Netflix tests language model as alternative to hand-built recommendation logic

Insight summary

• •Netflix developed GenRec, a language-model-based recommendation system outperforming its old methods with significantly less training data.
• •GenRec converts user behavior into plain text, analyzed by a fine-tuned open-weight model to score matching titles.
• •Offline tests and a live A/B experiment showed measurable improvements in recommendation quality, with a 1.6% offline ranking improvement.
• •The system needs about 40 times fewer labeled examples during training compared to the previous system.
• •GenRec uses aggressive filtering of user interactions to fit model input limits and avoids suggesting non-existent titles.
• •The approach shifts recommendation development from feature engineering to context engineering and uses GPU servers for LLM tooling.
• •Netflix considers GenRec a promising alternative but not yet a full replacement for existing recommendation models.