NxN E-valuation Certifies LLM Hypotheses to Combat Hallucination

Large Language Models excel at generating hypotheses but frequently suffer from hallucination, making their direct outputs unreliable. Existing verification methods, such as self-correction or held-out testing, often fall short. This new research proposes NxN E-valuation, an algorithm designed to certify LLM-generated hypotheses. The method operates by exploiting large existing datasets, using different samples within the dataset to serve as null hypotheses for one another. This approach effectively implements a conditional randomization test, providing a robust way to verify hypotheses without requiring the construction of a dedicated, case-specific null. It offers a potentially superior alternative to current LLM verification techniques, particularly for hypotheses applicable to individual data samples.