Multi‑agent large language model (LLM) systems distribute tasks across different models, so answering often requires knowledge residing in another agent’s context: a Sharer encodes information that a Receiver needs to complete its task. The usual approach is text exchange, which puts autoregressive decoding on the critical path and forces the message to be written without seeing the receiver’s state. Recent latent‑space protocols translate the Sharer’s key‑value (KV) cache directly into the Receiver’s: C2C supports heterogeneous models but requires both to read the same input, while LCF‑X removes the shared‑context requirement via position‑free Sharer‑cache pooling. LCF‑X still has three limitations: it compresses only the Sharer, supplies the same layer‑local summary to every Receiver position, and assumes matched layer count and KV geometry. We introduce XKV, which lifts all three constraints. XKV uses learned‑query attention to pool both caches; self‑attention over receiver‑aligned layer tokens, together with a learned layer map that reconciles different depths, mixes the pooled summaries into a compact joint memory; a shared position decoder lets each raw Receiver cache position retrieve its own per‑head‑gated residual in the Receiver’s native KV geometry. All models remain frozen and may differ in family, depth, KV‑head count, head dimension, and tokenizer; only the translator is trained. Across 45 dataset‑model pair settings (six heterogeneous and three same‑model ordered pairings, five datasets), XKV achieves the highest macro score and best average rank, improving over LCF‑X on every dataset (by 4.6 exact‑match and 4.2 F1 points on ROPES) and surpassing text communication on four of the five. XKV trains 76 % fewer parameters and translates a cache pair 10.3× faster (5.8 ms vs 59.9 ms); end‑to‑end it is 26 % faster than LCF‑X and 6.8× faster than text communication.
Blogger's Review: XKV’s clever handling of heterogeneous cache alignment dramatically boosts cross‑model collaboration efficiency, marking a significant milestone for latent‑space communication between LLMs.