Researchers introduce Joint Subspace Sparse Attention, a method that improves long-context decoding efficiency by learning shared query-key subspaces and using gather-free kernels for token-level selection.
Carried by: arXiv, HF Daily Papers. First seen: .