<ul><li>The study introduces PoisonSwarm, a novel framework for synthesizing harmful information by utilizing model crowdsourcing.</li><li>PoisonSwarm aims to address challenges in generation reliability and content diversity faced by Large Language Models (LLMs) in synthesizing harmful data.</li><li>The framework generates diverse harmful data by employing a model crowdsourcing strategy, utilizing abundant benign data as templates, and performing unit-by-unit toxification.</li><li>Experimental results show that PoisonSwarm outperforms existing methods in synthesizing various categories of harmful data with scalability and diversity.</li></ul>

PoisonSwarm: Universal Harmful Information Synthesis via Model Crowdsourcing

Discover more