TLDR: This research introduces a framework that builds and expands knowledge graphs during inference to improve the factual accuracy of Large Language Models (LLMs). By combining internal LLM knowledge with external information from sources like Wikipedia and Google Search, the method effectively corrects inaccuracies and fills knowledge gaps, leading to more reliable answers in factual question-answering tasks.
Large Language Models (LLMs) have shown impressive capabilities in understanding, generating, and reasoning with natural language. However, they often struggle to produce factually consistent answers, a problem commonly referred to as “hallucination.” This limitation stems from the inherent constraints of their parametric memory, which can be incomplete or imprecise.
Traditional methods like Retrieval-Augmented Generation (RAG) attempt to address this by incorporating external knowledge. However, these methods typically treat knowledge as unstructured text, which can limit their ability to perform complex reasoning or identify factual inconsistencies effectively.
A New Approach: Dynamic Knowledge Graph Construction
A recent research paper, “Improving Factuality in LLMs via Inference-Time Knowledge Graph Construction,” proposes a novel framework to tackle these challenges. The core idea is to dynamically build and expand knowledge graphs (KGs) during the inference process itself. This framework integrates both the internal knowledge latent within LLMs and external information retrieved from trusted sources.
The process begins by extracting a “seed” knowledge graph directly from the user’s question using the LLM. This initial graph is then iteratively expanded using the LLM’s own internal knowledge. Crucially, the graph is then refined and enhanced through external retrieval, drawing information from sources like Wikipedia and Google Search. This external grounding helps to correct inaccuracies and fill in any missing factual details, ensuring a more robust and accurate knowledge base for answering questions.
How the Framework Works
The pipeline involves several key steps:
Graph Initialization: The LLM first parses the input question to extract relevant entities and relations, forming an initial knowledge graph.
Graph Expansion: This initial graph is then expanded in a breadth-first manner. The LLM iteratively generates new relations and objects from selected entities, building a larger, more comprehensive graph based on its internal knowledge.
External Retrieval: If the LLM’s internal knowledge proves insufficient or inaccurate, this step comes into play. The LLM can select specific triplets from the graph and choose to either “Correct” them (if they contain errors) or “Expand” them (to add more detail). Search queries are constructed from these triplets, and relevant information is retrieved from external sources. This external context is then used to refine and update the knowledge graph.
Answering on the Graph: Finally, with the refined and externally validated knowledge graph, the LLM generates an answer to the original question. This structured approach ensures that the answer is grounded in verifiable facts.
Key Findings and Impact
The researchers evaluated their approach on three diverse factual question-answering benchmarks: Complex WebQuestions (KBQA), HotpotQA (Document-based QA), and SimpleQA (Expert-curated QA). The results demonstrated consistent and substantial improvements in factual accuracy and answer precision across various LLMs, including GPT-4o, Deepseek-V3, Gemini-2.5-flash, Qwen2.5-32B, and Llama-4-scout.
A significant finding was that while constructing KGs from internal LLM knowledge alone did not always lead to consistent improvements (and sometimes even degraded performance due to incomplete or imprecise internal knowledge), the integration of external retrieval consistently boosted performance. This hybrid approach proved particularly effective in bridging the capability gap between smaller and larger LLMs, allowing models with fewer parameters to achieve performance comparable to or even surpassing much larger models on certain tasks.
The study also explored the impact of reasoning complexity (number of “hops” in a question) and model scale. They observed that more complex questions led to larger knowledge graphs but a decline in accuracy, highlighting the challenge of pinpointing the correct answer within a vast graph. Larger models generally performed better, but the external grounding mechanism significantly enhanced the robustness of all models.
Also Read:
- A New Approach to Integrating Knowledge Graphs with Large Language Models for Enhanced Completion
- RTQA: A Framework for Answering Complex Time-Based Questions on Knowledge Graphs
Future Directions and Limitations
While promising, the work acknowledges certain limitations. There’s a risk of generating “hallucinated” content when constructing KGs solely from the LLM’s internal knowledge. Additionally, a gap remains between the recall of the constructed graphs and the exact match performance of the answers, suggesting that graph retrieval mechanisms could still be optimized.
This research offers a promising direction for enhancing LLM factuality in a structured, interpretable, and scalable manner. By dynamically building and refining knowledge graphs with both internal and external knowledge, LLMs can provide more reliable and accurate responses, which is crucial for high-stakes applications. You can find the full research paper here.
