Search papers, labs, and topics across Lattice.
This paper introduces Recommendation-as-Generation (RaG), a novel framework that generates personalized videos on demand by leveraging inferred user interests, thereby overcoming the limitations of traditional video recommendation systems. By utilizing shared semantic IDs (SIDs) to disentangle video representations into content and style semantics, RaG enables fine-grained modeling of user preferences and controllable video generation. The implementation of Video Generation Agents (VGAs) and a cross-domain reward learning mechanism resulted in a significant 1.87% increase in ad revenue in a large-scale deployment with over 400 million daily active users, showcasing the framework's effectiveness in enhancing user engagement and monetization.
Generating personalized videos on demand can boost ad revenue by nearly 2% in high-traffic environments, transforming traditional recommendation systems.
Traditional short-video recommendation systems match user interest to a fixed pool of pre-produced videos, which limits their ability to capture fine-grained and dynamic preferences. We propose Recommendation-as-Generation (RaG), a new paradigm that generates personalized videos on demand from inferred user interest. Our framework unifies generative recommendation and video generation through shared semantic IDs (SIDs), which disentangle video representation into content semantics and creative style semantics, enabling both fine-grained modeling of user interest and controllable generation of interest-aligned videos. We further develop Video Generation Agents (VGAs) that are conditioned on inferred SIDs to drive hierarchical planning and refinement for video creation, including visual composition, audio alignment, and artistic effect enhancement. To optimize the framework, we effectively introduce a synergistic cross-domain reward learning mechanism that jointly enforces interest alignment, user feedback, and video quality assessment. We deploy RaG on an industrial-scale platform with over 400 million daily active users and evaluate it in a revenue-critical advertising scenario. Online A/B tests show up to 1.87% ad revenue improvement compared to a strong production GRM baseline, demonstrating its effectiveness in driving further revenue gains beyond generative recommendation. Our results highlight a closed-loop generative system as a promising paradigm for integrating personalized video generation into recommendation.
