dart and Swift. Small specialized LLMs trained with synthetic same-language examples produce idiomatic high-level code from x86-64 assembly, compared against augmentation with human-written examples. Opens reverse engineering capabilities for mobile and cross-platform apps built in non-C languages that traditional decompilers handle poorly.