OxygenREC: Instruction-Following Generative Framework for E-commerce Recommendation

This paper introduces OxygenREC, an industrial recommendation system designed to address limitations in existing Generative Recommendation (GR) systems. It leverages a Fast-Slow Thinking architecture to balance deep reasoning capabilities with real-time performance requirements. The key contributions are a semantic alignment mechanism for instruction-enhanced generation and a multi-scenario scalability solution using controllable instructions and policy optimization. The paper aims to improve recommendation accuracy and efficiency in real-world e-commerce environments.