B3: overlapping block denoising speeds up diffusion LMs 11.5-14.5% training-free
orlitany · x · 2026-11-12
arXiv paper [2611.14387] introduces B3 (Blurring the Block Boundaries): block-diffusion LMs are bottlenecked by their slowest token per block, so B3 refines two blocks in parallel within a shared forward pass and self-calibrates when to open the next block. Training-free and hyperparameter-free, it matches generation quality while delivering 11.5-14.5% speedups across 10 datasets, pushing the Pareto frontier for DLM inference.
More from Research
- IFM unveils PAN 2, an open foundation world model for physical intelligence — heyshrutimishra · 2026-11-12
- Axiom Math formalization finds broken step in finite simple group classification proof — jiqizhixin · 2026-11-12
- SimEX: agents learn manipulation skills in imperfect simulation and transfer them to real robots — chris_j_paxton · 2026-11-12
- Lie bracket visualized: how second-order residue lets wheeled robots move without sliding — burny_tech · 2026-11-12
- On PostTrainBench, AI post-training now beats the human-tuned SmolLM3-3B from the same base — maksym_andr · 2026-11-12
- FigBrains: models are outpacing our ability to build training environments by hand — vsikka · 2026-11-12