PIVOT Speeds Up Sparse Attention 4x Without Retraining

PIVOT groups adjacent queries that share ~90% of top-k token selections, running one proxy scan per group instead of per query. This reduces indexer complexity from O(L²) to O(L²/g), achieving a 4x speedup on DeepSeek-V3.2 and GLM-5.1. The method is training-free and offers two modes: PIVOT-Reuse for max speed and PIVOT-Refine for matching dense indexer accuracy.