In this paper we survey the methods for control and cre- ative interaction with pre-trained generative models for au- dio and music. By using reduced (lossy) encoding and sym- bolization steps we are able to examine the level of informa- tion that is passing between the environment (the musician) and the agent (machine improvisation). We further use the concept of music information dynamics to find an optimal symbolization in terms of predictive information measure. Methods and strategies for generative models are surveyed in this paper and their implications for creative interaction with the machine are discussed in the musical improvisation framework.