Kuaishou's UniFormer unifies model scaling, boosting watch time 1%+

UniFormer, developed by Kuaishou's research team, addresses the limitations of component-centric scaling in industrial recommendation systems. Previous approaches scaled individual modules (behavior, interaction, task) independently or co-scaled selectively, but lacked a unified framework. UniFormer decomposes the overall modeling space into feature and task spaces, each modeled by stacked interaction modules. A semantic-based tokenization scheme enables user-item decoupling, allowing request-level inference acceleration. To avoid preference collapse, it employs multi-sequence cross-attention to capture heterogeneous behavior patterns, followed by self-attention for enhanced interaction. Dedicated multi-view FFNs enable flexible, scalable parameter scaling across components.

Extensive online A/B testing across two production scenarios—Kuaishou and Kuaishou Lite—demonstrates consistent improvements in user engagement. App Stay Time increased by +0.101% and +0.260%, while Watch Time rose by +0.729% and +1.113%, respectively. These gains, though modest in percentage, translate to significant user experience improvements at Kuaishou's scale. The framework's efficiency stems from its unified design, which reduces computational overhead while maintaining model capacity. For practitioners, UniFormer offers a practical blueprint for scaling recommendation models holistically without ballooning latency or resource costs.