IndexCache: Accelerating Sparse Attention via Cross-Layer Index Reuse

Benchmark Model Rank Results
mathematical-reasoning-on-aime-2025DSA (30B) + IndexCache (1/4, searched pattern)#17Accuracy: 92.6
code-generation-on-livecodebench-v6DSA (30B) + IndexCache (1/2, training-aware)#30Pass@1: 72.2
multiple-choice-question-answering-mcqa-on-gpqa-diamondDSA (30B) + IndexCache (1/4, searched pattern)#59Accuracy: 78.6