less-is-more-sparse-attention

Improves inference efficiency through training-free sparse attention using global token selection patterns aggregated from local attention heads for unified cross-head ranking.

Independent GitHub snapshot · Automated classification · Original documentation governs