GenRecs Tokenization: Transforming Ads Recommendation at Meta with Stable Model Vocabulary
Meta's GenRecs Tokenization applies LLM semantic vocabulary to ads, yielding +0.35% conversion lift
“A newly launched "trail shoe" inherits what the model has already learned about that category, rather than starting a new representation with no signal.”
Meta has shipped GenRecs Tokenization, replacing per-item ID embeddings in its ads stack with a stable semantic token vocabulary drawn from LLM architecture, allowing cold-start ads to bootstrap from world knowledge rather than engagement history. Paired with Meta's GEM and Adaptive Ranking Model, the system delivered a measurable +0.35% conversion lift across surfaces. This is a notable production signal that LLM-style tokenization is now powering hyperscale commercial recommendation at Meta.