Source
From User Sequences to Scaling Laws: A Multi-Stage Architecture for Meta’s Ads Ranking
Meta developed a multi-stage sequence model and dense tokenization techniques for its Generative Ads Recommendation Model (GEM). This architecture separates offline user behavior modeling from real-time online ranking to improve efficiency and scalability.
Why it matters
This system improves ad relevance, which increases the likelihood of conversions for advertisers and provides users with more personalized content.
The details
- Conversions increased 6% on Instagram and 3% on Facebook.
- Facebook ad clicks saw a cumulative lift of 3.5%.
- Performance scales predictably based on compute, similar to large language models.
Get the weekly recap
The stories like this one, picked and explained — once a week, straight to your inbox.