@paullammers@eldritch.cafe

It does not seem like a contradiction. Although the paper is poorly done. Stating 0% in the abstract was misleading. The benchmarks are against other neural systems it looks like. They are at least decompiler tools I never heard of, which is suspicious. And the data is particular to similar coding styles and a lot of potential hidden issues, as deep neural networks tend to have.

But anyway, you can look up what a decompiler is, and how one works, for a refresher maybe? Decompilation has been an existing security technique for decades. It is just traditional code compilation in reverse, which does not allow for much ambiguity. You can use objdump on Linux for some easy assembler code, or a tool like this to get C code.
https://sourceforge.net/projects/decompiler/

What is your background? Are you interested in getting into security or programming language theory? Compiler writing is probably one of the most underated computer fields, considering how diverse computation hardware is becoming. And large compilers do have many security bugs. There is also attempts by SE researchers to make auto documenting code. This paper is kind of similar.