Multi-Hop Retrieval Failures Cluster Predictably, and Nine Structural Features Predict Them Without an Extra LLM Call

arXiv 2609.22056·medium signalXBlueskyLinkedInCopy link

This paper proves two results about multi-hop retrieval: confident-failure reduction is achievable only if retrieval features carry mutual information about success, which holds for LLM-judge pipelines but is much weaker for dense-only setups, and no single ANN score feature dominates across failure regimes (query length wins on MuSiQue, hop-1 concentration on HoVer). RegimeAbstain instantiates this as a Retrieval Confidence Score, a logistic function over up to nine query-ANN structural features that require no additional LLM call, driving a calibrated abstention policy. Across MuSiQue, 2WikiMultiHopQA and HoVer with two retrieval architectures, the Confident-Wrong-Answer Rate spans 14.5% to 62.1% and RCS is best or co-best on AUC-AC in all five conditions.