GuardAgent: A Memory-Augmented Agent Framework for Adaptive Detection of Adversarial Promotion in LLM-Based Recommender Systems
GuardAgent: A Memory-Augmented Agentic Framework for Adaptive Detection of Adversarial Promotions in LLM-Based Recommendation
AI summary
Researchers propose GuardAgent, an LLM agent framework that operates through perception, action, reflection and memory loops to adaptively detect adversarial promotion in LLM-based recommender systems. The authors call it the first systematic study of this class of attacks; unlike static, one-shot LLM detectors, GuardAgent continuously learns and self-evolves to adapt to unseen threats, generating interpretable reasoning traces. Experiments on real datasets show its generalization performance reaches the state of the art.