GEMs: Breaking the Long-Sequence Barrier in Generative Recommendation with a Multi-Stream Decoder

Researchers have unveiled GEMs, a new AI framework that breaks a major barrier in generative recommendation systems by processing user behavior sequences exceeding 100,000 interactions—far beyond current limits. It splits user history into Recent, Mid-term, and Lifecycle streams for efficient analysis, achieving state-of-the-art accuracy. This is the first 'lifelong' framework successfully deployed in a high-concurrency industrial environment, enabling models to capture a user's complete digital history for the first time.