I don't know if I'd call the output of modern de-compilers "very good", not for native code anyway. They're just a little better than raw disassembly. Even state of the art de-compilers struggle to reconstruct control flow, distinguish data from code, identify the presence of a variable, let alone its type, and they fundamentally lack context. If a LLM could be used even just to reliably reconstruct symbol information it would be game-changing.