Uniform Diffusion Scales Better for LLMs

While masked diffusion models excel at small scales, new research suggests that uniform diffusion demonstrates superior scaling performance for large language models.