A new paper on arXiv introduces Stratified Sparse Attention, an approach aimed at reducing the gap between sparse and dense attention when sparsity is extreme. The authors frame the problem around long-context inference, where the cost of attending to a growing Key-Value cache becomes a major bottleneck.
Sparse attention is a known remedy: it accesses only a fraction of the cache, lowering cost. But the abstract notes that existing sparse methods still struggle at extreme sparsity, where the quality gap relative to dense attention widens. The paper's proposed strategy is to address that gap directly.
Because only the abstract is available in the source, the article does not detail the mechanism or experimental results. The significance, as stated, is in targeting the sparse-to-dense gap at the very sparsity levels where prior approaches degrade.