MIRAGE: Defending Long-Form RAG Against Misinformation Pollution
2026-07-06 • Computation and Language
Computation and Language
AI summaryⓘ
The authors show that when language models use information from other sources to answer questions, the information can sometimes be wrong or misleading. They introduce MIRAGE, a method that checks if different pieces of information agree before using them to generate answers, or decides to rely on the model's own knowledge if the information is inconsistent. They tested MIRAGE on several question-answering tasks and found it helps keep answers accurate even when the retrieved information is polluted with errors. Their approach works without changing the original model and is better than previous methods designed for similar problems.
Retrieval-Augmented GenerationLarge Language ModelsNatural Language InferenceClaim GraphFactualityPollution in RetrievalLong-Form Question AnsweringRobustnessModel-Agnostic MethodsInformation Consistency
Authors
Saadeldine Eletter, Ruihong Zeng, Yuxia Wang, Maxim Panov, Aleksandr Rubashevskii, Preslav Nakov
Abstract
Retrieval-Augmented Generation (RAG) improves factuality by grounding LLMs in external evidence, but real-world retrieval is often polluted: semantically relevant passages may contain subtle misinformation, misleading framings, or fabrications. We introduce MIRAGE, a training-free, model-agnostic defense for long-form RAG. MIRAGE builds an NLI-based cross-document claim graph and applies a Defended-Claims Gate to either condition generation on a consistent, multi-source supported subset or to block retrieval and answer parametrically. We also release a minimal-edit pollution protocol spanning four perturbation families (Unambiguous, Conflicting, Misleading, Fabricated) to construct matched clean, mixed, and fully polluted evaluation regimes. Across four long-form QA benchmarks and multiple commercial and open-weight LLMs, pollution severely degrades vanilla RAG, while MIRAGE consistently restores factuality under mixed and fully polluted evidence and outperforms prior robust-RAG methods. Our implementation and datasets are available at https://github.com/SaadElDine/MIRAGE.