Double-P
Hierarchical top-p sparse attention for long-context LLMs
Advisors: Prof. Jishen Zhao and Prof. Zhijian Liu · 2025/09 – 2026/01
- Proposed a hierarchical sparse attention framework that jointly optimizes top-p accuracy, selection overhead, and sparse attention cost.
- Performed coarse-grained top-p estimation at the cluster level using size-weighted centroids, then adaptively refined computation through a second top-p stage.
- Achieved near-zero accuracy drop with up to 1.8× attention computation reduction and 1.3× end-to-end decoding speedup over state-of-the-art methods.
Paper: arXiv:2602.05191