An original approach to multi-objective optimization is introduced, using a message-passing algorithm to sample the Pareto set, i.e. the set of Pareto-non-dominated solutions. Several heuristics are proposed and tested on a simple bi-objective 3-SAT problem. The first one is based on a straightforward deformation of the Survey-Propagation (SP) equation to locally encode a Pareto trade-off. A simple heuristic is then tested, which combines an elimination procedure of clauses with the usual decimation of variables used in the SP algorithm, and is able to sample different regions of the Pareto-front. We study in more details the compliance of these deformed equations with basic Belief-Propagation (BP) properties. This first leads to an explicit Markov Random Field (MRF) of valid warning configuration, for which the SP equations are basic BP equations. This observation is then generalized to the multi-objective context. Numerical experiments on artificial problems up to 100000 variables are presented and discussed.