Problem
A streaming homepage is a two-dimensional layout of rows, and each row holds items. It is usually built from separate candidate generation and ranking models for rows and for items. Such a stack has many models with misaligned objectives and heavy feature engineering. It also scores items one by one, so it misses page-level effects such as diversity, or a Continue Watching row at the top lowering how much of the page people browse.
Example from the article: the Netflix homepage is almost entirely personalized, and the team replaced the multi-stage production recommender with one generative model.
Solution in brief
GenPage turns user context into a prompt of discrete tokens and generates the whole homepage with a decoder-only transformer. Each row and item is one token, so business rules become token masks at decoding time. The model is pretrained by next-token prediction on pages with positive feedback, then post-trained toward a reward by weighted binary classification or reinforcement learning.
Questions this case answers
- How to generate a whole structured page of recommendations with one autoregressive model so that interactions between rows and items are optimized together
- How to enforce hard business constraints on the output of a generative ranking model at inference time
- How to keep a large sequence model current and handle new items under a daily-changing catalog without full retraining
- How to build a personalized homepage (rows, items and layout) with one model instead of a multi-stage ranking stack?
Tools and methods
- Domain-specific tokenizer — one token per item and row; short sequences, rules map to tokens
- Decoder-only transformer — generates the page autoregressively
- Next-token-prediction pretraining — imitates good production pages
- Weighted binary classification (WBC) post-training — token-level value prediction
- Reward model and Dr. GRPO — RL on a page-level reward
- Constrained and hybrid row decoding — rules via logit masks; faster rows
Google or a one-time email code. No password.