Metraipolis › archive › dc_ff82614d5ce6

In RAG We Trust? Measuring Robustness of Retrieval-Augmented Generation Under Document Poisoning (2026) · DOI 10.32388/7hxkqe

Source
Metraipolis archive
Original
https://doi.org/10.32388/7hxkqe
License (as archived)
CC-BY-4.0
Archived
2026-10-07T09:07:58.125Z
SHA-256 of the archived text
0f493b81441d1df88c2846d8cb720e1d676474990ec02d17485122ccade68021
Stable URL of this page
https://metraipolis.com/documento/dc_ff82614d5ce6

How to cite

In RAG We Trust? Measuring Robustness of Retrieval-Augmented Generation Under Document Poisoning (2026) · DOI 10.32388/7hxkqe — Metraipolis archive, https://doi.org/10.32388/7hxkqe (CC-BY-4.0); archived copy: https://metraipolis.com/documento/dc_ff82614d5ce6 (Metraipolis archive dc_ff82614d5ce6, sha256 0f493b81441d1df88c2846d8cb720e1d676474990ec02d17485122ccade68021)
BibTeX
@misc{dcff82614d5ce6,
  title = {In RAG We Trust? Measuring Robustness of Retrieval-Augmented Generation Under Document Poisoning (2026) · DOI 10.32388/7hxkqe},
  howpublished = {Metraipolis archive},
  url = {https://doi.org/10.32388/7hxkqe},
  note = {License CC-BY-4.0; archived copy https://metraipolis.com/documento/dc_ff82614d5ce6; sha256 0f493b81441d1df88c2846d8cb720e1d676474990ec02d17485122ccade68021},
  year = {2026}
}

Archived text

Reproduced under CC-BY-4.0. Attribution: Metraipolis archive, https://doi.org/10.32388/7hxkqe. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Articolo recente con licenza aperta: scheda dai metadati Crossref (posted-content).
Titolo: In RAG We Trust? Measuring Robustness of Retrieval-Augmented Generation Under Document Poisoning
Autori: Iliano Fasolino
Pubblicato: 2026-08-11 · editore: Qeios Ltd
DOI: https://doi.org/10.32388/7hxkqe
Licenza dell'articolo: CC-BY-4.0

Abstract (testo degli autori, CC-BY-4.0):
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 is the most damaging in terms of answers flipped from correct to wrong, and number-based corruption shows a sharp threshold once poisoned passages form the majority. The model rarely invents new falsehoods; instead its dominant reaction is to abstain, and hallucination actually drops under attack. These results quantify a practical weakness of RAG and identify abstention, not fabrication, as the main behaviour to plan for.

Fonte: https://api.crossref.org/works/10.32388%2F7hxkqe — metadati Crossref, consultati il 2026-10-06. Interrogazione: «retrieval augmented generation».