Weird result on selecting number of topics for LDA

I can give you an informed answer with a reproducible example, called a reprex

Just reading the tea leafs suggests that whatever the y axis represents decreases for Arun2010 as the number of topics increase; it increases for CaoJuan2009 with some troubling local minima.

For the lower graph Deveraau2018's sweet spot is 10 and Griffiths2004 65.