Apple's ICML paper: learned support functions speed up RAG retrieval up to 8x
大模型之路 · wechat · 2026-11-19
An Apple ML paper accepted at ICML 2026 replaces centroid-based IVF routing with learned support functions. Exploiting convex analysis — a cluster's best achievable score is a convex function of the query whose gradient points to the optimal key — it trains two small networks offline: SupportNet (a router scoring each cluster's upper bound) and KeyNet (a query rewriter feeding existing indexes like FAISS untouched). On Natural Questions (2.6M passages), SupportNet hits 95% recall 1.9x faster; KeyNet reaches 99% recall up to 8x faster scanning 11x fewer keys.
Caveats: the largest network is 20% of database size, results are vendor-reported on one dataset, and the method only works when queries follow the training distribution — requiring real query logs and stable query patterns. HN commenters half-argue that BM25/grep is often the better answer for LLM agents. Code: apple/ml-amips (arXiv 2603.08001).
More from Infra
- AWS emulator passed 66 of 72 hands-on checks, but Terraform exposed gaps — JeremyCMorgan · 2026-11-19
- LLM autoresearch boosts Qwen3.8 decode speed 2.4x on RTX 5060 Ti in 20 hours — bobaburger · 2026-11-19
- Vercel CEO: pitched an inference neocloud every 2 days as AI Gateway tokens grow 40% MoM — Kyrannio · 2026-11-19
- Open-weight models now handle 81.2% of OpenRouter tokens — the price of generic intelligence is open source — markjeffrey · 2026-11-19
- NVIDIA pitches AI factory digital twins: test designs before construction, boost tokens per watt — luisceze · 2026-11-19
- NYT: Singapore Firm Megaspeed Bought $4B+ in Nvidia Chips Within Reach of China's Military — Gregory_C_Allen · 2026-11-19