GenRec: Towards LLM-Native Recommendation at Netflix
Netflix replaced years of feature engineering with an LLM that verbalizes user history as text, matching their production recommender with 40x less training data while opening the door to natural-language steering.
Read Original Summary used for search
TLDR
• GenRec beats Netflix's mature production ranker using 40x fewer labeled examples by treating recommendation as a language modeling problem—verbalizing user interactions, item metadata, and context as natural language prompts
• Two-phase training: Phase 1 adapts an open-source LLM on Netflix data for content understanding, Phase 2 post-trains on ranking objectives with reward-weighted losses that optimize for long-term satisfaction, not just clicks
• "Context engineering" replaces feature engineering—the prompt becomes the feature vector, with careful decisions about what history to include, compress, or drop within a token budget
• Serves cost-efficiently via prefill-only inference (no autoregressive decoding), aggressive context compaction (1/3 original tokens with no quality loss), and smaller distilled models
• Achieved statistically significant gains on both short-term and long-term online metrics in A/B test covering ~10% of Netflix traffic
In Detail
Netflix's GenRec represents a fundamental shift from traditional recommendation systems built on thousands of hand-crafted features to an LLM-native approach. The system post-trains an internal foundation LLM specifically for ranking by verbalizing user histories, item metadata, and context as natural language. This "context engineering" approach treats the prompt as the new feature vector, with modeling effort shifting to decisions about what signals to include, how to compress repetitive behavior, and how to balance information density against token budgets.
The training follows a two-phase framework. Phase 1 adapts an open-source LLM on proprietary Netflix data to learn content understanding and member behavior patterns—this foundation is updated infrequently and shared across applications. Phase 2 post-trains specifically for ranking using interaction logs converted into conversational format, combining a catalog-aware ranking objective with language modeling objectives and reward-weighted losses. The reward weighting is crucial: it aligns the model with long-term member satisfaction and business goals (like balancing content types) rather than optimizing for immediate engagement metrics that could lead to over-recommending binge content.
GenRec achieves remarkable data efficiency and performance. With roughly 40x fewer Phase-2 labeled examples than their production system, it delivers +1.6% improvement in offline MRR and statistically significant gains in online A/B tests. Cost optimization comes through three strategies: smaller/distilled models, aggressive context compaction (reducing tokens to 1/3 with negligible quality loss), and prefill-only inference that scores the entire candidate set in a single forward pass. The system demonstrates clear scaling laws—both data and model size improvements translate predictably to quality gains—bringing recommendation systems into the broader LLM paradigm where investment decisions can be guided by scaling behavior rather than task-specific engineering intuition.