Year: 2026
Type: peer-review
Abstract: The problem the paper takes on is the right one.Corpus poisoning that can outnumber clean neighbors is exactly where majority voting and simple aggregation break, and Table 1 makes that failure visible for RobustRAG Keyword, InstructRAG ICL, and ASTUTE RAG under PoisonedRAG.Framing TrustRAG as a training-free, post-retrieval lter that sits after the retriever, with code at https://github.com/HuichiZhou/TrustRAG, is a clean engineering posture.The two-stage split-K-means with ROUGE-L n-gram preservation, then con ict removal that mixes internal knowledge, consolidation, and self-assessment-is also a coherent design.My concern is that several strong claims in the abstract and §6 sit ahead of what the threat model and the Stage-1 numbers can carry.
Source: Qeios
DOI: 10.32388/iv9ww4
Author: Iliano Fasolino
Institution: University of Milan
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7214967442