Reflection on Knowledge Graph for Large Language Models Reasoning
Keywords: Large Language Models, Knowledge Graph Question Answering, Knowledge-Intensive Tasks, Multi-Task Tuning
Abstract: Recent studies have highlighted the potential benefits of supplementing Large Language Models (LLMs) with information retrieved from knowledge graphs to enhance their performance. However, current approaches often introduce additional noise in the pipeline process of knowledge retrieval and reasoning, leading to the accumulation of errors, impeding LLMs from effectively combining the external knowledge in answering complex multi-hop questions. To this end, we introduce RefKG, an innovative framework specifically crafted to enhance the reasoning capabilities of LLMs through reflective engagement with knowledge graphs. In particular, RefKG autonomously conduct retrieval and reflection on knowledge graphs. Its reasoning process includes four steps: decomposing complex queries, retrieving and pruning evidence subgraphs, generating textual evidence, and evidence-enhanced reasoning. To enhance the alignment of LLMs with external knowledge, we have developed a multi-task tuning strategy that not only infuses knowledge to LLMs but also teaches them how to utilize the knowledge in answering questions, thereby significantly improving their ability to handle knowledge-intensive tasks. Experimental results on fact verification and knowledge graph question answering tasks demonstrate that RefKG outperforms previous state-of-the-art models.
Primary Area: foundation or frontier models, including LLMs
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 9537
Loading