DELA Enables Controllable Draft Refinement in Diffusion Language Models
October 26, 2026
DELA is a plug-in framework for frozen Diffusion Language Models that enables explicit control over Insert, Delete, Remask, and Stop actions. It separates high-level editing decisions from token-level denoising to allow for variable-length draft refinement without disrupting prefix stability.
HOW THIS AFFECTS YOU
●
builderYou can use this adapter approach to add precise text-editing capabilities to existing diffusion-based language models.
●
researcherYou can explore this method to decouple high-level edit control from low-level generative denoising.