Imagine writing a program with a keyboard that only lets you append. You can think before typing, but once a token lands, the next token must live with it. This is how ordinary autoregressive language generation works. The model can later produce a correction, but it cannot silently rewrite the answer already emitted. Text diffusion changes that workflow. It starts with an incomplete or corrupted sequence and constructs an answer through repeated denoising. Multiple positions can become words during the same step. The opportunity is faster generation and more flexible editing. The challenge is making those parallel decisions agree without spending the speed advantage on extra computation. One distinction matters immediately: diffusion is not the opposite of a transformer. A transformer is a neural network architecture. Autoregression and diffusion specify how a model learns and generates. Many text diffusion models, including LLaDA, use transformers. We are comparing two ways to...
learn more
