A Case-Based Verification Framework for Detecting and Reducing Hallucinations in Generative AI

Abstract

Generative artificial intelligence has become an important technology for knowledge creation, education, healthcare, finance, and organizational decision-making. However, its practical adoption is limited by hallucinations, where generated responses contain fabricated, unsupported, outdated, contextually inappropriate, logically inconsistent, or incorrectly cited information. Existing verification approaches frequently assess responses independently, rely on imperfect evidence retrieval, and lack mechanisms for reusing previously verified knowledge, resulting in limited adaptability and explainability. This study proposes a Case-Based Verification Framework (CBVF) to improve the reliability of generated responses through experience-driven verification. The framework employs the four stages of Case-Based Reasoning—Retrieve, Reuse, Revise, and Retain—and integrates atomic-claim decomposition, semantic case retrieval, evidence entailment, source-reliability weighting, semantic-entropy estimation, hallucination-risk calibration, selective verification, explainable correction, and incremental case-base maintenance. The framework is evaluated using approximately 2,000 prompt–response instances constructed from six publicly available benchmark datasets spanning general knowledge, multi-hop reasoning, scientific verification, citation verification, temporal reasoning, and hallucination evaluation. Performance is compared with unverified generation, self-consistency verification, retrieval-augmented generation, and retrieval-augmented generation combined with rule-based fact-checking using claim groundedness, evidence support, calibration quality, hallucination reduction, selective verification, correction safety, retrieval effectiveness, and Bayesian multilevel analysis. The experimental evaluation demonstrates that the proposed framework consistently outperforms the baseline approaches by improving factual grounding, evidence alignment, calibration accuracy, and retrieval quality while substantially reducing hallucinations and preserving response relevance and semantic meaning. The Bayesian analysis further confirms statistically robust improvements across multiple benchmark domains. The findings indicate that integrating Case-Based Reasoning with evidence-driven verification provides an adaptive, explainable, and continuously improving mechanism for enhancing the trustworthiness of generative artificial intelligence in applications requiring reliable and evidence-supported information.