Hyper-RAG: combating LLM hallucinations using hypergraph-driven retrieval-augmented generation

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Natural Language Processing

Background:

• Large Language Models (LLMs) offer transformative potential across sectors, including medicine.
• Medical LLM integration faces challenges due to "hallucinations"—inaccurate generated content posing risks.
• Existing retrieval-augmented generation methods struggle with complex knowledge correlations.

Purpose of the Study:

• To introduce Hyper-RAG, a hypergraph-driven Retrieval-Augmented Generation method.
• To mitigate LLM hallucinations by capturing pairwise and beyond-pairwise knowledge correlations.
• To enhance the reliability and accuracy of LLMs in high-stakes domains like medicine.

Main Methods:

• Developed Hyper-RAG, utilizing hypergraphs to model intricate relationships in domain-specific knowledge.
• Implemented a hypergraph-driven approach to augment LLM retrieval and generation processes.
• Conducted experiments on the NeurologyCrop dataset and nine diverse datasets using prominent LLMs.

Main Results:

• Hyper-RAG improved LLM accuracy by an average of 12.3% over direct use and outperformed GraphRAG (6.3%) and LightRAG (6.0%).
• Hyper-RAG maintained stable performance with increasing query complexity, unlike other methods.
• Hyper-RAG-Lite achieved double the retrieval speed and a 3.3% performance increase over LightRAG.

Conclusions:

• Hyper-RAG effectively enhances LLM reliability and reduces hallucinations in medical applications.
• The method demonstrates robustness across various datasets and query complexities.
• Hyper-RAG presents a promising solution for safe and accurate AI in critical fields like medical diagnostics.