CQC-RAG boosts RAG reliability with cross-query consistency check

CQC-RAG addresses a core weakness in Retrieval-Augmented Generation (RAG): answers can vary dramatically based on how a question is phrased or which documents are retrieved. The paper introduces the Cross-Query Consistency Hypothesis — correct answers maintain high confidence across semantically equivalent but syntactically diverse queries, while hallucinations show unstable confidence. The framework first rewrites the original question into multiple meaning-preserving variants, then reranks a shared document pool to build query-specific reasoning contexts. An evidence-grounded protocol extracts answer-evidence pairs, and selection is based on confidence stability across these contexts. This design enables self-evaluation without external supervision and does not require expanding retrieval coverage.

Experiments on four open-domain QA benchmarks confirm the approach's effectiveness. On TriviaQA, CQC-RAG achieves a +4.76 percentage point improvement in exact match over the strongest prior multi-query baseline; on the more challenging MuSiQue dataset, the gain reaches +9.12 pp EM. The method uses only standard LLM inference and retrieval — no additional training or external knowledge bases are needed. By making RAG systems more robust to retrieval noise and query phrasing, CQC-RAG offers a practical path to reducing hallucinations in production applications, from enterprise chatbots to research assistants.