Breaking Diffusion: Parallel Reasoning with Draft Refinement

One-line takeaway

Microsoft's Draft Refinement Language Models use a single denoiser both to propose parallel revisions and to evaluate a candidate draft, letting a reasoning model reconsider each decision in context; trained to reconstruct every token from uniformly corrupted drafts, it learns preservation and repair without a diffusion reverse transition.

微软的草稿修订语言模型（DRLM）用一个去噪器并行提出修改并评估整份草稿，让推理模型能把每个决定放在整体里重新考虑；训练时从均匀损坏的草稿重建每个 token，同时学会保留和修复，不再走扩散的反向过程。