The paper maps two energy-based methods to each other. I'm not sure what the innovation is.
The novelty is in identifying spurious samples at the memorization-generalization threshold that act like memorized samples even though they don't appear in the training data.