HallGraph: A Zero-Resource Approach for Hallucination Detection in LLMs
Abstract
Hallucination in large language models, the generation of fluent but factually unsupported text, remains a barrier to their safe deployment. Existing zero-resource detectors such as SelfCheckGPT measure consistency across stochastic samples but treat all claims uniformly and draw a fixed, often excessive, number of generations. We present HallGraph, a zero-resource detection framework built on three components. First, typed claim weighting decomposes a response into atomic claims, assigns each a semantic type, and scales its contribution by the empirical detectability of hallucination for that type. Second, adaptive sampling monitors the running variance of perclaim consistency and stops once the estimate is stable, reducing Natural Language Inference (NLI) calls by roughly sixty percent without loss of accuracy. Third, a graph divergence score builds a typed-claim interaction graph for the response and for its samples and measures the structural distance between them as an additional signal. Evaluated on HotpotQA, TruthfulQA, and WikiBio against SelfCheck-BERTScore and SelfCheck-NGram, HallGraph leads on three of four settings, reaching an AUC-ROC of 0.928 and AUC-PR of 0.996 on passage-level WikiBio, with consistently higher rank correlations.