DnD RAG Poisoning Spreads a False Answer Across Stylistically Diverse Passages Plus One Doubt Passage to Evade Clustering Defenses
arXiv·low signal
arXiv 2609.27090 (22 Sep) avoids the correlated near-duplicate passages that similarity- and conflict-aware defenses filter. It spreads support for the target answer across stylistically varied passages and adds one passage casting doubt on the true answer's evidence. Across two QA datasets, three LLMs and nine RAG configurations it matches or beats prior attacks in most settings, with its largest gains against clustering- and conflict-aware defenses.