Citation-Consistent Voting for Permutation-Robust Retrieval-Augmented Generation

Abstract

Retrieval-augmented generation (RAG) systems are sensitive to the ordering of retrieved documents, with language models exhibiting position bias that causes inconsistent outputs across different document permutations. We propose Citation-Consistent Voting (CCV), a training-free method that improves RAG robustness by aggregating answers based on document-ID agreement rather than answer frequency. CCV prompts the generator to cite supporting documents, maps citations to permutation-invariant document identifiers, and selects answers with the highest citation consistency across multiple permutations. On NaturalQuestions with Qwen3-8B and Contriever retrieval, CCV achieves 46.37% SubEM, outperforming majority voting by +0.19 points at $K=20$ permutations. The improvement scales monotonically with the number of permutations, and diagnostic analysis confirms that citation agreement correlates significantly with answer correctness ($p = 1.14 \times 10^{-5}$). CCV requires no additional training and is compatible with any RAG system that can produce structured citations.