Haoyu Guo

DefensiveKV: Taming the Fragility of KV Cache Eviction in LLM Inference

Identify Critical KV Cache in LLM Inference from an Output Perturbation Perspective