Most multi-agent systems pass text, forcing every agent to decode and re-encode messages, which is lossy and expensive. This work instead communicates through KV caches, the transformer key-value states that store context, and tackles the harder case where sender and receiver are different models.
The proposed method learns a dense cross-model cache transformation with two-phase training: first reconstruct the receiver's cache, then train for generation-oriented communication. Across Qwen3-4B, 8B, and 14B, all six sender-receiver directions, and six benchmarks, it outperforms prior heterogeneous latent baselines and matches or exceeds text communication in context-aware settings at roughly 2 to 3 times lower compute. In context-unaware settings, where the receiver sees no input, prior methods collapse but this approach remains effective.