【摘要】 针对通用大语言模型在军事、医疗、金融等专业问答场景中幻觉严重、专业性不足的问题，以及检索增强生成（retrieval-augmented generation,RAG）技术缓解模型幻觉过程中易受检索文档虚假信息误导的新问题，提出一种可信度感知多策略图RAG(credibility-aware multistrategy graph RAG,CAMG-RAG)框架，构建“可信知识生成—高可信信息检索—注意力权重优化”三级协同架构。首先，通过改进的知识图谱链接器生成含实体、关系、三元组及可信度评分的图构件；其次，基于可信度评估机制，实现多策略图检索；最后，定位大模型中对知识图谱元素敏感的“有影响力注意力头”，依据可信度评分修正其注意力权重，引导模型优先聚焦高价值知识。实验结果表明，在军事领域QAonMilitaryKG数据集上，CAMG-RAG能够显著提高大模型对抗幻觉的能力，基于Qwen3-4B模型，其忠实度、答案正确性分别达到0.746、0.863，比微调大模型提高14.6%、5.9%。更多还原

【Abstract】 To address the severe hallucinations and insufficient professionalism problems of general large language models in professional question-answering scenarios such as the military, medical, and financial fields, as well as the new problem that retrieval-augmented generation(RAG) technology is susceptible to being misled by false information in retrieved documents during the process of mitigating model hallucinations, a credibility-aware multi-strategy graph RAG(CAMG-RAG) framework was proposed to construct a three-level collaborative architecture of “trustworthy knowledge generation-highcredibility information retrieval-attention weight optimization”. Firstly, an improved knowledge graph linker was used to generate graph components containing entities, relations, triples, and credibility scores. Secondly, multi-strategy graph retrieval was realized based on a credibility evaluation mechanism. Finally, “the influential attention heads” that are sensitive to knowledge graph elements in the large language model were identified, and their attention weights were revised according to credibility scores to guide the model to prioritize high-value knowledge. Experimental results show that on the QAonMilitaryKG dataset in the military field, CAMG-RAG can significantly enhance the antihallucination capability of large language models. Based on the Qwen3-4B model, the faithfulness and answer correctness reach 0. 746 and 0. 863, respectively, which are 14. 6% and 5. 9% higher than those of the fine-tuned large model.更多还原