KVCMAS: Efficient KV cache Correction for Shared Context in Multi-Agent Systems

Researchers introduced KVCMAS, a method that corrects the KV cache for shared context in prompt-specialized multi-agent systems, where agent-specific prefixes otherwise force each agent to repeatedly prefill the same growing context. The approach targets the redundant prefill cost that arises when multiple agents share a model but generate different KV caches for identical context.

Prompt-specialized multi-agent systems enable multiple agents to share a model while performing complementary roles to solve complex tasks. However, agent-specific prefixes change the KV cache generated for the same shared context, causing each agent to repeatedly prefill the growing context and con