Systematic Evaluation of Black-Box Uncertainty Methods for Large Language Models

Large Language Models (LLMs) are powerful but can produce unreliable outputs, including hallucinations. To build more trustworthy LLMs, accurately estimating their uncertainty is crucial. Many mainstream LLMs are only accessible via APIs, meaning internal signals like logits are unavailable, making "black-box" uncertainty estimation particularly important. This research addresses the fragmented state of black-box uncertainty estimation by providing a systematic review and unified empirical comparison. The authors categorize existing methods into five groups: verbalization-based, sampling-based, explanation-based, multi-agent, and hybrid. They then benchmark 24 representative methods across four different LLMs and four dataset settings. The findings indicate that no single method is universally superior. However, methods that involve reasoning over and comparing multiple candidate answers tend to be effective. Additionally, hybrid methods, which combine various uncertainty signals, generally perform well across most conditions. The researchers are releasing their benchmark data and evaluation framework to encourage future research and provide practical guidance for developing more robust black-box uncertainty estimation techniques.