OneTrans-V2：用单一 Transformer 统一工业推荐召回、粗排与精排
原标题：OneTrans-V2: Unifying Retrieval, Pre-rank, and Fine-rank with One Transformer in Industrial Recommender
AI 摘要
该论文提出 OneTrans-V2，用单个 Transformer 统一工业推荐系统中的召回、粗排和精排三个阶段。它只编码一次用户行为序列作为共享上下文，各阶段保留自身候选特征与计算，并通过联合训练实现精排到粗排的模型内知识蒸馏。模型采用稀疏 MoE 扩展共享骨干、P-style 参数化稳定训练，并引入 DCGR 与 SNT。部署于大规模工业推荐系统后，GMV 提升 9.74%，在相同硬件预算下达到原级联的吞吐量。
正文节选
OneTrans-V2: Unifying Retrieval, Pre-rank, and Fine-rank with One Transformer in Industrial Recommender Abstract Industrial recommendation systems typically operate as a cascade of retrieval, pre-rank, and fine-rank, but these stages are usually trained and served as separate models, causing repeated user-sequence encoding, isolated optimization, and duplicated engineering effort. Building on OneTrans’ model-level unification, we present OneTrans-V2, one Transformer that unifies the entire casca