Memory-Augmented Knowledge Fusion with Safety-Aware Decoding for Domain-Adaptive Question Answering

Domain-specific question answering (QA) systemsfor services face unique challenges in integrating heterogeneousknowledge sources while ensuring both accuracy and safety.Existing large language models often struggle with factual consistencyand context alignment in sensitive domains such ashealthcare policies and government welfare. In this work, weintroduce Knowledge-Aware Reasoning and Memory-AugmentedAdaptation (KARMA), a novel framework designed to enhanceQA performance in care scenarios. KARMA incorporates adual-encoder architecture to fuse structured and unstructuredknowledge sources, a gated memory unit to dynamically regulateexternal knowledge integration, and a safety-aware controllabledecoder that mitigates unsafe outputs using safety classificationand guided generation techniques. Extensive experiments on aproprietary QA dataset demonstrate that KARMA outperformsstrong baselines in both answer quality and safety. This studyoffers a comprehensive solution for building trustworthy andadaptive QA systems in service contexts.