Recursive Training Loops in LLMs: How training data properties modulate distribution shift in generated data?

Recursive Training Loops in LLMs: How training data properties modulate distribution shift in generated data?