독서리서치레이더투자 프레임워크
로그인 / 회원가입
← 읽기로 돌아가기
뉴스주제 · WeGR生成式检索框架중대1 분 소요DisconfirmAI 편집부

WeChat Deploys WeGR Generative Retrieval: Orders Up 2.5%, Compute Down Over 25%

WeChat E-commerce Search deployed the WeGR framework, showing a 2.5% increase in orders and over 25% reduction in CPU/GPU usage in an A/B test.

2026-10-14 12:00 GMT+8

The WeChat E-commerce Search system has deployed a unified generative retrieval framework named WeGR. According to an arXiv preprint submitted on October 12, 2026, a two-month online A/B test demonstrated relative improvements of 2.539% in orders, 1.508% in click-through rate, and 1.381% in conversion rate. Simultaneously, online CPU and GPU usage decreased by 25.5% and 26.1%, respectively.

WeGR addresses challenges in existing generative retrieval methods, specifically the difficulty of incorporating collaborative signals into Semantic ID tokenizers and the sparsity of reinforcement learning rewards. The framework includes the CoCo-VAE tokenizer and Dynamic Horizon Policy Optimization (DHPO). The system is currently deployed across multiple search scenarios in WeChat, serving billions of users with tens of millions of daily page views.

출처

  • arxiv.org↗
주제

WeGR生成式检索框架

腾讯微信电商搜索部署了名为WeGR的统一生成式检索框架,旨在通过自回归模型直接生成候选商品ID,替代传统多阶段级联架构。该框架包含CoCo-VAE分词器和DHPO优化策略两个核心组件,分别解决语义ID的协同信号缺失与强化学习奖励稀疏问题。

2 단계0 건업데이트됨 2026-10-16 03:37

현재까지의 과정

로그인 / 회원가입
독서리서치레이더투자 프레임워크
  • 2026-10-12腾讯团队向arXiv提交WeGR论文,介绍其在微信电商搜索中的统一生成式检索框架及部署情况。arxiv.org ↗
  • 2026-10-15新闻发布,确认WeGR在两个月A/B测试中带来订单增2.5%且算力降超25%的效果。arxiv.org ↗
  • 아직 해결되지 않은 문제

    • WeGR移除80%其他渠道候选项后,对长尾或新上架商品的召回效果是否有负面影响?
    • 与百度ICEGR相比,CoCo-VAE在处理查询意图一致性方面是否存在特定短板?
    • 生成式检索在更大规模流量下的稳定性及延迟表现如何?
    전체 주제 읽기 →
    리서치 계속하기 →