Draft-then-Refine Decoding Boosts Diffusion Language Models

A new inference method called draft-then-refine decoding improves the performance of diffusion language models (DLMs) while reducing latency. The approach, introduced in a paper on arXiv, involves generating a complete draft response and then refining it using bidirectional diffusion. The study evaluates two configurations: Flash-Flash, where the same model (LLaDA2.1-Flash) serves as both drafter and refiner, and Mini-Flash, where a smaller model (LLaDA2.1-Mini) drafts and Flash revises. In Flash-Flash, accuracy on GSM8K-384 improved from 0.848 to 0.899, and on MBPP-384 from 0.545 to 0.693, while running 1.20 times faster than the baseline. The Mini-Flash configuration, inspired by speculative decoding, is termed speculative correction. The paper is available at arXiv:2608.02625.