Netflix Tests AI Language Model To Replace Hand-Built Recommendation Logic

Netflix is testing a new recommendation system called GenRec that uses a language model instead of relying primarily on the thousands of hand-crafted features that have powered its recommendation technology for years. The early results suggest the approach can improve recommendation quality while requiring substantially less labeled training data than the existing system.

The experiment represents a potentially important change in how large consumer platforms build recommendation engines. Rather than manually designing separate features to describe users, titles and interactions, Netflix is converting viewing behavior into plain text and allowing a fine-tuned language model to identify patterns. Both offline testing and a live A/B experiment showed statistically significant improvements, although Netflix describes GenRec as an early-stage system rather than a replacement for its production recommendation infrastructure.

Netflix GenRec At A Glance

Particular	Details
Company	Netflix
Recommendation system	GenRec
Technology	Language model
Model type	Fine-tuned open-weight model
Primary input	User viewing behavior converted to text
Existing approach	Thousands of hand-crafted features
Offline improvement	~1.6% better ranking quality
Labeled-data requirement	~40x fewer examples in Phase 2
Live A/B test	~10% of traffic
A/B test duration	4 weeks
Short-term metric improvement	0.115%
Long-term core metric improvement	0.006%
Serving infrastructure	vLLM
Output	Scores for catalog candidates
Current status	Early-stage test, not full replacement

The figures come from Netflix’s own technical work and reporting on the GenRec experiment.

Why Netflix Is Rethinking Its Recommendation System

Netflix’s traditional recommendation infrastructure has been developed and refined over many years.

It relies on a large number of manually engineered features describing users, titles and interactions. Those features help the system determine what content a particular subscriber is likely to watch.

The problem is that such systems can become increasingly complicated as the company adds new types of content and new places where recommendations appear.

Netflix says the complexity makes it more expensive to onboard new content types such as games, live formats and podcasts, as well as expand recommendations into new areas of its interface.

Traditional Recommendation Model Vs GenRec

Area	Traditional System	GenRec
User representation	Hand-crafted features	Natural-language history
Feature creation	Extensive manual engineering	Reduced feature engineering
Content representation	Structured features	Language-model representation
New use cases	Can require new engineering	Potentially more adaptable
Model	Custom recommendation architecture	Fine-tuned language model
Training	Large labeled datasets	Much smaller Phase 2 dataset
Ranking	Specialized pipeline	Language model scores candidates

The goal is not simply to replace one algorithm with another. Netflix is testing whether a general-purpose language model can reduce the amount of specialized engineering required to build recommendation systems.

What Is Netflix’s GenRec?

GenRec is a language-model-based recommendation system developed internally by Netflix.

The system takes user behavior and converts it into natural-language text.

Instead of giving the model a large collection of manually designed numerical features, Netflix describes interactions in a textual format that the model can interpret.

User Behavior
     ↓
Viewing History
     ↓
Likes / Dislikes
     ↓
Watch Duration
     ↓
List Additions
     ↓
Drop-Offs
     ↓
Natural-Language Context
     ↓
GenRec
     ↓
Ranked Titles

The approach allows the model to identify patterns in behavior without Netflix explicitly encoding every possible preference as a manually engineered feature.

Netflix Turns Viewing History Into Plain Text

One of GenRec’s most interesting changes is how it represents user behavior.

Netflix converts activities such as plays, watch duration, thumbs-up or thumbs-down ratings, list additions and abandonment into a textual representation.

This effectively creates a behavioral narrative that the language model can process.

Signals GenRec Can Consider

User Signal	Example
Play	User started a title
Watch duration	User watched most of an episode
Rating	User gave positive feedback
List addition	User saved a title
Drop-off	User stopped watching early
Viewing sequence	Titles watched over time
Recent activity	Latest interests
Changing behavior	Shifts in preferences

The model can then identify patterns such as recurring genre interests or changing preferences without Netflix needing to manually define every pattern.

Why Netflix Cannot Simply Feed Everything Into The Model

There is an obvious problem with converting a user’s entire viewing history into text: context length.

A Netflix subscriber can accumulate years of viewing activity.

If every interaction were included in a single prompt, the amount of information could quickly become too large and expensive to process.

Netflix therefore filters the behavioral history.

High-signal events receive greater attention, while low-value interactions can be removed or compressed.

Complete Viewing History
          ↓
      Filtering
          ↓
 ┌────────┼────────┐
 ↓        ↓        ↓
High     Medium    Low
Signal   Signal    Signal
 ↓        ↓        ↓
Keep     Compress   Drop
          ↓
      Model Context

For example, a long viewing session can contain more useful information than a brief tap or quick scroll.

Netflix Uses Context Engineering Instead Of Feature Engineering

This is one of the broader ideas behind GenRec.

Traditional recommendation systems often require engineers to determine which features should be created and how they should be combined.

GenRec shifts part of that responsibility toward context engineering.

Instead of asking:

Which feature should represent this behavior?

The system can ask:

Which pieces of behavioral context should be included in the model’s input?

Feature Engineering Vs Context Engineering

Traditional Approach	GenRec Approach
Create numerical features	Create textual context
Manually define relationships	Model learns relationships
Add new features for new behaviors	Modify context representation
Specialized architectures	General-purpose language model
Engineering-heavy	Context-selection-heavy

Netflix describes this as part of a broader movement toward LLM-native recommendation systems.

GenRec Uses Two Training Stages

Netflix trains GenRec in two stages.

First, an open-weight language model is fine-tuned using Netflix’s data so that it understands the company’s catalog and user behavior.

A second training stage then specializes the model as a recommendation ranker.

Netflix can update this second stage more frequently because user preferences and the content catalog change quickly.

Open-Weight Language Model
          ↓
      Phase 1
          ↓
Netflix-Specific
Model Adaptation
          ↓
      Phase 2
          ↓
Recommendation
Ranking Fine-Tuning
          ↓
        GenRec

This separation allows Netflix to maintain a general base model while updating the recommendation-specific component more frequently.

Phase 2 Needs Far Less Labeled Data

One of the strongest results from the experiment was the amount of training data required.

GenRec achieved approximately 1.6% better ranking quality offline than Netflix’s production system while requiring roughly 40 times fewer labeled examples in its second training phase.

The comparison is specifically about the Phase 2 training data and should not be interpreted as meaning GenRec used 40 times less data across the entire development process.

Offline Test Results

Metric	GenRec Result
Ranking-quality improvement	~1.6%
Phase 2 labeled data requirement	~40x lower
Comparison	Existing production ranking model
Test type	Offline evaluation

A 1.6% improvement may sound modest, but recommendation systems operate at enormous scale, so relatively small improvements can potentially influence a large number of user interactions.

Netflix Also Tested GenRec With Real Users

Netflix did not stop at offline testing.

The company conducted a four-week A/B experiment involving approximately 10% of traffic on recommendation surfaces that use pre-computed recommendations.

The test produced measurable improvements in both short- and long-term metrics.

Live A/B Test Results

Metric	Improvement
Traffic included	~10%
Test duration	4 weeks
Short-term home-screen metric	+0.115%
Long-term core metric	+0.006%
Statistical significance	Netflix says both were significant

The improvements were relatively small, but Netflix says both were too large to be explained by chance.

GenRec
  ↓
10% Netflix Traffic
  ↓
4-Week A/B Test
  ↓
Short-Term Metric
+0.115%
  +
Long-Term Metric
+0.006%
  ↓
Statistically Significant
Improvement

Why Small Recommendation Gains Matter

Recommendation systems influence what users see when they open Netflix.

A small increase in the likelihood that users find something interesting can affect viewing behavior across millions of subscribers.

This makes recommendation quality an important part of Netflix’s product experience.

The company has spent years optimizing this area, which makes a measurable improvement from a new architecture particularly notable.

GenRec Still Needs A Separate Catalog Component

Language models have an important weakness when used directly for recommendation.

A general-purpose model might suggest a title that sounds plausible but does not actually exist in Netflix’s catalog.

Netflix therefore uses a separate component that scores only real catalog candidates.

User Context
     ↓
GenRec
     ↓
Recommendation Scores
     ↓
Real Netflix Catalog
     ↓
Candidate Ranking
     ↓
Final Recommendations

This prevents the model from simply generating fictional titles and ensures the ranking system operates over actual content available in Netflix’s catalog.

Netflix Avoids Generating Text During Ranking

Another important technical decision helps keep GenRec practical.

The system runs on vLLM in prefill-only mode.

Instead of generating a natural-language response, the model reads the user context once and produces scores for candidate titles.

This allows Netflix to use the language model for ranking without paying the computational cost of generating long text responses.

Traditional LLM Generation Vs GenRec

LLM Operation	GenRec
Reads context	Yes
Generates text	No
Scores candidates	Yes
Uses prefill-only processing	Yes
Output	Recommendation scores
Primary objective	Ranking

This distinction is crucial because a recommendation system must process large numbers of candidates efficiently.

Recommendation Models Become Outdated Quickly

Netflix’s experiment also highlights an important problem with recommendation systems: staleness.

User interests change.

New movies and shows arrive.

Old titles become less relevant.

A base model trained on older information may therefore lose accuracy over time.

Netflix found that the recommendation-specific Phase 2 fine-tuning added 35% to 50% on top of the base model’s performance. When the base model was already two weeks old, the difference widened to roughly 80%, because it no longer reflected new titles and changing preferences.

Model Freshness Effect

Model State	Phase 2 Advantage
Recently updated base model	+35%-50%
Base model ~2 weeks old	~80%
Reason	New content + changing user preferences

This shows why recommendation systems need frequent updates even when the underlying language model itself remains unchanged.

AI Could Make Recommendation Systems More Flexible

Netflix has historically built specialized systems for different recommendation problems.

A language-model-based approach could potentially make it easier to reuse a common model across different recommendation surfaces and content categories.

             GenRec
               ↓
     ┌─────────┼─────────┐
     ↓         ↓         ↓
   Movies     Shows     Games
     ↓         ↓         ↓
   Live      Podcasts   New Formats

The company believes this approach could support a broader shift toward general-purpose models instead of custom architectures for every individual recommendation task.

New Content Types Could Become Easier To Add

Netflix has increasingly expanded beyond conventional movies and television.

Its product ecosystem can include games, live programming and other forms of content.

Traditional recommendation systems may require engineers to create new features for each category.

A more general language-model architecture could potentially reduce that burden.

Potential Expansion Areas

Content Type	Recommendation Challenge
Movies	Genre, actors, themes
TV shows	Series progression
Games	Different engagement patterns
Live content	Time-sensitive relevance
Podcasts	Different consumption behavior
New formats	Limited historical data

This flexibility could become increasingly valuable as Netflix expands its entertainment ecosystem.

GenRec Does Not Mean Netflix Is Abandoning Its Existing System

Despite the promising results, Netflix is not replacing its entire recommendation infrastructure with GenRec.

The company describes GenRec as an “early but promising step” and a strong alternative to traditional recommendation models.

This is important because production recommendation systems have complex requirements beyond raw ranking accuracy.

They must handle:

• Reliability
• Latency
• Cost
• Catalog constraints
• Business rules
• Freshness
• Large-scale traffic
• Experimentation

Netflix’s existing infrastructure has been optimized around these requirements for years.

Off-The-Shelf Language Models Are Not Enough

Netflix’s experiment also demonstrates that simply putting a general-purpose AI model in charge of recommendations does not automatically solve the problem.

The company says standard language models can:

• Over-recommend popular content
• Hallucinate titles
• Ignore business rules
• Fail to optimize for recommendation-specific objectives

GenRec is therefore not simply an off-the-shelf chatbot being connected to Netflix’s catalog.

It is specifically adapted to Netflix data and then trained for ranking.

Why A General Chatbot Is Not A Recommendation Engine

General LLM Problem	GenRec Solution
Popularity bias	Recommendation-specific training
Fake titles	Real-catalog scoring component
Generic responses	Ranking objective
Business-rule limitations	Dedicated recommendation architecture
Outdated preferences	Frequent Phase 2 updates

The distinction could become important for other companies considering LLM-based recommendation systems.

The Experiment Could Reduce Engineering Complexity

One of the most significant long-term possibilities is reducing the amount of specialized engineering needed to maintain recommendation systems.

Instead of continuously creating new features, teams could focus more on:

• Selecting useful behavioral signals
• Designing model context
• Updating training data
• Controlling candidate sets
• Monitoring performance

This could change the role of recommendation engineers.

Traditional
Feature Engineering
       ↓
Many Custom Features
       ↓
Specialized Models
       ↓
Separate Systems


GenRec Approach
       ↓
Context Engineering
       ↓
General Model
       ↓
Task-Specific Fine-Tuning
       ↓
Multiple Recommendation Uses

Netflix sees this as part of a wider transition from custom-built architectures toward general-purpose language models.

Netflix Is Not New To AI-Powered Recommendations

Netflix has used machine learning and recommendation technology for years.

Its recommendation research covers areas including personalization and large-scale ranking systems. The company’s long history in this field means GenRec represents an evolution of an established AI infrastructure rather than Netflix suddenly adopting AI for the first time.

Netflix previously used technologies such as knowledge graphs and similarity maps to understand content and audiences, while earlier applications of language models focused on specific parts of the recommendation pipeline rather than replacing the core ranking logic.

The Bigger Picture

Netflix’s GenRec experiment shows how large technology companies are beginning to use language models for tasks that traditionally depended on highly specialized machine-learning architectures. Instead of manually creating thousands of features describing users, titles and interactions, Netflix converts viewing behavior into natural-language context and uses a fine-tuned language model to rank real catalog titles. The approach produced about a 1.6% improvement in offline ranking quality while requiring roughly 40 times fewer labeled examples in the second training phase.

The live experiment is arguably more important. Netflix tested GenRec on around 10% of its traffic for four weeks and reported a 0.115% improvement in a short-term home-screen metric and a 0.006% improvement in a long-term core metric, with both results described as statistically significant. The gains are small, but recommendation systems operate at massive scale, making even incremental improvements potentially meaningful.

The experiment also highlights the limits of simply applying general-purpose AI to recommendation problems. Netflix still needs mechanisms to ensure only real catalog titles are scored, manage context length, account for business rules and keep recommendation-specific training current. GenRec is therefore better understood as a new architecture for recommendation rather than a simple replacement of Netflix’s existing system.

Looking Ahead

Netflix’s experiment could mark an important step toward LLM-native recommendation systems, where general-purpose language models handle a wider range of personalization tasks. If the approach scales successfully, recommendation teams may spend less time designing individual features and specialized architectures and more time engineering context, maintaining fresh training data and controlling how models interact with real-world catalogs.

For now, Netflix is not planning a wholesale replacement of its established recommendation infrastructure. GenRec remains an early-stage system, but its combination of lower labeled-data requirements, measurable offline gains and statistically significant improvements in a live experiment gives Netflix a strong reason to continue testing the approach. The broader implication is that language models may increasingly become the underlying intelligence for recommendation systems across entertainment, commerce and other consumer platforms—not because they eliminate traditional recommendation technology overnight, but because they can potentially make those systems more flexible and easier to adapt