<ul><li>Large language models (LLMs) can generate fluent but factually incorrect outputs, known as hallucinations, affecting reliability in real-world applications.</li><li>A framework decomposing LLM uncertainty into four distinct sources is presented in this paper.</li><li>A source-specific estimation pipeline is developed to quantify these uncertainty types across tasks and models.</li><li>Experiments show that the proposed uncertainty-aware selection strategy consistently outperforms baseline strategies in selecting appropriate models or uncertainty metrics for more reliable deployment.</li></ul>

Uncertainty Profiles for LLMs: Uncertainty Source Decomposition and Adaptive Model-Metric Selection

Discover more