KVCMAS: Efficient KV cache Correction for Shared Context in Multi-Agent Systems
KVCMAS is presented, an online KV cache correction framework that represents cross-agent cache deviations using compact low-rank states and seamlessly chains corrections along the agent workflow without an additional reference prefill, establishing KVCMAS as an accurate and scalable KV cache sharing approach for prompt...
Hyesung Jeon, Hyeongju Ha, Seoyoun Lee et al.
· 0 citations