1 paper · 1 filter
Zhicheng Du, Lan Ma
Gaussian-corrupted sentence embeddings have no direct linguistic interpretation, yet continuous diffusion language models can generate fluent text from them. We study this puzzle t…