Meta engineers describe a multi-stage sequence-modeling architecture for ads ranking that decouples heavy offline user modeling from lightweight, latency-sensitive online ranking, using dense tokenization and target-aware attention to learn feature interactions directly from behavioral data instead of hand-engineered sparse features. The architecture reportedly exhibits LLM-style log-linear scaling laws, where model quality improves predictably with added compute, and it now underpins Meta’s Generative Ads Recommendation Model (GEM). The team reports a cumulative 6% lift in Instagram conversions and 3% in Facebook conversions plus 3.5% more ad clicks on Facebook, achieved without a proportional increase in serving resources.
