标签
本文扩展了MST-Direct方法（一种使用Sinkhorn最优传输进行多元地质统计模拟的方法），从双变量无条件小网格设置扩展到多元、条件和大规模设置，精确保持联合分布，并优于现有方法。
This paper presents Block-Wise Differentiable Sinkhorn Attention, a method for efficient long-context balanced entropic optimal transport attention on TPU hardware. It introduces a tail-refinement surrogate for exact differentiation, proving an efficient backward pass schedule and demonstrating significant improvements in Pfam sequence alignment reconstruction.