Netflix tests language model as alternative to hand-built recommendation logic
decoder ·
Netflix developed a new recommendation system called GenRec that uses a language model instead of its traditional hand-built feature-based engine. The company reports that GenRec achieved better ranking quality in offline tests while requiring far less labeled training data, and a live A/B test on about ten percent of traffic showed small but statistically significant metric improvements. GenRec converts user watch history and context into plain text, and it filters and condenses interaction events to fit the model's context window. A fine-tuned open-weight model is trained in two stages: first to understand Netflix's catalog and behavior, then specifically as a ranking model that scores all available titles in a single pass. Netflix views this as part of a broader trend toward using general-purpose language models instead of custom-built recommendation architectures, though it calls the system an early step rather than a full replacement.
decoder