Think-on-Graph: Deep and Responsible Reasoning of Large Language Model with Knowledge Graph

Large language models (LLMs) have made significant strides in various tasks, yet they often struggle with complex reasoning and exhibit poor performance in scenarios where knowledge traceability, timeliness, and accuracy are crucial. To address these limitations, we present Think-on-Graph (ToG), a novel framework that leverages knowledge graphs to enhance LLMs’ ability for deep and responsible reasoning. By employing ToG, we can identify entities relevant to a given question and conduct exploration and reasoning to retrieve related triples from an external knowledge database. This iterative procedure generates multiple reasoning pathways consisting of sequentially connected triplets until sufficient information is gathered to answer the question or the maximum depth is reached. Through experiments on complex multi-hop reasoning question-answering tasks, we demonstrate that ToG outperforms existing methods, effectively addressing the aforementioned limitations of LLMs without incurring additional training costs.

Introduction. Large language models (LLMs) Ouyang et al. (2022); OpenAI (2023); Thoppilan et al. (2022); Brown et al. (2020); Chowdhery et al. (2022) have demonstrated remarkable performance across various natural language processing tasks. These models capitalize on pre-training techniques applied to vast text corpora to generate responses that are coherent and contextually appropriate. Despite their impressive performance, LLMs face substantial challenges when confronted with knowledge-intensive tasks Petroni et al. (2021); Talmor et al. (2019); Talmor and Berant (2018) that require traceability, timeliness, and accuracy of knowledge, and therefore have the risk of generating hallucination or toxic text without responsible reasoning. Moreover, fine-tuning is nontrivial for LLMs since their parameters are inherently difficult to update or modify Yao et al. (2023), thereby constraining their ability to provide accurate answers to questions requiring specialized knowledge beyond what was included in the pre-training phase.

Discussion / Conclusion. We propose a novel framework ToG, which integrates chain of thought reasoning and knowledge graphs to answer knowledge-intensive questions. Our framework draws inspiration from human-like iterative information retrieval and generates multiple high-probability reasoning paths. Experimental results demonstrate that ToG significantly enhances existing prompt methods without additional training costs and mitigates the hallucination issue in LLMs, showcasing the potential of integrating LLMs with knowledge graphs for reasoning tasks.