CQC-RAG: Enhancing the Robustness of RAG Systems via Cross-Query Consistency

This article introduces the CQC-RAG framework, which aims to address the hallucination problem in RAG systems. Based on the cross-query consistency hypothesis, this framework implements a self-evaluation mechanism without external supervision, achieves significant performance improvements in multiple question-answering benchmarks, and provides a new path for enhancing the robustness of RAG systems.

RAG is a mainstream technology for improving the factual accuracy of large language models, but it has retrieval sensitivity issues (semantically equivalent queries may lead to different results) and noise-induced hallucinations. Existing multi-path reasoning methods have limitations: crude diversity injection (relying on decoding randomness) and narrow answer evaluation perspective (single query view).

The core hypothesis is cross-query consistency—correct answers have stable confidence across semantically equivalent query variants, while hallucinated answers fluctuate greatly. The framework flow includes: 1. Query Rewriting (generating semantically equivalent query variants); 2. Document Re-ranking (constructing query-conditional reasoning context); 3. Answer Extraction (generating candidate answers with evidence); 4. Stability Evaluation (selecting the answer with the most stable confidence).

1. Self-evaluation mechanism without external supervision; 2. No reliance on expanding retrieval coverage; 3. Controllable query-level diversity (more reliable than decoding randomness).

On the TriviaQA dataset, compared to the strongest multi-query baseline, it improved the exact match score by 4.76 percentage points; on the MuSiQue multi-hop QA dataset, it improved by 9.12 percentage points. These improvements were achieved without external supervision and without expanding retrieval coverage.

CQC-RAG provides a new paradigm for improving RAG reliability, indicating that answer quality can be enhanced through query strategy and evaluation mechanism design. In the future, it can be extended to multi-document summarization, fact-checking, and other scenarios, and we can also explore combining with other uncertainty quantification methods to further improve the accuracy of consistency evaluation.