More modern choices are JADX (https://github.com/skylot/jadx) or Vineflower (https://github.com/Vineflower/vineflower). If you want a paid, higher-quality option, try JEB (https://www.pnfsoftware.com/).
I took a quick peek at Vineflower first, and it's a fork of Fernflower. So would recommend that for anyone who might stumble on this in the future who is looking for a decompiler.
when debugging symbols are included (sort of the default) the local variables are already present; LLM would be the last thing I'd consider
then again, who needs reveng when you can use said LLMs to write new software "just in time" with the same API.
reveng also was one of those industries that always had a very suspicious crowd of people - i dont mean malicious, i mean... a lot of them drew a disturbing amount of pleasure from doing incredibly labourious work, sort of like someone who enjoys putting together an airfix model over many months with a microscopic brush and tweezers.
so i wonder if a lot of them perversely enjoy starting at reams of bytes and putting together this 10,000 piece puzzle, and having an llm solve it for them is a deep affront to their tastes.
That said there are probably ways a LLM could improve a decompiler in a way that does not impact its correctness. Like deriving class and variables names based on context, when symbols are missing or obfuscated.
why doens't someone train an LLM on predicting source code given a sequence of input machine code tokens? that is so obvious. why does nobody do it?
or maybe its not even neccesary, and doing something akin to fuzzing syscalls 'but smartly' probably yields more results.