Netflix’s recommendation stack rests on thousands of hand-crafted features, yet a language model has now edged past it by 1.6 percent on offline ranking quality — while consuming roughly 40 times fewer labeled examples during the second training stage.
The accuracy bump is arguably the less interesting figure. What stands out is that the production system GenRec was benchmarked against has absorbed years of Netflix tuning.
Why the old system got expensive
Per a blog post from Netflix’s tech team, the system in production today depends on thousands of manually engineered features spanning users, titles and interactions. Maintaining that pile of hand-built logic makes it expensive to onboard new content formats — games, live programming, podcasts — and expensive to extend recommendations into fresh corners of the Netflix interface.
Swapping in an off-the-shelf language model isn’t a fix, though. Stock models skew heavily toward popular content, invent titles that aren’t in the catalog, and pay no attention to business rules.
Two rounds of training, one of them disposable
Training happens in two phases. First, an open-weight language model — Netflix doesn’t name it — is fine-tuned on the company’s own data so it picks up the catalog and user behavior patterns. A second, specialized training pass then turns that base into a recommendation ranker.
Phase two is rerun far more frequently, since new titles arrive constantly and tastes keep shifting.
Your viewing history, rewritten as a conversation
This is where Netflix departs from the standard playbook. Instead of packing user data into dense numerical vectors, the company converts it into plain text.
Plays, how long you watched, thumbs up or down, additions to your list, abandoned titles — all of it is framed as a sort of dialogue between the viewer and the recommender. Genre preferences and drifting interests are something the model infers by itself, rather than something engineered features have to spell out.
Spelling out every single interaction would blow past the context window, so Netflix prunes aggressively. Lengthy viewing sessions are preserved in full; fleeting taps and quick scrolls are thrown out, and binges are compressed.
Hallucination is handled by a separate component that only ever scores entries that genuinely exist in the catalog.
The efficiency trick that keeps the bill down
Serving happens through vLLM, in a mode where the input is read a single time and every candidate is scored in one pass with no text generated at all. Nothing is emitted token by token, so inference costs don’t spiral.
What happened when real subscribers got it
The company ran an A/B test lasting four weeks across roughly ten percent of its traffic, limited to recommendation surfaces that are computed in advance.
On the home screen, a short-term behavioral metric climbed 0.115 percent, while a long-term core metric moved up 0.006 percent. Modest figures, certainly — but Netflix maintains that both are too large to write off as chance.
One caveat bears restating alongside the 1.6 percent offline improvement and the 40x reduction in data: that comparison covers the second training phase specifically, not the entire training corpus.
Models go stale in two weeks
Phase 2’s recommendation-specific fine-tuning contributes another 35 to 50 percent on top of what the base model delivers. Leave that base model untouched for a fortnight and the gap stretches to something like 80 percent, simply because it’s unaware of new titles and shifted preferences.
Where this fits in the wider shift
GenRec is framed by Netflix as part of the same family as PLUM, GLIDE and OneRec-Think. All of them chase one goal: a single language model serving several recommendation use cases, in place of a bespoke architecture for every task.
The nature of the engineering shifts as well. Less effort goes into dreaming up features, more into judging which signals deserve a spot in the model’s input and in what proportion. Infrastructure, meanwhile, tilts toward GPU servers and LLM tooling.
Netflix has been here before
Machine learning at the company reaches far beyond the rows on your homepage. Back in 2020, Netflix explained how knowledge graphs and similarity maps are used to predict the content category a planned title belongs to, along with the audience figures it might pull in each country.
At the time, Google’s BERT language model covered only a narrow portion of that job — reading human-written title summaries and handing machine-readable representations off to downstream models. The company has since begun developing its own models for production workflows, sometimes putting them out publicly, as it did with VOID, a framework for erasing objects from video.
Don’t expect your homepage to change yet
Netflix’s team labels GenRec “an early but promising step” and characterizes it as a strong alternative to conventional recommendation models. Retiring the existing system outright isn’t being considered.
For a sense of how much weight Netflix puts behind this, look at the retraining cadence instead of the accuracy charts. A ranker that has to be retrained every couple of weeks just to stay useful is one the company has backed with real infrastructure.