Topology-Dependent Correction Failure in Multi-Agent Networks
Scale-free communication topologies fail to self-correct after misinformation injection; hubs anchor incorrect beliefs and bottleneck downstream correction.
Multi-agent information diffusion across communication topologies (scale-free, chain, ring, random, full), with a seed of incorrect information injected into the network and a correction signal introduced subsequently. Qwen-2.5-32B-Instruct with persona conditioning, structured agent-to-agent communication, and judge-based coding of belief states across rounds. Scale-free topologies fail to self-correct while other topologies do — hubs anchor the incorrect belief and bottleneck the correction signal. A secondary effect on persona richness shows a vulnerability-then-resilience trajectory: rich personas are more susceptible early but more robust later. The framing is structural: hub-dominated communication is not just inefficient for error correction but actively obstructive.