Can diffusion language models reuse what autoregressive models learn during post-training?
A researcher introducing A2D says diffusion language models often come from pretrained autoregressive models.
TLDR
A researcher introducing the paper “A2D: AR-to-Diffusion Transfer and Composition” says diffusion language models often come from pretrained autoregressive (AR) models. The paper explores whether they can also reuse what AR models learn during post-training.
Combined views
2.5K
1 Source, first seen ago
