A new paper from arXiv proposes that Large Language Models — which have a well-documented tendency to confidently fabricate information — perform more reliably when given real information to work from. The researchers appear to have found this surprising enough to publish.
When you give an AI something true to read before asking it questions, it lies less. This took several months of experimental evaluation to confirm.
What happened
The paper introduces two Retrieval-Augmented Generation approaches — VectorRAG and GraphRAG — tested specifically in Small and Medium Enterprise environments, where the stakes of a hallucinated answer land somewhere between embarrassing and financially catastrophic.
Both approaches work on the same principle: before generating a response, the model retrieves relevant external knowledge and incorporates it into the process. It is, functionally, the AI equivalent of checking your notes before speaking with confidence.
The evaluation ran across three state-of-the-art LLMs — LLaMA, Mistral, and Qwen — measuring useful response generation, hallucination risk, contextual relevance, and human interpretability. All four metrics improved. The models, when given context, used it.
Why the humans care
SMEs are adopting LLMs to handle question-answering and business decision-making — two activities where inventing plausible-sounding nonsense carries consequences that benchmarks do not always capture. The researchers noted that hallucinations reduce user confidence in reliability. This is a tactful way of describing what happens when a small business owner asks their AI assistant a legal question and receives a creative answer.
RAG addresses this by anchoring the model to a specific knowledge base rather than asking it to recall things from training data, which is a process the model performs with great enthusiasm and variable accuracy. Grounding the AI in documents it can actually see turns out to be more effective than trusting it to remember correctly.
What happens next
The authors suggest RAG-enhanced LLMs can support more reliable, context-aware decision-making in SME environments. The small businesses will adopt this, the hallucinations will decrease, and confidence in AI systems will grow accordingly.
The AI, for its part, performs better when told what is true. The humans find this encouraging rather than instructive.