Large language models often produce text that reads fluently but is factually wrong. A new arXiv paper describes a class of detectors that aim to catch such hallucinations by looking at the token-probability statistics the model emits during generation.
The approach is called gray-box because it sits between black-box and white-box methods. It does not require full white-box access to the model's internals, only the probability statistics that come out during generation. This could make hallucination detection more practical for models where internal weights are not available.
The abstract provides limited detail, so specifics on the method's architecture, evaluation, or performance are not yet available from this source. The paper is listed as arXiv:2610.19996v1.