What Are Diffusion Language Models, and Why Are They Different From GPT?

Then you train the neural network to reverse this process, commonly using a mean squared error objective.