June 28, 2026 Compute Only Once: UG-Separation for Efficient Large Recommendation Models UG-Sep 在 TokenMixer 架构中解耦用户侧和物品侧信息流，实现用户侧计算复用，推理延迟降低 20%，已部署于字节跳动多个业务场景 efficient-inference recommendation-systems token-mixer computation-reuse industrial-deployment