Netflix is testing GenRec, a language-model recommendation system that converts selected viewing behavior into natural language instead of relying on thousands of hand-crafted features. Plays, viewing duration, ratings, list additions, and drop-offs become a text-based history from which the model can infer patterns.
The proprietary system starts with an unnamed open-weight model fine-tuned on Netflix data, followed by a more frequently updated ranking stage. To stay within the context window, it preserves high-signal events, removes brief interactions, and condenses binge sessions. A separate component restricts results to real catalog titles, while vLLM lets the model score candidates in one pass without generating text.
In offline testing, GenRec improved ranking quality by about 1.6% over Netflix’s established production system. Its second training stage used roughly 40 times fewer labeled examples, although that comparison does not cover all training data. In a four-week experiment involving about 10% of traffic on precomputed recommendation surfaces, a short-term home-screen metric rose 0.115% and a long-term core metric increased 0.006%. Netflix said both changes were statistically significant.
The company describes GenRec as an early but promising step, not a planned full replacement for its current system. The approach shifts work from manually engineering features toward choosing which context signals the model should receive and how much history to include.