KOEN
Read Original
#AI

GenPage: Towards End-to-End Generative Homepage Construction at Netflix

GenPage: Towards End-to-End Generative Homepage Construction at Netflix
01

Summary

Netflix's New Play: Generating the Entire Homepage with a Single Transformer

Inside GenPage—the end-to-end generative model that slashed latency by 20% and redefined recommendation architecture.

Netflix is moving beyond traditional ranking stacks to a generative paradigm where the homepage is 'written' like a language response. This article explores how they leveraged GenPage, a transformer-based model, to replace complex pipelines and optimize for whole-page satisfaction using Reinforcement Learning.

  • 01Consolidation of multi-stage recommender stacks into a single generative transformer
  • 02Use of a domain-specific custom tokenizer for efficient handling of user history and layouts
  • 03Advanced post-training pipeline combining Weighted Binary Classification (WBC) and RL
  • 04Introduction of Dr. GRPO for optimizing aggregate page-level rewards and diversity
  • 05Proven 20% reduction in serving latency with significant gains in core engagement metrics

RECOMMENDATION

A must-read for ML engineers looking to simplify recommendation complexity and leverage GenAI scalability. It provides practical insights for applying LLM architectures in high-traffic environments.

The Problem

The Netflix homepage is a complex 2D structure, where traditional multi-stage pipelines suffer from high maintenance overhead and misaligned optimization objectives across components.

The Solution

Netflix introduced 'GenPage,' a single generative transformer that constructs the homepage autoregressively, and optimized it using Reinforcement Learning (Dr. GRPO) for page-level user satisfaction.

The Result

Online A/B tests showed statistically significant gains in user engagement metrics and a 20% reduction in end-to-end serving latency compared to the legacy system.

Trade-off

Challenges include meeting strict real-time latency constraints and handling entity cold starts, while manual context summarization still introduces a form of prompt engineering.

03

Key Concepts

Concept · 01

GenPage

A generative recommender model that builds the Netflix homepage autoregressively, conditioning each row and entity on previous selections and user context.

  • Replaces complex multi-stage pipelines with a single decoder-only model
  • Treats user engagement history and request context as prompts for page generation
Concept · 02

Dr. GRPO

A specialized variant of the Group Relative Policy Optimization algorithm used to align generative models with page-level user satisfaction.

  • Optimizes for page-level rewards such as diversity and stopping power
  • Mitigates training biases while preserving autoregressive flexibility
Concept · 03

Custom Domain Tokenization

A strategy of representing domain-specific items like movies and user actions as discrete tokens instead of using general text tokenizers.

  • Significantly reduces sequence length and lower inference cost
  • Enables direct mapping between tokens and product concepts for better control