This paper introduces a controlled testbed to measure how goal-directed persuaders shift stances within LLM agent networks based on real-world ego-network topologies. Testing across four model backbones and five graph structures reveals that persuasion dynamics are driven by the interplay of topology, competition, topic, and model priors. The study confirms that direct exposure is the primary predictor of stance change, while peer relays provide smaller but measurable influence.
- Persuasion success depends on the mix of network topology, competition, topic, and model prior.
- Direct exposure to persuaders reliably predicts stance changes in subsequent rounds.
- Peer relays exert smaller but statistically measurable influence on agent stances.
- Agent-to-agent persuasion is quantifiable using ego-network grounded testbeds.
- Results hold across four distinct LLM backbones and five different graph structures.
