A Geometric Method to Spot Hallucinations Without an LLM Judge

A novel geometric method called Displacement Consistency (DC) can detect LLM hallucinations without relying on another LLM as a judge. This technique analyzes the vector embeddings of question-answer pairs, observing that the displacement vectors for grounded responses within a specific domain point in a consistent direction. Hallucinated answers create displacement vectors that do not align with this local pattern, making them geometrically identifiable. While highly effective, the method is domain-specific, requiring a small, matched calibration set for each application area. This suggests that grounding is not a universal property but a locally coherent structure within the embedding space, similar to a fiber bundle in geometry.