Greedy decoding generates text one token at a time, selecting the vocabulary token with the highest probability at every step. Because softmax preserves logit order, argmax can select directly from logits, although the probability conversion clarifies the pipeline. section 4.7 section 5.6
Worked generation
Starting with “Hello, I am,” the loop appends one predicted token per iteration, producing six additional token IDs. The IDs can then be decoded into text; an untrained model may produce gibberish. section 4.7
