research
The Diffusion Model Revised the Whole Draft
A plug-in decoding step improved reported math and code scores while preserving useful speed trade-offs.
Summary
A plug-in decoding step improved reported math and code scores while preserving useful speed trade-offs.
A preprint lets diffusion language models generate a complete draft and then revise it bidirectionally. With LLaDA2.1, same-model draft-and-refine raised reported GSM8K accuracy from 0.848 to 0.899 and MBPP from 0.545 to 0.693; a smaller drafter with a larger refiner offered faster trade-offs rather than uniform quality parity. The evidence is limited to the tested models and benchmarks.
Why it matters
A plug-in decoding step improved reported math and code scores while preserving useful speed trade-offs.
Limits and context
No additional limitation was separately recorded.
Key claims
A plug-in decoding step improved reported math and code scores while preserving useful speed trade-offs.
Evidence: source-2026-08-05-018
Sources
- arXiv preprint 2608.02625arXiv · primary research
Corrections
No corrections have been recorded for this story.