CDLMs sound promising for smoother, more coherent text generation. Excited to see how they tackle the token-level discontinuities.