Knowledge-Geometry Decoupling: Refreshable Pretrained Transfer for Streaming Recommendation

  • 类型：arxiv
  • 标识：2608.02738
  • 链接：https://arxiv.org/abs/2608.02738
  • 主分类：engineering
  • 形态：method
  • 被引：0
  • 被引来源：Semantic Scholar
  • S2被引：0
  • 影响力被引：0
  • TLDR：Knowledge-Geometry Decoupling (KGD) is proposed and Behavioral Multi-Token Prediction (BMTP) is introduced to retain only collaboratively or semantically related future items as supervision, yielding cleaner and more transferable behavioral knowledge.
  • 待LLM分类：否
  • 标题中文：知识-几何解耦：面向流式推荐的可刷新预训练迁移
  • TLDR中文：提出 Knowledge-Geometry Decoupling (KGD) 并引入 Behavioral Multi-Token Prediction (BMTP)，仅将协作或语义相关的未来项作为监督，从而得到更干净、更可迁移的行为知识。
  • 来源文件：
  • /inbox/tom/_candidates/2026-08-05-agent-rag-longcontext-candidates.json
  • [S2 enrich]