Neurosymbolic Framework for Ternary Claim Verification

A new neurosymbolic framework called inference-time argumentation (ITA) enables ternary claim verification—true, false, or uncertain—by combining formal argumentation semantics with LLM training. The method uses argumentation to guide LLMs in generating and scoring arguments, then computes predictions from those scored arguments. This approach addresses high-stakes domains like health and finance where incomplete or conflicting information requires nuanced answers and faithful explanations. The framework is trainable, optimizing argument generation and scoring based on the quality of induced predictions. The paper is published on arXiv.