Diffusion-LM Improves Controllable Text Generation

Controlling the behavior of language models (LMs) without re-training is a major open problem in natural language generation. We develop a new non-autoregressive language model based on continuous diffusions that we call Diffusion-LM.