DEAM: Knowledge-Aware Retrieval with Dynamic Evidence Grounding

To mitigate the long-standing issue of hallucinations in large language models (LLMs) for question-answering systems and enhance response interpretability, we propose the Dynamic Evidence Augmentation Mechanism (DEAM), a framework applicable to zero-shot and few-shot scenarios. The framework synergistically integrates the implicit knowledge of LLMs with structured external knowledge from knowledge graphs. Its core innovation lies in the dynamic reasoning supplementation mechanism, which utilizes retrieved knowledge for knowledge filtering and evidence-based reasoning to effectively reduce interference from redundant information while ensuring comprehensive knowledge integration. Experiments on the CMCQA dataset demonstrate that DEAM achieves an accuracy of 0.9444 under the BERTScore evaluation metric, outperforming state-of-the-art baseline models. Ablation studies further validate the effectiveness of the dynamic reasoning supplementation mechanism. To assess factual consistency, we introduces a knowledge graph-based verification method to quantify the factuality of model-generated answers. DEAM achieves 0.9347 under this criterion, demonstrating a significant improvement over baseline models.