阿里发布 SPAR 论文：自进化销售智能体在 1688 平台完成 20 万商家实测
阿里在 1688 上跑了 20 多万商家的 A/B 测试，销售智能体靠记忆和归因把转化率提了 15.8%，做智能体的可以看看归因那一步怎么设计的。
阿里团队发表 SPAR 论文，该智能体在 1688 平台上运行 B2B 销售触达，覆盖 200,000+ 商家。实测中转化率提升 15.8%，高意向商家增长 35.8%。SPAR 不更新模型权重，通过内循环维护每个商家的情景记忆，外循环维护以可测试假设形式编写的共享销售策略手册。关键在于归因步骤：改写策略前先判断丢单源于策略缺陷还是外部因素。领域专家给其销售质量打 4.62/5 分，为所有受测方法中最高。
Great paper from Alibaba on self-evolving agents, with a live A/B test across 200,000+ merchants. (bookmark it) This is another brilliant example of leveraging episodic memory with agents and the importance of attribution. SPAR runs B2B sales outreach on Alibaba's 1688 platform. Conversion rose 15.8%, and high-intent merchant growth rose 35.8%. The agent learns at two levels without updating weights. An inner loop keeps episodic memory for each merchant across sessions. An outer loop maintains a shared playbook of selling strategies, each written as a testable hypothesis. The key step is attribution. Before changing the playbook, SPAR decides whether a lost deal came from a bad strategy or an outside factor, so a deal lost to outside factors doesn't change a strategy that works. Domain experts rated it 4.62 out of 5 on sales quality, the highest of all methods tested. If you are building agents that learn from production outcomes, it might be worth adding an attribution step before writing any lesson to memory. arxiv.org/abs/2611.01695 Chat with Paper: academy.dair.ai/papers/spar-ex… 💬 22 🔄 4 ❤️ 45 👀 2980 📊 29 ⚡
