TL;DR
A roadmap survey that structures LLM-KG integration into three paradigms — KG-enhanced LLMs, LLM-augmented KGs, and Synergized LLM+KG — and identifies open challenges for each direction.
Summary
LLMs suffer from hallucination and knowledge staleness; knowledge graphs are expensive to build and hard to query flexibly. This survey defines three integration paradigms: KG-enhanced LLMs inject structured knowledge to reduce hallucination; LLM-augmented KGs use language models for automated KG construction and completion; and Synergized LLM+KG enables bidirectional mutual enhancement. Methods, datasets, and benchmarks are reviewed for each paradigm. Open challenges include temporal knowledge drift, scalability of symbolic grounding, and reliable retrieval-augmented factual reasoning over large KGs.
Key contributions
- Proposes a three-paradigm taxonomy (KG-enhanced LLMs, LLM-augmented KGs, Synergized LLM+KG) as a unifying framework for the field.
- Establishes KG grounding as the most practical current path to reducing LLM hallucination in factual tasks.
- Demonstrates that LLMs can automate KG construction pipelines, lowering the cost of maintaining structured knowledge.
- Identifies temporal reasoning and multi-hop RAG over large KGs as the weakest evaluation areas and primary open challenges.
When to cite
- When positioning a system that combines LLMs and knowledge graphs and needing a taxonomy to classify the approach.
- When justifying KG grounding as a hallucination-reduction strategy for factual generation tasks.
- When arguing that LLMs can reduce the cost of KG construction through automated extraction or completion.
- When identifying gaps in benchmark coverage for temporal or multi-hop KG reasoning.