Intel / article

Speculative Correction: Draft-then-Refine Decoding for Diffusion Language Models

Source: arxiv.org
Author: Brian K Chen, Chong Wu, Kenji Kawaguchi
Date: Published Aug 5, 2026

Context

A diffusion can revise text using on both sides of a position. This study tests whether that ability works better when the model first produces a complete draft, then revises the response as a whole. The comparison is with generating the answer block by block. Method.

Using LLaDA2.1-Flash for both drafting and revision raised accuracy on the tested GSM8K-384 arithmetic problems from 84.8% to 89.9%, while running 1.20 times faster than the selected block-generation baseline. The authors also report gains after targeted tuning of comparison settings. The result shows that changing how an existing model generates and revises an answer can improve its speed and accuracy without retraining it. Reported evaluation.

Key takeaways · AI-distilled

Draft-then-refine is a plug-and-play decoding pattern for diffusion LMs: generate a complete block-autoregressive draft, then let bidirectional diffusion revise the whole response.
Same-model refinement (Flash-Flash) lifts GSM8K from 0.848 to 0.899 while running 1.20x FASTER than the tuned block-autoregressive baseline, and MBPP from 0.545 to 0.693.
Speculative correction - a small Mini model drafts, Flash edits the draft as an initialization - trades quality for speed usefully: 0.294 vs 0.300 on MATH at 2.17x faster.
Ablations show completed drafts are what matter: refining from a fully masked span performs poorly, while global refinement adds most on GSM8K and local refinement captures most of the gain on MBPP and MATH.

Terms in this piece · Glossary

LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
model routing — Sending each request to a model chosen by the difficulty of the task, rather than using one model for everything.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.

Key quotes

“Flash-Flash improves GSM8K-384 accuracy from 0.848 to 0.899 while running 1.20 times faster than the selected Flash block-autoregressive baseline, and improves MBPP-384 from 0.545 to 0.693.”

“Mini-Flash provides useful quality-latency trade-offs, including MATH-384 performance of 0.294 versus 0.300 for Flash while running 2.17 times faster.”

Read the source arxiv.org