From Argumentation to Labeled Logic Program for LLM Verification

Abstract

Large language models (LLMs) often generate fluent but in-correct or unsupported statements, commonly referred to as hallucinations. We propose a hallucination detection frame-work based on a Labeled Logic Program (LLP) architecture that integrates multiple reasoning paradigms, including logic programming, argumentation, probabilistic inference, and ab-ductive explanation. By enriching symbolic rules with seman-tic, epistemic, and contextual labels and applying discourse-aware weighting, the system prioritizes nucleus claims over peripheral statements during verification. Experiments on three benchmark datasets and a challenging clinical narrative dataset show that LLP consistently outperforms classical symbolic validators, achieving the highest detection accuracy when combined with discourse modeling. A human evaluation further demonstrates that logic-assisted explanations im-prove both hallucination detection accuracy and user trust. The results suggest that labeled symbolic reasoning with dis-course awareness provides a robust and interpretable ap-proach to LLM verification in safety-critical domains.