Without any experiment showing that language modeling performance actually continues to improve past 32k tokens using this scheme, how are we supposed to tell w

Without any experiment showing that language modeling performance actually continues to improve past 32k tokens using this scheme, how are we supposed to tell whether this is actually viable?