Pinned post
New paper: diffusion LMs often come from pretrained AR models. If we already reuse AR pretraining, can we also reuse what AR models learn during post-training?
We explore this in A2D: AR-to-Diffusion Transfer and Composition. ๐งต
We explore this in A2D: AR-to-Diffusion Transfer and Composition. ๐งต