UR$^2$: A Reinforcement Learning Framework Unifying RAG and Reasoning

A recent research article presents UR$^2$ (Unified RAG and Reasoning), a versatile reinforcement learning framework that effectively synchronizes retrieval and reasoning within large language models. This framework tackles the limitations of previous unification efforts, which often focus on open-domain QA with predetermined retrieval parameters. UR$^2$ features two innovative elements: a difficulty-aware curriculum that triggers retrieval for only the most complex cases, and a hybrid knowledge access method that merges domain-specific offline data with real-time summaries generated by LLMs. These features work together to balance retrieval and reasoning. The paper can be found on arXiv with the identifier 2508.06165.