GenRec (Netflix)

GenRec is an experimental large language model (LLM) backed recommendation ranker built at Netflix. It turns a member's viewing history, item metadata and request context into text, feeds that text to a Netflix-adapted foundation LLM, and scores the full title catalog in a single forward pass through a catalog-aware ranking head. Netflix described the system in a Netflix Tech Blog post on 30 July 2026 and in the arXiv paper "GenRec: An LLM-Backed Recommendation Ranker at Netflix" (arXiv 2608.10257, v1 10 August 2026, v2 21 August 2026).^[1][4] The authors frame it as an exploration: the abstract says Netflix is "exploring this direction through GenRec", and the paper calls the system "our first step toward transitioning from a traditional recommendation stack to an LLM-native system."^[2]

In a four-week online A/B test on about 10% of Netflix traffic, GenRec produced statistically significant gains over Netflix's mature production ranker on short-term and long-term online metrics, while being trained on far fewer Phase-2 labeled examples and input signals. Offline, a GenRec model trained with about 40 times fewer Phase-2 labeled examples than the production model improved Mean Reciprocal Rank (MRR) by about 1.6% relative.^[2] The work drew wide attention in late September 2026 after a viral post on X claimed that Netflix had "replaced" its recommendation algorithm with an LLM, a framing the paper does not support.^[9][2]

GenRec should not be confused with an unrelated 2023 academic paper of the same name, "GenRec: Large Language Model for Generative Recommendation" by Jianchao Ji and co-authors (arXiv 2307.00457).^[8]

Overview

Background: the production ranker

According to the paper, Netflix's production recommender systems generate personalized title recommendations from a "large number of hand-crafted features spanning user, item, and interaction signals, along with extensive higher-order feature interactions," and use customized architectures such as feature-interaction networks, transformer-based sequence models and multi-task learning architectures.^[2] The accompanying blog post says the baseline ranker "relies on thousands of engineered dense and embedding features, as well as custom architectures for modeling feature interactions and sequences."^[4] The paper describes this baseline as "a mature discriminative model tuned over many years."^[2]

The authors say the problem with this stack is not quality but the cost of change. Onboarding a new content type or recommendation surface "often requires substantial work in feature engineering, model architecture design, infrastructure changes, and experimentation," which makes it hard to keep pace with the rapid growth of business needs at Netflix, whose recommenders already span movies, series, games, live events and podcasts.^[2]

They also list why an off-the-shelf LLM is not a usable recommender: such models "tend to over-recommend globally popular content, hallucinate out-of-catalog titles, ignore nuanced business constraints, and offer limited personalization."^[2] GenRec's design addresses each of these points in turn: post-training on Netflix data for personalization, a catalog-constrained scoring head against hallucination of titles, and reward weighting for business constraints.

Authors and versions

The arXiv v1 PDF lists six authors: Ying Li, Shradha Sehgal, Arjun Rao, Rein Houthooft, Yaochen Zhu and Ashish Rastogi.^[3] The v2 PDF lists twelve, adding Yunan Hu, Sourabh Medapati, Yun Li, Linas Baltrunas, Grace Huang and Kamelia Aryafar.^[2] The arXiv abstract page still showed the original six names as of 30 September 2026.^[1] Sehgal and Houthooft are marked "Work done while employed at Netflix"; Houthooft's listed affiliation is Amazon.^[2] The Netflix Tech Blog post is credited to Ying Li, Arjun Rao and Shradha Sehgal.^[4]

The paper's acknowledgments list 39 contributors across three groups: AI for Members (22 names, including Justin Basilico and Moumita Bhattacharya), AI Platform and Serving (13 names) and Product (4 names).^[2]

Two-phase framework

GenRec is trained in two phases.^[2]

• Phase 1 trains a foundation LLM "on top of OSS models with proprietary Netflix data to develop deep understanding of Netflix user and content." It balances broad capabilities such as world knowledge, personalization, content understanding and language ability. The blog post describes Phase 1 as a shared, Netflix-aware backbone for many applications that is updated relatively infrequently.^[2][4]
• Phase 2 post-trains this foundation model on ranking-specific data and objectives. The result is GenRec itself. The paper focuses on this phase.^[2]

The paper separates the two phases along three axes:^[2]

The paper names neither the open-source model that Phase 1 starts from nor the size of the Phase-1 model, and it does not describe the Phase-1 training data beyond "proprietary Netflix data."^[2]

Post-training data

Netflix members generate "hundreds of billions of interaction events" across movies, shows, games, live events and other content, with signals such as viewing, playing, thumbs ratings and adding to a list.^[2] For Phase 2, Netflix converts a member's history into a single-turn or multi-turn "conversation" between a user and a recommender. Each conversation carries:^[2]

The blog post adds that at inference time Netflix does not decode assistant messages; the conversational format mainly supports the language-modeling objective during training.^[4]

Input verbalization and context engineering

Instead of hand-crafted features or dense embeddings, GenRec "verbalizes rich user interaction histories and context as natural language or lightly structured text" and relies on the LLM to learn item relationships, temporal dynamics and shifting interests.^[2] Because a member's full history can exceed the token budget, the team applies what it calls context engineering, citing Anthropic's 2025 engineering post on the topic.^[2] The rules are:^[2]

Within the fixed budget, short- to medium-term history is kept at higher granularity, and older history is omitted or compressed into a brief user-interest summary.^[2] The blog post adds that the prompt is structured "to maximize shared prefixes for better prefix caching."^[4] The paper's pipeline figure uses an illustrative verbalization that, according to its footnote, "does not reflect actual user data in production."^[2]

Training objectives

Phase-2 training combines two main objectives:^[2]

1. Catalog-aware ranking objective. Positive labels are high-quality engagements (for example, long plays or strong explicit feedback), with denoising logic and thresholds that vary by content type. These labels feed a cross-entropy loss over the item catalog or candidate set.
2. Language-modeling objective. A next-token loss over the verbalized inputs and outputs, such as predicted titles, preserves the backbone's language ability. The authors say this matters for reading natural-language histories and for "recommendation steering via prompts."

The combined loss is a weighted sum, L = α·L_ranking + β·L_language + γ·L_miscellaneous, with α + β + γ = 1 and the weights tuned offline.^[2] At inference Netflix currently uses only the ranking output; the language objective is kept to improve ranking, preserve responsiveness to prompt-based steering and "keep open the possibility of future text generation use cases (e.g., natural-language explanations)."^[2]

Reward signals

Training on raw interaction logs alone, the paper says, risks a model that over-recommends binge-watching over discovery, favors videos over games, or chases immediate clicks at the expense of exploration and retention.^[2] GenRec therefore folds in outputs from separate reward models drawn from an existing reward-modeling framework (the paper cites Tang et al., "Reward innovation for long-term member satisfaction," RecSys 2023). The rewards fall into two groups:^[2]

• Long-term satisfaction proxies, which estimate how strongly a short-term engagement correlates with long-term outcomes such as returning to the service, exploring more of the catalog, or sustained engagement.
• Behavior rebalancing, which adjusts across content types (movies, games, live, podcasts) and launch stages (pre-launch, newly launched, evergreen) to meet goals such as exposure fairness across content types.

Each training example receives a scalar weight derived from these signals, and its ranking loss is scaled by that weight.^[2] The authors chose this reward-weighted loss over full reinforcement learning for simplicity, stability and cost. In preliminary experiments, RL-style methods such as GRPO "showed additional gains over supervised finetuning," but their training overhead led the team to leave them for future work.^[2]

Model architecture

GenRec's backbone "closely follows the foundational LLM: a decoder-only Transformer trained with next-token-prediction style objectives," augmented with a catalog-aware ranking head.^[2] Scoring runs in three stages:^[2]

1. Verbalization. A verbalizer maps the interaction history, context and item metadata into one text sequence.
2. Pooled representation. The LLM encodes the sequence and takes the hidden state at a pooling position as a d-dimensional summary of the user's preferences and context.
3. Catalog-aware scoring. A scoring head combines that hidden state with a learned d-dimensional embedding for each item. The blog post gives a dot product or a small MLP as examples.^[4]

The LLM weights, scoring head and item embeddings are trained jointly. A softmax over the catalog turns scores into a probability distribution, from which the ranking is derived; for catalogs too large to score exhaustively, sampled softmax can be used.^[2] Because only items with learned embeddings can be scored, the model cannot recommend titles outside the Netflix catalog.^[2]

The paper contrasts this with most LLM-based generative retrieval systems, which use autoregressive decoding with beam search over item identifiers and pay a latency cost that "can be prohibitive at scale, especially for large candidate sets."^[2]

Serving and cost

GenRec runs on Netflix's internal LLM serving stack using vLLM.^[2] The paper states that inference cost "is roughly proportional to the product of model size and context length," and pursues three levers:^[2]

• Smaller and distilled models trained on larger or more targeted datasets (the paper cites MobileLLM-R1 in this context), aiming to recover much of a larger model's quality at lower per-request cost. See knowledge distillation.
• Context compaction through the verbalization rules above.
• Prefill-only inference. Although the backbone can decode autoregressively, Netflix deploys GenRec so that "the model consumes the input context once, and produces ranks for the full candidate set in a single forward pass." The authors say this is what makes it feasible to serve high-volume workloads within their compute budget.

Netflix's separate Tech Blog post on its in-house LLM serving platform (17 July 2026) describes the underlying stack: vLLM chosen as the "paved-path engine" (replacing TensorRT-LLM) after a re-benchmark prompted by workload changes the post dates to summer 2025, running inside NVIDIA Triton Inference Server behind a shared Model Scoring Service. That post lists "prefill-only inference for ranking and retrieval" among the workloads that drove the engine choice, without naming GenRec.^[5]

Results

Comparison with the production ranker

Offline, GenRec beat the production model on MRR while using less Phase-2 training data and fewer input signals. The paper states: "with about 40× less Phase-2 labeled training examples than the production model, GenRec achieves an offline lift of approximately +1.6% relative in MRR over the baseline," and offline metrics kept improving as training data and input signals were scaled up.^[2] The blog post puts the data advantage as a range: starting from a strong Phase-1 model, GenRec matches or exceeds the production ranker with 10 to 40 times fewer Phase-2 labeled examples, depending on configuration.^[4]

Online, Netflix ran a large-scale A/B test "focusing on key batch-compute surfaces," allocating about 10% of Netflix traffic for 4 weeks. The GenRec model was tested "under a low-data, low-signals configuration."^[2] The paper's Figure 3 reports the average treatment effect against production:^[2]

The paper describes the +0.006% relative gain on the core online metric as "statistically meaningful at Netflix scale." It does not name the underlying metrics.^[2]

Data and model scaling

Netflix trained GenRec on Phase-2 datasets ranging from 1x to 20x its smallest configuration, for backbones of roughly 1B and 10B parameters. For both sizes, offline MRR improved monotonically with more data; the ~1B model scored lower in absolute terms but followed a similar curve.^[2] Under a fixed training budget (same GPU configuration and similar wall-clock time), larger backbones consistently achieved higher offline MRR.^[2] Because larger models and more data raise both training and serving costs, the team uses the scaling curves and context-length ablations to pick a "sweet spot" on the quality-cost Pareto frontier.^[2] The authors present this as evidence that scaling laws can guide recommender design.^[2]

Contribution of each phase

Source: paper Section 5.3 and Table 1.^[2] The authors attribute the growing Phase-2 gain partly to task-specific adaptation and partly to the Phase-1 backbone going stale as content popularity and member interests shift.^[2]

Context length

The team tuned context length in three steps: cleaning and compressing events, sweeping the number of retained engagement events to find an "elbow point" beyond which extra tokens add little, and testing prompts at different levels of detail.^[2] Context could be cut to "roughly a third of the original token budget (for example, from about 5,000 tokens down to about 1,700 tokens) with only negligible degradation in offline ranking metrics." Because GenRec is largely compute-bound, serving cost fell to roughly a third as well.^[2]

Discussion and stated limitations

The paper's discussion section argues that LLM-backed recommenders change recommender engineering in five ways: from feature engineering to context engineering ("The 'prompt' becomes the new feature vector"), from bespoke architectures such as two-tower and DLRM-style networks to shared foundation model backbones, toward scaling laws as a design guide, toward pre-training plus post-training and reward alignment, and from recommender infrastructure toward LLM infrastructure built on GPUs, vLLM or Triton, KV caching, prefix caching and prefill-only inference.^[2]

The paper is hedged in several places:

• The results "suggest that LLM-backed rankers are a viable and promising alternative to traditional models, at least on the recommendation surfaces tested."^[2]
• The online test covered batch-compute surfaces, and the conclusion describes a ranker "suitable for production traffic on batch-compute surfaces."^[2]
• The conclusion says the approach is viable "in at least some scenarios, provided we continue to manage cost and infrastructure complexity carefully and to evaluate quality trade-offs rigorously against strong non-LLM baselines."^[2]
• RL-based alignment, natural-language explanations and prompt-based steering are described as possibilities or future work, not as shipped features.^[2]

The paper also leaves out several details: it does not name the base open-source LLM, the online metrics, the absolute size of the training sets, or the size of the Netflix catalog scored at inference.^[2]

Relation to other work

Industry LLM recommenders. The paper places GenRec alongside Google's PLUM (adapting a pretrained LLM at YouTube scale through Semantic-ID tokenization), Spotify's GLIDE (podcast discovery as instruction following over a Semantic-ID catalog) and Kuaishou's OneRec-Think (OneRec with a Qwen3 backbone and reasoning steps).^[2] Those systems generate item identifiers autoregressively; GenRec instead verbalizes history as text and scores items discriminatively in one prefill pass.^[2] Devoteam, a consultancy, described this split in an explainer as two "schools": semantic-ID generative retrieval versus natural-language verbalization with discriminative scoring.^[10]

Netflix's recommendation foundation model. In March 2025 Netflix described a separate "Foundation Model for Personalized Recommendation," a transformer trained with next-token prediction over tokenized member interaction sequences (not natural language), with learnable item ID and metadata embeddings and a scaling-law plot of its own.^[6] The GenRec paper cites that post in its related-work section on generative recommenders built with special tokens, and treats GenRec's text-based approach as a different line of work.^[2]

GenPage. A month before the GenRec blog post, Lequn Wang, Jiangwei Pan and Linas Baltrunas posted "GenPage: Towards End-to-End Generative Homepage Construction at Netflix" (arXiv 2606.31031, 30 June 2026), a single transformer that autoregressively generates an entire multi-row homepage. They reported a lift on Netflix's core engagement metric and a 20% cut in end-to-end serving latency in online A/B tests.^[7]

Earlier recommender techniques. The paper positions GenRec against the feature-engineered lineage of large-scale recommenders, citing Koren, Bell and Volinsky's 2009 matrix factorization paper, Wide & Deep (Cheng et al., 2016), YouTube's deep neural network recommender (Covington et al., 2016) and DLRM (Naumov et al., 2019) as prior systems in line with Netflix's production stack.^[2] For the older history of Netflix recommendation research, see collaborative filtering and the Netflix Prize.

Reception

Coverage after the blog post and paper mostly summarized Netflix's own figures. StartupHub.ai reported on the blog post on 30 July 2026.^[12] ZenML's LLMOps Database called GenRec "a significant production deployment of an LLM-native recommendation system," a stronger framing than Netflix's own "first step" language.^[11][2]

On 28 September 2026 the X account @thesupermannx published a long post with a screenshot of the paper's first page, opening "Netflix replaced their 15 years old recommendation algorithm with an LLM." By 30 September it had about 361,000 views and 4,100 likes.^[9] Its main claims, checked against the paper and Netflix's blog post:

References

1. ^1 ^2 ^3 ^4 ^5"[2608.10257] GenRec: An LLM-Backed Recommendation Ranker at Netflix." arXiv. arxiv.org/...2608.10257
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35 ^36 ^37 ^38 ^39 ^40 ^41 ^42 ^43 ^44 ^45 ^46 ^47 ^48 ^49 ^50 ^51 ^52 ^53 ^54 ^55 ^56 ^57 ^58 ^59 ^60 ^61 ^62 ^63 ^64 ^65 ^66 ^67 ^68 ^69 ^70 ^71 ^72 ^73 ^74 ^75Li, Ying; Sehgal, Shradha; Rao, Arjun; Houthooft, Rein; Hu, Yunan; Zhu, Yaochen; Medapati, Sourabh; Li, Yun; Baltrunas, Linas; Huang, Grace; Rastogi, Ashish; Aryafar, Kamelia. "GenRec: An LLM-Backed Recommendation Ranker at Netflix." arXiv:2608.10257v2, 21 August 2026. arxiv.org/...2608.10257v2
3. ^1 ^2 ^3Li, Ying; Sehgal, Shradha; Rao, Arjun; Houthooft, Rein; Zhu, Yaochen; Rastogi, Ashish. "GenRec: An LLM-Backed Recommendation Ranker at Netflix." arXiv:2608.10257v1, 10 August 2026. arxiv.org/...2608.10257v1
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13Li, Ying; Rao, Arjun; Sehgal, Shradha. "GenRec: Towards LLM-Native Recommendation at Netflix." Netflix Technology Blog, 30 July 2026. netflixtechblog.com/...ion-at-netflix-f20be6f643e3
5. ^Netflix AI Platform Model Runtime and Inference teams. "In-House LLM Serving at Netflix." Netflix Technology Blog, 17 July 2026. netflixtechblog.com/...ing-at-netflix-a5a8e799ea2c
6. ^Hsiao, Ko-Jen; Feng, Yesu; Lamkhede, Sudarshan. "Foundation Model for Personalized Recommendation." Netflix Technology Blog, 21 March 2025. netflixtechblog.com/...recommendation-1a0bd8e02d39
7. ^Wang, Lequn; Pan, Jiangwei; Baltrunas, Linas. "GenPage: Towards End-to-End Generative Homepage Construction at Netflix." arXiv:2606.31031, 30 June 2026. arxiv.org/...2606.31031
8. ^Ji, Jianchao; Li, Zelong; Xu, Shuyuan; Hua, Wenyue; Ge, Yingqiang; Tan, Juntao; Zhang, Yongfeng. "GenRec: Large Language Model for Generative Recommendation." arXiv:2307.00457, 2 July 2023. arxiv.org/...2307.00457
9. ^1 ^2 ^3@thesupermannx. Post on X, 28 September 2026 (retrieved via the fxtwitter API, 30 September 2026). x.com/...2104579598105379159
10. ^"GenRec explained: Netflix's move to LLM-native recommendation." Devoteam. devoteam.com/...ix-genrec-native-llm-recommendtion
11. ^"Netflix: LLM-Native Recommendation System at Scale." ZenML LLMOps Database. zenml.io/...-native-recommendation-system-at-scale
12. ^Singer, Daniel. "Netflix Bets on LLMs for Smarter Recommendations." StartupHub.ai, 30 July 2026. startuphub.ai/...-llms-for-smarter-recommendations