Consistency-Driven Iterative Retrieval-Augmented Generation Using Cross-Evidence Agreement as a Stopping Signal
Year: 2026
Type: article
Abstract: Iterative retrieval-augmented generation improves multi-hop question answering, but deciding when to stop retrieving remains an open problem. Fixed retrieval depths spend the same effort on every question, and confidence-based triggers inherit the overconfidence of the generator. This paper studies a stopping test that needs no training and no access to token probabilities. After each retrieval round the generator answers once from all retrieved passages and again from leave-one-group-out subsets of them. Retrieval stops only when every subset reproduces the full answer. Otherwise the facts that the generator reports as missing become the next queries. A pilot study on 100 HotpotQA questions shows that partial agreement between subsets carries no usable signal, which motivates the unanimity rule. The method is then evaluated on 300 held-out HotpotQA questions against single-pass retrieval, fixed-depth retrieval, and a self-consistency twin that makes the same generator calls per round but resamples answers from unchanged evidence. The method matches fixed-depth retrieval on answer F1, 61.3 against 61.2 percent, while making 9.1 generator calls per question against 3.0, and no difference in F1 between the method and either iterative baseline has a confidence interval that excludes zero. Unanimous agreement identifies reliable answers under both controllers, but resampling flags 167 of 252 answers at 69.5 percent precision where evidence removal flags 41 at 75.6 percent, so varying the evidence yields a stricter test rather than a better one.
Authors: Jain, Adish, Sah, Krishan Kumar
Institution: Chitkara University
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7235871053
