Large Language Models and Knowledge Graphs: Recent Advances, Synergies, and Future Directions

Introduction In the past year, large language model (LLM) technology has advanced dramatically, ushering in a new stage of AI research and presenting fresh opportunities and challenges for knowledge graph (KG) technology. This talk focuses on the latest LLM research, its assistance to knowledge engineering, how KGs help evaluate and apply LLMs, and future prospects for KG‑LLM interaction.

1. Comparison between 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. Knowledge graphs, on the other hand, provide explicit, structured, highly interpretable knowledge at the cost of high construction effort and weaker NLP performance. The two are complementary.

2. LLM‑Assisted Knowledge Extraction By issuing specific prompts, LLMs can extract entities, relations, and events from text. Examples include InstructUIE (Fudan University) and KnowLM (Zhejiang University), which can be further fine‑tuned with supervised‑feature‑tuning (SFT) to improve extraction quality.

3. LLM‑Assisted Knowledge Completion Techniques extract latent factual knowledge from LLM parameters to enrich KGs. While promising, these methods must mitigate hallucinations through targeted fine‑tuning, leaving ample research space.

4. KG‑Assisted LLM Evaluation Knowledge graphs provide rigorous benchmarks for LLMs. The KoLA benchmark (Tsinghua University) evaluates memory, understanding, application, and innovation of world knowledge. Findings show that larger, instruction‑tuned models improve high‑level abilities but may suffer “alignment tax” on low‑level tasks.

5. KG‑Assisted LLM Deployment KGs can improve the factual accuracy, safety, and consistency of LLM outputs. They act as post‑hoc verification tools (e.g., checking Aristotle’s lifespan) and help filter unsafe content, thereby enhancing reliability.

6. KG‑Enhanced Complex Reasoning For multi‑hop reasoning, KGs enable precise logical steps. Projects such as KoPL (Tsinghua) translate natural language queries into compositional programs, allowing LLMs to perform accurate, explainable reasoning over structured knowledge.

7. Interactive Fusion of KG and LLM Combining the explicit structure of KGs with the expressive power of LLMs through iterative collaboration can yield deeper semantic understanding, richer knowledge representation, and stronger reasoning capabilities.

Conclusion The synergistic mode of KGs and LLMs is expected to become a breakthrough for integrating neural and symbolic AI, opening new pathways toward general AI and creating abundant research opportunities.