Researchers propose a geometric method for detecting LLM hallucinations by analyzing the mathematical structure of text embeddings. Instead of using another LLM as a fallible judge, this approach examines the displacement vectors between questions and answers in embedding space. For factual responses within a domain, these vectors show consistent directional patterns, while hallucinations break this geometric consistency.

The method uses a reference set of verified question-answer pairs to establish a domain-specific baseline. When a new response is generated, its displacement vector is compared against the mean direction of neighboring examples in the reference set. Grounded responses show high alignment scores, while hallucinations produce anomalous vectors with low scores, revealing the fabrication without requiring additional LLM inference.

This technique achieves near-perfect discrimination (AUROC 1.0) across multiple benchmarks and embedding models, including MPNet and BGE variants. However, it requires domain-specific calibration—a one-time offline cost—and doesn't transfer between domains like law and medicine. The approach offers a practical alternative to costly LLM-as-judge systems, adding minimal latency while avoiding recursive uncertainty.