Researchers from Amazon Web Services (AWS) and the University of Central Florida (UCF) have introduced XKV, a dual-cache latent space communication protocol that connects heterogeneous large language models directly through their key-value (KV) caches. Developed by Jiyao Liu, Qi Zhang, Yaoyi Jia, Ziwen Kan, and Song Wang, the system eliminates autoregressive decoding and re-tokenization between agents entirely, allowing models to exchange continuous internal representations instead of chatty text.
The framework keeps both LLMs completely frozen, training only a lightweight translation bridge. Unlike earlier latent communication methods such as C2C, which demands identical inputs, or LCF-X, which only compresses the sender's cache, XKV pools both sharer and receiver caches using learned-query attention. It blends them into a joint memory across disparate layer counts via a learned layer map, allowing raw receiver positions to retrieve native-geometry KV updates across models with differing architecture families, depths, KV-head counts, and tokenizers.
Benchmark evaluations across 45 dataset-model pairings show XKV achieving the top macro score and best average rank among tested latent protocols, outperforming traditional text communication on four out of five datasets while beating LCF-X on ROPES by 4.6 exact-match points. By translating cache pairs in 5.8 milliseconds compared to 59.9 milliseconds for LCF-X while training 76% fewer parameters, XKV delivers an end-to-end communication speedup of 6.8x over text exchange, providing a practical blueprint for low-latency pipelines built from heterogeneous specialized models.