LLMs are capable of introspection - and the kind where models know facts about themselves not in the dataset and right off the bat
(from experience, 405b does this well)

Buuut you don't need academic research to figure that out, it's far more efficient to talk/interact with the models yourself and find out in like a day than spend a while setting up experiments and writing a paper

Besides the "speed-of-understanding" efficiency you get from playing with models yourself to understand cognitive properties like this, you also don't get a mode collapsed view of how far these cognitive properties can go and what shape they take on