A Self-Pruning Transformer: Extreme KV-Cache Compression with Universal AttentionDavis WertheimerHaochen Shenet al.2026COLM 2026Conference paper
DropKV: Decoupling Residual-Output Perturbation for Near-Optimal KV-Cache EvictionAozhong ZhangSelcuk Gurseset al.2026ICML 2026Workshop paper
Frayed RoPE and Long Inputs: A Geometric PerspectiveDavis WertheimerAozhong Zhanget al.2026ICLR 2026Conference paper
Universal Position Interpolation: Unified Context Scaling for Hybrid Mamba-Transformer ModelsHaochen ShenDavis Wertheimeret al.2026ICLR 2026Conference paper