Blurry Window Attention

In three linesBlurry Window Attention (BLA) is a bounded-memory attention method inspired by State-Space Models. It reconstructs a blurry KV history via Dirichlet kernel interpolation, generalizing Sliding Window Attention. On MQAR, BLA achieves 8× better state efficiency than SWA and is competitive with linear attention models.

Summary generated by Claude — human-verified