Input sequence mutation >>> novel token generation in LLMs. Why? IDK, there must be a good theory article someone could point me to.