Leveraging Low-Level Symbolic Competences for Unsupervised Grounding in Hallucination Detection

The researchers proposed a supervised hallucination detection method based on low-level symbolic capabilities. This method enables large language models to build SQL databases from reference documents and infer the relationship between the reference content and sampled responses based on these databases, thereby providing neural symbolic checks. Experiments on the RAGTruth and DiaHalu datasets show that this approach is superior to direct prediction and comparable to state-of-the-art methods, without the need for domain-specific fine-tuning, relying only on the low-level general capabilities already available in large language models.