← Quit0 / 0 correct
LLMs & AIHard1 of 647

How does speculative decoding speed up generation without changing the output distribution?