SAMPLING OF LARGE LANGUAGE MODELS WITH EQUIVALENCE TESTING

Abstract

Generative pre-trained large language models (LLMs) can generate domain-specific text responses in various formats such as JSON, XML, HTML, SQL, or programming languages. However, LLMs can "hallucinate," generating false or nonsensical responses that deviate from reality, thereby undermining confidence in their results or worse. The revealed techniques employ a sampling-based approach and an equivalence checker. Multiple responses (samples) to a prompt are generated by the LLM; if they are equivalent, the LLM is likely responding correctly. If the samples are inconsistent or contradictory, it is more likely that the LLM is hallucinating or that the prompt is ambiguous.An automated reasoning equivalence checker is used to verify the functional equivalence of the samples, thereby providing a procedure to detect and potentially resolve problems with hallucinations in LLM-generated responses.