Beyond Semantic Equivalence: Logical Graphs for LLM Uncertainty Quantification

This paper introduces Logical Graph Uncertainty (LGU), an uncertainty-estimation framework that models entailment and incompatibility relations among candidate answers. It aggregates probability mass along entailment chains, computes entropy over logically maximal hypotheses, and penalizes mutual incompatibility. According to the abstract, LGU consistently improves uncertainty estimation across multiple question-answering benchmarks, with gains over semantic entropy of up to 7.1% in AUROC and 3.5% in AUARC. The approach targets cases where answers differ in specificity or wording but remain logically compatible.