I like John’s idea too — it might be fun to implement. I also think there’s a lot of room just for optimizing the snarXiv to sound more like the arXiv. Currently, its vocabulary and sentence structures are limited to whatever popped into my head while I was writing the grammar. There’s no sense in which the adjectives that appear, for instance, represent a complete list of common adjectives in arXiv titles. And the abstracts grammar definitely doesn’t have enough flexibility to mimic everything that people tend to write. My sense is that many who figured out how to beat the snarXiv simply learned what it could spit out and what it couldn’t, and that was basically enough to distinguish it from the arXiv every time.