You can definitely do a lot of relabeling that way. It may be also worth trying a loop of "fix until it matches binary" for separate files... But I haven't seen anyone actually write it up.

There are attempts like this https://github.com/louisgthier/decompai that are related, but not quite the same as this project.

Edit: just gave it a go, and guessing reasonable variable names works extremely well when there's partial information already available. Especially nice for things like naming all the counters and temporaries when you know what you're iterating over (so it's just boring manual and trivial work), but can also figure out the meaning of larger patterns for function names.

I'm using an agent to port a game. I have the source. It's not going well. Lots of rabbit holes that are self-inflicted because the LLM doesn't want to port a lot of libraries because it's too much work for one round. It does a lot of stubbing and makes assumptions and that breaks the whole thing.

Fairly well. They aren't perfect, but they save a lot of time.

They are also downright superhuman at recognizing common library functions or spotting well known algorithms, even if badly mangled by compilation and decompilation.

I’m not sure if EFF has made any statements about it, but I would be concerned with the copyright aspects. Decompilation works (legally) because a new creative work is being produced.

It would be easy to argue LLMs are producing derivative, non-transformative works. AI companies have not made that any easier by paying publishers huge royalties to license training data.

I personally would stay away from it to avoid the risk, but I’d imagine if it became “easy” enough to produce a matching decomp with an LLM, we would get a more specific legal precedent pretty quickly.

I’ve not experimented but I thought they might be valuable for isolated variable / function renaming