Gradient Sensitivity Method Detects LLM Stubborn Hallucinations

Researchers propose Embedding-Perturbed Gradient Sensitivity (EPGS) to detect 'Stubborn Hallucinations' in large language models, where models are confidently wrong. The method exploits the geometric property that robust facts lie in flat minima while stubborn hallucinations occupy sharp minima, supported by brittle memorization. EPGS perturbs input embeddings with Gaussian noise and measures the resulting gradient magnitude spike, serving as an efficient proxy for the Hessian spectrum. Experiments show EPGS significantly outperforms entropy-based and representation-based baselines in identifying high-confidence factual errors.