Code-MUE: Measuring Code LLM Uncertainty via Execution-Based Semantic Interaction Graphs

Code-MUE is a purely black-box framework that measures Code LLM uncertainty using execution-based semantic interaction graphs, and quantifies global semantic diversity with the Von Neumann entropy of the solution space. On 8 SOTA LLMs, its Spearman correlation with functional correctness reaches as high as -0.98, significantly outperforming lexical and embedding-vector baselines, and it supports risk detection and selective prediction.