SIREN (Luring LLMs onto the Rocks): PAIR-Driven Preference Manipulation in Web-RAG Recommenders

The document is an arXiv computer-science information-retrieval feed containing research papers on recommender systems, multilingual reranking, retrieval-augmented generation, conversational search, legal retrieval, content exploration, and related applications. The primary security-relevant item is SIREN, which demonstrates automated content poisoning and preference manipulation against web-RAG LLM recommenders by editing already retrieved webpages to promote a chosen entity to rank one. The study reports successful rank-one manipulation in 62 of 124 trials and an average reproduction rate of