Parallax: A Parameterized Local Linear Attention That Keeps Softmax and Adds a Learned Covariance Correction Branch 发表评论 / 作者: / 6 月 1, 2026