A new decentralized federated learning method has been proposed for heterogeneous multi-task semantic communication networks. The approach aims to prevent negative transfer and overconsensus bias by separating task-specific features from shared representations at the node level and calibrating a consensus matrix across the network. This is achieved through a policy-driven routing mechanism and a 'communication-while-aggregation' protocol. The method has been evaluated on several datasets, including NYU-v2, Taskonomy, and imperfect wireless links, showing improved performance over existing methods. --- Why it matters: This research matters to engineers working in AI because it addresses the challenge of decentralized federated learning in heterogeneous multi-task environments, which is a critical issue for large-scale semantic communication networks. The proposed method can improve the accuracy and efficiency of these networks by reducing negative transfer and overconsensus bias. Source: https://arxiv.org/abs/2608.15256