SepLRM：面向推荐的基础模型范式
原标题：SepLRM: A Foundation Model Paradigm for Recommendation
AI 摘要
论文提出 SepLRM，一种推荐基础模型（RFM）范式，将多个端到端大推荐模型（LRM）重组为一个共享 RFM 和多个独立优化的垂直模型（VM）。RFM 采用编码器-解码器架构，导出与候选无关的 KV cache 作为共享接口，以保留细粒度行为信息并实现高保真能力迁移。在三个大规模数据集上取得 SOTA 离线性能，迁移率约 100%，并在字节跳动两个大规模工业场景部署，提升了序列扩展的 ROI 且在线 A/B 测试获得一致收益。
正文节选
SepLRM: A Foundation Model Paradigm for RecommendationThanks: * Equal contributions. \(\dagger\) Corresponding authors. Abstract. Large recommendation models (LRMs) improve recommendation performance by scaling behavior sequence length, model capacity, and training compute. However, the return on investment (ROI) of continued scaling deteriorates as marginal gains diminish and long-sequence encoding costs rise, especially when this computation is repeatedly executed by distinct models for divers