标题： 探索推理：一种在图上进行检索和生成的统一方法 显示英文标题

标题： Reasoning by Exploration: A Unified Approach to Retrieval and Generation over Graphs

摘要： 在结构化图上进行推理仍然是大型语言模型（LLMs）的一个基本挑战，尤其是在扩展到大型图时。现有方法通常遵循检索增强生成（RAG）范式：首先检索与查询相关的子图，然后在检索到的子图基础上生成答案。然而，这种两阶段的流程往往难以忠实地融入图结构，因为生成过程最终受限于检索到的子图的质量和完整性。尽管最近提出了许多先进的检索器来缓解这个问题，但它们通常针对训练图进行优化，在未见过的图上泛化能力较差，这限制了它们的实际应用。在本工作中，我们提出了探索推理（RoE），这是一种新颖的方法，通过将图上的推理视为图探索过程，统一了检索和生成。在每一步，LLM会选择要探索的候选节点和边，逐步构建推理路径并在过程中生成答案。为了实现有效的探索，RoE分为两个阶段进行训练：首先在黄金推理路径上进行监督微调（SFT），然后通过强化学习（RL）来提高探索效果和泛化能力。在基准数据集上的实验表明，RoE在基线方法上实现了显著的整体改进，同时也能有效地泛化到未见过的图。 显示英文摘要

摘要： Reasoning over structured graphs remains a fundamental challenge for Large Language Models (LLMs), particularly when scaling to large graphs. Existing approaches typically follow the retrieval-augmented generation (RAG) paradigm: first retrieving subgraphs relevant to the query and then generating answers conditioned on the retrieved subgraphs. However, such two-phase pipelines often struggle to faithfully incorporate graph structure, since the generation process is ultimately constrained by the quality and completeness of the retrieved subgraph. Although many advanced retrievers have been proposed recently to mitigate this issue, they are usually tailored to the training graphs and generalize poorly to unseen graphs, which limits their practical applicability. In this work, we propose Reasoning by Exploration (RoE), a novel approach that unifies retrieval and generation by framing reasoning over graphs as a process of graph exploration. At each step, the LLM selects candidate nodes and edges to explore, gradually constructing reasoning paths and generating answers along the way. To enable effective exploration, RoE is trained in two stages: supervised fine-tuning (SFT) on gold reasoning paths, followed by reinforcement learning (RL) to enhance exploration effectiveness and generalization. Experiments on benchmark datasets demonstrate that RoE achieves substantial overall improvements over baselines, while also generalizing effectively to unseen graphs.