Radar ·
Log-linear block selection for sparse attention
Published on September 28, 2026, the article describes a hierarchical block-selection method for sparse attention in long contexts, reducing selection cost from O(N²) to O(N log N).
Published on September 28, 2026, the article presents a hierarchical method for selecting blocks in sparse attention. It scores large context regions with LogSumExp, then recursively narrows selection to regions considered promising. According to the summary, this reduces block-selection cost from O(N²) to O(N log N) for models with long contexts.
The result concerns selection cost, not necessarily the model's total execution cost. To check the claim's scope, consult the original article and verify how N is defined, the evaluation conditions, and the experimental results. If you use AI to study or apply the method, avoid sending internal or identifiable data without authorization, and check conclusions against the original text.
Original source: Block-Sparse Attention com Seleção de Blocos Log-Linear