DraftPal

LLMs write text via probabilistic next-token prediction and sampling

Large language models generate text through a process of next-token prediction, where the model calculates probabilities for every possible next word based on the preceding text. Instead of always selecting the highest …