Test-Time Collective Classification over Multi-Agent Networks

Motivated by the challenges of joint training in heterogeneous multi-agent systems and by the benefits of collective decision-making observed in the social sciences, we propose a framework for test-time collaboration among independently trained agents. We study a distributed binary classification task in which agents, potentially differing in architecture, feature space, and modality, must coordinate to produce collective predictions. This coordination is achieved at test time by exchanging local beliefs through a decentralized decision-making protocol. We analyze the generalization performance of this collective classification framework and establish theoretical error bounds. The results quantify the cost of independent training, demonstrate the benefits of collective action, and reveal how network structure and aggregation rules shape classification accuracy.