An Explainable AI Framework Hallucination Mitigation in Large Language Models

Abstract

Machine learning (ML) systems are rapidly becoming used in high stakes decisions in fields like finance, healthcare, criminal justice, and autonomous systems, but they are becoming increasingly complex, thus becoming opaque black boxes that are hard to interpret or trust by the stakeholders. The paper introduces Large language models (LLMs) are now used to power search, tutoring, clinical support and software engineering, but they continue to generate statements that are fluent but backed by no evidence or even their own input. This is commonly referred to as hallucination, and it constrains faith in high-stakes deployment and end-to-end Explainable Artificial Intelligence (XAI) system that maximizes the transparency, responsibility, and reliability of the ML models without significantly affecting the predictive performance. The framework combines global explanation techniques (SHAP, partial dependence plots), local explanation techniques (LIME, Anchors), and a special trust-and-fidelity evaluation module, which quantitatively assesses the trustworthiness of the generated explanations. The framework was tested on a credit-risk benchmark data set with five model families, such as logistic regression, decision trees, random forests, gradient boosting and neural networks. The protocol records the rate of hallucinations with latency, such that the accuracy gain can be traded off with cost. Resulting in the lowest rates of hallucinations, retrieval and verification when combined working with a reference configuration and lightweight prompting when used provided moderate benefits with minimal costs. We leave to unresolved issues, such as robust uncertainty estimation, the assessment of long outputs, and multilingual mitigation The experimental findings indicate that the suggested framework has a predictive accuracy of 90.2 and increases the mean user trust score by 91.3 out of 100 points as compared to the 63.5 at the baseline. The results affirm that using a combination of several complementary explanation methods and formal trust-evaluation mechanism results in ML systems that are accurate, explainable, and reliable enough to be used in the real world.