To address the limitations of large-scale language models (LLMs) that tend to generate hallucinatory or anachronistic responses due to static internal knowledge, this paper proposes an improved reinforcement learning (RL)-based retrieval-augmented generation (RAG) framework, called RAG-R1. RAG-R1 is designed to adaptively utilize internal and external knowledge during the inference process, and extends the generation and retrieval processes from single-query mode to multi-query parallel processing to shorten the inference time and improve the model performance. Experimental results on seven question-answering benchmarks show that the proposed method outperforms the state-of-the-art baseline models by up to 13.2% and reduces the inference time by 11.1%.
