Embedding Graph Kernel Entropy (EGKE): Fine-grained Hallucination Detection for LLMs from Embeddings
EGKE is a method for detecting hallucinations in Large Language Models by replacing Natural Language Inference (NLI) with a Radial Basis Function (RBF) kernel applied to dense sentence embeddings. It computes pairwise response similarities directly in embedding space, creating a symmetric graph structure. This structure is then used to calculate the von Neumann entropy of a heat kernel, providing a computationally efficient and consistent way to measure semantic uncertainty.