🤖 AI Summary
Hallucination detection in large language models (LLMs) faces key bottlenecks: high annotation costs, dataset model specificity, and reliance on white-box access or supervised signals. Method: We propose a zero-resource, black-box hallucination detection paradigm. Our approach introduces the first automated framework for constructing hallucination datasets from fact-checking corpora, integrating prompt-engineering–driven self-consistency verification, black-box response analysis, and cross-model hallucination pattern comparison—requiring neither human annotation nor internal model access. Contribution/Results: Experiments demonstrate significant improvements over state-of-the-art baselines across major open- and closed-source LLMs. Crucially, our method systematically uncovers structural differences across models in hallucination types and prevalence—revealing previously uncharacterized variation. This enables scalable, model-agnostic evaluation of LLM reliability, advancing trustworthy AI assessment.
📝 Abstract
While Large language models (LLMs) have garnered widespread applications across various domains due to their powerful language understanding and generation capabilities, the detection of non-factual or hallucinatory content generated by LLMs remains scarce. Currently, one significant challenge in hallucination detection is the laborious task of time-consuming and expensive manual annotation of the hallucinatory generation. To address this issue, this paper first introduces a method for automatically constructing model-specific hallucination datasets based on existing fact-checking datasets called AutoHall. Furthermore, we propose a zero-resource and black-box hallucination detection method based on self-contradiction. We conduct experiments towards prevalent open-/closed-source LLMs, achieving superior hallucination detection performance compared to extant baselines. Moreover, our experiments reveal variations in hallucination proportions and types among different models.
