Automated SELF-KNOWLEDGE prompt optimization

Large Language Models are gradually being adopted in high-risk domains such as healthcare, finance, insurance, and law. Although they are capable of learning highly fluent answers, they continue to have the problem of hallucinations. Work in the past has proposed the SELF-KNOWLEDGE approach for prompting, in which the model begins by indicating its familiarity with a query before it proceeds to answer it. This can easily minimize hallucinations without having to retrain the model itself. Although techniques for SELF-KNOWLEDGE prompts are available, the problem of scalability arises because such prompts are manually designed. In this paper, we propose an automated approach to optimize the prompts of the SELF-KNOWLEDGE system. To this end, we develop and describe three complementary techniques that differ in these goals and may be used separately or combined: (1) the evolution of prompts based on genetic algorithms aiming to search the vast space of possible prompts and control hallucinations effectively, (2) domain-specific optimization methods to include specific concepts, regulatory needs, and risks into the structure of the self-knowledge system, and (3) dynamic prompting to dynamically alter the language of the prompts and the threshold of prompting and decision abstinence according to the observed confidence and behavior. Finally, we describe the design of experiments to optimize the techniques of all the mentioned problems simultaneously.