CacheBack Uses Receiver Needs to Compress Multi-Agent KV Caches

Summary

Multi-agent systems often communicate by sending text or latent key-value (KV) caches, but a full KV cache grows with both each agent’s context and the number of coordinating agents. The paper introduces receiver-conditioned communication: the receiving agent describes its local information needs so the sender can filter and compress the state it transmits. CacheBack implements this idea without additional training by using the sender’s attention weights to select relevant parts of the KV cache. On FanOutQA, CacheBack with Qwen 3 removes 75% of the state that would otherwise be sent, raises accuracy by 14.7 percentage points, and reduces median task-completion latency by 3.2 times compared with text communication. The reported gains are comparable across dense Transformer models, Mamba-attention hybrids, and sliding-window attention models.