Netflix tests language model as alternative to hand-built recommendation logic

Netflix pitted its years-old recommendation engine against a language model and says it got better results. The system, called GenRec, needed a fraction of the labeled training data the old system required.

Netflix's current recommendation system relies on thousands of hand-crafted features about users, titles, and interactions, according to a blog post from the Netflix tech team. That complexity makes it expensive to onboard new content types like games, live formats, or podcasts, and to expand into new areas of the Netflix interface. But off-the-shelf language models aren't ready for recommendations either. They over-index on popular content, hallucinate titles that don't exist in the catalog, and ignore business rules.

GenRec is designed to fill that gap. Netflix trains a proprietary model in two stages. First, an unnamed open-weight language model gets fine-tuned on Netflix data so it understands the catalog and user behavior. Then a second round of specialized training turns that base model into a recommendation ranker. This second stage gets updated more often to account for new titles and shifting preferences.

Watch history becomes plain text

Instead of encoding user data as dense numerical vectors, Netflix converts it into plain text. Plays, watch durations, thumbs up or down, list additions, and drop-offs become a kind of dialogue between the user and the recommendation system. The model picks up on patterns like genre preferences or shifting interests on its own, rather than having them spelled out through manually engineered features.

A full text version of every interaction would blow past the model's context window, so Netflix filters aggressively. High-signal events like long watch sessions stay in full detail, while brief taps or quick scrolls get dropped and binge sessions get condensed. To keep the model from suggesting titles that don't actually exist, Netflix adds a separate component that only scores real catalog entries.

GenRec runs on vLLM in a mode where the model reads the input once and scores all candidates in a single pass without generating any text. That keeps costs manageable.

Small gains, but statistically solid

Compared to the production system that has been tuned over many years, GenRec delivered about 1.6 percent better ranking quality offline. It needed roughly 40 times fewer labeled examples in the second training stage to get there. That comparison applies to this specific phase, not to all training data.

For the online test, Netflix ran a four-week A/B experiment on about ten percent of its traffic, limited to recommendation surfaces that get pre-computed. A short-term metric tracking user behavior on the home screen rose by 0.115 percent, and a long-term core metric improved by 0.006 percent. Both gains are too large to be explained by chance, according to Netflix.

The recommendation-specific fine-tuning in Phase 2 adds another 35 to 50 percent on top of the base model's performance. When the base model is already two weeks old, that gap widens to about 80 percent, because the base model no longer reflects new titles and changed preferences. Recommendation models go stale fast.

Context engineering replaces feature engineering

Netflix sees GenRec as part of a broader shift that's also showing up in work like PLUM, GLIDE, and OneRec-Think. Instead of building custom architectures for every recommendation task, a single language model handles multiple use cases. The work shifts from building ever more features to deciding which signals belong in the model's input and how much of them to include. Infrastructure is moving toward GPU servers and LLM tooling as well.

The Netflix team calls GenRec "an early but promising step" and describes the system as a strong alternative to traditional recommendation models. A full replacement of the existing system isn't on the table yet.

Netflix has been using machine learning beyond its recommendation lists for years. Back in 2020, the company described how knowledge graphs and similarity maps predict which content category a planned title fits into and what audience numbers it might reach in each country. At the time, Google's BERT language model only processed human-written title summaries and fed machine-readable representations to downstream models. Netflix has also started building its own models for production workflows and sometimes releases them publicly, like the VOID framework for removing objects from video.