A/B Agent Automates Recommendation Strategy Optimization

Industrial recommendation systems heavily rely on A/B testing to iterate and optimize strategies. Traditionally, this process is labor-intensive, requiring experts to manually design experiments, analyze results, and adjust parameters. Furthermore, valuable insights from past experiments often remain fragmented and difficult to reuse systematically. Existing RAG agents offer partial solutions but typically lack a hierarchical understanding of experimental contexts, leading to suboptimal retrieval and limited knowledge transfer. To address these challenges, researchers propose A/B Agent, a closed-loop system designed for autonomous recommendation strategy optimization. This framework consists of three main components: a hierarchical organization of historical strategy knowledge, an autonomous target-aware strategy generation module using multi-path Tree-RAG, and an experiment-guided strategy self-evolution mechanism. The agent continuously analyzes online A/B feedback to refine strategies and update its knowledge base, demonstrating significant improvements in real-world e-commerce recommendation systems, including a 4.829% increase in GMV.