最低跨度置信度：LLM幻觉检测的零样本黑盒指标

Hallucinations in Large Language Models (LLMs), characterized by the generation of plausible but non-factual content, present a significant hurdle for their reliable deployment in high-stakes environments. Existing hallucination detection methodologies often rely on unrealistic assumptions, such as demanding expensive intensive sampling strategies for consistency checks or requiring access to white-box LLM states, which are frequently unavailable or inefficient in practical settings. This work introduces Lowest Span Confidence (LSC), a zero-shot hallucination detection metric designed to overcome these limitations. LSC assesses the potential for hallucination by evaluating the semantic spans within LLM-generated text that exhibit the lowest confidence scores. This approach operates entirely in a black-box manner, requiring no fine-tuning of the LLM or access to its internal parameters. The fundamental premise of LSC is that hallucinatory content often manifests with localized low generation confidence, even if the overall text appears coherent. By identifying and quantifying these least confident semantic units, LSC effectively captures the model's uncertainty and potential inaccuracies. The metric can be directly applied to any pre-trained LLM without the need for additional labeled data or computational resources. LSC's key advantages lie in its efficiency and universality, positioning it as a practical tool for evaluating and mitigating LLM hallucinations across various applications, especially in contexts demanding high factual accuracy.