We study a cross-attention message router that
selects targets by combining capability match, performance weighting, and reliability scoring, with multi-head routing and a KV
cache for repeated queries. Against round-robin and capabilityonly baselines, cross-attention improves capability satisfaction,
end-to-end latency, and reliability-weighted success. We also
quantify routing decision time benefits from the KV routing
cache.
#WuqingXinhaoLiandao / bgilbert1984