Matt sat on the couch LM p(· | context) couch 2.24 0.414 chair 1.73 0.329 floor 1.01 0.249 cat −1.98 0.006 table −3.66 0.001 potato −6.23 0.000 logits probabilities

At each step the LM emits a logit per vocabulary token; softmax turns logits into a probability distribution. Sampling picks one token (here, the top one).