📝 SummarXiv

Abstract

Large Language Models (LLMs) often struggle with producing factually consistent answers due to limitations in their parametric memory. Retrieval-Augmented Generation (RAG) methods address this issue by incorporating external knowledge from trusted sources at inference time. However, such methods typically treat knowledge as unstructured text, which limits their ability to support compositional reasoning and identify factual inconsistencies. To overcome these limitations, we propose a novel framework that dynamically constructs and expands knowledge graphs (KGs) during inference, integrating both internal knowledge extracted from LLMs and external information retrieved from external sources. Our method begins by extracting a seed KG from the question via prompting, followed by iterative expansion using the LLM's latent knowledge. The graph is then selectively refined through external retrieval, enhancing factual coverage and correcting inaccuracies. We evaluate our approach on three diverse factual QA benchmarks, demonstrating consistent improvements in factual accuracy, answer precision, and interpretability over baseline prompting and static KG-augmented methods. Our findings suggest that inference-time KG construction is a promising direction for enhancing LLM factuality in a structured, interpretable, and scalable manner.