Researchers have developed a new method for heterogeneous language models to communicate with each other. The approach, called XKV, allows different models to share information without needing to read the same input or having matched layer count and key-value geometry. This is achieved through learned-query attention pools, self-attention over receiver-aligned layer tokens, and a shared position decoder. In experiments, XKV outperformed previous methods on several datasets, improving accuracy by up to 4.6 exact-match points and reducing training time by 76%. The method requires less parameters and translates cache pairs 10.3x faster than previous approaches. --- Why it matters: This matters because it enables more efficient communication between different language models, which can be crucial for applications like multi-agent systems and knowledge sharing across various tasks. Source: https://arxiv.org/abs/2608.20617