金融证据拥挤：诊断并缓解RAG中约束诱导的证据位移
原标题：Financial Evidence Crowding: Diagnosing and Mitigating Constraint-Induced Displacement in Retrieval-Augmented Generation
AI 摘要
论文研究金融问答中RAG的排序失败问题，提出「金融证据拥挤」概念：与查询主题匹配但期间、分部、指标范围或表格范围冲突的候选段落会挤占有限上下文槽位，导致支撑证据被挤出。作者构建FinDeCrowd-Stress受控诊断，在FinDER测试集上发现不兼容候选池使Recall@10下降0.147；并提出FinDeCrowd-RAG，通过类型化兼容性与局部词法竞争的学习分数校正加身份门控重排，在相同候选上将Recall@10从0.757提升至0.902，在非注入索引上从0.420提升至0.492，并改善FinanceBench与FinQA的答案准确率和引用召回。
正文节选
Financial Evidence Crowding: Diagnosing and Mitigating Constraint-Induced Displacement in Retrieval-Augmented Generation Abstract Retrieval-augmented generation (RAG) retrieves candidate evidence and sends only a limited top-ranked subset, the top- context, to a generator. In financial question answering, passages can match a query’s topic while conflicting with its period, segment, metric scope, or table scope. We study the resulting set-level ordering failure, which we call financial evidence