Netflix's GenPage generates entire homepage with single transformer, boosts engagement 0.24%

Netflix has unveiled GenPage, an end-to-end generative system that replaces the traditional multi-stage recommender pipeline with a single transformer model. Instead of sequentially ranking content, GenPage treats the user and request context as a prompt and autoregressively generates the entire structured, multi-row homepage as a single response. The training recipe mirrors LLM practices: pretraining on production pages followed by post-training using weighted binary classification (WBC) or reinforcement learning (RL). Key engineering innovations tackle cold start, model freshness, business-rule enforcement, and serving efficiency at scale.

In online A/B tests against Netflix's mature, highly optimized production homepage recommender, the WBC variant of GenPage delivered a statistically significant +0.24% lift on the core user engagement metric used for launch decisions, while simultaneously reducing end-to-end serving latency by 20%. Offline analysis revealed two key findings: enriching the prompt yields a larger improvement than scaling model capacity in the current regime, and RL post-training increases homepage diversity even though diversity is not explicitly optimized. This marks a paradigm shift from complex multi-stage recommendation to a single generative model, with implications for personalization across the streaming industry.