Meta's GenRecs Tokenization Improves Ad Recommendations with LLM-based Semantic Vocabulary

Generative Recommenders at Meta

Meta has implemented Generative Recommenders (GenRecs) to transform its ad recommendation systems. This new paradigm shifts from scoring individual ad candidates to using large language model (LLM) architecture to reason about ad content and user preferences simultaneously. This allows for more intelligent and precise ad matching by leveraging pre-trained world knowledge and semantic understanding.

Introducing GenRecs Tokenization

GenRecs Tokenization is the foundational data representation enabling this shift. Instead of assigning a unique ID to every item, it represents each item using a set of tokens from a stable semantic vocabulary. This token-based approach allows the model to understand and relate new products or ads to existing categories, even without prior engagement data.

Impact on Advertiser Performance

Paired with Meta’s Generative Ads Model (GEM) and Adaptive Ranking Model, GenRecs Tokenization has resulted in a +0.35% lift in conversions across various surfaces for advertisers. This improvement stems from the system's ability to predict the most relevant ad for each person, including for brand-new products with no engagement history.

Semantic Understanding for Recommendations

The use of semantic tokens means that the model can infer characteristics of a new item, such as a 'trail shoe,' based on what it has already learned about that category. This allows new items to inherit existing knowledge rather than starting with no signal, improving the accuracy and relevance of recommendations from the outset. This structured vocabulary approach supports both retrieval and ranking across all item types, including ads, products, pages, and pixels.