We are working on the techniques which iteratively learn the formal models from black box implementations by testing. The novelty of the approach addressed here is our processing of the long counterexamples. There is a possibility that the counterexamples generated by a counterexample generator include needless sub sequences. We address the techniques which are developed to avoid the impact of such unwanted sequences on the learning process. The gain of the proposed algorithm is confirmed by considering a comprehensive set of experiments on the finite sate machines.