Predictable Failure in Multi-Hop Retrieval: Score-Distributional Confidence Scoring and Abstention

In 30 Sekunden

Was passiert ist

Research shows multi-hop retrieval failures cluster predictably by query structure, enabling confidence scoring without extra LLM calls to reduce confident-wrong answers.

Warum es zählt

Engineers building retrieval systems need this when balancing accuracy and cost in multi-hop question answering across dense and LLM-judge architectures.

Achtung

The method's feature importance varies by dataset; models trained on one benchmark show minimal transfer loss but optimal features differ across datasets.