Speculative correction lets diffusion language models draft, then revise

A new arXiv paper studies speculative correction, a decoding pattern for diffusion language models that first generates a full draft and then revises it bidirectionally. Diffusion language models can edit tokens in both directions, but many decoding methods still force them into block-by-block generation.

The authors tested LLaDA2.1-Flash and LLaDA2.1-Mini in two setups. In one, the same model drafted and refined its own answer. In another, the smaller model drafted while the larger model revised the result.

Reported gains included GSM8K-384 accuracy rising from 0.848 to 0.899 while running 1.20 times faster than the selected baseline, plus an MBPP-384 improvement from 0.545 to 0.693.

The results suggest that completed drafts can be useful starting points for global refinement. The approach is still benchmark evidence, but it points to a path where diffusion models use their editing ability more directly instead of imitating left-to-right generation.