Apple researchers propose SupportNet and KeyNet for faster retrieval

Key facts

Apple researchers presented a paper at ICML 2026 proposing SupportNet and KeyNet for faster Maximum Inner Product Search.

SupportNet learns the support function using input convex neural networks, while KeyNet directly outputs the best-matching key vector.

The methods aim to speed up retrieval in AI systems such as RAG and recommendation engines.

Apple researchers have presented a paper at ICML 2026 proposing two new neural network architectures — SupportNet and KeyNet — designed to accelerate Maximum Inner Product Search (MIPS) for retrieval-augmented generation and recommendation systems. The paper, titled “Amortizing Maximum Inner Product Search with Learned Support Functions”, was published today by Apple Machine Learning Research.

The methods involve an upfront amortisation cost, learning from the query distribution to improve speed and accuracy at inference. SupportNet learns the support function using input convex neural networks (ICNNs), while KeyNet learns to directly output the best-matching key vector.

MIPS is a core operation in recommendation systems, search and RAG. Conventional approaches trade off accuracy for speed using clustering and quantisation; the Apple researchers’ method instead learns the support function of the set of key vectors, whose gradient returns the best-matching key.

Why it matters

Our take: improving MIPS directly attacks a bottleneck in RAG, which Apple is increasingly relying on for on-device AI. The work follows a similar vein to the SSR multimodal agent research Apple published earlier this year, which cut agent latency by 90 per cent.

Key facts

Why it matters

Sources