📄 中文摘要
块扩散技术在扩散语言模型中实现了高效的并行优化，但其解码行为对块大小的依赖性极强。现有的块大小策略通常依赖于固定规则或启发式信号，未能考虑决定哪些标记可以安全地一起优化的依赖几何。这促使了对扩散解码的几何视角的提出：强因果顺序的区域需要顺序更新，而语义一致的区域则允许并行优化。GeoBlock 是一种几何感知的块推断框架，能够直接从基于注意力的依赖几何中确定块粒度。GeoBlock 不依赖于预定义的调度或局部置信度启发式，而是通过几何信息实现更为灵活和高效的块优化。
GeoBlock：从依赖几何推断扩散语言模型中的块粒度
📄 English Summary
GeoBlock: Inferring Block Granularity from Dependency Geometry in Diffusion Language Models
Block diffusion enables efficient parallel refinement in diffusion language models, but decoding behavior is critically dependent on block size. Existing block-sizing strategies rely on fixed rules or heuristic signals, failing to account for the dependency geometry that dictates which tokens can be safely refined together. This leads to a geometric perspective on diffusion decoding: regions with strong causal ordering necessitate sequential updates, while semantically cohesive regions allow for parallel refinement. GeoBlock is introduced as a geometry-aware block inference framework that determines block granularity directly from attention-derived dependency geometry. By moving away from predefined schedules or local confidence heuristics, GeoBlock facilitates more flexible and efficient block optimization.