What it is
Training-free decoder for block-wise diffusion LMs that parallelizes inter-block generation with wavefront scheduling while preserving or improving accuracy. Parallelize block-wise diffusion LM decoding without retraining by starting downstream blocks from informative drafts, then jointly refining and committing them in order.
Key results
- Up to 2.95× / 4.01× TPS vs LLaDA-2.1 / 2.0 in batched serving
- Cuts latency by up to 53.6% / 77.1% vs LLaDA-2.1 / 2.0
- Improves average accuracy by 1.3 points across 8 math/code benchmarks
Shows that self-correcting diffusion LMs can break strict inter-block serial decoding without retraining.