收录解读
ELF proposes Embedded Language Flows, a diffusion/flow language-modeling approach that operates primarily in continuous embedding space rather than over discrete token states.
The method uses continuous-time Flow Matching and delays the projection back to discrete tokens until the final step through a shared-weight network, making language generation structurally closer to image-domain diffusion workflows.
This formulation allows established diffusion techniques such as classifier-free guidance to transfer more naturally into language modeling, while reducing the sampling burden compared with prior diffusion language models.
For this repository, the paper is valuable because it contributes a reusable generative-modeling primitive for non-autoregressive language modeling and strengthens the broader continuous-generation path for text.