UDLM: Uniform Discrete Diffusion Language Model
Reference: arXiv:2412.10193
UDLM uses uniform noise corruption with a novel continuous-time variational lower bound, enabling state-of-the-art performance among uniform noising methods. The forward process corrupts tokens towards a uniform distribution, and the model learns to reverse this process by minimizing a continuous-time ELBO.
Usage
Train on text8: