View original
GenRec: An LLM-Backed Recommendation Ranker at Netflix
Paper

GenRec: An LLM-Backed Recommendation Ranker at Netflix

Arjun Rao
2026.09.29
·Arxiv·by Homin.Lee
#GenRec#LLM#Netflix#Ranker#Recommendation System

Key Points

  • 1Netflix developed GenRec, an LLM-backed recommendation ranker that utilizes a two-phase training framework to transition from traditional feature engineering to natural-language-based context engineering.
  • 2By implementing a prefill-only inference approach and reward-weighted ranking loss, the model effectively balances business objectives and long-term user satisfaction within strict cost constraints.
  • 3Empirical results demonstrate that GenRec achieves superior ranking performance compared to mature production baselines while requiring significantly fewer input signals and Phase-2 training examples.

GenRec is an LLM-backed recommendation ranker developed at Netflix that transitions from traditional feature-engineered ranking models to a generative, context-aware foundation model architecture. The system operates on a two-phase training framework: Phase 1 establishes an "in-house" foundation model by adapting an open-source LLM to proprietary Netflix data for deep content and user understanding; Phase 2 performs high-cadence post-training focused on specific ranking objectives, business alignment, and long-term member satisfaction.

The core methodology revolves around replacing traditional feature engineering with "context engineering." Rather than processing hand-crafted features, GenRec verbalizes user interaction histories (viewing, playing, thumbs-up events) and contextual signals (time, device, surface) as natural language prompts. To address the challenge of limited token budgets and high inference costs, Netflix implements aggressive verbalization strategies: full retention of high-signal events, removal of low-signal interactions, and compression/summarization of repetitive binge-watching behaviors.

The model architecture utilizes a decoder-only Transformer backbone augmented with a catalog-aware ranking head. During inference, GenRec operates in a "prefill-only" mode within the vLLM framework; it processes the verbalized input and produces ranking scores for the entire item catalog in a single forward pass, bypassing expensive autoregressive beam search.

The training objective is a multi-objective loss function designed to balance ranking accuracy, language modeling, and business constraints:

L=α⋅Lranking+β⋅Llanguage+γ⋅LmiscellaneousL = \alpha \cdot L_{ranking} + \beta \cdot L_{language} + \gamma \cdot L_{miscellaneous}

s.t. α+β+γ=1,α,β,γ≥0\text{s.t. } \alpha + \beta + \gamma = 1, \quad \alpha, \beta, \gamma \geq 0

To ensure recommendations align with business goals (e.g., content variety and long-term retention), Netflix integrates reward signals via a reward-weighted ranking loss. This steers the model toward long-term value by scaling the cross-entropy loss of training examples based on scalar weights derived from external reward models. This approach mimics the benefits of reinforcement learning while maintaining the stability and deployment simplicity of supervised fine-tuning.

Experimental results from large-scale A/B testing demonstrate that GenRec achieves statistically significant gains in offline and online metrics compared to the mature production baseline, despite using significantly fewer Phase-2 labeled training examples and input signals. The study concludes that LLM-backed ranking represents a paradigm shift from manual feature engineering toward context-driven, foundation-backbone architectures that are more adaptable to new content types and business requirements.