SYSTEM Cited by 1 source
Netflix GenRec¶
GenRec is Netflix's LLM-backed recommendation ranker — a decoder-only transformer post-trained on Netflix-specific data and objectives to perform full-catalog ranking. It represents Netflix's step toward an LLM-native recommendation stack where the model's world knowledge and language understanding replace thousands of hand-crafted features.
Architecture¶
GenRec is a decoder-only Transformer augmented with a catalog-aware ranking head:
- Verbalizer: Serializes user history, context (device, surface, locale, time), and item metadata into a single text sequence.
- LLM backbone: Processes the text and produces a pooled hidden state h summarizing user preferences.
- Scoring head: Each catalog item i has a learned embedding eᵢ. A scoring function φ(h, eᵢ) produces per-item scores; softmax yields a probability distribution converted to a ranking.
All parameters — backbone, scoring head, item embeddings — are trained jointly.
Two-Phase Training¶
- Phase 1 (Foundation): Adapts an open-source LLM on proprietary Netflix corpora for content understanding + member behavior patterns. Updated infrequently. Shared backbone for many applications.
- Phase 2 (GenRec): Post-trains the foundation model with ranking-specific data and objectives. Refreshed more frequently. Incorporates reward-weighted losses for alignment.
Training Objectives (Multi-Objective Loss)¶
- Catalog-aware ranking objective — cross-entropy over candidate set, positives labeled from high-value engagements.
- Language modeling objective — preserves language understanding over verbalized inputs.
- Reward-weighted loss — separate reward models provide per-example scalar weights for long-term satisfaction proxies and behavior rebalancing across content types.
Serving¶
- Runs on Netflix MSS using vLLM in prefill-only mode.
- No autoregressive decoding — the model processes the prompt once and scores the entire candidate set in a single forward pass.
- Three cost-control levers: smaller/distilled models, aggressive context compaction (~1/3 tokens with negligible quality loss), prefill-only inference.
Results¶
- A/B test: ~10% of Netflix traffic, ~4 weeks, statistically significant gains on both short-term and long-term online metrics vs. the production ranker.
- Data efficiency: 10–40× fewer Phase-2 labeled examples than the production baseline.
- Offline MRR: +1.6% in the low-data configuration; scales further with more data and richer signals.
Relationship to Other Netflix Systems¶
| System | Relationship |
|---|---|
| systems/netflix-genpage | Sibling generative recommender — GenPage generates the full homepage via autoregressive decoding; GenRec scores the full catalog via a scoring head (no decoding). |
| systems/netflix-ranker | The production baseline GenRec is benchmarked against — thousands of hand-engineered features, custom architectures. |
| systems/netflix-mss | Serving substrate — GenRec is deployed on MSS via vLLM in prefill-only mode. |
| systems/vllm | Inference engine — provides the serving runtime (no autoregressive decoding needed). |