MIRAGE: Defending Long-Form RAG Against Misinformation Pollution

2026-07-06

AI Summary

This work addresses the significant degradation in factual accuracy of long-context retrieval-augmented generation (RAG) systems when exposed to polluted retrieval results containing errors, misinformation, or fabricated content. The authors propose a training-free, model-agnostic defense mechanism that constructs a cross-document claim graph via natural language inference and employs defensive claim gating to select multi-source consistent information for generation or, under severe pollution, falls back to parametric knowledge. They introduce, for the first time, a training-free multi-source consistency verification framework and design a minimally edited contamination protocol covering four perturbation types to enable fine-grained evaluation. Experiments demonstrate that the method substantially restores factual accuracy across four long-context question-answering benchmarks and consistently outperforms existing robust RAG approaches on both commercial and open-source large language models.

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.

Problem

• Retrieval-Augmented Generation
• misinformation pollution
• factuality
• long-form QA
• evidence reliability

Innovation

• Retrieval-Augmented Generation
• misinformation defense
• natural language inference
• claim consistency
• training-free robustness