XKV Enables Faster Latent Communication Between Diverse LLMs

Multi-agent LLM systems often require agents to share information, but traditional text-based communication is slow and lacks awareness of the receiver's internal state. Previous latent protocols, like C2C and LCF-X, attempted to translate a sharer's key-value (KV) cache into a receiver's, but had limitations such as requiring shared input contexts or assuming matched model architectures. A new protocol called XKV addresses these limitations. It uses learned-query attention to pool both the sharer's and receiver's caches, then mixes these pooled summaries into a compact joint memory using self-attention over receiver-aligned layer tokens. A learned layer map reconciles different model depths, and a shared position decoder allows each receiver cache position to retrieve its own gated residual in its native KV geometry. XKV allows models to remain frozen and differ significantly in family, depth, and KV geometry, with only the translator requiring training. In tests across 45 dataset-model pairs, XKV achieved the highest macro scores and average ranks, outperforming LCF-X on every dataset and surpassing text communication on most. It also trains with 76% fewer parameters and translates cache pairs 10.3 times faster, resulting in an end-to-end speedup of 26% over LCF-X and 6.8 times over text communication.