GenRec: Towards LLM-Native Recommendation at Netflix
Engineers at Netflix developed GenRec, an LLM-native recommendation system that outperforms its mature production counterpart while relying on fewer labeled examples and input signals. GenRec verbalizes user histories and item metadata into text, post-trains a foundation LLM on Netflix-specific data, and adds a catalog-aware scoring head to produce personalized rankings. In a large-scale A/B test, GenRec achieved significant improvements in short-term and long-term online metrics, reducing the reliance on hand-engineered features and shifting focus from feature engineering to context engineering. The system's two-phase training framework leverages Phase 1 to adapt a foundation LLM on Netflix data and Phase 2 to post-train on ranking-specific data and objectives. GenRec training uses conversations between users and a recommender, converted from log data, to support both language-modeling and ranking objectives. The system employs verbalization and context engineering techniques to balance token budgets and ranking quality, applying rewards-weighted training to align with long-term member