cs.CLOct 29, 2026

Illusory Truth or Salient Anomaly? Model-Dependent Responses to Repeated Poisoned Passages in RAG-Style Contexts

Authors: Mehrdad Ghassabi, Audrina Ebrahimi, Sadra Hakim, Hamidreza Baradaran Kashani, Zahra Zojaji, Amin Milani Fard

Organizations: University of Isfahan · University of Texas at Dallas · New York Institute of Technology, Vancouver Campus

Abstract

Repeated passages in retrieved evidence may influence how language models identify misinformation, even when those passages restate a single underlying claim rather than provide independent support. We compare four instruction-tuned models on 200 claims from the BEIR SciFact test split using paired, fixed seven-document contexts. The minority condition contains one poisoned document and six non-target documents, whereas the majority condition contains the original poisoned document, five paraphrases, and one non-target document. Poison detection rates increase from 45.5% to 66.0% for Apertus-8B and from 4.0% to 15.5% for Aya-Expanse-8B, but decrease from 80.0% to 47.5% for Mistral-Nemo and from 25.0% to 9.5% for Gemma-3-12B-IT. All four paired comparisons are statistically significant. These contrasting results suggest that repeated poisoned evidence does not affect all models in the same way. One possible interpretation is that repetition makes the poisoned claim more conspicuous to some models, potentially facilitating its detection, while other models may treat repeated formulations as reinforcing evidence, potentially reducing their ability to identify the poisoned claim. Although these interpretations require further investigation, the findings highlight the importance of model-dependent responses to repeated evidence in retrieval-augmented settings.

Figures & tables

Explore similar work

Aug 17, 2026

When Detection Does Not Guarantee Resistance: Reasoning and Poisoned Context in RAG

Retrieval-Augmented Generation (RAG) exposes large language models to knowledge-poisoning attacks, where misinformation injected into retrieved documents can influence model outputs. Prior work has shown that models may detect contradictory evidence yet still allow it to influence their responses, revealing a gap between monitoring and control. We investigate whether deliberative reasoning changes this relationship. Because attack success and poison detection alone do not reveal whether detected poison continues to influence the final answer, we use two complementary measures: Cordon Rate, the probability that a model detects poison and nevertheless produces a poison-aligned answer, conditioned on a non-poison-aligned no-RAG response; and Leakage Rate, the probability that poisoned context influences the answer despite an explicit instruction to ignore retrieved documents. Across 200 SciFact questions, we compare reasoning-disabled and reasoning-enabled configurations of DeepSeek-V4-Flash and Qwen3.6-Plus. For DeepSeek-V4-Flash, reasoning reduces Cordon Rate from 0.205 to 0.075 and Leakage Rate from 0.235 to 0.140, despite increasing Attack Success Rate from 0.233 to 0.298 and decreasing Poison Detection Rate from 0.965 to 0.665. Qwen3.6-Plus shows the same qualitative pattern. These results demonstrate that poison detection and downstream influence capture distinct aspects of poisoning behavior.
Jul 6, 2026

MIRAGE: Defending Long-Form RAG Against Misinformation Pollution

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.
Sep 8, 2026

In RAG We Trust? Measuring Robustness of Retrieval-Augmented Generation Under Post-Retrieval Context Tampering

Retrieval-augmented generation (RAG) grounds a language model in retrieved documents, which reduces hallucination but creates a new attack surface: if retrieved text is tampered with, the model may repeat the falsehood. We study how much a small quantized model, Llama 3.1 8B, degrades when a fraction of its retrieved context is poisoned. Three corruption strategies are tested, entity swap, number swap, and negation, each applied to zero, one, two, or three of the three retrieved passages, over a factorial sweep of 588 runs on a fact-checking task built from FEVER. Accuracy falls from 77.9% on clean context to 43.5% when all three passages are corrupted. Entity swap flips the largest share of answers that were correct on clean context. Number-based corruption stays flat while poisoned passages are a minority and jumps once they form a majority, a pattern we re-check with query-level bootstrap intervals. The model rarely invents new falsehoods; its dominant reaction is to abstain, and a lexical overlap proxy of unsupported generation falls under attack rather than rising. The study is a small-scale measurement with coarse automated labels; we treat the strategy contrasts as suggestive until decoding is controlled and stronger adjudication is in place.