A detailed description of the algorithm implemented in BLIP can be found in the paper Learning Bayesian networks with thousands of variables.