Foundations of Top-$k$ Decoding for Language Models | Read Paper on Bytez