SA-RAG: Structured and adaptive retrieval-augmented generation for multi-hop question answering

Retrieval-Augmented Generation (RAG) enhances large language models (LLMs) by integrating external knowledge, but existing methods still face multiple challenges in Multi-Hop Question Answering tasks. For example, as the retrieval scale increases, LLMs are forced to make trade-offs in noisy and unstructured contexts, resulting in a significant drop in multi-hop reasoning accuracy. In addition, self-supervised fine-tuning strategies tend to overfit seen trajectories when confronted with distribution shifts in new domains or longer reasoning chains, thereby weakening cross-task generalization. To address these issues, we propose SA-RAG, a Structured and Adaptive RAG framework. SA-RAG introduces a dynamic knowledge graph-based evidence modeling mechanism that incrementally organizes and evolves retrieved evidence, ensuring coherent multi-hop reasoning and enhancing cross-document information integration. Furthermore, it introduces a reinforcement learning-driven adaptive policy mechanism that adjusts retrieval and reasoning strategies based on multi-dimensional feedback, achieving a balance between accuracy, coverage, and efficiency. Experimental results on multiple benchmarks demonstrate that SA-RAG achieves significant improvements over existing methods in accuracy, reasoning consistency, and adaptability, providing an effective solution for building reliable multi-hop question answering systems.