Kuaishou's GLAN uses Transformers to boost landing page personalization by 0.15%

Kuaishou, the Chinese short-video giant, has open-sourced a new framework called GLAN (Generative Landing-page Adaptive Navigator) for optimizing which page users land on when they open the app. Published on arXiv, the paper tackles Personalized Landing-Page Modeling (PLPM) — the task of deciding the most engaging first page for each user. Previous approaches relied on Conservative Q-Learning (CQL), a reinforcement learning method that assumes user behavior is Markovian (memoryless). That assumption fails in practice: users show strong temporal dependencies (e.g., patterns across days) and RL bootstrapping accumulates errors under delayed rewards.

GLAN replaces bootstrapping with a sequence modeling paradigm using Decision Transformer. It introduces two key modules: Local-Return-to-Go (L-RTG) captures inter-day consumption dynamics to provide global guidance, and Hourly Reward Module (HRM) decomposes session-level feedback into fine-grained signals for local supervision. This dual perspective allows the model to learn optimal page assignments without the cumulative bias of TD learning. In live A/B tests on the Kuaishou platform, GLAN achieved +0.158% gains in Daily Active Users (DAU) and +0.108% in user Lifetime (LT). While these may seem small, they represent significant uplifts at Kuaishou's scale (hundreds of millions of users). The paper suggests that generative sequence modeling can outperform traditional RL for user-facing personalization tasks.