The LLM is sampled to make only one-token continuation from the context. Presented a sequence of tokens, a single token is drawn with th