Comment by helloplanets

I guess it could be generalized to filling up the context window with any token, but just making sure none of the tokens repeat.

An interesting twist could be making sure a specific token is an anagram of the token N tokens back. This could possibly measure how much a model can actually plan forwards.