Make It Long, Keep It Fast: End-to-End 10K Long User Behavior Sequence Modeling for Billion-Scale Douyin Recommendation
STCA 将历史自注意力替换为目标-历史交叉注意力，复杂度从 O(L²) 降至 O(L)，RLB 在请求级复用用户编码，实现 10K 长序列端到端训练，已部署于抖音全量流量
All tags
industrial-deployment
4 posts tagged with "industrial-deployment"
Better Generalization with Semantic IDs: A Case Study in Ranking for Recommendations
在 YouTube 排序模型中验证 Semantic ID 替代随机视频 ID，通过 RQ-VAE 量化内容嵌入为离散语义 token，提升冷启动和长尾物品的泛化能力
Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations
Generative Recommenders (GR) 将推荐重新定义为序列转导任务，HSTU 编码器比 FlashAttention2 Transformer 快 5.3x-15.2x，1.5 万亿参数模型在线 A/B 测试提升 12.4%
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