Netflix just ran a real-world experiment pitting its meticulously crafted, years-old recommendation engine against a fine-tuned language model — and the LLM won. The system, called GenRec, delivered about 1.6 percent better ranking quality offline while needing roughly 40 times fewer labeled examples in its second training stage. That’s not a rounding error. It’s a signal that the era of hand-built recommendation architectures may be winding down.
The approach is genuinely different from what Netflix has done before. Instead of encoding user behavior as dense numerical vectors built from thousands of hand-crafted features, GenRec converts watch history into plain text. Plays, thumbs up, list additions, and drop-offs become a kind of dialogue between the user and the system. The model picks up on genre preferences and shifting interests on its own — no manual feature engineering required. Netflix does filter aggressively to fit context windows: high-signal events like long watch sessions stay detailed, brief taps get dropped, and binge sessions get condensed.
There’s a two-stage training pipeline. First, an unnamed open-weight model gets fine-tuned on Netflix data to understand the catalog. Then a second phase turns it into a ranker, and that phase gets updated frequently because recommendation models go stale fast. The team found the recommendation-specific fine-tuning adds 35 to 50 percent on top of the base model’s performance — and that gap widens to about 80 percent when the base model is just two weeks old. To prevent hallucinated titles, a separate component only scores real catalog entries. The system runs on vLLM in a mode that reads input once and scores all candidates in a single pass, keeping costs manageable.
In a four-week A/B test on about ten percent of traffic, a short-term home screen metric rose 0.115 percent and a long-term core metric improved 0.006 percent — both statistically significant. Those are modest gains, but the fact that a general-purpose language model can outperform a system tuned over many years, with far less data, is the real story. Netflix calls GenRec “an early but promising step” and isn’t replacing its production system yet. But this aligns with a broader shift seen in work like PLUM, GLIDE, and OneRec-Think: one model handling multiple recommendation tasks instead of custom architectures for each. The work shifts from feature engineering to context engineering — deciding which signals belong in the input and how much of them to include.
