GenRec: Netflix's LLM-native recommendation ranker
Netflix replaced its recommendation ranker's traditional feature-engineering stack with GenRec, an LLM-based system that scores titles from natural-language "context" — verbalized user history and catalog metadata — instead of thousands of hand-crafted features. A two-phase pipeline first adapts a foundation model on Netflix data, then post-trains it for ranking with a catalog-constrained scoring head (so it can't recommend a title that doesn't exist) and reward signals tuned for long-term satisfaction, not just click-through. It runs prefill-only on vLLM, scoring the whole catalog in a single forward pass without autoregressive decoding, and beat the production baseline offline (+1.6% Mean Reciprocal Rank) using roughly 40x fewer labeled examples, then held up with statistically significant gains in an online A/B test across about 10% of Netflix's traffic.
Sources & depth
- netflixtechblog.comGenRec: Towards LLM-Native Recommendation at Netflix2026-08-16 IST