HybridKV: Query-Agnostic KV Cache Compression via Head-Dominant Context Reconstruction

Changwoo Baek, Kyeongbo Kong† — Pusan National University · ICMLw 2026

🌐 Project page: served from docs/ via GitHub Pages.

HybridKV pre-selects salient attention heads offline and applies token-level context-reconstruction pruning only within them, matching full reconstruction accuracy while cutting compression-only overhead by up to 36%.

Credits

Page design inspired by WorldKV.