tl;dr it uses the standard encoder-combiner-MLP-output results, but with a lot of variability in the process.