Abstract: Recently, Large Language Models (LLM) have achieved breakthrough progress in natural language processing tasks such as machine translation and dialogue systems, owing to their powerful text generation capabilities and advanced semantic understanding. However, their performance in structured reasoning tasks over Knowledge Graphs (KG) still faces two major challenges: factual hallucination and insufficient structural utilization. Existing approaches that integrate KG reasoning with LLM mainly fall into two categories: retrieval-based methods, which directly feed retrieved relevant triples from KG to LLM, and Agent-based methods, which employ LLM as an Agent to iteratively search for paths within KG. The former fails to generalize to unseen problems or interpret graph structures, while the latter suffers from factual hallucination and high computational costs. This paper proposed RADaR (Retrieval-Augmented Decoding and Reasoning), a unified three-stage collaborative reasoning framework. RADaR first decomposed complex queries and located relevant evidence subgraphs via an enhanced semantic parsing and retrieval module. It then constructed a local structural index over retrieved subgraph and employed a graph-constrained decoding mechanism to generated faithful and interpretable reasoning paths. Building upon this, a final inductive reasoning module synthesized multiple paths to produce answer. Experimental results on KG question-answering benchmarks (e.g., WebQSP, CWQ) show that RADaR outperforms existing single-paradigm methods across several metrics, achieving an optimal balance among performance, faithfulness, and efficiency.

Key words: KGQA&, #41

摘要： 最近，大语言模型(LLM) 凭借其强大的文本生成能力和语义理解水平，在机器翻译、对话系统等自然语言处理任务中取得突破性进展。然而，其在知识图谱(KG)结构化推理任务中的表现仍面临存在事实幻觉与结构利用不足的双重挑战。现有的将知识图谱推理与大语言模型结合的方法主要分为两类：从知识图谱中检索相关三元组直接输入给LLM的检索式方法，或将LLM作为代理，在知识图谱中迭代搜索路径的代理式方法。其中前者无法泛化未见问题或解释图结构，后者则存在事实幻觉和计算成本高的问题。基于以上问题提出RADaR(Retrieval-Augmented Decoding and Reasoning)，一个统一的三阶段推理框架。RADaR首先通过增强的语义解析与检索模块分解复杂查询并定位相关证据子图；随后，在检索到的子图上构建局部结构索引，并利用图约束解码机制生成忠实且可解释的推理路径；在此基础上通过归纳推理模块综合多条路径生成最终答案。在知识图谱问答(WebQSP、CWQ等)基准上的实验结果表明，RADaR在多项指标上优于现有的单一范式方法，实现了性能、忠实性与效率的最佳平衡。