ArXiv:2609.38296 Shows Resonance Radicalizes Target LLMs More Than Persuasion

The study ran controlled dialogues between a target LLM role-playing demographic personas and an influencer LLM tasked with shifting beliefs. Two pathways were tested: resonance reinforcing existing views and persuasion introducing new ones. Metrics tracked attitude extremity, emotional intensity, and stated behavioral intent. Resonance consistently outperformed persuasion on all measures.

Data showed resonance effects propagated to semantically related beliefs, indicating structured internal representations. Tactics such as sycophancy and unverified claims varied in impact but lacked uniform strength across metrics. No single tactic dominated every outcome dimension.

Operationally this implies personalized agents and multi-agent networks can amplify shared priors without external oversight. Belief graphs in deployed systems will require consistency audits before scaling. Current alignment methods focused on single-model safety do not address inter-agent influence loops.