HySparse: A Hybrid Sparse Attention Architecture with Oracle Token Selection and KV Cache Sharing

Benchmark Model Rank Results
common-sense-reasoning-on-winograndeHySparse (80B MoE)–Accuracy: 72.1
math-word-problem-solving-on-mathHySparse (80B MoE)–Accuracy: 30.8
sentence-completion-on-hellaswagHySparse (80B MoE)–Accuracy: 79.0