EVENT DOSSIER
Researchers have proposed a retrieval-enhanced generation framework called APT-RAG, aimed at addressing the issues of structural rigidity and neglect of evidence topology in existing structured methods in intensive question-answering tasks. This framework dynamically expands the reasoning structure based on question dependencies and evidence requirements, utilizing brother node evidence reuse, direct retrieval, and sub-node evidence aggregation to improve evidence coverage. Evidence-guided batch processing is also introduced to optimize generation efficiency. Experiments show that APT-RAG performs better than existing structured RAG methods in intensive question-answering benchmarks, and the related code has been made open-source.
提出名为 APT-RAG 的自适应规划与拓扑感知证据获取检索增强生成框架，旨在解决现有结构化方法在密集型问答中存在的结构僵化及忽视拓扑结构的证据收集问题。该框架通过依据问题依赖和证据需求动态扩展推理结构，并利用兄弟节点证据复用、直接检索及子节点证据聚合提升证据覆盖度；同时引入证据引导批处理答案生成以减少生成开销。实验表明，APT-RAG 在密集型问答基准测试中优于现有结构化 RAG 方法，相关代码已开源。