Uncertainty estimation is critical for deploying large language models (LLMs) in safety-sensitive and decision-critical applications. Recent approaches estimate semantic uncertainty by clustering multiple sampled responses into equivalence classes and measuring their diversity via entropy-based criteria. However, existing methods typically rely on greedy hard clustering and von Neumann entropy, which suffer from sensitivity to clustering order, noise in semantic equivalence judgments, and limited control over spectral contributions. In this work, we propose a principled information-theoretic framework for LLM semantic uncertainty estimation based on soft semantic communities and kernel Rényi entropy. Given multiple generations for a query, we construct a weighted semantic graph using pairwise semantic similarity scores and infer soft community assignments via weighted graph community detection. These soft assignments induce a positive semi-definite semantic kernel that captures the distribution of semantic modes without enforcing hard equivalence relations. Uncertainty is then quantified by the Rényi entropy of the kernel spectrum, yielding a tunable measure that interpolates between sensitivity to dominant semantic modes and long-tail semantic diversity. Compared to prior von Neumann entropy-based estimators, the proposed Rényi spectral uncertainty offers improved robustness to semantic noise, reduced dependence on clustering heuristics, and greater flexibility through its order parameter. Extensive experiments on question answering tasks demonstrate that our method provides more stable and discriminative uncertainty estimates, particularly under limited sampling budgets and noisy semantic judgments.
Li et al. (Tue,) studied this question.