New decentralized AI framework enables agents to learn dynamic communication patterns without central coordination

BayesG framework allows networked multi-agent systems to operate under local observability with limited communication

Uses Bayesian variational inference to learn sparse, context-aware interaction structures dynamically rather than assuming static neighborhoods

Each agent samples latent communication masks over its local ego-graph to guide message passing and policy decisions independently

Addresses key limitation of existing centralized approaches by eliminating need for global state access and centralized infrastructure