Authors: Ying Li, Arjun Rao, Shradha Sehgal

Recommendations are central to the Netflix experience. Current production models depend on thousands of hand-crafted features and specialized architectures, making it costly to onboard new use cases. Large language models (LLMs) offer a new paradigm by representing user histories and item metadata directly as text, enabling natural-language prompting and shared semantic understanding. However, off-the-shelf LLMs are not production-ready recommenders—they over-recommend popular content, hallucinate out-of-catalog items, and lack personalization.

To address this, Netflix built GenRec, an LLM-backed recommendation ranker. GenRec post-trains an internal foundation LLM on Netflix-specific data and objectives. It verbalizes user histories, item metadata, and context as text, adds a catalog-aware scoring head, and aligns with long-term member value using reward signals. GenRec runs in prefill-only mode for cost efficiency.

In a large-scale A/B test, GenRec achieved statistically significant improvements in both short-term and long-term metrics while using only a fraction of labeled data and input signals. It reduces reliance on hand-engineered features, shifting focus to context engineering. GenRec follows a two-phase framework: Phase 1 adapts a foundation LLM on Netflix corpora, and Phase 2 post-trains on ranking-specific data. Training data is converted into conversational formats to support language modeling and ranking objectives. Context engineering manages token budgets by retaining high-signal interactions and compressing repetitive ones.