Researchers propose draft-refine decoding for diffusion LMs
New 'draft-then-refine' decoding boosts diffusion language models by 1.2x with 5% accuracy gains
Deep Dive
Brian K Chen, Chong Wu, and Kenji Kawaguchi introduce "speculative correction," a draft-then-refine decoding method for diffusion language models. Instead of generating text block by block left-to-right, the model first produces a complete draft, then refines it with bidirectional diffusion. Using LLaDA2.1-Flash and LLaDA2.1-Mini, the same-model Flash-Flash setup boosts GSM8K-384 accuracy from 0
Key Points
- Flash-Flash improves GSM8K accuracy from 0.848 to 0.899 while running 1.20x faster than baseline
- Mini-Flash achieves 2.17x speedup on MATH-384 with just 0.006 accuracy drop (0.294 vs 0.300)
- Method uses training-free 'speculative correction' where a draft model initializes refinement by a stronger model
Why It Matters
Could make diffusion language models 2x faster while improving accuracy, potentially unlocking real-time bidirectional AI generation.