Sampling from multi-modal and high-dimensional target distributions is a challenging task that is often required in order to perform Bayesian inference. Parallel tempering (PT) methods address this problem by constructing a Markov chain on an expanded state space that simultaneously samples from a sequence of distributions lying on an annealing path from the prior to the target. In this work we consider generalized annealing paths that start from a variational reference. The reference distribution is tuned to minimize an appropriate notion of distance to the target distribution, maximizing a quantity related to the effective sample size. We apply the method to several posterior inference problems, finding that PT with a variational reference can greatly improve performance. The proposed methodology is particularly useful in cases where the prior and posterior are almost mutually singular and the geometry of the posterior is complex.