Current LLMs decohere rapidly, and as far as I am aware, this problem seems like a fundamental to the architecture one.