The authors systematically analyzed the construction of poison sets used in prior evaluations of per-document RAG poisoning defenses.
Query-time RAG poisoning defenses exploit artifact cues, not true detection
The authors show that removing construction artifacts from poison sets, such as verbatim target questions or length differences, collapses detection performance while the ranking of candidates remains intact.
Independent
Pushkal Kumar · Arvin Sahaym
Research Digest··3 min read
This paper re-evaluates query-time filters for defending retrieval-augmented generation (RAG) against data poisoning.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.