Netflix is testing a different way to decide what viewers should see next. Instead of relying only on a long-running recommendation engine built around thousands of hand-crafted signals, the company has built GenRec, a language-model-based system designed to understand viewing behavior through text.
The result, according to Netflix, is not a full replacement for its existing system. But it is a meaningful signal: a general-purpose language model, trained and constrained for recommendations, can compete with a highly tuned production system while using much less labeled data in one key training phase.
Why Netflix Is Looking Beyond Hand-Built Logic
Netflix's current recommendation system depends on thousands of hand-crafted features covering users, titles, and interactions. That kind of engineering can be powerful, especially when it has been refined over many years. It also creates complexity.
The challenge becomes sharper when Netflix wants to support newer kinds of content or recommendation surfaces. The source article names games, live formats, and podcasts as examples of areas where the older approach can become expensive to extend.
Off-the-shelf language models do not solve the problem on their own. Netflix found that general models can lean too heavily toward popular titles, produce titles that are not actually in the catalog, and miss business rules that matter for a real recommendation product.
GenRec is Netflix's attempt to sit between those two extremes. It uses language-model capabilities, but it is trained on Netflix data and shaped for ranking real titles rather than generating free-form answers.
How GenRec Turns Viewing History Into Text
The core design choice is simple to explain: Netflix converts user behavior into plain text. Plays, watch durations, thumbs up or down, list additions, and drop-offs become a readable history for the model to analyze.
That changes the job of the recommendation model. Instead of requiring engineers to predefine every useful feature, the model can identify patterns in the sequence of user behavior. It may infer genre preferences or notice that a person's interests have shifted, without those patterns being spelled out in a rigid feature list.
Netflix still has to be selective. A full text record of every interaction would exceed the model's context window, so the system filters the input. High-signal activity, such as long watch sessions, stays in detail. Brief taps and quick scrolls can be removed, while binge sessions can be condensed.
Another guardrail addresses a common language-model weakness: hallucination. Netflix adds a separate component that only scores real catalog entries, so the recommendation process is tied to titles that actually exist in the service.
The Training Approach Behind GenRec
Netflix trains GenRec in two stages. First, an unnamed open-weight language model is fine-tuned on Netflix data so it can understand the company's catalog and user behavior. Then a second specialized training stage turns that base model into a recommendation ranker.
That second stage is updated more frequently. The reason is practical: new titles arrive and user preferences change. A recommendation model that performs well at one point can become less useful if it no longer reflects what is currently available or what viewers now want.
GenRec runs on vLLM in a mode where the model reads the input once and scores all matching candidates in a single pass. It does not generate text during that process. That design helps keep the cost of using a language model for recommendations manageable.
What Netflix Says the Tests Showed
Netflix compared GenRec with its production recommendation system, which has been tuned over many years. In offline testing, GenRec delivered about 1.6 percent better ranking quality. In the second training stage, it needed roughly 40 times fewer labeled examples to reach that result.
The source article is careful about the scope of that data point. The roughly 40 times reduction applies to that specific training phase, not to every part of the overall training process.
Netflix also ran a four-week A/B experiment on about ten percent of its traffic. The test was limited to recommendation surfaces that get pre-computed. In that live setting, a short-term metric tracking user behavior on the home screen improved by 0.115 percent, while a long-term core metric improved by 0.006 percent.
According to Netflix, both gains were large enough that they could not be explained by chance. The numbers are small, but in a mature recommendation system, even small measured gains can matter because the baseline is already heavily optimized.
The specialized second training phase also appears important. Netflix found that recommendation-specific fine-tuning 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 it no longer captures new titles and changed preferences as well.
From Feature Engineering to Context Engineering
GenRec points to a broader shift in how recommendation systems may be built. Instead of creating separate custom architectures for every recommendation task, Netflix sees room for a single language model to support multiple use cases.
That does not remove engineering work. It changes where the work happens. The emphasis moves from manually building more features to deciding which signals should enter the model's context, how those signals should be represented, and how much history should be included.
The infrastructure changes too. The source article describes a move toward GPU servers and LLM tooling. That fits the broader technical direction implied by GenRec: recommendation systems become less like narrowly tailored ranking stacks and more like controlled deployments of large language models.
Netflix describes GenRec as "an early but promising step" and as a strong alternative to traditional recommendation models. At the same time, a full replacement of the existing system is not currently on the table.
The company has used machine learning outside recommendation lists before. In 2020, Netflix described work with knowledge graphs and similarity maps to predict which content category a planned title fits into and what audience numbers it might reach in each country. At that time, Google's BERT language model processed human-written title summaries and sent machine-readable representations to downstream models.
Netflix has also built its own models for production workflows and has sometimes released them publicly, including the VOID framework for removing objects from video. GenRec adds another example of the company applying machine learning to core media operations, this time by testing whether language models can reshape how recommendations are ranked.