Mitigating Implicit Hallucinations in Large Language Models Based on Progressive Prompt Chains

The phenomenon of hallucinations in large language models (LLMs) refers to the generation of text that appears contextually coherent yet contains factual inaccuracies or logical inconsistencies. While explicit hallucinations typically refer to direct and verifiable, implicit hallucinations – subtle inaccuracies based on commonsense logic – prove more challenging to discern. This paper categorizes hallucinations into explicit and implicit types, with a focus on mitigating the critical issue of implicit hallucinations. To enhance LLMs’ self-awareness in detecting subtler inaccuracies, we begin with crowdsourced experiments aimed at identifying the most challenging type of hallucination: commonsense logic hallucinations. Based on these insights, we propose a novel framework utilizing progressive prompt chains. This framework has two interconnected phases. In the first phase, we extract core entities from user queries and link them to external knowledge bases to retrieve relevant contextual summaries. In the second phase, we implement multi-phase self-validation prompting, which enables iterative refinement of responses by applying truthfulness discrimination. Our extensive experiments across ten LLMs and two benchmark datasets demonstrate that the progressive prompt chains framework achieves an impressive 93% accuracy in self-assessing answer authenticity. Additionally, it improves answer generation quality accuracy by 41.6% compared to traditional prompting methods.