Apple Machine Learning Research · Learned MIPS Routing

Amortizing Maximum Inner Product Search with Learned Support Functions

Apple researchers propose training small neural networks offline, once for a fixed database of keys, to route queries to their best-matching keys faster at inference time. The SupportNet model learns per-cluster support functions using input convex neural networks, and KeyNet directly outputs the best-matching key vector. Both are trained on samples from the expected query distribution, which suits short user questions in RAG. The page excerpt does not give benchmark figures, and the paper was presented at ICML 2026.