多跳检索可预测失败：分数分布置信度评分与弃答
原标题：Predictable Failure in Multi-Hop Retrieval: Score-Distributional Confidence Scoring and Abstention
AI 摘要
该论文研究多跳检索中的失败预测问题，提出CWAR可归约性定理和特征机制互补性命题，证明检索失败在结构上可预测且不同失败机制下主导特征不同。作者提出RegimeAbstain方法，利用最多九个查询-ANN结构特征计算检索置信度分数，无需额外LLM调用即可实现校准弃答策略。在MuSiQue、2WikiMultiHopQA、HoVer三个基准和两种检索架构上，该方法在五种失败机制下均取得最佳或并列最佳AUC-AC，MuSiQue上50%覆盖率时将CWAR从39.5%降至20.6%，且模型可跨数据集迁移。
正文节选
Predictable Failure in Multi-Hop Retrieval: Score-Distributional Confidence Scoring and Abstention Abstract Multi-hop retrieval failures are not uniformly distributed across queries: they cluster in structurally predictable subpopulations. We prove two results formalizing this structure. First (CWAR Reducibility, Theorem 1): confident-failure reduction is achievable if and only if retrieval features carry mutual information about success — a condition satisfied by LLM-judge pipelines but substan