MIRAGE defends long-form RAG against misinformation pollution

The authors introduce MIRAGE, a training-free and model-agnostic defense mechanism designed to protect long-form Retrieval-Augmented Generation (RAG) systems from misinformation in retrieved evidence. MIRAGE constructs an NLI-based cross-document claim graph and utilizes a Defended-Claims Gate to either condition generation on consistent, multi-source supported claims or block retrieval entirely.

The system is evaluated alongside a minimal-edit pollution protocol covering four perturbation families: Unambiguous, Conflicting, Misleading, and Fabricated. Testing across four long-form QA benchmarks and multiple commercial and open-weight LLMs demonstrates that while pollution severely degrades vanilla RAG, MIRAGE consistently restores factuality under mixed and fully polluted evidence.

MIRAGE outperforms prior robust-RAG methods in maintaining accuracy despite the presence of subtle misinformation or fabrications in the retrieved passages. The implementation and datasets are.