We study a network formation interaction in which the formatted network serves the agents to form beliefs about an unknown state via the DeGroot naive learning process. Consequently, the value of the network captures the ability of the network to aggregate information. We focus on the notion of \textit{stable networks} in which agents do not want to create or erase any of their links. We precisely characterize the costs for which a stable network exists. Whenever a stable network exists, either the cycle network, the line network, or the empty network is stable. We further provide structural properties for every stable network: in every stable network, the degree of any two vertices differs by at most one. As a corollary we get that any stable network aggregates information well and a price of anarchy upper bound is provided. Advisor: Prof. Udi Yariv