Netflix Replaces Manual Feature Engineering with LLM-Powered GenRec

Netflix has deployed GenRec, a new recommendation ranker powered by large language models (LLMs) that fundamentally changes how the platform predicts user preferences. The system marks a transition from traditional manual feature engineering to a "context engineering" approach, leveraging natural language to understand user behavior.

According to the Netflix Technology Blog and a technical paper published on arXiv, GenRec replaces a traditional discriminative ranker that previously relied on thousands of hand-crafted features tracking users, items, and interactions. Instead of these rigid data points, GenRec uses verbalized user histories and content descriptions processed by an in-house foundational LLM. Large-scale A/B testing confirmed that GenRec achieved statistically significant gains in both offline and online metrics compared to the previous production ranker, even while utilizing fewer labeled training examples.

The Two-Phase Architecture

To achieve this performance, Netflix implemented a two-phase training framework. In the first phase, the company adapted an open-source LLM using Netflix-specific data to build a deep understanding of its content catalog and member behavior. The second phase involved post-training the model with ranking-specific data and reward signals, specifically designed to align the system's outputs with long-term member satisfaction.

To handle the immense computational demands of serving an LLM at scale, the system runs on Netflix's internal LLM stack using vLLM. To maintain operational efficiency and control costs, the company employs a "prefill-only" inference approach, optimizing how the model processes input tokens without the overhead of traditional generative decoding.

A Shift in Recommendation Logic

For years, the industry standard for recommendation engines involved complex, bespoke architectures where engineers manually defined which data points—such as time of day or genre frequency—mattered most. GenRec suggests that foundation models can instead synthesize this information through natural language descriptions, allowing the system to generalize more effectively across diverse content types and evolving user behaviors without constant manual updates.

This shift represents a broader move toward LLM-native systems in the streaming and e-commerce sectors. By reducing the reliance on domain-specific feature engineering, Netflix is demonstrating that general-purpose models can outperform specialized ranking architectures when properly aligned with platform-specific rewards.

Industry Implications

If GenRec continues to scale successfully, it provides a blueprint for other tech giants to dismantle legacy ranking systems. The ability to achieve superior metrics with less labeled training data suggests that the inherent reasoning capabilities of LLMs can compensate for the lack of massive, manually curated datasets.

Industry observers will now be watching to see if this "context engineering" approach can be applied to other areas of the user experience, such as personalized search or dynamic interface adjustments, and whether the cost-constrained inference model can remain sustainable as the user base grows.