New VerTox framework poisons neural ranking models with deceptive AI-generated documents

Researchers have developed VerTox, a novel framework that uses verifiable reward-guided reinforcement learning to poison the corpora used by neural ranking models. This method trains compact LLMs to generate malicious documents that can distort ranking behavior and degrade downstream applications like retrieval-augmented generation (RAG). Experiments show VerTox achieves high attack success rates, producing fluent and difficult-to-detect adversarial documents that outperform target documents across various ranking architectures and a commercial embedding model. AI

IMPACT Introduces a new attack vector against retrieval systems, potentially impacting the reliability of AI-powered information access.

RANK_REASON Academic paper detailing a new method for attacking AI systems. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• LLMs
• retrieval-augmented generation
• RLVR
• VerTox

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →