We reformulate the task vector coefficient prediction as a variational inference problem and propose a novel composition framework.