Synergy Between Large Language Models and Knowledge Graphs: Recent Advances, Evaluation, and Future Integration

Introduction In the past year, large language model (LLM) technology has advanced dramatically, heralding a new era for artificial intelligence and presenting both opportunities and challenges for knowledge graph (KG) research. The presentation focuses on recent LLM research, how LLMs assist knowledge engineering, how KGs aid LLM evaluation and applications, and future directions for their interactive integration.

1. Comparison of LLMs and Knowledge Graphs LLMs, built on deep neural networks, excel at natural language understanding and generation but store knowledge implicitly as parameters, leading to hallucinations and limited interpretability. In contrast, KGs embody symbolic, explicit, highly interpretable knowledge structures, though they are costly to construct, often incomplete, and less effective for raw NLP tasks.

2. LLMs Empower Knowledge Extraction By issuing specific instructions, LLMs can perform entity, relation, and event extraction. Notable examples include InstructUIE (Fudan University) and KnowLM (Zhejiang University), which use instruction tuning and optional supervised fine‑tuning (SFT) to achieve high‑quality extraction across multiple tasks.

3. LLMs Aid Knowledge Completion Researchers extract parametric knowledge from LLMs to populate KGs, improving graph completeness. However, due to potential hallucinations, constrained fine‑tuning or alignment techniques are required, leaving substantial research space.

4. Knowledge Graphs Support LLM Evaluation KGs provide benchmarks such as the KoLA suite (Tsinghua University) that assess LLMs on memory, understanding, application, and innovation. Findings reveal that larger, unaligned models excel at factual recall, while instruction‑tuned, human‑aligned models improve high‑order abilities but may suffer a “alignment tax” on low‑level tasks.

5. Knowledge Graphs Enhance LLM Deployment KGs act as external tools to improve factual accuracy, safety, and consistency of LLM outputs. Examples include post‑hoc verification of generated facts (e.g., Aristotle’s lifespan) and filtering unsafe or illegal content by cross‑referencing KG facts.

6. Knowledge Graphs Boost Complex Reasoning Structured KG data enables multi‑hop reasoning and precise inference. The KoPL programming language (Tsinghua) translates natural language queries into composable functions for complex question answering, improving interpretability and extensibility.

7. Interactive Fusion of LLMs and KGs Future research envisions iterative collaboration where LLMs provide rich textual understanding while KGs supply explicit, verifiable structures, leading to deeper semantic comprehension and stronger reasoning capabilities.

Conclusion The collaborative mode between knowledge graphs and large language models is poised to become a key breakthrough for neural‑symbolic AI, opening new pathways toward artificial general intelligence and offering abundant research opportunities across safety, accuracy, and explainability.