From User Sequences to Scaling Laws: A Multi-Stage Architecture for Meta’s Ads Ranking
Meta achieves LLM-style scaling laws for ads ranking with 6% Instagram conversion lift
“a flexible production strategy that helps generalize sequence learning models and establish an LLM-style scaling law that predictably balances model performance with compute”
Meta Engineering describes two architectural breakthroughs in their ads ranking stack: a multi-stage model that decouples heavy offline user modeling from lightweight online inference, and a dense tokenization with target-aware attention technique that eliminates manual feature engineering. The system achieves LLM-style predictable scaling laws and has delivered measurable business results — 6% conversion lift on Instagram and 3% on Facebook. It matters because it demonstrates that transformer scaling intuitions from LLM research are transferable to latency-constrained production recommendation systems at Meta's scale.