NI Sampling: Accelerating Discrete Diffusion Sampling by Token Order Optimization

The core insight isn't just speed: discrete diffusion models generate tokens in parallel rather than left-to-right, which means the order in which tokens are committed during sampling is a free variable that prior work largely ignored. Neural Indicator Sampling treats that ordering as something learnable rather than arbitrary.

This sits inside a cluster of inference efficiency work Modelwire has been tracking across April. The most direct neighbor is UDM-GRPO (covered the same day, April 20), which applies reinforcement learning to uniform discrete diffusion models to stabilize training. Together, the two papers suggest discrete diffusion is entering a phase where researchers are attacking both the training side and the sampling side simultaneously. Further out, the token-level efficiency framing connects to K-Token Merging from April 16, which compresses token sequences at the embedding level to reduce inference cost in standard LLMs. The approaches are architecturally different, but both are responding to the same pressure: parallel decoding workflows waste compute on tokens that are already effectively decided.

The paper claims order-of-magnitude reductions in sampling iterations without accuracy loss; the test is whether those gains replicate on longer-form generation tasks (document or code completion) where token interdependence is higher and early commitment errors compound.