Netflix GenRec: An LLM-Native Ranker Beyond Feature Engineering
Netflix's GenRec replaces feature engineering with an LLM-native ranker, matching production performance using far less labeled data.
Netflix has introduced GenRec, an LLM-based recommendation ranker built to replace much of the hand-engineered feature stack behind its production recommender system. Instead of thousands of manual features for users, items, and interactions, GenRec verbalizes user history, item metadata, and context as natural-language prompts fed into a Netflix-adapted foundation LLM.
The system follows a two-phase training approach: Phase 1 adapts an open-source LLM on proprietary Netflix data to build general content and member-behavior understanding, while Phase 2 post-trains the model specifically for ranking, combining a catalog-aware scoring objective, a language-modeling objective, and reward-weighted losses aligned with long-term member value. To manage token budgets at Netflix's scale, the team applies context engineering — selectively retaining, summarizing, or dropping interaction history based on signal strength.
In a large-scale A/B test, GenRec matched or exceeded a mature production ranker on both short- and long-term online metrics while using only a fraction of the labeled data and input signals the existing system requires, and it runs in prefill-only mode on Netflix's LLM serving stack to control inference costs.
For engineers, GenRec is a concrete, production-validated example of recommendation systems shifting from feature engineering to context engineering, and from specialized architectures to general-purpose LLM backbones.