This new technique could improve AI output accuracy by 80% – and tackle hallucinations once and for all

A new data refinement technique could help improve the accuracy of AI outputs and help build trust in the technology, according to researchers.

The UK’s National Innovation Centre for Data (NICD) at Newcastle University has developed a technique that can boost accuracy by as much as 80% – a major improvement, although AI answers still remain imperfect.

Researchers achieved this by combining large language models (LLMs) with graph-based knowledge databases, a storage method which creates networks of data entities to provide deeper semantic meaning and context.

This combination could help reduce “hallucinations” and omissions in AI outputs.

"The work of AI experts at the UK National Innovation Centre for Data has shown that integration with a graph database can significantly reduce the two most significant problems holding back the exploitation of LLMs for real-world applications: hallucinations and omissions," said Paul Watson, the director of the NICD.

"This is especially important for organizations deploying LLMs in applications where regulatory compliance and the avoidance of reputational or financial damage is key."

How it works

Select models were tested against 510 questions designed to check reasoning across different sources. Combining an AI model with graph database technology led to the best results, significantly outperforming AI models that rely on their training data as well as retrieval augmentation generation (RAG) approaches.

The researchers' technique, which combined RAG with a simple graph system, increased the rate of accurate answers from 29% to 66%.

A knowledge graph is a structured set of information that can be labelled and highlights the relationship between ideas, and could provide vital in helping to improve context for AI models.

"Rather than operating over unstructured, chunked text documents, a knowledge graph (KG) is either harnessed directly or created using an LLM from a set of related documents," the researchers explain in their paper.

Challenges remain, however, most notably in creating knowledge graphs that work well with AI. Too complex, and they risk filling an AI's context window; too simple, and though "easily digestible" by an AI, the results aren't as good.

Still, researchers said the combined RAG-knowledge graph approach was a "promising direction" for improving AI accuracy. Plus, that approach uses fewer tokens for better quality results, they noted.

Hallucination risks

Improving the accuracy of AI will help reduce the significant risks to businesses, noted Dr Jim Webber, the chief scientist at Neo4j.

"In an agentic world, where autonomous systems can make significant decisions over time, this cannot hold," he said.

Webber noted that the work comes with another benefit: cost reduction. Improving accuracy in AI normally involves larger models with more data, which raises costs significantly – a fact seen in recent price increases.

Notably, the NICD's technique leads to better results at lower costs and "shows that better models alone cannot solve the problem, but that the right data at the right time can provide significant benefits in terms of accuracy, responsiveness, and cost."

"They have shown that enterprises no longer need to compromise on the quality of their AI, while they can compromise on cost. This is an unusual and highly welcome finding,” he added.

FOLLOW US ON SOCIAL MEDIA

Follow ITPro on Google News and add us as a preferred source to keep tabs on all our latest news, analysis, views, and reviews.

You can also follow ITPro on LinkedIn, X, Facebook, and BlueSky.