What changed
Apple researchers have introduced a new approach to maximum inner product search (MIPS) that enhances retrieval efficiency for recommendation systems and retrieval-augmented generation (RAG). The method involves training small neural networks to learn how to route queries toward their best match more quickly. This approach trades an upfront amortization cost for faster, more accurate retrieval during inference.
What this means in practice
The new method, detailed in a paper presented at ICML 2026, leverages the mathematical structure of MIPS by approximating the support function of a set of keys using neural networks. This allows for more efficient routing of queries by learning how typical queries tend to be matched. The SupportNet architecture, which is based on input convex neural networks (ICNNs), can mimic the ground-truth support function for a lower computational cost. KeyNet models, on the other hand, are trained to directly output the best-matching key vector when given a query, with a consistency loss derived from Euler’s theorem for homogeneous functions to ensure geometric accuracy.
Limitations and availability
The method is designed for use in production environments where the query distribution is stable and well-characterized. It requires a precomputed dataset of training examples formed by sampling input vectors and pairing them with the relevant output signal from MIPS. The approach is not yet available as a public API or downloadable model, and the paper does not specify the exact hardware requirements or licensing conditions for its implementation.
