Netflix's GenRec LLM ranker beats production with fewer labels

Netflix has shared early findings on GenRec, an LLM-backed recommendation ranker built atop an in-house foundational LLM. Instead of traditional discriminative rankers that rely on thousands of hand-crafted features, GenRec uses verbalized user histories and context as inputs. The system follows a two-phase framework: Phase 1 adapts an open-source LLM to Netflix catalog and member behavior, while Phase 2 post-trains the model with recommendation-specific data, labels, and reward signals aligned to business goals and long-term member satisfaction. This paper focuses on Phase 2 and the engineering transition: input verbalization, post-training data construction, reward integration, and a cost-constrained serving design based on a prefill-only inference approach. The authors describe this as a paradigm shift from feature engineering to context engineering and from bespoke architectures to shared foundation backbones.

The team evaluated GenRec in a large-scale A/B test against Netflix's current production ranker. Results show that GenRec, trained with substantially fewer Phase-2 labeled examples and input signals, achieved statistically significant gains in both offline and online metrics. This is notable because it suggests LLM-backed recommenders can match or exceed heavily tuned production systems while relying more on natural language understanding of user context and less on manually engineered features. The paper also discusses practical lessons for serving such systems under real-world resource constraints, particularly the prefill-only inference trick to manage costs. For the broader industry, GenRec points toward a future where recommendation systems are built on general-purpose foundation models, with teams focusing on context design and reward modeling rather than feature pipelines.