[2607.17719] SR-Agent: An Expertise-Pushed Agentic Framework for Publish-Rating Technique Refinement in E-Commerce Suggestion

View a PDF of the paper titled SR-Agent: An Expertise-Pushed Agentic Framework for Publish-Rating Technique Refinement in E-Commerce Suggestion, by Hanchen Yang and 9 different authors

View PDF
HTML (experimental)

Summary:Consumer expertise is a first-class goal in industrial e-commerce recommender methods (RS). Publish-ranking methods, which govern variety, similarity, and publicity over a ranked record, are broadly deployed in industrial RS for his or her simplicity and low serving price. Nonetheless, as the net advice surroundings evolves repeatedly, these statically configured methods step by step turn out to be stale, thereby degrading the consumer expertise. Refining them sometimes depends on guide inspection, analysis, and updates, making it sluggish, expensive, and tough to scale or reuse. Though latest LLM-based brokers (e.g., RecUserSim, SimUSER, and Self-EvolveRec) supply promising instructions, none of them shut the complete loop of automated, self-evolving technique refinement. To bridge this hole, we introduce SR-Agent, which, to the very best of our data, is the primary agentic framework deployed to refine post-ranking methods in industrial RS. SR-Agent unifies three elements: (i) a UserSim agent that applies inspection expertise to floor user-perceived unhealthy instances; (ii) an Evaluation agent that consolidates recurring unhealthy instances into structured, reusable diagnoses; and (iii) a constrained Technique Refinement Harness that maps diagnoses to typed and bounded actions, gated by a four-stage reward pipeline with reversible rollback. Deployed on the Kuaishou e-commerce platform, SR-Agent repeatedly runs this refinement loop and, in a one-month on-line A/B take a look at, will increase order quantity by 0.71%, shopping depth by 0.34%, and clicked-category variety by 0.48%, whereas markedly shortening the refinement cycle and decreasing operational price.