Next-Token Prediction

A paradigm in which language models generate text serially token by token, predicting the probability distribution of the next word based on preceding context; it is the underlying generation method of current mainstream large language models.