This paper proposes R$^2$ec, an integrated large-scale recommendation model with inference capabilities for the recommendation field. R$^2$ec introduces a dual-head architecture that supports inference chain generation and efficient item prediction in a single model, significantly reducing inference time. To address the lack of annotated inference data, we design RecPO, a reinforcement learning framework that jointly optimizes inference and recommendation through a novel fusion reward mechanism. Extensive experiments on three datasets demonstrate that R$^2$ec outperforms conventional, LLM-based, and inference-augmented recommendation models, demonstrating competitive efficiency among conventional LLM-based recommendation models and demonstrating strong adaptability to various recommendation scenarios.
